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_wait.h>
36#include <linux/interrupt.h>
37#include <linux/sched.h>
38#include <linux/sched/debug.h>
39#include <linux/nmi.h>
40#include <linux/atomic.h>
41#include <linux/bitops.h>
42#include <linux/export.h>
43#include <linux/completion.h>
44#include <linux/moduleparam.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 <uapi/linux/sched/types.h>
54#include <linux/prefetch.h>
55#include <linux/delay.h>
56#include <linux/stop_machine.h>
57#include <linux/random.h>
58#include <linux/trace_events.h>
59#include <linux/suspend.h>
60#include <linux/ftrace.h>
61
62#include "tree.h"
63#include "rcu.h"
64
65#ifdef MODULE_PARAM_PREFIX
66#undef MODULE_PARAM_PREFIX
67#endif
68#define MODULE_PARAM_PREFIX "rcutree."
69
70/* Data structures. */
71
72/*
73 * In order to export the rcu_state name to the tracing tools, it
74 * needs to be added in the __tracepoint_string section.
75 * This requires defining a separate variable tp_<sname>_varname
76 * that points to the string being used, and this will allow
77 * the tracing userspace tools to be able to decipher the string
78 * address to the matching string.
79 */
80#ifdef CONFIG_TRACING
81# define DEFINE_RCU_TPS(sname) \
82static char sname##_varname[] = #sname; \
83static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
84# define RCU_STATE_NAME(sname) sname##_varname
85#else
86# define DEFINE_RCU_TPS(sname)
87# define RCU_STATE_NAME(sname) __stringify(sname)
88#endif
89
90#define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
91DEFINE_RCU_TPS(sname) \
92static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
93struct rcu_state sname##_state = { \
94 .level = { &sname##_state.node[0] }, \
95 .rda = &sname##_data, \
96 .call = cr, \
97 .gp_state = RCU_GP_IDLE, \
98 .gpnum = 0UL - 300UL, \
99 .completed = 0UL - 300UL, \
100 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
101 .name = RCU_STATE_NAME(sname), \
102 .abbr = sabbr, \
103 .exp_mutex = __MUTEX_INITIALIZER(sname##_state.exp_mutex), \
104 .exp_wake_mutex = __MUTEX_INITIALIZER(sname##_state.exp_wake_mutex), \
105}
106
107RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
108RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
109
110static struct rcu_state *const rcu_state_p;
111LIST_HEAD(rcu_struct_flavors);
112
113/* Dump rcu_node combining tree at boot to verify correct setup. */
114static bool dump_tree;
115module_param(dump_tree, bool, 0444);
116/* Control rcu_node-tree auto-balancing at boot time. */
117static bool rcu_fanout_exact;
118module_param(rcu_fanout_exact, bool, 0444);
119/* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
120static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
121module_param(rcu_fanout_leaf, int, 0444);
122int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
123/* Number of rcu_nodes at specified level. */
124int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
125int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
126/* panic() on RCU Stall sysctl. */
127int sysctl_panic_on_rcu_stall __read_mostly;
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_init_new_rnp(struct rcu_node *rnp_leaf);
159static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
160static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
161static void invoke_rcu_core(void);
162static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
163static void rcu_report_exp_rdp(struct rcu_state *rsp,
164 struct rcu_data *rdp, bool wake);
165static void sync_sched_exp_online_cleanup(int cpu);
166
167/* rcuc/rcub kthread realtime priority */
168static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
169module_param(kthread_prio, int, 0644);
170
171/* Delay in jiffies for grace-period initialization delays, debug only. */
172
173static int gp_preinit_delay;
174module_param(gp_preinit_delay, int, 0444);
175static int gp_init_delay;
176module_param(gp_init_delay, int, 0444);
177static int gp_cleanup_delay;
178module_param(gp_cleanup_delay, int, 0444);
179
180/*
181 * Number of grace periods between delays, normalized by the duration of
182 * the delay. The longer the delay, the more the grace periods between
183 * each delay. The reason for this normalization is that it means that,
184 * for non-zero delays, the overall slowdown of grace periods is constant
185 * regardless of the duration of the delay. This arrangement balances
186 * the need for long delays to increase some race probabilities with the
187 * need for fast grace periods to increase other race probabilities.
188 */
189#define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
190
191/*
192 * Track the rcutorture test sequence number and the update version
193 * number within a given test. The rcutorture_testseq is incremented
194 * on every rcutorture module load and unload, so has an odd value
195 * when a test is running. The rcutorture_vernum is set to zero
196 * when rcutorture starts and is incremented on each rcutorture update.
197 * These variables enable correlating rcutorture output with the
198 * RCU tracing information.
199 */
200unsigned long rcutorture_testseq;
201unsigned long rcutorture_vernum;
202
203/*
204 * Compute the mask of online CPUs for the specified rcu_node structure.
205 * This will not be stable unless the rcu_node structure's ->lock is
206 * held, but the bit corresponding to the current CPU will be stable
207 * in most contexts.
208 */
209unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
210{
211 return READ_ONCE(rnp->qsmaskinitnext);
212}
213
214/*
215 * Return true if an RCU grace period is in progress. The READ_ONCE()s
216 * permit this function to be invoked without holding the root rcu_node
217 * structure's ->lock, but of course results can be subject to change.
218 */
219static int rcu_gp_in_progress(struct rcu_state *rsp)
220{
221 return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum);
222}
223
224/*
225 * Note a quiescent state. Because we do not need to know
226 * how many quiescent states passed, just if there was at least
227 * one since the start of the grace period, this just sets a flag.
228 * The caller must have disabled preemption.
229 */
230void rcu_sched_qs(void)
231{
232 RCU_LOCKDEP_WARN(preemptible(), "rcu_sched_qs() invoked with preemption enabled!!!");
233 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.s))
234 return;
235 trace_rcu_grace_period(TPS("rcu_sched"),
236 __this_cpu_read(rcu_sched_data.gpnum),
237 TPS("cpuqs"));
238 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
239 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
240 return;
241 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
242 rcu_report_exp_rdp(&rcu_sched_state,
243 this_cpu_ptr(&rcu_sched_data), true);
244}
245
246void rcu_bh_qs(void)
247{
248 RCU_LOCKDEP_WARN(preemptible(), "rcu_bh_qs() invoked with preemption enabled!!!");
249 if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) {
250 trace_rcu_grace_period(TPS("rcu_bh"),
251 __this_cpu_read(rcu_bh_data.gpnum),
252 TPS("cpuqs"));
253 __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false);
254 }
255}
256
257/*
258 * Steal a bit from the bottom of ->dynticks for idle entry/exit
259 * control. Initially this is for TLB flushing.
260 */
261#define RCU_DYNTICK_CTRL_MASK 0x1
262#define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
263#ifndef rcu_eqs_special_exit
264#define rcu_eqs_special_exit() do { } while (0)
265#endif
266
267static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
268 .dynticks_nesting = 1,
269 .dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE,
270 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
271};
272
273/*
274 * Record entry into an extended quiescent state. This is only to be
275 * called when not already in an extended quiescent state.
276 */
277static void rcu_dynticks_eqs_enter(void)
278{
279 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
280 int seq;
281
282 /*
283 * CPUs seeing atomic_add_return() must see prior RCU read-side
284 * critical sections, and we also must force ordering with the
285 * next idle sojourn.
286 */
287 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
288 /* Better be in an extended quiescent state! */
289 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
290 (seq & RCU_DYNTICK_CTRL_CTR));
291 /* Better not have special action (TLB flush) pending! */
292 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
293 (seq & RCU_DYNTICK_CTRL_MASK));
294}
295
296/*
297 * Record exit from an extended quiescent state. This is only to be
298 * called from an extended quiescent state.
299 */
300static void rcu_dynticks_eqs_exit(void)
301{
302 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
303 int seq;
304
305 /*
306 * CPUs seeing atomic_add_return() must see prior idle sojourns,
307 * and we also must force ordering with the next RCU read-side
308 * critical section.
309 */
310 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
311 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
312 !(seq & RCU_DYNTICK_CTRL_CTR));
313 if (seq & RCU_DYNTICK_CTRL_MASK) {
314 atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdtp->dynticks);
315 smp_mb__after_atomic(); /* _exit after clearing mask. */
316 /* Prefer duplicate flushes to losing a flush. */
317 rcu_eqs_special_exit();
318 }
319}
320
321/*
322 * Reset the current CPU's ->dynticks counter to indicate that the
323 * newly onlined CPU is no longer in an extended quiescent state.
324 * This will either leave the counter unchanged, or increment it
325 * to the next non-quiescent value.
326 *
327 * The non-atomic test/increment sequence works because the upper bits
328 * of the ->dynticks counter are manipulated only by the corresponding CPU,
329 * or when the corresponding CPU is offline.
330 */
331static void rcu_dynticks_eqs_online(void)
332{
333 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
334
335 if (atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR)
336 return;
337 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
338}
339
340/*
341 * Is the current CPU in an extended quiescent state?
342 *
343 * No ordering, as we are sampling CPU-local information.
344 */
345bool rcu_dynticks_curr_cpu_in_eqs(void)
346{
347 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
348
349 return !(atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR);
350}
351
352/*
353 * Snapshot the ->dynticks counter with full ordering so as to allow
354 * stable comparison of this counter with past and future snapshots.
