Loading...
1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
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>
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/random.h>
47#include <linux/trace_events.h>
48#include <linux/suspend.h>
49#include <linux/ftrace.h>
50#include <linux/tick.h>
51#include <linux/sysrq.h>
52#include <linux/kprobes.h>
53#include <linux/gfp.h>
54#include <linux/oom.h>
55#include <linux/smpboot.h>
56#include <linux/jiffies.h>
57#include <linux/slab.h>
58#include <linux/sched/isolation.h>
59#include <linux/sched/clock.h>
60#include <linux/vmalloc.h>
61#include <linux/mm.h>
62#include <linux/kasan.h>
63#include "../time/tick-internal.h"
64
65#include "tree.h"
66#include "rcu.h"
67
68#ifdef MODULE_PARAM_PREFIX
69#undef MODULE_PARAM_PREFIX
70#endif
71#define MODULE_PARAM_PREFIX "rcutree."
72
73#ifndef data_race
74#define data_race(expr) \
75 ({ \
76 expr; \
77 })
78#endif
79#ifndef ASSERT_EXCLUSIVE_WRITER
80#define ASSERT_EXCLUSIVE_WRITER(var) do { } while (0)
81#endif
82#ifndef ASSERT_EXCLUSIVE_ACCESS
83#define ASSERT_EXCLUSIVE_ACCESS(var) do { } while (0)
84#endif
85
86/* Data structures. */
87
88/*
89 * Steal a bit from the bottom of ->dynticks for idle entry/exit
90 * control. Initially this is for TLB flushing.
91 */
92#define RCU_DYNTICK_CTRL_MASK 0x1
93#define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
94
95static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = {
96 .dynticks_nesting = 1,
97 .dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE,
98 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
99};
100static struct rcu_state rcu_state = {
101 .level = { &rcu_state.node[0] },
102 .gp_state = RCU_GP_IDLE,
103 .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT,
104 .barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex),
105 .name = RCU_NAME,
106 .abbr = RCU_ABBR,
107 .exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
108 .exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
109 .ofl_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.ofl_lock),
110};
111
112/* Dump rcu_node combining tree at boot to verify correct setup. */
113static bool dump_tree;
114module_param(dump_tree, bool, 0444);
115/* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */
116static bool use_softirq = true;
117module_param(use_softirq, bool, 0444);
118/* Control rcu_node-tree auto-balancing at boot time. */
119static bool rcu_fanout_exact;
120module_param(rcu_fanout_exact, bool, 0444);
121/* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
122static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
123module_param(rcu_fanout_leaf, int, 0444);
124int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
125/* Number of rcu_nodes at specified level. */
126int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
127int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
128
129/*
130 * The rcu_scheduler_active variable is initialized to the value
131 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
132 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
133 * RCU can assume that there is but one task, allowing RCU to (for example)
134 * optimize synchronize_rcu() to a simple barrier(). When this variable
135 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
136 * to detect real grace periods. This variable is also used to suppress
137 * boot-time false positives from lockdep-RCU error checking. Finally, it
138 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
139 * is fully initialized, including all of its kthreads having been spawned.
140 */
141int rcu_scheduler_active __read_mostly;
142EXPORT_SYMBOL_GPL(rcu_scheduler_active);
143
144/*
145 * The rcu_scheduler_fully_active variable transitions from zero to one
146 * during the early_initcall() processing, which is after the scheduler
147 * is capable of creating new tasks. So RCU processing (for example,
148 * creating tasks for RCU priority boosting) must be delayed until after
149 * rcu_scheduler_fully_active transitions from zero to one. We also
150 * currently delay invocation of any RCU callbacks until after this point.
151 *
152 * It might later prove better for people registering RCU callbacks during
153 * early boot to take responsibility for these callbacks, but one step at
154 * a time.
155 */
156static int rcu_scheduler_fully_active __read_mostly;
157
158static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
159 unsigned long gps, unsigned long flags);
160static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
161static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
162static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
163static void invoke_rcu_core(void);
164static void rcu_report_exp_rdp(struct rcu_data *rdp);
165static void sync_sched_exp_online_cleanup(int cpu);
166static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp);
167
168/* rcuc/rcub kthread realtime priority */
169static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
170module_param(kthread_prio, int, 0444);
171
172/* Delay in jiffies for grace-period initialization delays, debug only. */
173
174static int gp_preinit_delay;
175module_param(gp_preinit_delay, int, 0444);
176static int gp_init_delay;
177module_param(gp_init_delay, int, 0444);
178static int gp_cleanup_delay;
179module_param(gp_cleanup_delay, int, 0444);
180
181/*
182 * This rcu parameter is runtime-read-only. It reflects
183 * a minimum allowed number of objects which can be cached
184 * per-CPU. Object size is equal to one page. This value
185 * can be changed at boot time.
186 */
187static int rcu_min_cached_objs = 2;
188module_param(rcu_min_cached_objs, int, 0444);
189
190/* Retrieve RCU kthreads priority for rcutorture */
191int rcu_get_gp_kthreads_prio(void)
192{
193 return kthread_prio;
194}
195EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio);
196
197/*
198 * Number of grace periods between delays, normalized by the duration of
199 * the delay. The longer the delay, the more the grace periods between
200 * each delay. The reason for this normalization is that it means that,
201 * for non-zero delays, the overall slowdown of grace periods is constant
202 * regardless of the duration of the delay. This arrangement balances
203 * the need for long delays to increase some race probabilities with the
204 * need for fast grace periods to increase other race probabilities.
205 */
206#define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
207
208/*
209 * Compute the mask of online CPUs for the specified rcu_node structure.
210 * This will not be stable unless the rcu_node structure's ->lock is
211 * held, but the bit corresponding to the current CPU will be stable
212 * in most contexts.
213 */
214static unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
215{
216 return READ_ONCE(rnp->qsmaskinitnext);
217}
218
219/*
220 * Return true if an RCU grace period is in progress. The READ_ONCE()s
221 * permit this function to be invoked without holding the root rcu_node
222 * structure's ->lock, but of course results can be subject to change.
223 */
224static int rcu_gp_in_progress(void)
225{
226 return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq));
227}
228
229/*
230 * Return the number of callbacks queued on the specified CPU.
231 * Handles both the nocbs and normal cases.
232 */
233static long rcu_get_n_cbs_cpu(int cpu)
234{
235 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
236
237 if (rcu_segcblist_is_enabled(&rdp->cblist))
238 return rcu_segcblist_n_cbs(&rdp->cblist);
239 return 0;
240}
241
242void rcu_softirq_qs(void)
243{
244 rcu_qs();
245 rcu_preempt_deferred_qs(current);
246}
247
248/*
249 * Record entry into an extended quiescent state. This is only to be
250 * called when not already in an extended quiescent state, that is,
251 * RCU is watching prior to the call to this function and is no longer
252 * watching upon return.
253 */
254static noinstr void rcu_dynticks_eqs_enter(void)
255{
256 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
257 int seq;
258
259 /*
260 * CPUs seeing atomic_add_return() must see prior RCU read-side
261 * critical sections, and we also must force ordering with the
262 * next idle sojourn.
263 */
264 rcu_dynticks_task_trace_enter(); // Before ->dynticks update!
265 seq = arch_atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
266 // RCU is no longer watching. Better be in extended quiescent state!
267 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
268 (seq & RCU_DYNTICK_CTRL_CTR));
269 /* Better not have special action (TLB flush) pending! */
270 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
271 (seq & RCU_DYNTICK_CTRL_MASK));
272}
273
274/*
275 * Record exit from an extended quiescent state. This is only to be
276 * called from an extended quiescent state, that is, RCU is not watching
277 * prior to the call to this function and is watching upon return.
278 */
279static noinstr void rcu_dynticks_eqs_exit(void)
280{
281 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
282 int seq;
283
284 /*
285 * CPUs seeing atomic_add_return() must see prior idle sojourns,
286 * and we also must force ordering with the next RCU read-side
287 * critical section.
288 */
289 seq = arch_atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
290 // RCU is now watching. Better not be in an extended quiescent state!
291 rcu_dynticks_task_trace_exit(); // After ->dynticks update!
292 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
293 !(seq & RCU_DYNTICK_CTRL_CTR));
294 if (seq & RCU_DYNTICK_CTRL_MASK) {
295 arch_atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdp->dynticks);
296 smp_mb__after_atomic(); /* _exit after clearing mask. */
297 }
298}
299
300/*
301 * Reset the current CPU's ->dynticks counter to indicate that the
302 * newly onlined CPU is no longer in an extended quiescent state.
303 * This will either leave the counter unchanged, or increment it
304 * to the next non-quiescent value.
305 *
306 * The non-atomic test/increment sequence works because the upper bits
307 * of the ->dynticks counter are manipulated only by the corresponding CPU,
308 * or when the corresponding CPU is offline.
309 */
310static void rcu_dynticks_eqs_online(void)
311{
312 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
313
314 if (atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR)
315 return;
316 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
317}
318
319/*
320 * Is the current CPU in an extended quiescent state?
321 *
322 * No ordering, as we are sampling CPU-local information.
323 */
324static __always_inline bool rcu_dynticks_curr_cpu_in_eqs(void)
325{
326 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
327
328 return !(arch_atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR);
329}
330
331/*
332 * Snapshot the ->dynticks counter with full ordering so as to allow
333 * stable comparison of this counter with past and future snapshots.
334 */
335static int rcu_dynticks_snap(struct rcu_data *rdp)
336{
337 int snap = atomic_add_return(0, &rdp->dynticks);
338
339 return snap & ~RCU_DYNTICK_CTRL_MASK;
340}
341
342/*
343 * Return true if the snapshot returned from rcu_dynticks_snap()
344 * indicates that RCU is in an extended quiescent state.
345 */
346static bool rcu_dynticks_in_eqs(int snap)
347{
348 return !(snap & RCU_DYNTICK_CTRL_CTR);
349}
350
351/*
352 * Return true if the CPU corresponding to the specified rcu_data
353 * structure has spent some time in an extended quiescent state since
354 * rcu_dynticks_snap() returned the specified snapshot.
355 */
356static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap)
357{
358 return snap != rcu_dynticks_snap(rdp);
359}
360
361/*
362 * Return true if the referenced integer is zero while the specified
363 * CPU remains within a single extended quiescent state.
364 */
365bool rcu_dynticks_zero_in_eqs(int cpu, int *vp)
366{
367 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
368 int snap;
369
370 // If not quiescent, force back to earlier extended quiescent state.
371 snap = atomic_read(&rdp->dynticks) & ~(RCU_DYNTICK_CTRL_MASK |
372 RCU_DYNTICK_CTRL_CTR);
373
374 smp_rmb(); // Order ->dynticks and *vp reads.
375 if (READ_ONCE(*vp))
376 return false; // Non-zero, so report failure;
377 smp_rmb(); // Order *vp read and ->dynticks re-read.
378
379 // If still in the same extended quiescent state, we are good!
380 return snap == (atomic_read(&rdp->dynticks) & ~RCU_DYNTICK_CTRL_MASK);
381}
382
383/*
384 * Set the special (bottom) bit of the specified CPU so that it
385 * will take special action (such as flushing its TLB) on the
386 * next exit from an extended quiescent state. Returns true if
387 * the bit was successfully set, or false if the CPU was not in
388 * an extended quiescent state.
389 */
390bool rcu_eqs_special_set(int cpu)
391{
392 int old;
393 int new;
394 int new_old;
395 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
396
397 new_old = atomic_read(&rdp->dynticks);
398 do {
399 old = new_old;
400 if (old & RCU_DYNTICK_CTRL_CTR)
401 return false;
402 new = old | RCU_DYNTICK_CTRL_MASK;
403 new_old = atomic_cmpxchg(&rdp->dynticks, old, new);
404 } while (new_old != old);
405 return true;
406}
407
408/*
409 * Let the RCU core know that this CPU has gone through the scheduler,
410 * which is a quiescent state. This is called when the need for a
411 * quiescent state is urgent, so we burn an atomic operation and full
412 * memory barriers to let the RCU core know about it, regardless of what
413 * this CPU might (or might not) do in the near future.
414 *
415 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
416 *
417 * The caller must have disabled interrupts and must not be idle.
418 */
419void rcu_momentary_dyntick_idle(void)
420{
421 int special;
422
423 raw_cpu_write(rcu_data.rcu_need_heavy_qs, false);
424 special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
425 &this_cpu_ptr(&rcu_data)->dynticks);
426 /* It is illegal to call this from idle state. */
427 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
428 rcu_preempt_deferred_qs(current);
429}
430EXPORT_SYMBOL_GPL(rcu_momentary_dyntick_idle);
431
432/**
433 * rcu_is_cpu_rrupt_from_idle - see if 'interrupted' from idle
434 *
435 * If the current CPU is idle and running at a first-level (not nested)
436 * interrupt, or directly, from idle, return true.
437 *
438 * The caller must have at least disabled IRQs.
439 */
440static int rcu_is_cpu_rrupt_from_idle(void)
441{
442 long nesting;
443
444 /*
445 * Usually called from the tick; but also used from smp_function_call()
446 * for expedited grace periods. This latter can result in running from
447 * the idle task, instead of an actual IPI.
448 */
449 lockdep_assert_irqs_disabled();
450
451 /* Check for counter underflows */
452 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) < 0,
453 "RCU dynticks_nesting counter underflow!");
454 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) <= 0,
455 "RCU dynticks_nmi_nesting counter underflow/zero!");
456
457 /* Are we at first interrupt nesting level? */
458 nesting = __this_cpu_read(rcu_data.dynticks_nmi_nesting);
459 if (nesting > 1)
460 return false;
461
462 /*
463 * If we're not in an interrupt, we must be in the idle task!
464 */
465 WARN_ON_ONCE(!nesting && !is_idle_task(current));
466
467 /* Does CPU appear to be idle from an RCU standpoint? */
468 return __this_cpu_read(rcu_data.dynticks_nesting) == 0;
469}
470
471#define DEFAULT_RCU_BLIMIT 10 /* Maximum callbacks per rcu_do_batch ... */
472#define DEFAULT_MAX_RCU_BLIMIT 10000 /* ... even during callback flood. */
473static long blimit = DEFAULT_RCU_BLIMIT;
474#define DEFAULT_RCU_QHIMARK 10000 /* If this many pending, ignore blimit. */
475static long qhimark = DEFAULT_RCU_QHIMARK;
476#define DEFAULT_RCU_QLOMARK 100 /* Once only this many pending, use blimit. */
477static long qlowmark = DEFAULT_RCU_QLOMARK;
478#define DEFAULT_RCU_QOVLD_MULT 2
479#define DEFAULT_RCU_QOVLD (DEFAULT_RCU_QOVLD_MULT * DEFAULT_RCU_QHIMARK)
480static long qovld = DEFAULT_RCU_QOVLD; /* If this many pending, hammer QS. */
481static long qovld_calc = -1; /* No pre-initialization lock acquisitions! */
482
483module_param(blimit, long, 0444);
484module_param(qhimark, long, 0444);
485module_param(qlowmark, long, 0444);
486module_param(qovld, long, 0444);
487
488static ulong jiffies_till_first_fqs = ULONG_MAX;
489static ulong jiffies_till_next_fqs = ULONG_MAX;
490static bool rcu_kick_kthreads;
491static int rcu_divisor = 7;
492module_param(rcu_divisor, int, 0644);
493
494/* Force an exit from rcu_do_batch() after 3 milliseconds. */
495static long rcu_resched_ns = 3 * NSEC_PER_MSEC;
496module_param(rcu_resched_ns, long, 0644);
497
498/*
499 * How long the grace period must be before we start recruiting
500 * quiescent-state help from rcu_note_context_switch().
501 */
502static ulong jiffies_till_sched_qs = ULONG_MAX;
503module_param(jiffies_till_sched_qs, ulong, 0444);
504static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */
505module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
506
507/*
508 * Make sure that we give the grace-period kthread time to detect any
509 * idle CPUs before taking active measures to force quiescent states.
510 * However, don't go below 100 milliseconds, adjusted upwards for really
511 * large systems.
512 */
513static void adjust_jiffies_till_sched_qs(void)
514{
515 unsigned long j;
516
517 /* If jiffies_till_sched_qs was specified, respect the request. */
518 if (jiffies_till_sched_qs != ULONG_MAX) {
519 WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
520 return;
521 }
522 /* Otherwise, set to third fqs scan, but bound below on large system. */
523 j = READ_ONCE(jiffies_till_first_fqs) +
524 2 * READ_ONCE(jiffies_till_next_fqs);
525 if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
526 j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
527 pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
528 WRITE_ONCE(jiffies_to_sched_qs, j);
529}
530
531static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
532{
533 ulong j;
534 int ret = kstrtoul(val, 0, &j);
535
536 if (!ret) {
537 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
538 adjust_jiffies_till_sched_qs();
539 }
540 return ret;
541}
542
543static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
544{
545 ulong j;
546 int ret = kstrtoul(val, 0, &j);
547
548 if (!ret) {
549 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
550 adjust_jiffies_till_sched_qs();
551 }
552 return ret;
553}
554
555static struct kernel_param_ops first_fqs_jiffies_ops = {
556 .set = param_set_first_fqs_jiffies,
557 .get = param_get_ulong,
558};
559
560static struct kernel_param_ops next_fqs_jiffies_ops = {
561 .set = param_set_next_fqs_jiffies,
562 .get = param_get_ulong,
563};
564
565module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
566module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
567module_param(rcu_kick_kthreads, bool, 0644);
568
569static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
570static int rcu_pending(int user);
571
572/*
573 * Return the number of RCU GPs completed thus far for debug & stats.
574 */
575unsigned long rcu_get_gp_seq(void)
576{
577 return READ_ONCE(rcu_state.gp_seq);
578}
579EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
580
581/*
582 * Return the number of RCU expedited batches completed thus far for
583 * debug & stats. Odd numbers mean that a batch is in progress, even
584 * numbers mean idle. The value returned will thus be roughly double
585 * the cumulative batches since boot.
586 */
587unsigned long rcu_exp_batches_completed(void)
588{
589 return rcu_state.expedited_sequence;
590}
591EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
592
593/*
594 * Return the root node of the rcu_state structure.
595 */
596static struct rcu_node *rcu_get_root(void)
597{
598 return &rcu_state.node[0];
599}
600
601/*
602 * Send along grace-period-related data for rcutorture diagnostics.
603 */
604void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
605 unsigned long *gp_seq)
606{
607 switch (test_type) {
608 case RCU_FLAVOR:
609 *flags = READ_ONCE(rcu_state.gp_flags);
610 *gp_seq = rcu_seq_current(&rcu_state.gp_seq);
611 break;
612 default:
613 break;
614 }
615}
616EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
617
618/*
619 * Enter an RCU extended quiescent state, which can be either the
620 * idle loop or adaptive-tickless usermode execution.
621 *
622 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
623 * the possibility of usermode upcalls having messed up our count
624 * of interrupt nesting level during the prior busy period.
625 */
626static noinstr void rcu_eqs_enter(bool user)
627{
628 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
629
630 WARN_ON_ONCE(rdp->dynticks_nmi_nesting != DYNTICK_IRQ_NONIDLE);
631 WRITE_ONCE(rdp->dynticks_nmi_nesting, 0);
632 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
633 rdp->dynticks_nesting == 0);
634 if (rdp->dynticks_nesting != 1) {
635 // RCU will still be watching, so just do accounting and leave.
