Loading...
1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * Read-Copy Update mechanism for mutual exclusion
4 *
5 * Copyright IBM Corporation, 2008
6 *
7 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
8 * Manfred Spraul <manfred@colorfullife.com>
9 * Paul E. McKenney <paulmck@linux.ibm.com> Hierarchical version
10 *
11 * Based on the original work by Paul McKenney <paulmck@linux.ibm.com>
12 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
13 *
14 * For detailed explanation of Read-Copy Update mechanism see -
15 * Documentation/RCU
16 */
17
18#define pr_fmt(fmt) "rcu: " fmt
19
20#include <linux/types.h>
21#include <linux/kernel.h>
22#include <linux/init.h>
23#include <linux/spinlock.h>
24#include <linux/smp.h>
25#include <linux/rcupdate_wait.h>
26#include <linux/interrupt.h>
27#include <linux/sched.h>
28#include <linux/sched/debug.h>
29#include <linux/nmi.h>
30#include <linux/atomic.h>
31#include <linux/bitops.h>
32#include <linux/export.h>
33#include <linux/completion.h>
34#include <linux/moduleparam.h>
35#include <linux/percpu.h>
36#include <linux/notifier.h>
37#include <linux/cpu.h>
38#include <linux/mutex.h>
39#include <linux/time.h>
40#include <linux/kernel_stat.h>
41#include <linux/wait.h>
42#include <linux/kthread.h>
43#include <uapi/linux/sched/types.h>
44#include <linux/prefetch.h>
45#include <linux/delay.h>
46#include <linux/stop_machine.h>
47#include <linux/random.h>
48#include <linux/trace_events.h>
49#include <linux/suspend.h>
50#include <linux/ftrace.h>
51#include <linux/tick.h>
52#include <linux/sysrq.h>
53#include <linux/kprobes.h>
54#include <linux/gfp.h>
55#include <linux/oom.h>
56#include <linux/smpboot.h>
57#include <linux/jiffies.h>
58#include <linux/sched/isolation.h>
59#include <linux/sched/clock.h>
60#include "../time/tick-internal.h"
61
62#include "tree.h"
63#include "rcu.h"
64
65#ifdef MODULE_PARAM_PREFIX
66#undef MODULE_PARAM_PREFIX
67#endif
68#define MODULE_PARAM_PREFIX "rcutree."
69
70/* Data structures. */
71
72/*
73 * Steal a bit from the bottom of ->dynticks for idle entry/exit
74 * control. Initially this is for TLB flushing.
75 */
76#define RCU_DYNTICK_CTRL_MASK 0x1
77#define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
78#ifndef rcu_eqs_special_exit
79#define rcu_eqs_special_exit() do { } while (0)
80#endif
81
82static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = {
83 .dynticks_nesting = 1,
84 .dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE,
85 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
86};
87struct rcu_state rcu_state = {
88 .level = { &rcu_state.node[0] },
89 .gp_state = RCU_GP_IDLE,
90 .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT,
91 .barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex),
92 .name = RCU_NAME,
93 .abbr = RCU_ABBR,
94 .exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
95 .exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
96 .ofl_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.ofl_lock),
97};
98
99/* Dump rcu_node combining tree at boot to verify correct setup. */
100static bool dump_tree;
101module_param(dump_tree, bool, 0444);
102/* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */
103static bool use_softirq = 1;
104module_param(use_softirq, bool, 0444);
105/* Control rcu_node-tree auto-balancing at boot time. */
106static bool rcu_fanout_exact;
107module_param(rcu_fanout_exact, bool, 0444);
108/* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
109static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
110module_param(rcu_fanout_leaf, int, 0444);
111int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
112/* Number of rcu_nodes at specified level. */
113int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
114int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
115
116/*
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.
127 */
128int rcu_scheduler_active __read_mostly;
129EXPORT_SYMBOL_GPL(rcu_scheduler_active);
130
131/*
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.
138 *
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
141 * a time.
142 */
143static int rcu_scheduler_fully_active __read_mostly;
144
145static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
146 unsigned long gps, unsigned long flags);
147static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
148static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
149static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
150static void invoke_rcu_core(void);
151static void rcu_report_exp_rdp(struct rcu_data *rdp);
152static void sync_sched_exp_online_cleanup(int cpu);
153
154/* rcuc/rcub kthread realtime priority */
155static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
156module_param(kthread_prio, int, 0444);
157
158/* Delay in jiffies for grace-period initialization delays, debug only. */
159
160static int gp_preinit_delay;
161module_param(gp_preinit_delay, int, 0444);
162static int gp_init_delay;
163module_param(gp_init_delay, int, 0444);
164static int gp_cleanup_delay;
165module_param(gp_cleanup_delay, int, 0444);
166
167/* Retrieve RCU kthreads priority for rcutorture */
168int rcu_get_gp_kthreads_prio(void)
169{
170 return kthread_prio;
171}
172EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio);
173
174/*
175 * Number of grace periods between delays, normalized by the duration of
176 * the delay. The longer the delay, the more the grace periods between
177 * each delay. The reason for this normalization is that it means that,
178 * for non-zero delays, the overall slowdown of grace periods is constant
179 * regardless of the duration of the delay. This arrangement balances
180 * the need for long delays to increase some race probabilities with the
181 * need for fast grace periods to increase other race probabilities.
182 */
183#define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
184
185/*
186 * Compute the mask of online CPUs for the specified rcu_node structure.
187 * This will not be stable unless the rcu_node structure's ->lock is
188 * held, but the bit corresponding to the current CPU will be stable
189 * in most contexts.
190 */
191unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
192{
193 return READ_ONCE(rnp->qsmaskinitnext);
194}
195
196/*
197 * Return true if an RCU grace period is in progress. The READ_ONCE()s
198 * permit this function to be invoked without holding the root rcu_node
199 * structure's ->lock, but of course results can be subject to change.
200 */
201static int rcu_gp_in_progress(void)
202{
203 return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq));
204}
205
206/*
207 * Return the number of callbacks queued on the specified CPU.
208 * Handles both the nocbs and normal cases.
209 */
210static long rcu_get_n_cbs_cpu(int cpu)
211{
212 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
213
214 if (rcu_segcblist_is_enabled(&rdp->cblist))
215 return rcu_segcblist_n_cbs(&rdp->cblist);
216 return 0;
217}
218
219void rcu_softirq_qs(void)
220{
221 rcu_qs();
222 rcu_preempt_deferred_qs(current);
223}
224
225/*
226 * Record entry into an extended quiescent state. This is only to be
227 * called when not already in an extended quiescent state.
228 */
229static void rcu_dynticks_eqs_enter(void)
230{
231 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
232 int seq;
233
234 /*
235 * CPUs seeing atomic_add_return() must see prior RCU read-side
236 * critical sections, and we also must force ordering with the
237 * next idle sojourn.
238 */
239 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
240 /* Better be in an extended quiescent state! */
241 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
242 (seq & RCU_DYNTICK_CTRL_CTR));
243 /* Better not have special action (TLB flush) pending! */
244 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
245 (seq & RCU_DYNTICK_CTRL_MASK));
246}
247
248/*
249 * Record exit from an extended quiescent state. This is only to be
250 * called from an extended quiescent state.
251 */
252static void rcu_dynticks_eqs_exit(void)
253{
254 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
255 int seq;
256
257 /*
258 * CPUs seeing atomic_add_return() must see prior idle sojourns,
259 * and we also must force ordering with the next RCU read-side
260 * critical section.
261 */
262 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
263 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
264 !(seq & RCU_DYNTICK_CTRL_CTR));
265 if (seq & RCU_DYNTICK_CTRL_MASK) {
266 atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdp->dynticks);
267 smp_mb__after_atomic(); /* _exit after clearing mask. */
268 /* Prefer duplicate flushes to losing a flush. */
269 rcu_eqs_special_exit();
270 }
271}
272
273/*
274 * Reset the current CPU's ->dynticks counter to indicate that the
275 * newly onlined CPU is no longer in an extended quiescent state.
276 * This will either leave the counter unchanged, or increment it
277 * to the next non-quiescent value.
278 *
279 * The non-atomic test/increment sequence works because the upper bits
280 * of the ->dynticks counter are manipulated only by the corresponding CPU,
281 * or when the corresponding CPU is offline.
282 */
283static void rcu_dynticks_eqs_online(void)
284{
285 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
286
287 if (atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR)
288 return;
289 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
290}
291
292/*
293 * Is the current CPU in an extended quiescent state?
294 *
295 * No ordering, as we are sampling CPU-local information.
296 */
297bool rcu_dynticks_curr_cpu_in_eqs(void)
298{
299 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
300
301 return !(atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR);
302}
303
304/*
305 * Snapshot the ->dynticks counter with full ordering so as to allow
306 * stable comparison of this counter with past and future snapshots.
307 */
308int rcu_dynticks_snap(struct rcu_data *rdp)
309{
310 int snap = atomic_add_return(0, &rdp->dynticks);
311
312 return snap & ~RCU_DYNTICK_CTRL_MASK;
313}
314
315/*
316 * Return true if the snapshot returned from rcu_dynticks_snap()
317 * indicates that RCU is in an extended quiescent state.
318 */
319static bool rcu_dynticks_in_eqs(int snap)
320{
321 return !(snap & RCU_DYNTICK_CTRL_CTR);
322}
323
324/*
325 * Return true if the CPU corresponding to the specified rcu_data
326 * structure has spent some time in an extended quiescent state since
327 * rcu_dynticks_snap() returned the specified snapshot.
328 */
329static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap)
330{
331 return snap != rcu_dynticks_snap(rdp);
332}
333
334/*
335 * Set the special (bottom) bit of the specified CPU so that it
336 * will take special action (such as flushing its TLB) on the
337 * next exit from an extended quiescent state. Returns true if
338 * the bit was successfully set, or false if the CPU was not in
339 * an extended quiescent state.
340 */
341bool rcu_eqs_special_set(int cpu)
342{
343 int old;
344 int new;
345 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
346
347 do {
348 old = atomic_read(&rdp->dynticks);
349 if (old & RCU_DYNTICK_CTRL_CTR)
350 return false;
351 new = old | RCU_DYNTICK_CTRL_MASK;
352 } while (atomic_cmpxchg(&rdp->dynticks, old, new) != old);
353 return true;
354}
355
356/*
357 * Let the RCU core know that this CPU has gone through the scheduler,
358 * which is a quiescent state. This is called when the need for a
359 * quiescent state is urgent, so we burn an atomic operation and full
360 * memory barriers to let the RCU core know about it, regardless of what
361 * this CPU might (or might not) do in the near future.
362 *
363 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
364 *
365 * The caller must have disabled interrupts and must not be idle.
366 */
367static void __maybe_unused rcu_momentary_dyntick_idle(void)
368{
369 int special;
370
371 raw_cpu_write(rcu_data.rcu_need_heavy_qs, false);
372 special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
373 &this_cpu_ptr(&rcu_data)->dynticks);
374 /* It is illegal to call this from idle state. */
375 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
376 rcu_preempt_deferred_qs(current);
377}
378
379/**
380 * rcu_is_cpu_rrupt_from_idle - see if interrupted from idle
381 *
382 * If the current CPU is idle and running at a first-level (not nested)
383 * interrupt from idle, return true. The caller must have at least
384 * disabled preemption.
385 */
386static int rcu_is_cpu_rrupt_from_idle(void)
387{
388 /* Called only from within the scheduling-clock interrupt */
389 lockdep_assert_in_irq();
390
391 /* Check for counter underflows */
392 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) < 0,
393 "RCU dynticks_nesting counter underflow!");
394 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) <= 0,
395 "RCU dynticks_nmi_nesting counter underflow/zero!");
396
397 /* Are we at first interrupt nesting level? */
398 if (__this_cpu_read(rcu_data.dynticks_nmi_nesting) != 1)
399 return false;
400
401 /* Does CPU appear to be idle from an RCU standpoint? */
402 return __this_cpu_read(rcu_data.dynticks_nesting) == 0;
403}
404
405#define DEFAULT_RCU_BLIMIT 10 /* Maximum callbacks per rcu_do_batch ... */
406#define DEFAULT_MAX_RCU_BLIMIT 10000 /* ... even during callback flood. */
407static long blimit = DEFAULT_RCU_BLIMIT;
408#define DEFAULT_RCU_QHIMARK 10000 /* If this many pending, ignore blimit. */
409static long qhimark = DEFAULT_RCU_QHIMARK;
410#define DEFAULT_RCU_QLOMARK 100 /* Once only this many pending, use blimit. */
411static long qlowmark = DEFAULT_RCU_QLOMARK;
412
413module_param(blimit, long, 0444);
414module_param(qhimark, long, 0444);
415module_param(qlowmark, long, 0444);
416
417static ulong jiffies_till_first_fqs = ULONG_MAX;
418static ulong jiffies_till_next_fqs = ULONG_MAX;
419static bool rcu_kick_kthreads;
420static int rcu_divisor = 7;
421module_param(rcu_divisor, int, 0644);
422
423/* Force an exit from rcu_do_batch() after 3 milliseconds. */
424static long rcu_resched_ns = 3 * NSEC_PER_MSEC;
425module_param(rcu_resched_ns, long, 0644);
426
427/*
428 * How long the grace period must be before we start recruiting
429 * quiescent-state help from rcu_note_context_switch().
430 */
431static ulong jiffies_till_sched_qs = ULONG_MAX;
432module_param(jiffies_till_sched_qs, ulong, 0444);
433static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */
434module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
435
436/*
437 * Make sure that we give the grace-period kthread time to detect any
438 * idle CPUs before taking active measures to force quiescent states.
439 * However, don't go below 100 milliseconds, adjusted upwards for really
440 * large systems.
441 */
442static void adjust_jiffies_till_sched_qs(void)
443{
444 unsigned long j;
445
446 /* If jiffies_till_sched_qs was specified, respect the request. */
447 if (jiffies_till_sched_qs != ULONG_MAX) {
448 WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
449 return;
450 }
451 /* Otherwise, set to third fqs scan, but bound below on large system. */
452 j = READ_ONCE(jiffies_till_first_fqs) +
453 2 * READ_ONCE(jiffies_till_next_fqs);
454 if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
455 j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
456 pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
457 WRITE_ONCE(jiffies_to_sched_qs, j);
458}
459
460static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
461{
462 ulong j;
463 int ret = kstrtoul(val, 0, &j);
464
465 if (!ret) {
466 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
467 adjust_jiffies_till_sched_qs();
468 }
469 return ret;
470}
471
472static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
473{
474 ulong j;
475 int ret = kstrtoul(val, 0, &j);
476
477 if (!ret) {
478 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
479 adjust_jiffies_till_sched_qs();
480 }
481 return ret;
482}
483
484static struct kernel_param_ops first_fqs_jiffies_ops = {
485 .set = param_set_first_fqs_jiffies,
486 .get = param_get_ulong,
487};
488
489static struct kernel_param_ops next_fqs_jiffies_ops = {
490 .set = param_set_next_fqs_jiffies,
491 .get = param_get_ulong,
492};
493
494module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
495module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
496module_param(rcu_kick_kthreads, bool, 0644);
497
498static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
499static int rcu_pending(void);
500
501/*
502 * Return the number of RCU GPs completed thus far for debug & stats.
503 */
504unsigned long rcu_get_gp_seq(void)
505{
506 return READ_ONCE(rcu_state.gp_seq);
507}
508EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
509
510/*
511 * Return the number of RCU expedited batches completed thus far for
512 * debug & stats. Odd numbers mean that a batch is in progress, even
513 * numbers mean idle. The value returned will thus be roughly double
514 * the cumulative batches since boot.
515 */
516unsigned long rcu_exp_batches_completed(void)
517{
518 return rcu_state.expedited_sequence;
519}
520EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
521
522/*
523 * Return the root node of the rcu_state structure.
524 */
525static struct rcu_node *rcu_get_root(void)
526{
527 return &rcu_state.node[0];
528}
529
530/*
531 * Convert a ->gp_state value to a character string.
532 */
533static const char *gp_state_getname(short gs)
534{
535 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
536 return "???";
537 return gp_state_names[gs];
538}
539
540/*
541 * Send along grace-period-related data for rcutorture diagnostics.
542 */
543void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
544 unsigned long *gp_seq)
545{
546 switch (test_type) {
547 case RCU_FLAVOR:
548 *flags = READ_ONCE(rcu_state.gp_flags);
549 *gp_seq = rcu_seq_current(&rcu_state.gp_seq);
550 break;
551 default:
552 break;
553 }
554}
555EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
556
557/*
558 * Enter an RCU extended quiescent state, which can be either the
559 * idle loop or adaptive-tickless usermode execution.
560 *
561 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
562 * the possibility of usermode upcalls having messed up our count
563 * of interrupt nesting level during the prior busy period.
564 */
565static void rcu_eqs_enter(bool user)
566{
567 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
568
569 WARN_ON_ONCE(rdp->dynticks_nmi_nesting != DYNTICK_IRQ_NONIDLE);
570 WRITE_ONCE(rdp->dynticks_nmi_nesting, 0);
571 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
572 rdp->dynticks_nesting == 0);
573 if (rdp->dynticks_nesting != 1) {
574 rdp->dynticks_nesting--;
575 return;
576 }
577
578 lockdep_assert_irqs_disabled();
579 trace_rcu_dyntick(TPS("Start"), rdp->dynticks_nesting, 0, rdp->dynticks);
580 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
581 rdp = this_cpu_ptr(&rcu_data);
582 do_nocb_deferred_wakeup(rdp);
583 rcu_prepare_for_idle();
584 rcu_preempt_deferred_qs(current);
585 WRITE_ONCE(rdp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
586 rcu_dynticks_eqs_enter();
587 rcu_dynticks_task_enter();
588}
589
590/**
591 * rcu_idle_enter - inform RCU that current CPU is entering idle
592 *
593 * Enter idle mode, in other words, -leave- the mode in which RCU
594 * read-side critical sections can occur. (Though RCU read-side
595 * critical sections can occur in irq handlers in idle, a possibility
596 * handled by irq_enter() and irq_exit().)
597 *
598 * If you add or remove a call to rcu_idle_enter(), be sure to test with
599 * CONFIG_RCU_EQS_DEBUG=y.
600 */
601void rcu_idle_enter(void)
602{
603 lockdep_assert_irqs_disabled();
604 rcu_eqs_enter(false);
605}
606
607#ifdef CONFIG_NO_HZ_FULL
608/**
609 * rcu_user_enter - inform RCU that we are resuming userspace.
610 *
611 * Enter RCU idle mode right before resuming userspace. No use of RCU
612 * is permitted between this call and rcu_user_exit(). This way the
613 * CPU doesn't need to maintain the tick for RCU maintenance purposes
614 * when the CPU runs in userspace.
615 *
616 * If you add or remove a call to rcu_user_enter(), be sure to test with
617 * CONFIG_RCU_EQS_DEBUG=y.
618 */
619void rcu_user_enter(void)
620{
621 lockdep_assert_irqs_disabled();
622 rcu_eqs_enter(true);
623}
624#endif /* CONFIG_NO_HZ_FULL */
625
626/*
627 * If we are returning from the outermost NMI handler that interrupted an
628 * RCU-idle period, update rdp->dynticks and rdp->dynticks_nmi_nesting
629 * to let the RCU grace-period handling know that the CPU is back to
630 * being RCU-idle.
631 *
632 * If you add or remove a call to rcu_nmi_exit_common(), be sure to test
633 * with CONFIG_RCU_EQS_DEBUG=y.
634 */
635static __always_inline void rcu_nmi_exit_common(bool irq)
636{
637 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
638
639 /*
640 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
641 * (We are exiting an NMI handler, so RCU better be paying attention
642 * to us!)
643 */
644 WARN_ON_ONCE(rdp->dynticks_nmi_nesting <= 0);
645 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
646
647 /*
648 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
649 * leave it in non-RCU-idle state.
650 */
651 if (rdp->dynticks_nmi_nesting != 1) {
652 trace_rcu_dyntick(TPS("--="), rdp->dynticks_nmi_nesting, rdp->dynticks_nmi_nesting - 2, rdp->dynticks);
653 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* No store tearing. */
654 rdp->dynticks_nmi_nesting - 2);
655 return;
656 }
657
658 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
659 trace_rcu_dyntick(TPS("Startirq"), rdp->dynticks_nmi_nesting, 0, rdp->dynticks);
660 WRITE_ONCE(rdp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
661
662 if (irq)
663 rcu_prepare_for_idle();
664
665 rcu_dynticks_eqs_enter();
666
667 if (irq)
668 rcu_dynticks_task_enter();
669}
670
671/**
672 * rcu_nmi_exit - inform RCU of exit from NMI context
673 *
674 * If you add or remove a call to rcu_nmi_exit(), be sure to test
675 * with CONFIG_RCU_EQS_DEBUG=y.
676 */
677void rcu_nmi_exit(void)
678{
679 rcu_nmi_exit_common(false);
680}
681
682/**
683 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
684 *
685 * Exit from an interrupt handler, which might possibly result in entering
686 * idle mode, in other words, leaving the mode in which read-side critical
687 * sections can occur. The caller must have disabled interrupts.
688 *
689 * This code assumes that the idle loop never does anything that might
690 * result in unbalanced calls to irq_enter() and irq_exit(). If your
691 * architecture's idle loop violates this assumption, RCU will give you what
692 * you deserve, good and hard. But very infrequently and irreproducibly.
693 *
694 * Use things like work queues to work around this limitation.
695 *
696 * You have been warned.
697 *
698 * If you add or remove a call to rcu_irq_exit(), be sure to test with
699 * CONFIG_RCU_EQS_DEBUG=y.
700 */
701void rcu_irq_exit(void)
702{
703 lockdep_assert_irqs_disabled();
704 rcu_nmi_exit_common(true);
705}
706
707/*
708 * Wrapper for rcu_irq_exit() where interrupts are enabled.
709 *
710 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
711 * with CONFIG_RCU_EQS_DEBUG=y.
712 */
713void rcu_irq_exit_irqson(void)
714{
715 unsigned long flags;
716
717 local_irq_save(flags);
718 rcu_irq_exit();
719 local_irq_restore(flags);
720}
721
722/*
723 * Exit an RCU extended quiescent state, which can be either the
724 * idle loop or adaptive-tickless usermode execution.
725 *
726 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
727 * allow for the possibility of usermode upcalls messing up our count of
728 * interrupt nesting level during the busy period that is just now starting.
729 */
730static void rcu_eqs_exit(bool user)
731{
732 struct rcu_data *rdp;
733 long oldval;
734
735 lockdep_assert_irqs_disabled();
736 rdp = this_cpu_ptr(&rcu_data);
737 oldval = rdp->dynticks_nesting;
738 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
739 if (oldval) {
740 rdp->dynticks_nesting++;
741 return;
742 }
743 rcu_dynticks_task_exit();
744 rcu_dynticks_eqs_exit();
745 rcu_cleanup_after_idle();
746 trace_rcu_dyntick(TPS("End"), rdp->dynticks_nesting, 1, rdp->dynticks);
747 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
748 WRITE_ONCE(rdp->dynticks_nesting, 1);
749 WARN_ON_ONCE(rdp->dynticks_nmi_nesting);
750 WRITE_ONCE(rdp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
751}
752
753/**
754 * rcu_idle_exit - inform RCU that current CPU is leaving idle
755 *
756 * Exit idle mode, in other words, -enter- the mode in which RCU
757 * read-side critical sections can occur.
758 *
759 * If you add or remove a call to rcu_idle_exit(), be sure to test with
760 * CONFIG_RCU_EQS_DEBUG=y.
761 */
762void rcu_idle_exit(void)
763{
764 unsigned long flags;
765
766 local_irq_save(flags);
767 rcu_eqs_exit(false);
768 local_irq_restore(flags);
769}
770
771#ifdef CONFIG_NO_HZ_FULL
772/**
773 * rcu_user_exit - inform RCU that we are exiting userspace.
774 *
775 * Exit RCU idle mode while entering the kernel because it can
776 * run a RCU read side critical section anytime.
777 *
778 * If you add or remove a call to rcu_user_exit(), be sure to test with
779 * CONFIG_RCU_EQS_DEBUG=y.
780 */
781void rcu_user_exit(void)
782{
783 rcu_eqs_exit(1);
784}
785#endif /* CONFIG_NO_HZ_FULL */
786
787/**
788 * rcu_nmi_enter_common - inform RCU of entry to NMI context
789 * @irq: Is this call from rcu_irq_enter?
790 *
791 * If the CPU was idle from RCU's viewpoint, update rdp->dynticks and
792 * rdp->dynticks_nmi_nesting to let the RCU grace-period handling know
793 * that the CPU is active. This implementation permits nested NMIs, as
794 * long as the nesting level does not overflow an int. (You will probably
795 * run out of stack space first.)
796 *
797 * If you add or remove a call to rcu_nmi_enter_common(), be sure to test
798 * with CONFIG_RCU_EQS_DEBUG=y.
799 */
800static __always_inline void rcu_nmi_enter_common(bool irq)
801{
802 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
803 long incby = 2;
804
805 /* Complain about underflow. */
806 WARN_ON_ONCE(rdp->dynticks_nmi_nesting < 0);
807
808 /*
809 * If idle from RCU viewpoint, atomically increment ->dynticks
810 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
811 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
812 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
813 * to be in the outermost NMI handler that interrupted an RCU-idle
814 * period (observation due to Andy Lutomirski).
815 */
816 if (rcu_dynticks_curr_cpu_in_eqs()) {
817
818 if (irq)
819 rcu_dynticks_task_exit();
820
821 rcu_dynticks_eqs_exit();
822
823 if (irq)
824 rcu_cleanup_after_idle();
825
826 incby = 1;
827 }
828 trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
829 rdp->dynticks_nmi_nesting,
830 rdp->dynticks_nmi_nesting + incby, rdp->dynticks);
831 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* Prevent store tearing. */
832 rdp->dynticks_nmi_nesting + incby);
833 barrier();
834}
835
836/**
837 * rcu_nmi_enter - inform RCU of entry to NMI context
838 */
839void rcu_nmi_enter(void)
840{
841 rcu_nmi_enter_common(false);
842}
843NOKPROBE_SYMBOL(rcu_nmi_enter);
844
845/**
846 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
847 *
848 * Enter an interrupt handler, which might possibly result in exiting
849 * idle mode, in other words, entering the mode in which read-side critical
850 * sections can occur. The caller must have disabled interrupts.
851 *
852 * Note that the Linux kernel is fully capable of entering an interrupt
853 * handler that it never exits, for example when doing upcalls to user mode!
854 * This code assumes that the idle loop never does upcalls to user mode.
855 * If your architecture's idle loop does do upcalls to user mode (or does
856 * anything else that results in unbalanced calls to the irq_enter() and
857 * irq_exit() functions), RCU will give you what you deserve, good and hard.
858 * But very infrequently and irreproducibly.
859 *
860 * Use things like work queues to work around this limitation.
861 *
862 * You have been warned.
863 *
864 * If you add or remove a call to rcu_irq_enter(), be sure to test with
865 * CONFIG_RCU_EQS_DEBUG=y.
866 */
867void rcu_irq_enter(void)
868{
869 lockdep_assert_irqs_disabled();
870 rcu_nmi_enter_common(true);
871}
872
873/*
874 * Wrapper for rcu_irq_enter() where interrupts are enabled.
875 *
876 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
877 * with CONFIG_RCU_EQS_DEBUG=y.
878 */
879void rcu_irq_enter_irqson(void)
880{
881 unsigned long flags;
882
883 local_irq_save(flags);
884 rcu_irq_enter();
885 local_irq_restore(flags);
886}
887
888/**
889 * rcu_is_watching - see if RCU thinks that the current CPU is not idle
890 *
891 * Return true if RCU is watching the running CPU, which means that this
892 * CPU can safely enter RCU read-side critical sections. In other words,
893 * if the current CPU is not in its idle loop or is in an interrupt or
894 * NMI handler, return true.
895 */
896bool notrace rcu_is_watching(void)
897{
898 bool ret;
899
900 preempt_disable_notrace();
901 ret = !rcu_dynticks_curr_cpu_in_eqs();
902 preempt_enable_notrace();
903 return ret;
904}
905EXPORT_SYMBOL_GPL(rcu_is_watching);
906
907/*
908 * If a holdout task is actually running, request an urgent quiescent
909 * state from its CPU. This is unsynchronized, so migrations can cause
910 * the request to go to the wrong CPU. Which is OK, all that will happen
911 * is that the CPU's next context switch will be a bit slower and next
912 * time around this task will generate another request.
913 */
914void rcu_request_urgent_qs_task(struct task_struct *t)
915{
916 int cpu;
917
918 barrier();
919 cpu = task_cpu(t);
920 if (!task_curr(t))
921 return; /* This task is not running on that CPU. */
922 smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
923}
924
925#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
926
927/*
928 * Is the current CPU online as far as RCU is concerned?
929 *
930 * Disable preemption to avoid false positives that could otherwise
931 * happen due to the current CPU number being sampled, this task being
932 * preempted, its old CPU being taken offline, resuming on some other CPU,
933 * then determining that its old CPU is now offline.