355 */
356int rcu_dynticks_snap(struct rcu_dynticks *rdtp)
357{
358 int snap = atomic_add_return(0, &rdtp->dynticks);
359
360 return snap & ~RCU_DYNTICK_CTRL_MASK;
361}
362
363/*
364 * Return true if the snapshot returned from rcu_dynticks_snap()
365 * indicates that RCU is in an extended quiescent state.
366 */
367static bool rcu_dynticks_in_eqs(int snap)
368{
369 return !(snap & RCU_DYNTICK_CTRL_CTR);
370}
371
372/*
373 * Return true if the CPU corresponding to the specified rcu_dynticks
374 * structure has spent some time in an extended quiescent state since
375 * rcu_dynticks_snap() returned the specified snapshot.
376 */
377static bool rcu_dynticks_in_eqs_since(struct rcu_dynticks *rdtp, int snap)
378{
379 return snap != rcu_dynticks_snap(rdtp);
380}
381
382/*
383 * Do a double-increment of the ->dynticks counter to emulate a
384 * momentary idle-CPU quiescent state.
385 */
386static void rcu_dynticks_momentary_idle(void)
387{
388 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
389 int special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
390 &rdtp->dynticks);
391
392 /* It is illegal to call this from idle state. */
393 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
394}
395
396/*
397 * Set the special (bottom) bit of the specified CPU so that it
398 * will take special action (such as flushing its TLB) on the
399 * next exit from an extended quiescent state. Returns true if
400 * the bit was successfully set, or false if the CPU was not in
401 * an extended quiescent state.
402 */
403bool rcu_eqs_special_set(int cpu)
404{
405 int old;
406 int new;
407 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
408
409 do {
410 old = atomic_read(&rdtp->dynticks);
411 if (old & RCU_DYNTICK_CTRL_CTR)
412 return false;
413 new = old | RCU_DYNTICK_CTRL_MASK;
414 } while (atomic_cmpxchg(&rdtp->dynticks, old, new) != old);
415 return true;
416}
417
418/*
419 * Let the RCU core know that this CPU has gone through the scheduler,
420 * which is a quiescent state. This is called when the need for a
421 * quiescent state is urgent, so we burn an atomic operation and full
422 * memory barriers to let the RCU core know about it, regardless of what
423 * this CPU might (or might not) do in the near future.
424 *
425 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
426 *
427 * The caller must have disabled interrupts.
428 */
429static void rcu_momentary_dyntick_idle(void)
430{
431 raw_cpu_write(rcu_dynticks.rcu_need_heavy_qs, false);
432 rcu_dynticks_momentary_idle();
433}
434
435/*
436 * Note a context switch. This is a quiescent state for RCU-sched,
437 * and requires special handling for preemptible RCU.
438 * The caller must have disabled interrupts.
439 */
440void rcu_note_context_switch(bool preempt)
441{
442 barrier(); /* Avoid RCU read-side critical sections leaking down. */
443 trace_rcu_utilization(TPS("Start context switch"));
444 rcu_sched_qs();
445 rcu_preempt_note_context_switch(preempt);
446 /* Load rcu_urgent_qs before other flags. */
447 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs)))
448 goto out;
449 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
450 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs)))
451 rcu_momentary_dyntick_idle();
452 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
453 if (!preempt)
454 rcu_note_voluntary_context_switch_lite(current);
455out:
456 trace_rcu_utilization(TPS("End context switch"));
457 barrier(); /* Avoid RCU read-side critical sections leaking up. */
458}
459EXPORT_SYMBOL_GPL(rcu_note_context_switch);
460
461/*
462 * Register a quiescent state for all RCU flavors. If there is an
463 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
464 * dyntick-idle quiescent state visible to other CPUs (but only for those
465 * RCU flavors in desperate need of a quiescent state, which will normally
466 * be none of them). Either way, do a lightweight quiescent state for
467 * all RCU flavors.
468 *
469 * The barrier() calls are redundant in the common case when this is
470 * called externally, but just in case this is called from within this
471 * file.
472 *
473 */
474void rcu_all_qs(void)
475{
476 unsigned long flags;
477
478 if (!raw_cpu_read(rcu_dynticks.rcu_urgent_qs))
479 return;
480 preempt_disable();
481 /* Load rcu_urgent_qs before other flags. */
482 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs))) {
483 preempt_enable();
484 return;
485 }
486 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
487 barrier(); /* Avoid RCU read-side critical sections leaking down. */
488 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs))) {
489 local_irq_save(flags);
490 rcu_momentary_dyntick_idle();
491 local_irq_restore(flags);
492 }
493 if (unlikely(raw_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)))
494 rcu_sched_qs();
495 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
496 barrier(); /* Avoid RCU read-side critical sections leaking up. */
497 preempt_enable();
498}
499EXPORT_SYMBOL_GPL(rcu_all_qs);
500
501#define DEFAULT_RCU_BLIMIT 10 /* Maximum callbacks per rcu_do_batch. */
502static long blimit = DEFAULT_RCU_BLIMIT;
503#define DEFAULT_RCU_QHIMARK 10000 /* If this many pending, ignore blimit. */
504static long qhimark = DEFAULT_RCU_QHIMARK;
505#define DEFAULT_RCU_QLOMARK 100 /* Once only this many pending, use blimit. */
506static long qlowmark = DEFAULT_RCU_QLOMARK;
507
508module_param(blimit, long, 0444);
509module_param(qhimark, long, 0444);
510module_param(qlowmark, long, 0444);
511
512static ulong jiffies_till_first_fqs = ULONG_MAX;
513static ulong jiffies_till_next_fqs = ULONG_MAX;
514static bool rcu_kick_kthreads;
515
516module_param(jiffies_till_first_fqs, ulong, 0644);
517module_param(jiffies_till_next_fqs, ulong, 0644);
518module_param(rcu_kick_kthreads, bool, 0644);
519
520/*
521 * How long the grace period must be before we start recruiting
522 * quiescent-state help from rcu_note_context_switch().
523 */
524static ulong jiffies_till_sched_qs = HZ / 10;
525module_param(jiffies_till_sched_qs, ulong, 0444);
526
527static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
528 struct rcu_data *rdp);
529static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp));
530static void force_quiescent_state(struct rcu_state *rsp);
531static int rcu_pending(void);
532
533/*
534 * Return the number of RCU batches started thus far for debug & stats.
535 */
536unsigned long rcu_batches_started(void)
537{
538 return rcu_state_p->gpnum;
539}
540EXPORT_SYMBOL_GPL(rcu_batches_started);
541
542/*
543 * Return the number of RCU-sched batches started thus far for debug & stats.
544 */
545unsigned long rcu_batches_started_sched(void)
546{
547 return rcu_sched_state.gpnum;
548}
549EXPORT_SYMBOL_GPL(rcu_batches_started_sched);
550
551/*
552 * Return the number of RCU BH batches started thus far for debug & stats.
553 */
554unsigned long rcu_batches_started_bh(void)
555{
556 return rcu_bh_state.gpnum;
557}
558EXPORT_SYMBOL_GPL(rcu_batches_started_bh);
559
560/*
561 * Return the number of RCU batches completed thus far for debug & stats.
562 */
563unsigned long rcu_batches_completed(void)
564{
565 return rcu_state_p->completed;
566}
567EXPORT_SYMBOL_GPL(rcu_batches_completed);
568
569/*
570 * Return the number of RCU-sched batches completed thus far for debug & stats.
571 */
572unsigned long rcu_batches_completed_sched(void)
573{
574 return rcu_sched_state.completed;
575}
576EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
577
578/*
579 * Return the number of RCU BH batches completed thus far for debug & stats.
580 */
581unsigned long rcu_batches_completed_bh(void)
582{
583 return rcu_bh_state.completed;
584}
585EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
586
587/*
588 * Return the number of RCU expedited batches completed thus far for
589 * debug & stats. Odd numbers mean that a batch is in progress, even
590 * numbers mean idle. The value returned will thus be roughly double
591 * the cumulative batches since boot.
592 */
593unsigned long rcu_exp_batches_completed(void)
594{
595 return rcu_state_p->expedited_sequence;
596}
597EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
598
599/*
600 * Return the number of RCU-sched expedited batches completed thus far
601 * for debug & stats. Similar to rcu_exp_batches_completed().
602 */
603unsigned long rcu_exp_batches_completed_sched(void)
604{
605 return rcu_sched_state.expedited_sequence;
606}
607EXPORT_SYMBOL_GPL(rcu_exp_batches_completed_sched);
608
609/*
610 * Force a quiescent state.
611 */
612void rcu_force_quiescent_state(void)
613{
614 force_quiescent_state(rcu_state_p);
615}
616EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
617
618/*
619 * Force a quiescent state for RCU BH.
620 */
621void rcu_bh_force_quiescent_state(void)
622{
623 force_quiescent_state(&rcu_bh_state);
624}
625EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
626
627/*
628 * Force a quiescent state for RCU-sched.
629 */
630void rcu_sched_force_quiescent_state(void)
631{
632 force_quiescent_state(&rcu_sched_state);
633}
634EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
635
636/*
637 * Show the state of the grace-period kthreads.
638 */
639void show_rcu_gp_kthreads(void)
640{
641 struct rcu_state *rsp;
642
643 for_each_rcu_flavor(rsp) {
644 pr_info("%s: wait state: %d ->state: %#lx\n",
645 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
646 /* sched_show_task(rsp->gp_kthread); */
647 }
648}
649EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
650
651/*
652 * Record the number of times rcutorture tests have been initiated and
653 * terminated. This information allows the debugfs tracing stats to be
654 * correlated to the rcutorture messages, even when the rcutorture module
655 * is being repeatedly loaded and unloaded. In other words, we cannot
656 * store this state in rcutorture itself.