636 rdp->dynticks_nesting--;
637 return;
638 }
639
640 lockdep_assert_irqs_disabled();
641 instrumentation_begin();
642 trace_rcu_dyntick(TPS("Start"), rdp->dynticks_nesting, 0, atomic_read(&rdp->dynticks));
643 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
644 rdp = this_cpu_ptr(&rcu_data);
645 do_nocb_deferred_wakeup(rdp);
646 rcu_prepare_for_idle();
647 rcu_preempt_deferred_qs(current);
648
649 // instrumentation for the noinstr rcu_dynticks_eqs_enter()
650 instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
651
652 instrumentation_end();
653 WRITE_ONCE(rdp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
654 // RCU is watching here ...
655 rcu_dynticks_eqs_enter();
656 // ... but is no longer watching here.
657 rcu_dynticks_task_enter();
658}
659
660/**
661 * rcu_idle_enter - inform RCU that current CPU is entering idle
662 *
663 * Enter idle mode, in other words, -leave- the mode in which RCU
664 * read-side critical sections can occur. (Though RCU read-side
665 * critical sections can occur in irq handlers in idle, a possibility
666 * handled by irq_enter() and irq_exit().)
667 *
668 * If you add or remove a call to rcu_idle_enter(), be sure to test with
669 * CONFIG_RCU_EQS_DEBUG=y.
670 */
671void rcu_idle_enter(void)
672{
673 lockdep_assert_irqs_disabled();
674 rcu_eqs_enter(false);
675}
676EXPORT_SYMBOL_GPL(rcu_idle_enter);
677
678#ifdef CONFIG_NO_HZ_FULL
679/**
680 * rcu_user_enter - inform RCU that we are resuming userspace.
681 *
682 * Enter RCU idle mode right before resuming userspace. No use of RCU
683 * is permitted between this call and rcu_user_exit(). This way the
684 * CPU doesn't need to maintain the tick for RCU maintenance purposes
685 * when the CPU runs in userspace.
686 *
687 * If you add or remove a call to rcu_user_enter(), be sure to test with
688 * CONFIG_RCU_EQS_DEBUG=y.
689 */
690noinstr void rcu_user_enter(void)
691{
692 lockdep_assert_irqs_disabled();
693 rcu_eqs_enter(true);
694}
695#endif /* CONFIG_NO_HZ_FULL */
696
697/**
698 * rcu_nmi_exit - inform RCU of exit from NMI context
699 *
700 * If we are returning from the outermost NMI handler that interrupted an
701 * RCU-idle period, update rdp->dynticks and rdp->dynticks_nmi_nesting
702 * to let the RCU grace-period handling know that the CPU is back to
703 * being RCU-idle.
704 *
705 * If you add or remove a call to rcu_nmi_exit(), be sure to test
706 * with CONFIG_RCU_EQS_DEBUG=y.
707 */
708noinstr void rcu_nmi_exit(void)
709{
710 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
711
712 instrumentation_begin();
713 /*
714 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
715 * (We are exiting an NMI handler, so RCU better be paying attention
716 * to us!)
717 */
718 WARN_ON_ONCE(rdp->dynticks_nmi_nesting <= 0);
719 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
720
721 /*
722 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
723 * leave it in non-RCU-idle state.
724 */
725 if (rdp->dynticks_nmi_nesting != 1) {
726 trace_rcu_dyntick(TPS("--="), rdp->dynticks_nmi_nesting, rdp->dynticks_nmi_nesting - 2,
727 atomic_read(&rdp->dynticks));
728 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* No store tearing. */
729 rdp->dynticks_nmi_nesting - 2);
730 instrumentation_end();
731 return;
732 }
733
734 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
735 trace_rcu_dyntick(TPS("Startirq"), rdp->dynticks_nmi_nesting, 0, atomic_read(&rdp->dynticks));
736 WRITE_ONCE(rdp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
737
738 if (!in_nmi())
739 rcu_prepare_for_idle();
740
741 // instrumentation for the noinstr rcu_dynticks_eqs_enter()
742 instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
743 instrumentation_end();
744
745 // RCU is watching here ...
746 rcu_dynticks_eqs_enter();
747 // ... but is no longer watching here.
748
749 if (!in_nmi())
750 rcu_dynticks_task_enter();
751}
752
753/**
754 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
755 *
756 * Exit from an interrupt handler, which might possibly result in entering
757 * idle mode, in other words, leaving the mode in which read-side critical
758 * sections can occur. The caller must have disabled interrupts.
759 *
760 * This code assumes that the idle loop never does anything that might
761 * result in unbalanced calls to irq_enter() and irq_exit(). If your
762 * architecture's idle loop violates this assumption, RCU will give you what
763 * you deserve, good and hard. But very infrequently and irreproducibly.
764 *
765 * Use things like work queues to work around this limitation.
766 *
767 * You have been warned.
768 *
769 * If you add or remove a call to rcu_irq_exit(), be sure to test with
770 * CONFIG_RCU_EQS_DEBUG=y.
771 */
772void noinstr rcu_irq_exit(void)
773{
774 lockdep_assert_irqs_disabled();
775 rcu_nmi_exit();
776}
777
778/**
779 * rcu_irq_exit_preempt - Inform RCU that current CPU is exiting irq
780 * towards in kernel preemption
781 *
782 * Same as rcu_irq_exit() but has a sanity check that scheduling is safe
783 * from RCU point of view. Invoked from return from interrupt before kernel
784 * preemption.
785 */
786void rcu_irq_exit_preempt(void)
787{
788 lockdep_assert_irqs_disabled();
789 rcu_nmi_exit();
790
791 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) <= 0,
792 "RCU dynticks_nesting counter underflow/zero!");
793 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) !=
794 DYNTICK_IRQ_NONIDLE,
795 "Bad RCU dynticks_nmi_nesting counter\n");
796 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
797 "RCU in extended quiescent state!");
798}
799
800#ifdef CONFIG_PROVE_RCU
801/**
802 * rcu_irq_exit_check_preempt - Validate that scheduling is possible
803 */
804void rcu_irq_exit_check_preempt(void)
805{
806 lockdep_assert_irqs_disabled();
807
808 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) <= 0,
809 "RCU dynticks_nesting counter underflow/zero!");
810 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) !=
811 DYNTICK_IRQ_NONIDLE,
812 "Bad RCU dynticks_nmi_nesting counter\n");
813 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
814 "RCU in extended quiescent state!");
815}
816#endif /* #ifdef CONFIG_PROVE_RCU */
817
818/*
819 * Wrapper for rcu_irq_exit() where interrupts are enabled.
820 *
821 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
822 * with CONFIG_RCU_EQS_DEBUG=y.
823 */
824void rcu_irq_exit_irqson(void)
825{
826 unsigned long flags;
827
828 local_irq_save(flags);
829 rcu_irq_exit();
830 local_irq_restore(flags);
831}
832
833/*
834 * Exit an RCU extended quiescent state, which can be either the
835 * idle loop or adaptive-tickless usermode execution.
836 *
837 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
838 * allow for the possibility of usermode upcalls messing up our count of
839 * interrupt nesting level during the busy period that is just now starting.
840 */
841static void noinstr rcu_eqs_exit(bool user)
842{
843 struct rcu_data *rdp;
844 long oldval;
845
846 lockdep_assert_irqs_disabled();
847 rdp = this_cpu_ptr(&rcu_data);
848 oldval = rdp->dynticks_nesting;
849 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
850 if (oldval) {
851 // RCU was already watching, so just do accounting and leave.
852 rdp->dynticks_nesting++;
853 return;
854 }
855 rcu_dynticks_task_exit();
856 // RCU is not watching here ...
857 rcu_dynticks_eqs_exit();
858 // ... but is watching here.
859 instrumentation_begin();
860
861 // instrumentation for the noinstr rcu_dynticks_eqs_exit()
862 instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
863
864 rcu_cleanup_after_idle();
865 trace_rcu_dyntick(TPS("End"), rdp->dynticks_nesting, 1, atomic_read(&rdp->dynticks));
866 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
867 WRITE_ONCE(rdp->dynticks_nesting, 1);
868 WARN_ON_ONCE(rdp->dynticks_nmi_nesting);
869 WRITE_ONCE(rdp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
870 instrumentation_end();
871}
872
873/**
874 * rcu_idle_exit - inform RCU that current CPU is leaving idle
875 *
876 * Exit idle mode, in other words, -enter- the mode in which RCU
877 * read-side critical sections can occur.
878 *
879 * If you add or remove a call to rcu_idle_exit(), be sure to test with
880 * CONFIG_RCU_EQS_DEBUG=y.
881 */
882void rcu_idle_exit(void)
883{
884 unsigned long flags;
885
886 local_irq_save(flags);
887 rcu_eqs_exit(false);
888 local_irq_restore(flags);
889}
890EXPORT_SYMBOL_GPL(rcu_idle_exit);
891
892#ifdef CONFIG_NO_HZ_FULL
893/**
894 * rcu_user_exit - inform RCU that we are exiting userspace.
895 *
896 * Exit RCU idle mode while entering the kernel because it can
897 * run a RCU read side critical section anytime.
898 *
899 * If you add or remove a call to rcu_user_exit(), be sure to test with
900 * CONFIG_RCU_EQS_DEBUG=y.
901 */
902void noinstr rcu_user_exit(void)
903{
904 rcu_eqs_exit(1);
905}
906
907/**
908 * __rcu_irq_enter_check_tick - Enable scheduler tick on CPU if RCU needs it.
909 *
910 * The scheduler tick is not normally enabled when CPUs enter the kernel
911 * from nohz_full userspace execution. After all, nohz_full userspace
912 * execution is an RCU quiescent state and the time executing in the kernel
913 * is quite short. Except of course when it isn't. And it is not hard to
914 * cause a large system to spend tens of seconds or even minutes looping
915 * in the kernel, which can cause a number of problems, include RCU CPU
916 * stall warnings.
917 *
918 * Therefore, if a nohz_full CPU fails to report a quiescent state
919 * in a timely manner, the RCU grace-period kthread sets that CPU's
920 * ->rcu_urgent_qs flag with the expectation that the next interrupt or
921 * exception will invoke this function, which will turn on the scheduler
922 * tick, which will enable RCU to detect that CPU's quiescent states,
923 * for example, due to cond_resched() calls in CONFIG_PREEMPT=n kernels.
924 * The tick will be disabled once a quiescent state is reported for
925 * this CPU.
926 *
927 * Of course, in carefully tuned systems, there might never be an
928 * interrupt or exception. In that case, the RCU grace-period kthread
929 * will eventually cause one to happen. However, in less carefully
930 * controlled environments, this function allows RCU to get what it
931 * needs without creating otherwise useless interruptions.
932 */
933void __rcu_irq_enter_check_tick(void)
934{
935 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
936
937 // Enabling the tick is unsafe in NMI handlers.
938 if (WARN_ON_ONCE(in_nmi()))
939 return;
940
941 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
942 "Illegal rcu_irq_enter_check_tick() from extended quiescent state");
943
944 if (!tick_nohz_full_cpu(rdp->cpu) ||
945 !READ_ONCE(rdp->rcu_urgent_qs) ||
946 READ_ONCE(rdp->rcu_forced_tick)) {
947 // RCU doesn't need nohz_full help from this CPU, or it is
948 // already getting that help.
949 return;
950 }
951
952 // We get here only when not in an extended quiescent state and
953 // from interrupts (as opposed to NMIs). Therefore, (1) RCU is
954 // already watching and (2) The fact that we are in an interrupt
955 // handler and that the rcu_node lock is an irq-disabled lock
956 // prevents self-deadlock. So we can safely recheck under the lock.
957 // Note that the nohz_full state currently cannot change.
958 raw_spin_lock_rcu_node(rdp->mynode);
959 if (rdp->rcu_urgent_qs && !rdp->rcu_forced_tick) {
960 // A nohz_full CPU is in the kernel and RCU needs a
961 // quiescent state. Turn on the tick!
962 WRITE_ONCE(rdp->rcu_forced_tick, true);
963 tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
964 }
965 raw_spin_unlock_rcu_node(rdp->mynode);
966}
967#endif /* CONFIG_NO_HZ_FULL */
968
969/**
970 * rcu_nmi_enter - inform RCU of entry to NMI context
971 *
972 * If the CPU was idle from RCU's viewpoint, update rdp->dynticks and
973 * rdp->dynticks_nmi_nesting to let the RCU grace-period handling know
974 * that the CPU is active. This implementation permits nested NMIs, as
975 * long as the nesting level does not overflow an int. (You will probably
976 * run out of stack space first.)
977 *
978 * If you add or remove a call to rcu_nmi_enter(), be sure to test
979 * with CONFIG_RCU_EQS_DEBUG=y.
980 */
981noinstr void rcu_nmi_enter(void)
982{
983 long incby = 2;
984 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
985
986 /* Complain about underflow. */
987 WARN_ON_ONCE(rdp->dynticks_nmi_nesting < 0);
988
989 /*
990 * If idle from RCU viewpoint, atomically increment ->dynticks
991 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
992 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
993 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
994 * to be in the outermost NMI handler that interrupted an RCU-idle
995 * period (observation due to Andy Lutomirski).
996 */
997 if (rcu_dynticks_curr_cpu_in_eqs()) {
998
999 if (!in_nmi())
1000 rcu_dynticks_task_exit();
1001
1002 // RCU is not watching here ...
1003 rcu_dynticks_eqs_exit();
1004 // ... but is watching here.
1005
1006 if (!in_nmi()) {
1007 instrumentation_begin();
1008 rcu_cleanup_after_idle();
1009 instrumentation_end();
1010 }
1011
1012 instrumentation_begin();
1013 // instrumentation for the noinstr rcu_dynticks_curr_cpu_in_eqs()
1014 instrument_atomic_read(&rdp->dynticks, sizeof(rdp->dynticks));
1015 // instrumentation for the noinstr rcu_dynticks_eqs_exit()
1016 instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
1017
1018 incby = 1;
1019 } else if (!in_nmi()) {
1020 instrumentation_begin();
1021 rcu_irq_enter_check_tick();
1022 instrumentation_end();
1023 } else {
1024 instrumentation_begin();
1025 }
1026
1027 trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
1028 rdp->dynticks_nmi_nesting,
1029 rdp->dynticks_nmi_nesting + incby, atomic_read(&rdp->dynticks));
1030 instrumentation_end();
1031 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* Prevent store tearing. */
1032 rdp->dynticks_nmi_nesting + incby);
1033 barrier();
1034}
1035
1036/**
1037 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
1038 *
1039 * Enter an interrupt handler, which might possibly result in exiting
1040 * idle mode, in other words, entering the mode in which read-side critical
1041 * sections can occur. The caller must have disabled interrupts.
1042 *
1043 * Note that the Linux kernel is fully capable of entering an interrupt
1044 * handler that it never exits, for example when doing upcalls to user mode!
1045 * This code assumes that the idle loop never does upcalls to user mode.
1046 * If your architecture's idle loop does do upcalls to user mode (or does
1047 * anything else that results in unbalanced calls to the irq_enter() and
1048 * irq_exit() functions), RCU will give you what you deserve, good and hard.
1049 * But very infrequently and irreproducibly.
1050 *
1051 * Use things like work queues to work around this limitation.
1052 *
1053 * You have been warned.
1054 *
1055 * If you add or remove a call to rcu_irq_enter(), be sure to test with
1056 * CONFIG_RCU_EQS_DEBUG=y.
1057 */
1058noinstr void rcu_irq_enter(void)
1059{
1060 lockdep_assert_irqs_disabled();
1061 rcu_nmi_enter();
1062}
1063
1064/*
1065 * Wrapper for rcu_irq_enter() where interrupts are enabled.
1066 *
1067 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
1068 * with CONFIG_RCU_EQS_DEBUG=y.
1069 */
1070void rcu_irq_enter_irqson(void)
1071{
1072 unsigned long flags;
1073
1074 local_irq_save(flags);
1075 rcu_irq_enter();
1076 local_irq_restore(flags);
1077}
1078
1079/*
1080 * If any sort of urgency was applied to the current CPU (for example,
1081 * the scheduler-clock interrupt was enabled on a nohz_full CPU) in order
1082 * to get to a quiescent state, disable it.
1083 */
1084static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp)
1085{
1086 raw_lockdep_assert_held_rcu_node(rdp->mynode);
1087 WRITE_ONCE(rdp->rcu_urgent_qs, false);
1088 WRITE_ONCE(rdp->rcu_need_heavy_qs, false);
1089 if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) {
1090 tick_dep_clear_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
1091 WRITE_ONCE(rdp->rcu_forced_tick, false);
1092 }
1093}
1094
1095noinstr bool __rcu_is_watching(void)
1096{
1097 return !rcu_dynticks_curr_cpu_in_eqs();
1098}
1099
1100/**
1101 * rcu_is_watching - see if RCU thinks that the current CPU is not idle
1102 *
1103 * Return true if RCU is watching the running CPU, which means that this
1104 * CPU can safely enter RCU read-side critical sections. In other words,
1105 * if the current CPU is not in its idle loop or is in an interrupt or
1106 * NMI handler, return true.
1107 */
1108bool rcu_is_watching(void)
1109{
1110 bool ret;
1111
1112 preempt_disable_notrace();
1113 ret = !rcu_dynticks_curr_cpu_in_eqs();
1114 preempt_enable_notrace();
1115 return ret;
1116}
1117EXPORT_SYMBOL_GPL(rcu_is_watching);
1118
1119/*
1120 * If a holdout task is actually running, request an urgent quiescent
1121 * state from its CPU. This is unsynchronized, so migrations can cause
1122 * the request to go to the wrong CPU. Which is OK, all that will happen
1123 * is that the CPU's next context switch will be a bit slower and next
1124 * time around this task will generate another request.
1125 */
1126void rcu_request_urgent_qs_task(struct task_struct *t)
1127{
1128 int cpu;
1129
1130 barrier();
1131 cpu = task_cpu(t);
1132 if (!task_curr(t))
1133 return; /* This task is not running on that CPU. */
1134 smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
1135}
1136
1137#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1138
1139/*
1140 * Is the current CPU online as far as RCU is concerned?
1141 *
1142 * Disable preemption to avoid false positives that could otherwise
1143 * happen due to the current CPU number being sampled, this task being
1144 * preempted, its old CPU being taken offline, resuming on some other CPU,
1145 * then determining that its old CPU is now offline.
1146 *
1147 * Disable checking if in an NMI handler because we cannot safely
1148 * report errors from NMI handlers anyway. In addition, it is OK to use
1149 * RCU on an offline processor during initial boot, hence the check for
1150 * rcu_scheduler_fully_active.
1151 */
1152bool rcu_lockdep_current_cpu_online(void)
1153{
1154 struct rcu_data *rdp;
1155 struct rcu_node *rnp;
1156 bool ret = false;
1157
1158 if (in_nmi() || !rcu_scheduler_fully_active)
1159 return true;
1160 preempt_disable_notrace();
1161 rdp = this_cpu_ptr(&rcu_data);
1162 rnp = rdp->mynode;
1163 if (rdp->grpmask & rcu_rnp_online_cpus(rnp))
1164 ret = true;
1165 preempt_enable_notrace();
1166 return ret;
1167}
1168EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1169
1170#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1171
1172/*
1173 * We are reporting a quiescent state on behalf of some other CPU, so
1174 * it is our responsibility to check for and handle potential overflow
1175 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
1176 * After all, the CPU might be in deep idle state, and thus executing no
1177 * code whatsoever.