934 *
935 * Disable checking if in an NMI handler because we cannot safely
936 * report errors from NMI handlers anyway. In addition, it is OK to use
937 * RCU on an offline processor during initial boot, hence the check for
938 * rcu_scheduler_fully_active.
939 */
940bool rcu_lockdep_current_cpu_online(void)
941{
942 struct rcu_data *rdp;
943 struct rcu_node *rnp;
944 bool ret = false;
945
946 if (in_nmi() || !rcu_scheduler_fully_active)
947 return true;
948 preempt_disable();
949 rdp = this_cpu_ptr(&rcu_data);
950 rnp = rdp->mynode;
951 if (rdp->grpmask & rcu_rnp_online_cpus(rnp))
952 ret = true;
953 preempt_enable();
954 return ret;
955}
956EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
957
958#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
959
960/*
961 * We are reporting a quiescent state on behalf of some other CPU, so
962 * it is our responsibility to check for and handle potential overflow
963 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
964 * After all, the CPU might be in deep idle state, and thus executing no
965 * code whatsoever.
966 */
967static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
968{
969 raw_lockdep_assert_held_rcu_node(rnp);
970 if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
971 rnp->gp_seq))
972 WRITE_ONCE(rdp->gpwrap, true);
973 if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
974 rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
975}
976
977/*
978 * Snapshot the specified CPU's dynticks counter so that we can later
979 * credit them with an implicit quiescent state. Return 1 if this CPU
980 * is in dynticks idle mode, which is an extended quiescent state.
981 */
982static int dyntick_save_progress_counter(struct rcu_data *rdp)
983{
984 rdp->dynticks_snap = rcu_dynticks_snap(rdp);
985 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
986 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
987 rcu_gpnum_ovf(rdp->mynode, rdp);
988 return 1;
989 }
990 return 0;
991}
992
993/*
994 * Return true if the specified CPU has passed through a quiescent
995 * state by virtue of being in or having passed through an dynticks
996 * idle state since the last call to dyntick_save_progress_counter()
997 * for this same CPU, or by virtue of having been offline.
998 */
999static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1000{
1001 unsigned long jtsq;
1002 bool *rnhqp;
1003 bool *ruqp;
1004 struct rcu_node *rnp = rdp->mynode;
1005
1006 /*
1007 * If the CPU passed through or entered a dynticks idle phase with
1008 * no active irq/NMI handlers, then we can safely pretend that the CPU
1009 * already acknowledged the request to pass through a quiescent
1010 * state. Either way, that CPU cannot possibly be in an RCU
1011 * read-side critical section that started before the beginning
1012 * of the current RCU grace period.
1013 */
1014 if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
1015 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1016 rcu_gpnum_ovf(rnp, rdp);
1017 return 1;
1018 }
1019
1020 /* If waiting too long on an offline CPU, complain. */
1021 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp)) &&
1022 time_after(jiffies, rcu_state.gp_start + HZ)) {
1023 bool onl;
1024 struct rcu_node *rnp1;
1025
1026 WARN_ON(1); /* Offline CPUs are supposed to report QS! */
1027 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
1028 __func__, rnp->grplo, rnp->grphi, rnp->level,
1029 (long)rnp->gp_seq, (long)rnp->completedqs);
1030 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
1031 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
1032 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
1033 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
1034 pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
1035 __func__, rdp->cpu, ".o"[onl],
1036 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
1037 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
1038 return 1; /* Break things loose after complaining. */
1039 }
1040
1041 /*
1042 * A CPU running for an extended time within the kernel can
1043 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
1044 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
1045 * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the
1046 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
1047 * variable are safe because the assignments are repeated if this
1048 * CPU failed to pass through a quiescent state. This code
1049 * also checks .jiffies_resched in case jiffies_to_sched_qs
1050 * is set way high.
1051 */
1052 jtsq = READ_ONCE(jiffies_to_sched_qs);
1053 ruqp = per_cpu_ptr(&rcu_data.rcu_urgent_qs, rdp->cpu);
1054 rnhqp = &per_cpu(rcu_data.rcu_need_heavy_qs, rdp->cpu);
1055 if (!READ_ONCE(*rnhqp) &&
1056 (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
1057 time_after(jiffies, rcu_state.jiffies_resched))) {
1058 WRITE_ONCE(*rnhqp, true);
1059 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1060 smp_store_release(ruqp, true);
1061 } else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
1062 WRITE_ONCE(*ruqp, true);
1063 }
1064
1065 /*
1066 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
1067 * The above code handles this, but only for straight cond_resched().
1068 * And some in-kernel loops check need_resched() before calling
1069 * cond_resched(), which defeats the above code for CPUs that are
1070 * running in-kernel with scheduling-clock interrupts disabled.
1071 * So hit them over the head with the resched_cpu() hammer!
1072 */
1073 if (tick_nohz_full_cpu(rdp->cpu) &&
1074 time_after(jiffies,
1075 READ_ONCE(rdp->last_fqs_resched) + jtsq * 3)) {
1076 resched_cpu(rdp->cpu);
1077 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1078 }
1079
1080 /*
1081 * If more than halfway to RCU CPU stall-warning time, invoke
1082 * resched_cpu() more frequently to try to loosen things up a bit.
1083 * Also check to see if the CPU is getting hammered with interrupts,
1084 * but only once per grace period, just to keep the IPIs down to
1085 * a dull roar.
1086 */
1087 if (time_after(jiffies, rcu_state.jiffies_resched)) {
1088 if (time_after(jiffies,
1089 READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
1090 resched_cpu(rdp->cpu);
1091 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1092 }
1093 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1094 !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
1095 (rnp->ffmask & rdp->grpmask)) {
1096 init_irq_work(&rdp->rcu_iw, rcu_iw_handler);
1097 rdp->rcu_iw_pending = true;
1098 rdp->rcu_iw_gp_seq = rnp->gp_seq;
1099 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1100 }
1101 }
1102
1103 return 0;
1104}
1105
1106/* Trace-event wrapper function for trace_rcu_future_grace_period. */
1107static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1108 unsigned long gp_seq_req, const char *s)
1109{
1110 trace_rcu_future_grace_period(rcu_state.name, rnp->gp_seq, gp_seq_req,
1111 rnp->level, rnp->grplo, rnp->grphi, s);
1112}
1113
1114/*
1115 * rcu_start_this_gp - Request the start of a particular grace period
1116 * @rnp_start: The leaf node of the CPU from which to start.
1117 * @rdp: The rcu_data corresponding to the CPU from which to start.
1118 * @gp_seq_req: The gp_seq of the grace period to start.
1119 *
1120 * Start the specified grace period, as needed to handle newly arrived
1121 * callbacks. The required future grace periods are recorded in each
1122 * rcu_node structure's ->gp_seq_needed field. Returns true if there
1123 * is reason to awaken the grace-period kthread.
1124 *
1125 * The caller must hold the specified rcu_node structure's ->lock, which
1126 * is why the caller is responsible for waking the grace-period kthread.
1127 *
1128 * Returns true if the GP thread needs to be awakened else false.
1129 */
1130static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
1131 unsigned long gp_seq_req)
1132{
1133 bool ret = false;
1134 struct rcu_node *rnp;
1135
1136 /*
1137 * Use funnel locking to either acquire the root rcu_node
1138 * structure's lock or bail out if the need for this grace period
1139 * has already been recorded -- or if that grace period has in
1140 * fact already started. If there is already a grace period in
1141 * progress in a non-leaf node, no recording is needed because the
1142 * end of the grace period will scan the leaf rcu_node structures.
1143 * Note that rnp_start->lock must not be released.
1144 */
1145 raw_lockdep_assert_held_rcu_node(rnp_start);
1146 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
1147 for (rnp = rnp_start; 1; rnp = rnp->parent) {
1148 if (rnp != rnp_start)
1149 raw_spin_lock_rcu_node(rnp);
1150 if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
1151 rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
1152 (rnp != rnp_start &&
1153 rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
1154 trace_rcu_this_gp(rnp, rdp, gp_seq_req,
1155 TPS("Prestarted"));
1156 goto unlock_out;
1157 }
1158 rnp->gp_seq_needed = gp_seq_req;
1159 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
1160 /*
1161 * We just marked the leaf or internal node, and a
1162 * grace period is in progress, which means that
1163 * rcu_gp_cleanup() will see the marking. Bail to
1164 * reduce contention.
1165 */
1166 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
1167 TPS("Startedleaf"));
1168 goto unlock_out;
1169 }
1170 if (rnp != rnp_start && rnp->parent != NULL)
1171 raw_spin_unlock_rcu_node(rnp);
1172 if (!rnp->parent)
1173 break; /* At root, and perhaps also leaf. */
1174 }
1175
1176 /* If GP already in progress, just leave, otherwise start one. */
1177 if (rcu_gp_in_progress()) {
1178 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
1179 goto unlock_out;
1180 }
1181 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
1182 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
1183 rcu_state.gp_req_activity = jiffies;
1184 if (!rcu_state.gp_kthread) {
1185 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
1186 goto unlock_out;
1187 }
1188 trace_rcu_grace_period(rcu_state.name, READ_ONCE(rcu_state.gp_seq), TPS("newreq"));
1189 ret = true; /* Caller must wake GP kthread. */
1190unlock_out:
1191 /* Push furthest requested GP to leaf node and rcu_data structure. */
1192 if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
1193 rnp_start->gp_seq_needed = rnp->gp_seq_needed;
1194 rdp->gp_seq_needed = rnp->gp_seq_needed;
1195 }
1196 if (rnp != rnp_start)
1197 raw_spin_unlock_rcu_node(rnp);
1198 return ret;
1199}
1200
1201/*
1202 * Clean up any old requests for the just-ended grace period. Also return
1203 * whether any additional grace periods have been requested.
1204 */
1205static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
1206{
1207 bool needmore;
1208 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1209
1210 needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1211 if (!needmore)
1212 rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1213 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1214 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1215 return needmore;
1216}
1217
1218/*
1219 * Awaken the grace-period kthread. Don't do a self-awaken (unless in
1220 * an interrupt or softirq handler), and don't bother awakening when there
1221 * is nothing for the grace-period kthread to do (as in several CPUs raced
1222 * to awaken, and we lost), and finally don't try to awaken a kthread that
1223 * has not yet been created. If all those checks are passed, track some
1224 * debug information and awaken.
1225 *
1226 * So why do the self-wakeup when in an interrupt or softirq handler
1227 * in the grace-period kthread's context? Because the kthread might have
1228 * been interrupted just as it was going to sleep, and just after the final
1229 * pre-sleep check of the awaken condition. In this case, a wakeup really
1230 * is required, and is therefore supplied.
1231 */
1232static void rcu_gp_kthread_wake(void)
1233{
1234 if ((current == rcu_state.gp_kthread &&
1235 !in_irq() && !in_serving_softirq()) ||
1236 !READ_ONCE(rcu_state.gp_flags) ||
1237 !rcu_state.gp_kthread)
1238 return;
1239 WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
1240 WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
1241 swake_up_one(&rcu_state.gp_wq);
1242}
1243
1244/*
1245 * If there is room, assign a ->gp_seq number to any callbacks on this
1246 * CPU that have not already been assigned. Also accelerate any callbacks
1247 * that were previously assigned a ->gp_seq number that has since proven
1248 * to be too conservative, which can happen if callbacks get assigned a
1249 * ->gp_seq number while RCU is idle, but with reference to a non-root
1250 * rcu_node structure. This function is idempotent, so it does not hurt
1251 * to call it repeatedly. Returns an flag saying that we should awaken
1252 * the RCU grace-period kthread.
1253 *
1254 * The caller must hold rnp->lock with interrupts disabled.
1255 */
1256static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1257{
1258 unsigned long gp_seq_req;
1259 bool ret = false;
1260
1261 rcu_lockdep_assert_cblist_protected(rdp);
1262 raw_lockdep_assert_held_rcu_node(rnp);
1263
1264 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1265 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1266 return false;
1267
1268 /*
1269 * Callbacks are often registered with incomplete grace-period
1270 * information. Something about the fact that getting exact
1271 * information requires acquiring a global lock... RCU therefore
1272 * makes a conservative estimate of the grace period number at which
1273 * a given callback will become ready to invoke. The following
1274 * code checks this estimate and improves it when possible, thus
1275 * accelerating callback invocation to an earlier grace-period
1276 * number.
1277 */
1278 gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1279 if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1280 ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1281
1282 /* Trace depending on how much we were able to accelerate. */
1283 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1284 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccWaitCB"));
1285 else
1286 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccReadyCB"));
1287 return ret;
1288}
1289
1290/*
1291 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1292 * rcu_node structure's ->lock be held. It consults the cached value
1293 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1294 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1295 * while holding the leaf rcu_node structure's ->lock.
1296 */
1297static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1298 struct rcu_data *rdp)
1299{
1300 unsigned long c;
1301 bool needwake;
1302
1303 rcu_lockdep_assert_cblist_protected(rdp);
1304 c = rcu_seq_snap(&rcu_state.gp_seq);
1305 if (!rdp->gpwrap && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1306 /* Old request still live, so mark recent callbacks. */
1307 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1308 return;
1309 }
1310 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1311 needwake = rcu_accelerate_cbs(rnp, rdp);
1312 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1313 if (needwake)
1314 rcu_gp_kthread_wake();
1315}
1316
1317/*
1318 * Move any callbacks whose grace period has completed to the
1319 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1320 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1321 * sublist. This function is idempotent, so it does not hurt to
1322 * invoke it repeatedly. As long as it is not invoked -too- often...
1323 * Returns true if the RCU grace-period kthread needs to be awakened.
1324 *
1325 * The caller must hold rnp->lock with interrupts disabled.
1326 */
1327static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1328{
1329 rcu_lockdep_assert_cblist_protected(rdp);
1330 raw_lockdep_assert_held_rcu_node(rnp);
1331
1332 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1333 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1334 return false;
1335
1336 /*
1337 * Find all callbacks whose ->gp_seq numbers indicate that they
1338 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1339 */
1340 rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1341
1342 /* Classify any remaining callbacks. */
1343 return rcu_accelerate_cbs(rnp, rdp);
1344}
1345
1346/*
1347 * Move and classify callbacks, but only if doing so won't require
1348 * that the RCU grace-period kthread be awakened.
1349 */
1350static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
1351 struct rcu_data *rdp)
1352{
1353 rcu_lockdep_assert_cblist_protected(rdp);
1354 if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) ||
1355 !raw_spin_trylock_rcu_node(rnp))
1356 return;
1357 WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
1358 raw_spin_unlock_rcu_node(rnp);
1359}
1360
1361/*
1362 * Update CPU-local rcu_data state to record the beginnings and ends of
1363 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1364 * structure corresponding to the current CPU, and must have irqs disabled.
1365 * Returns true if the grace-period kthread needs to be awakened.
1366 */
1367static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
1368{
1369 bool ret = false;
1370 bool need_gp;
1371 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
1372 rcu_segcblist_is_offloaded(&rdp->cblist);
1373
1374 raw_lockdep_assert_held_rcu_node(rnp);
1375
1376 if (rdp->gp_seq == rnp->gp_seq)
1377 return false; /* Nothing to do. */
1378
1379 /* Handle the ends of any preceding grace periods first. */
1380 if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1381 unlikely(READ_ONCE(rdp->gpwrap))) {
1382 if (!offloaded)
1383 ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
1384 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1385 } else {
1386 if (!offloaded)
1387 ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
1388 }
1389
1390 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1391 if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1392 unlikely(READ_ONCE(rdp->gpwrap))) {
1393 /*
1394 * If the current grace period is waiting for this CPU,
1395 * set up to detect a quiescent state, otherwise don't
1396 * go looking for one.
1397 */
1398 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1399 need_gp = !!(rnp->qsmask & rdp->grpmask);
1400 rdp->cpu_no_qs.b.norm = need_gp;
1401 rdp->core_needs_qs = need_gp;
1402 zero_cpu_stall_ticks(rdp);
1403 }
1404 rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
1405 if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
1406 rdp->gp_seq_needed = rnp->gp_seq_needed;
1407 WRITE_ONCE(rdp->gpwrap, false);
1408 rcu_gpnum_ovf(rnp, rdp);
1409 return ret;
1410}
1411
1412static void note_gp_changes(struct rcu_data *rdp)
1413{
1414 unsigned long flags;
1415 bool needwake;
1416 struct rcu_node *rnp;
1417
1418 local_irq_save(flags);
1419 rnp = rdp->mynode;
1420 if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1421 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1422 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1423 local_irq_restore(flags);
1424 return;
1425 }
1426 needwake = __note_gp_changes(rnp, rdp);
1427 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1428 if (needwake)
1429 rcu_gp_kthread_wake();
1430}
1431
1432static void rcu_gp_slow(int delay)
1433{
1434 if (delay > 0 &&
1435 !(rcu_seq_ctr(rcu_state.gp_seq) %
1436 (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1437 schedule_timeout_uninterruptible(delay);
1438}
1439
1440/*
1441 * Initialize a new grace period. Return false if no grace period required.
1442 */
1443static bool rcu_gp_init(void)
1444{
1445 unsigned long flags;
1446 unsigned long oldmask;
1447 unsigned long mask;
1448 struct rcu_data *rdp;
1449 struct rcu_node *rnp = rcu_get_root();
1450
1451 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1452 raw_spin_lock_irq_rcu_node(rnp);
1453 if (!READ_ONCE(rcu_state.gp_flags)) {
1454 /* Spurious wakeup, tell caller to go back to sleep. */
1455 raw_spin_unlock_irq_rcu_node(rnp);
1456 return false;
1457 }
1458 WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1459
1460 if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1461 /*
1462 * Grace period already in progress, don't start another.
1463 * Not supposed to be able to happen.
1464 */
1465 raw_spin_unlock_irq_rcu_node(rnp);
1466 return false;
1467 }
1468
1469 /* Advance to a new grace period and initialize state. */
1470 record_gp_stall_check_time();
1471 /* Record GP times before starting GP, hence rcu_seq_start(). */
1472 rcu_seq_start(&rcu_state.gp_seq);
1473 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
1474 raw_spin_unlock_irq_rcu_node(rnp);
1475
1476 /*
1477 * Apply per-leaf buffered online and offline operations to the
1478 * rcu_node tree. Note that this new grace period need not wait
1479 * for subsequent online CPUs, and that quiescent-state forcing
1480 * will handle subsequent offline CPUs.
1481 */
1482 rcu_state.gp_state = RCU_GP_ONOFF;
1483 rcu_for_each_leaf_node(rnp) {
1484 raw_spin_lock(&rcu_state.ofl_lock);
1485 raw_spin_lock_irq_rcu_node(rnp);
1486 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1487 !rnp->wait_blkd_tasks) {
1488 /* Nothing to do on this leaf rcu_node structure. */
1489 raw_spin_unlock_irq_rcu_node(rnp);
1490 raw_spin_unlock(&rcu_state.ofl_lock);
1491 continue;
1492 }
1493
1494 /* Record old state, apply changes to ->qsmaskinit field. */
1495 oldmask = rnp->qsmaskinit;
1496 rnp->qsmaskinit = rnp->qsmaskinitnext;
1497
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);
1507 }
1508 }
1509
1510 /*
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.
1517 */
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);
1523 }
1524
1525 raw_spin_unlock_irq_rcu_node(rnp);
1526 raw_spin_unlock(&rcu_state.ofl_lock);
1527 }
1528 rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1529
1530 /*
1531 * Set the quiescent-state-needed bits in all the rcu_node
1532 * structures for all currently online CPUs in breadth-first
1533 * order, starting from the root rcu_node structure, relying on the
1534 * layout of the tree within the rcu_state.node[] array. Note that
1535 * other CPUs will access only the leaves of the hierarchy, thus
1536 * seeing that no grace period is in progress, at least until the
1537 * corresponding leaf node has been initialized.
1538 *
1539 * The grace period cannot complete until the initialization
1540 * process finishes, because this kthread handles both.
1541 */
1542 rcu_state.gp_state = RCU_GP_INIT;
1543 rcu_for_each_node_breadth_first(rnp) {
1544 rcu_gp_slow(gp_init_delay);
1545 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1546 rdp = this_cpu_ptr(&rcu_data);
1547 rcu_preempt_check_blocked_tasks(rnp);
1548 rnp->qsmask = rnp->qsmaskinit;
1549 WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
1550 if (rnp == rdp->mynode)
1551 (void)__note_gp_changes(rnp, rdp);
1552 rcu_preempt_boost_start_gp(rnp);
1553 trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1554 rnp->level, rnp->grplo,
1555 rnp->grphi, rnp->qsmask);
1556 /* Quiescent states for tasks on any now-offline CPUs. */
1557 mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1558 rnp->rcu_gp_init_mask = mask;
1559 if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1560 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1561 else
1562 raw_spin_unlock_irq_rcu_node(rnp);
1563 cond_resched_tasks_rcu_qs();
1564 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1565 }
1566
1567 return true;
1568}
1569
1570/*
1571 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1572 * time.
1573 */
1574static bool rcu_gp_fqs_check_wake(int *gfp)
1575{
1576 struct rcu_node *rnp = rcu_get_root();
1577
1578 /* Someone like call_rcu() requested a force-quiescent-state scan. */
1579 *gfp = READ_ONCE(rcu_state.gp_flags);
1580 if (*gfp & RCU_GP_FLAG_FQS)
1581 return true;
1582
1583 /* The current grace period has completed. */
1584 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1585 return true;
1586
1587 return false;
1588}
1589
1590/*
1591 * Do one round of quiescent-state forcing.
1592 */
1593static void rcu_gp_fqs(bool first_time)
1594{
1595 struct rcu_node *rnp = rcu_get_root();
1596
1597 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1598 rcu_state.n_force_qs++;
1599 if (first_time) {
1600 /* Collect dyntick-idle snapshots. */
1601 force_qs_rnp(dyntick_save_progress_counter);
1602 } else {
1603 /* Handle dyntick-idle and offline CPUs. */
1604 force_qs_rnp(rcu_implicit_dynticks_qs);
1605 }
1606 /* Clear flag to prevent immediate re-entry. */
1607 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
1608 raw_spin_lock_irq_rcu_node(rnp);
1609 WRITE_ONCE(rcu_state.gp_flags,
1610 READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
1611 raw_spin_unlock_irq_rcu_node(rnp);
1612 }
1613}
1614
1615/*
1616 * Loop doing repeated quiescent-state forcing until the grace period ends.
1617 */
1618static void rcu_gp_fqs_loop(void)
1619{
1620 bool first_gp_fqs;
1621 int gf;
1622 unsigned long j;
1623 int ret;
1624 struct rcu_node *rnp = rcu_get_root();
1625
1626 first_gp_fqs = true;
1627 j = READ_ONCE(jiffies_till_first_fqs);
1628 ret = 0;
1629 for (;;) {
1630 if (!ret) {
1631 rcu_state.jiffies_force_qs = jiffies + j;
1632 WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
1633 jiffies + (j ? 3 * j : 2));
1634 }
1635 trace_rcu_grace_period(rcu_state.name,
1636 READ_ONCE(rcu_state.gp_seq),
1637 TPS("fqswait"));
1638 rcu_state.gp_state = RCU_GP_WAIT_FQS;
1639 ret = swait_event_idle_timeout_exclusive(
1640 rcu_state.gp_wq, rcu_gp_fqs_check_wake(&gf), j);
1641 rcu_state.gp_state = RCU_GP_DOING_FQS;
1642 /* Locking provides needed memory barriers. */
1643 /* If grace period done, leave loop. */
1644 if (!READ_ONCE(rnp->qsmask) &&
1645 !rcu_preempt_blocked_readers_cgp(rnp))
1646 break;
1647 /* If time for quiescent-state forcing, do it. */
1648 if (ULONG_CMP_GE(jiffies, rcu_state.jiffies_force_qs) ||
1649 (gf & RCU_GP_FLAG_FQS)) {
1650 trace_rcu_grace_period(rcu_state.name,
1651 READ_ONCE(rcu_state.gp_seq),
1652 TPS("fqsstart"));
1653 rcu_gp_fqs(first_gp_fqs);
1654 first_gp_fqs = false;
1655 trace_rcu_grace_period(rcu_state.name,
1656 READ_ONCE(rcu_state.gp_seq),
1657 TPS("fqsend"));
1658 cond_resched_tasks_rcu_qs();
1659 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1660 ret = 0; /* Force full wait till next FQS. */
1661 j = READ_ONCE(jiffies_till_next_fqs);
1662 } else {
1663 /* Deal with stray signal. */
1664 cond_resched_tasks_rcu_qs();
1665 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1666 WARN_ON(signal_pending(current));
1667 trace_rcu_grace_period(rcu_state.name,
1668 READ_ONCE(rcu_state.gp_seq),
1669 TPS("fqswaitsig"));
1670 ret = 1; /* Keep old FQS timing. */
1671 j = jiffies;
1672 if (time_after(jiffies, rcu_state.jiffies_force_qs))
1673 j = 1;
1674 else
1675 j = rcu_state.jiffies_force_qs - j;
1676 }
1677 }
1678}
1679
1680/*
1681 * Clean up after the old grace period.
1682 */
1683static void rcu_gp_cleanup(void)
1684{
1685 unsigned long gp_duration;
1686 bool needgp = false;
1687 unsigned long new_gp_seq;
1688 bool offloaded;
1689 struct rcu_data *rdp;
1690 struct rcu_node *rnp = rcu_get_root();
1691 struct swait_queue_head *sq;
1692
1693 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1694 raw_spin_lock_irq_rcu_node(rnp);
1695 rcu_state.gp_end = jiffies;
1696 gp_duration = rcu_state.gp_end - rcu_state.gp_start;
1697 if (gp_duration > rcu_state.gp_max)
1698 rcu_state.gp_max = gp_duration;
1699
1700 /*
1701 * We know the grace period is complete, but to everyone else
1702 * it appears to still be ongoing. But it is also the case
1703 * that to everyone else it looks like there is nothing that
1704 * they can do to advance the grace period. It is therefore
1705 * safe for us to drop the lock in order to mark the grace
1706 * period as completed in all of the rcu_node structures.
1707 */
1708 raw_spin_unlock_irq_rcu_node(rnp);
1709
1710 /*
1711 * Propagate new ->gp_seq value to rcu_node structures so that
1712 * other CPUs don't have to wait until the start of the next grace
1713 * period to process their callbacks. This also avoids some nasty
1714 * RCU grace-period initialization races by forcing the end of
1715 * the current grace period to be completely recorded in all of
1716 * the rcu_node structures before the beginning of the next grace
1717 * period is recorded in any of the rcu_node structures.
1718 */
1719 new_gp_seq = rcu_state.gp_seq;
1720 rcu_seq_end(&new_gp_seq);
1721 rcu_for_each_node_breadth_first(rnp) {
1722 raw_spin_lock_irq_rcu_node(rnp);
1723 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
1724 dump_blkd_tasks(rnp, 10);
1725 WARN_ON_ONCE(rnp->qsmask);
1726 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
1727 rdp = this_cpu_ptr(&rcu_data);
1728 if (rnp == rdp->mynode)
1729 needgp = __note_gp_changes(rnp, rdp) || needgp;
1730 /* smp_mb() provided by prior unlock-lock pair. */
1731 needgp = rcu_future_gp_cleanup(rnp) || needgp;
1732 sq = rcu_nocb_gp_get(rnp);
1733 raw_spin_unlock_irq_rcu_node(rnp);
1734 rcu_nocb_gp_cleanup(sq);
1735 cond_resched_tasks_rcu_qs();
1736 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1737 rcu_gp_slow(gp_cleanup_delay);
1738 }
1739 rnp = rcu_get_root();
1740 raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
1741
1742 /* Declare grace period done, trace first to use old GP number. */
1743 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
1744 rcu_seq_end(&rcu_state.gp_seq);
1745 rcu_state.gp_state = RCU_GP_IDLE;
1746 /* Check for GP requests since above loop. */
1747 rdp = this_cpu_ptr(&rcu_data);
1748 if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
1749 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
1750 TPS("CleanupMore"));
1751 needgp = true;
1752 }
1753 /* Advance CBs to reduce false positives below. */
1754 offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
1755 rcu_segcblist_is_offloaded(&rdp->cblist);
1756 if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
1757 WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
1758 rcu_state.gp_req_activity = jiffies;
1759 trace_rcu_grace_period(rcu_state.name,
1760 READ_ONCE(rcu_state.gp_seq),
1761 TPS("newreq"));
1762 } else {
1763 WRITE_ONCE(rcu_state.gp_flags,
1764 rcu_state.gp_flags & RCU_GP_FLAG_INIT);
1765 }
1766 raw_spin_unlock_irq_rcu_node(rnp);
1767}
1768
1769/*
1770 * Body of kthread that handles grace periods.