657 */
658void rcutorture_record_test_transition(void)
659{
660 rcutorture_testseq++;
661 rcutorture_vernum = 0;
662}
663EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
664
665/*
666 * Send along grace-period-related data for rcutorture diagnostics.
667 */
668void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
669 unsigned long *gpnum, unsigned long *completed)
670{
671 struct rcu_state *rsp = NULL;
672
673 switch (test_type) {
674 case RCU_FLAVOR:
675 rsp = rcu_state_p;
676 break;
677 case RCU_BH_FLAVOR:
678 rsp = &rcu_bh_state;
679 break;
680 case RCU_SCHED_FLAVOR:
681 rsp = &rcu_sched_state;
682 break;
683 default:
684 break;
685 }
686 if (rsp == NULL)
687 return;
688 *flags = READ_ONCE(rsp->gp_flags);
689 *gpnum = READ_ONCE(rsp->gpnum);
690 *completed = READ_ONCE(rsp->completed);
691}
692EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
693
694/*
695 * Record the number of writer passes through the current rcutorture test.
696 * This is also used to correlate debugfs tracing stats with the rcutorture
697 * messages.
698 */
699void rcutorture_record_progress(unsigned long vernum)
700{
701 rcutorture_vernum++;
702}
703EXPORT_SYMBOL_GPL(rcutorture_record_progress);
704
705/*
706 * Return the root node of the specified rcu_state structure.
707 */
708static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
709{
710 return &rsp->node[0];
711}
712
713/*
714 * Is there any need for future grace periods?
715 * Interrupts must be disabled. If the caller does not hold the root
716 * rnp_node structure's ->lock, the results are advisory only.
717 */
718static int rcu_future_needs_gp(struct rcu_state *rsp)
719{
720 struct rcu_node *rnp = rcu_get_root(rsp);
721 int idx = (READ_ONCE(rnp->completed) + 1) & 0x1;
722 int *fp = &rnp->need_future_gp[idx];
723
724 lockdep_assert_irqs_disabled();
725 return READ_ONCE(*fp);
726}
727
728/*
729 * Does the current CPU require a not-yet-started grace period?
730 * The caller must have disabled interrupts to prevent races with
731 * normal callback registry.
732 */
733static bool
734cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
735{
736 lockdep_assert_irqs_disabled();
737 if (rcu_gp_in_progress(rsp))
738 return false; /* No, a grace period is already in progress. */
739 if (rcu_future_needs_gp(rsp))
740 return true; /* Yes, a no-CBs CPU needs one. */
741 if (!rcu_segcblist_is_enabled(&rdp->cblist))
742 return false; /* No, this is a no-CBs (or offline) CPU. */
743 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
744 return true; /* Yes, CPU has newly registered callbacks. */
745 if (rcu_segcblist_future_gp_needed(&rdp->cblist,
746 READ_ONCE(rsp->completed)))
747 return true; /* Yes, CBs for future grace period. */
748 return false; /* No grace period needed. */
749}
750
751/*
752 * Enter an RCU extended quiescent state, which can be either the
753 * idle loop or adaptive-tickless usermode execution.
754 *
755 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
756 * the possibility of usermode upcalls having messed up our count
757 * of interrupt nesting level during the prior busy period.
758 */
759static void rcu_eqs_enter(bool user)
760{
761 struct rcu_state *rsp;
762 struct rcu_data *rdp;
763 struct rcu_dynticks *rdtp;
764
765 rdtp = this_cpu_ptr(&rcu_dynticks);
766 WRITE_ONCE(rdtp->dynticks_nmi_nesting, 0);
767 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
768 rdtp->dynticks_nesting == 0);
769 if (rdtp->dynticks_nesting != 1) {
770 rdtp->dynticks_nesting--;
771 return;
772 }
773
774 lockdep_assert_irqs_disabled();
775 trace_rcu_dyntick(TPS("Start"), rdtp->dynticks_nesting, 0, rdtp->dynticks);
776 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
777 for_each_rcu_flavor(rsp) {
778 rdp = this_cpu_ptr(rsp->rda);
779 do_nocb_deferred_wakeup(rdp);
780 }
781 rcu_prepare_for_idle();
782 WRITE_ONCE(rdtp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
783 rcu_dynticks_eqs_enter();
784 rcu_dynticks_task_enter();
785}
786
787/**
788 * rcu_idle_enter - inform RCU that current CPU is entering idle
789 *
790 * Enter idle mode, in other words, -leave- the mode in which RCU
791 * read-side critical sections can occur. (Though RCU read-side
792 * critical sections can occur in irq handlers in idle, a possibility
793 * handled by irq_enter() and irq_exit().)
794 *
795 * If you add or remove a call to rcu_idle_enter(), be sure to test with
796 * CONFIG_RCU_EQS_DEBUG=y.
797 */
798void rcu_idle_enter(void)
799{
800 lockdep_assert_irqs_disabled();
801 rcu_eqs_enter(false);
802}
803
804#ifdef CONFIG_NO_HZ_FULL
805/**
806 * rcu_user_enter - inform RCU that we are resuming userspace.
807 *
808 * Enter RCU idle mode right before resuming userspace. No use of RCU
809 * is permitted between this call and rcu_user_exit(). This way the
810 * CPU doesn't need to maintain the tick for RCU maintenance purposes
811 * when the CPU runs in userspace.
812 *
813 * If you add or remove a call to rcu_user_enter(), be sure to test with
814 * CONFIG_RCU_EQS_DEBUG=y.
815 */
816void rcu_user_enter(void)
817{
818 lockdep_assert_irqs_disabled();
819 rcu_eqs_enter(true);
820}
821#endif /* CONFIG_NO_HZ_FULL */
822
823/**
824 * rcu_nmi_exit - inform RCU of exit from NMI context
825 *
826 * If we are returning from the outermost NMI handler that interrupted an
827 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
828 * to let the RCU grace-period handling know that the CPU is back to
829 * being RCU-idle.
830 *
831 * If you add or remove a call to rcu_nmi_exit(), be sure to test
832 * with CONFIG_RCU_EQS_DEBUG=y.
833 */
834void rcu_nmi_exit(void)
835{
836 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
837
838 /*
839 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
840 * (We are exiting an NMI handler, so RCU better be paying attention
841 * to us!)
842 */
843 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
844 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
845
846 /*
847 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
848 * leave it in non-RCU-idle state.
849 */
850 if (rdtp->dynticks_nmi_nesting != 1) {
851 trace_rcu_dyntick(TPS("--="), rdtp->dynticks_nmi_nesting, rdtp->dynticks_nmi_nesting - 2, rdtp->dynticks);
852 WRITE_ONCE(rdtp->dynticks_nmi_nesting, /* No store tearing. */
853 rdtp->dynticks_nmi_nesting - 2);
854 return;
855 }
856
857 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
858 trace_rcu_dyntick(TPS("Startirq"), rdtp->dynticks_nmi_nesting, 0, rdtp->dynticks);
859 WRITE_ONCE(rdtp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
860 rcu_dynticks_eqs_enter();
861}
862
863/**
864 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
865 *
866 * Exit from an interrupt handler, which might possibly result in entering
867 * idle mode, in other words, leaving the mode in which read-side critical
868 * sections can occur. The caller must have disabled interrupts.
869 *
870 * This code assumes that the idle loop never does anything that might
871 * result in unbalanced calls to irq_enter() and irq_exit(). If your
872 * architecture's idle loop violates this assumption, RCU will give you what
873 * you deserve, good and hard. But very infrequently and irreproducibly.
874 *
875 * Use things like work queues to work around this limitation.
876 *
877 * You have been warned.
878 *
879 * If you add or remove a call to rcu_irq_exit(), be sure to test with
880 * CONFIG_RCU_EQS_DEBUG=y.
881 */
882void rcu_irq_exit(void)
883{
884 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
885
886 lockdep_assert_irqs_disabled();
887 if (rdtp->dynticks_nmi_nesting == 1)
888 rcu_prepare_for_idle();
889 rcu_nmi_exit();
890 if (rdtp->dynticks_nmi_nesting == 0)
891 rcu_dynticks_task_enter();
892}
893
894/*
895 * Wrapper for rcu_irq_exit() where interrupts are enabled.
896 *
897 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
898 * with CONFIG_RCU_EQS_DEBUG=y.
899 */
900void rcu_irq_exit_irqson(void)
901{
902 unsigned long flags;
903
904 local_irq_save(flags);
905 rcu_irq_exit();
906 local_irq_restore(flags);
907}
908
909/*
910 * Exit an RCU extended quiescent state, which can be either the
911 * idle loop or adaptive-tickless usermode execution.
912 *
913 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
914 * allow for the possibility of usermode upcalls messing up our count of
915 * interrupt nesting level during the busy period that is just now starting.
916 */
917static void rcu_eqs_exit(bool user)
918{
919 struct rcu_dynticks *rdtp;
920 long oldval;
921
922 lockdep_assert_irqs_disabled();
923 rdtp = this_cpu_ptr(&rcu_dynticks);
924 oldval = rdtp->dynticks_nesting;
925 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
926 if (oldval) {
927 rdtp->dynticks_nesting++;
928 return;
929 }
930 rcu_dynticks_task_exit();
931 rcu_dynticks_eqs_exit();
932 rcu_cleanup_after_idle();
933 trace_rcu_dyntick(TPS("End"), rdtp->dynticks_nesting, 1, rdtp->dynticks);
934 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
935 WRITE_ONCE(rdtp->dynticks_nesting, 1);
936 WRITE_ONCE(rdtp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
937}
938
939/**
940 * rcu_idle_exit - inform RCU that current CPU is leaving idle
941 *
942 * Exit idle mode, in other words, -enter- the mode in which RCU
943 * read-side critical sections can occur.