1178 */
1179static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
1180{
1181 raw_lockdep_assert_held_rcu_node(rnp);
1182 if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
1183 rnp->gp_seq))
1184 WRITE_ONCE(rdp->gpwrap, true);
1185 if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
1186 rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
1187}
1188
1189/*
1190 * Snapshot the specified CPU's dynticks counter so that we can later
1191 * credit them with an implicit quiescent state. Return 1 if this CPU
1192 * is in dynticks idle mode, which is an extended quiescent state.
1193 */
1194static int dyntick_save_progress_counter(struct rcu_data *rdp)
1195{
1196 rdp->dynticks_snap = rcu_dynticks_snap(rdp);
1197 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1198 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1199 rcu_gpnum_ovf(rdp->mynode, rdp);
1200 return 1;
1201 }
1202 return 0;
1203}
1204
1205/*
1206 * Return true if the specified CPU has passed through a quiescent
1207 * state by virtue of being in or having passed through an dynticks
1208 * idle state since the last call to dyntick_save_progress_counter()
1209 * for this same CPU, or by virtue of having been offline.
1210 */
1211static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1212{
1213 unsigned long jtsq;
1214 bool *rnhqp;
1215 bool *ruqp;
1216 struct rcu_node *rnp = rdp->mynode;
1217
1218 /*
1219 * If the CPU passed through or entered a dynticks idle phase with
1220 * no active irq/NMI handlers, then we can safely pretend that the CPU
1221 * already acknowledged the request to pass through a quiescent
1222 * state. Either way, that CPU cannot possibly be in an RCU
1223 * read-side critical section that started before the beginning
1224 * of the current RCU grace period.
1225 */
1226 if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
1227 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1228 rcu_gpnum_ovf(rnp, rdp);
1229 return 1;
1230 }
1231
1232 /* If waiting too long on an offline CPU, complain. */
1233 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp)) &&
1234 time_after(jiffies, rcu_state.gp_start + HZ)) {
1235 bool onl;
1236 struct rcu_node *rnp1;
1237
1238 WARN_ON(1); /* Offline CPUs are supposed to report QS! */
1239 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
1240 __func__, rnp->grplo, rnp->grphi, rnp->level,
1241 (long)rnp->gp_seq, (long)rnp->completedqs);
1242 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
1243 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
1244 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
1245 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
1246 pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
1247 __func__, rdp->cpu, ".o"[onl],
1248 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
1249 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
1250 return 1; /* Break things loose after complaining. */
1251 }
1252
1253 /*
1254 * A CPU running for an extended time within the kernel can
1255 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
1256 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
1257 * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the
1258 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
1259 * variable are safe because the assignments are repeated if this
1260 * CPU failed to pass through a quiescent state. This code
1261 * also checks .jiffies_resched in case jiffies_to_sched_qs
1262 * is set way high.
1263 */
1264 jtsq = READ_ONCE(jiffies_to_sched_qs);
1265 ruqp = per_cpu_ptr(&rcu_data.rcu_urgent_qs, rdp->cpu);
1266 rnhqp = &per_cpu(rcu_data.rcu_need_heavy_qs, rdp->cpu);
1267 if (!READ_ONCE(*rnhqp) &&
1268 (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
1269 time_after(jiffies, rcu_state.jiffies_resched) ||
1270 rcu_state.cbovld)) {
1271 WRITE_ONCE(*rnhqp, true);
1272 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1273 smp_store_release(ruqp, true);
1274 } else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
1275 WRITE_ONCE(*ruqp, true);
1276 }
1277
1278 /*
1279 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
1280 * The above code handles this, but only for straight cond_resched().
1281 * And some in-kernel loops check need_resched() before calling
1282 * cond_resched(), which defeats the above code for CPUs that are
1283 * running in-kernel with scheduling-clock interrupts disabled.
1284 * So hit them over the head with the resched_cpu() hammer!
1285 */
1286 if (tick_nohz_full_cpu(rdp->cpu) &&
1287 (time_after(jiffies, READ_ONCE(rdp->last_fqs_resched) + jtsq * 3) ||
1288 rcu_state.cbovld)) {
1289 WRITE_ONCE(*ruqp, true);
1290 resched_cpu(rdp->cpu);
1291 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1292 }
1293
1294 /*
1295 * If more than halfway to RCU CPU stall-warning time, invoke
1296 * resched_cpu() more frequently to try to loosen things up a bit.
1297 * Also check to see if the CPU is getting hammered with interrupts,
1298 * but only once per grace period, just to keep the IPIs down to
1299 * a dull roar.
1300 */
1301 if (time_after(jiffies, rcu_state.jiffies_resched)) {
1302 if (time_after(jiffies,
1303 READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
1304 resched_cpu(rdp->cpu);
1305 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1306 }
1307 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1308 !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
1309 (rnp->ffmask & rdp->grpmask)) {
1310 init_irq_work(&rdp->rcu_iw, rcu_iw_handler);
1311 atomic_set(&rdp->rcu_iw.flags, IRQ_WORK_HARD_IRQ);
1312 rdp->rcu_iw_pending = true;
1313 rdp->rcu_iw_gp_seq = rnp->gp_seq;
1314 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1315 }
1316 }
1317
1318 return 0;
1319}
1320
1321/* Trace-event wrapper function for trace_rcu_future_grace_period. */
1322static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1323 unsigned long gp_seq_req, const char *s)
1324{
1325 trace_rcu_future_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
1326 gp_seq_req, rnp->level,
1327 rnp->grplo, rnp->grphi, s);
1328}
1329
1330/*
1331 * rcu_start_this_gp - Request the start of a particular grace period
1332 * @rnp_start: The leaf node of the CPU from which to start.
1333 * @rdp: The rcu_data corresponding to the CPU from which to start.
1334 * @gp_seq_req: The gp_seq of the grace period to start.
1335 *
1336 * Start the specified grace period, as needed to handle newly arrived
1337 * callbacks. The required future grace periods are recorded in each
1338 * rcu_node structure's ->gp_seq_needed field. Returns true if there
1339 * is reason to awaken the grace-period kthread.
1340 *
1341 * The caller must hold the specified rcu_node structure's ->lock, which
1342 * is why the caller is responsible for waking the grace-period kthread.
1343 *
1344 * Returns true if the GP thread needs to be awakened else false.
1345 */
1346static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
1347 unsigned long gp_seq_req)
1348{
1349 bool ret = false;
1350 struct rcu_node *rnp;
1351
1352 /*
1353 * Use funnel locking to either acquire the root rcu_node
1354 * structure's lock or bail out if the need for this grace period
1355 * has already been recorded -- or if that grace period has in
1356 * fact already started. If there is already a grace period in
1357 * progress in a non-leaf node, no recording is needed because the
1358 * end of the grace period will scan the leaf rcu_node structures.
1359 * Note that rnp_start->lock must not be released.
1360 */
1361 raw_lockdep_assert_held_rcu_node(rnp_start);
1362 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
1363 for (rnp = rnp_start; 1; rnp = rnp->parent) {
1364 if (rnp != rnp_start)
1365 raw_spin_lock_rcu_node(rnp);
1366 if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
1367 rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
1368 (rnp != rnp_start &&
1369 rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
1370 trace_rcu_this_gp(rnp, rdp, gp_seq_req,
1371 TPS("Prestarted"));
1372 goto unlock_out;
1373 }
1374 WRITE_ONCE(rnp->gp_seq_needed, gp_seq_req);
1375 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
1376 /*
1377 * We just marked the leaf or internal node, and a
1378 * grace period is in progress, which means that
1379 * rcu_gp_cleanup() will see the marking. Bail to
1380 * reduce contention.
1381 */
1382 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
1383 TPS("Startedleaf"));
1384 goto unlock_out;
1385 }
1386 if (rnp != rnp_start && rnp->parent != NULL)
1387 raw_spin_unlock_rcu_node(rnp);
1388 if (!rnp->parent)
1389 break; /* At root, and perhaps also leaf. */
1390 }
1391
1392 /* If GP already in progress, just leave, otherwise start one. */
1393 if (rcu_gp_in_progress()) {
1394 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
1395 goto unlock_out;
1396 }
1397 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
1398 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
1399 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
1400 if (!READ_ONCE(rcu_state.gp_kthread)) {
1401 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
1402 goto unlock_out;
1403 }
1404 trace_rcu_grace_period(rcu_state.name, data_race(rcu_state.gp_seq), TPS("newreq"));
1405 ret = true; /* Caller must wake GP kthread. */
1406unlock_out:
1407 /* Push furthest requested GP to leaf node and rcu_data structure. */
1408 if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
1409 WRITE_ONCE(rnp_start->gp_seq_needed, rnp->gp_seq_needed);
1410 WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1411 }
1412 if (rnp != rnp_start)
1413 raw_spin_unlock_rcu_node(rnp);
1414 return ret;
1415}
1416
1417/*
1418 * Clean up any old requests for the just-ended grace period. Also return
1419 * whether any additional grace periods have been requested.
1420 */
1421static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
1422{
1423 bool needmore;
1424 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1425
1426 needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1427 if (!needmore)
1428 rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1429 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1430 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1431 return needmore;
1432}
1433
1434/*
1435 * Awaken the grace-period kthread. Don't do a self-awaken (unless in an
1436 * interrupt or softirq handler, in which case we just might immediately
1437 * sleep upon return, resulting in a grace-period hang), and don't bother
1438 * awakening when there is nothing for the grace-period kthread to do
1439 * (as in several CPUs raced to awaken, we lost), and finally don't try
1440 * to awaken a kthread that has not yet been created. If all those checks
1441 * are passed, track some debug information and awaken.
1442 *
1443 * So why do the self-wakeup when in an interrupt or softirq handler
1444 * in the grace-period kthread's context? Because the kthread might have
1445 * been interrupted just as it was going to sleep, and just after the final
1446 * pre-sleep check of the awaken condition. In this case, a wakeup really
1447 * is required, and is therefore supplied.
1448 */
1449static void rcu_gp_kthread_wake(void)
1450{
1451 struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
1452
1453 if ((current == t && !in_irq() && !in_serving_softirq()) ||
1454 !READ_ONCE(rcu_state.gp_flags) || !t)
1455 return;
1456 WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
1457 WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
1458 swake_up_one(&rcu_state.gp_wq);
1459}
1460
1461/*
1462 * If there is room, assign a ->gp_seq number to any callbacks on this
1463 * CPU that have not already been assigned. Also accelerate any callbacks
1464 * that were previously assigned a ->gp_seq number that has since proven
1465 * to be too conservative, which can happen if callbacks get assigned a
1466 * ->gp_seq number while RCU is idle, but with reference to a non-root
1467 * rcu_node structure. This function is idempotent, so it does not hurt
1468 * to call it repeatedly. Returns an flag saying that we should awaken
1469 * the RCU grace-period kthread.
1470 *
1471 * The caller must hold rnp->lock with interrupts disabled.
1472 */
1473static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1474{
1475 unsigned long gp_seq_req;
1476 bool ret = false;
1477
1478 rcu_lockdep_assert_cblist_protected(rdp);
1479 raw_lockdep_assert_held_rcu_node(rnp);
1480
1481 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1482 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1483 return false;
1484
1485 /*
1486 * Callbacks are often registered with incomplete grace-period
1487 * information. Something about the fact that getting exact
1488 * information requires acquiring a global lock... RCU therefore
1489 * makes a conservative estimate of the grace period number at which
1490 * a given callback will become ready to invoke. The following
1491 * code checks this estimate and improves it when possible, thus
1492 * accelerating callback invocation to an earlier grace-period
1493 * number.
1494 */
1495 gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1496 if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1497 ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1498
1499 /* Trace depending on how much we were able to accelerate. */
1500 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1501 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccWaitCB"));
1502 else
1503 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccReadyCB"));
1504 return ret;
1505}
1506
1507/*
1508 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1509 * rcu_node structure's ->lock be held. It consults the cached value
1510 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1511 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1512 * while holding the leaf rcu_node structure's ->lock.
1513 */
1514static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1515 struct rcu_data *rdp)
1516{
1517 unsigned long c;
1518 bool needwake;
1519
1520 rcu_lockdep_assert_cblist_protected(rdp);
1521 c = rcu_seq_snap(&rcu_state.gp_seq);
1522 if (!READ_ONCE(rdp->gpwrap) && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1523 /* Old request still live, so mark recent callbacks. */
1524 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1525 return;
1526 }
1527 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1528 needwake = rcu_accelerate_cbs(rnp, rdp);
1529 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1530 if (needwake)
1531 rcu_gp_kthread_wake();
1532}
1533
1534/*
1535 * Move any callbacks whose grace period has completed to the
1536 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1537 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1538 * sublist. This function is idempotent, so it does not hurt to
1539 * invoke it repeatedly. As long as it is not invoked -too- often...
1540 * Returns true if the RCU grace-period kthread needs to be awakened.
1541 *
1542 * The caller must hold rnp->lock with interrupts disabled.
1543 */
1544static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1545{
1546 rcu_lockdep_assert_cblist_protected(rdp);
1547 raw_lockdep_assert_held_rcu_node(rnp);
1548
1549 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1550 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1551 return false;
1552
1553 /*
1554 * Find all callbacks whose ->gp_seq numbers indicate that they
1555 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1556 */
1557 rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1558
1559 /* Classify any remaining callbacks. */
1560 return rcu_accelerate_cbs(rnp, rdp);
1561}
1562
1563/*
1564 * Move and classify callbacks, but only if doing so won't require
1565 * that the RCU grace-period kthread be awakened.
1566 */
1567static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
1568 struct rcu_data *rdp)
1569{
1570 rcu_lockdep_assert_cblist_protected(rdp);
1571 if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) ||
1572 !raw_spin_trylock_rcu_node(rnp))
1573 return;
1574 WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
1575 raw_spin_unlock_rcu_node(rnp);
1576}
1577
1578/*
1579 * Update CPU-local rcu_data state to record the beginnings and ends of
1580 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1581 * structure corresponding to the current CPU, and must have irqs disabled.
1582 * Returns true if the grace-period kthread needs to be awakened.
1583 */
1584static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
1585{
1586 bool ret = false;
1587 bool need_qs;
1588 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
1589 rcu_segcblist_is_offloaded(&rdp->cblist);
1590
1591 raw_lockdep_assert_held_rcu_node(rnp);
1592
1593 if (rdp->gp_seq == rnp->gp_seq)
1594 return false; /* Nothing to do. */
1595
1596 /* Handle the ends of any preceding grace periods first. */
1597 if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1598 unlikely(READ_ONCE(rdp->gpwrap))) {
1599 if (!offloaded)
1600 ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
1601 rdp->core_needs_qs = false;
1602 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1603 } else {
1604 if (!offloaded)
1605 ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
1606 if (rdp->core_needs_qs)
1607 rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask);
1608 }
1609
1610 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1611 if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1612 unlikely(READ_ONCE(rdp->gpwrap))) {
1613 /*
1614 * If the current grace period is waiting for this CPU,
1615 * set up to detect a quiescent state, otherwise don't
1616 * go looking for one.
1617 */
1618 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1619 need_qs = !!(rnp->qsmask & rdp->grpmask);
1620 rdp->cpu_no_qs.b.norm = need_qs;
1621 rdp->core_needs_qs = need_qs;
1622 zero_cpu_stall_ticks(rdp);
1623 }
1624 rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
1625 if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
1626 WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1627 WRITE_ONCE(rdp->gpwrap, false);
1628 rcu_gpnum_ovf(rnp, rdp);
1629 return ret;
1630}
1631
1632static void note_gp_changes(struct rcu_data *rdp)
1633{
1634 unsigned long flags;
1635 bool needwake;
1636 struct rcu_node *rnp;
1637
1638 local_irq_save(flags);
1639 rnp = rdp->mynode;
1640 if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1641 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1642 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1643 local_irq_restore(flags);
1644 return;
1645 }
1646 needwake = __note_gp_changes(rnp, rdp);
1647 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1648 if (needwake)
1649 rcu_gp_kthread_wake();
1650}
1651
1652static void rcu_gp_slow(int delay)
1653{
1654 if (delay > 0 &&
1655 !(rcu_seq_ctr(rcu_state.gp_seq) %
1656 (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1657 schedule_timeout_idle(delay);
1658}
1659
1660static unsigned long sleep_duration;
1661
1662/* Allow rcutorture to stall the grace-period kthread. */
1663void rcu_gp_set_torture_wait(int duration)
1664{
1665 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST) && duration > 0)
1666 WRITE_ONCE(sleep_duration, duration);
1667}
1668EXPORT_SYMBOL_GPL(rcu_gp_set_torture_wait);
1669
1670/* Actually implement the aforementioned wait. */
1671static void rcu_gp_torture_wait(void)
1672{
1673 unsigned long duration;
1674
1675 if (!IS_ENABLED(CONFIG_RCU_TORTURE_TEST))
1676 return;
1677 duration = xchg(&sleep_duration, 0UL);
1678 if (duration > 0) {
1679 pr_alert("%s: Waiting %lu jiffies\n", __func__, duration);
1680 schedule_timeout_idle(duration);
1681 pr_alert("%s: Wait complete\n", __func__);
1682 }
1683}
1684
1685/*
1686 * Initialize a new grace period. Return false if no grace period required.
1687 */
1688static bool rcu_gp_init(void)
1689{
1690 unsigned long flags;
1691 unsigned long oldmask;
1692 unsigned long mask;
1693 struct rcu_data *rdp;
1694 struct rcu_node *rnp = rcu_get_root();
1695
1696 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1697 raw_spin_lock_irq_rcu_node(rnp);
1698 if (!READ_ONCE(rcu_state.gp_flags)) {
1699 /* Spurious wakeup, tell caller to go back to sleep. */
1700 raw_spin_unlock_irq_rcu_node(rnp);
1701 return false;
1702 }
1703 WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1704
1705 if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1706 /*
1707 * Grace period already in progress, don't start another.
1708 * Not supposed to be able to happen.
1709 */
1710 raw_spin_unlock_irq_rcu_node(rnp);
1711 return false;
1712 }
1713
1714 /* Advance to a new grace period and initialize state. */
1715 record_gp_stall_check_time();
1716 /* Record GP times before starting GP, hence rcu_seq_start(). */
1717 rcu_seq_start(&rcu_state.gp_seq);
1718 ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
1719 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
1720 raw_spin_unlock_irq_rcu_node(rnp);
1721
1722 /*
1723 * Apply per-leaf buffered online and offline operations to the
1724 * rcu_node tree. Note that this new grace period need not wait
1725 * for subsequent online CPUs, and that quiescent-state forcing
1726 * will handle subsequent offline CPUs.