1771 */
1772static int __noreturn rcu_gp_kthread(void *unused)
1773{
1774 rcu_bind_gp_kthread();
1775 for (;;) {
1776
1777 /* Handle grace-period start. */
1778 for (;;) {
1779 trace_rcu_grace_period(rcu_state.name,
1780 READ_ONCE(rcu_state.gp_seq),
1781 TPS("reqwait"));
1782 rcu_state.gp_state = RCU_GP_WAIT_GPS;
1783 swait_event_idle_exclusive(rcu_state.gp_wq,
1784 READ_ONCE(rcu_state.gp_flags) &
1785 RCU_GP_FLAG_INIT);
1786 rcu_state.gp_state = RCU_GP_DONE_GPS;
1787 /* Locking provides needed memory barrier. */
1788 if (rcu_gp_init())
1789 break;
1790 cond_resched_tasks_rcu_qs();
1791 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1792 WARN_ON(signal_pending(current));
1793 trace_rcu_grace_period(rcu_state.name,
1794 READ_ONCE(rcu_state.gp_seq),
1795 TPS("reqwaitsig"));
1796 }
1797
1798 /* Handle quiescent-state forcing. */
1799 rcu_gp_fqs_loop();
1800
1801 /* Handle grace-period end. */
1802 rcu_state.gp_state = RCU_GP_CLEANUP;
1803 rcu_gp_cleanup();
1804 rcu_state.gp_state = RCU_GP_CLEANED;
1805 }
1806}
1807
1808/*
1809 * Report a full set of quiescent states to the rcu_state data structure.
1810 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
1811 * another grace period is required. Whether we wake the grace-period
1812 * kthread or it awakens itself for the next round of quiescent-state
1813 * forcing, that kthread will clean up after the just-completed grace
1814 * period. Note that the caller must hold rnp->lock, which is released
1815 * before return.
1816 */
1817static void rcu_report_qs_rsp(unsigned long flags)
1818 __releases(rcu_get_root()->lock)
1819{
1820 raw_lockdep_assert_held_rcu_node(rcu_get_root());
1821 WARN_ON_ONCE(!rcu_gp_in_progress());
1822 WRITE_ONCE(rcu_state.gp_flags,
1823 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
1824 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
1825 rcu_gp_kthread_wake();
1826}
1827
1828/*
1829 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1830 * Allows quiescent states for a group of CPUs to be reported at one go
1831 * to the specified rcu_node structure, though all the CPUs in the group
1832 * must be represented by the same rcu_node structure (which need not be a
1833 * leaf rcu_node structure, though it often will be). The gps parameter
1834 * is the grace-period snapshot, which means that the quiescent states
1835 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
1836 * must be held upon entry, and it is released before return.
1837 *
1838 * As a special case, if mask is zero, the bit-already-cleared check is
1839 * disabled. This allows propagating quiescent state due to resumed tasks
1840 * during grace-period initialization.
1841 */
1842static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
1843 unsigned long gps, unsigned long flags)
1844 __releases(rnp->lock)
1845{
1846 unsigned long oldmask = 0;
1847 struct rcu_node *rnp_c;
1848
1849 raw_lockdep_assert_held_rcu_node(rnp);
1850
1851 /* Walk up the rcu_node hierarchy. */
1852 for (;;) {
1853 if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
1854
1855 /*
1856 * Our bit has already been cleared, or the
1857 * relevant grace period is already over, so done.
1858 */
1859 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1860 return;
1861 }
1862 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
1863 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
1864 rcu_preempt_blocked_readers_cgp(rnp));
1865 rnp->qsmask &= ~mask;
1866 trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
1867 mask, rnp->qsmask, rnp->level,
1868 rnp->grplo, rnp->grphi,
1869 !!rnp->gp_tasks);
1870 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1871
1872 /* Other bits still set at this level, so done. */
1873 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1874 return;
1875 }
1876 rnp->completedqs = rnp->gp_seq;
1877 mask = rnp->grpmask;
1878 if (rnp->parent == NULL) {
1879
1880 /* No more levels. Exit loop holding root lock. */
1881
1882 break;
1883 }
1884 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1885 rnp_c = rnp;
1886 rnp = rnp->parent;
1887 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1888 oldmask = rnp_c->qsmask;
1889 }
1890
1891 /*
1892 * Get here if we are the last CPU to pass through a quiescent
1893 * state for this grace period. Invoke rcu_report_qs_rsp()
1894 * to clean up and start the next grace period if one is needed.
1895 */
1896 rcu_report_qs_rsp(flags); /* releases rnp->lock. */
1897}
1898
1899/*
1900 * Record a quiescent state for all tasks that were previously queued
1901 * on the specified rcu_node structure and that were blocking the current
1902 * RCU grace period. The caller must hold the corresponding rnp->lock with
1903 * irqs disabled, and this lock is released upon return, but irqs remain
1904 * disabled.
1905 */
1906static void __maybe_unused
1907rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
1908 __releases(rnp->lock)
1909{
1910 unsigned long gps;
1911 unsigned long mask;
1912 struct rcu_node *rnp_p;
1913
1914 raw_lockdep_assert_held_rcu_node(rnp);
1915 if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPTION)) ||
1916 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
1917 rnp->qsmask != 0) {
1918 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1919 return; /* Still need more quiescent states! */
1920 }
1921
1922 rnp->completedqs = rnp->gp_seq;
1923 rnp_p = rnp->parent;
1924 if (rnp_p == NULL) {
1925 /*
1926 * Only one rcu_node structure in the tree, so don't
1927 * try to report up to its nonexistent parent!
1928 */
1929 rcu_report_qs_rsp(flags);
1930 return;
1931 }
1932
1933 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
1934 gps = rnp->gp_seq;
1935 mask = rnp->grpmask;
1936 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1937 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
1938 rcu_report_qs_rnp(mask, rnp_p, gps, flags);
1939}
1940
1941/*
1942 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1943 * structure. This must be called from the specified CPU.
1944 */
1945static void
1946rcu_report_qs_rdp(int cpu, struct rcu_data *rdp)
1947{
1948 unsigned long flags;
1949 unsigned long mask;
1950 bool needwake = false;
1951 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
1952 rcu_segcblist_is_offloaded(&rdp->cblist);
1953 struct rcu_node *rnp;
1954
1955 rnp = rdp->mynode;
1956 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1957 if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
1958 rdp->gpwrap) {
1959
1960 /*
1961 * The grace period in which this quiescent state was
1962 * recorded has ended, so don't report it upwards.
1963 * We will instead need a new quiescent state that lies
1964 * within the current grace period.
1965 */
1966 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
1967 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1968 return;
1969 }
1970 mask = rdp->grpmask;
1971 rdp->core_needs_qs = false;
1972 if ((rnp->qsmask & mask) == 0) {
1973 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1974 } else {
1975 /*
1976 * This GP can't end until cpu checks in, so all of our
1977 * callbacks can be processed during the next GP.
1978 */
1979 if (!offloaded)
1980 needwake = rcu_accelerate_cbs(rnp, rdp);
1981
1982 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1983 /* ^^^ Released rnp->lock */
1984 if (needwake)
1985 rcu_gp_kthread_wake();
1986 }
1987}
1988
1989/*
1990 * Check to see if there is a new grace period of which this CPU
1991 * is not yet aware, and if so, set up local rcu_data state for it.
1992 * Otherwise, see if this CPU has just passed through its first
1993 * quiescent state for this grace period, and record that fact if so.
1994 */
1995static void
1996rcu_check_quiescent_state(struct rcu_data *rdp)
1997{
1998 /* Check for grace-period ends and beginnings. */
1999 note_gp_changes(rdp);
2000
2001 /*
2002 * Does this CPU still need to do its part for current grace period?
2003 * If no, return and let the other CPUs do their part as well.
2004 */
2005 if (!rdp->core_needs_qs)
2006 return;
2007
2008 /*
2009 * Was there a quiescent state since the beginning of the grace
2010 * period? If no, then exit and wait for the next call.
2011 */
2012 if (rdp->cpu_no_qs.b.norm)
2013 return;
2014
2015 /*
2016 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2017 * judge of that).
2018 */
2019 rcu_report_qs_rdp(rdp->cpu, rdp);
2020}
2021
2022/*
2023 * Near the end of the offline process. Trace the fact that this CPU
2024 * is going offline.
2025 */
2026int rcutree_dying_cpu(unsigned int cpu)
2027{
2028 bool blkd;
2029 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
2030 struct rcu_node *rnp = rdp->mynode;
2031
2032 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2033 return 0;
2034
2035 blkd = !!(rnp->qsmask & rdp->grpmask);
2036 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq,
2037 blkd ? TPS("cpuofl") : TPS("cpuofl-bgp"));
2038 return 0;
2039}
2040
2041/*
2042 * All CPUs for the specified rcu_node structure have gone offline,
2043 * and all tasks that were preempted within an RCU read-side critical
2044 * section while running on one of those CPUs have since exited their RCU
2045 * read-side critical section. Some other CPU is reporting this fact with
2046 * the specified rcu_node structure's ->lock held and interrupts disabled.
2047 * This function therefore goes up the tree of rcu_node structures,
2048 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2049 * the leaf rcu_node structure's ->qsmaskinit field has already been
2050 * updated.
2051 *
2052 * This function does check that the specified rcu_node structure has
2053 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2054 * prematurely. That said, invoking it after the fact will cost you
2055 * a needless lock acquisition. So once it has done its work, don't
2056 * invoke it again.
2057 */
2058static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2059{
2060 long mask;
2061 struct rcu_node *rnp = rnp_leaf;
2062
2063 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2064 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2065 WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
2066 WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
2067 return;
2068 for (;;) {
2069 mask = rnp->grpmask;
2070 rnp = rnp->parent;
2071 if (!rnp)
2072 break;
2073 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2074 rnp->qsmaskinit &= ~mask;
2075 /* Between grace periods, so better already be zero! */
2076 WARN_ON_ONCE(rnp->qsmask);
2077 if (rnp->qsmaskinit) {
2078 raw_spin_unlock_rcu_node(rnp);
2079 /* irqs remain disabled. */
2080 return;
2081 }
2082 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2083 }
2084}
2085
2086/*
2087 * The CPU has been completely removed, and some other CPU is reporting
2088 * this fact from process context. Do the remainder of the cleanup.
2089 * There can only be one CPU hotplug operation at a time, so no need for
2090 * explicit locking.
2091 */
2092int rcutree_dead_cpu(unsigned int cpu)
2093{
2094 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2095 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2096
2097 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2098 return 0;
2099
2100 /* Adjust any no-longer-needed kthreads. */
2101 rcu_boost_kthread_setaffinity(rnp, -1);
2102 /* Do any needed no-CB deferred wakeups from this CPU. */
2103 do_nocb_deferred_wakeup(per_cpu_ptr(&rcu_data, cpu));
2104 return 0;
2105}
2106
2107/*
2108 * Invoke any RCU callbacks that have made it to the end of their grace
2109 * period. Thottle as specified by rdp->blimit.
2110 */
2111static void rcu_do_batch(struct rcu_data *rdp)
2112{
2113 unsigned long flags;
2114 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2115 rcu_segcblist_is_offloaded(&rdp->cblist);
2116 struct rcu_head *rhp;
2117 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2118 long bl, count;
2119 long pending, tlimit = 0;
2120
2121 /* If no callbacks are ready, just return. */
2122 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2123 trace_rcu_batch_start(rcu_state.name,
2124 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2125 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2126 trace_rcu_batch_end(rcu_state.name, 0,
2127 !rcu_segcblist_empty(&rdp->cblist),
2128 need_resched(), is_idle_task(current),
2129 rcu_is_callbacks_kthread());
2130 return;
2131 }
2132
2133 /*
2134 * Extract the list of ready callbacks, disabling to prevent
2135 * races with call_rcu() from interrupt handlers. Leave the
2136 * callback counts, as rcu_barrier() needs to be conservative.
2137 */
2138 local_irq_save(flags);
2139 rcu_nocb_lock(rdp);
2140 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2141 pending = rcu_segcblist_n_cbs(&rdp->cblist);
2142 bl = max(rdp->blimit, pending >> rcu_divisor);
2143 if (unlikely(bl > 100))
2144 tlimit = local_clock() + rcu_resched_ns;
2145 trace_rcu_batch_start(rcu_state.name,
2146 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2147 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2148 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2149 if (offloaded)
2150 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2151 rcu_nocb_unlock_irqrestore(rdp, flags);
2152
2153 /* Invoke callbacks. */
2154 rhp = rcu_cblist_dequeue(&rcl);
2155 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2156 debug_rcu_head_unqueue(rhp);
2157 if (__rcu_reclaim(rcu_state.name, rhp))
2158 rcu_cblist_dequeued_lazy(&rcl);
2159 /*
2160 * Stop only if limit reached and CPU has something to do.
2161 * Note: The rcl structure counts down from zero.
2162 */
2163 if (-rcl.len >= bl && !offloaded &&
2164 (need_resched() ||
2165 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2166 break;
2167 if (unlikely(tlimit)) {
2168 /* only call local_clock() every 32 callbacks */
2169 if (likely((-rcl.len & 31) || local_clock() < tlimit))
2170 continue;
2171 /* Exceeded the time limit, so leave. */
2172 break;
2173 }
2174 if (offloaded) {
2175 WARN_ON_ONCE(in_serving_softirq());
2176 local_bh_enable();
2177 lockdep_assert_irqs_enabled();
2178 cond_resched_tasks_rcu_qs();
2179 lockdep_assert_irqs_enabled();
2180 local_bh_disable();
2181 }
2182 }
2183
2184 local_irq_save(flags);
2185 rcu_nocb_lock(rdp);
2186 count = -rcl.len;
2187 trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2188 is_idle_task(current), rcu_is_callbacks_kthread());
2189
2190 /* Update counts and requeue any remaining callbacks. */
2191 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2192 smp_mb(); /* List handling before counting for rcu_barrier(). */
2193 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2194
2195 /* Reinstate batch limit if we have worked down the excess. */
2196 count = rcu_segcblist_n_cbs(&rdp->cblist);
2197 if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
2198 rdp->blimit = blimit;
2199
2200 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2201 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2202 rdp->qlen_last_fqs_check = 0;
2203 rdp->n_force_qs_snap = rcu_state.n_force_qs;
2204 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2205 rdp->qlen_last_fqs_check = count;
2206
2207 /*
2208 * The following usually indicates a double call_rcu(). To track
2209 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2210 */
2211 WARN_ON_ONCE(count == 0 && !rcu_segcblist_empty(&rdp->cblist));
2212 WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2213 count != 0 && rcu_segcblist_empty(&rdp->cblist));
2214
2215 rcu_nocb_unlock_irqrestore(rdp, flags);
2216
2217 /* Re-invoke RCU core processing if there are callbacks remaining. */
2218 if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist))
2219 invoke_rcu_core();
2220}
2221
2222/*
2223 * This function is invoked from each scheduling-clock interrupt,
2224 * and checks to see if this CPU is in a non-context-switch quiescent
2225 * state, for example, user mode or idle loop. It also schedules RCU
2226 * core processing. If the current grace period has gone on too long,
2227 * it will ask the scheduler to manufacture a context switch for the sole
2228 * purpose of providing a providing the needed quiescent state.
2229 */
2230void rcu_sched_clock_irq(int user)
2231{
2232 trace_rcu_utilization(TPS("Start scheduler-tick"));
2233 raw_cpu_inc(rcu_data.ticks_this_gp);
2234 /* The load-acquire pairs with the store-release setting to true. */
2235 if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
2236 /* Idle and userspace execution already are quiescent states. */
2237 if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2238 set_tsk_need_resched(current);
2239 set_preempt_need_resched();
2240 }
2241 __this_cpu_write(rcu_data.rcu_urgent_qs, false);
2242 }
2243 rcu_flavor_sched_clock_irq(user);
2244 if (rcu_pending())
2245 invoke_rcu_core();
2246
2247 trace_rcu_utilization(TPS("End scheduler-tick"));
2248}
2249
2250/*
2251 * Scan the leaf rcu_node structures. For each structure on which all
2252 * CPUs have reported a quiescent state and on which there are tasks
2253 * blocking the current grace period, initiate RCU priority boosting.
2254 * Otherwise, invoke the specified function to check dyntick state for
2255 * each CPU that has not yet reported a quiescent state.
2256 */
2257static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
2258{
2259 int cpu;
2260 unsigned long flags;
2261 unsigned long mask;
2262 struct rcu_node *rnp;
2263
2264 rcu_for_each_leaf_node(rnp) {
2265 cond_resched_tasks_rcu_qs();
2266 mask = 0;
2267 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2268 if (rnp->qsmask == 0) {
2269 if (!IS_ENABLED(CONFIG_PREEMPTION) ||
2270 rcu_preempt_blocked_readers_cgp(rnp)) {
2271 /*
2272 * No point in scanning bits because they
2273 * are all zero. But we might need to
2274 * priority-boost blocked readers.
2275 */
2276 rcu_initiate_boost(rnp, flags);
2277 /* rcu_initiate_boost() releases rnp->lock */
2278 continue;
2279 }
2280 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2281 continue;
2282 }
2283 for_each_leaf_node_possible_cpu(rnp, cpu) {
2284 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
2285 if ((rnp->qsmask & bit) != 0) {
2286 if (f(per_cpu_ptr(&rcu_data, cpu)))
2287 mask |= bit;
2288 }
2289 }
2290 if (mask != 0) {
2291 /* Idle/offline CPUs, report (releases rnp->lock). */
2292 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2293 } else {
2294 /* Nothing to do here, so just drop the lock. */
2295 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2296 }
2297 }
2298}
2299
2300/*
2301 * Force quiescent states on reluctant CPUs, and also detect which
2302 * CPUs are in dyntick-idle mode.
2303 */
2304void rcu_force_quiescent_state(void)
2305{
2306 unsigned long flags;
2307 bool ret;
2308 struct rcu_node *rnp;
2309 struct rcu_node *rnp_old = NULL;
2310
2311 /* Funnel through hierarchy to reduce memory contention. */
2312 rnp = __this_cpu_read(rcu_data.mynode);
2313 for (; rnp != NULL; rnp = rnp->parent) {
2314 ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2315 !raw_spin_trylock(&rnp->fqslock);
2316 if (rnp_old != NULL)
2317 raw_spin_unlock(&rnp_old->fqslock);
2318 if (ret)
2319 return;
2320 rnp_old = rnp;
2321 }
2322 /* rnp_old == rcu_get_root(), rnp == NULL. */
2323
2324 /* Reached the root of the rcu_node tree, acquire lock. */
2325 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2326 raw_spin_unlock(&rnp_old->fqslock);
2327 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2328 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2329 return; /* Someone beat us to it. */
2330 }
2331 WRITE_ONCE(rcu_state.gp_flags,
2332 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2333 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2334 rcu_gp_kthread_wake();
2335}
2336EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
2337
2338/* Perform RCU core processing work for the current CPU. */
2339static __latent_entropy void rcu_core(void)
2340{
2341 unsigned long flags;
2342 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2343 struct rcu_node *rnp = rdp->mynode;
2344 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2345 rcu_segcblist_is_offloaded(&rdp->cblist);
2346
2347 if (cpu_is_offline(smp_processor_id()))
2348 return;
2349 trace_rcu_utilization(TPS("Start RCU core"));
2350 WARN_ON_ONCE(!rdp->beenonline);
2351
2352 /* Report any deferred quiescent states if preemption enabled. */
2353 if (!(preempt_count() & PREEMPT_MASK)) {
2354 rcu_preempt_deferred_qs(current);
2355 } else if (rcu_preempt_need_deferred_qs(current)) {
2356 set_tsk_need_resched(current);
2357 set_preempt_need_resched();
2358 }
2359
2360 /* Update RCU state based on any recent quiescent states. */
2361 rcu_check_quiescent_state(rdp);
2362
2363 /* No grace period and unregistered callbacks? */
2364 if (!rcu_gp_in_progress() &&
2365 rcu_segcblist_is_enabled(&rdp->cblist) && !offloaded) {
2366 local_irq_save(flags);
2367 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2368 rcu_accelerate_cbs_unlocked(rnp, rdp);
2369 local_irq_restore(flags);
2370 }
2371
2372 rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
2373
2374 /* If there are callbacks ready, invoke them. */
2375 if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist) &&
2376 likely(READ_ONCE(rcu_scheduler_fully_active)))
2377 rcu_do_batch(rdp);
2378
2379 /* Do any needed deferred wakeups of rcuo kthreads. */
2380 do_nocb_deferred_wakeup(rdp);
2381 trace_rcu_utilization(TPS("End RCU core"));
2382}
2383
2384static void rcu_core_si(struct softirq_action *h)
2385{
2386 rcu_core();
2387}
2388
2389static void rcu_wake_cond(struct task_struct *t, int status)
2390{
2391 /*
2392 * If the thread is yielding, only wake it when this
2393 * is invoked from idle
2394 */
2395 if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
2396 wake_up_process(t);
2397}
2398
2399static void invoke_rcu_core_kthread(void)
2400{
2401 struct task_struct *t;
2402 unsigned long flags;
2403
2404 local_irq_save(flags);
2405 __this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
2406 t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
2407 if (t != NULL && t != current)
2408 rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
2409 local_irq_restore(flags);
2410}
2411
2412/*
2413 * Wake up this CPU's rcuc kthread to do RCU core processing.
2414 */
2415static void invoke_rcu_core(void)
2416{
2417 if (!cpu_online(smp_processor_id()))
2418 return;
2419 if (use_softirq)
2420 raise_softirq(RCU_SOFTIRQ);
2421 else
2422 invoke_rcu_core_kthread();
2423}
2424
2425static void rcu_cpu_kthread_park(unsigned int cpu)
2426{
2427 per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
2428}
2429
2430static int rcu_cpu_kthread_should_run(unsigned int cpu)
2431{
2432 return __this_cpu_read(rcu_data.rcu_cpu_has_work);
2433}
2434
2435/*
2436 * Per-CPU kernel thread that invokes RCU callbacks. This replaces
2437 * the RCU softirq used in configurations of RCU that do not support RCU
2438 * priority boosting.
2439 */
2440static void rcu_cpu_kthread(unsigned int cpu)
2441{
2442 unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
2443 char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
2444 int spincnt;
2445
2446 for (spincnt = 0; spincnt < 10; spincnt++) {
2447 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
2448 local_bh_disable();
2449 *statusp = RCU_KTHREAD_RUNNING;
2450 local_irq_disable();
2451 work = *workp;
2452 *workp = 0;
2453 local_irq_enable();
2454 if (work)
2455 rcu_core();
2456 local_bh_enable();
2457 if (*workp == 0) {
2458 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2459 *statusp = RCU_KTHREAD_WAITING;
2460 return;
2461 }
2462 }
2463 *statusp = RCU_KTHREAD_YIELDING;
2464 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2465 schedule_timeout_interruptible(2);
2466 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2467 *statusp = RCU_KTHREAD_WAITING;
2468}
2469
2470static struct smp_hotplug_thread rcu_cpu_thread_spec = {
2471 .store = &rcu_data.rcu_cpu_kthread_task,
2472 .thread_should_run = rcu_cpu_kthread_should_run,
2473 .thread_fn = rcu_cpu_kthread,
2474 .thread_comm = "rcuc/%u",
2475 .setup = rcu_cpu_kthread_setup,
2476 .park = rcu_cpu_kthread_park,
2477};
2478
2479/*
2480 * Spawn per-CPU RCU core processing kthreads.
2481 */
2482static int __init rcu_spawn_core_kthreads(void)
2483{
2484 int cpu;
2485
2486 for_each_possible_cpu(cpu)
2487 per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
2488 if (!IS_ENABLED(CONFIG_RCU_BOOST) && use_softirq)
2489 return 0;
2490 WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
2491 "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
2492 return 0;
2493}
2494early_initcall(rcu_spawn_core_kthreads);
2495
2496/*
2497 * Handle any core-RCU processing required by a call_rcu() invocation.
2498 */
2499static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2500 unsigned long flags)
2501{
2502 /*
2503 * If called from an extended quiescent state, invoke the RCU
2504 * core in order to force a re-evaluation of RCU's idleness.
2505 */
2506 if (!rcu_is_watching())
2507 invoke_rcu_core();
2508
2509 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2510 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2511 return;
2512
2513 /*
2514 * Force the grace period if too many callbacks or too long waiting.
2515 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2516 * if some other CPU has recently done so. Also, don't bother
2517 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2518 * is the only one waiting for a grace period to complete.
2519 */
2520 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2521 rdp->qlen_last_fqs_check + qhimark)) {
2522
2523 /* Are we ignoring a completed grace period? */
2524 note_gp_changes(rdp);
2525
2526 /* Start a new grace period if one not already started. */
2527 if (!rcu_gp_in_progress()) {
2528 rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
2529 } else {
2530 /* Give the grace period a kick. */
2531 rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
2532 if (rcu_state.n_force_qs == rdp->n_force_qs_snap &&
2533 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2534 rcu_force_quiescent_state();
2535 rdp->n_force_qs_snap = rcu_state.n_force_qs;
2536 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2537 }
2538 }
2539}
2540
2541/*
2542 * RCU callback function to leak a callback.
2543 */
2544static void rcu_leak_callback(struct rcu_head *rhp)
2545{
2546}
2547
2548/*
2549 * Helper function for call_rcu() and friends. The cpu argument will
2550 * normally be -1, indicating "currently running CPU". It may specify
2551 * a CPU only if that CPU is a no-CBs CPU. Currently, only rcu_barrier()
2552 * is expected to specify a CPU.
2553 */
2554static void
2555__call_rcu(struct rcu_head *head, rcu_callback_t func, bool lazy)
2556{
2557 unsigned long flags;
2558 struct rcu_data *rdp;
2559 bool was_alldone;
2560
2561 /* Misaligned rcu_head! */
2562 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2563
2564 if (debug_rcu_head_queue(head)) {
2565 /*
2566 * Probable double call_rcu(), so leak the callback.
2567 * Use rcu:rcu_callback trace event to find the previous
2568 * time callback was passed to __call_rcu().
2569 */
2570 WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pS()!!!\n",
2571 head, head->func);
2572 WRITE_ONCE(head->func, rcu_leak_callback);
2573 return;
2574 }
2575 head->func = func;
2576 head->next = NULL;
2577 local_irq_save(flags);
2578 rdp = this_cpu_ptr(&rcu_data);
2579
2580 /* Add the callback to our list. */
2581 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
2582 // This can trigger due to call_rcu() from offline CPU:
2583 WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
2584 WARN_ON_ONCE(!rcu_is_watching());
2585 // Very early boot, before rcu_init(). Initialize if needed
2586 // and then drop through to queue the callback.
2587 if (rcu_segcblist_empty(&rdp->cblist))
2588 rcu_segcblist_init(&rdp->cblist);
2589 }
2590 if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags))
2591 return; // Enqueued onto ->nocb_bypass, so just leave.
2592 /* If we get here, rcu_nocb_try_bypass() acquired ->nocb_lock. */
2593 rcu_segcblist_enqueue(&rdp->cblist, head, lazy);
2594 if (__is_kfree_rcu_offset((unsigned long)func))
2595 trace_rcu_kfree_callback(rcu_state.name, head,
2596 (unsigned long)func,
2597 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2598 rcu_segcblist_n_cbs(&rdp->cblist));
2599 else
2600 trace_rcu_callback(rcu_state.name, head,
2601 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2602 rcu_segcblist_n_cbs(&rdp->cblist));
2603
2604 /* Go handle any RCU core processing required. */
2605 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2606 unlikely(rcu_segcblist_is_offloaded(&rdp->cblist))) {
2607 __call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */
2608 } else {
2609 __call_rcu_core(rdp, head, flags);
2610 local_irq_restore(flags);
2611 }
2612}
2613
2614/**
2615 * call_rcu() - Queue an RCU callback for invocation after a grace period.
2616 * @head: structure to be used for queueing the RCU updates.
2617 * @func: actual callback function to be invoked after the grace period
2618 *
2619 * The callback function will be invoked some time after a full grace
2620 * period elapses, in other words after all pre-existing RCU read-side
2621 * critical sections have completed. However, the callback function
2622 * might well execute concurrently with RCU read-side critical sections
2623 * that started after call_rcu() was invoked. RCU read-side critical
2624 * sections are delimited by rcu_read_lock() and rcu_read_unlock(), and
2625 * may be nested. In addition, regions of code across which interrupts,
2626 * preemption, or softirqs have been disabled also serve as RCU read-side
2627 * critical sections. This includes hardware interrupt handlers, softirq
2628 * handlers, and NMI handlers.
2629 *
2630 * Note that all CPUs must agree that the grace period extended beyond
2631 * all pre-existing RCU read-side critical section. On systems with more
2632 * than one CPU, this means that when "func()" is invoked, each CPU is
2633 * guaranteed to have executed a full memory barrier since the end of its
2634 * last RCU read-side critical section whose beginning preceded the call
2635 * to call_rcu(). It also means that each CPU executing an RCU read-side
2636 * critical section that continues beyond the start of "func()" must have
2637 * executed a memory barrier after the call_rcu() but before the beginning
2638 * of that RCU read-side critical section. Note that these guarantees
2639 * include CPUs that are offline, idle, or executing in user mode, as
2640 * well as CPUs that are executing in the kernel.
2641 *
2642 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
2643 * resulting RCU callback function "func()", then both CPU A and CPU B are
2644 * guaranteed to execute a full memory barrier during the time interval
2645 * between the call to call_rcu() and the invocation of "func()" -- even
2646 * if CPU A and CPU B are the same CPU (but again only if the system has
2647 * more than one CPU).