944 *
945 * If you add or remove a call to rcu_idle_exit(), be sure to test with
946 * CONFIG_RCU_EQS_DEBUG=y.
947 */
948void rcu_idle_exit(void)
949{
950 unsigned long flags;
951
952 local_irq_save(flags);
953 rcu_eqs_exit(false);
954 local_irq_restore(flags);
955}
956
957#ifdef CONFIG_NO_HZ_FULL
958/**
959 * rcu_user_exit - inform RCU that we are exiting userspace.
960 *
961 * Exit RCU idle mode while entering the kernel because it can
962 * run a RCU read side critical section anytime.
963 *
964 * If you add or remove a call to rcu_user_exit(), be sure to test with
965 * CONFIG_RCU_EQS_DEBUG=y.
966 */
967void rcu_user_exit(void)
968{
969 rcu_eqs_exit(1);
970}
971#endif /* CONFIG_NO_HZ_FULL */
972
973/**
974 * rcu_nmi_enter - inform RCU of entry to NMI context
975 *
976 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
977 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
978 * that the CPU is active. This implementation permits nested NMIs, as
979 * long as the nesting level does not overflow an int. (You will probably
980 * run out of stack space first.)
981 *
982 * If you add or remove a call to rcu_nmi_enter(), be sure to test
983 * with CONFIG_RCU_EQS_DEBUG=y.
984 */
985void rcu_nmi_enter(void)
986{
987 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
988 long incby = 2;
989
990 /* Complain about underflow. */
991 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
992
993 /*
994 * If idle from RCU viewpoint, atomically increment ->dynticks
995 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
996 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
997 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
998 * to be in the outermost NMI handler that interrupted an RCU-idle
999 * period (observation due to Andy Lutomirski).
1000 */
1001 if (rcu_dynticks_curr_cpu_in_eqs()) {
1002 rcu_dynticks_eqs_exit();
1003 incby = 1;
1004 }
1005 trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
1006 rdtp->dynticks_nmi_nesting,
1007 rdtp->dynticks_nmi_nesting + incby, rdtp->dynticks);
1008 WRITE_ONCE(rdtp->dynticks_nmi_nesting, /* Prevent store tearing. */
1009 rdtp->dynticks_nmi_nesting + incby);
1010 barrier();
1011}
1012
1013/**
1014 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
1015 *
1016 * Enter an interrupt handler, which might possibly result in exiting
1017 * idle mode, in other words, entering the mode in which read-side critical
1018 * sections can occur. The caller must have disabled interrupts.
1019 *
1020 * Note that the Linux kernel is fully capable of entering an interrupt
1021 * handler that it never exits, for example when doing upcalls to user mode!
1022 * This code assumes that the idle loop never does upcalls to user mode.
1023 * If your architecture's idle loop does do upcalls to user mode (or does
1024 * anything else that results in unbalanced calls to the irq_enter() and
1025 * irq_exit() functions), RCU will give you what you deserve, good and hard.
1026 * But very infrequently and irreproducibly.
1027 *
1028 * Use things like work queues to work around this limitation.
1029 *
1030 * You have been warned.
1031 *
1032 * If you add or remove a call to rcu_irq_enter(), be sure to test with
1033 * CONFIG_RCU_EQS_DEBUG=y.
1034 */
1035void rcu_irq_enter(void)
1036{
1037 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1038
1039 lockdep_assert_irqs_disabled();
1040 if (rdtp->dynticks_nmi_nesting == 0)
1041 rcu_dynticks_task_exit();
1042 rcu_nmi_enter();
1043 if (rdtp->dynticks_nmi_nesting == 1)
1044 rcu_cleanup_after_idle();
1045}
1046
1047/*
1048 * Wrapper for rcu_irq_enter() where interrupts are enabled.
1049 *
1050 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
1051 * with CONFIG_RCU_EQS_DEBUG=y.
1052 */
1053void rcu_irq_enter_irqson(void)
1054{
1055 unsigned long flags;
1056
1057 local_irq_save(flags);
1058 rcu_irq_enter();
1059 local_irq_restore(flags);
1060}
1061
1062/**
1063 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1064 *
1065 * Return true if RCU is watching the running CPU, which means that this
1066 * CPU can safely enter RCU read-side critical sections. In other words,
1067 * if the current CPU is in its idle loop and is neither in an interrupt
1068 * or NMI handler, return true.
1069 */
1070bool notrace rcu_is_watching(void)
1071{
1072 bool ret;
1073
1074 preempt_disable_notrace();
1075 ret = !rcu_dynticks_curr_cpu_in_eqs();
1076 preempt_enable_notrace();
1077 return ret;
1078}
1079EXPORT_SYMBOL_GPL(rcu_is_watching);
1080
1081/*
1082 * If a holdout task is actually running, request an urgent quiescent
1083 * state from its CPU. This is unsynchronized, so migrations can cause
1084 * the request to go to the wrong CPU. Which is OK, all that will happen
1085 * is that the CPU's next context switch will be a bit slower and next
1086 * time around this task will generate another request.
1087 */
1088void rcu_request_urgent_qs_task(struct task_struct *t)
1089{
1090 int cpu;
1091
1092 barrier();
1093 cpu = task_cpu(t);
1094 if (!task_curr(t))
1095 return; /* This task is not running on that CPU. */
1096 smp_store_release(per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, cpu), true);
1097}
1098
1099#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1100
1101/*
1102 * Is the current CPU online? Disable preemption to avoid false positives
1103 * that could otherwise happen due to the current CPU number being sampled,
1104 * this task being preempted, its old CPU being taken offline, resuming
1105 * on some other CPU, then determining that its old CPU is now offline.
1106 * It is OK to use RCU on an offline processor during initial boot, hence
1107 * the check for rcu_scheduler_fully_active. Note also that it is OK
1108 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1109 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1110 * offline to continue to use RCU for one jiffy after marking itself
1111 * offline in the cpu_online_mask. This leniency is necessary given the
1112 * non-atomic nature of the online and offline processing, for example,
1113 * the fact that a CPU enters the scheduler after completing the teardown
1114 * of the CPU.
1115 *
1116 * This is also why RCU internally marks CPUs online during in the
1117 * preparation phase and offline after the CPU has been taken down.
1118 *
1119 * Disable checking if in an NMI handler because we cannot safely report
1120 * errors from NMI handlers anyway.
1121 */
1122bool rcu_lockdep_current_cpu_online(void)
1123{
1124 struct rcu_data *rdp;
1125 struct rcu_node *rnp;
1126 bool ret;
1127
1128 if (in_nmi())
1129 return true;
1130 preempt_disable();
1131 rdp = this_cpu_ptr(&rcu_sched_data);
1132 rnp = rdp->mynode;
1133 ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
1134 !rcu_scheduler_fully_active;
1135 preempt_enable();
1136 return ret;
1137}
1138EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1139
1140#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1141
1142/**
1143 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1144 *
1145 * If the current CPU is idle or running at a first-level (not nested)
1146 * interrupt from idle, return true. The caller must have at least
1147 * disabled preemption.
1148 */
1149static int rcu_is_cpu_rrupt_from_idle(void)
1150{
1151 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 0 &&
1152 __this_cpu_read(rcu_dynticks.dynticks_nmi_nesting) <= 1;
1153}
1154
1155/*
1156 * We are reporting a quiescent state on behalf of some other CPU, so
1157 * it is our responsibility to check for and handle potential overflow
1158 * of the rcu_node ->gpnum counter with respect to the rcu_data counters.
1159 * After all, the CPU might be in deep idle state, and thus executing no
1160 * code whatsoever.
1161 */
1162static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
1163{
1164 raw_lockdep_assert_held_rcu_node(rnp);
1165 if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4, rnp->gpnum))
1166 WRITE_ONCE(rdp->gpwrap, true);
1167 if (ULONG_CMP_LT(rdp->rcu_iw_gpnum + ULONG_MAX / 4, rnp->gpnum))
1168 rdp->rcu_iw_gpnum = rnp->gpnum + ULONG_MAX / 4;
1169}
1170
1171/*
1172 * Snapshot the specified CPU's dynticks counter so that we can later
1173 * credit them with an implicit quiescent state. Return 1 if this CPU
1174 * is in dynticks idle mode, which is an extended quiescent state.
1175 */
1176static int dyntick_save_progress_counter(struct rcu_data *rdp)
1177{
1178 rdp->dynticks_snap = rcu_dynticks_snap(rdp->dynticks);
1179 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1180 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1181 rcu_gpnum_ovf(rdp->mynode, rdp);
1182 return 1;
1183 }
1184 return 0;
1185}
1186
1187/*
1188 * Handler for the irq_work request posted when a grace period has
1189 * gone on for too long, but not yet long enough for an RCU CPU
1190 * stall warning. Set state appropriately, but just complain if
1191 * there is unexpected state on entry.