1727 */
1728 rcu_state.gp_state = RCU_GP_ONOFF;
1729 rcu_for_each_leaf_node(rnp) {
1730 raw_spin_lock(&rcu_state.ofl_lock);
1731 raw_spin_lock_irq_rcu_node(rnp);
1732 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1733 !rnp->wait_blkd_tasks) {
1734 /* Nothing to do on this leaf rcu_node structure. */
1735 raw_spin_unlock_irq_rcu_node(rnp);
1736 raw_spin_unlock(&rcu_state.ofl_lock);
1737 continue;
1738 }
1739
1740 /* Record old state, apply changes to ->qsmaskinit field. */
1741 oldmask = rnp->qsmaskinit;
1742 rnp->qsmaskinit = rnp->qsmaskinitnext;
1743
1744 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1745 if (!oldmask != !rnp->qsmaskinit) {
1746 if (!oldmask) { /* First online CPU for rcu_node. */
1747 if (!rnp->wait_blkd_tasks) /* Ever offline? */
1748 rcu_init_new_rnp(rnp);
1749 } else if (rcu_preempt_has_tasks(rnp)) {
1750 rnp->wait_blkd_tasks = true; /* blocked tasks */
1751 } else { /* Last offline CPU and can propagate. */
1752 rcu_cleanup_dead_rnp(rnp);
1753 }
1754 }
1755
1756 /*
1757 * If all waited-on tasks from prior grace period are
1758 * done, and if all this rcu_node structure's CPUs are
1759 * still offline, propagate up the rcu_node tree and
1760 * clear ->wait_blkd_tasks. Otherwise, if one of this
1761 * rcu_node structure's CPUs has since come back online,
1762 * simply clear ->wait_blkd_tasks.
1763 */
1764 if (rnp->wait_blkd_tasks &&
1765 (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1766 rnp->wait_blkd_tasks = false;
1767 if (!rnp->qsmaskinit)
1768 rcu_cleanup_dead_rnp(rnp);
1769 }
1770
1771 raw_spin_unlock_irq_rcu_node(rnp);
1772 raw_spin_unlock(&rcu_state.ofl_lock);
1773 }
1774 rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1775
1776 /*
1777 * Set the quiescent-state-needed bits in all the rcu_node
1778 * structures for all currently online CPUs in breadth-first
1779 * order, starting from the root rcu_node structure, relying on the
1780 * layout of the tree within the rcu_state.node[] array. Note that
1781 * other CPUs will access only the leaves of the hierarchy, thus
1782 * seeing that no grace period is in progress, at least until the
1783 * corresponding leaf node has been initialized.
1784 *
1785 * The grace period cannot complete until the initialization
1786 * process finishes, because this kthread handles both.
1787 */
1788 rcu_state.gp_state = RCU_GP_INIT;
1789 rcu_for_each_node_breadth_first(rnp) {
1790 rcu_gp_slow(gp_init_delay);
1791 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1792 rdp = this_cpu_ptr(&rcu_data);
1793 rcu_preempt_check_blocked_tasks(rnp);
1794 rnp->qsmask = rnp->qsmaskinit;
1795 WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
1796 if (rnp == rdp->mynode)
1797 (void)__note_gp_changes(rnp, rdp);
1798 rcu_preempt_boost_start_gp(rnp);
1799 trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1800 rnp->level, rnp->grplo,
1801 rnp->grphi, rnp->qsmask);
1802 /* Quiescent states for tasks on any now-offline CPUs. */
1803 mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1804 rnp->rcu_gp_init_mask = mask;
1805 if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1806 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1807 else
1808 raw_spin_unlock_irq_rcu_node(rnp);
1809 cond_resched_tasks_rcu_qs();
1810 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1811 }
1812
1813 return true;
1814}
1815
1816/*
1817 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1818 * time.
1819 */
1820static bool rcu_gp_fqs_check_wake(int *gfp)
1821{
1822 struct rcu_node *rnp = rcu_get_root();
1823
1824 // If under overload conditions, force an immediate FQS scan.
1825 if (*gfp & RCU_GP_FLAG_OVLD)
1826 return true;
1827
1828 // Someone like call_rcu() requested a force-quiescent-state scan.
1829 *gfp = READ_ONCE(rcu_state.gp_flags);
1830 if (*gfp & RCU_GP_FLAG_FQS)
1831 return true;
1832
1833 // The current grace period has completed.
1834 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1835 return true;
1836
1837 return false;
1838}
1839
1840/*
1841 * Do one round of quiescent-state forcing.
1842 */
1843static void rcu_gp_fqs(bool first_time)
1844{
1845 struct rcu_node *rnp = rcu_get_root();
1846
1847 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1848 rcu_state.n_force_qs++;
1849 if (first_time) {
1850 /* Collect dyntick-idle snapshots. */
1851 force_qs_rnp(dyntick_save_progress_counter);
1852 } else {
1853 /* Handle dyntick-idle and offline CPUs. */
1854 force_qs_rnp(rcu_implicit_dynticks_qs);
1855 }
1856 /* Clear flag to prevent immediate re-entry. */
1857 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
1858 raw_spin_lock_irq_rcu_node(rnp);
1859 WRITE_ONCE(rcu_state.gp_flags,
1860 READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
1861 raw_spin_unlock_irq_rcu_node(rnp);
1862 }
1863}
1864
1865/*
1866 * Loop doing repeated quiescent-state forcing until the grace period ends.
1867 */
1868static void rcu_gp_fqs_loop(void)
1869{
1870 bool first_gp_fqs;
1871 int gf = 0;
1872 unsigned long j;
1873 int ret;
1874 struct rcu_node *rnp = rcu_get_root();
1875
1876 first_gp_fqs = true;
1877 j = READ_ONCE(jiffies_till_first_fqs);
1878 if (rcu_state.cbovld)
1879 gf = RCU_GP_FLAG_OVLD;
1880 ret = 0;
1881 for (;;) {
1882 if (!ret) {
1883 rcu_state.jiffies_force_qs = jiffies + j;
1884 WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
1885 jiffies + (j ? 3 * j : 2));
1886 }
1887 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1888 TPS("fqswait"));
1889 rcu_state.gp_state = RCU_GP_WAIT_FQS;
1890 ret = swait_event_idle_timeout_exclusive(
1891 rcu_state.gp_wq, rcu_gp_fqs_check_wake(&gf), j);
1892 rcu_gp_torture_wait();
1893 rcu_state.gp_state = RCU_GP_DOING_FQS;
1894 /* Locking provides needed memory barriers. */
1895 /* If grace period done, leave loop. */
1896 if (!READ_ONCE(rnp->qsmask) &&
1897 !rcu_preempt_blocked_readers_cgp(rnp))
1898 break;
1899 /* If time for quiescent-state forcing, do it. */
1900 if (!time_after(rcu_state.jiffies_force_qs, jiffies) ||
1901 (gf & RCU_GP_FLAG_FQS)) {
1902 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1903 TPS("fqsstart"));
1904 rcu_gp_fqs(first_gp_fqs);
1905 gf = 0;
1906 if (first_gp_fqs) {
1907 first_gp_fqs = false;
1908 gf = rcu_state.cbovld ? RCU_GP_FLAG_OVLD : 0;
1909 }
1910 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1911 TPS("fqsend"));
1912 cond_resched_tasks_rcu_qs();
1913 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1914 ret = 0; /* Force full wait till next FQS. */
1915 j = READ_ONCE(jiffies_till_next_fqs);
1916 } else {
1917 /* Deal with stray signal. */
1918 cond_resched_tasks_rcu_qs();
1919 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1920 WARN_ON(signal_pending(current));
1921 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1922 TPS("fqswaitsig"));
1923 ret = 1; /* Keep old FQS timing. */
1924 j = jiffies;
1925 if (time_after(jiffies, rcu_state.jiffies_force_qs))
1926 j = 1;
1927 else
1928 j = rcu_state.jiffies_force_qs - j;
1929 gf = 0;
1930 }
1931 }
1932}
1933
1934/*
1935 * Clean up after the old grace period.
1936 */
1937static void rcu_gp_cleanup(void)
1938{
1939 int cpu;
1940 bool needgp = false;
1941 unsigned long gp_duration;
1942 unsigned long new_gp_seq;
1943 bool offloaded;
1944 struct rcu_data *rdp;
1945 struct rcu_node *rnp = rcu_get_root();
1946 struct swait_queue_head *sq;
1947
1948 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1949 raw_spin_lock_irq_rcu_node(rnp);
1950 rcu_state.gp_end = jiffies;
1951 gp_duration = rcu_state.gp_end - rcu_state.gp_start;
1952 if (gp_duration > rcu_state.gp_max)
1953 rcu_state.gp_max = gp_duration;
1954
1955 /*
1956 * We know the grace period is complete, but to everyone else
1957 * it appears to still be ongoing. But it is also the case
1958 * that to everyone else it looks like there is nothing that
1959 * they can do to advance the grace period. It is therefore
1960 * safe for us to drop the lock in order to mark the grace
1961 * period as completed in all of the rcu_node structures.
1962 */
1963 raw_spin_unlock_irq_rcu_node(rnp);
1964
1965 /*
1966 * Propagate new ->gp_seq value to rcu_node structures so that
1967 * other CPUs don't have to wait until the start of the next grace
1968 * period to process their callbacks. This also avoids some nasty
1969 * RCU grace-period initialization races by forcing the end of
1970 * the current grace period to be completely recorded in all of
1971 * the rcu_node structures before the beginning of the next grace
1972 * period is recorded in any of the rcu_node structures.
1973 */
1974 new_gp_seq = rcu_state.gp_seq;
1975 rcu_seq_end(&new_gp_seq);
1976 rcu_for_each_node_breadth_first(rnp) {
1977 raw_spin_lock_irq_rcu_node(rnp);
1978 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
1979 dump_blkd_tasks(rnp, 10);
1980 WARN_ON_ONCE(rnp->qsmask);
1981 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
1982 rdp = this_cpu_ptr(&rcu_data);
1983 if (rnp == rdp->mynode)
1984 needgp = __note_gp_changes(rnp, rdp) || needgp;
1985 /* smp_mb() provided by prior unlock-lock pair. */
1986 needgp = rcu_future_gp_cleanup(rnp) || needgp;
1987 // Reset overload indication for CPUs no longer overloaded
1988 if (rcu_is_leaf_node(rnp))
1989 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->cbovldmask) {
1990 rdp = per_cpu_ptr(&rcu_data, cpu);
1991 check_cb_ovld_locked(rdp, rnp);
1992 }
1993 sq = rcu_nocb_gp_get(rnp);
1994 raw_spin_unlock_irq_rcu_node(rnp);
1995 rcu_nocb_gp_cleanup(sq);
1996 cond_resched_tasks_rcu_qs();
1997 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1998 rcu_gp_slow(gp_cleanup_delay);
1999 }
2000 rnp = rcu_get_root();
2001 raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
2002
2003 /* Declare grace period done, trace first to use old GP number. */
2004 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
2005 rcu_seq_end(&rcu_state.gp_seq);
2006 ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
2007 rcu_state.gp_state = RCU_GP_IDLE;
2008 /* Check for GP requests since above loop. */
2009 rdp = this_cpu_ptr(&rcu_data);
2010 if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
2011 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
2012 TPS("CleanupMore"));
2013 needgp = true;
2014 }
2015 /* Advance CBs to reduce false positives below. */
2016 offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2017 rcu_segcblist_is_offloaded(&rdp->cblist);
2018 if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
2019 WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
2020 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
2021 trace_rcu_grace_period(rcu_state.name,
2022 rcu_state.gp_seq,
2023 TPS("newreq"));
2024 } else {
2025 WRITE_ONCE(rcu_state.gp_flags,
2026 rcu_state.gp_flags & RCU_GP_FLAG_INIT);
2027 }
2028 raw_spin_unlock_irq_rcu_node(rnp);
2029}
2030
2031/*
2032 * Body of kthread that handles grace periods.
2033 */
2034static int __noreturn rcu_gp_kthread(void *unused)
2035{
2036 rcu_bind_gp_kthread();
2037 for (;;) {
2038
2039 /* Handle grace-period start. */
2040 for (;;) {
2041 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2042 TPS("reqwait"));
2043 rcu_state.gp_state = RCU_GP_WAIT_GPS;
2044 swait_event_idle_exclusive(rcu_state.gp_wq,
2045 READ_ONCE(rcu_state.gp_flags) &
2046 RCU_GP_FLAG_INIT);
2047 rcu_gp_torture_wait();
2048 rcu_state.gp_state = RCU_GP_DONE_GPS;
2049 /* Locking provides needed memory barrier. */
2050 if (rcu_gp_init())
2051 break;
2052 cond_resched_tasks_rcu_qs();
2053 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2054 WARN_ON(signal_pending(current));
2055 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2056 TPS("reqwaitsig"));
2057 }
2058
2059 /* Handle quiescent-state forcing. */
2060 rcu_gp_fqs_loop();
2061
2062 /* Handle grace-period end. */
2063 rcu_state.gp_state = RCU_GP_CLEANUP;
2064 rcu_gp_cleanup();
2065 rcu_state.gp_state = RCU_GP_CLEANED;
2066 }
2067}
2068
2069/*
2070 * Report a full set of quiescent states to the rcu_state data structure.
2071 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
2072 * another grace period is required. Whether we wake the grace-period
2073 * kthread or it awakens itself for the next round of quiescent-state
2074 * forcing, that kthread will clean up after the just-completed grace
2075 * period. Note that the caller must hold rnp->lock, which is released
2076 * before return.
2077 */
2078static void rcu_report_qs_rsp(unsigned long flags)
2079 __releases(rcu_get_root()->lock)
2080{
2081 raw_lockdep_assert_held_rcu_node(rcu_get_root());
2082 WARN_ON_ONCE(!rcu_gp_in_progress());
2083 WRITE_ONCE(rcu_state.gp_flags,
2084 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2085 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
2086 rcu_gp_kthread_wake();
2087}
2088
2089/*
2090 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2091 * Allows quiescent states for a group of CPUs to be reported at one go
2092 * to the specified rcu_node structure, though all the CPUs in the group
2093 * must be represented by the same rcu_node structure (which need not be a
2094 * leaf rcu_node structure, though it often will be). The gps parameter
2095 * is the grace-period snapshot, which means that the quiescent states
2096 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
2097 * must be held upon entry, and it is released before return.
2098 *
2099 * As a special case, if mask is zero, the bit-already-cleared check is
2100 * disabled. This allows propagating quiescent state due to resumed tasks
2101 * during grace-period initialization.
2102 */
2103static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
2104 unsigned long gps, unsigned long flags)
2105 __releases(rnp->lock)
2106{
2107 unsigned long oldmask = 0;
2108 struct rcu_node *rnp_c;
2109
2110 raw_lockdep_assert_held_rcu_node(rnp);
2111
2112 /* Walk up the rcu_node hierarchy. */
2113 for (;;) {
2114 if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
2115
2116 /*
2117 * Our bit has already been cleared, or the
2118 * relevant grace period is already over, so done.
2119 */
2120 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2121 return;
2122 }
2123 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2124 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
2125 rcu_preempt_blocked_readers_cgp(rnp));
2126 WRITE_ONCE(rnp->qsmask, rnp->qsmask & ~mask);
2127 trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
2128 mask, rnp->qsmask, rnp->level,
2129 rnp->grplo, rnp->grphi,
2130 !!rnp->gp_tasks);
2131 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2132
2133 /* Other bits still set at this level, so done. */
2134 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2135 return;
2136 }
2137 rnp->completedqs = rnp->gp_seq;
2138 mask = rnp->grpmask;
2139 if (rnp->parent == NULL) {
2140
2141 /* No more levels. Exit loop holding root lock. */
2142
2143 break;
2144 }
2145 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2146 rnp_c = rnp;
2147 rnp = rnp->parent;
2148 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2149 oldmask = READ_ONCE(rnp_c->qsmask);
2150 }
2151
2152 /*
2153 * Get here if we are the last CPU to pass through a quiescent
2154 * state for this grace period. Invoke rcu_report_qs_rsp()
2155 * to clean up and start the next grace period if one is needed.
2156 */
2157 rcu_report_qs_rsp(flags); /* releases rnp->lock. */
2158}
2159
2160/*
2161 * Record a quiescent state for all tasks that were previously queued
2162 * on the specified rcu_node structure and that were blocking the current
2163 * RCU grace period. The caller must hold the corresponding rnp->lock with
2164 * irqs disabled, and this lock is released upon return, but irqs remain
2165 * disabled.
2166 */
2167static void __maybe_unused
2168rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
2169 __releases(rnp->lock)
2170{
2171 unsigned long gps;
2172 unsigned long mask;
2173 struct rcu_node *rnp_p;
2174
2175 raw_lockdep_assert_held_rcu_node(rnp);
2176 if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RCU)) ||
2177 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
2178 rnp->qsmask != 0) {
2179 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2180 return; /* Still need more quiescent states! */
2181 }
2182
2183 rnp->completedqs = rnp->gp_seq;
2184 rnp_p = rnp->parent;
2185 if (rnp_p == NULL) {
2186 /*
2187 * Only one rcu_node structure in the tree, so don't
2188 * try to report up to its nonexistent parent!
2189 */
2190 rcu_report_qs_rsp(flags);
2191 return;
2192 }
2193
2194 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
2195 gps = rnp->gp_seq;
2196 mask = rnp->grpmask;
2197 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2198 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2199 rcu_report_qs_rnp(mask, rnp_p, gps, flags);
2200}
2201
2202/*
2203 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2204 * structure. This must be called from the specified CPU.
2205 */
2206static void
2207rcu_report_qs_rdp(int cpu, struct rcu_data *rdp)
2208{
2209 unsigned long flags;
2210 unsigned long mask;
2211 bool needwake = false;
2212 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2213 rcu_segcblist_is_offloaded(&rdp->cblist);
2214 struct rcu_node *rnp;
2215
2216 rnp = rdp->mynode;
2217 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2218 if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
2219 rdp->gpwrap) {
2220
2221 /*
2222 * The grace period in which this quiescent state was
2223 * recorded has ended, so don't report it upwards.
2224 * We will instead need a new quiescent state that lies
2225 * within the current grace period.
2226 */
2227 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2228 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2229 return;
2230 }
2231 mask = rdp->grpmask;
2232 if (rdp->cpu == smp_processor_id())
2233 rdp->core_needs_qs = false;
2234 if ((rnp->qsmask & mask) == 0) {
2235 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2236 } else {
2237 /*
2238 * This GP can't end until cpu checks in, so all of our
2239 * callbacks can be processed during the next GP.
2240 */
2241 if (!offloaded)
2242 needwake = rcu_accelerate_cbs(rnp, rdp);
2243
2244 rcu_disable_urgency_upon_qs(rdp);
2245 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2246 /* ^^^ Released rnp->lock */
2247 if (needwake)
2248 rcu_gp_kthread_wake();
2249 }
2250}
2251
2252/*
2253 * Check to see if there is a new grace period of which this CPU
2254 * is not yet aware, and if so, set up local rcu_data state for it.
2255 * Otherwise, see if this CPU has just passed through its first
2256 * quiescent state for this grace period, and record that fact if so.
2257 */
2258static void
2259rcu_check_quiescent_state(struct rcu_data *rdp)
2260{
2261 /* Check for grace-period ends and beginnings. */
2262 note_gp_changes(rdp);
2263
2264 /*
2265 * Does this CPU still need to do its part for current grace period?
2266 * If no, return and let the other CPUs do their part as well.