2648 */
2649void call_rcu(struct rcu_head *head, rcu_callback_t func)
2650{
2651 __call_rcu(head, func, 0);
2652}
2653EXPORT_SYMBOL_GPL(call_rcu);
2654
2655/*
2656 * Queue an RCU callback for lazy invocation after a grace period.
2657 * This will likely be later named something like "call_rcu_lazy()",
2658 * but this change will require some way of tagging the lazy RCU
2659 * callbacks in the list of pending callbacks. Until then, this
2660 * function may only be called from __kfree_rcu().
2661 */
2662void kfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
2663{
2664 __call_rcu(head, func, 1);
2665}
2666EXPORT_SYMBOL_GPL(kfree_call_rcu);
2667
2668/*
2669 * During early boot, any blocking grace-period wait automatically
2670 * implies a grace period. Later on, this is never the case for PREEMPT.
2671 *
2672 * Howevr, because a context switch is a grace period for !PREEMPT, any
2673 * blocking grace-period wait automatically implies a grace period if
2674 * there is only one CPU online at any point time during execution of
2675 * either synchronize_rcu() or synchronize_rcu_expedited(). It is OK to
2676 * occasionally incorrectly indicate that there are multiple CPUs online
2677 * when there was in fact only one the whole time, as this just adds some
2678 * overhead: RCU still operates correctly.
2679 */
2680static int rcu_blocking_is_gp(void)
2681{
2682 int ret;
2683
2684 if (IS_ENABLED(CONFIG_PREEMPTION))
2685 return rcu_scheduler_active == RCU_SCHEDULER_INACTIVE;
2686 might_sleep(); /* Check for RCU read-side critical section. */
2687 preempt_disable();
2688 ret = num_online_cpus() <= 1;
2689 preempt_enable();
2690 return ret;
2691}
2692
2693/**
2694 * synchronize_rcu - wait until a grace period has elapsed.
2695 *
2696 * Control will return to the caller some time after a full grace
2697 * period has elapsed, in other words after all currently executing RCU
2698 * read-side critical sections have completed. Note, however, that
2699 * upon return from synchronize_rcu(), the caller might well be executing
2700 * concurrently with new RCU read-side critical sections that began while
2701 * synchronize_rcu() was waiting. RCU read-side critical sections are
2702 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
2703 * In addition, regions of code across which interrupts, preemption, or
2704 * softirqs have been disabled also serve as RCU read-side critical
2705 * sections. This includes hardware interrupt handlers, softirq handlers,
2706 * and NMI handlers.
2707 *
2708 * Note that this guarantee implies further memory-ordering guarantees.
2709 * On systems with more than one CPU, when synchronize_rcu() returns,
2710 * each CPU is guaranteed to have executed a full memory barrier since
2711 * the end of its last RCU read-side critical section whose beginning
2712 * preceded the call to synchronize_rcu(). In addition, each CPU having
2713 * an RCU read-side critical section that extends beyond the return from
2714 * synchronize_rcu() is guaranteed to have executed a full memory barrier
2715 * after the beginning of synchronize_rcu() and before the beginning of
2716 * that RCU read-side critical section. Note that these guarantees include
2717 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2718 * that are executing in the kernel.
2719 *
2720 * Furthermore, if CPU A invoked synchronize_rcu(), which returned
2721 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2722 * to have executed a full memory barrier during the execution of
2723 * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
2724 * again only if the system has more than one CPU).
2725 */
2726void synchronize_rcu(void)
2727{
2728 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
2729 lock_is_held(&rcu_lock_map) ||
2730 lock_is_held(&rcu_sched_lock_map),
2731 "Illegal synchronize_rcu() in RCU read-side critical section");
2732 if (rcu_blocking_is_gp())
2733 return;
2734 if (rcu_gp_is_expedited())
2735 synchronize_rcu_expedited();
2736 else
2737 wait_rcu_gp(call_rcu);
2738}
2739EXPORT_SYMBOL_GPL(synchronize_rcu);
2740
2741/**
2742 * get_state_synchronize_rcu - Snapshot current RCU state
2743 *
2744 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
2745 * to determine whether or not a full grace period has elapsed in the
2746 * meantime.
2747 */
2748unsigned long get_state_synchronize_rcu(void)
2749{
2750 /*
2751 * Any prior manipulation of RCU-protected data must happen
2752 * before the load from ->gp_seq.
2753 */
2754 smp_mb(); /* ^^^ */
2755 return rcu_seq_snap(&rcu_state.gp_seq);
2756}
2757EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
2758
2759/**
2760 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
2761 *
2762 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
2763 *
2764 * If a full RCU grace period has elapsed since the earlier call to
2765 * get_state_synchronize_rcu(), just return. Otherwise, invoke
2766 * synchronize_rcu() to wait for a full grace period.
2767 *
2768 * Yes, this function does not take counter wrap into account. But
2769 * counter wrap is harmless. If the counter wraps, we have waited for
2770 * more than 2 billion grace periods (and way more on a 64-bit system!),
2771 * so waiting for one additional grace period should be just fine.
2772 */
2773void cond_synchronize_rcu(unsigned long oldstate)
2774{
2775 if (!rcu_seq_done(&rcu_state.gp_seq, oldstate))
2776 synchronize_rcu();
2777 else
2778 smp_mb(); /* Ensure GP ends before subsequent accesses. */
2779}
2780EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
2781
2782/*
2783 * Check to see if there is any immediate RCU-related work to be done by
2784 * the current CPU, returning 1 if so and zero otherwise. The checks are
2785 * in order of increasing expense: checks that can be carried out against
2786 * CPU-local state are performed first. However, we must check for CPU
2787 * stalls first, else we might not get a chance.
2788 */
2789static int rcu_pending(void)
2790{
2791 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
2792 struct rcu_node *rnp = rdp->mynode;
2793
2794 /* Check for CPU stalls, if enabled. */
2795 check_cpu_stall(rdp);
2796
2797 /* Does this CPU need a deferred NOCB wakeup? */
2798 if (rcu_nocb_need_deferred_wakeup(rdp))
2799 return 1;
2800
2801 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
2802 if (rcu_nohz_full_cpu())
2803 return 0;
2804
2805 /* Is the RCU core waiting for a quiescent state from this CPU? */
2806 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm)
2807 return 1;
2808
2809 /* Does this CPU have callbacks ready to invoke? */
2810 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2811 return 1;
2812
2813 /* Has RCU gone idle with this CPU needing another grace period? */
2814 if (!rcu_gp_in_progress() &&
2815 rcu_segcblist_is_enabled(&rdp->cblist) &&
2816 (!IS_ENABLED(CONFIG_RCU_NOCB_CPU) ||
2817 !rcu_segcblist_is_offloaded(&rdp->cblist)) &&
2818 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2819 return 1;
2820
2821 /* Have RCU grace period completed or started? */
2822 if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
2823 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
2824 return 1;
2825
2826 /* nothing to do */
2827 return 0;
2828}
2829
2830/*
2831 * Helper function for rcu_barrier() tracing. If tracing is disabled,
2832 * the compiler is expected to optimize this away.
2833 */
2834static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
2835{
2836 trace_rcu_barrier(rcu_state.name, s, cpu,
2837 atomic_read(&rcu_state.barrier_cpu_count), done);
2838}
2839
2840/*
2841 * RCU callback function for rcu_barrier(). If we are last, wake
2842 * up the task executing rcu_barrier().
2843 */
2844static void rcu_barrier_callback(struct rcu_head *rhp)
2845{
2846 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
2847 rcu_barrier_trace(TPS("LastCB"), -1,
2848 rcu_state.barrier_sequence);
2849 complete(&rcu_state.barrier_completion);
2850 } else {
2851 rcu_barrier_trace(TPS("CB"), -1, rcu_state.barrier_sequence);
2852 }
2853}
2854
2855/*
2856 * Called with preemption disabled, and from cross-cpu IRQ context.
2857 */
2858static void rcu_barrier_func(void *unused)
2859{
2860 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2861
2862 rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
2863 rdp->barrier_head.func = rcu_barrier_callback;
2864 debug_rcu_head_queue(&rdp->barrier_head);
2865 rcu_nocb_lock(rdp);
2866 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
2867 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head, 0)) {
2868 atomic_inc(&rcu_state.barrier_cpu_count);
2869 } else {
2870 debug_rcu_head_unqueue(&rdp->barrier_head);
2871 rcu_barrier_trace(TPS("IRQNQ"), -1,
2872 rcu_state.barrier_sequence);
2873 }
2874 rcu_nocb_unlock(rdp);
2875}
2876
2877/**
2878 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
2879 *
2880 * Note that this primitive does not necessarily wait for an RCU grace period
2881 * to complete. For example, if there are no RCU callbacks queued anywhere
2882 * in the system, then rcu_barrier() is within its rights to return
2883 * immediately, without waiting for anything, much less an RCU grace period.
2884 */
2885void rcu_barrier(void)
2886{
2887 int cpu;
2888 struct rcu_data *rdp;
2889 unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
2890
2891 rcu_barrier_trace(TPS("Begin"), -1, s);
2892
2893 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2894 mutex_lock(&rcu_state.barrier_mutex);
2895
2896 /* Did someone else do our work for us? */
2897 if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
2898 rcu_barrier_trace(TPS("EarlyExit"), -1,
2899 rcu_state.barrier_sequence);
2900 smp_mb(); /* caller's subsequent code after above check. */
2901 mutex_unlock(&rcu_state.barrier_mutex);
2902 return;
2903 }
2904
2905 /* Mark the start of the barrier operation. */
2906 rcu_seq_start(&rcu_state.barrier_sequence);
2907 rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
2908
2909 /*
2910 * Initialize the count to one rather than to zero in order to
2911 * avoid a too-soon return to zero in case of a short grace period
2912 * (or preemption of this task). Exclude CPU-hotplug operations
2913 * to ensure that no offline CPU has callbacks queued.
2914 */
2915 init_completion(&rcu_state.barrier_completion);
2916 atomic_set(&rcu_state.barrier_cpu_count, 1);
2917 get_online_cpus();
2918
2919 /*
2920 * Force each CPU with callbacks to register a new callback.
2921 * When that callback is invoked, we will know that all of the
2922 * corresponding CPU's preceding callbacks have been invoked.
2923 */
2924 for_each_possible_cpu(cpu) {
2925 rdp = per_cpu_ptr(&rcu_data, cpu);
2926 if (!cpu_online(cpu) &&
2927 !rcu_segcblist_is_offloaded(&rdp->cblist))
2928 continue;
2929 if (rcu_segcblist_n_cbs(&rdp->cblist)) {
2930 rcu_barrier_trace(TPS("OnlineQ"), cpu,
2931 rcu_state.barrier_sequence);
2932 smp_call_function_single(cpu, rcu_barrier_func, NULL, 1);
2933 } else {
2934 rcu_barrier_trace(TPS("OnlineNQ"), cpu,
2935 rcu_state.barrier_sequence);
2936 }
2937 }
2938 put_online_cpus();
2939
2940 /*
2941 * Now that we have an rcu_barrier_callback() callback on each
2942 * CPU, and thus each counted, remove the initial count.
2943 */
2944 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count))
2945 complete(&rcu_state.barrier_completion);
2946
2947 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2948 wait_for_completion(&rcu_state.barrier_completion);
2949
2950 /* Mark the end of the barrier operation. */
2951 rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
2952 rcu_seq_end(&rcu_state.barrier_sequence);
2953
2954 /* Other rcu_barrier() invocations can now safely proceed. */
2955 mutex_unlock(&rcu_state.barrier_mutex);
2956}
2957EXPORT_SYMBOL_GPL(rcu_barrier);
2958
2959/*
2960 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
2961 * first CPU in a given leaf rcu_node structure coming online. The caller
2962 * must hold the corresponding leaf rcu_node ->lock with interrrupts
2963 * disabled.
2964 */
2965static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
2966{
2967 long mask;
2968 long oldmask;
2969 struct rcu_node *rnp = rnp_leaf;
2970
2971 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2972 WARN_ON_ONCE(rnp->wait_blkd_tasks);
2973 for (;;) {
2974 mask = rnp->grpmask;
2975 rnp = rnp->parent;
2976 if (rnp == NULL)
2977 return;
2978 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
2979 oldmask = rnp->qsmaskinit;
2980 rnp->qsmaskinit |= mask;
2981 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
2982 if (oldmask)
2983 return;
2984 }
2985}
2986
2987/*
2988 * Do boot-time initialization of a CPU's per-CPU RCU data.
2989 */
2990static void __init
2991rcu_boot_init_percpu_data(int cpu)
2992{
2993 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2994
2995 /* Set up local state, ensuring consistent view of global state. */
2996 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
2997 WARN_ON_ONCE(rdp->dynticks_nesting != 1);
2998 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp)));
2999 rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
3000 rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
3001 rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
3002 rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
3003 rdp->cpu = cpu;
3004 rcu_boot_init_nocb_percpu_data(rdp);
3005}
3006
3007/*
3008 * Invoked early in the CPU-online process, when pretty much all services
3009 * are available. The incoming CPU is not present.
3010 *
3011 * Initializes a CPU's per-CPU RCU data. Note that only one online or
3012 * offline event can be happening at a given time. Note also that we can
3013 * accept some slop in the rsp->gp_seq access due to the fact that this
3014 * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
3015 * And any offloaded callbacks are being numbered elsewhere.
3016 */
3017int rcutree_prepare_cpu(unsigned int cpu)
3018{
3019 unsigned long flags;
3020 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3021 struct rcu_node *rnp = rcu_get_root();
3022
3023 /* Set up local state, ensuring consistent view of global state. */
3024 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3025 rdp->qlen_last_fqs_check = 0;
3026 rdp->n_force_qs_snap = rcu_state.n_force_qs;
3027 rdp->blimit = blimit;
3028 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3029 !rcu_segcblist_is_offloaded(&rdp->cblist))
3030 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
3031 rdp->dynticks_nesting = 1; /* CPU not up, no tearing. */
3032 rcu_dynticks_eqs_online();
3033 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3034
3035 /*
3036 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3037 * propagation up the rcu_node tree will happen at the beginning
3038 * of the next grace period.
3039 */
3040 rnp = rdp->mynode;
3041 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3042 rdp->beenonline = true; /* We have now been online. */
3043 rdp->gp_seq = rnp->gp_seq;
3044 rdp->gp_seq_needed = rnp->gp_seq;
3045 rdp->cpu_no_qs.b.norm = true;
3046 rdp->core_needs_qs = false;
3047 rdp->rcu_iw_pending = false;
3048 rdp->rcu_iw_gp_seq = rnp->gp_seq - 1;
3049 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
3050 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3051 rcu_prepare_kthreads(cpu);
3052 rcu_spawn_cpu_nocb_kthread(cpu);
3053
3054 return 0;
3055}
3056
3057/*
3058 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3059 */
3060static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3061{
3062 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3063
3064 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3065}
3066
3067/*
3068 * Near the end of the CPU-online process. Pretty much all services
3069 * enabled, and the CPU is now very much alive.
3070 */
3071int rcutree_online_cpu(unsigned int cpu)
3072{
3073 unsigned long flags;
3074 struct rcu_data *rdp;
3075 struct rcu_node *rnp;
3076
3077 rdp = per_cpu_ptr(&rcu_data, cpu);
3078 rnp = rdp->mynode;
3079 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3080 rnp->ffmask |= rdp->grpmask;
3081 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3082 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
3083 return 0; /* Too early in boot for scheduler work. */
3084 sync_sched_exp_online_cleanup(cpu);
3085 rcutree_affinity_setting(cpu, -1);
3086 return 0;
3087}
3088
3089/*
3090 * Near the beginning of the process. The CPU is still very much alive
3091 * with pretty much all services enabled.
3092 */
3093int rcutree_offline_cpu(unsigned int cpu)
3094{
3095 unsigned long flags;
3096 struct rcu_data *rdp;
3097 struct rcu_node *rnp;
3098
3099 rdp = per_cpu_ptr(&rcu_data, cpu);
3100 rnp = rdp->mynode;
3101 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3102 rnp->ffmask &= ~rdp->grpmask;
3103 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3104
3105 rcutree_affinity_setting(cpu, cpu);
3106 return 0;
3107}
3108
3109static DEFINE_PER_CPU(int, rcu_cpu_started);
3110
3111/*
3112 * Mark the specified CPU as being online so that subsequent grace periods
3113 * (both expedited and normal) will wait on it. Note that this means that
3114 * incoming CPUs are not allowed to use RCU read-side critical sections
3115 * until this function is called. Failing to observe this restriction
3116 * will result in lockdep splats.
3117 *
3118 * Note that this function is special in that it is invoked directly
3119 * from the incoming CPU rather than from the cpuhp_step mechanism.
3120 * This is because this function must be invoked at a precise location.
3121 */
3122void rcu_cpu_starting(unsigned int cpu)
3123{
3124 unsigned long flags;
3125 unsigned long mask;
3126 int nbits;
3127 unsigned long oldmask;
3128 struct rcu_data *rdp;
3129 struct rcu_node *rnp;
3130
3131 if (per_cpu(rcu_cpu_started, cpu))
3132 return;
3133
3134 per_cpu(rcu_cpu_started, cpu) = 1;
3135
3136 rdp = per_cpu_ptr(&rcu_data, cpu);
3137 rnp = rdp->mynode;
3138 mask = rdp->grpmask;
3139 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3140 rnp->qsmaskinitnext |= mask;
3141 oldmask = rnp->expmaskinitnext;
3142 rnp->expmaskinitnext |= mask;
3143 oldmask ^= rnp->expmaskinitnext;
3144 nbits = bitmap_weight(&oldmask, BITS_PER_LONG);
3145 /* Allow lockless access for expedited grace periods. */
3146 smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + nbits); /* ^^^ */
3147 rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
3148 rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
3149 rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
3150 if (rnp->qsmask & mask) { /* RCU waiting on incoming CPU? */
3151 /* Report QS -after- changing ->qsmaskinitnext! */
3152 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
3153 } else {
3154 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3155 }
3156 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
3157}
3158
3159#ifdef CONFIG_HOTPLUG_CPU
3160/*
3161 * The outgoing function has no further need of RCU, so remove it from
3162 * the rcu_node tree's ->qsmaskinitnext bit masks.
3163 *
3164 * Note that this function is special in that it is invoked directly
3165 * from the outgoing CPU rather than from the cpuhp_step mechanism.
3166 * This is because this function must be invoked at a precise location.
3167 */
3168void rcu_report_dead(unsigned int cpu)
3169{
3170 unsigned long flags;
3171 unsigned long mask;
3172 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3173 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3174
3175 /* QS for any half-done expedited grace period. */
3176 preempt_disable();
3177 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
3178 preempt_enable();
3179 rcu_preempt_deferred_qs(current);
3180
3181 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3182 mask = rdp->grpmask;
3183 raw_spin_lock(&rcu_state.ofl_lock);
3184 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3185 rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
3186 rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
3187 if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
3188 /* Report quiescent state -before- changing ->qsmaskinitnext! */
3189 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
3190 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3191 }
3192 rnp->qsmaskinitnext &= ~mask;
3193 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3194 raw_spin_unlock(&rcu_state.ofl_lock);
3195
3196 per_cpu(rcu_cpu_started, cpu) = 0;
3197}
3198
3199/*
3200 * The outgoing CPU has just passed through the dying-idle state, and we
3201 * are being invoked from the CPU that was IPIed to continue the offline
3202 * operation. Migrate the outgoing CPU's callbacks to the current CPU.
3203 */
3204void rcutree_migrate_callbacks(int cpu)
3205{
3206 unsigned long flags;
3207 struct rcu_data *my_rdp;
3208 struct rcu_node *my_rnp;
3209 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3210 bool needwake;
3211
3212 if (rcu_segcblist_is_offloaded(&rdp->cblist) ||
3213 rcu_segcblist_empty(&rdp->cblist))
3214 return; /* No callbacks to migrate. */
3215
3216 local_irq_save(flags);
3217 my_rdp = this_cpu_ptr(&rcu_data);
3218 my_rnp = my_rdp->mynode;
3219 rcu_nocb_lock(my_rdp); /* irqs already disabled. */
3220 WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies));
3221 raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
3222 /* Leverage recent GPs and set GP for new callbacks. */
3223 needwake = rcu_advance_cbs(my_rnp, rdp) ||
3224 rcu_advance_cbs(my_rnp, my_rdp);
3225 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
3226 needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
3227 rcu_segcblist_disable(&rdp->cblist);
3228 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
3229 !rcu_segcblist_n_cbs(&my_rdp->cblist));
3230 if (rcu_segcblist_is_offloaded(&my_rdp->cblist)) {
3231 raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
3232 __call_rcu_nocb_wake(my_rdp, true, flags);
3233 } else {
3234 rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
3235 raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags);
3236 }
3237 if (needwake)
3238 rcu_gp_kthread_wake();
3239 lockdep_assert_irqs_enabled();
3240 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
3241 !rcu_segcblist_empty(&rdp->cblist),
3242 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
3243 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
3244 rcu_segcblist_first_cb(&rdp->cblist));
3245}
3246#endif
3247
3248/*
3249 * On non-huge systems, use expedited RCU grace periods to make suspend
3250 * and hibernation run faster.
3251 */
3252static int rcu_pm_notify(struct notifier_block *self,
3253 unsigned long action, void *hcpu)
3254{
3255 switch (action) {
3256 case PM_HIBERNATION_PREPARE:
3257 case PM_SUSPEND_PREPARE:
3258 rcu_expedite_gp();
3259 break;
3260 case PM_POST_HIBERNATION:
3261 case PM_POST_SUSPEND:
3262 rcu_unexpedite_gp();
3263 break;
3264 default:
3265 break;
3266 }
3267 return NOTIFY_OK;
3268}
3269
3270/*
3271 * Spawn the kthreads that handle RCU's grace periods.
3272 */
3273static int __init rcu_spawn_gp_kthread(void)
3274{
3275 unsigned long flags;
3276 int kthread_prio_in = kthread_prio;
3277 struct rcu_node *rnp;
3278 struct sched_param sp;
3279 struct task_struct *t;
3280
3281 /* Force priority into range. */
3282 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
3283 && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
3284 kthread_prio = 2;
3285 else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
3286 kthread_prio = 1;
3287 else if (kthread_prio < 0)
3288 kthread_prio = 0;
3289 else if (kthread_prio > 99)
3290 kthread_prio = 99;
3291
3292 if (kthread_prio != kthread_prio_in)
3293 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3294 kthread_prio, kthread_prio_in);
3295
3296 rcu_scheduler_fully_active = 1;
3297 t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
3298 if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
3299 return 0;
3300 if (kthread_prio) {
3301 sp.sched_priority = kthread_prio;
3302 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3303 }
3304 rnp = rcu_get_root();
3305 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3306 rcu_state.gp_kthread = t;
3307 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3308 wake_up_process(t);
3309 rcu_spawn_nocb_kthreads();
3310 rcu_spawn_boost_kthreads();
3311 return 0;
3312}
3313early_initcall(rcu_spawn_gp_kthread);
3314
3315/*
3316 * This function is invoked towards the end of the scheduler's
3317 * initialization process. Before this is called, the idle task might
3318 * contain synchronous grace-period primitives (during which time, this idle
3319 * task is booting the system, and such primitives are no-ops). After this
3320 * function is called, any synchronous grace-period primitives are run as
3321 * expedited, with the requesting task driving the grace period forward.
3322 * A later core_initcall() rcu_set_runtime_mode() will switch to full
3323 * runtime RCU functionality.
3324 */
3325void rcu_scheduler_starting(void)
3326{
3327 WARN_ON(num_online_cpus() != 1);
3328 WARN_ON(nr_context_switches() > 0);
3329 rcu_test_sync_prims();
3330 rcu_scheduler_active = RCU_SCHEDULER_INIT;
3331 rcu_test_sync_prims();
3332}
3333
3334/*
3335 * Helper function for rcu_init() that initializes the rcu_state structure.
3336 */
3337static void __init rcu_init_one(void)
3338{
3339 static const char * const buf[] = RCU_NODE_NAME_INIT;
3340 static const char * const fqs[] = RCU_FQS_NAME_INIT;
3341 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
3342 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
3343
3344 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
3345 int cpustride = 1;
3346 int i;
3347 int j;
3348 struct rcu_node *rnp;
3349
3350 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3351
3352 /* Silence gcc 4.8 false positive about array index out of range. */
3353 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
3354 panic("rcu_init_one: rcu_num_lvls out of range");
3355
3356 /* Initialize the level-tracking arrays. */
3357
3358 for (i = 1; i < rcu_num_lvls; i++)
3359 rcu_state.level[i] =
3360 rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
3361 rcu_init_levelspread(levelspread, num_rcu_lvl);
3362
3363 /* Initialize the elements themselves, starting from the leaves. */
3364
3365 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3366 cpustride *= levelspread[i];
3367 rnp = rcu_state.level[i];
3368 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
3369 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
3370 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
3371 &rcu_node_class[i], buf[i]);
3372 raw_spin_lock_init(&rnp->fqslock);
3373 lockdep_set_class_and_name(&rnp->fqslock,
3374 &rcu_fqs_class[i], fqs[i]);
3375 rnp->gp_seq = rcu_state.gp_seq;
3376 rnp->gp_seq_needed = rcu_state.gp_seq;
3377 rnp->completedqs = rcu_state.gp_seq;
3378 rnp->qsmask = 0;
3379 rnp->qsmaskinit = 0;
3380 rnp->grplo = j * cpustride;
3381 rnp->grphi = (j + 1) * cpustride - 1;
3382 if (rnp->grphi >= nr_cpu_ids)
3383 rnp->grphi = nr_cpu_ids - 1;
3384 if (i == 0) {
3385 rnp->grpnum = 0;
3386 rnp->grpmask = 0;
3387 rnp->parent = NULL;
3388 } else {
3389 rnp->grpnum = j % levelspread[i - 1];
3390 rnp->grpmask = BIT(rnp->grpnum);
3391 rnp->parent = rcu_state.level[i - 1] +
3392 j / levelspread[i - 1];
3393 }
3394 rnp->level = i;
3395 INIT_LIST_HEAD(&rnp->blkd_tasks);
3396 rcu_init_one_nocb(rnp);
3397 init_waitqueue_head(&rnp->exp_wq[0]);
3398 init_waitqueue_head(&rnp->exp_wq[1]);
3399 init_waitqueue_head(&rnp->exp_wq[2]);
3400 init_waitqueue_head(&rnp->exp_wq[3]);
3401 spin_lock_init(&rnp->exp_lock);
3402 }
3403 }
3404
3405 init_swait_queue_head(&rcu_state.gp_wq);
3406 init_swait_queue_head(&rcu_state.expedited_wq);
3407 rnp = rcu_first_leaf_node();
3408 for_each_possible_cpu(i) {
3409 while (i > rnp->grphi)
3410 rnp++;
3411 per_cpu_ptr(&rcu_data, i)->mynode = rnp;
3412 rcu_boot_init_percpu_data(i);
3413 }
3414}
3415
3416/*
3417 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3418 * replace the definitions in tree.h because those are needed to size
3419 * the ->node array in the rcu_state structure.
3420 */
3421static void __init rcu_init_geometry(void)
3422{
3423 ulong d;
3424 int i;
3425 int rcu_capacity[RCU_NUM_LVLS];
3426
3427 /*
3428 * Initialize any unspecified boot parameters.
3429 * The default values of jiffies_till_first_fqs and
3430 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3431 * value, which is a function of HZ, then adding one for each
3432 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3433 */
3434 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3435 if (jiffies_till_first_fqs == ULONG_MAX)
3436 jiffies_till_first_fqs = d;
3437 if (jiffies_till_next_fqs == ULONG_MAX)
3438 jiffies_till_next_fqs = d;
3439 adjust_jiffies_till_sched_qs();
3440
3441 /* If the compile-time values are accurate, just leave. */
3442 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
3443 nr_cpu_ids == NR_CPUS)
3444 return;
3445 pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
3446 rcu_fanout_leaf, nr_cpu_ids);
3447
3448 /*
3449 * The boot-time rcu_fanout_leaf parameter must be at least two
3450 * and cannot exceed the number of bits in the rcu_node masks.
3451 * Complain and fall back to the compile-time values if this
3452 * limit is exceeded.
3453 */
3454 if (rcu_fanout_leaf < 2 ||
3455 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
3456 rcu_fanout_leaf = RCU_FANOUT_LEAF;
3457 WARN_ON(1);
3458 return;
3459 }
3460
3461 /*
3462 * Compute number of nodes that can be handled an rcu_node tree
3463 * with the given number of levels.
3464 */
3465 rcu_capacity[0] = rcu_fanout_leaf;
3466 for (i = 1; i < RCU_NUM_LVLS; i++)
3467 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
3468
3469 /*
3470 * The tree must be able to accommodate the configured number of CPUs.
3471 * If this limit is exceeded, fall back to the compile-time values.