1192 */
1193static void rcu_iw_handler(struct irq_work *iwp)
1194{
1195 struct rcu_data *rdp;
1196 struct rcu_node *rnp;
1197
1198 rdp = container_of(iwp, struct rcu_data, rcu_iw);
1199 rnp = rdp->mynode;
1200 raw_spin_lock_rcu_node(rnp);
1201 if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
1202 rdp->rcu_iw_gpnum = rnp->gpnum;
1203 rdp->rcu_iw_pending = false;
1204 }
1205 raw_spin_unlock_rcu_node(rnp);
1206}
1207
1208/*
1209 * Return true if the specified CPU has passed through a quiescent
1210 * state by virtue of being in or having passed through an dynticks
1211 * idle state since the last call to dyntick_save_progress_counter()
1212 * for this same CPU, or by virtue of having been offline.
1213 */
1214static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1215{
1216 unsigned long jtsq;
1217 bool *rnhqp;
1218 bool *ruqp;
1219 struct rcu_node *rnp = rdp->mynode;
1220
1221 /*
1222 * If the CPU passed through or entered a dynticks idle phase with
1223 * no active irq/NMI handlers, then we can safely pretend that the CPU
1224 * already acknowledged the request to pass through a quiescent
1225 * state. Either way, that CPU cannot possibly be in an RCU
1226 * read-side critical section that started before the beginning
1227 * of the current RCU grace period.
1228 */
1229 if (rcu_dynticks_in_eqs_since(rdp->dynticks, rdp->dynticks_snap)) {
1230 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1231 rdp->dynticks_fqs++;
1232 rcu_gpnum_ovf(rnp, rdp);
1233 return 1;
1234 }
1235
1236 /*
1237 * Has this CPU encountered a cond_resched_rcu_qs() since the
1238 * beginning of the grace period? For this to be the case,
1239 * the CPU has to have noticed the current grace period. This
1240 * might not be the case for nohz_full CPUs looping in the kernel.
1241 */
1242 jtsq = jiffies_till_sched_qs;
1243 ruqp = per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, rdp->cpu);
1244 if (time_after(jiffies, rdp->rsp->gp_start + jtsq) &&
1245 READ_ONCE(rdp->rcu_qs_ctr_snap) != per_cpu(rcu_dynticks.rcu_qs_ctr, rdp->cpu) &&
1246 READ_ONCE(rdp->gpnum) == rnp->gpnum && !rdp->gpwrap) {
1247 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("rqc"));
1248 rcu_gpnum_ovf(rnp, rdp);
1249 return 1;
1250 } else if (time_after(jiffies, rdp->rsp->gp_start + jtsq)) {
1251 /* Load rcu_qs_ctr before store to rcu_urgent_qs. */
1252 smp_store_release(ruqp, true);
1253 }
1254
1255 /* Check for the CPU being offline. */
1256 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp))) {
1257 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
1258 rdp->offline_fqs++;
1259 rcu_gpnum_ovf(rnp, rdp);
1260 return 1;
1261 }
1262
1263 /*
1264 * A CPU running for an extended time within the kernel can
1265 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1266 * even context-switching back and forth between a pair of
1267 * in-kernel CPU-bound tasks cannot advance grace periods.
1268 * So if the grace period is old enough, make the CPU pay attention.
1269 * Note that the unsynchronized assignments to the per-CPU
1270 * rcu_need_heavy_qs variable are safe. Yes, setting of
1271 * bits can be lost, but they will be set again on the next
1272 * force-quiescent-state pass. So lost bit sets do not result
1273 * in incorrect behavior, merely in a grace period lasting
1274 * a few jiffies longer than it might otherwise. Because
1275 * there are at most four threads involved, and because the
1276 * updates are only once every few jiffies, the probability of
1277 * lossage (and thus of slight grace-period extension) is
1278 * quite low.
1279 */
1280 rnhqp = &per_cpu(rcu_dynticks.rcu_need_heavy_qs, rdp->cpu);
1281 if (!READ_ONCE(*rnhqp) &&
1282 (time_after(jiffies, rdp->rsp->gp_start + jtsq) ||
1283 time_after(jiffies, rdp->rsp->jiffies_resched))) {
1284 WRITE_ONCE(*rnhqp, true);
1285 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1286 smp_store_release(ruqp, true);
1287 rdp->rsp->jiffies_resched += jtsq; /* Re-enable beating. */
1288 }
1289
1290 /*
1291 * If more than halfway to RCU CPU stall-warning time, do a
1292 * resched_cpu() to try to loosen things up a bit. Also check to
1293 * see if the CPU is getting hammered with interrupts, but only
1294 * once per grace period, just to keep the IPIs down to a dull roar.
1295 */
1296 if (jiffies - rdp->rsp->gp_start > rcu_jiffies_till_stall_check() / 2) {
1297 resched_cpu(rdp->cpu);
1298 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1299 !rdp->rcu_iw_pending && rdp->rcu_iw_gpnum != rnp->gpnum &&
1300 (rnp->ffmask & rdp->grpmask)) {
1301 init_irq_work(&rdp->rcu_iw, rcu_iw_handler);
1302 rdp->rcu_iw_pending = true;
1303 rdp->rcu_iw_gpnum = rnp->gpnum;
1304 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1305 }
1306 }
1307
1308 return 0;
1309}
1310
1311static void record_gp_stall_check_time(struct rcu_state *rsp)
1312{
1313 unsigned long j = jiffies;
1314 unsigned long j1;
1315
1316 rsp->gp_start = j;
1317 smp_wmb(); /* Record start time before stall time. */
1318 j1 = rcu_jiffies_till_stall_check();
1319 WRITE_ONCE(rsp->jiffies_stall, j + j1);
1320 rsp->jiffies_resched = j + j1 / 2;
1321 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1322}
1323
1324/*
1325 * Convert a ->gp_state value to a character string.
1326 */
1327static const char *gp_state_getname(short gs)
1328{
1329 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
1330 return "???";
1331 return gp_state_names[gs];
1332}
1333
1334/*
1335 * Complain about starvation of grace-period kthread.
1336 */
1337static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1338{
1339 unsigned long gpa;
1340 unsigned long j;
1341
1342 j = jiffies;
1343 gpa = READ_ONCE(rsp->gp_activity);
1344 if (j - gpa > 2 * HZ) {
1345 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x %s(%d) ->state=%#lx ->cpu=%d\n",
1346 rsp->name, j - gpa,
1347 rsp->gpnum, rsp->completed,
1348 rsp->gp_flags,
1349 gp_state_getname(rsp->gp_state), rsp->gp_state,
1350 rsp->gp_kthread ? rsp->gp_kthread->state : ~0,
1351 rsp->gp_kthread ? task_cpu(rsp->gp_kthread) : -1);
1352 if (rsp->gp_kthread) {
1353 pr_err("RCU grace-period kthread stack dump:\n");
1354 sched_show_task(rsp->gp_kthread);
1355 wake_up_process(rsp->gp_kthread);
1356 }
1357 }
1358}
1359
1360/*
1361 * Dump stacks of all tasks running on stalled CPUs. First try using
1362 * NMIs, but fall back to manual remote stack tracing on architectures
1363 * that don't support NMI-based stack dumps. The NMI-triggered stack
1364 * traces are more accurate because they are printed by the target CPU.
1365 */
1366static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1367{
1368 int cpu;
1369 unsigned long flags;
1370 struct rcu_node *rnp;
1371
1372 rcu_for_each_leaf_node(rsp, rnp) {
1373 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1374 for_each_leaf_node_possible_cpu(rnp, cpu)
1375 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
1376 if (!trigger_single_cpu_backtrace(cpu))
1377 dump_cpu_task(cpu);
1378 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1379 }
1380}
1381
1382/*
1383 * If too much time has passed in the current grace period, and if
1384 * so configured, go kick the relevant kthreads.
1385 */
1386static void rcu_stall_kick_kthreads(struct rcu_state *rsp)
1387{
1388 unsigned long j;
1389
1390 if (!rcu_kick_kthreads)
1391 return;
1392 j = READ_ONCE(rsp->jiffies_kick_kthreads);
1393 if (time_after(jiffies, j) && rsp->gp_kthread &&
1394 (rcu_gp_in_progress(rsp) || READ_ONCE(rsp->gp_flags))) {
1395 WARN_ONCE(1, "Kicking %s grace-period kthread\n", rsp->name);
1396 rcu_ftrace_dump(DUMP_ALL);
1397 wake_up_process(rsp->gp_kthread);
1398 WRITE_ONCE(rsp->jiffies_kick_kthreads, j + HZ);
1399 }
1400}
1401
1402static inline void panic_on_rcu_stall(void)
1403{
1404 if (sysctl_panic_on_rcu_stall)
1405 panic("RCU Stall\n");
1406}
1407
1408static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
1409{
1410 int cpu;
1411 long delta;
1412 unsigned long flags;
1413 unsigned long gpa;
1414 unsigned long j;
1415 int ndetected = 0;
1416 struct rcu_node *rnp = rcu_get_root(rsp);
1417 long totqlen = 0;
1418
1419 /* Kick and suppress, if so configured. */
1420 rcu_stall_kick_kthreads(rsp);
1421 if (rcu_cpu_stall_suppress)
1422 return;
1423
1424 /* Only let one CPU complain about others per time interval. */
1425
1426 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1427 delta = jiffies - READ_ONCE(rsp->jiffies_stall);
1428 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1429 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1430 return;
1431 }
1432 WRITE_ONCE(rsp->jiffies_stall,
1433 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1434 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1435
1436 /*
1437 * OK, time to rat on our buddy...