2267 */
2268 if (!rdp->core_needs_qs)
2269 return;
2270
2271 /*
2272 * Was there a quiescent state since the beginning of the grace
2273 * period? If no, then exit and wait for the next call.
2274 */
2275 if (rdp->cpu_no_qs.b.norm)
2276 return;
2277
2278 /*
2279 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2280 * judge of that).
2281 */
2282 rcu_report_qs_rdp(rdp->cpu, rdp);
2283}
2284
2285/*
2286 * Near the end of the offline process. Trace the fact that this CPU
2287 * is going offline.
2288 */
2289int rcutree_dying_cpu(unsigned int cpu)
2290{
2291 bool blkd;
2292 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
2293 struct rcu_node *rnp = rdp->mynode;
2294
2295 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2296 return 0;
2297
2298 blkd = !!(rnp->qsmask & rdp->grpmask);
2299 trace_rcu_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
2300 blkd ? TPS("cpuofl") : TPS("cpuofl-bgp"));
2301 return 0;
2302}
2303
2304/*
2305 * All CPUs for the specified rcu_node structure have gone offline,
2306 * and all tasks that were preempted within an RCU read-side critical
2307 * section while running on one of those CPUs have since exited their RCU
2308 * read-side critical section. Some other CPU is reporting this fact with
2309 * the specified rcu_node structure's ->lock held and interrupts disabled.
2310 * This function therefore goes up the tree of rcu_node structures,
2311 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2312 * the leaf rcu_node structure's ->qsmaskinit field has already been
2313 * updated.
2314 *
2315 * This function does check that the specified rcu_node structure has
2316 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2317 * prematurely. That said, invoking it after the fact will cost you
2318 * a needless lock acquisition. So once it has done its work, don't
2319 * invoke it again.
2320 */
2321static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2322{
2323 long mask;
2324 struct rcu_node *rnp = rnp_leaf;
2325
2326 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2327 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2328 WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
2329 WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
2330 return;
2331 for (;;) {
2332 mask = rnp->grpmask;
2333 rnp = rnp->parent;
2334 if (!rnp)
2335 break;
2336 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2337 rnp->qsmaskinit &= ~mask;
2338 /* Between grace periods, so better already be zero! */
2339 WARN_ON_ONCE(rnp->qsmask);
2340 if (rnp->qsmaskinit) {
2341 raw_spin_unlock_rcu_node(rnp);
2342 /* irqs remain disabled. */
2343 return;
2344 }
2345 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2346 }
2347}
2348
2349/*
2350 * The CPU has been completely removed, and some other CPU is reporting
2351 * this fact from process context. Do the remainder of the cleanup.
2352 * There can only be one CPU hotplug operation at a time, so no need for
2353 * explicit locking.
2354 */
2355int rcutree_dead_cpu(unsigned int cpu)
2356{
2357 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2358 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2359
2360 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2361 return 0;
2362
2363 /* Adjust any no-longer-needed kthreads. */
2364 rcu_boost_kthread_setaffinity(rnp, -1);
2365 /* Do any needed no-CB deferred wakeups from this CPU. */
2366 do_nocb_deferred_wakeup(per_cpu_ptr(&rcu_data, cpu));
2367
2368 // Stop-machine done, so allow nohz_full to disable tick.
2369 tick_dep_clear(TICK_DEP_BIT_RCU);
2370 return 0;
2371}
2372
2373/*
2374 * Invoke any RCU callbacks that have made it to the end of their grace
2375 * period. Thottle as specified by rdp->blimit.
2376 */
2377static void rcu_do_batch(struct rcu_data *rdp)
2378{
2379 unsigned long flags;
2380 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2381 rcu_segcblist_is_offloaded(&rdp->cblist);
2382 struct rcu_head *rhp;
2383 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2384 long bl, count;
2385 long pending, tlimit = 0;
2386
2387 /* If no callbacks are ready, just return. */
2388 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2389 trace_rcu_batch_start(rcu_state.name,
2390 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2391 trace_rcu_batch_end(rcu_state.name, 0,
2392 !rcu_segcblist_empty(&rdp->cblist),
2393 need_resched(), is_idle_task(current),
2394 rcu_is_callbacks_kthread());
2395 return;
2396 }
2397
2398 /*
2399 * Extract the list of ready callbacks, disabling to prevent
2400 * races with call_rcu() from interrupt handlers. Leave the
2401 * callback counts, as rcu_barrier() needs to be conservative.
2402 */
2403 local_irq_save(flags);
2404 rcu_nocb_lock(rdp);
2405 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2406 pending = rcu_segcblist_n_cbs(&rdp->cblist);
2407 bl = max(rdp->blimit, pending >> rcu_divisor);
2408 if (unlikely(bl > 100))
2409 tlimit = local_clock() + rcu_resched_ns;
2410 trace_rcu_batch_start(rcu_state.name,
2411 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2412 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2413 if (offloaded)
2414 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2415 rcu_nocb_unlock_irqrestore(rdp, flags);
2416
2417 /* Invoke callbacks. */
2418 tick_dep_set_task(current, TICK_DEP_BIT_RCU);
2419 rhp = rcu_cblist_dequeue(&rcl);
2420 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2421 rcu_callback_t f;
2422
2423 debug_rcu_head_unqueue(rhp);
2424
2425 rcu_lock_acquire(&rcu_callback_map);
2426 trace_rcu_invoke_callback(rcu_state.name, rhp);
2427
2428 f = rhp->func;
2429 WRITE_ONCE(rhp->func, (rcu_callback_t)0L);
2430 f(rhp);
2431
2432 rcu_lock_release(&rcu_callback_map);
2433
2434 /*
2435 * Stop only if limit reached and CPU has something to do.
2436 * Note: The rcl structure counts down from zero.
2437 */
2438 if (-rcl.len >= bl && !offloaded &&
2439 (need_resched() ||
2440 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2441 break;
2442 if (unlikely(tlimit)) {
2443 /* only call local_clock() every 32 callbacks */
2444 if (likely((-rcl.len & 31) || local_clock() < tlimit))
2445 continue;
2446 /* Exceeded the time limit, so leave. */
2447 break;
2448 }
2449 if (offloaded) {
2450 WARN_ON_ONCE(in_serving_softirq());
2451 local_bh_enable();
2452 lockdep_assert_irqs_enabled();
2453 cond_resched_tasks_rcu_qs();
2454 lockdep_assert_irqs_enabled();
2455 local_bh_disable();
2456 }
2457 }
2458
2459 local_irq_save(flags);
2460 rcu_nocb_lock(rdp);
2461 count = -rcl.len;
2462 rdp->n_cbs_invoked += count;
2463 trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2464 is_idle_task(current), rcu_is_callbacks_kthread());
2465
2466 /* Update counts and requeue any remaining callbacks. */
2467 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2468 smp_mb(); /* List handling before counting for rcu_barrier(). */
2469 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2470
2471 /* Reinstate batch limit if we have worked down the excess. */
2472 count = rcu_segcblist_n_cbs(&rdp->cblist);
2473 if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
2474 rdp->blimit = blimit;
2475
2476 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2477 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2478 rdp->qlen_last_fqs_check = 0;
2479 rdp->n_force_qs_snap = rcu_state.n_force_qs;
2480 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2481 rdp->qlen_last_fqs_check = count;
2482
2483 /*
2484 * The following usually indicates a double call_rcu(). To track
2485 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2486 */
2487 WARN_ON_ONCE(count == 0 && !rcu_segcblist_empty(&rdp->cblist));
2488 WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2489 count != 0 && rcu_segcblist_empty(&rdp->cblist));
2490
2491 rcu_nocb_unlock_irqrestore(rdp, flags);
2492
2493 /* Re-invoke RCU core processing if there are callbacks remaining. */
2494 if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist))
2495 invoke_rcu_core();
2496 tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
2497}
2498
2499/*
2500 * This function is invoked from each scheduling-clock interrupt,
2501 * and checks to see if this CPU is in a non-context-switch quiescent
2502 * state, for example, user mode or idle loop. It also schedules RCU
2503 * core processing. If the current grace period has gone on too long,
2504 * it will ask the scheduler to manufacture a context switch for the sole
2505 * purpose of providing a providing the needed quiescent state.
2506 */
2507void rcu_sched_clock_irq(int user)
2508{
2509 trace_rcu_utilization(TPS("Start scheduler-tick"));
2510 raw_cpu_inc(rcu_data.ticks_this_gp);
2511 /* The load-acquire pairs with the store-release setting to true. */
2512 if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
2513 /* Idle and userspace execution already are quiescent states. */
2514 if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2515 set_tsk_need_resched(current);
2516 set_preempt_need_resched();
2517 }
2518 __this_cpu_write(rcu_data.rcu_urgent_qs, false);
2519 }
2520 rcu_flavor_sched_clock_irq(user);
2521 if (rcu_pending(user))
2522 invoke_rcu_core();
2523
2524 trace_rcu_utilization(TPS("End scheduler-tick"));
2525}
2526
2527/*
2528 * Scan the leaf rcu_node structures. For each structure on which all
2529 * CPUs have reported a quiescent state and on which there are tasks
2530 * blocking the current grace period, initiate RCU priority boosting.
2531 * Otherwise, invoke the specified function to check dyntick state for
2532 * each CPU that has not yet reported a quiescent state.
2533 */
2534static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
2535{
2536 int cpu;
2537 unsigned long flags;
2538 unsigned long mask;
2539 struct rcu_data *rdp;
2540 struct rcu_node *rnp;
2541
2542 rcu_state.cbovld = rcu_state.cbovldnext;
2543 rcu_state.cbovldnext = false;
2544 rcu_for_each_leaf_node(rnp) {
2545 cond_resched_tasks_rcu_qs();
2546 mask = 0;
2547 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2548 rcu_state.cbovldnext |= !!rnp->cbovldmask;
2549 if (rnp->qsmask == 0) {
2550 if (!IS_ENABLED(CONFIG_PREEMPT_RCU) ||
2551 rcu_preempt_blocked_readers_cgp(rnp)) {
2552 /*
2553 * No point in scanning bits because they
2554 * are all zero. But we might need to
2555 * priority-boost blocked readers.
2556 */
2557 rcu_initiate_boost(rnp, flags);
2558 /* rcu_initiate_boost() releases rnp->lock */
2559 continue;
2560 }
2561 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2562 continue;
2563 }
2564 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->qsmask) {
2565 rdp = per_cpu_ptr(&rcu_data, cpu);
2566 if (f(rdp)) {
2567 mask |= rdp->grpmask;
2568 rcu_disable_urgency_upon_qs(rdp);
2569 }
2570 }
2571 if (mask != 0) {
2572 /* Idle/offline CPUs, report (releases rnp->lock). */
2573 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2574 } else {
2575 /* Nothing to do here, so just drop the lock. */
2576 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2577 }
2578 }
2579}
2580
2581/*
2582 * Force quiescent states on reluctant CPUs, and also detect which
2583 * CPUs are in dyntick-idle mode.
2584 */
2585void rcu_force_quiescent_state(void)
2586{
2587 unsigned long flags;
2588 bool ret;
2589 struct rcu_node *rnp;
2590 struct rcu_node *rnp_old = NULL;
2591
2592 /* Funnel through hierarchy to reduce memory contention. */
2593 rnp = __this_cpu_read(rcu_data.mynode);
2594 for (; rnp != NULL; rnp = rnp->parent) {
2595 ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2596 !raw_spin_trylock(&rnp->fqslock);
2597 if (rnp_old != NULL)
2598 raw_spin_unlock(&rnp_old->fqslock);
2599 if (ret)
2600 return;
2601 rnp_old = rnp;
2602 }
2603 /* rnp_old == rcu_get_root(), rnp == NULL. */
2604
2605 /* Reached the root of the rcu_node tree, acquire lock. */
2606 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2607 raw_spin_unlock(&rnp_old->fqslock);
2608 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2609 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2610 return; /* Someone beat us to it. */
2611 }
2612 WRITE_ONCE(rcu_state.gp_flags,
2613 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2614 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2615 rcu_gp_kthread_wake();
2616}
2617EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
2618
2619/* Perform RCU core processing work for the current CPU. */
2620static __latent_entropy void rcu_core(void)
2621{
2622 unsigned long flags;
2623 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2624 struct rcu_node *rnp = rdp->mynode;
2625 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2626 rcu_segcblist_is_offloaded(&rdp->cblist);
2627
2628 if (cpu_is_offline(smp_processor_id()))
2629 return;
2630 trace_rcu_utilization(TPS("Start RCU core"));
2631 WARN_ON_ONCE(!rdp->beenonline);
2632
2633 /* Report any deferred quiescent states if preemption enabled. */
2634 if (!(preempt_count() & PREEMPT_MASK)) {
2635 rcu_preempt_deferred_qs(current);
2636 } else if (rcu_preempt_need_deferred_qs(current)) {
2637 set_tsk_need_resched(current);
2638 set_preempt_need_resched();
2639 }
2640
2641 /* Update RCU state based on any recent quiescent states. */
2642 rcu_check_quiescent_state(rdp);
2643
2644 /* No grace period and unregistered callbacks? */
2645 if (!rcu_gp_in_progress() &&
2646 rcu_segcblist_is_enabled(&rdp->cblist) && !offloaded) {
2647 local_irq_save(flags);
2648 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2649 rcu_accelerate_cbs_unlocked(rnp, rdp);
2650 local_irq_restore(flags);
2651 }
2652
2653 rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
2654
2655 /* If there are callbacks ready, invoke them. */
2656 if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist) &&
2657 likely(READ_ONCE(rcu_scheduler_fully_active)))
2658 rcu_do_batch(rdp);
2659
2660 /* Do any needed deferred wakeups of rcuo kthreads. */
2661 do_nocb_deferred_wakeup(rdp);
2662 trace_rcu_utilization(TPS("End RCU core"));
2663}
2664
2665static void rcu_core_si(struct softirq_action *h)
2666{
2667 rcu_core();
2668}
2669
2670static void rcu_wake_cond(struct task_struct *t, int status)
2671{
2672 /*
2673 * If the thread is yielding, only wake it when this
2674 * is invoked from idle
2675 */
2676 if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
2677 wake_up_process(t);
2678}
2679
2680static void invoke_rcu_core_kthread(void)
2681{
2682 struct task_struct *t;
2683 unsigned long flags;
2684
2685 local_irq_save(flags);
2686 __this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
2687 t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
2688 if (t != NULL && t != current)
2689 rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
2690 local_irq_restore(flags);
2691}
2692
2693/*
2694 * Wake up this CPU's rcuc kthread to do RCU core processing.
2695 */
2696static void invoke_rcu_core(void)
2697{
2698 if (!cpu_online(smp_processor_id()))
2699 return;
2700 if (use_softirq)
2701 raise_softirq(RCU_SOFTIRQ);
2702 else
2703 invoke_rcu_core_kthread();
2704}
2705
2706static void rcu_cpu_kthread_park(unsigned int cpu)
2707{
2708 per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
2709}
2710
2711static int rcu_cpu_kthread_should_run(unsigned int cpu)
2712{
2713 return __this_cpu_read(rcu_data.rcu_cpu_has_work);
2714}
2715
2716/*
2717 * Per-CPU kernel thread that invokes RCU callbacks. This replaces
2718 * the RCU softirq used in configurations of RCU that do not support RCU
2719 * priority boosting.
2720 */
2721static void rcu_cpu_kthread(unsigned int cpu)
2722{
2723 unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
2724 char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
2725 int spincnt;
2726
2727 trace_rcu_utilization(TPS("Start CPU kthread@rcu_run"));
2728 for (spincnt = 0; spincnt < 10; spincnt++) {
2729 local_bh_disable();
2730 *statusp = RCU_KTHREAD_RUNNING;
2731 local_irq_disable();
2732 work = *workp;
2733 *workp = 0;
2734 local_irq_enable();
2735 if (work)
2736 rcu_core();
2737 local_bh_enable();
2738 if (*workp == 0) {
2739 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2740 *statusp = RCU_KTHREAD_WAITING;
2741 return;
2742 }
2743 }
2744 *statusp = RCU_KTHREAD_YIELDING;
2745 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2746 schedule_timeout_idle(2);
2747 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2748 *statusp = RCU_KTHREAD_WAITING;
2749}
2750
2751static struct smp_hotplug_thread rcu_cpu_thread_spec = {
2752 .store = &rcu_data.rcu_cpu_kthread_task,
2753 .thread_should_run = rcu_cpu_kthread_should_run,
2754 .thread_fn = rcu_cpu_kthread,
2755 .thread_comm = "rcuc/%u",
2756 .setup = rcu_cpu_kthread_setup,
2757 .park = rcu_cpu_kthread_park,
2758};
2759
2760/*
2761 * Spawn per-CPU RCU core processing kthreads.
2762 */
2763static int __init rcu_spawn_core_kthreads(void)
2764{
2765 int cpu;
2766
2767 for_each_possible_cpu(cpu)
2768 per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
2769 if (!IS_ENABLED(CONFIG_RCU_BOOST) && use_softirq)
2770 return 0;
2771 WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
2772 "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
2773 return 0;
2774}
2775early_initcall(rcu_spawn_core_kthreads);
2776
2777/*
2778 * Handle any core-RCU processing required by a call_rcu() invocation.
2779 */
2780static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2781 unsigned long flags)
2782{
2783 /*
2784 * If called from an extended quiescent state, invoke the RCU
2785 * core in order to force a re-evaluation of RCU's idleness.
2786 */
2787 if (!rcu_is_watching())
2788 invoke_rcu_core();
2789
2790 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2791 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2792 return;
2793
2794 /*
2795 * Force the grace period if too many callbacks or too long waiting.
2796 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2797 * if some other CPU has recently done so. Also, don't bother
2798 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2799 * is the only one waiting for a grace period to complete.
2800 */
2801 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2802 rdp->qlen_last_fqs_check + qhimark)) {
2803
2804 /* Are we ignoring a completed grace period? */
2805 note_gp_changes(rdp);
2806
2807 /* Start a new grace period if one not already started. */
2808 if (!rcu_gp_in_progress()) {
2809 rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
2810 } else {
2811 /* Give the grace period a kick. */
2812 rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
2813 if (rcu_state.n_force_qs == rdp->n_force_qs_snap &&
2814 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2815 rcu_force_quiescent_state();
2816 rdp->n_force_qs_snap = rcu_state.n_force_qs;
2817 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2818 }
2819 }
2820}
2821
2822/*
2823 * RCU callback function to leak a callback.
2824 */
2825static void rcu_leak_callback(struct rcu_head *rhp)
2826{
2827}
2828
2829/*
2830 * Check and if necessary update the leaf rcu_node structure's
2831 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2832 * number of queued RCU callbacks. The caller must hold the leaf rcu_node
2833 * structure's ->lock.
2834 */
2835static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp)
2836{
2837 raw_lockdep_assert_held_rcu_node(rnp);
2838 if (qovld_calc <= 0)
2839 return; // Early boot and wildcard value set.
2840 if (rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc)
2841 WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask | rdp->grpmask);
2842 else
2843 WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask & ~rdp->grpmask);
2844}
2845
2846/*
2847 * Check and if necessary update the leaf rcu_node structure's
2848 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2849 * number of queued RCU callbacks. No locks need be held, but the
2850 * caller must have disabled interrupts.