3472 */
3473 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
3474 rcu_fanout_leaf = RCU_FANOUT_LEAF;
3475 WARN_ON(1);
3476 return;
3477 }
3478
3479 /* Calculate the number of levels in the tree. */
3480 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
3481 }
3482 rcu_num_lvls = i + 1;
3483
3484 /* Calculate the number of rcu_nodes at each level of the tree. */
3485 for (i = 0; i < rcu_num_lvls; i++) {
3486 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
3487 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
3488 }
3489
3490 /* Calculate the total number of rcu_node structures. */
3491 rcu_num_nodes = 0;
3492 for (i = 0; i < rcu_num_lvls; i++)
3493 rcu_num_nodes += num_rcu_lvl[i];
3494}
3495
3496/*
3497 * Dump out the structure of the rcu_node combining tree associated
3498 * with the rcu_state structure.
3499 */
3500static void __init rcu_dump_rcu_node_tree(void)
3501{
3502 int level = 0;
3503 struct rcu_node *rnp;
3504
3505 pr_info("rcu_node tree layout dump\n");
3506 pr_info(" ");
3507 rcu_for_each_node_breadth_first(rnp) {
3508 if (rnp->level != level) {
3509 pr_cont("\n");
3510 pr_info(" ");
3511 level = rnp->level;
3512 }
3513 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
3514 }
3515 pr_cont("\n");
3516}
3517
3518struct workqueue_struct *rcu_gp_wq;
3519struct workqueue_struct *rcu_par_gp_wq;
3520
3521void __init rcu_init(void)
3522{
3523 int cpu;
3524
3525 rcu_early_boot_tests();
3526
3527 rcu_bootup_announce();
3528 rcu_init_geometry();
3529 rcu_init_one();
3530 if (dump_tree)
3531 rcu_dump_rcu_node_tree();
3532 if (use_softirq)
3533 open_softirq(RCU_SOFTIRQ, rcu_core_si);
3534
3535 /*
3536 * We don't need protection against CPU-hotplug here because
3537 * this is called early in boot, before either interrupts
3538 * or the scheduler are operational.
3539 */
3540 pm_notifier(rcu_pm_notify, 0);
3541 for_each_online_cpu(cpu) {
3542 rcutree_prepare_cpu(cpu);
3543 rcu_cpu_starting(cpu);
3544 rcutree_online_cpu(cpu);
3545 }
3546
3547 /* Create workqueue for expedited GPs and for Tree SRCU. */
3548 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
3549 WARN_ON(!rcu_gp_wq);
3550 rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
3551 WARN_ON(!rcu_par_gp_wq);
3552 srcu_init();
3553}
3554
3555#include "tree_stall.h"
3556#include "tree_exp.h"
3557#include "tree_plugin.h"
1/*
2 * Read-Copy Update mechanism for mutual exclusion
3 *
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
17 *
18 * Copyright IBM Corporation, 2008
19 *
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
23 *
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
26 *
27 * For detailed explanation of Read-Copy Update mechanism see -
28 * Documentation/RCU
29 */
30#include <linux/types.h>
31#include <linux/kernel.h>
32#include <linux/init.h>
33#include <linux/spinlock.h>
34#include <linux/smp.h>
35#include <linux/rcupdate.h>
36#include <linux/interrupt.h>
37#include <linux/sched.h>
38#include <linux/nmi.h>
39#include <linux/atomic.h>
40#include <linux/bitops.h>
41#include <linux/export.h>
42#include <linux/completion.h>
43#include <linux/moduleparam.h>
44#include <linux/module.h>
45#include <linux/percpu.h>
46#include <linux/notifier.h>
47#include <linux/cpu.h>
48#include <linux/mutex.h>
49#include <linux/time.h>
50#include <linux/kernel_stat.h>
51#include <linux/wait.h>
52#include <linux/kthread.h>
53#include <linux/prefetch.h>
54#include <linux/delay.h>
55#include <linux/stop_machine.h>
56#include <linux/random.h>
57#include <linux/ftrace_event.h>
58#include <linux/suspend.h>
59
60#include "tree.h"
61#include "rcu.h"
62
63MODULE_ALIAS("rcutree");
64#ifdef MODULE_PARAM_PREFIX
65#undef MODULE_PARAM_PREFIX
66#endif
67#define MODULE_PARAM_PREFIX "rcutree."
68
69/* Data structures. */
70
71static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
72static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
73
74/*
75 * In order to export the rcu_state name to the tracing tools, it
76 * needs to be added in the __tracepoint_string section.
77 * This requires defining a separate variable tp_<sname>_varname
78 * that points to the string being used, and this will allow
79 * the tracing userspace tools to be able to decipher the string
80 * address to the matching string.
81 */
82#define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
83static char sname##_varname[] = #sname; \
84static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \
85struct rcu_state sname##_state = { \
86 .level = { &sname##_state.node[0] }, \
87 .call = cr, \
88 .fqs_state = RCU_GP_IDLE, \
89 .gpnum = 0UL - 300UL, \
90 .completed = 0UL - 300UL, \
91 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
92 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
93 .orphan_donetail = &sname##_state.orphan_donelist, \
94 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
95 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
96 .name = sname##_varname, \
97 .abbr = sabbr, \
98}; \
99DEFINE_PER_CPU(struct rcu_data, sname##_data)
100
101RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
102RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
103
104static struct rcu_state *rcu_state;
105LIST_HEAD(rcu_struct_flavors);
106
107/* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
108static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
109module_param(rcu_fanout_leaf, int, 0444);
110int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
111static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
112 NUM_RCU_LVL_0,
113 NUM_RCU_LVL_1,
114 NUM_RCU_LVL_2,
115 NUM_RCU_LVL_3,
116 NUM_RCU_LVL_4,
117};
118int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
119
120/*
121 * The rcu_scheduler_active variable transitions from zero to one just
122 * before the first task is spawned. So when this variable is zero, RCU
123 * can assume that there is but one task, allowing RCU to (for example)
124 * optimize synchronize_sched() to a simple barrier(). When this variable
125 * is one, RCU must actually do all the hard work required to detect real
126 * grace periods. This variable is also used to suppress boot-time false
127 * positives from lockdep-RCU error checking.
128 */
129int rcu_scheduler_active __read_mostly;
130EXPORT_SYMBOL_GPL(rcu_scheduler_active);
131
132/*
133 * The rcu_scheduler_fully_active variable transitions from zero to one
134 * during the early_initcall() processing, which is after the scheduler
135 * is capable of creating new tasks. So RCU processing (for example,
136 * creating tasks for RCU priority boosting) must be delayed until after
137 * rcu_scheduler_fully_active transitions from zero to one. We also
138 * currently delay invocation of any RCU callbacks until after this point.
139 *
140 * It might later prove better for people registering RCU callbacks during
141 * early boot to take responsibility for these callbacks, but one step at
142 * a time.
143 */
144static int rcu_scheduler_fully_active __read_mostly;
145
146#ifdef CONFIG_RCU_BOOST
147
148/*
149 * Control variables for per-CPU and per-rcu_node kthreads. These
150 * handle all flavors of RCU.
151 */
152static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
153DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
154DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
155DEFINE_PER_CPU(char, rcu_cpu_has_work);
156
157#endif /* #ifdef CONFIG_RCU_BOOST */
158
159static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
160static void invoke_rcu_core(void);
161static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
162
163/*
164 * Track the rcutorture test sequence number and the update version
165 * number within a given test. The rcutorture_testseq is incremented
166 * on every rcutorture module load and unload, so has an odd value
167 * when a test is running. The rcutorture_vernum is set to zero
168 * when rcutorture starts and is incremented on each rcutorture update.
169 * These variables enable correlating rcutorture output with the
170 * RCU tracing information.
171 */
172unsigned long rcutorture_testseq;
173unsigned long rcutorture_vernum;
174
175/*
176 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
177 * permit this function to be invoked without holding the root rcu_node
178 * structure's ->lock, but of course results can be subject to change.
179 */
180static int rcu_gp_in_progress(struct rcu_state *rsp)
181{
182 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
183}
184
185/*
186 * Note a quiescent state. Because we do not need to know
187 * how many quiescent states passed, just if there was at least
188 * one since the start of the grace period, this just sets a flag.
189 * The caller must have disabled preemption.
190 */
191void rcu_sched_qs(int cpu)
192{
193 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
194
195 if (rdp->passed_quiesce == 0)
196 trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs"));
197 rdp->passed_quiesce = 1;
198}
199
200void rcu_bh_qs(int cpu)
201{
202 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
203
204 if (rdp->passed_quiesce == 0)
205 trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs"));
206 rdp->passed_quiesce = 1;
207}
208
209/*
210 * Note a context switch. This is a quiescent state for RCU-sched,
211 * and requires special handling for preemptible RCU.
212 * The caller must have disabled preemption.
213 */
214void rcu_note_context_switch(int cpu)
215{
216 trace_rcu_utilization(TPS("Start context switch"));
217 rcu_sched_qs(cpu);
218 rcu_preempt_note_context_switch(cpu);
219 trace_rcu_utilization(TPS("End context switch"));
220}
221EXPORT_SYMBOL_GPL(rcu_note_context_switch);
222
223static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
224 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
225 .dynticks = ATOMIC_INIT(1),
226#ifdef CONFIG_NO_HZ_FULL_SYSIDLE
227 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
228 .dynticks_idle = ATOMIC_INIT(1),
229#endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
230};
231
232static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
233static long qhimark = 10000; /* If this many pending, ignore blimit. */
234static long qlowmark = 100; /* Once only this many pending, use blimit. */
235
236module_param(blimit, long, 0444);
237module_param(qhimark, long, 0444);
238module_param(qlowmark, long, 0444);
239
240static ulong jiffies_till_first_fqs = ULONG_MAX;
241static ulong jiffies_till_next_fqs = ULONG_MAX;
242
243module_param(jiffies_till_first_fqs, ulong, 0644);
244module_param(jiffies_till_next_fqs, ulong, 0644);
245
246static void rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
247 struct rcu_data *rdp);
248static void force_qs_rnp(struct rcu_state *rsp,
249 int (*f)(struct rcu_data *rsp, bool *isidle,
250 unsigned long *maxj),
251 bool *isidle, unsigned long *maxj);
252static void force_quiescent_state(struct rcu_state *rsp);
253static int rcu_pending(int cpu);
254
255/*
256 * Return the number of RCU-sched batches processed thus far for debug & stats.
257 */
258long rcu_batches_completed_sched(void)
259{
260 return rcu_sched_state.completed;
261}
262EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
263
264/*
265 * Return the number of RCU BH batches processed thus far for debug & stats.
266 */
267long rcu_batches_completed_bh(void)
268{
269 return rcu_bh_state.completed;
270}
271EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
272
273/*
274 * Force a quiescent state for RCU BH.
275 */
276void rcu_bh_force_quiescent_state(void)
277{
278 force_quiescent_state(&rcu_bh_state);
279}
280EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
281
282/*
283 * Record the number of times rcutorture tests have been initiated and
284 * terminated. This information allows the debugfs tracing stats to be
285 * correlated to the rcutorture messages, even when the rcutorture module
286 * is being repeatedly loaded and unloaded. In other words, we cannot
287 * store this state in rcutorture itself.
288 */
289void rcutorture_record_test_transition(void)
290{
291 rcutorture_testseq++;
292 rcutorture_vernum = 0;
293}
294EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
295
296/*
297 * Record the number of writer passes through the current rcutorture test.
298 * This is also used to correlate debugfs tracing stats with the rcutorture
299 * messages.
300 */
301void rcutorture_record_progress(unsigned long vernum)
302{
303 rcutorture_vernum++;
304}
305EXPORT_SYMBOL_GPL(rcutorture_record_progress);
306
307/*
308 * Force a quiescent state for RCU-sched.
309 */
310void rcu_sched_force_quiescent_state(void)
311{
312 force_quiescent_state(&rcu_sched_state);
313}
314EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
315
316/*
317 * Does the CPU have callbacks ready to be invoked?
318 */
319static int
320cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
321{
322 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
323 rdp->nxttail[RCU_DONE_TAIL] != NULL;
324}
325
326/*
327 * Does the current CPU require a not-yet-started grace period?
328 * The caller must have disabled interrupts to prevent races with
329 * normal callback registry.
330 */
331static int
332cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
333{
334 int i;
335
336 if (rcu_gp_in_progress(rsp))
337 return 0; /* No, a grace period is already in progress. */
338 if (rcu_nocb_needs_gp(rsp))
339 return 1; /* Yes, a no-CBs CPU needs one. */
340 if (!rdp->nxttail[RCU_NEXT_TAIL])
341 return 0; /* No, this is a no-CBs (or offline) CPU. */
342 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
343 return 1; /* Yes, this CPU has newly registered callbacks. */
344 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
345 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
346 ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
347 rdp->nxtcompleted[i]))
348 return 1; /* Yes, CBs for future grace period. */
349 return 0; /* No grace period needed. */
350}
351
352/*
353 * Return the root node of the specified rcu_state structure.
354 */
355static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
356{
357 return &rsp->node[0];
358}
359
360/*
361 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
362 *
363 * If the new value of the ->dynticks_nesting counter now is zero,
364 * we really have entered idle, and must do the appropriate accounting.
365 * The caller must have disabled interrupts.
366 */
367static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
368 bool user)
369{
370 struct rcu_state *rsp;
371 struct rcu_data *rdp;
372
373 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
374 if (!user && !is_idle_task(current)) {
375 struct task_struct *idle __maybe_unused =
376 idle_task(smp_processor_id());
377
378 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
379 ftrace_dump(DUMP_ORIG);
380 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
381 current->pid, current->comm,
382 idle->pid, idle->comm); /* must be idle task! */
383 }
384 for_each_rcu_flavor(rsp) {
385 rdp = this_cpu_ptr(rsp->rda);
386 do_nocb_deferred_wakeup(rdp);
387 }
388 rcu_prepare_for_idle(smp_processor_id());
389 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
390 smp_mb__before_atomic_inc(); /* See above. */
391 atomic_inc(&rdtp->dynticks);
392 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
393 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
394
395 /*
396 * It is illegal to enter an extended quiescent state while
397 * in an RCU read-side critical section.
398 */
399 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
400 "Illegal idle entry in RCU read-side critical section.");
401 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
402 "Illegal idle entry in RCU-bh read-side critical section.");
403 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
404 "Illegal idle entry in RCU-sched read-side critical section.");
405}
406
407/*
408 * Enter an RCU extended quiescent state, which can be either the
409 * idle loop or adaptive-tickless usermode execution.
410 */
411static void rcu_eqs_enter(bool user)
412{
413 long long oldval;
414 struct rcu_dynticks *rdtp;
415
416 rdtp = this_cpu_ptr(&rcu_dynticks);
417 oldval = rdtp->dynticks_nesting;
418 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
419 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
420 rdtp->dynticks_nesting = 0;
421 rcu_eqs_enter_common(rdtp, oldval, user);
422 } else {
423 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
424 }
425}
426
427/**
428 * rcu_idle_enter - inform RCU that current CPU is entering idle
429 *
430 * Enter idle mode, in other words, -leave- the mode in which RCU
431 * read-side critical sections can occur. (Though RCU read-side
432 * critical sections can occur in irq handlers in idle, a possibility
433 * handled by irq_enter() and irq_exit().)
434 *
435 * We crowbar the ->dynticks_nesting field to zero to allow for
436 * the possibility of usermode upcalls having messed up our count
437 * of interrupt nesting level during the prior busy period.
438 */
439void rcu_idle_enter(void)
440{
441 unsigned long flags;
442
443 local_irq_save(flags);
444 rcu_eqs_enter(false);
445 rcu_sysidle_enter(this_cpu_ptr(&rcu_dynticks), 0);
446 local_irq_restore(flags);
447}
448EXPORT_SYMBOL_GPL(rcu_idle_enter);
449
450#ifdef CONFIG_RCU_USER_QS
451/**
452 * rcu_user_enter - inform RCU that we are resuming userspace.
453 *
454 * Enter RCU idle mode right before resuming userspace. No use of RCU
455 * is permitted between this call and rcu_user_exit(). This way the
456 * CPU doesn't need to maintain the tick for RCU maintenance purposes
457 * when the CPU runs in userspace.
458 */
459void rcu_user_enter(void)
460{
461 rcu_eqs_enter(1);
462}
463#endif /* CONFIG_RCU_USER_QS */
464
465/**
466 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
467 *
468 * Exit from an interrupt handler, which might possibly result in entering
469 * idle mode, in other words, leaving the mode in which read-side critical
470 * sections can occur.
471 *
472 * This code assumes that the idle loop never does anything that might
473 * result in unbalanced calls to irq_enter() and irq_exit(). If your
474 * architecture violates this assumption, RCU will give you what you
475 * deserve, good and hard. But very infrequently and irreproducibly.
476 *
477 * Use things like work queues to work around this limitation.
478 *
479 * You have been warned.
480 */
481void rcu_irq_exit(void)
482{
483 unsigned long flags;
484 long long oldval;
485 struct rcu_dynticks *rdtp;
486
487 local_irq_save(flags);
488 rdtp = this_cpu_ptr(&rcu_dynticks);
489 oldval = rdtp->dynticks_nesting;
490 rdtp->dynticks_nesting--;
491 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
492 if (rdtp->dynticks_nesting)
493 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
494 else
495 rcu_eqs_enter_common(rdtp, oldval, true);
496 rcu_sysidle_enter(rdtp, 1);
497 local_irq_restore(flags);
498}
499
500/*
501 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
502 *
503 * If the new value of the ->dynticks_nesting counter was previously zero,
504 * we really have exited idle, and must do the appropriate accounting.
505 * The caller must have disabled interrupts.
506 */
507static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
508 int user)
509{
510 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
511 atomic_inc(&rdtp->dynticks);
512 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
513 smp_mb__after_atomic_inc(); /* See above. */
514 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
515 rcu_cleanup_after_idle(smp_processor_id());
516 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
517 if (!user && !is_idle_task(current)) {
518 struct task_struct *idle __maybe_unused =
519 idle_task(smp_processor_id());
520
521 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
522 oldval, rdtp->dynticks_nesting);
523 ftrace_dump(DUMP_ORIG);
524 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
525 current->pid, current->comm,
526 idle->pid, idle->comm); /* must be idle task! */
527 }
528}
529
530/*
531 * Exit an RCU extended quiescent state, which can be either the
532 * idle loop or adaptive-tickless usermode execution.
533 */
534static void rcu_eqs_exit(bool user)
535{
536 struct rcu_dynticks *rdtp;
537 long long oldval;
538
539 rdtp = this_cpu_ptr(&rcu_dynticks);
540 oldval = rdtp->dynticks_nesting;
541 WARN_ON_ONCE(oldval < 0);
542 if (oldval & DYNTICK_TASK_NEST_MASK) {
543 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
544 } else {
545 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
546 rcu_eqs_exit_common(rdtp, oldval, user);
547 }
548}
549
550/**
551 * rcu_idle_exit - inform RCU that current CPU is leaving idle
552 *
553 * Exit idle mode, in other words, -enter- the mode in which RCU
554 * read-side critical sections can occur.
555 *
556 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
557 * allow for the possibility of usermode upcalls messing up our count
558 * of interrupt nesting level during the busy period that is just
559 * now starting.
560 */
561void rcu_idle_exit(void)
562{
563 unsigned long flags;
564
565 local_irq_save(flags);
566 rcu_eqs_exit(false);
567 rcu_sysidle_exit(this_cpu_ptr(&rcu_dynticks), 0);
568 local_irq_restore(flags);
569}
570EXPORT_SYMBOL_GPL(rcu_idle_exit);
571
572#ifdef CONFIG_RCU_USER_QS
573/**
574 * rcu_user_exit - inform RCU that we are exiting userspace.
575 *
576 * Exit RCU idle mode while entering the kernel because it can
577 * run a RCU read side critical section anytime.
578 */
579void rcu_user_exit(void)
580{
581 rcu_eqs_exit(1);
582}
583#endif /* CONFIG_RCU_USER_QS */
584
585/**
586 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
587 *
588 * Enter an interrupt handler, which might possibly result in exiting
589 * idle mode, in other words, entering the mode in which read-side critical
590 * sections can occur.
591 *
592 * Note that the Linux kernel is fully capable of entering an interrupt
593 * handler that it never exits, for example when doing upcalls to
594 * user mode! This code assumes that the idle loop never does upcalls to
595 * user mode. If your architecture does do upcalls from the idle loop (or
596 * does anything else that results in unbalanced calls to the irq_enter()
597 * and irq_exit() functions), RCU will give you what you deserve, good
598 * and hard. But very infrequently and irreproducibly.
599 *
600 * Use things like work queues to work around this limitation.
601 *
602 * You have been warned.
603 */
604void rcu_irq_enter(void)
605{
606 unsigned long flags;
607 struct rcu_dynticks *rdtp;
608 long long oldval;
609
610 local_irq_save(flags);
611 rdtp = this_cpu_ptr(&rcu_dynticks);
612 oldval = rdtp->dynticks_nesting;
613 rdtp->dynticks_nesting++;
614 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
615 if (oldval)
616 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
617 else
618 rcu_eqs_exit_common(rdtp, oldval, true);
619 rcu_sysidle_exit(rdtp, 1);
620 local_irq_restore(flags);
621}
622
623/**
624 * rcu_nmi_enter - inform RCU of entry to NMI context
625 *
626 * If the CPU was idle with dynamic ticks active, and there is no
627 * irq handler running, this updates rdtp->dynticks_nmi to let the
628 * RCU grace-period handling know that the CPU is active.
629 */
630void rcu_nmi_enter(void)
631{
632 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
633
634 if (rdtp->dynticks_nmi_nesting == 0 &&
635 (atomic_read(&rdtp->dynticks) & 0x1))
636 return;
637 rdtp->dynticks_nmi_nesting++;
638 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
639 atomic_inc(&rdtp->dynticks);
640 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
641 smp_mb__after_atomic_inc(); /* See above. */
642 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
643}
644
645/**
646 * rcu_nmi_exit - inform RCU of exit from NMI context
647 *
648 * If the CPU was idle with dynamic ticks active, and there is no
649 * irq handler running, this updates rdtp->dynticks_nmi to let the
650 * RCU grace-period handling know that the CPU is no longer active.
651 */
652void rcu_nmi_exit(void)
653{
654 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
655
656 if (rdtp->dynticks_nmi_nesting == 0 ||
657 --rdtp->dynticks_nmi_nesting != 0)
658 return;
659 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
660 smp_mb__before_atomic_inc(); /* See above. */
661 atomic_inc(&rdtp->dynticks);
662 smp_mb__after_atomic_inc(); /* Force delay to next write. */
663 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
664}
665
666/**
667 * __rcu_is_watching - are RCU read-side critical sections safe?
668 *
669 * Return true if RCU is watching the running CPU, which means that
670 * this CPU can safely enter RCU read-side critical sections. Unlike
671 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
672 * least disabled preemption.
673 */
674bool notrace __rcu_is_watching(void)
675{
676 return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
677}
678
679/**
680 * rcu_is_watching - see if RCU thinks that the current CPU is idle
681 *
682 * If the current CPU is in its idle loop and is neither in an interrupt
683 * or NMI handler, return true.
684 */
685bool notrace rcu_is_watching(void)
686{
687 int ret;
688
689 preempt_disable();
690 ret = __rcu_is_watching();
691 preempt_enable();
692 return ret;
693}
694EXPORT_SYMBOL_GPL(rcu_is_watching);
695
696#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
697
698/*
699 * Is the current CPU online? Disable preemption to avoid false positives
700 * that could otherwise happen due to the current CPU number being sampled,
701 * this task being preempted, its old CPU being taken offline, resuming
702 * on some other CPU, then determining that its old CPU is now offline.
703 * It is OK to use RCU on an offline processor during initial boot, hence
704 * the check for rcu_scheduler_fully_active. Note also that it is OK
705 * for a CPU coming online to use RCU for one jiffy prior to marking itself
706 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
707 * offline to continue to use RCU for one jiffy after marking itself
708 * offline in the cpu_online_mask. This leniency is necessary given the
709 * non-atomic nature of the online and offline processing, for example,
710 * the fact that a CPU enters the scheduler after completing the CPU_DYING
711 * notifiers.
712 *
713 * This is also why RCU internally marks CPUs online during the
714 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
715 *
716 * Disable checking if in an NMI handler because we cannot safely report
717 * errors from NMI handlers anyway.
718 */
719bool rcu_lockdep_current_cpu_online(void)
720{
721 struct rcu_data *rdp;
722 struct rcu_node *rnp;
723 bool ret;
724
725 if (in_nmi())
726 return true;
727 preempt_disable();
728 rdp = this_cpu_ptr(&rcu_sched_data);
729 rnp = rdp->mynode;
730 ret = (rdp->grpmask & rnp->qsmaskinit) ||
731 !rcu_scheduler_fully_active;
732 preempt_enable();
733 return ret;
734}
735EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
736
737#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
738
739/**
740 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
741 *
742 * If the current CPU is idle or running at a first-level (not nested)
743 * interrupt from idle, return true. The caller must have at least
744 * disabled preemption.
745 */
746static int rcu_is_cpu_rrupt_from_idle(void)
747{
748 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
749}
750
751/*
752 * Snapshot the specified CPU's dynticks counter so that we can later
753 * credit them with an implicit quiescent state. Return 1 if this CPU
754 * is in dynticks idle mode, which is an extended quiescent state.
755 */
756static int dyntick_save_progress_counter(struct rcu_data *rdp,
757 bool *isidle, unsigned long *maxj)
758{
759 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
760 rcu_sysidle_check_cpu(rdp, isidle, maxj);
761 return (rdp->dynticks_snap & 0x1) == 0;
762}
763
764/*
765 * This function really isn't for public consumption, but RCU is special in
766 * that context switches can allow the state machine to make progress.
767 */
768extern void resched_cpu(int cpu);
769
770/*
771 * Return true if the specified CPU has passed through a quiescent
772 * state by virtue of being in or having passed through an dynticks
773 * idle state since the last call to dyntick_save_progress_counter()
774 * for this same CPU, or by virtue of having been offline.
775 */
776static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
777 bool *isidle, unsigned long *maxj)
778{
779 unsigned int curr;
780 unsigned int snap;
781
782 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
783 snap = (unsigned int)rdp->dynticks_snap;
784
785 /*
786 * If the CPU passed through or entered a dynticks idle phase with
787 * no active irq/NMI handlers, then we can safely pretend that the CPU
788 * already acknowledged the request to pass through a quiescent
789 * state. Either way, that CPU cannot possibly be in an RCU
790 * read-side critical section that started before the beginning
791 * of the current RCU grace period.
792 */
793 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
794 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
795 rdp->dynticks_fqs++;
796 return 1;
797 }
798
799 /*
800 * Check for the CPU being offline, but only if the grace period
801 * is old enough. We don't need to worry about the CPU changing
802 * state: If we see it offline even once, it has been through a
803 * quiescent state.
804 *
805 * The reason for insisting that the grace period be at least
806 * one jiffy old is that CPUs that are not quite online and that
807 * have just gone offline can still execute RCU read-side critical
808 * sections.
809 */
810 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
811 return 0; /* Grace period is not old enough. */
812 barrier();
813 if (cpu_is_offline(rdp->cpu)) {
814 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
815 rdp->offline_fqs++;
816 return 1;
817 }
818
819 /*
820 * There is a possibility that a CPU in adaptive-ticks state
821 * might run in the kernel with the scheduling-clock tick disabled
822 * for an extended time period. Invoke rcu_kick_nohz_cpu() to
823 * force the CPU to restart the scheduling-clock tick in this
824 * CPU is in this state.
825 */
826 rcu_kick_nohz_cpu(rdp->cpu);
827
828 /*
829 * Alternatively, the CPU might be running in the kernel
830 * for an extended period of time without a quiescent state.
831 * Attempt to force the CPU through the scheduler to gain the
832 * needed quiescent state, but only if the grace period has gone
833 * on for an uncommonly long time. If there are many stuck CPUs,
834 * we will beat on the first one until it gets unstuck, then move
835 * to the next. Only do this for the primary flavor of RCU.
836 */
837 if (rdp->rsp == rcu_state &&
838 ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
839 rdp->rsp->jiffies_resched += 5;
840 resched_cpu(rdp->cpu);
841 }
842
843 return 0;
844}
845
846static void record_gp_stall_check_time(struct rcu_state *rsp)
847{
848 unsigned long j = jiffies;
849 unsigned long j1;
850
851 rsp->gp_start = j;
852 smp_wmb(); /* Record start time before stall time. */
853 j1 = rcu_jiffies_till_stall_check();
854 rsp->jiffies_stall = j + j1;
855 rsp->jiffies_resched = j + j1 / 2;
856}
857
858/*
859 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
860 * for architectures that do not implement trigger_all_cpu_backtrace().
861 * The NMI-triggered stack traces are more accurate because they are
862 * printed by the target CPU.