1438 * See Documentation/RCU/stallwarn.txt for info on how to debug
1439 * RCU CPU stall warnings.
1440 */
1441 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1442 rsp->name);
1443 print_cpu_stall_info_begin();
1444 rcu_for_each_leaf_node(rsp, rnp) {
1445 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1446 ndetected += rcu_print_task_stall(rnp);
1447 if (rnp->qsmask != 0) {
1448 for_each_leaf_node_possible_cpu(rnp, cpu)
1449 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
1450 print_cpu_stall_info(rsp, cpu);
1451 ndetected++;
1452 }
1453 }
1454 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1455 }
1456
1457 print_cpu_stall_info_end();
1458 for_each_possible_cpu(cpu)
1459 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1460 cpu)->cblist);
1461 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1462 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1463 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1464 if (ndetected) {
1465 rcu_dump_cpu_stacks(rsp);
1466
1467 /* Complain about tasks blocking the grace period. */
1468 rcu_print_detail_task_stall(rsp);
1469 } else {
1470 if (READ_ONCE(rsp->gpnum) != gpnum ||
1471 READ_ONCE(rsp->completed) == gpnum) {
1472 pr_err("INFO: Stall ended before state dump start\n");
1473 } else {
1474 j = jiffies;
1475 gpa = READ_ONCE(rsp->gp_activity);
1476 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1477 rsp->name, j - gpa, j, gpa,
1478 jiffies_till_next_fqs,
1479 rcu_get_root(rsp)->qsmask);
1480 /* In this case, the current CPU might be at fault. */
1481 sched_show_task(current);
1482 }
1483 }
1484
1485 rcu_check_gp_kthread_starvation(rsp);
1486
1487 panic_on_rcu_stall();
1488
1489 force_quiescent_state(rsp); /* Kick them all. */
1490}
1491
1492static void print_cpu_stall(struct rcu_state *rsp)
1493{
1494 int cpu;
1495 unsigned long flags;
1496 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1497 struct rcu_node *rnp = rcu_get_root(rsp);
1498 long totqlen = 0;
1499
1500 /* Kick and suppress, if so configured. */
1501 rcu_stall_kick_kthreads(rsp);
1502 if (rcu_cpu_stall_suppress)
1503 return;
1504
1505 /*
1506 * OK, time to rat on ourselves...
1507 * See Documentation/RCU/stallwarn.txt for info on how to debug
1508 * RCU CPU stall warnings.
1509 */
1510 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1511 print_cpu_stall_info_begin();
1512 raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
1513 print_cpu_stall_info(rsp, smp_processor_id());
1514 raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
1515 print_cpu_stall_info_end();
1516 for_each_possible_cpu(cpu)
1517 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1518 cpu)->cblist);
1519 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1520 jiffies - rsp->gp_start,
1521 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1522
1523 rcu_check_gp_kthread_starvation(rsp);
1524
1525 rcu_dump_cpu_stacks(rsp);
1526
1527 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1528 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1529 WRITE_ONCE(rsp->jiffies_stall,
1530 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1531 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1532
1533 panic_on_rcu_stall();
1534
1535 /*
1536 * Attempt to revive the RCU machinery by forcing a context switch.
1537 *
1538 * A context switch would normally allow the RCU state machine to make
1539 * progress and it could be we're stuck in kernel space without context
1540 * switches for an entirely unreasonable amount of time.
1541 */
1542 resched_cpu(smp_processor_id());
1543}
1544
1545static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1546{
1547 unsigned long completed;
1548 unsigned long gpnum;
1549 unsigned long gps;
1550 unsigned long j;
1551 unsigned long js;
1552 struct rcu_node *rnp;
1553
1554 if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) ||
1555 !rcu_gp_in_progress(rsp))
1556 return;
1557 rcu_stall_kick_kthreads(rsp);
1558 j = jiffies;
1559
1560 /*
1561 * Lots of memory barriers to reject false positives.
1562 *
1563 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1564 * then rsp->gp_start, and finally rsp->completed. These values
1565 * are updated in the opposite order with memory barriers (or
1566 * equivalent) during grace-period initialization and cleanup.
1567 * Now, a false positive can occur if we get an new value of
1568 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1569 * the memory barriers, the only way that this can happen is if one
1570 * grace period ends and another starts between these two fetches.
1571 * Detect this by comparing rsp->completed with the previous fetch
1572 * from rsp->gpnum.
1573 *
1574 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1575 * and rsp->gp_start suffice to forestall false positives.
1576 */
1577 gpnum = READ_ONCE(rsp->gpnum);
1578 smp_rmb(); /* Pick up ->gpnum first... */
1579 js = READ_ONCE(rsp->jiffies_stall);
1580 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1581 gps = READ_ONCE(rsp->gp_start);
1582 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1583 completed = READ_ONCE(rsp->completed);
1584 if (ULONG_CMP_GE(completed, gpnum) ||
1585 ULONG_CMP_LT(j, js) ||
1586 ULONG_CMP_GE(gps, js))
1587 return; /* No stall or GP completed since entering function. */
1588 rnp = rdp->mynode;
1589 if (rcu_gp_in_progress(rsp) &&
1590 (READ_ONCE(rnp->qsmask) & rdp->grpmask)) {
1591
1592 /* We haven't checked in, so go dump stack. */
1593 print_cpu_stall(rsp);
1594
1595 } else if (rcu_gp_in_progress(rsp) &&
1596 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1597
1598 /* They had a few time units to dump stack, so complain. */
1599 print_other_cpu_stall(rsp, gpnum);
1600 }
1601}
1602
1603/**
1604 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1605 *
1606 * Set the stall-warning timeout way off into the future, thus preventing
1607 * any RCU CPU stall-warning messages from appearing in the current set of
1608 * RCU grace periods.
1609 *
1610 * The caller must disable hard irqs.
1611 */
1612void rcu_cpu_stall_reset(void)
1613{
1614 struct rcu_state *rsp;
1615
1616 for_each_rcu_flavor(rsp)
1617 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1618}
1619
1620/*
1621 * Determine the value that ->completed will have at the end of the
1622 * next subsequent grace period. This is used to tag callbacks so that
1623 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1624 * been dyntick-idle for an extended period with callbacks under the
1625 * influence of RCU_FAST_NO_HZ.
1626 *
1627 * The caller must hold rnp->lock with interrupts disabled.
1628 */
1629static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1630 struct rcu_node *rnp)
1631{
1632 raw_lockdep_assert_held_rcu_node(rnp);
1633
1634 /*
1635 * If RCU is idle, we just wait for the next grace period.
1636 * But we can only be sure that RCU is idle if we are looking
1637 * at the root rcu_node structure -- otherwise, a new grace
1638 * period might have started, but just not yet gotten around
1639 * to initializing the current non-root rcu_node structure.
1640 */
1641 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1642 return rnp->completed + 1;
1643
1644 /*
1645 * Otherwise, wait for a possible partial grace period and
1646 * then the subsequent full grace period.
1647 */
1648 return rnp->completed + 2;
1649}
1650
1651/*
1652 * Trace-event helper function for rcu_start_future_gp() and
1653 * rcu_nocb_wait_gp().
1654 */
1655static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1656 unsigned long c, const char *s)
1657{
1658 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1659 rnp->completed, c, rnp->level,
1660 rnp->grplo, rnp->grphi, s);
1661}
1662
1663/*
1664 * Start some future grace period, as needed to handle newly arrived
1665 * callbacks. The required future grace periods are recorded in each
1666 * rcu_node structure's ->need_future_gp field. Returns true if there
1667 * is reason to awaken the grace-period kthread.
1668 *
1669 * The caller must hold the specified rcu_node structure's ->lock.
1670 */
1671static bool __maybe_unused
1672rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1673 unsigned long *c_out)
1674{
1675 unsigned long c;
1676 bool ret = false;
1677 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1678
1679 raw_lockdep_assert_held_rcu_node(rnp);
1680
1681 /*
1682 * Pick up grace-period number for new callbacks. If this
1683 * grace period is already marked as needed, return to the caller.
1684 */
1685 c = rcu_cbs_completed(rdp->rsp, rnp);
1686 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1687 if (rnp->need_future_gp[c & 0x1]) {
1688 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1689 goto out;
1690 }
1691
1692 /*
1693 * If either this rcu_node structure or the root rcu_node structure
1694 * believe that a grace period is in progress, then we must wait
1695 * for the one following, which is in "c". Because our request
1696 * will be noticed at the end of the current grace period, we don't
1697 * need to explicitly start one. We only do the lockless check
1698 * of rnp_root's fields if the current rcu_node structure thinks
1699 * there is no grace period in flight, and because we hold rnp->lock,
1700 * the only possible change is when rnp_root's two fields are
1701 * equal, in which case rnp_root->gpnum might be concurrently
1702 * incremented. But that is OK, as it will just result in our
1703 * doing some extra useless work.
1704 */
1705 if (rnp->gpnum != rnp->completed ||
1706 READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) {
1707 rnp->need_future_gp[c & 0x1]++;
1708 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1709 goto out;
1710 }
1711
1712 /*
1713 * There might be no grace period in progress. If we don't already
1714 * hold it, acquire the root rcu_node structure's lock in order to
1715 * start one (if needed).
1716 */
1717 if (rnp != rnp_root)
1718 raw_spin_lock_rcu_node(rnp_root);
1719
1720 /*
1721 * Get a new grace-period number. If there really is no grace
1722 * period in progress, it will be smaller than the one we obtained
1723 * earlier. Adjust callbacks as needed.