2851 *
2852 * Note that this function ignores the possibility that there are a lot
2853 * of callbacks all of which have already seen the end of their respective
2854 * grace periods. This omission is due to the need for no-CBs CPUs to
2855 * be holding ->nocb_lock to do this check, which is too heavy for a
2856 * common-case operation.
2857 */
2858static void check_cb_ovld(struct rcu_data *rdp)
2859{
2860 struct rcu_node *const rnp = rdp->mynode;
2861
2862 if (qovld_calc <= 0 ||
2863 ((rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) ==
2864 !!(READ_ONCE(rnp->cbovldmask) & rdp->grpmask)))
2865 return; // Early boot wildcard value or already set correctly.
2866 raw_spin_lock_rcu_node(rnp);
2867 check_cb_ovld_locked(rdp, rnp);
2868 raw_spin_unlock_rcu_node(rnp);
2869}
2870
2871/* Helper function for call_rcu() and friends. */
2872static void
2873__call_rcu(struct rcu_head *head, rcu_callback_t func)
2874{
2875 unsigned long flags;
2876 struct rcu_data *rdp;
2877 bool was_alldone;
2878
2879 /* Misaligned rcu_head! */
2880 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2881
2882 if (debug_rcu_head_queue(head)) {
2883 /*
2884 * Probable double call_rcu(), so leak the callback.
2885 * Use rcu:rcu_callback trace event to find the previous
2886 * time callback was passed to __call_rcu().
2887 */
2888 WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pS()!!!\n",
2889 head, head->func);
2890 WRITE_ONCE(head->func, rcu_leak_callback);
2891 return;
2892 }
2893 head->func = func;
2894 head->next = NULL;
2895 local_irq_save(flags);
2896 kasan_record_aux_stack(head);
2897 rdp = this_cpu_ptr(&rcu_data);
2898
2899 /* Add the callback to our list. */
2900 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
2901 // This can trigger due to call_rcu() from offline CPU:
2902 WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
2903 WARN_ON_ONCE(!rcu_is_watching());
2904 // Very early boot, before rcu_init(). Initialize if needed
2905 // and then drop through to queue the callback.
2906 if (rcu_segcblist_empty(&rdp->cblist))
2907 rcu_segcblist_init(&rdp->cblist);
2908 }
2909
2910 check_cb_ovld(rdp);
2911 if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags))
2912 return; // Enqueued onto ->nocb_bypass, so just leave.
2913 // If no-CBs CPU gets here, rcu_nocb_try_bypass() acquired ->nocb_lock.
2914 rcu_segcblist_enqueue(&rdp->cblist, head);
2915 if (__is_kvfree_rcu_offset((unsigned long)func))
2916 trace_rcu_kvfree_callback(rcu_state.name, head,
2917 (unsigned long)func,
2918 rcu_segcblist_n_cbs(&rdp->cblist));
2919 else
2920 trace_rcu_callback(rcu_state.name, head,
2921 rcu_segcblist_n_cbs(&rdp->cblist));
2922
2923 /* Go handle any RCU core processing required. */
2924 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2925 unlikely(rcu_segcblist_is_offloaded(&rdp->cblist))) {
2926 __call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */
2927 } else {
2928 __call_rcu_core(rdp, head, flags);
2929 local_irq_restore(flags);
2930 }
2931}
2932
2933/**
2934 * call_rcu() - Queue an RCU callback for invocation after a grace period.
2935 * @head: structure to be used for queueing the RCU updates.
2936 * @func: actual callback function to be invoked after the grace period
2937 *
2938 * The callback function will be invoked some time after a full grace
2939 * period elapses, in other words after all pre-existing RCU read-side
2940 * critical sections have completed. However, the callback function
2941 * might well execute concurrently with RCU read-side critical sections
2942 * that started after call_rcu() was invoked. RCU read-side critical
2943 * sections are delimited by rcu_read_lock() and rcu_read_unlock(), and
2944 * may be nested. In addition, regions of code across which interrupts,
2945 * preemption, or softirqs have been disabled also serve as RCU read-side
2946 * critical sections. This includes hardware interrupt handlers, softirq
2947 * handlers, and NMI handlers.
2948 *
2949 * Note that all CPUs must agree that the grace period extended beyond
2950 * all pre-existing RCU read-side critical section. On systems with more
2951 * than one CPU, this means that when "func()" is invoked, each CPU is
2952 * guaranteed to have executed a full memory barrier since the end of its
2953 * last RCU read-side critical section whose beginning preceded the call
2954 * to call_rcu(). It also means that each CPU executing an RCU read-side
2955 * critical section that continues beyond the start of "func()" must have
2956 * executed a memory barrier after the call_rcu() but before the beginning
2957 * of that RCU read-side critical section. Note that these guarantees
2958 * include CPUs that are offline, idle, or executing in user mode, as
2959 * well as CPUs that are executing in the kernel.
2960 *
2961 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
2962 * resulting RCU callback function "func()", then both CPU A and CPU B are
2963 * guaranteed to execute a full memory barrier during the time interval
2964 * between the call to call_rcu() and the invocation of "func()" -- even
2965 * if CPU A and CPU B are the same CPU (but again only if the system has
2966 * more than one CPU).
2967 */
2968void call_rcu(struct rcu_head *head, rcu_callback_t func)
2969{
2970 __call_rcu(head, func);
2971}
2972EXPORT_SYMBOL_GPL(call_rcu);
2973
2974
2975/* Maximum number of jiffies to wait before draining a batch. */
2976#define KFREE_DRAIN_JIFFIES (HZ / 50)
2977#define KFREE_N_BATCHES 2
2978#define FREE_N_CHANNELS 2
2979
2980/**
2981 * struct kvfree_rcu_bulk_data - single block to store kvfree_rcu() pointers
2982 * @nr_records: Number of active pointers in the array
2983 * @next: Next bulk object in the block chain
2984 * @records: Array of the kvfree_rcu() pointers
2985 */
2986struct kvfree_rcu_bulk_data {
2987 unsigned long nr_records;
2988 struct kvfree_rcu_bulk_data *next;
2989 void *records[];
2990};
2991
2992/*
2993 * This macro defines how many entries the "records" array
2994 * will contain. It is based on the fact that the size of
2995 * kvfree_rcu_bulk_data structure becomes exactly one page.
2996 */
2997#define KVFREE_BULK_MAX_ENTR \
2998 ((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *))
2999
3000/**
3001 * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
3002 * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
3003 * @head_free: List of kfree_rcu() objects waiting for a grace period
3004 * @bkvhead_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
3005 * @krcp: Pointer to @kfree_rcu_cpu structure
3006 */
3007
3008struct kfree_rcu_cpu_work {
3009 struct rcu_work rcu_work;
3010 struct rcu_head *head_free;
3011 struct kvfree_rcu_bulk_data *bkvhead_free[FREE_N_CHANNELS];
3012 struct kfree_rcu_cpu *krcp;
3013};
3014
3015/**
3016 * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
3017 * @head: List of kfree_rcu() objects not yet waiting for a grace period
3018 * @bkvhead: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period
3019 * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
3020 * @lock: Synchronize access to this structure
3021 * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES
3022 * @monitor_todo: Tracks whether a @monitor_work delayed work is pending
3023 * @initialized: The @rcu_work fields have been initialized
3024 * @count: Number of objects for which GP not started
3025 *
3026 * This is a per-CPU structure. The reason that it is not included in
3027 * the rcu_data structure is to permit this code to be extracted from
3028 * the RCU files. Such extraction could allow further optimization of
3029 * the interactions with the slab allocators.
3030 */
3031struct kfree_rcu_cpu {
3032 struct rcu_head *head;
3033 struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS];
3034 struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
3035 raw_spinlock_t lock;
3036 struct delayed_work monitor_work;
3037 bool monitor_todo;
3038 bool initialized;
3039 int count;
3040
3041 /*
3042 * A simple cache list that contains objects for
3043 * reuse purpose. In order to save some per-cpu
3044 * space the list is singular. Even though it is
3045 * lockless an access has to be protected by the
3046 * per-cpu lock.
3047 */
3048 struct llist_head bkvcache;
3049 int nr_bkv_objs;
3050};
3051
3052static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc) = {
3053 .lock = __RAW_SPIN_LOCK_UNLOCKED(krc.lock),
3054};
3055
3056static __always_inline void
3057debug_rcu_bhead_unqueue(struct kvfree_rcu_bulk_data *bhead)
3058{
3059#ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
3060 int i;
3061
3062 for (i = 0; i < bhead->nr_records; i++)
3063 debug_rcu_head_unqueue((struct rcu_head *)(bhead->records[i]));
3064#endif
3065}
3066
3067static inline struct kfree_rcu_cpu *
3068krc_this_cpu_lock(unsigned long *flags)
3069{
3070 struct kfree_rcu_cpu *krcp;
3071
3072 local_irq_save(*flags); // For safely calling this_cpu_ptr().
3073 krcp = this_cpu_ptr(&krc);
3074 raw_spin_lock(&krcp->lock);
3075
3076 return krcp;
3077}
3078
3079static inline void
3080krc_this_cpu_unlock(struct kfree_rcu_cpu *krcp, unsigned long flags)
3081{
3082 raw_spin_unlock(&krcp->lock);
3083 local_irq_restore(flags);
3084}
3085
3086static inline struct kvfree_rcu_bulk_data *
3087get_cached_bnode(struct kfree_rcu_cpu *krcp)
3088{
3089 if (!krcp->nr_bkv_objs)
3090 return NULL;
3091
3092 krcp->nr_bkv_objs--;
3093 return (struct kvfree_rcu_bulk_data *)
3094 llist_del_first(&krcp->bkvcache);
3095}
3096
3097static inline bool
3098put_cached_bnode(struct kfree_rcu_cpu *krcp,
3099 struct kvfree_rcu_bulk_data *bnode)
3100{
3101 // Check the limit.
3102 if (krcp->nr_bkv_objs >= rcu_min_cached_objs)
3103 return false;
3104
3105 llist_add((struct llist_node *) bnode, &krcp->bkvcache);
3106 krcp->nr_bkv_objs++;
3107 return true;
3108
3109}
3110
3111/*
3112 * This function is invoked in workqueue context after a grace period.
3113 * It frees all the objects queued on ->bhead_free or ->head_free.
3114 */
3115static void kfree_rcu_work(struct work_struct *work)
3116{
3117 unsigned long flags;
3118 struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS], *bnext;
3119 struct rcu_head *head, *next;
3120 struct kfree_rcu_cpu *krcp;
3121 struct kfree_rcu_cpu_work *krwp;
3122 int i, j;
3123
3124 krwp = container_of(to_rcu_work(work),
3125 struct kfree_rcu_cpu_work, rcu_work);
3126 krcp = krwp->krcp;
3127
3128 raw_spin_lock_irqsave(&krcp->lock, flags);
3129 // Channels 1 and 2.
3130 for (i = 0; i < FREE_N_CHANNELS; i++) {
3131 bkvhead[i] = krwp->bkvhead_free[i];
3132 krwp->bkvhead_free[i] = NULL;
3133 }
3134
3135 // Channel 3.
3136 head = krwp->head_free;
3137 krwp->head_free = NULL;
3138 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3139
3140 // Handle two first channels.
3141 for (i = 0; i < FREE_N_CHANNELS; i++) {
3142 for (; bkvhead[i]; bkvhead[i] = bnext) {
3143 bnext = bkvhead[i]->next;
3144 debug_rcu_bhead_unqueue(bkvhead[i]);
3145
3146 rcu_lock_acquire(&rcu_callback_map);
3147 if (i == 0) { // kmalloc() / kfree().
3148 trace_rcu_invoke_kfree_bulk_callback(
3149 rcu_state.name, bkvhead[i]->nr_records,
3150 bkvhead[i]->records);
3151
3152 kfree_bulk(bkvhead[i]->nr_records,
3153 bkvhead[i]->records);
3154 } else { // vmalloc() / vfree().
3155 for (j = 0; j < bkvhead[i]->nr_records; j++) {
3156 trace_rcu_invoke_kvfree_callback(
3157 rcu_state.name,
3158 bkvhead[i]->records[j], 0);
3159
3160 vfree(bkvhead[i]->records[j]);
3161 }
3162 }
3163 rcu_lock_release(&rcu_callback_map);
3164
3165 krcp = krc_this_cpu_lock(&flags);
3166 if (put_cached_bnode(krcp, bkvhead[i]))
3167 bkvhead[i] = NULL;
3168 krc_this_cpu_unlock(krcp, flags);
3169
3170 if (bkvhead[i])
3171 free_page((unsigned long) bkvhead[i]);
3172
3173 cond_resched_tasks_rcu_qs();
3174 }
3175 }
3176
3177 /*
3178 * Emergency case only. It can happen under low memory
3179 * condition when an allocation gets failed, so the "bulk"
3180 * path can not be temporary maintained.
3181 */
3182 for (; head; head = next) {
3183 unsigned long offset = (unsigned long)head->func;
3184 void *ptr = (void *)head - offset;
3185
3186 next = head->next;
3187 debug_rcu_head_unqueue((struct rcu_head *)ptr);
3188 rcu_lock_acquire(&rcu_callback_map);
3189 trace_rcu_invoke_kvfree_callback(rcu_state.name, head, offset);
3190
3191 if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset)))
3192 kvfree(ptr);
3193
3194 rcu_lock_release(&rcu_callback_map);
3195 cond_resched_tasks_rcu_qs();
3196 }
3197}
3198
3199/*
3200 * Schedule the kfree batch RCU work to run in workqueue context after a GP.
3201 *
3202 * This function is invoked by kfree_rcu_monitor() when the KFREE_DRAIN_JIFFIES
3203 * timeout has been reached.
3204 */
3205static inline bool queue_kfree_rcu_work(struct kfree_rcu_cpu *krcp)
3206{
3207 struct kfree_rcu_cpu_work *krwp;
3208 bool repeat = false;
3209 int i, j;
3210
3211 lockdep_assert_held(&krcp->lock);
3212
3213 for (i = 0; i < KFREE_N_BATCHES; i++) {
3214 krwp = &(krcp->krw_arr[i]);
3215
3216 /*
3217 * Try to detach bkvhead or head and attach it over any
3218 * available corresponding free channel. It can be that
3219 * a previous RCU batch is in progress, it means that
3220 * immediately to queue another one is not possible so
3221 * return false to tell caller to retry.
3222 */
3223 if ((krcp->bkvhead[0] && !krwp->bkvhead_free[0]) ||
3224 (krcp->bkvhead[1] && !krwp->bkvhead_free[1]) ||
3225 (krcp->head && !krwp->head_free)) {
3226 // Channel 1 corresponds to SLAB ptrs.
3227 // Channel 2 corresponds to vmalloc ptrs.
3228 for (j = 0; j < FREE_N_CHANNELS; j++) {
3229 if (!krwp->bkvhead_free[j]) {
3230 krwp->bkvhead_free[j] = krcp->bkvhead[j];
3231 krcp->bkvhead[j] = NULL;
3232 }
3233 }
3234
3235 // Channel 3 corresponds to emergency path.
3236 if (!krwp->head_free) {
3237 krwp->head_free = krcp->head;
3238 krcp->head = NULL;
3239 }
3240
3241 WRITE_ONCE(krcp->count, 0);
3242
3243 /*
3244 * One work is per one batch, so there are three
3245 * "free channels", the batch can handle. It can
3246 * be that the work is in the pending state when
3247 * channels have been detached following by each
3248 * other.
3249 */
3250 queue_rcu_work(system_wq, &krwp->rcu_work);
3251 }
3252
3253 // Repeat if any "free" corresponding channel is still busy.
3254 if (krcp->bkvhead[0] || krcp->bkvhead[1] || krcp->head)
3255 repeat = true;
3256 }
3257
3258 return !repeat;
3259}
3260
3261static inline void kfree_rcu_drain_unlock(struct kfree_rcu_cpu *krcp,
3262 unsigned long flags)
3263{
3264 // Attempt to start a new batch.
3265 krcp->monitor_todo = false;
3266 if (queue_kfree_rcu_work(krcp)) {
3267 // Success! Our job is done here.
3268 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3269 return;
3270 }
3271
3272 // Previous RCU batch still in progress, try again later.
3273 krcp->monitor_todo = true;
3274 schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES);
3275 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3276}
3277
3278/*
3279 * This function is invoked after the KFREE_DRAIN_JIFFIES timeout.
3280 * It invokes kfree_rcu_drain_unlock() to attempt to start another batch.
3281 */
3282static void kfree_rcu_monitor(struct work_struct *work)
3283{
3284 unsigned long flags;
3285 struct kfree_rcu_cpu *krcp = container_of(work, struct kfree_rcu_cpu,
3286 monitor_work.work);
3287
3288 raw_spin_lock_irqsave(&krcp->lock, flags);
3289 if (krcp->monitor_todo)
3290 kfree_rcu_drain_unlock(krcp, flags);
3291 else
3292 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3293}
3294
3295static inline bool
3296kvfree_call_rcu_add_ptr_to_bulk(struct kfree_rcu_cpu *krcp, void *ptr)
3297{
3298 struct kvfree_rcu_bulk_data *bnode;
3299 int idx;
3300
3301 if (unlikely(!krcp->initialized))
3302 return false;
3303
3304 lockdep_assert_held(&krcp->lock);
3305 idx = !!is_vmalloc_addr(ptr);
3306
3307 /* Check if a new block is required. */
3308 if (!krcp->bkvhead[idx] ||
3309 krcp->bkvhead[idx]->nr_records == KVFREE_BULK_MAX_ENTR) {
3310 bnode = get_cached_bnode(krcp);
3311 if (!bnode) {
3312 /*
3313 * To keep this path working on raw non-preemptible
3314 * sections, prevent the optional entry into the
3315 * allocator as it uses sleeping locks. In fact, even
3316 * if the caller of kfree_rcu() is preemptible, this
3317 * path still is not, as krcp->lock is a raw spinlock.
3318 * With additional page pre-allocation in the works,
3319 * hitting this return is going to be much less likely.
3320 */
3321 if (IS_ENABLED(CONFIG_PREEMPT_RT))
3322 return false;
3323
3324 /*
3325 * NOTE: For one argument of kvfree_rcu() we can
3326 * drop the lock and get the page in sleepable
3327 * context. That would allow to maintain an array
3328 * for the CONFIG_PREEMPT_RT as well if no cached
3329 * pages are available.
3330 */
3331 bnode = (struct kvfree_rcu_bulk_data *)
3332 __get_free_page(GFP_NOWAIT | __GFP_NOWARN);
3333 }
3334
3335 /* Switch to emergency path. */
3336 if (unlikely(!bnode))
3337 return false;
3338
3339 /* Initialize the new block. */
3340 bnode->nr_records = 0;
3341 bnode->next = krcp->bkvhead[idx];
3342
3343 /* Attach it to the head. */
3344 krcp->bkvhead[idx] = bnode;
3345 }
3346
3347 /* Finally insert. */
3348 krcp->bkvhead[idx]->records
3349 [krcp->bkvhead[idx]->nr_records++] = ptr;
3350
3351 return true;
3352}
3353
3354/*
3355 * Queue a request for lazy invocation of appropriate free routine after a
3356 * grace period. Please note there are three paths are maintained, two are the
3357 * main ones that use array of pointers interface and third one is emergency
3358 * one, that is used only when the main path can not be maintained temporary,
3359 * due to memory pressure.