863 */
864static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
865{
866 int cpu;
867 unsigned long flags;
868 struct rcu_node *rnp;
869
870 rcu_for_each_leaf_node(rsp, rnp) {
871 raw_spin_lock_irqsave(&rnp->lock, flags);
872 if (rnp->qsmask != 0) {
873 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
874 if (rnp->qsmask & (1UL << cpu))
875 dump_cpu_task(rnp->grplo + cpu);
876 }
877 raw_spin_unlock_irqrestore(&rnp->lock, flags);
878 }
879}
880
881static void print_other_cpu_stall(struct rcu_state *rsp)
882{
883 int cpu;
884 long delta;
885 unsigned long flags;
886 int ndetected = 0;
887 struct rcu_node *rnp = rcu_get_root(rsp);
888 long totqlen = 0;
889
890 /* Only let one CPU complain about others per time interval. */
891
892 raw_spin_lock_irqsave(&rnp->lock, flags);
893 delta = jiffies - rsp->jiffies_stall;
894 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
895 raw_spin_unlock_irqrestore(&rnp->lock, flags);
896 return;
897 }
898 rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
899 raw_spin_unlock_irqrestore(&rnp->lock, flags);
900
901 /*
902 * OK, time to rat on our buddy...
903 * See Documentation/RCU/stallwarn.txt for info on how to debug
904 * RCU CPU stall warnings.
905 */
906 pr_err("INFO: %s detected stalls on CPUs/tasks:",
907 rsp->name);
908 print_cpu_stall_info_begin();
909 rcu_for_each_leaf_node(rsp, rnp) {
910 raw_spin_lock_irqsave(&rnp->lock, flags);
911 ndetected += rcu_print_task_stall(rnp);
912 if (rnp->qsmask != 0) {
913 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
914 if (rnp->qsmask & (1UL << cpu)) {
915 print_cpu_stall_info(rsp,
916 rnp->grplo + cpu);
917 ndetected++;
918 }
919 }
920 raw_spin_unlock_irqrestore(&rnp->lock, flags);
921 }
922
923 /*
924 * Now rat on any tasks that got kicked up to the root rcu_node
925 * due to CPU offlining.
926 */
927 rnp = rcu_get_root(rsp);
928 raw_spin_lock_irqsave(&rnp->lock, flags);
929 ndetected += rcu_print_task_stall(rnp);
930 raw_spin_unlock_irqrestore(&rnp->lock, flags);
931
932 print_cpu_stall_info_end();
933 for_each_possible_cpu(cpu)
934 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
935 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
936 smp_processor_id(), (long)(jiffies - rsp->gp_start),
937 rsp->gpnum, rsp->completed, totqlen);
938 if (ndetected == 0)
939 pr_err("INFO: Stall ended before state dump start\n");
940 else if (!trigger_all_cpu_backtrace())
941 rcu_dump_cpu_stacks(rsp);
942
943 /* Complain about tasks blocking the grace period. */
944
945 rcu_print_detail_task_stall(rsp);
946
947 force_quiescent_state(rsp); /* Kick them all. */
948}
949
950/*
951 * This function really isn't for public consumption, but RCU is special in
952 * that context switches can allow the state machine to make progress.
953 */
954extern void resched_cpu(int cpu);
955
956static void print_cpu_stall(struct rcu_state *rsp)
957{
958 int cpu;
959 unsigned long flags;
960 struct rcu_node *rnp = rcu_get_root(rsp);
961 long totqlen = 0;
962
963 /*
964 * OK, time to rat on ourselves...
965 * See Documentation/RCU/stallwarn.txt for info on how to debug
966 * RCU CPU stall warnings.
967 */
968 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
969 print_cpu_stall_info_begin();
970 print_cpu_stall_info(rsp, smp_processor_id());
971 print_cpu_stall_info_end();
972 for_each_possible_cpu(cpu)
973 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
974 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
975 jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
976 if (!trigger_all_cpu_backtrace())
977 dump_stack();
978
979 raw_spin_lock_irqsave(&rnp->lock, flags);
980 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
981 rsp->jiffies_stall = jiffies +
982 3 * rcu_jiffies_till_stall_check() + 3;
983 raw_spin_unlock_irqrestore(&rnp->lock, flags);
984
985 /*
986 * Attempt to revive the RCU machinery by forcing a context switch.
987 *
988 * A context switch would normally allow the RCU state machine to make
989 * progress and it could be we're stuck in kernel space without context
990 * switches for an entirely unreasonable amount of time.
991 */
992 resched_cpu(smp_processor_id());
993}
994
995static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
996{
997 unsigned long completed;
998 unsigned long gpnum;
999 unsigned long gps;
1000 unsigned long j;
1001 unsigned long js;
1002 struct rcu_node *rnp;
1003
1004 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
1005 return;
1006 j = jiffies;
1007
1008 /*
1009 * Lots of memory barriers to reject false positives.
1010 *
1011 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1012 * then rsp->gp_start, and finally rsp->completed. These values
1013 * are updated in the opposite order with memory barriers (or
1014 * equivalent) during grace-period initialization and cleanup.
1015 * Now, a false positive can occur if we get an new value of
1016 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1017 * the memory barriers, the only way that this can happen is if one
1018 * grace period ends and another starts between these two fetches.
1019 * Detect this by comparing rsp->completed with the previous fetch
1020 * from rsp->gpnum.
1021 *
1022 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1023 * and rsp->gp_start suffice to forestall false positives.
1024 */
1025 gpnum = ACCESS_ONCE(rsp->gpnum);
1026 smp_rmb(); /* Pick up ->gpnum first... */
1027 js = ACCESS_ONCE(rsp->jiffies_stall);
1028 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1029 gps = ACCESS_ONCE(rsp->gp_start);
1030 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1031 completed = ACCESS_ONCE(rsp->completed);
1032 if (ULONG_CMP_GE(completed, gpnum) ||
1033 ULONG_CMP_LT(j, js) ||
1034 ULONG_CMP_GE(gps, js))
1035 return; /* No stall or GP completed since entering function. */
1036 rnp = rdp->mynode;
1037 if (rcu_gp_in_progress(rsp) &&
1038 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) {
1039
1040 /* We haven't checked in, so go dump stack. */
1041 print_cpu_stall(rsp);
1042
1043 } else if (rcu_gp_in_progress(rsp) &&
1044 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1045
1046 /* They had a few time units to dump stack, so complain. */
1047 print_other_cpu_stall(rsp);
1048 }
1049}
1050
1051/**
1052 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1053 *
1054 * Set the stall-warning timeout way off into the future, thus preventing
1055 * any RCU CPU stall-warning messages from appearing in the current set of
1056 * RCU grace periods.
1057 *
1058 * The caller must disable hard irqs.
1059 */
1060void rcu_cpu_stall_reset(void)
1061{
1062 struct rcu_state *rsp;
1063
1064 for_each_rcu_flavor(rsp)
1065 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
1066}
1067
1068/*
1069 * Initialize the specified rcu_data structure's callback list to empty.
1070 */
1071static void init_callback_list(struct rcu_data *rdp)
1072{
1073 int i;
1074
1075 if (init_nocb_callback_list(rdp))
1076 return;
1077 rdp->nxtlist = NULL;
1078 for (i = 0; i < RCU_NEXT_SIZE; i++)
1079 rdp->nxttail[i] = &rdp->nxtlist;
1080}
1081
1082/*
1083 * Determine the value that ->completed will have at the end of the
1084 * next subsequent grace period. This is used to tag callbacks so that
1085 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1086 * been dyntick-idle for an extended period with callbacks under the
1087 * influence of RCU_FAST_NO_HZ.
1088 *
1089 * The caller must hold rnp->lock with interrupts disabled.
1090 */
1091static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1092 struct rcu_node *rnp)
1093{
1094 /*
1095 * If RCU is idle, we just wait for the next grace period.
1096 * But we can only be sure that RCU is idle if we are looking
1097 * at the root rcu_node structure -- otherwise, a new grace
1098 * period might have started, but just not yet gotten around
1099 * to initializing the current non-root rcu_node structure.
1100 */
1101 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1102 return rnp->completed + 1;
1103
1104 /*
1105 * Otherwise, wait for a possible partial grace period and
1106 * then the subsequent full grace period.
1107 */
1108 return rnp->completed + 2;
1109}
1110
1111/*
1112 * Trace-event helper function for rcu_start_future_gp() and
1113 * rcu_nocb_wait_gp().
1114 */
1115static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1116 unsigned long c, const char *s)
1117{
1118 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1119 rnp->completed, c, rnp->level,
1120 rnp->grplo, rnp->grphi, s);
1121}
1122
1123/*
1124 * Start some future grace period, as needed to handle newly arrived
1125 * callbacks. The required future grace periods are recorded in each
1126 * rcu_node structure's ->need_future_gp field.
1127 *
1128 * The caller must hold the specified rcu_node structure's ->lock.
1129 */
1130static unsigned long __maybe_unused
1131rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
1132{
1133 unsigned long c;
1134 int i;
1135 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1136
1137 /*
1138 * Pick up grace-period number for new callbacks. If this
1139 * grace period is already marked as needed, return to the caller.
1140 */
1141 c = rcu_cbs_completed(rdp->rsp, rnp);
1142 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1143 if (rnp->need_future_gp[c & 0x1]) {
1144 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1145 return c;
1146 }
1147
1148 /*
1149 * If either this rcu_node structure or the root rcu_node structure
1150 * believe that a grace period is in progress, then we must wait
1151 * for the one following, which is in "c". Because our request
1152 * will be noticed at the end of the current grace period, we don't
1153 * need to explicitly start one.
1154 */
1155 if (rnp->gpnum != rnp->completed ||
1156 ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
1157 rnp->need_future_gp[c & 0x1]++;
1158 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1159 return c;
1160 }
1161
1162 /*
1163 * There might be no grace period in progress. If we don't already
1164 * hold it, acquire the root rcu_node structure's lock in order to
1165 * start one (if needed).
1166 */
1167 if (rnp != rnp_root) {
1168 raw_spin_lock(&rnp_root->lock);
1169 smp_mb__after_unlock_lock();
1170 }
1171
1172 /*
1173 * Get a new grace-period number. If there really is no grace
1174 * period in progress, it will be smaller than the one we obtained
1175 * earlier. Adjust callbacks as needed. Note that even no-CBs
1176 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1177 */
1178 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1179 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1180 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1181 rdp->nxtcompleted[i] = c;
1182
1183 /*
1184 * If the needed for the required grace period is already
1185 * recorded, trace and leave.
1186 */
1187 if (rnp_root->need_future_gp[c & 0x1]) {
1188 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1189 goto unlock_out;
1190 }
1191
1192 /* Record the need for the future grace period. */
1193 rnp_root->need_future_gp[c & 0x1]++;
1194
1195 /* If a grace period is not already in progress, start one. */
1196 if (rnp_root->gpnum != rnp_root->completed) {
1197 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1198 } else {
1199 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1200 rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1201 }
1202unlock_out:
1203 if (rnp != rnp_root)
1204 raw_spin_unlock(&rnp_root->lock);
1205 return c;
1206}
1207
1208/*
1209 * Clean up any old requests for the just-ended grace period. Also return
1210 * whether any additional grace periods have been requested. Also invoke
1211 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1212 * waiting for this grace period to complete.
1213 */
1214static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1215{
1216 int c = rnp->completed;
1217 int needmore;
1218 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1219
1220 rcu_nocb_gp_cleanup(rsp, rnp);
1221 rnp->need_future_gp[c & 0x1] = 0;
1222 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1223 trace_rcu_future_gp(rnp, rdp, c,
1224 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1225 return needmore;
1226}
1227
1228/*
1229 * If there is room, assign a ->completed number to any callbacks on
1230 * this CPU that have not already been assigned. Also accelerate any
1231 * callbacks that were previously assigned a ->completed number that has
1232 * since proven to be too conservative, which can happen if callbacks get
1233 * assigned a ->completed number while RCU is idle, but with reference to
1234 * a non-root rcu_node structure. This function is idempotent, so it does
1235 * not hurt to call it repeatedly.
1236 *
1237 * The caller must hold rnp->lock with interrupts disabled.
1238 */
1239static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1240 struct rcu_data *rdp)
1241{
1242 unsigned long c;
1243 int i;
1244
1245 /* If the CPU has no callbacks, nothing to do. */
1246 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1247 return;
1248
1249 /*
1250 * Starting from the sublist containing the callbacks most
1251 * recently assigned a ->completed number and working down, find the
1252 * first sublist that is not assignable to an upcoming grace period.
1253 * Such a sublist has something in it (first two tests) and has
1254 * a ->completed number assigned that will complete sooner than
1255 * the ->completed number for newly arrived callbacks (last test).
1256 *
1257 * The key point is that any later sublist can be assigned the
1258 * same ->completed number as the newly arrived callbacks, which
1259 * means that the callbacks in any of these later sublist can be
1260 * grouped into a single sublist, whether or not they have already
1261 * been assigned a ->completed number.
1262 */
1263 c = rcu_cbs_completed(rsp, rnp);
1264 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1265 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1266 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1267 break;
1268
1269 /*
1270 * If there are no sublist for unassigned callbacks, leave.
1271 * At the same time, advance "i" one sublist, so that "i" will
1272 * index into the sublist where all the remaining callbacks should
1273 * be grouped into.
1274 */
1275 if (++i >= RCU_NEXT_TAIL)
1276 return;
1277
1278 /*
1279 * Assign all subsequent callbacks' ->completed number to the next
1280 * full grace period and group them all in the sublist initially
1281 * indexed by "i".
1282 */
1283 for (; i <= RCU_NEXT_TAIL; i++) {
1284 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1285 rdp->nxtcompleted[i] = c;
1286 }
1287 /* Record any needed additional grace periods. */
1288 rcu_start_future_gp(rnp, rdp);
1289
1290 /* Trace depending on how much we were able to accelerate. */
1291 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1292 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1293 else
1294 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1295}
1296
1297/*
1298 * Move any callbacks whose grace period has completed to the
1299 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1300 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1301 * sublist. This function is idempotent, so it does not hurt to
1302 * invoke it repeatedly. As long as it is not invoked -too- often...
1303 *
1304 * The caller must hold rnp->lock with interrupts disabled.
1305 */
1306static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1307 struct rcu_data *rdp)
1308{
1309 int i, j;
1310
1311 /* If the CPU has no callbacks, nothing to do. */
1312 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1313 return;
1314
1315 /*
1316 * Find all callbacks whose ->completed numbers indicate that they
1317 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1318 */
1319 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1320 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1321 break;
1322 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1323 }
1324 /* Clean up any sublist tail pointers that were misordered above. */
1325 for (j = RCU_WAIT_TAIL; j < i; j++)
1326 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1327
1328 /* Copy down callbacks to fill in empty sublists. */
1329 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1330 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1331 break;
1332 rdp->nxttail[j] = rdp->nxttail[i];
1333 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1334 }
1335
1336 /* Classify any remaining callbacks. */
1337 rcu_accelerate_cbs(rsp, rnp, rdp);
1338}
1339
1340/*
1341 * Update CPU-local rcu_data state to record the beginnings and ends of
1342 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1343 * structure corresponding to the current CPU, and must have irqs disabled.
1344 */
1345static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1346{
1347 /* Handle the ends of any preceding grace periods first. */
1348 if (rdp->completed == rnp->completed) {
1349
1350 /* No grace period end, so just accelerate recent callbacks. */
1351 rcu_accelerate_cbs(rsp, rnp, rdp);
1352
1353 } else {
1354
1355 /* Advance callbacks. */
1356 rcu_advance_cbs(rsp, rnp, rdp);
1357
1358 /* Remember that we saw this grace-period completion. */
1359 rdp->completed = rnp->completed;
1360 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1361 }
1362
1363 if (rdp->gpnum != rnp->gpnum) {
1364 /*
1365 * If the current grace period is waiting for this CPU,
1366 * set up to detect a quiescent state, otherwise don't
1367 * go looking for one.
1368 */
1369 rdp->gpnum = rnp->gpnum;
1370 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1371 rdp->passed_quiesce = 0;
1372 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1373 zero_cpu_stall_ticks(rdp);
1374 }
1375}
1376
1377static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1378{
1379 unsigned long flags;
1380 struct rcu_node *rnp;
1381
1382 local_irq_save(flags);
1383 rnp = rdp->mynode;
1384 if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1385 rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */
1386 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1387 local_irq_restore(flags);
1388 return;
1389 }
1390 smp_mb__after_unlock_lock();
1391 __note_gp_changes(rsp, rnp, rdp);
1392 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1393}
1394
1395/*
1396 * Initialize a new grace period. Return 0 if no grace period required.
1397 */
1398static int rcu_gp_init(struct rcu_state *rsp)
1399{
1400 struct rcu_data *rdp;
1401 struct rcu_node *rnp = rcu_get_root(rsp);
1402
1403 rcu_bind_gp_kthread();
1404 raw_spin_lock_irq(&rnp->lock);
1405 smp_mb__after_unlock_lock();
1406 if (rsp->gp_flags == 0) {
1407 /* Spurious wakeup, tell caller to go back to sleep. */
1408 raw_spin_unlock_irq(&rnp->lock);
1409 return 0;
1410 }
1411 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1412
1413 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1414 /*
1415 * Grace period already in progress, don't start another.
1416 * Not supposed to be able to happen.
1417 */
1418 raw_spin_unlock_irq(&rnp->lock);
1419 return 0;
1420 }
1421
1422 /* Advance to a new grace period and initialize state. */
1423 record_gp_stall_check_time(rsp);
1424 /* Record GP times before starting GP, hence smp_store_release(). */
1425 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1426 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1427 raw_spin_unlock_irq(&rnp->lock);
1428
1429 /* Exclude any concurrent CPU-hotplug operations. */
1430 mutex_lock(&rsp->onoff_mutex);
1431 smp_mb__after_unlock_lock(); /* ->gpnum increment before GP! */
1432
1433 /*
1434 * Set the quiescent-state-needed bits in all the rcu_node
1435 * structures for all currently online CPUs in breadth-first order,
1436 * starting from the root rcu_node structure, relying on the layout
1437 * of the tree within the rsp->node[] array. Note that other CPUs
1438 * will access only the leaves of the hierarchy, thus seeing that no
1439 * grace period is in progress, at least until the corresponding
1440 * leaf node has been initialized. In addition, we have excluded
1441 * CPU-hotplug operations.
1442 *
1443 * The grace period cannot complete until the initialization
1444 * process finishes, because this kthread handles both.
1445 */
1446 rcu_for_each_node_breadth_first(rsp, rnp) {
1447 raw_spin_lock_irq(&rnp->lock);
1448 smp_mb__after_unlock_lock();
1449 rdp = this_cpu_ptr(rsp->rda);
1450 rcu_preempt_check_blocked_tasks(rnp);
1451 rnp->qsmask = rnp->qsmaskinit;
1452 ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1453 WARN_ON_ONCE(rnp->completed != rsp->completed);
1454 ACCESS_ONCE(rnp->completed) = rsp->completed;
1455 if (rnp == rdp->mynode)
1456 __note_gp_changes(rsp, rnp, rdp);
1457 rcu_preempt_boost_start_gp(rnp);
1458 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1459 rnp->level, rnp->grplo,
1460 rnp->grphi, rnp->qsmask);
1461 raw_spin_unlock_irq(&rnp->lock);
1462#ifdef CONFIG_PROVE_RCU_DELAY
1463 if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 &&
1464 system_state == SYSTEM_RUNNING)
1465 udelay(200);
1466#endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1467 cond_resched();
1468 }
1469
1470 mutex_unlock(&rsp->onoff_mutex);
1471 return 1;
1472}
1473
1474/*
1475 * Do one round of quiescent-state forcing.
1476 */
1477static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1478{
1479 int fqs_state = fqs_state_in;
1480 bool isidle = false;
1481 unsigned long maxj;
1482 struct rcu_node *rnp = rcu_get_root(rsp);
1483
1484 rsp->n_force_qs++;
1485 if (fqs_state == RCU_SAVE_DYNTICK) {
1486 /* Collect dyntick-idle snapshots. */
1487 if (is_sysidle_rcu_state(rsp)) {
1488 isidle = 1;
1489 maxj = jiffies - ULONG_MAX / 4;
1490 }
1491 force_qs_rnp(rsp, dyntick_save_progress_counter,
1492 &isidle, &maxj);
1493 rcu_sysidle_report_gp(rsp, isidle, maxj);
1494 fqs_state = RCU_FORCE_QS;
1495 } else {
1496 /* Handle dyntick-idle and offline CPUs. */
1497 isidle = 0;
1498 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1499 }
1500 /* Clear flag to prevent immediate re-entry. */
1501 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1502 raw_spin_lock_irq(&rnp->lock);
1503 smp_mb__after_unlock_lock();
1504 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1505 raw_spin_unlock_irq(&rnp->lock);
1506 }
1507 return fqs_state;
1508}
1509
1510/*
1511 * Clean up after the old grace period.
1512 */
1513static void rcu_gp_cleanup(struct rcu_state *rsp)
1514{
1515 unsigned long gp_duration;
1516 int nocb = 0;
1517 struct rcu_data *rdp;
1518 struct rcu_node *rnp = rcu_get_root(rsp);
1519
1520 raw_spin_lock_irq(&rnp->lock);
1521 smp_mb__after_unlock_lock();
1522 gp_duration = jiffies - rsp->gp_start;
1523 if (gp_duration > rsp->gp_max)
1524 rsp->gp_max = gp_duration;
1525
1526 /*
1527 * We know the grace period is complete, but to everyone else
1528 * it appears to still be ongoing. But it is also the case
1529 * that to everyone else it looks like there is nothing that
1530 * they can do to advance the grace period. It is therefore
1531 * safe for us to drop the lock in order to mark the grace
1532 * period as completed in all of the rcu_node structures.
1533 */
1534 raw_spin_unlock_irq(&rnp->lock);
1535
1536 /*
1537 * Propagate new ->completed value to rcu_node structures so
1538 * that other CPUs don't have to wait until the start of the next
1539 * grace period to process their callbacks. This also avoids
1540 * some nasty RCU grace-period initialization races by forcing
1541 * the end of the current grace period to be completely recorded in
1542 * all of the rcu_node structures before the beginning of the next
1543 * grace period is recorded in any of the rcu_node structures.
1544 */
1545 rcu_for_each_node_breadth_first(rsp, rnp) {
1546 raw_spin_lock_irq(&rnp->lock);
1547 smp_mb__after_unlock_lock();
1548 ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1549 rdp = this_cpu_ptr(rsp->rda);
1550 if (rnp == rdp->mynode)
1551 __note_gp_changes(rsp, rnp, rdp);
1552 /* smp_mb() provided by prior unlock-lock pair. */
1553 nocb += rcu_future_gp_cleanup(rsp, rnp);
1554 raw_spin_unlock_irq(&rnp->lock);
1555 cond_resched();
1556 }
1557 rnp = rcu_get_root(rsp);
1558 raw_spin_lock_irq(&rnp->lock);
1559 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
1560 rcu_nocb_gp_set(rnp, nocb);
1561
1562 /* Declare grace period done. */
1563 ACCESS_ONCE(rsp->completed) = rsp->gpnum;
1564 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
1565 rsp->fqs_state = RCU_GP_IDLE;
1566 rdp = this_cpu_ptr(rsp->rda);
1567 rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */
1568 if (cpu_needs_another_gp(rsp, rdp)) {
1569 rsp->gp_flags = RCU_GP_FLAG_INIT;
1570 trace_rcu_grace_period(rsp->name,
1571 ACCESS_ONCE(rsp->gpnum),
1572 TPS("newreq"));
1573 }
1574 raw_spin_unlock_irq(&rnp->lock);
1575}
1576
1577/*
1578 * Body of kthread that handles grace periods.
1579 */
1580static int __noreturn rcu_gp_kthread(void *arg)
1581{
1582 int fqs_state;
1583 int gf;
1584 unsigned long j;
1585 int ret;
1586 struct rcu_state *rsp = arg;
1587 struct rcu_node *rnp = rcu_get_root(rsp);
1588
1589 for (;;) {
1590
1591 /* Handle grace-period start. */
1592 for (;;) {
1593 trace_rcu_grace_period(rsp->name,
1594 ACCESS_ONCE(rsp->gpnum),
1595 TPS("reqwait"));
1596 wait_event_interruptible(rsp->gp_wq,
1597 ACCESS_ONCE(rsp->gp_flags) &
1598 RCU_GP_FLAG_INIT);
1599 /* Locking provides needed memory barrier. */
1600 if (rcu_gp_init(rsp))
1601 break;
1602 cond_resched();
1603 flush_signals(current);
1604 trace_rcu_grace_period(rsp->name,
1605 ACCESS_ONCE(rsp->gpnum),
1606 TPS("reqwaitsig"));
1607 }
1608
1609 /* Handle quiescent-state forcing. */
1610 fqs_state = RCU_SAVE_DYNTICK;
1611 j = jiffies_till_first_fqs;
1612 if (j > HZ) {
1613 j = HZ;
1614 jiffies_till_first_fqs = HZ;
1615 }
1616 ret = 0;
1617 for (;;) {
1618 if (!ret)
1619 rsp->jiffies_force_qs = jiffies + j;
1620 trace_rcu_grace_period(rsp->name,
1621 ACCESS_ONCE(rsp->gpnum),
1622 TPS("fqswait"));
1623 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1624 ((gf = ACCESS_ONCE(rsp->gp_flags)) &
1625 RCU_GP_FLAG_FQS) ||
1626 (!ACCESS_ONCE(rnp->qsmask) &&
1627 !rcu_preempt_blocked_readers_cgp(rnp)),
1628 j);
1629 /* Locking provides needed memory barriers. */
1630 /* If grace period done, leave loop. */
1631 if (!ACCESS_ONCE(rnp->qsmask) &&
1632 !rcu_preempt_blocked_readers_cgp(rnp))
1633 break;
1634 /* If time for quiescent-state forcing, do it. */
1635 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
1636 (gf & RCU_GP_FLAG_FQS)) {
1637 trace_rcu_grace_period(rsp->name,
1638 ACCESS_ONCE(rsp->gpnum),
1639 TPS("fqsstart"));
1640 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1641 trace_rcu_grace_period(rsp->name,
1642 ACCESS_ONCE(rsp->gpnum),
1643 TPS("fqsend"));
1644 cond_resched();
1645 } else {
1646 /* Deal with stray signal. */
1647 cond_resched();
1648 flush_signals(current);
1649 trace_rcu_grace_period(rsp->name,
1650 ACCESS_ONCE(rsp->gpnum),
1651 TPS("fqswaitsig"));
1652 }
1653 j = jiffies_till_next_fqs;
1654 if (j > HZ) {
1655 j = HZ;
1656 jiffies_till_next_fqs = HZ;
1657 } else if (j < 1) {
1658 j = 1;
1659 jiffies_till_next_fqs = 1;
1660 }
1661 }
1662
1663 /* Handle grace-period end. */
1664 rcu_gp_cleanup(rsp);
1665 }
1666}
1667
1668static void rsp_wakeup(struct irq_work *work)
1669{
1670 struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work);
1671
1672 /* Wake up rcu_gp_kthread() to start the grace period. */
1673 wake_up(&rsp->gp_wq);
1674}
1675
1676/*
1677 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1678 * in preparation for detecting the next grace period. The caller must hold
1679 * the root node's ->lock and hard irqs must be disabled.
1680 *
1681 * Note that it is legal for a dying CPU (which is marked as offline) to
1682 * invoke this function. This can happen when the dying CPU reports its
1683 * quiescent state.
1684 */
1685static void
1686rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1687 struct rcu_data *rdp)
1688{
1689 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1690 /*
1691 * Either we have not yet spawned the grace-period
1692 * task, this CPU does not need another grace period,
1693 * or a grace period is already in progress.
1694 * Either way, don't start a new grace period.
1695 */
1696 return;
1697 }
1698 rsp->gp_flags = RCU_GP_FLAG_INIT;
1699 trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum),
1700 TPS("newreq"));
1701
1702 /*
1703 * We can't do wakeups while holding the rnp->lock, as that
1704 * could cause possible deadlocks with the rq->lock. Defer
1705 * the wakeup to interrupt context. And don't bother waking
1706 * up the running kthread.
1707 */
1708 if (current != rsp->gp_kthread)
1709 irq_work_queue(&rsp->wakeup_work);
1710}
1711
1712/*
1713 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1714 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1715 * is invoked indirectly from rcu_advance_cbs(), which would result in
1716 * endless recursion -- or would do so if it wasn't for the self-deadlock
1717 * that is encountered beforehand.
1718 */
1719static void
1720rcu_start_gp(struct rcu_state *rsp)
1721{
1722 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1723 struct rcu_node *rnp = rcu_get_root(rsp);
1724
1725 /*
1726 * If there is no grace period in progress right now, any
1727 * callbacks we have up to this point will be satisfied by the
1728 * next grace period. Also, advancing the callbacks reduces the
1729 * probability of false positives from cpu_needs_another_gp()
1730 * resulting in pointless grace periods. So, advance callbacks
1731 * then start the grace period!
1732 */
1733 rcu_advance_cbs(rsp, rnp, rdp);
1734 rcu_start_gp_advanced(rsp, rnp, rdp);
1735}
1736
1737/*
1738 * Report a full set of quiescent states to the specified rcu_state
1739 * data structure. This involves cleaning up after the prior grace
1740 * period and letting rcu_start_gp() start up the next grace period
1741 * if one is needed. Note that the caller must hold rnp->lock, which
1742 * is released before return.