1724 */
1725 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1726 if (!rcu_is_nocb_cpu(rdp->cpu))
1727 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1728
1729 /*
1730 * If the needed for the required grace period is already
1731 * recorded, trace and leave.
1732 */
1733 if (rnp_root->need_future_gp[c & 0x1]) {
1734 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1735 goto unlock_out;
1736 }
1737
1738 /* Record the need for the future grace period. */
1739 rnp_root->need_future_gp[c & 0x1]++;
1740
1741 /* If a grace period is not already in progress, start one. */
1742 if (rnp_root->gpnum != rnp_root->completed) {
1743 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1744 } else {
1745 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1746 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1747 }
1748unlock_out:
1749 if (rnp != rnp_root)
1750 raw_spin_unlock_rcu_node(rnp_root);
1751out:
1752 if (c_out != NULL)
1753 *c_out = c;
1754 return ret;
1755}
1756
1757/*
1758 * Clean up any old requests for the just-ended grace period. Also return
1759 * whether any additional grace periods have been requested.
1760 */
1761static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1762{
1763 int c = rnp->completed;
1764 int needmore;
1765 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1766
1767 rnp->need_future_gp[c & 0x1] = 0;
1768 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1769 trace_rcu_future_gp(rnp, rdp, c,
1770 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1771 return needmore;
1772}
1773
1774/*
1775 * Awaken the grace-period kthread for the specified flavor of RCU.
1776 * Don't do a self-awaken, and don't bother awakening when there is
1777 * nothing for the grace-period kthread to do (as in several CPUs
1778 * raced to awaken, and we lost), and finally don't try to awaken
1779 * a kthread that has not yet been created.
1780 */
1781static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1782{
1783 if (current == rsp->gp_kthread ||
1784 !READ_ONCE(rsp->gp_flags) ||
1785 !rsp->gp_kthread)
1786 return;
1787 swake_up(&rsp->gp_wq);
1788}
1789
1790/*
1791 * If there is room, assign a ->completed number to any callbacks on
1792 * this CPU that have not already been assigned. Also accelerate any
1793 * callbacks that were previously assigned a ->completed number that has
1794 * since proven to be too conservative, which can happen if callbacks get
1795 * assigned a ->completed number while RCU is idle, but with reference to
1796 * a non-root rcu_node structure. This function is idempotent, so it does
1797 * not hurt to call it repeatedly. Returns an flag saying that we should
1798 * awaken the RCU grace-period kthread.
1799 *
1800 * The caller must hold rnp->lock with interrupts disabled.
1801 */
1802static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1803 struct rcu_data *rdp)
1804{
1805 bool ret = false;
1806
1807 raw_lockdep_assert_held_rcu_node(rnp);
1808
1809 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1810 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1811 return false;
1812
1813 /*
1814 * Callbacks are often registered with incomplete grace-period
1815 * information. Something about the fact that getting exact
1816 * information requires acquiring a global lock... RCU therefore
1817 * makes a conservative estimate of the grace period number at which
1818 * a given callback will become ready to invoke. The following
1819 * code checks this estimate and improves it when possible, thus
1820 * accelerating callback invocation to an earlier grace-period
1821 * number.
1822 */
1823 if (rcu_segcblist_accelerate(&rdp->cblist, rcu_cbs_completed(rsp, rnp)))
1824 ret = rcu_start_future_gp(rnp, rdp, NULL);
1825
1826 /* Trace depending on how much we were able to accelerate. */
1827 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1828 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1829 else
1830 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1831 return ret;
1832}
1833
1834/*
1835 * Move any callbacks whose grace period has completed to the
1836 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1837 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1838 * sublist. This function is idempotent, so it does not hurt to
1839 * invoke it repeatedly. As long as it is not invoked -too- often...
1840 * Returns true if the RCU grace-period kthread needs to be awakened.
1841 *
1842 * The caller must hold rnp->lock with interrupts disabled.
1843 */
1844static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1845 struct rcu_data *rdp)
1846{
1847 raw_lockdep_assert_held_rcu_node(rnp);
1848
1849 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1850 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1851 return false;
1852
1853 /*
1854 * Find all callbacks whose ->completed numbers indicate that they
1855 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1856 */
1857 rcu_segcblist_advance(&rdp->cblist, rnp->completed);
1858
1859 /* Classify any remaining callbacks. */
1860 return rcu_accelerate_cbs(rsp, rnp, rdp);
1861}
1862
1863/*
1864 * Update CPU-local rcu_data state to record the beginnings and ends of
1865 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1866 * structure corresponding to the current CPU, and must have irqs disabled.
1867 * Returns true if the grace-period kthread needs to be awakened.
1868 */
1869static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1870 struct rcu_data *rdp)
1871{
1872 bool ret;
1873 bool need_gp;
1874
1875 raw_lockdep_assert_held_rcu_node(rnp);
1876
1877 /* Handle the ends of any preceding grace periods first. */
1878 if (rdp->completed == rnp->completed &&
1879 !unlikely(READ_ONCE(rdp->gpwrap))) {
1880
1881 /* No grace period end, so just accelerate recent callbacks. */
1882 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1883
1884 } else {
1885
1886 /* Advance callbacks. */
1887 ret = rcu_advance_cbs(rsp, rnp, rdp);
1888
1889 /* Remember that we saw this grace-period completion. */
1890 rdp->completed = rnp->completed;
1891 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1892 }
1893
1894 if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) {
1895 /*
1896 * If the current grace period is waiting for this CPU,
1897 * set up to detect a quiescent state, otherwise don't
1898 * go looking for one.
1899 */
1900 rdp->gpnum = rnp->gpnum;
1901 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1902 need_gp = !!(rnp->qsmask & rdp->grpmask);
1903 rdp->cpu_no_qs.b.norm = need_gp;
1904 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
1905 rdp->core_needs_qs = need_gp;
1906 zero_cpu_stall_ticks(rdp);
1907 WRITE_ONCE(rdp->gpwrap, false);
1908 rcu_gpnum_ovf(rnp, rdp);
1909 }
1910 return ret;
1911}
1912
1913static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1914{
1915 unsigned long flags;
1916 bool needwake;
1917 struct rcu_node *rnp;
1918
1919 local_irq_save(flags);
1920 rnp = rdp->mynode;
1921 if ((rdp->gpnum == READ_ONCE(rnp->gpnum) &&
1922 rdp->completed == READ_ONCE(rnp->completed) &&
1923 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1924 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1925 local_irq_restore(flags);
1926 return;
1927 }
1928 needwake = __note_gp_changes(rsp, rnp, rdp);
1929 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1930 if (needwake)
1931 rcu_gp_kthread_wake(rsp);
1932}
1933
1934static void rcu_gp_slow(struct rcu_state *rsp, int delay)
1935{
1936 if (delay > 0 &&
1937 !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1938 schedule_timeout_uninterruptible(delay);
1939}
1940
1941/*
1942 * Initialize a new grace period. Return false if no grace period required.
1943 */
1944static bool rcu_gp_init(struct rcu_state *rsp)
1945{
1946 unsigned long oldmask;
1947 struct rcu_data *rdp;
1948 struct rcu_node *rnp = rcu_get_root(rsp);
1949
1950 WRITE_ONCE(rsp->gp_activity, jiffies);
1951 raw_spin_lock_irq_rcu_node(rnp);
1952 if (!READ_ONCE(rsp->gp_flags)) {
1953 /* Spurious wakeup, tell caller to go back to sleep. */
1954 raw_spin_unlock_irq_rcu_node(rnp);
1955 return false;
1956 }
1957 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
1958
1959 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1960 /*
1961 * Grace period already in progress, don't start another.
1962 * Not supposed to be able to happen.
1963 */
1964 raw_spin_unlock_irq_rcu_node(rnp);
1965 return false;
1966 }
1967
1968 /* Advance to a new grace period and initialize state. */
1969 record_gp_stall_check_time(rsp);
1970 /* Record GP times before starting GP, hence smp_store_release(). */
1971 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1972 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1973 raw_spin_unlock_irq_rcu_node(rnp);
1974
1975 /*
1976 * Apply per-leaf buffered online and offline operations to the
1977 * rcu_node tree. Note that this new grace period need not wait
1978 * for subsequent online CPUs, and that quiescent-state forcing
1979 * will handle subsequent offline CPUs.
1980 */
1981 rcu_for_each_leaf_node(rsp, rnp) {
1982 rcu_gp_slow(rsp, gp_preinit_delay);
1983 raw_spin_lock_irq_rcu_node(rnp);
1984 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1985 !rnp->wait_blkd_tasks) {
1986 /* Nothing to do on this leaf rcu_node structure. */
1987 raw_spin_unlock_irq_rcu_node(rnp);
1988 continue;
1989 }
1990
1991 /* Record old state, apply changes to ->qsmaskinit field. */
1992 oldmask = rnp->qsmaskinit;
1993 rnp->qsmaskinit = rnp->qsmaskinitnext;
1994
1995 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1996 if (!oldmask != !rnp->qsmaskinit) {
1997 if (!oldmask) /* First online CPU for this rcu_node. */
1998 rcu_init_new_rnp(rnp);
1999 else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */
2000 rnp->wait_blkd_tasks = true;
2001 else /* Last offline CPU and can propagate. */
2002 rcu_cleanup_dead_rnp(rnp);
2003 }
2004
2005 /*
2006 * If all waited-on tasks from prior grace period are
2007 * done, and if all this rcu_node structure's CPUs are
2008 * still offline, propagate up the rcu_node tree and
2009 * clear ->wait_blkd_tasks. Otherwise, if one of this
2010 * rcu_node structure's CPUs has since come back online,
2011 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
2012 * checks for this, so just call it unconditionally).