3360 *
3361 * Each kvfree_call_rcu() request is added to a batch. The batch will be drained
3362 * every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch will
3363 * be free'd in workqueue context. This allows us to: batch requests together to
3364 * reduce the number of grace periods during heavy kfree_rcu()/kvfree_rcu() load.
3365 */
3366void kvfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
3367{
3368 unsigned long flags;
3369 struct kfree_rcu_cpu *krcp;
3370 bool success;
3371 void *ptr;
3372
3373 if (head) {
3374 ptr = (void *) head - (unsigned long) func;
3375 } else {
3376 /*
3377 * Please note there is a limitation for the head-less
3378 * variant, that is why there is a clear rule for such
3379 * objects: it can be used from might_sleep() context
3380 * only. For other places please embed an rcu_head to
3381 * your data.
3382 */
3383 might_sleep();
3384 ptr = (unsigned long *) func;
3385 }
3386
3387 krcp = krc_this_cpu_lock(&flags);
3388
3389 // Queue the object but don't yet schedule the batch.
3390 if (debug_rcu_head_queue(ptr)) {
3391 // Probable double kfree_rcu(), just leak.
3392 WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n",
3393 __func__, head);
3394
3395 // Mark as success and leave.
3396 success = true;
3397 goto unlock_return;
3398 }
3399
3400 /*
3401 * Under high memory pressure GFP_NOWAIT can fail,
3402 * in that case the emergency path is maintained.
3403 */
3404 success = kvfree_call_rcu_add_ptr_to_bulk(krcp, ptr);
3405 if (!success) {
3406 if (head == NULL)
3407 // Inline if kvfree_rcu(one_arg) call.
3408 goto unlock_return;
3409
3410 head->func = func;
3411 head->next = krcp->head;
3412 krcp->head = head;
3413 success = true;
3414 }
3415
3416 WRITE_ONCE(krcp->count, krcp->count + 1);
3417
3418 // Set timer to drain after KFREE_DRAIN_JIFFIES.
3419 if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
3420 !krcp->monitor_todo) {
3421 krcp->monitor_todo = true;
3422 schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES);
3423 }
3424
3425unlock_return:
3426 krc_this_cpu_unlock(krcp, flags);
3427
3428 /*
3429 * Inline kvfree() after synchronize_rcu(). We can do
3430 * it from might_sleep() context only, so the current
3431 * CPU can pass the QS state.
3432 */
3433 if (!success) {
3434 debug_rcu_head_unqueue((struct rcu_head *) ptr);
3435 synchronize_rcu();
3436 kvfree(ptr);
3437 }
3438}
3439EXPORT_SYMBOL_GPL(kvfree_call_rcu);
3440
3441static unsigned long
3442kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
3443{
3444 int cpu;
3445 unsigned long count = 0;
3446
3447 /* Snapshot count of all CPUs */
3448 for_each_online_cpu(cpu) {
3449 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3450
3451 count += READ_ONCE(krcp->count);
3452 }
3453
3454 return count;
3455}
3456
3457static unsigned long
3458kfree_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
3459{
3460 int cpu, freed = 0;
3461 unsigned long flags;
3462
3463 for_each_online_cpu(cpu) {
3464 int count;
3465 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3466
3467 count = krcp->count;
3468 raw_spin_lock_irqsave(&krcp->lock, flags);
3469 if (krcp->monitor_todo)
3470 kfree_rcu_drain_unlock(krcp, flags);
3471 else
3472 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3473
3474 sc->nr_to_scan -= count;
3475 freed += count;
3476
3477 if (sc->nr_to_scan <= 0)
3478 break;
3479 }
3480
3481 return freed == 0 ? SHRINK_STOP : freed;
3482}
3483
3484static struct shrinker kfree_rcu_shrinker = {
3485 .count_objects = kfree_rcu_shrink_count,
3486 .scan_objects = kfree_rcu_shrink_scan,
3487 .batch = 0,
3488 .seeks = DEFAULT_SEEKS,
3489};
3490
3491void __init kfree_rcu_scheduler_running(void)
3492{
3493 int cpu;
3494 unsigned long flags;
3495
3496 for_each_online_cpu(cpu) {
3497 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3498
3499 raw_spin_lock_irqsave(&krcp->lock, flags);
3500 if (!krcp->head || krcp->monitor_todo) {
3501 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3502 continue;
3503 }
3504 krcp->monitor_todo = true;
3505 schedule_delayed_work_on(cpu, &krcp->monitor_work,
3506 KFREE_DRAIN_JIFFIES);
3507 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3508 }
3509}
3510
3511/*
3512 * During early boot, any blocking grace-period wait automatically
3513 * implies a grace period. Later on, this is never the case for PREEMPTION.
3514 *
3515 * Howevr, because a context switch is a grace period for !PREEMPTION, any
3516 * blocking grace-period wait automatically implies a grace period if
3517 * there is only one CPU online at any point time during execution of
3518 * either synchronize_rcu() or synchronize_rcu_expedited(). It is OK to
3519 * occasionally incorrectly indicate that there are multiple CPUs online
3520 * when there was in fact only one the whole time, as this just adds some
3521 * overhead: RCU still operates correctly.
3522 */
3523static int rcu_blocking_is_gp(void)
3524{
3525 int ret;
3526
3527 if (IS_ENABLED(CONFIG_PREEMPTION))
3528 return rcu_scheduler_active == RCU_SCHEDULER_INACTIVE;
3529 might_sleep(); /* Check for RCU read-side critical section. */
3530 preempt_disable();
3531 ret = num_online_cpus() <= 1;
3532 preempt_enable();
3533 return ret;
3534}
3535
3536/**
3537 * synchronize_rcu - wait until a grace period has elapsed.
3538 *
3539 * Control will return to the caller some time after a full grace
3540 * period has elapsed, in other words after all currently executing RCU
3541 * read-side critical sections have completed. Note, however, that
3542 * upon return from synchronize_rcu(), the caller might well be executing
3543 * concurrently with new RCU read-side critical sections that began while
3544 * synchronize_rcu() was waiting. RCU read-side critical sections are
3545 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
3546 * In addition, regions of code across which interrupts, preemption, or
3547 * softirqs have been disabled also serve as RCU read-side critical
3548 * sections. This includes hardware interrupt handlers, softirq handlers,
3549 * and NMI handlers.
3550 *
3551 * Note that this guarantee implies further memory-ordering guarantees.
3552 * On systems with more than one CPU, when synchronize_rcu() returns,
3553 * each CPU is guaranteed to have executed a full memory barrier since
3554 * the end of its last RCU read-side critical section whose beginning
3555 * preceded the call to synchronize_rcu(). In addition, each CPU having
3556 * an RCU read-side critical section that extends beyond the return from
3557 * synchronize_rcu() is guaranteed to have executed a full memory barrier
3558 * after the beginning of synchronize_rcu() and before the beginning of
3559 * that RCU read-side critical section. Note that these guarantees include
3560 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3561 * that are executing in the kernel.
3562 *
3563 * Furthermore, if CPU A invoked synchronize_rcu(), which returned
3564 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3565 * to have executed a full memory barrier during the execution of
3566 * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
3567 * again only if the system has more than one CPU).
3568 */
3569void synchronize_rcu(void)
3570{
3571 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3572 lock_is_held(&rcu_lock_map) ||
3573 lock_is_held(&rcu_sched_lock_map),
3574 "Illegal synchronize_rcu() in RCU read-side critical section");
3575 if (rcu_blocking_is_gp())
3576 return;
3577 if (rcu_gp_is_expedited())
3578 synchronize_rcu_expedited();
3579 else
3580 wait_rcu_gp(call_rcu);
3581}
3582EXPORT_SYMBOL_GPL(synchronize_rcu);
3583
3584/**
3585 * get_state_synchronize_rcu - Snapshot current RCU state
3586 *
3587 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3588 * to determine whether or not a full grace period has elapsed in the
3589 * meantime.
3590 */
3591unsigned long get_state_synchronize_rcu(void)
3592{
3593 /*
3594 * Any prior manipulation of RCU-protected data must happen
3595 * before the load from ->gp_seq.
3596 */
3597 smp_mb(); /* ^^^ */
3598 return rcu_seq_snap(&rcu_state.gp_seq);
3599}
3600EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3601
3602/**
3603 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3604 *
3605 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3606 *
3607 * If a full RCU grace period has elapsed since the earlier call to
3608 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3609 * synchronize_rcu() to wait for a full grace period.
3610 *
3611 * Yes, this function does not take counter wrap into account. But
3612 * counter wrap is harmless. If the counter wraps, we have waited for
3613 * more than 2 billion grace periods (and way more on a 64-bit system!),
3614 * so waiting for one additional grace period should be just fine.
3615 */
3616void cond_synchronize_rcu(unsigned long oldstate)
3617{
3618 if (!rcu_seq_done(&rcu_state.gp_seq, oldstate))
3619 synchronize_rcu();
3620 else
3621 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3622}
3623EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3624
3625/*
3626 * Check to see if there is any immediate RCU-related work to be done by
3627 * the current CPU, returning 1 if so and zero otherwise. The checks are
3628 * in order of increasing expense: checks that can be carried out against
3629 * CPU-local state are performed first. However, we must check for CPU
3630 * stalls first, else we might not get a chance.
3631 */
3632static int rcu_pending(int user)
3633{
3634 bool gp_in_progress;
3635 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
3636 struct rcu_node *rnp = rdp->mynode;
3637
3638 /* Check for CPU stalls, if enabled. */
3639 check_cpu_stall(rdp);
3640
3641 /* Does this CPU need a deferred NOCB wakeup? */
3642 if (rcu_nocb_need_deferred_wakeup(rdp))
3643 return 1;
3644
3645 /* Is this a nohz_full CPU in userspace or idle? (Ignore RCU if so.) */
3646 if ((user || rcu_is_cpu_rrupt_from_idle()) && rcu_nohz_full_cpu())
3647 return 0;
3648
3649 /* Is the RCU core waiting for a quiescent state from this CPU? */
3650 gp_in_progress = rcu_gp_in_progress();
3651 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm && gp_in_progress)
3652 return 1;
3653
3654 /* Does this CPU have callbacks ready to invoke? */
3655 if (rcu_segcblist_ready_cbs(&rdp->cblist))
3656 return 1;
3657
3658 /* Has RCU gone idle with this CPU needing another grace period? */
3659 if (!gp_in_progress && rcu_segcblist_is_enabled(&rdp->cblist) &&
3660 (!IS_ENABLED(CONFIG_RCU_NOCB_CPU) ||
3661 !rcu_segcblist_is_offloaded(&rdp->cblist)) &&
3662 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
3663 return 1;
3664
3665 /* Have RCU grace period completed or started? */
3666 if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
3667 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
3668 return 1;
3669
3670 /* nothing to do */
3671 return 0;
3672}
3673
3674/*
3675 * Helper function for rcu_barrier() tracing. If tracing is disabled,
3676 * the compiler is expected to optimize this away.
3677 */
3678static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
3679{
3680 trace_rcu_barrier(rcu_state.name, s, cpu,
3681 atomic_read(&rcu_state.barrier_cpu_count), done);
3682}
3683
3684/*
3685 * RCU callback function for rcu_barrier(). If we are last, wake
3686 * up the task executing rcu_barrier().
3687 *
3688 * Note that the value of rcu_state.barrier_sequence must be captured
3689 * before the atomic_dec_and_test(). Otherwise, if this CPU is not last,
3690 * other CPUs might count the value down to zero before this CPU gets
3691 * around to invoking rcu_barrier_trace(), which might result in bogus
3692 * data from the next instance of rcu_barrier().
3693 */
3694static void rcu_barrier_callback(struct rcu_head *rhp)
3695{
3696 unsigned long __maybe_unused s = rcu_state.barrier_sequence;
3697
3698 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
3699 rcu_barrier_trace(TPS("LastCB"), -1, s);
3700 complete(&rcu_state.barrier_completion);
3701 } else {
3702 rcu_barrier_trace(TPS("CB"), -1, s);
3703 }
3704}
3705
3706/*
3707 * Called with preemption disabled, and from cross-cpu IRQ context.
3708 */
3709static void rcu_barrier_func(void *cpu_in)
3710{
3711 uintptr_t cpu = (uintptr_t)cpu_in;
3712 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3713
3714 rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
3715 rdp->barrier_head.func = rcu_barrier_callback;
3716 debug_rcu_head_queue(&rdp->barrier_head);
3717 rcu_nocb_lock(rdp);
3718 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
3719 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head)) {
3720 atomic_inc(&rcu_state.barrier_cpu_count);
3721 } else {
3722 debug_rcu_head_unqueue(&rdp->barrier_head);
3723 rcu_barrier_trace(TPS("IRQNQ"), -1,
3724 rcu_state.barrier_sequence);
3725 }
3726 rcu_nocb_unlock(rdp);
3727}
3728
3729/**
3730 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
3731 *
3732 * Note that this primitive does not necessarily wait for an RCU grace period
3733 * to complete. For example, if there are no RCU callbacks queued anywhere
3734 * in the system, then rcu_barrier() is within its rights to return
3735 * immediately, without waiting for anything, much less an RCU grace period.
3736 */
3737void rcu_barrier(void)
3738{
3739 uintptr_t cpu;
3740 struct rcu_data *rdp;
3741 unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
3742
3743 rcu_barrier_trace(TPS("Begin"), -1, s);
3744
3745 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3746 mutex_lock(&rcu_state.barrier_mutex);
3747
3748 /* Did someone else do our work for us? */
3749 if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
3750 rcu_barrier_trace(TPS("EarlyExit"), -1,
3751 rcu_state.barrier_sequence);
3752 smp_mb(); /* caller's subsequent code after above check. */
3753 mutex_unlock(&rcu_state.barrier_mutex);
3754 return;
3755 }
3756
3757 /* Mark the start of the barrier operation. */
3758 rcu_seq_start(&rcu_state.barrier_sequence);
3759 rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
3760
3761 /*
3762 * Initialize the count to two rather than to zero in order
3763 * to avoid a too-soon return to zero in case of an immediate
3764 * invocation of the just-enqueued callback (or preemption of
3765 * this task). Exclude CPU-hotplug operations to ensure that no
3766 * offline non-offloaded CPU has callbacks queued.
3767 */
3768 init_completion(&rcu_state.barrier_completion);
3769 atomic_set(&rcu_state.barrier_cpu_count, 2);
3770 get_online_cpus();
3771
3772 /*
3773 * Force each CPU with callbacks to register a new callback.
3774 * When that callback is invoked, we will know that all of the
3775 * corresponding CPU's preceding callbacks have been invoked.
3776 */
3777 for_each_possible_cpu(cpu) {
3778 rdp = per_cpu_ptr(&rcu_data, cpu);
3779 if (cpu_is_offline(cpu) &&
3780 !rcu_segcblist_is_offloaded(&rdp->cblist))
3781 continue;
3782 if (rcu_segcblist_n_cbs(&rdp->cblist) && cpu_online(cpu)) {
3783 rcu_barrier_trace(TPS("OnlineQ"), cpu,
3784 rcu_state.barrier_sequence);
3785 smp_call_function_single(cpu, rcu_barrier_func, (void *)cpu, 1);
3786 } else if (rcu_segcblist_n_cbs(&rdp->cblist) &&
3787 cpu_is_offline(cpu)) {
3788 rcu_barrier_trace(TPS("OfflineNoCBQ"), cpu,
3789 rcu_state.barrier_sequence);
3790 local_irq_disable();
3791 rcu_barrier_func((void *)cpu);
3792 local_irq_enable();
3793 } else if (cpu_is_offline(cpu)) {
3794 rcu_barrier_trace(TPS("OfflineNoCBNoQ"), cpu,
3795 rcu_state.barrier_sequence);
3796 } else {
3797 rcu_barrier_trace(TPS("OnlineNQ"), cpu,
3798 rcu_state.barrier_sequence);
3799 }
3800 }
3801 put_online_cpus();
3802
3803 /*
3804 * Now that we have an rcu_barrier_callback() callback on each
3805 * CPU, and thus each counted, remove the initial count.
3806 */
3807 if (atomic_sub_and_test(2, &rcu_state.barrier_cpu_count))
3808 complete(&rcu_state.barrier_completion);
3809
3810 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3811 wait_for_completion(&rcu_state.barrier_completion);
3812
3813 /* Mark the end of the barrier operation. */
3814 rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
3815 rcu_seq_end(&rcu_state.barrier_sequence);
3816
3817 /* Other rcu_barrier() invocations can now safely proceed. */
3818 mutex_unlock(&rcu_state.barrier_mutex);
3819}
3820EXPORT_SYMBOL_GPL(rcu_barrier);
3821
3822/*
3823 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3824 * first CPU in a given leaf rcu_node structure coming online. The caller
3825 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3826 * disabled.
3827 */
3828static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3829{
3830 long mask;
3831 long oldmask;
3832 struct rcu_node *rnp = rnp_leaf;
3833
3834 raw_lockdep_assert_held_rcu_node(rnp_leaf);
3835 WARN_ON_ONCE(rnp->wait_blkd_tasks);
3836 for (;;) {
3837 mask = rnp->grpmask;
3838 rnp = rnp->parent;
3839 if (rnp == NULL)
3840 return;
3841 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3842 oldmask = rnp->qsmaskinit;
3843 rnp->qsmaskinit |= mask;
3844 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3845 if (oldmask)
3846 return;
3847 }
3848}
3849
3850/*
3851 * Do boot-time initialization of a CPU's per-CPU RCU data.
3852 */
3853static void __init
3854rcu_boot_init_percpu_data(int cpu)
3855{
3856 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3857
3858 /* Set up local state, ensuring consistent view of global state. */
3859 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3860 WARN_ON_ONCE(rdp->dynticks_nesting != 1);
3861 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp)));
3862 rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
3863 rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
3864 rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
3865 rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
3866 rdp->cpu = cpu;
3867 rcu_boot_init_nocb_percpu_data(rdp);
3868}
3869
3870/*
3871 * Invoked early in the CPU-online process, when pretty much all services
3872 * are available. The incoming CPU is not present.
3873 *
3874 * Initializes a CPU's per-CPU RCU data. Note that only one online or
3875 * offline event can be happening at a given time. Note also that we can
3876 * accept some slop in the rsp->gp_seq access due to the fact that this
3877 * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
3878 * And any offloaded callbacks are being numbered elsewhere.
3879 */
3880int rcutree_prepare_cpu(unsigned int cpu)
3881{
3882 unsigned long flags;
3883 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3884 struct rcu_node *rnp = rcu_get_root();
3885
3886 /* Set up local state, ensuring consistent view of global state. */
3887 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3888 rdp->qlen_last_fqs_check = 0;
3889 rdp->n_force_qs_snap = rcu_state.n_force_qs;
3890 rdp->blimit = blimit;
3891 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3892 !rcu_segcblist_is_offloaded(&rdp->cblist))
3893 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
3894 rdp->dynticks_nesting = 1; /* CPU not up, no tearing. */
3895 rcu_dynticks_eqs_online();
3896 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3897
3898 /*
3899 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3900 * propagation up the rcu_node tree will happen at the beginning
3901 * of the next grace period.