1743 */
1744static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1745 __releases(rcu_get_root(rsp)->lock)
1746{
1747 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1748 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1749 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1750}
1751
1752/*
1753 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1754 * Allows quiescent states for a group of CPUs to be reported at one go
1755 * to the specified rcu_node structure, though all the CPUs in the group
1756 * must be represented by the same rcu_node structure (which need not be
1757 * a leaf rcu_node structure, though it often will be). That structure's
1758 * lock must be held upon entry, and it is released before return.
1759 */
1760static void
1761rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1762 struct rcu_node *rnp, unsigned long flags)
1763 __releases(rnp->lock)
1764{
1765 struct rcu_node *rnp_c;
1766
1767 /* Walk up the rcu_node hierarchy. */
1768 for (;;) {
1769 if (!(rnp->qsmask & mask)) {
1770
1771 /* Our bit has already been cleared, so done. */
1772 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1773 return;
1774 }
1775 rnp->qsmask &= ~mask;
1776 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1777 mask, rnp->qsmask, rnp->level,
1778 rnp->grplo, rnp->grphi,
1779 !!rnp->gp_tasks);
1780 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1781
1782 /* Other bits still set at this level, so done. */
1783 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1784 return;
1785 }
1786 mask = rnp->grpmask;
1787 if (rnp->parent == NULL) {
1788
1789 /* No more levels. Exit loop holding root lock. */
1790
1791 break;
1792 }
1793 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1794 rnp_c = rnp;
1795 rnp = rnp->parent;
1796 raw_spin_lock_irqsave(&rnp->lock, flags);
1797 smp_mb__after_unlock_lock();
1798 WARN_ON_ONCE(rnp_c->qsmask);
1799 }
1800
1801 /*
1802 * Get here if we are the last CPU to pass through a quiescent
1803 * state for this grace period. Invoke rcu_report_qs_rsp()
1804 * to clean up and start the next grace period if one is needed.
1805 */
1806 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1807}
1808
1809/*
1810 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1811 * structure. This must be either called from the specified CPU, or
1812 * called when the specified CPU is known to be offline (and when it is
1813 * also known that no other CPU is concurrently trying to help the offline
1814 * CPU). The lastcomp argument is used to make sure we are still in the
1815 * grace period of interest. We don't want to end the current grace period
1816 * based on quiescent states detected in an earlier grace period!
1817 */
1818static void
1819rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1820{
1821 unsigned long flags;
1822 unsigned long mask;
1823 struct rcu_node *rnp;
1824
1825 rnp = rdp->mynode;
1826 raw_spin_lock_irqsave(&rnp->lock, flags);
1827 smp_mb__after_unlock_lock();
1828 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1829 rnp->completed == rnp->gpnum) {
1830
1831 /*
1832 * The grace period in which this quiescent state was
1833 * recorded has ended, so don't report it upwards.
1834 * We will instead need a new quiescent state that lies
1835 * within the current grace period.
1836 */
1837 rdp->passed_quiesce = 0; /* need qs for new gp. */
1838 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1839 return;
1840 }
1841 mask = rdp->grpmask;
1842 if ((rnp->qsmask & mask) == 0) {
1843 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1844 } else {
1845 rdp->qs_pending = 0;
1846
1847 /*
1848 * This GP can't end until cpu checks in, so all of our
1849 * callbacks can be processed during the next GP.
1850 */
1851 rcu_accelerate_cbs(rsp, rnp, rdp);
1852
1853 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1854 }
1855}
1856
1857/*
1858 * Check to see if there is a new grace period of which this CPU
1859 * is not yet aware, and if so, set up local rcu_data state for it.
1860 * Otherwise, see if this CPU has just passed through its first
1861 * quiescent state for this grace period, and record that fact if so.
1862 */
1863static void
1864rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1865{
1866 /* Check for grace-period ends and beginnings. */
1867 note_gp_changes(rsp, rdp);
1868
1869 /*
1870 * Does this CPU still need to do its part for current grace period?
1871 * If no, return and let the other CPUs do their part as well.
1872 */
1873 if (!rdp->qs_pending)
1874 return;
1875
1876 /*
1877 * Was there a quiescent state since the beginning of the grace
1878 * period? If no, then exit and wait for the next call.
1879 */
1880 if (!rdp->passed_quiesce)
1881 return;
1882
1883 /*
1884 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1885 * judge of that).
1886 */
1887 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1888}
1889
1890#ifdef CONFIG_HOTPLUG_CPU
1891
1892/*
1893 * Send the specified CPU's RCU callbacks to the orphanage. The
1894 * specified CPU must be offline, and the caller must hold the
1895 * ->orphan_lock.
1896 */
1897static void
1898rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1899 struct rcu_node *rnp, struct rcu_data *rdp)
1900{
1901 /* No-CBs CPUs do not have orphanable callbacks. */
1902 if (rcu_is_nocb_cpu(rdp->cpu))
1903 return;
1904
1905 /*
1906 * Orphan the callbacks. First adjust the counts. This is safe
1907 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1908 * cannot be running now. Thus no memory barrier is required.
1909 */
1910 if (rdp->nxtlist != NULL) {
1911 rsp->qlen_lazy += rdp->qlen_lazy;
1912 rsp->qlen += rdp->qlen;
1913 rdp->n_cbs_orphaned += rdp->qlen;
1914 rdp->qlen_lazy = 0;
1915 ACCESS_ONCE(rdp->qlen) = 0;
1916 }
1917
1918 /*
1919 * Next, move those callbacks still needing a grace period to
1920 * the orphanage, where some other CPU will pick them up.
1921 * Some of the callbacks might have gone partway through a grace
1922 * period, but that is too bad. They get to start over because we
1923 * cannot assume that grace periods are synchronized across CPUs.
1924 * We don't bother updating the ->nxttail[] array yet, instead
1925 * we just reset the whole thing later on.
1926 */
1927 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1928 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1929 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1930 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1931 }
1932
1933 /*
1934 * Then move the ready-to-invoke callbacks to the orphanage,
1935 * where some other CPU will pick them up. These will not be
1936 * required to pass though another grace period: They are done.
1937 */
1938 if (rdp->nxtlist != NULL) {
1939 *rsp->orphan_donetail = rdp->nxtlist;
1940 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1941 }
1942
1943 /* Finally, initialize the rcu_data structure's list to empty. */
1944 init_callback_list(rdp);
1945}
1946
1947/*
1948 * Adopt the RCU callbacks from the specified rcu_state structure's
1949 * orphanage. The caller must hold the ->orphan_lock.
1950 */
1951static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
1952{
1953 int i;
1954 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1955
1956 /* No-CBs CPUs are handled specially. */
1957 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
1958 return;
1959
1960 /* Do the accounting first. */
1961 rdp->qlen_lazy += rsp->qlen_lazy;
1962 rdp->qlen += rsp->qlen;
1963 rdp->n_cbs_adopted += rsp->qlen;
1964 if (rsp->qlen_lazy != rsp->qlen)
1965 rcu_idle_count_callbacks_posted();
1966 rsp->qlen_lazy = 0;
1967 rsp->qlen = 0;
1968
1969 /*
1970 * We do not need a memory barrier here because the only way we
1971 * can get here if there is an rcu_barrier() in flight is if
1972 * we are the task doing the rcu_barrier().
1973 */
1974
1975 /* First adopt the ready-to-invoke callbacks. */
1976 if (rsp->orphan_donelist != NULL) {
1977 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1978 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1979 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1980 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1981 rdp->nxttail[i] = rsp->orphan_donetail;
1982 rsp->orphan_donelist = NULL;
1983 rsp->orphan_donetail = &rsp->orphan_donelist;
1984 }
1985
1986 /* And then adopt the callbacks that still need a grace period. */
1987 if (rsp->orphan_nxtlist != NULL) {
1988 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1989 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1990 rsp->orphan_nxtlist = NULL;
1991 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1992 }
1993}
1994
1995/*
1996 * Trace the fact that this CPU is going offline.
1997 */
1998static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1999{
2000 RCU_TRACE(unsigned long mask);
2001 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
2002 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
2003
2004 RCU_TRACE(mask = rdp->grpmask);
2005 trace_rcu_grace_period(rsp->name,
2006 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2007 TPS("cpuofl"));
2008}
2009
2010/*
2011 * The CPU has been completely removed, and some other CPU is reporting
2012 * this fact from process context. Do the remainder of the cleanup,
2013 * including orphaning the outgoing CPU's RCU callbacks, and also
2014 * adopting them. There can only be one CPU hotplug operation at a time,
2015 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2016 */
2017static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2018{
2019 unsigned long flags;
2020 unsigned long mask;
2021 int need_report = 0;
2022 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2023 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2024
2025 /* Adjust any no-longer-needed kthreads. */
2026 rcu_boost_kthread_setaffinity(rnp, -1);
2027
2028 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
2029
2030 /* Exclude any attempts to start a new grace period. */
2031 mutex_lock(&rsp->onoff_mutex);
2032 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2033
2034 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2035 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2036 rcu_adopt_orphan_cbs(rsp, flags);
2037
2038 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
2039 mask = rdp->grpmask; /* rnp->grplo is constant. */
2040 do {
2041 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2042 smp_mb__after_unlock_lock();
2043 rnp->qsmaskinit &= ~mask;
2044 if (rnp->qsmaskinit != 0) {
2045 if (rnp != rdp->mynode)
2046 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2047 break;
2048 }
2049 if (rnp == rdp->mynode)
2050 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
2051 else
2052 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2053 mask = rnp->grpmask;
2054 rnp = rnp->parent;
2055 } while (rnp != NULL);
2056
2057 /*
2058 * We still hold the leaf rcu_node structure lock here, and
2059 * irqs are still disabled. The reason for this subterfuge is
2060 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
2061 * held leads to deadlock.
2062 */
2063 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
2064 rnp = rdp->mynode;
2065 if (need_report & RCU_OFL_TASKS_NORM_GP)
2066 rcu_report_unblock_qs_rnp(rnp, flags);
2067 else
2068 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2069 if (need_report & RCU_OFL_TASKS_EXP_GP)
2070 rcu_report_exp_rnp(rsp, rnp, true);
2071 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2072 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2073 cpu, rdp->qlen, rdp->nxtlist);
2074 init_callback_list(rdp);
2075 /* Disallow further callbacks on this CPU. */
2076 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2077 mutex_unlock(&rsp->onoff_mutex);
2078}
2079
2080#else /* #ifdef CONFIG_HOTPLUG_CPU */
2081
2082static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2083{
2084}
2085
2086static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2087{
2088}
2089
2090#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2091
2092/*
2093 * Invoke any RCU callbacks that have made it to the end of their grace
2094 * period. Thottle as specified by rdp->blimit.
2095 */
2096static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2097{
2098 unsigned long flags;
2099 struct rcu_head *next, *list, **tail;
2100 long bl, count, count_lazy;
2101 int i;
2102
2103 /* If no callbacks are ready, just return. */
2104 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2105 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2106 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
2107 need_resched(), is_idle_task(current),
2108 rcu_is_callbacks_kthread());
2109 return;
2110 }
2111
2112 /*
2113 * Extract the list of ready callbacks, disabling to prevent
2114 * races with call_rcu() from interrupt handlers.
2115 */
2116 local_irq_save(flags);
2117 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2118 bl = rdp->blimit;
2119 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2120 list = rdp->nxtlist;
2121 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2122 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2123 tail = rdp->nxttail[RCU_DONE_TAIL];
2124 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2125 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2126 rdp->nxttail[i] = &rdp->nxtlist;
2127 local_irq_restore(flags);
2128
2129 /* Invoke callbacks. */
2130 count = count_lazy = 0;
2131 while (list) {
2132 next = list->next;
2133 prefetch(next);
2134 debug_rcu_head_unqueue(list);
2135 if (__rcu_reclaim(rsp->name, list))
2136 count_lazy++;
2137 list = next;
2138 /* Stop only if limit reached and CPU has something to do. */
2139 if (++count >= bl &&
2140 (need_resched() ||
2141 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2142 break;
2143 }
2144
2145 local_irq_save(flags);
2146 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2147 is_idle_task(current),
2148 rcu_is_callbacks_kthread());
2149
2150 /* Update count, and requeue any remaining callbacks. */
2151 if (list != NULL) {
2152 *tail = rdp->nxtlist;
2153 rdp->nxtlist = list;
2154 for (i = 0; i < RCU_NEXT_SIZE; i++)
2155 if (&rdp->nxtlist == rdp->nxttail[i])
2156 rdp->nxttail[i] = tail;
2157 else
2158 break;
2159 }
2160 smp_mb(); /* List handling before counting for rcu_barrier(). */
2161 rdp->qlen_lazy -= count_lazy;
2162 ACCESS_ONCE(rdp->qlen) -= count;
2163 rdp->n_cbs_invoked += count;
2164
2165 /* Reinstate batch limit if we have worked down the excess. */
2166 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2167 rdp->blimit = blimit;
2168
2169 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2170 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2171 rdp->qlen_last_fqs_check = 0;
2172 rdp->n_force_qs_snap = rsp->n_force_qs;
2173 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2174 rdp->qlen_last_fqs_check = rdp->qlen;
2175 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2176
2177 local_irq_restore(flags);
2178
2179 /* Re-invoke RCU core processing if there are callbacks remaining. */
2180 if (cpu_has_callbacks_ready_to_invoke(rdp))
2181 invoke_rcu_core();
2182}
2183
2184/*
2185 * Check to see if this CPU is in a non-context-switch quiescent state
2186 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2187 * Also schedule RCU core processing.
2188 *
2189 * This function must be called from hardirq context. It is normally
2190 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2191 * false, there is no point in invoking rcu_check_callbacks().
2192 */
2193void rcu_check_callbacks(int cpu, int user)
2194{
2195 trace_rcu_utilization(TPS("Start scheduler-tick"));
2196 increment_cpu_stall_ticks();
2197 if (user || rcu_is_cpu_rrupt_from_idle()) {
2198
2199 /*
2200 * Get here if this CPU took its interrupt from user
2201 * mode or from the idle loop, and if this is not a
2202 * nested interrupt. In this case, the CPU is in
2203 * a quiescent state, so note it.
2204 *
2205 * No memory barrier is required here because both
2206 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2207 * variables that other CPUs neither access nor modify,
2208 * at least not while the corresponding CPU is online.
2209 */
2210
2211 rcu_sched_qs(cpu);
2212 rcu_bh_qs(cpu);
2213
2214 } else if (!in_softirq()) {
2215
2216 /*
2217 * Get here if this CPU did not take its interrupt from
2218 * softirq, in other words, if it is not interrupting
2219 * a rcu_bh read-side critical section. This is an _bh
2220 * critical section, so note it.
2221 */
2222
2223 rcu_bh_qs(cpu);
2224 }
2225 rcu_preempt_check_callbacks(cpu);
2226 if (rcu_pending(cpu))
2227 invoke_rcu_core();
2228 trace_rcu_utilization(TPS("End scheduler-tick"));
2229}
2230
2231/*
2232 * Scan the leaf rcu_node structures, processing dyntick state for any that
2233 * have not yet encountered a quiescent state, using the function specified.
2234 * Also initiate boosting for any threads blocked on the root rcu_node.
2235 *
2236 * The caller must have suppressed start of new grace periods.
2237 */
2238static void force_qs_rnp(struct rcu_state *rsp,
2239 int (*f)(struct rcu_data *rsp, bool *isidle,
2240 unsigned long *maxj),
2241 bool *isidle, unsigned long *maxj)
2242{
2243 unsigned long bit;
2244 int cpu;
2245 unsigned long flags;
2246 unsigned long mask;
2247 struct rcu_node *rnp;
2248
2249 rcu_for_each_leaf_node(rsp, rnp) {
2250 cond_resched();
2251 mask = 0;
2252 raw_spin_lock_irqsave(&rnp->lock, flags);
2253 smp_mb__after_unlock_lock();
2254 if (!rcu_gp_in_progress(rsp)) {
2255 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2256 return;
2257 }
2258 if (rnp->qsmask == 0) {
2259 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2260 continue;
2261 }
2262 cpu = rnp->grplo;
2263 bit = 1;
2264 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2265 if ((rnp->qsmask & bit) != 0) {
2266 if ((rnp->qsmaskinit & bit) != 0)
2267 *isidle = 0;
2268 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2269 mask |= bit;
2270 }
2271 }
2272 if (mask != 0) {
2273
2274 /* rcu_report_qs_rnp() releases rnp->lock. */
2275 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2276 continue;
2277 }
2278 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2279 }
2280 rnp = rcu_get_root(rsp);
2281 if (rnp->qsmask == 0) {
2282 raw_spin_lock_irqsave(&rnp->lock, flags);
2283 smp_mb__after_unlock_lock();
2284 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2285 }
2286}
2287
2288/*
2289 * Force quiescent states on reluctant CPUs, and also detect which
2290 * CPUs are in dyntick-idle mode.
2291 */
2292static void force_quiescent_state(struct rcu_state *rsp)
2293{
2294 unsigned long flags;
2295 bool ret;
2296 struct rcu_node *rnp;
2297 struct rcu_node *rnp_old = NULL;
2298
2299 /* Funnel through hierarchy to reduce memory contention. */
2300 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2301 for (; rnp != NULL; rnp = rnp->parent) {
2302 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2303 !raw_spin_trylock(&rnp->fqslock);
2304 if (rnp_old != NULL)
2305 raw_spin_unlock(&rnp_old->fqslock);
2306 if (ret) {
2307 ACCESS_ONCE(rsp->n_force_qs_lh)++;
2308 return;
2309 }
2310 rnp_old = rnp;
2311 }
2312 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2313
2314 /* Reached the root of the rcu_node tree, acquire lock. */
2315 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2316 smp_mb__after_unlock_lock();
2317 raw_spin_unlock(&rnp_old->fqslock);
2318 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2319 ACCESS_ONCE(rsp->n_force_qs_lh)++;
2320 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2321 return; /* Someone beat us to it. */
2322 }
2323 rsp->gp_flags |= RCU_GP_FLAG_FQS;
2324 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2325 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
2326}
2327
2328/*
2329 * This does the RCU core processing work for the specified rcu_state
2330 * and rcu_data structures. This may be called only from the CPU to
2331 * whom the rdp belongs.
2332 */
2333static void
2334__rcu_process_callbacks(struct rcu_state *rsp)
2335{
2336 unsigned long flags;
2337 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2338
2339 WARN_ON_ONCE(rdp->beenonline == 0);
2340
2341 /* Update RCU state based on any recent quiescent states. */
2342 rcu_check_quiescent_state(rsp, rdp);
2343
2344 /* Does this CPU require a not-yet-started grace period? */
2345 local_irq_save(flags);
2346 if (cpu_needs_another_gp(rsp, rdp)) {
2347 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2348 rcu_start_gp(rsp);
2349 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2350 } else {
2351 local_irq_restore(flags);
2352 }
2353
2354 /* If there are callbacks ready, invoke them. */
2355 if (cpu_has_callbacks_ready_to_invoke(rdp))
2356 invoke_rcu_callbacks(rsp, rdp);
2357
2358 /* Do any needed deferred wakeups of rcuo kthreads. */
2359 do_nocb_deferred_wakeup(rdp);
2360}
2361
2362/*
2363 * Do RCU core processing for the current CPU.
2364 */
2365static void rcu_process_callbacks(struct softirq_action *unused)
2366{
2367 struct rcu_state *rsp;
2368
2369 if (cpu_is_offline(smp_processor_id()))
2370 return;
2371 trace_rcu_utilization(TPS("Start RCU core"));
2372 for_each_rcu_flavor(rsp)
2373 __rcu_process_callbacks(rsp);
2374 trace_rcu_utilization(TPS("End RCU core"));
2375}
2376
2377/*
2378 * Schedule RCU callback invocation. If the specified type of RCU
2379 * does not support RCU priority boosting, just do a direct call,
2380 * otherwise wake up the per-CPU kernel kthread. Note that because we
2381 * are running on the current CPU with interrupts disabled, the
2382 * rcu_cpu_kthread_task cannot disappear out from under us.
2383 */
2384static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2385{
2386 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2387 return;
2388 if (likely(!rsp->boost)) {
2389 rcu_do_batch(rsp, rdp);
2390 return;
2391 }
2392 invoke_rcu_callbacks_kthread();
2393}
2394
2395static void invoke_rcu_core(void)
2396{
2397 if (cpu_online(smp_processor_id()))
2398 raise_softirq(RCU_SOFTIRQ);
2399}
2400
2401/*
2402 * Handle any core-RCU processing required by a call_rcu() invocation.
2403 */
2404static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2405 struct rcu_head *head, unsigned long flags)
2406{
2407 /*
2408 * If called from an extended quiescent state, invoke the RCU
2409 * core in order to force a re-evaluation of RCU's idleness.
2410 */
2411 if (!rcu_is_watching() && cpu_online(smp_processor_id()))
2412 invoke_rcu_core();
2413
2414 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2415 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2416 return;
2417
2418 /*
2419 * Force the grace period if too many callbacks or too long waiting.
2420 * Enforce hysteresis, and don't invoke force_quiescent_state()
2421 * if some other CPU has recently done so. Also, don't bother
2422 * invoking force_quiescent_state() if the newly enqueued callback
2423 * is the only one waiting for a grace period to complete.
2424 */
2425 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2426
2427 /* Are we ignoring a completed grace period? */
2428 note_gp_changes(rsp, rdp);
2429
2430 /* Start a new grace period if one not already started. */
2431 if (!rcu_gp_in_progress(rsp)) {
2432 struct rcu_node *rnp_root = rcu_get_root(rsp);
2433
2434 raw_spin_lock(&rnp_root->lock);
2435 smp_mb__after_unlock_lock();
2436 rcu_start_gp(rsp);
2437 raw_spin_unlock(&rnp_root->lock);
2438 } else {
2439 /* Give the grace period a kick. */
2440 rdp->blimit = LONG_MAX;
2441 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2442 *rdp->nxttail[RCU_DONE_TAIL] != head)
2443 force_quiescent_state(rsp);
2444 rdp->n_force_qs_snap = rsp->n_force_qs;
2445 rdp->qlen_last_fqs_check = rdp->qlen;
2446 }
2447 }
2448}
2449
2450/*
2451 * RCU callback function to leak a callback.
2452 */
2453static void rcu_leak_callback(struct rcu_head *rhp)
2454{
2455}
2456
2457/*
2458 * Helper function for call_rcu() and friends. The cpu argument will
2459 * normally be -1, indicating "currently running CPU". It may specify
2460 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2461 * is expected to specify a CPU.
2462 */
2463static void
2464__call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2465 struct rcu_state *rsp, int cpu, bool lazy)
2466{
2467 unsigned long flags;
2468 struct rcu_data *rdp;
2469
2470 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2471 if (debug_rcu_head_queue(head)) {
2472 /* Probable double call_rcu(), so leak the callback. */
2473 ACCESS_ONCE(head->func) = rcu_leak_callback;
2474 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2475 return;
2476 }
2477 head->func = func;
2478 head->next = NULL;
2479
2480 /*
2481 * Opportunistically note grace-period endings and beginnings.
2482 * Note that we might see a beginning right after we see an
2483 * end, but never vice versa, since this CPU has to pass through
2484 * a quiescent state betweentimes.
2485 */
2486 local_irq_save(flags);
2487 rdp = this_cpu_ptr(rsp->rda);
2488
2489 /* Add the callback to our list. */
2490 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2491 int offline;
2492
2493 if (cpu != -1)
2494 rdp = per_cpu_ptr(rsp->rda, cpu);
2495 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
2496 WARN_ON_ONCE(offline);
2497 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2498 local_irq_restore(flags);
2499 return;
2500 }
2501 ACCESS_ONCE(rdp->qlen)++;
2502 if (lazy)
2503 rdp->qlen_lazy++;
2504 else
2505 rcu_idle_count_callbacks_posted();
2506 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2507 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2508 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2509
2510 if (__is_kfree_rcu_offset((unsigned long)func))
2511 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2512 rdp->qlen_lazy, rdp->qlen);
2513 else
2514 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2515
2516 /* Go handle any RCU core processing required. */
2517 __call_rcu_core(rsp, rdp, head, flags);
2518 local_irq_restore(flags);
2519}
2520
2521/*
2522 * Queue an RCU-sched callback for invocation after a grace period.
2523 */
2524void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2525{
2526 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2527}
2528EXPORT_SYMBOL_GPL(call_rcu_sched);
2529
2530/*
2531 * Queue an RCU callback for invocation after a quicker grace period.
2532 */
2533void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2534{
2535 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2536}
2537EXPORT_SYMBOL_GPL(call_rcu_bh);
2538
2539/*
2540 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2541 * any blocking grace-period wait automatically implies a grace period
2542 * if there is only one CPU online at any point time during execution
2543 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2544 * occasionally incorrectly indicate that there are multiple CPUs online
2545 * when there was in fact only one the whole time, as this just adds
2546 * some overhead: RCU still operates correctly.
2547 */
2548static inline int rcu_blocking_is_gp(void)
2549{
2550 int ret;
2551
2552 might_sleep(); /* Check for RCU read-side critical section. */
2553 preempt_disable();
2554 ret = num_online_cpus() <= 1;
2555 preempt_enable();
2556 return ret;
2557}
2558
2559/**
2560 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2561 *
2562 * Control will return to the caller some time after a full rcu-sched
2563 * grace period has elapsed, in other words after all currently executing
2564 * rcu-sched read-side critical sections have completed. These read-side
2565 * critical sections are delimited by rcu_read_lock_sched() and
2566 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2567 * local_irq_disable(), and so on may be used in place of
2568 * rcu_read_lock_sched().
2569 *
2570 * This means that all preempt_disable code sequences, including NMI and
2571 * non-threaded hardware-interrupt handlers, in progress on entry will
2572 * have completed before this primitive returns. However, this does not
2573 * guarantee that softirq handlers will have completed, since in some
2574 * kernels, these handlers can run in process context, and can block.
2575 *
2576 * Note that this guarantee implies further memory-ordering guarantees.
2577 * On systems with more than one CPU, when synchronize_sched() returns,
2578 * each CPU is guaranteed to have executed a full memory barrier since the
2579 * end of its last RCU-sched read-side critical section whose beginning
2580 * preceded the call to synchronize_sched(). In addition, each CPU having
2581 * an RCU read-side critical section that extends beyond the return from
2582 * synchronize_sched() is guaranteed to have executed a full memory barrier
2583 * after the beginning of synchronize_sched() and before the beginning of
2584 * that RCU read-side critical section. Note that these guarantees include
2585 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2586 * that are executing in the kernel.
2587 *
2588 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2589 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2590 * to have executed a full memory barrier during the execution of
2591 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2592 * again only if the system has more than one CPU).
2593 *
2594 * This primitive provides the guarantees made by the (now removed)
2595 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2596 * guarantees that rcu_read_lock() sections will have completed.
2597 * In "classic RCU", these two guarantees happen to be one and
2598 * the same, but can differ in realtime RCU implementations.
2599 */
2600void synchronize_sched(void)
2601{
2602 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2603 !lock_is_held(&rcu_lock_map) &&
2604 !lock_is_held(&rcu_sched_lock_map),
2605 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2606 if (rcu_blocking_is_gp())
2607 return;
2608 if (rcu_expedited)
2609 synchronize_sched_expedited();
2610 else
2611 wait_rcu_gp(call_rcu_sched);
2612}
2613EXPORT_SYMBOL_GPL(synchronize_sched);
2614
2615/**
2616 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2617 *
2618 * Control will return to the caller some time after a full rcu_bh grace
2619 * period has elapsed, in other words after all currently executing rcu_bh
2620 * read-side critical sections have completed. RCU read-side critical
2621 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2622 * and may be nested.
2623 *
2624 * See the description of synchronize_sched() for more detailed information
2625 * on memory ordering guarantees.
2626 */
2627void synchronize_rcu_bh(void)
2628{
2629 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2630 !lock_is_held(&rcu_lock_map) &&
2631 !lock_is_held(&rcu_sched_lock_map),
2632 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2633 if (rcu_blocking_is_gp())
2634 return;
2635 if (rcu_expedited)
2636 synchronize_rcu_bh_expedited();
2637 else
2638 wait_rcu_gp(call_rcu_bh);
2639}
2640EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2641
2642/**
2643 * get_state_synchronize_rcu - Snapshot current RCU state
2644 *
2645 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
2646 * to determine whether or not a full grace period has elapsed in the
2647 * meantime.
2648 */
2649unsigned long get_state_synchronize_rcu(void)
2650{
2651 /*
2652 * Any prior manipulation of RCU-protected data must happen
2653 * before the load from ->gpnum.
2654 */
2655 smp_mb(); /* ^^^ */
2656
2657 /*
2658 * Make sure this load happens before the purportedly
2659 * time-consuming work between get_state_synchronize_rcu()
2660 * and cond_synchronize_rcu().
2661 */
2662 return smp_load_acquire(&rcu_state->gpnum);
2663}
2664EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
2665
2666/**
2667 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
2668 *
2669 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
2670 *
2671 * If a full RCU grace period has elapsed since the earlier call to
2672 * get_state_synchronize_rcu(), just return. Otherwise, invoke
2673 * synchronize_rcu() to wait for a full grace period.
2674 *
2675 * Yes, this function does not take counter wrap into account. But
2676 * counter wrap is harmless. If the counter wraps, we have waited for
2677 * more than 2 billion grace periods (and way more on a 64-bit system!),
2678 * so waiting for one additional grace period should be just fine.
2679 */
2680void cond_synchronize_rcu(unsigned long oldstate)
2681{
2682 unsigned long newstate;
2683
2684 /*
2685 * Ensure that this load happens before any RCU-destructive
2686 * actions the caller might carry out after we return.
2687 */
2688 newstate = smp_load_acquire(&rcu_state->completed);
2689 if (ULONG_CMP_GE(oldstate, newstate))
2690 synchronize_rcu();
2691}
2692EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
2693
2694static int synchronize_sched_expedited_cpu_stop(void *data)
2695{
2696 /*
2697 * There must be a full memory barrier on each affected CPU
2698 * between the time that try_stop_cpus() is called and the
2699 * time that it returns.
2700 *
2701 * In the current initial implementation of cpu_stop, the
2702 * above condition is already met when the control reaches
2703 * this point and the following smp_mb() is not strictly
2704 * necessary. Do smp_mb() anyway for documentation and
2705 * robustness against future implementation changes.
2706 */
2707 smp_mb(); /* See above comment block. */
2708 return 0;
2709}
2710
2711/**
2712 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2713 *
2714 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2715 * approach to force the grace period to end quickly. This consumes
2716 * significant time on all CPUs and is unfriendly to real-time workloads,
2717 * so is thus not recommended for any sort of common-case code. In fact,
2718 * if you are using synchronize_sched_expedited() in a loop, please
2719 * restructure your code to batch your updates, and then use a single
2720 * synchronize_sched() instead.
2721 *
2722 * Note that it is illegal to call this function while holding any lock
2723 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2724 * to call this function from a CPU-hotplug notifier. Failing to observe
2725 * these restriction will result in deadlock.
2726 *
2727 * This implementation can be thought of as an application of ticket
2728 * locking to RCU, with sync_sched_expedited_started and
2729 * sync_sched_expedited_done taking on the roles of the halves
2730 * of the ticket-lock word. Each task atomically increments
2731 * sync_sched_expedited_started upon entry, snapshotting the old value,
2732 * then attempts to stop all the CPUs. If this succeeds, then each
2733 * CPU will have executed a context switch, resulting in an RCU-sched
2734 * grace period. We are then done, so we use atomic_cmpxchg() to
2735 * update sync_sched_expedited_done to match our snapshot -- but
2736 * only if someone else has not already advanced past our snapshot.
2737 *
2738 * On the other hand, if try_stop_cpus() fails, we check the value
2739 * of sync_sched_expedited_done. If it has advanced past our
2740 * initial snapshot, then someone else must have forced a grace period
2741 * some time after we took our snapshot. In this case, our work is
2742 * done for us, and we can simply return. Otherwise, we try again,
2743 * but keep our initial snapshot for purposes of checking for someone
2744 * doing our work for us.
2745 *
2746 * If we fail too many times in a row, we fall back to synchronize_sched().
2747 */
2748void synchronize_sched_expedited(void)
2749{
2750 long firstsnap, s, snap;
2751 int trycount = 0;
2752 struct rcu_state *rsp = &rcu_sched_state;
2753
2754 /*
2755 * If we are in danger of counter wrap, just do synchronize_sched().
2756 * By allowing sync_sched_expedited_started to advance no more than
2757 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2758 * that more than 3.5 billion CPUs would be required to force a
2759 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2760 * course be required on a 64-bit system.
2761 */
2762 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2763 (ulong)atomic_long_read(&rsp->expedited_done) +
2764 ULONG_MAX / 8)) {
2765 synchronize_sched();
2766 atomic_long_inc(&rsp->expedited_wrap);
2767 return;
2768 }
2769
2770 /*
2771 * Take a ticket. Note that atomic_inc_return() implies a
2772 * full memory barrier.
2773 */
2774 snap = atomic_long_inc_return(&rsp->expedited_start);
2775 firstsnap = snap;
2776 get_online_cpus();
2777 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2778
2779 /*
2780 * Each pass through the following loop attempts to force a
2781 * context switch on each CPU.
2782 */
2783 while (try_stop_cpus(cpu_online_mask,
2784 synchronize_sched_expedited_cpu_stop,
2785 NULL) == -EAGAIN) {
2786 put_online_cpus();
2787 atomic_long_inc(&rsp->expedited_tryfail);
2788
2789 /* Check to see if someone else did our work for us. */
2790 s = atomic_long_read(&rsp->expedited_done);
2791 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2792 /* ensure test happens before caller kfree */
2793 smp_mb__before_atomic_inc(); /* ^^^ */
2794 atomic_long_inc(&rsp->expedited_workdone1);
2795 return;
2796 }
2797
2798 /* No joy, try again later. Or just synchronize_sched(). */
2799 if (trycount++ < 10) {
2800 udelay(trycount * num_online_cpus());
2801 } else {
2802 wait_rcu_gp(call_rcu_sched);
2803 atomic_long_inc(&rsp->expedited_normal);
2804 return;
2805 }
2806
2807 /* Recheck to see if someone else did our work for us. */
2808 s = atomic_long_read(&rsp->expedited_done);
2809 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2810 /* ensure test happens before caller kfree */
2811 smp_mb__before_atomic_inc(); /* ^^^ */
2812 atomic_long_inc(&rsp->expedited_workdone2);
2813 return;
2814 }
2815
2816 /*
2817 * Refetching sync_sched_expedited_started allows later
2818 * callers to piggyback on our grace period. We retry
2819 * after they started, so our grace period works for them,
2820 * and they started after our first try, so their grace
2821 * period works for us.
2822 */
2823 get_online_cpus();
2824 snap = atomic_long_read(&rsp->expedited_start);
2825 smp_mb(); /* ensure read is before try_stop_cpus(). */
2826 }
2827 atomic_long_inc(&rsp->expedited_stoppedcpus);
2828
2829 /*
2830 * Everyone up to our most recent fetch is covered by our grace
2831 * period. Update the counter, but only if our work is still
2832 * relevant -- which it won't be if someone who started later
2833 * than we did already did their update.
2834 */
2835 do {
2836 atomic_long_inc(&rsp->expedited_done_tries);
2837 s = atomic_long_read(&rsp->expedited_done);
2838 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2839 /* ensure test happens before caller kfree */
2840 smp_mb__before_atomic_inc(); /* ^^^ */
2841 atomic_long_inc(&rsp->expedited_done_lost);
2842 break;
2843 }
2844 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2845 atomic_long_inc(&rsp->expedited_done_exit);
2846
2847 put_online_cpus();
2848}
2849EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2850
2851/*
2852 * Check to see if there is any immediate RCU-related work to be done
2853 * by the current CPU, for the specified type of RCU, returning 1 if so.
2854 * The checks are in order of increasing expense: checks that can be
2855 * carried out against CPU-local state are performed first. However,
2856 * we must check for CPU stalls first, else we might not get a chance.
2857 */
2858static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2859{
2860 struct rcu_node *rnp = rdp->mynode;
2861
2862 rdp->n_rcu_pending++;
2863
2864 /* Check for CPU stalls, if enabled. */
2865 check_cpu_stall(rsp, rdp);
2866
2867 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
2868 if (rcu_nohz_full_cpu(rsp))
2869 return 0;
2870
2871 /* Is the RCU core waiting for a quiescent state from this CPU? */
2872 if (rcu_scheduler_fully_active &&
2873 rdp->qs_pending && !rdp->passed_quiesce) {
2874 rdp->n_rp_qs_pending++;
2875 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2876 rdp->n_rp_report_qs++;
2877 return 1;
2878 }
2879
2880 /* Does this CPU have callbacks ready to invoke? */
2881 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2882 rdp->n_rp_cb_ready++;
2883 return 1;
2884 }
2885
2886 /* Has RCU gone idle with this CPU needing another grace period? */
2887 if (cpu_needs_another_gp(rsp, rdp)) {
2888 rdp->n_rp_cpu_needs_gp++;
2889 return 1;
2890 }
2891
2892 /* Has another RCU grace period completed? */
2893 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2894 rdp->n_rp_gp_completed++;
2895 return 1;
2896 }
2897
2898 /* Has a new RCU grace period started? */
2899 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2900 rdp->n_rp_gp_started++;
2901 return 1;
2902 }
2903
2904 /* Does this CPU need a deferred NOCB wakeup? */
2905 if (rcu_nocb_need_deferred_wakeup(rdp)) {
2906 rdp->n_rp_nocb_defer_wakeup++;
2907 return 1;
2908 }
2909
2910 /* nothing to do */
2911 rdp->n_rp_need_nothing++;
2912 return 0;
2913}
2914
2915/*
2916 * Check to see if there is any immediate RCU-related work to be done
2917 * by the current CPU, returning 1 if so. This function is part of the
2918 * RCU implementation; it is -not- an exported member of the RCU API.
2919 */
2920static int rcu_pending(int cpu)
2921{
2922 struct rcu_state *rsp;
2923
2924 for_each_rcu_flavor(rsp)
2925 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2926 return 1;
2927 return 0;
2928}
2929
2930/*
2931 * Return true if the specified CPU has any callback. If all_lazy is
2932 * non-NULL, store an indication of whether all callbacks are lazy.
2933 * (If there are no callbacks, all of them are deemed to be lazy.)
2934 */
2935static int __maybe_unused rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
2936{
2937 bool al = true;
2938 bool hc = false;
2939 struct rcu_data *rdp;
2940 struct rcu_state *rsp;
2941
2942 for_each_rcu_flavor(rsp) {
2943 rdp = per_cpu_ptr(rsp->rda, cpu);
2944 if (!rdp->nxtlist)
2945 continue;
2946 hc = true;
2947 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
2948 al = false;
2949 break;
2950 }
2951 }
2952 if (all_lazy)
2953 *all_lazy = al;
2954 return hc;
2955}
2956
2957/*
2958 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2959 * the compiler is expected to optimize this away.
2960 */
2961static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
2962 int cpu, unsigned long done)
2963{
2964 trace_rcu_barrier(rsp->name, s, cpu,
2965 atomic_read(&rsp->barrier_cpu_count), done);
2966}
2967
2968/*
2969 * RCU callback function for _rcu_barrier(). If we are last, wake
2970 * up the task executing _rcu_barrier().
2971 */
2972static void rcu_barrier_callback(struct rcu_head *rhp)
2973{
2974 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2975 struct rcu_state *rsp = rdp->rsp;
2976
2977 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2978 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2979 complete(&rsp->barrier_completion);
2980 } else {
2981 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2982 }
2983}
2984
2985/*
2986 * Called with preemption disabled, and from cross-cpu IRQ context.
2987 */
2988static void rcu_barrier_func(void *type)
2989{
2990 struct rcu_state *rsp = type;
2991 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2992
2993 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2994 atomic_inc(&rsp->barrier_cpu_count);
2995 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2996}
2997
2998/*
2999 * Orchestrate the specified type of RCU barrier, waiting for all
3000 * RCU callbacks of the specified type to complete.
3001 */
3002static void _rcu_barrier(struct rcu_state *rsp)
3003{
3004 int cpu;
3005 struct rcu_data *rdp;
3006 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
3007 unsigned long snap_done;
3008
3009 _rcu_barrier_trace(rsp, "Begin", -1, snap);
3010
3011 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3012 mutex_lock(&rsp->barrier_mutex);
3013
3014 /*
3015 * Ensure that all prior references, including to ->n_barrier_done,
3016 * are ordered before the _rcu_barrier() machinery.
3017 */
3018 smp_mb(); /* See above block comment. */
3019
3020 /*
3021 * Recheck ->n_barrier_done to see if others did our work for us.
3022 * This means checking ->n_barrier_done for an even-to-odd-to-even
3023 * transition. The "if" expression below therefore rounds the old
3024 * value up to the next even number and adds two before comparing.
3025 */
3026 snap_done = rsp->n_barrier_done;
3027 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
3028
3029 /*
3030 * If the value in snap is odd, we needed to wait for the current
3031 * rcu_barrier() to complete, then wait for the next one, in other
3032 * words, we need the value of snap_done to be three larger than
3033 * the value of snap. On the other hand, if the value in snap is
3034 * even, we only had to wait for the next rcu_barrier() to complete,
3035 * in other words, we need the value of snap_done to be only two
3036 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3037 * this for us (thank you, Linus!).
3038 */
3039 if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) {
3040 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
3041 smp_mb(); /* caller's subsequent code after above check. */
3042 mutex_unlock(&rsp->barrier_mutex);
3043 return;
3044 }
3045
3046 /*
3047 * Increment ->n_barrier_done to avoid duplicate work. Use
3048 * ACCESS_ONCE() to prevent the compiler from speculating
3049 * the increment to precede the early-exit check.
3050 */
3051 ACCESS_ONCE(rsp->n_barrier_done)++;
3052 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
3053 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
3054 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3055
3056 /*
3057 * Initialize the count to one rather than to zero in order to
3058 * avoid a too-soon return to zero in case of a short grace period
3059 * (or preemption of this task). Exclude CPU-hotplug operations
3060 * to ensure that no offline CPU has callbacks queued.
3061 */
3062 init_completion(&rsp->barrier_completion);
3063 atomic_set(&rsp->barrier_cpu_count, 1);
3064 get_online_cpus();
3065
3066 /*
3067 * Force each CPU with callbacks to register a new callback.
3068 * When that callback is invoked, we will know that all of the
3069 * corresponding CPU's preceding callbacks have been invoked.
3070 */
3071 for_each_possible_cpu(cpu) {
3072 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3073 continue;
3074 rdp = per_cpu_ptr(rsp->rda, cpu);
3075 if (rcu_is_nocb_cpu(cpu)) {
3076 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
3077 rsp->n_barrier_done);
3078 atomic_inc(&rsp->barrier_cpu_count);
3079 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
3080 rsp, cpu, 0);
3081 } else if (ACCESS_ONCE(rdp->qlen)) {
3082 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
3083 rsp->n_barrier_done);
3084 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3085 } else {
3086 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
3087 rsp->n_barrier_done);
3088 }
3089 }
3090 put_online_cpus();
3091
3092 /*
3093 * Now that we have an rcu_barrier_callback() callback on each
3094 * CPU, and thus each counted, remove the initial count.
3095 */
3096 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3097 complete(&rsp->barrier_completion);
3098
3099 /* Increment ->n_barrier_done to prevent duplicate work. */
3100 smp_mb(); /* Keep increment after above mechanism. */
3101 ACCESS_ONCE(rsp->n_barrier_done)++;
3102 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
3103 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
3104 smp_mb(); /* Keep increment before caller's subsequent code. */
3105
3106 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3107 wait_for_completion(&rsp->barrier_completion);
3108
3109 /* Other rcu_barrier() invocations can now safely proceed. */
3110 mutex_unlock(&rsp->barrier_mutex);
3111}
3112
3113/**
3114 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3115 */
3116void rcu_barrier_bh(void)
3117{
3118 _rcu_barrier(&rcu_bh_state);
3119}
3120EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3121
3122/**
3123 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3124 */
3125void rcu_barrier_sched(void)
3126{
3127 _rcu_barrier(&rcu_sched_state);
3128}
3129EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3130
3131/*
3132 * Do boot-time initialization of a CPU's per-CPU RCU data.
3133 */
3134static void __init
3135rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3136{
3137 unsigned long flags;
3138 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3139 struct rcu_node *rnp = rcu_get_root(rsp);
3140
3141 /* Set up local state, ensuring consistent view of global state. */
3142 raw_spin_lock_irqsave(&rnp->lock, flags);
3143 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
3144 init_callback_list(rdp);
3145 rdp->qlen_lazy = 0;
3146 ACCESS_ONCE(rdp->qlen) = 0;
3147 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3148 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3149 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
3150 rdp->cpu = cpu;
3151 rdp->rsp = rsp;
3152 rcu_boot_init_nocb_percpu_data(rdp);
3153 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3154}
3155
3156/*
3157 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3158 * offline event can be happening at a given time. Note also that we
3159 * can accept some slop in the rsp->completed access due to the fact
3160 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3161 */
3162static void
3163rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
3164{
3165 unsigned long flags;
3166 unsigned long mask;
3167 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3168 struct rcu_node *rnp = rcu_get_root(rsp);
3169
3170 /* Exclude new grace periods. */
3171 mutex_lock(&rsp->onoff_mutex);
3172
3173 /* Set up local state, ensuring consistent view of global state. */
3174 raw_spin_lock_irqsave(&rnp->lock, flags);
3175 rdp->beenonline = 1; /* We have now been online. */
3176 rdp->preemptible = preemptible;
3177 rdp->qlen_last_fqs_check = 0;
3178 rdp->n_force_qs_snap = rsp->n_force_qs;
3179 rdp->blimit = blimit;
3180 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3181 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3182 rcu_sysidle_init_percpu_data(rdp->dynticks);
3183 atomic_set(&rdp->dynticks->dynticks,
3184 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
3185 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
3186
3187 /* Add CPU to rcu_node bitmasks. */
3188 rnp = rdp->mynode;
3189 mask = rdp->grpmask;
3190 do {
3191 /* Exclude any attempts to start a new GP on small systems. */
3192 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
3193 rnp->qsmaskinit |= mask;
3194 mask = rnp->grpmask;
3195 if (rnp == rdp->mynode) {
3196 /*
3197 * If there is a grace period in progress, we will
3198 * set up to wait for it next time we run the
3199 * RCU core code.
3200 */
3201 rdp->gpnum = rnp->completed;
3202 rdp->completed = rnp->completed;
3203 rdp->passed_quiesce = 0;
3204 rdp->qs_pending = 0;
3205 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3206 }
3207 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
3208 rnp = rnp->parent;
3209 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
3210 local_irq_restore(flags);
3211
3212 mutex_unlock(&rsp->onoff_mutex);
3213}
3214
3215static void rcu_prepare_cpu(int cpu)
3216{
3217 struct rcu_state *rsp;
3218
3219 for_each_rcu_flavor(rsp)
3220 rcu_init_percpu_data(cpu, rsp,
3221 strcmp(rsp->name, "rcu_preempt") == 0);
3222}
3223
3224/*
3225 * Handle CPU online/offline notification events.
3226 */
3227static int rcu_cpu_notify(struct notifier_block *self,
3228 unsigned long action, void *hcpu)
3229{
3230 long cpu = (long)hcpu;
3231 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
3232 struct rcu_node *rnp = rdp->mynode;
3233 struct rcu_state *rsp;
3234
3235 trace_rcu_utilization(TPS("Start CPU hotplug"));
3236 switch (action) {
3237 case CPU_UP_PREPARE:
3238 case CPU_UP_PREPARE_FROZEN:
3239 rcu_prepare_cpu(cpu);
3240 rcu_prepare_kthreads(cpu);
3241 break;
3242 case CPU_ONLINE:
3243 case CPU_DOWN_FAILED:
3244 rcu_boost_kthread_setaffinity(rnp, -1);
3245 break;
3246 case CPU_DOWN_PREPARE:
3247 rcu_boost_kthread_setaffinity(rnp, cpu);
3248 break;
3249 case CPU_DYING:
3250 case CPU_DYING_FROZEN:
3251 for_each_rcu_flavor(rsp)
3252 rcu_cleanup_dying_cpu(rsp);
3253 break;
3254 case CPU_DEAD:
3255 case CPU_DEAD_FROZEN:
3256 case CPU_UP_CANCELED:
3257 case CPU_UP_CANCELED_FROZEN:
3258 for_each_rcu_flavor(rsp)
3259 rcu_cleanup_dead_cpu(cpu, rsp);
3260 break;
3261 default:
3262 break;
3263 }
3264 trace_rcu_utilization(TPS("End CPU hotplug"));
3265 return NOTIFY_OK;
3266}
3267
3268static int rcu_pm_notify(struct notifier_block *self,
3269 unsigned long action, void *hcpu)
3270{
3271 switch (action) {
3272 case PM_HIBERNATION_PREPARE:
3273 case PM_SUSPEND_PREPARE:
3274 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3275 rcu_expedited = 1;
3276 break;
3277 case PM_POST_HIBERNATION:
3278 case PM_POST_SUSPEND:
3279 rcu_expedited = 0;
3280 break;
3281 default:
3282 break;
3283 }
3284 return NOTIFY_OK;
3285}
3286
3287/*
3288 * Spawn the kthread that handles this RCU flavor's grace periods.
3289 */
3290static int __init rcu_spawn_gp_kthread(void)
3291{
3292 unsigned long flags;
3293 struct rcu_node *rnp;
3294 struct rcu_state *rsp;
3295 struct task_struct *t;
3296
3297 for_each_rcu_flavor(rsp) {
3298 t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
3299 BUG_ON(IS_ERR(t));
3300 rnp = rcu_get_root(rsp);
3301 raw_spin_lock_irqsave(&rnp->lock, flags);
3302 rsp->gp_kthread = t;
3303 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3304 rcu_spawn_nocb_kthreads(rsp);
3305 }
3306 return 0;
3307}
3308early_initcall(rcu_spawn_gp_kthread);
3309
3310/*
3311 * This function is invoked towards the end of the scheduler's initialization
3312 * process. Before this is called, the idle task might contain
3313 * RCU read-side critical sections (during which time, this idle
3314 * task is booting the system). After this function is called, the
3315 * idle tasks are prohibited from containing RCU read-side critical
3316 * sections. This function also enables RCU lockdep checking.
3317 */
3318void rcu_scheduler_starting(void)
3319{
3320 WARN_ON(num_online_cpus() != 1);
3321 WARN_ON(nr_context_switches() > 0);
3322 rcu_scheduler_active = 1;
3323}
3324
3325/*
3326 * Compute the per-level fanout, either using the exact fanout specified
3327 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3328 */
3329#ifdef CONFIG_RCU_FANOUT_EXACT
3330static void __init rcu_init_levelspread(struct rcu_state *rsp)
3331{
3332 int i;
3333
3334 rsp->levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
3335 for (i = rcu_num_lvls - 2; i >= 0; i--)
3336 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3337}
3338#else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3339static void __init rcu_init_levelspread(struct rcu_state *rsp)
3340{
3341 int ccur;
3342 int cprv;
3343 int i;
3344
3345 cprv = nr_cpu_ids;
3346 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3347 ccur = rsp->levelcnt[i];
3348 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3349 cprv = ccur;
3350 }
3351}
3352#endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3353
3354/*
3355 * Helper function for rcu_init() that initializes one rcu_state structure.
3356 */
3357static void __init rcu_init_one(struct rcu_state *rsp,
3358 struct rcu_data __percpu *rda)
3359{
3360 static char *buf[] = { "rcu_node_0",
3361 "rcu_node_1",
3362 "rcu_node_2",
3363 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3364 static char *fqs[] = { "rcu_node_fqs_0",
3365 "rcu_node_fqs_1",
3366 "rcu_node_fqs_2",
3367 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3368 int cpustride = 1;
3369 int i;
3370 int j;
3371 struct rcu_node *rnp;
3372
3373 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3374
3375 /* Silence gcc 4.8 warning about array index out of range. */
3376 if (rcu_num_lvls > RCU_NUM_LVLS)
3377 panic("rcu_init_one: rcu_num_lvls overflow");
3378
3379 /* Initialize the level-tracking arrays. */
3380
3381 for (i = 0; i < rcu_num_lvls; i++)
3382 rsp->levelcnt[i] = num_rcu_lvl[i];
3383 for (i = 1; i < rcu_num_lvls; i++)
3384 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3385 rcu_init_levelspread(rsp);
3386
3387 /* Initialize the elements themselves, starting from the leaves. */
3388
3389 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3390 cpustride *= rsp->levelspread[i];
3391 rnp = rsp->level[i];
3392 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3393 raw_spin_lock_init(&rnp->lock);
3394 lockdep_set_class_and_name(&rnp->lock,
3395 &rcu_node_class[i], buf[i]);
3396 raw_spin_lock_init(&rnp->fqslock);
3397 lockdep_set_class_and_name(&rnp->fqslock,
3398 &rcu_fqs_class[i], fqs[i]);
3399 rnp->gpnum = rsp->gpnum;
3400 rnp->completed = rsp->completed;
3401 rnp->qsmask = 0;
3402 rnp->qsmaskinit = 0;
3403 rnp->grplo = j * cpustride;
3404 rnp->grphi = (j + 1) * cpustride - 1;
3405 if (rnp->grphi >= NR_CPUS)
3406 rnp->grphi = NR_CPUS - 1;
3407 if (i == 0) {
3408 rnp->grpnum = 0;
3409 rnp->grpmask = 0;
3410 rnp->parent = NULL;
3411 } else {
3412 rnp->grpnum = j % rsp->levelspread[i - 1];
3413 rnp->grpmask = 1UL << rnp->grpnum;
3414 rnp->parent = rsp->level[i - 1] +
3415 j / rsp->levelspread[i - 1];
3416 }
3417 rnp->level = i;
3418 INIT_LIST_HEAD(&rnp->blkd_tasks);
3419 rcu_init_one_nocb(rnp);
3420 }
3421 }
3422
3423 rsp->rda = rda;
3424 init_waitqueue_head(&rsp->gp_wq);
3425 init_irq_work(&rsp->wakeup_work, rsp_wakeup);
3426 rnp = rsp->level[rcu_num_lvls - 1];
3427 for_each_possible_cpu(i) {
3428 while (i > rnp->grphi)
3429 rnp++;
3430 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3431 rcu_boot_init_percpu_data(i, rsp);
3432 }
3433 list_add(&rsp->flavors, &rcu_struct_flavors);
3434}
3435
3436/*
3437 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3438 * replace the definitions in tree.h because those are needed to size
3439 * the ->node array in the rcu_state structure.
3440 */
3441static void __init rcu_init_geometry(void)
3442{
3443 ulong d;
3444 int i;
3445 int j;
3446 int n = nr_cpu_ids;
3447 int rcu_capacity[MAX_RCU_LVLS + 1];
3448
3449 /*
3450 * Initialize any unspecified boot parameters.
3451 * The default values of jiffies_till_first_fqs and
3452 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3453 * value, which is a function of HZ, then adding one for each
3454 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3455 */
3456 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3457 if (jiffies_till_first_fqs == ULONG_MAX)
3458 jiffies_till_first_fqs = d;
3459 if (jiffies_till_next_fqs == ULONG_MAX)
3460 jiffies_till_next_fqs = d;
3461
3462 /* If the compile-time values are accurate, just leave. */
3463 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3464 nr_cpu_ids == NR_CPUS)
3465 return;
3466 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
3467 rcu_fanout_leaf, nr_cpu_ids);
3468
3469 /*
3470 * Compute number of nodes that can be handled an rcu_node tree
3471 * with the given number of levels. Setting rcu_capacity[0] makes
3472 * some of the arithmetic easier.
3473 */
3474 rcu_capacity[0] = 1;
3475 rcu_capacity[1] = rcu_fanout_leaf;
3476 for (i = 2; i <= MAX_RCU_LVLS; i++)
3477 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3478
3479 /*
3480 * The boot-time rcu_fanout_leaf parameter is only permitted
3481 * to increase the leaf-level fanout, not decrease it. Of course,
3482 * the leaf-level fanout cannot exceed the number of bits in
3483 * the rcu_node masks. Finally, the tree must be able to accommodate
3484 * the configured number of CPUs. Complain and fall back to the
3485 * compile-time values if these limits are exceeded.
3486 */
3487 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3488 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3489 n > rcu_capacity[MAX_RCU_LVLS]) {
3490 WARN_ON(1);
3491 return;
3492 }
3493
3494 /* Calculate the number of rcu_nodes at each level of the tree. */
3495 for (i = 1; i <= MAX_RCU_LVLS; i++)
3496 if (n <= rcu_capacity[i]) {
3497 for (j = 0; j <= i; j++)
3498 num_rcu_lvl[j] =
3499 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3500 rcu_num_lvls = i;
3501 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3502 num_rcu_lvl[j] = 0;
3503 break;
3504 }
3505
3506 /* Calculate the total number of rcu_node structures. */
3507 rcu_num_nodes = 0;
3508 for (i = 0; i <= MAX_RCU_LVLS; i++)
3509 rcu_num_nodes += num_rcu_lvl[i];
3510 rcu_num_nodes -= n;
3511}
3512
3513void __init rcu_init(void)
3514{
3515 int cpu;
3516
3517 rcu_bootup_announce();
3518 rcu_init_geometry();
3519 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3520 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3521 __rcu_init_preempt();
3522 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3523
3524 /*
3525 * We don't need protection against CPU-hotplug here because
3526 * this is called early in boot, before either interrupts
3527 * or the scheduler are operational.
3528 */
3529 cpu_notifier(rcu_cpu_notify, 0);
3530 pm_notifier(rcu_pm_notify, 0);
3531 for_each_online_cpu(cpu)
3532 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3533}
3534
3535#include "tree_plugin.h"