3902 */
3903 rnp = rdp->mynode;
3904 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3905 rdp->beenonline = true; /* We have now been online. */
3906 rdp->gp_seq = READ_ONCE(rnp->gp_seq);
3907 rdp->gp_seq_needed = rdp->gp_seq;
3908 rdp->cpu_no_qs.b.norm = true;
3909 rdp->core_needs_qs = false;
3910 rdp->rcu_iw_pending = false;
3911 rdp->rcu_iw_gp_seq = rdp->gp_seq - 1;
3912 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
3913 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3914 rcu_prepare_kthreads(cpu);
3915 rcu_spawn_cpu_nocb_kthread(cpu);
3916
3917 return 0;
3918}
3919
3920/*
3921 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3922 */
3923static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3924{
3925 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3926
3927 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3928}
3929
3930/*
3931 * Near the end of the CPU-online process. Pretty much all services
3932 * enabled, and the CPU is now very much alive.
3933 */
3934int rcutree_online_cpu(unsigned int cpu)
3935{
3936 unsigned long flags;
3937 struct rcu_data *rdp;
3938 struct rcu_node *rnp;
3939
3940 rdp = per_cpu_ptr(&rcu_data, cpu);
3941 rnp = rdp->mynode;
3942 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3943 rnp->ffmask |= rdp->grpmask;
3944 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3945 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
3946 return 0; /* Too early in boot for scheduler work. */
3947 sync_sched_exp_online_cleanup(cpu);
3948 rcutree_affinity_setting(cpu, -1);
3949
3950 // Stop-machine done, so allow nohz_full to disable tick.
3951 tick_dep_clear(TICK_DEP_BIT_RCU);
3952 return 0;
3953}
3954
3955/*
3956 * Near the beginning of the process. The CPU is still very much alive
3957 * with pretty much all services enabled.
3958 */
3959int rcutree_offline_cpu(unsigned int cpu)
3960{
3961 unsigned long flags;
3962 struct rcu_data *rdp;
3963 struct rcu_node *rnp;
3964
3965 rdp = per_cpu_ptr(&rcu_data, cpu);
3966 rnp = rdp->mynode;
3967 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3968 rnp->ffmask &= ~rdp->grpmask;
3969 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3970
3971 rcutree_affinity_setting(cpu, cpu);
3972
3973 // nohz_full CPUs need the tick for stop-machine to work quickly
3974 tick_dep_set(TICK_DEP_BIT_RCU);
3975 return 0;
3976}
3977
3978static DEFINE_PER_CPU(int, rcu_cpu_started);
3979
3980/*
3981 * Mark the specified CPU as being online so that subsequent grace periods
3982 * (both expedited and normal) will wait on it. Note that this means that
3983 * incoming CPUs are not allowed to use RCU read-side critical sections
3984 * until this function is called. Failing to observe this restriction
3985 * will result in lockdep splats.
3986 *
3987 * Note that this function is special in that it is invoked directly
3988 * from the incoming CPU rather than from the cpuhp_step mechanism.
3989 * This is because this function must be invoked at a precise location.
3990 */
3991void rcu_cpu_starting(unsigned int cpu)
3992{
3993 unsigned long flags;
3994 unsigned long mask;
3995 struct rcu_data *rdp;
3996 struct rcu_node *rnp;
3997 bool newcpu;
3998
3999 if (per_cpu(rcu_cpu_started, cpu))
4000 return;
4001
4002 per_cpu(rcu_cpu_started, cpu) = 1;
4003
4004 rdp = per_cpu_ptr(&rcu_data, cpu);
4005 rnp = rdp->mynode;
4006 mask = rdp->grpmask;
4007 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4008 WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask);
4009 newcpu = !(rnp->expmaskinitnext & mask);
4010 rnp->expmaskinitnext |= mask;
4011 /* Allow lockless access for expedited grace periods. */
4012 smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + newcpu); /* ^^^ */
4013 ASSERT_EXCLUSIVE_WRITER(rcu_state.ncpus);
4014 rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
4015 rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4016 rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4017 if (rnp->qsmask & mask) { /* RCU waiting on incoming CPU? */
4018 rcu_disable_urgency_upon_qs(rdp);
4019 /* Report QS -after- changing ->qsmaskinitnext! */
4020 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4021 } else {
4022 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4023 }
4024 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
4025}
4026
4027#ifdef CONFIG_HOTPLUG_CPU
4028/*
4029 * The outgoing function has no further need of RCU, so remove it from
4030 * the rcu_node tree's ->qsmaskinitnext bit masks.
4031 *
4032 * Note that this function is special in that it is invoked directly
4033 * from the outgoing CPU rather than from the cpuhp_step mechanism.
4034 * This is because this function must be invoked at a precise location.
4035 */
4036void rcu_report_dead(unsigned int cpu)
4037{
4038 unsigned long flags;
4039 unsigned long mask;
4040 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4041 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
4042
4043 /* QS for any half-done expedited grace period. */
4044 preempt_disable();
4045 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
4046 preempt_enable();
4047 rcu_preempt_deferred_qs(current);
4048
4049 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
4050 mask = rdp->grpmask;
4051 raw_spin_lock(&rcu_state.ofl_lock);
4052 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
4053 rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4054 rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4055 if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
4056 /* Report quiescent state -before- changing ->qsmaskinitnext! */
4057 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4058 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4059 }
4060 WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask);
4061 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4062 raw_spin_unlock(&rcu_state.ofl_lock);
4063
4064 per_cpu(rcu_cpu_started, cpu) = 0;
4065}
4066
4067/*
4068 * The outgoing CPU has just passed through the dying-idle state, and we
4069 * are being invoked from the CPU that was IPIed to continue the offline
4070 * operation. Migrate the outgoing CPU's callbacks to the current CPU.
4071 */
4072void rcutree_migrate_callbacks(int cpu)
4073{
4074 unsigned long flags;
4075 struct rcu_data *my_rdp;
4076 struct rcu_node *my_rnp;
4077 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4078 bool needwake;
4079
4080 if (rcu_segcblist_is_offloaded(&rdp->cblist) ||
4081 rcu_segcblist_empty(&rdp->cblist))
4082 return; /* No callbacks to migrate. */
4083
4084 local_irq_save(flags);
4085 my_rdp = this_cpu_ptr(&rcu_data);
4086 my_rnp = my_rdp->mynode;
4087 rcu_nocb_lock(my_rdp); /* irqs already disabled. */
4088 WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies));
4089 raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
4090 /* Leverage recent GPs and set GP for new callbacks. */
4091 needwake = rcu_advance_cbs(my_rnp, rdp) ||
4092 rcu_advance_cbs(my_rnp, my_rdp);
4093 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
4094 needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
4095 rcu_segcblist_disable(&rdp->cblist);
4096 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
4097 !rcu_segcblist_n_cbs(&my_rdp->cblist));
4098 if (rcu_segcblist_is_offloaded(&my_rdp->cblist)) {
4099 raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
4100 __call_rcu_nocb_wake(my_rdp, true, flags);
4101 } else {
4102 rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
4103 raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags);
4104 }
4105 if (needwake)
4106 rcu_gp_kthread_wake();
4107 lockdep_assert_irqs_enabled();
4108 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
4109 !rcu_segcblist_empty(&rdp->cblist),
4110 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
4111 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
4112 rcu_segcblist_first_cb(&rdp->cblist));
4113}
4114#endif
4115
4116/*
4117 * On non-huge systems, use expedited RCU grace periods to make suspend
4118 * and hibernation run faster.
4119 */
4120static int rcu_pm_notify(struct notifier_block *self,
4121 unsigned long action, void *hcpu)
4122{
4123 switch (action) {
4124 case PM_HIBERNATION_PREPARE:
4125 case PM_SUSPEND_PREPARE:
4126 rcu_expedite_gp();
4127 break;
4128 case PM_POST_HIBERNATION:
4129 case PM_POST_SUSPEND:
4130 rcu_unexpedite_gp();
4131 break;
4132 default:
4133 break;
4134 }
4135 return NOTIFY_OK;
4136}
4137
4138/*
4139 * Spawn the kthreads that handle RCU's grace periods.
4140 */
4141static int __init rcu_spawn_gp_kthread(void)
4142{
4143 unsigned long flags;
4144 int kthread_prio_in = kthread_prio;
4145 struct rcu_node *rnp;
4146 struct sched_param sp;
4147 struct task_struct *t;
4148
4149 /* Force priority into range. */
4150 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
4151 && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
4152 kthread_prio = 2;
4153 else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4154 kthread_prio = 1;
4155 else if (kthread_prio < 0)
4156 kthread_prio = 0;
4157 else if (kthread_prio > 99)
4158 kthread_prio = 99;
4159
4160 if (kthread_prio != kthread_prio_in)
4161 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4162 kthread_prio, kthread_prio_in);
4163
4164 rcu_scheduler_fully_active = 1;
4165 t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
4166 if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
4167 return 0;
4168 if (kthread_prio) {
4169 sp.sched_priority = kthread_prio;
4170 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4171 }
4172 rnp = rcu_get_root();
4173 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4174 WRITE_ONCE(rcu_state.gp_activity, jiffies);
4175 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
4176 // Reset .gp_activity and .gp_req_activity before setting .gp_kthread.
4177 smp_store_release(&rcu_state.gp_kthread, t); /* ^^^ */
4178 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4179 wake_up_process(t);
4180 rcu_spawn_nocb_kthreads();
4181 rcu_spawn_boost_kthreads();
4182 return 0;
4183}
4184early_initcall(rcu_spawn_gp_kthread);
4185
4186/*
4187 * This function is invoked towards the end of the scheduler's
4188 * initialization process. Before this is called, the idle task might
4189 * contain synchronous grace-period primitives (during which time, this idle
4190 * task is booting the system, and such primitives are no-ops). After this
4191 * function is called, any synchronous grace-period primitives are run as
4192 * expedited, with the requesting task driving the grace period forward.
4193 * A later core_initcall() rcu_set_runtime_mode() will switch to full
4194 * runtime RCU functionality.
4195 */
4196void rcu_scheduler_starting(void)
4197{
4198 WARN_ON(num_online_cpus() != 1);
4199 WARN_ON(nr_context_switches() > 0);
4200 rcu_test_sync_prims();
4201 rcu_scheduler_active = RCU_SCHEDULER_INIT;
4202 rcu_test_sync_prims();
4203}
4204
4205/*
4206 * Helper function for rcu_init() that initializes the rcu_state structure.
4207 */
4208static void __init rcu_init_one(void)
4209{
4210 static const char * const buf[] = RCU_NODE_NAME_INIT;
4211 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4212 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4213 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4214
4215 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4216 int cpustride = 1;
4217 int i;
4218 int j;
4219 struct rcu_node *rnp;
4220
4221 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4222
4223 /* Silence gcc 4.8 false positive about array index out of range. */
4224 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4225 panic("rcu_init_one: rcu_num_lvls out of range");
4226
4227 /* Initialize the level-tracking arrays. */
4228
4229 for (i = 1; i < rcu_num_lvls; i++)
4230 rcu_state.level[i] =
4231 rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
4232 rcu_init_levelspread(levelspread, num_rcu_lvl);
4233
4234 /* Initialize the elements themselves, starting from the leaves. */
4235
4236 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4237 cpustride *= levelspread[i];
4238 rnp = rcu_state.level[i];
4239 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
4240 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4241 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4242 &rcu_node_class[i], buf[i]);
4243 raw_spin_lock_init(&rnp->fqslock);
4244 lockdep_set_class_and_name(&rnp->fqslock,
4245 &rcu_fqs_class[i], fqs[i]);
4246 rnp->gp_seq = rcu_state.gp_seq;
4247 rnp->gp_seq_needed = rcu_state.gp_seq;
4248 rnp->completedqs = rcu_state.gp_seq;
4249 rnp->qsmask = 0;
4250 rnp->qsmaskinit = 0;
4251 rnp->grplo = j * cpustride;
4252 rnp->grphi = (j + 1) * cpustride - 1;
4253 if (rnp->grphi >= nr_cpu_ids)
4254 rnp->grphi = nr_cpu_ids - 1;
4255 if (i == 0) {
4256 rnp->grpnum = 0;
4257 rnp->grpmask = 0;
4258 rnp->parent = NULL;
4259 } else {
4260 rnp->grpnum = j % levelspread[i - 1];
4261 rnp->grpmask = BIT(rnp->grpnum);
4262 rnp->parent = rcu_state.level[i - 1] +
4263 j / levelspread[i - 1];
4264 }
4265 rnp->level = i;
4266 INIT_LIST_HEAD(&rnp->blkd_tasks);
4267 rcu_init_one_nocb(rnp);
4268 init_waitqueue_head(&rnp->exp_wq[0]);
4269 init_waitqueue_head(&rnp->exp_wq[1]);
4270 init_waitqueue_head(&rnp->exp_wq[2]);
4271 init_waitqueue_head(&rnp->exp_wq[3]);
4272 spin_lock_init(&rnp->exp_lock);
4273 }
4274 }
4275
4276 init_swait_queue_head(&rcu_state.gp_wq);
4277 init_swait_queue_head(&rcu_state.expedited_wq);
4278 rnp = rcu_first_leaf_node();
4279 for_each_possible_cpu(i) {
4280 while (i > rnp->grphi)
4281 rnp++;
4282 per_cpu_ptr(&rcu_data, i)->mynode = rnp;
4283 rcu_boot_init_percpu_data(i);
4284 }
4285}
4286
4287/*
4288 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4289 * replace the definitions in tree.h because those are needed to size
4290 * the ->node array in the rcu_state structure.
4291 */
4292static void __init rcu_init_geometry(void)
4293{
4294 ulong d;
4295 int i;
4296 int rcu_capacity[RCU_NUM_LVLS];
4297
4298 /*
4299 * Initialize any unspecified boot parameters.
4300 * The default values of jiffies_till_first_fqs and
4301 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4302 * value, which is a function of HZ, then adding one for each
4303 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4304 */
4305 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4306 if (jiffies_till_first_fqs == ULONG_MAX)
4307 jiffies_till_first_fqs = d;
4308 if (jiffies_till_next_fqs == ULONG_MAX)
4309 jiffies_till_next_fqs = d;
4310 adjust_jiffies_till_sched_qs();
4311
4312 /* If the compile-time values are accurate, just leave. */
4313 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4314 nr_cpu_ids == NR_CPUS)
4315 return;
4316 pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4317 rcu_fanout_leaf, nr_cpu_ids);
4318
4319 /*
4320 * The boot-time rcu_fanout_leaf parameter must be at least two
4321 * and cannot exceed the number of bits in the rcu_node masks.
4322 * Complain and fall back to the compile-time values if this
4323 * limit is exceeded.
4324 */
4325 if (rcu_fanout_leaf < 2 ||
4326 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4327 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4328 WARN_ON(1);
4329 return;
4330 }
4331
4332 /*
4333 * Compute number of nodes that can be handled an rcu_node tree
4334 * with the given number of levels.
4335 */
4336 rcu_capacity[0] = rcu_fanout_leaf;
4337 for (i = 1; i < RCU_NUM_LVLS; i++)
4338 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4339
4340 /*
4341 * The tree must be able to accommodate the configured number of CPUs.
4342 * If this limit is exceeded, fall back to the compile-time values.
4343 */
4344 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4345 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4346 WARN_ON(1);
4347 return;
4348 }
4349
4350 /* Calculate the number of levels in the tree. */
4351 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4352 }
4353 rcu_num_lvls = i + 1;
4354
4355 /* Calculate the number of rcu_nodes at each level of the tree. */
4356 for (i = 0; i < rcu_num_lvls; i++) {
4357 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4358 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4359 }
4360
4361 /* Calculate the total number of rcu_node structures. */
4362 rcu_num_nodes = 0;
4363 for (i = 0; i < rcu_num_lvls; i++)
4364 rcu_num_nodes += num_rcu_lvl[i];
4365}
4366
4367/*
4368 * Dump out the structure of the rcu_node combining tree associated
4369 * with the rcu_state structure.
4370 */
4371static void __init rcu_dump_rcu_node_tree(void)
4372{
4373 int level = 0;
4374 struct rcu_node *rnp;
4375
4376 pr_info("rcu_node tree layout dump\n");
4377 pr_info(" ");
4378 rcu_for_each_node_breadth_first(rnp) {
4379 if (rnp->level != level) {
4380 pr_cont("\n");
4381 pr_info(" ");
4382 level = rnp->level;
4383 }
4384 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4385 }
4386 pr_cont("\n");
4387}
4388
4389struct workqueue_struct *rcu_gp_wq;
4390struct workqueue_struct *rcu_par_gp_wq;
4391
4392static void __init kfree_rcu_batch_init(void)
4393{
4394 int cpu;
4395 int i;
4396
4397 for_each_possible_cpu(cpu) {
4398 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
4399 struct kvfree_rcu_bulk_data *bnode;
4400
4401 for (i = 0; i < KFREE_N_BATCHES; i++) {
4402 INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
4403 krcp->krw_arr[i].krcp = krcp;
4404 }
4405
4406 for (i = 0; i < rcu_min_cached_objs; i++) {
4407 bnode = (struct kvfree_rcu_bulk_data *)
4408 __get_free_page(GFP_NOWAIT | __GFP_NOWARN);
4409
4410 if (bnode)
4411 put_cached_bnode(krcp, bnode);
4412 else
4413 pr_err("Failed to preallocate for %d CPU!\n", cpu);
4414 }
4415
4416 INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
4417 krcp->initialized = true;
4418 }
4419 if (register_shrinker(&kfree_rcu_shrinker))
4420 pr_err("Failed to register kfree_rcu() shrinker!\n");
4421}
4422
4423void __init rcu_init(void)
4424{
4425 int cpu;
4426
4427 rcu_early_boot_tests();
4428
4429 kfree_rcu_batch_init();
4430 rcu_bootup_announce();
4431 rcu_init_geometry();
4432 rcu_init_one();
4433 if (dump_tree)
4434 rcu_dump_rcu_node_tree();
4435 if (use_softirq)
4436 open_softirq(RCU_SOFTIRQ, rcu_core_si);
4437
4438 /*
4439 * We don't need protection against CPU-hotplug here because
4440 * this is called early in boot, before either interrupts
4441 * or the scheduler are operational.
4442 */
4443 pm_notifier(rcu_pm_notify, 0);
4444 for_each_online_cpu(cpu) {
4445 rcutree_prepare_cpu(cpu);
4446 rcu_cpu_starting(cpu);
4447 rcutree_online_cpu(cpu);
4448 }
4449
4450 /* Create workqueue for expedited GPs and for Tree SRCU. */
4451 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
4452 WARN_ON(!rcu_gp_wq);
4453 rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
4454 WARN_ON(!rcu_par_gp_wq);
4455 srcu_init();
4456
4457 /* Fill in default value for rcutree.qovld boot parameter. */
4458 /* -After- the rcu_node ->lock fields are initialized! */
4459 if (qovld < 0)
4460 qovld_calc = DEFAULT_RCU_QOVLD_MULT * qhimark;
4461 else
4462 qovld_calc = qovld;
4463}
4464
4465#include "tree_stall.h"
4466#include "tree_exp.h"
4467#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"