2013 */
2014 if (rnp->wait_blkd_tasks &&
2015 (!rcu_preempt_has_tasks(rnp) ||
2016 rnp->qsmaskinit)) {
2017 rnp->wait_blkd_tasks = false;
2018 rcu_cleanup_dead_rnp(rnp);
2019 }
2020
2021 raw_spin_unlock_irq_rcu_node(rnp);
2022 }
2023
2024 /*
2025 * Set the quiescent-state-needed bits in all the rcu_node
2026 * structures for all currently online CPUs in breadth-first order,
2027 * starting from the root rcu_node structure, relying on the layout
2028 * of the tree within the rsp->node[] array. Note that other CPUs
2029 * will access only the leaves of the hierarchy, thus seeing that no
2030 * grace period is in progress, at least until the corresponding
2031 * leaf node has been initialized.
2032 *
2033 * The grace period cannot complete until the initialization
2034 * process finishes, because this kthread handles both.
2035 */
2036 rcu_for_each_node_breadth_first(rsp, rnp) {
2037 rcu_gp_slow(rsp, gp_init_delay);
2038 raw_spin_lock_irq_rcu_node(rnp);
2039 rdp = this_cpu_ptr(rsp->rda);
2040 rcu_preempt_check_blocked_tasks(rnp);
2041 rnp->qsmask = rnp->qsmaskinit;
2042 WRITE_ONCE(rnp->gpnum, rsp->gpnum);
2043 if (WARN_ON_ONCE(rnp->completed != rsp->completed))
2044 WRITE_ONCE(rnp->completed, rsp->completed);
2045 if (rnp == rdp->mynode)
2046 (void)__note_gp_changes(rsp, rnp, rdp);
2047 rcu_preempt_boost_start_gp(rnp);
2048 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
2049 rnp->level, rnp->grplo,
2050 rnp->grphi, rnp->qsmask);
2051 raw_spin_unlock_irq_rcu_node(rnp);
2052 cond_resched_rcu_qs();
2053 WRITE_ONCE(rsp->gp_activity, jiffies);
2054 }
2055
2056 return true;
2057}
2058
2059/*
2060 * Helper function for swait_event_idle() wakeup at force-quiescent-state
2061 * time.
2062 */
2063static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
2064{
2065 struct rcu_node *rnp = rcu_get_root(rsp);
2066
2067 /* Someone like call_rcu() requested a force-quiescent-state scan. */
2068 *gfp = READ_ONCE(rsp->gp_flags);
2069 if (*gfp & RCU_GP_FLAG_FQS)
2070 return true;
2071
2072 /* The current grace period has completed. */
2073 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
2074 return true;
2075
2076 return false;
2077}
2078
2079/*
2080 * Do one round of quiescent-state forcing.
2081 */
2082static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time)
2083{
2084 struct rcu_node *rnp = rcu_get_root(rsp);
2085
2086 WRITE_ONCE(rsp->gp_activity, jiffies);
2087 rsp->n_force_qs++;
2088 if (first_time) {
2089 /* Collect dyntick-idle snapshots. */
2090 force_qs_rnp(rsp, dyntick_save_progress_counter);
2091 } else {
2092 /* Handle dyntick-idle and offline CPUs. */
2093 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
2094 }
2095 /* Clear flag to prevent immediate re-entry. */
2096 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2097 raw_spin_lock_irq_rcu_node(rnp);
2098 WRITE_ONCE(rsp->gp_flags,
2099 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
2100 raw_spin_unlock_irq_rcu_node(rnp);
2101 }
2102}
2103
2104/*
2105 * Clean up after the old grace period.
2106 */
2107static void rcu_gp_cleanup(struct rcu_state *rsp)
2108{
2109 unsigned long gp_duration;
2110 bool needgp = false;
2111 int nocb = 0;
2112 struct rcu_data *rdp;
2113 struct rcu_node *rnp = rcu_get_root(rsp);
2114 struct swait_queue_head *sq;
2115
2116 WRITE_ONCE(rsp->gp_activity, jiffies);
2117 raw_spin_lock_irq_rcu_node(rnp);
2118 gp_duration = jiffies - rsp->gp_start;
2119 if (gp_duration > rsp->gp_max)
2120 rsp->gp_max = gp_duration;
2121
2122 /*
2123 * We know the grace period is complete, but to everyone else
2124 * it appears to still be ongoing. But it is also the case
2125 * that to everyone else it looks like there is nothing that
2126 * they can do to advance the grace period. It is therefore
2127 * safe for us to drop the lock in order to mark the grace
2128 * period as completed in all of the rcu_node structures.
2129 */
2130 raw_spin_unlock_irq_rcu_node(rnp);
2131
2132 /*
2133 * Propagate new ->completed value to rcu_node structures so
2134 * that other CPUs don't have to wait until the start of the next
2135 * grace period to process their callbacks. This also avoids
2136 * some nasty RCU grace-period initialization races by forcing
2137 * the end of the current grace period to be completely recorded in
2138 * all of the rcu_node structures before the beginning of the next
2139 * grace period is recorded in any of the rcu_node structures.
2140 */
2141 rcu_for_each_node_breadth_first(rsp, rnp) {
2142 raw_spin_lock_irq_rcu_node(rnp);
2143 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
2144 WARN_ON_ONCE(rnp->qsmask);
2145 WRITE_ONCE(rnp->completed, rsp->gpnum);
2146 rdp = this_cpu_ptr(rsp->rda);
2147 if (rnp == rdp->mynode)
2148 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2149 /* smp_mb() provided by prior unlock-lock pair. */
2150 nocb += rcu_future_gp_cleanup(rsp, rnp);
2151 sq = rcu_nocb_gp_get(rnp);
2152 raw_spin_unlock_irq_rcu_node(rnp);
2153 rcu_nocb_gp_cleanup(sq);
2154 cond_resched_rcu_qs();
2155 WRITE_ONCE(rsp->gp_activity, jiffies);
2156 rcu_gp_slow(rsp, gp_cleanup_delay);
2157 }
2158 rnp = rcu_get_root(rsp);
2159 raw_spin_lock_irq_rcu_node(rnp); /* Order GP before ->completed update. */
2160 rcu_nocb_gp_set(rnp, nocb);
2161
2162 /* Declare grace period done. */
2163 WRITE_ONCE(rsp->completed, rsp->gpnum);
2164 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
2165 rsp->gp_state = RCU_GP_IDLE;
2166 rdp = this_cpu_ptr(rsp->rda);
2167 /* Advance CBs to reduce false positives below. */
2168 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
2169 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
2170 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2171 trace_rcu_grace_period(rsp->name,
2172 READ_ONCE(rsp->gpnum),
2173 TPS("newreq"));
2174 }
2175 raw_spin_unlock_irq_rcu_node(rnp);
2176}
2177
2178/*
2179 * Body of kthread that handles grace periods.
2180 */
2181static int __noreturn rcu_gp_kthread(void *arg)
2182{
2183 bool first_gp_fqs;
2184 int gf;
2185 unsigned long j;
2186 int ret;
2187 struct rcu_state *rsp = arg;
2188 struct rcu_node *rnp = rcu_get_root(rsp);
2189
2190 rcu_bind_gp_kthread();
2191 for (;;) {
2192
2193 /* Handle grace-period start. */
2194 for (;;) {
2195 trace_rcu_grace_period(rsp->name,
2196 READ_ONCE(rsp->gpnum),
2197 TPS("reqwait"));
2198 rsp->gp_state = RCU_GP_WAIT_GPS;
2199 swait_event_idle(rsp->gp_wq, READ_ONCE(rsp->gp_flags) &
2200 RCU_GP_FLAG_INIT);
2201 rsp->gp_state = RCU_GP_DONE_GPS;
2202 /* Locking provides needed memory barrier. */
2203 if (rcu_gp_init(rsp))
2204 break;
2205 cond_resched_rcu_qs();
2206 WRITE_ONCE(rsp->gp_activity, jiffies);
2207 WARN_ON(signal_pending(current));
2208 trace_rcu_grace_period(rsp->name,
2209 READ_ONCE(rsp->gpnum),
2210 TPS("reqwaitsig"));
2211 }
2212
2213 /* Handle quiescent-state forcing. */
2214 first_gp_fqs = true;
2215 j = jiffies_till_first_fqs;
2216 if (j > HZ) {
2217 j = HZ;
2218 jiffies_till_first_fqs = HZ;
2219 }
2220 ret = 0;
2221 for (;;) {
2222 if (!ret) {
2223 rsp->jiffies_force_qs = jiffies + j;
2224 WRITE_ONCE(rsp->jiffies_kick_kthreads,
2225 jiffies + 3 * j);
2226 }
2227 trace_rcu_grace_period(rsp->name,
2228 READ_ONCE(rsp->gpnum),
2229 TPS("fqswait"));
2230 rsp->gp_state = RCU_GP_WAIT_FQS;
2231 ret = swait_event_idle_timeout(rsp->gp_wq,
2232 rcu_gp_fqs_check_wake(rsp, &gf), j);
2233 rsp->gp_state = RCU_GP_DOING_FQS;
2234 /* Locking provides needed memory barriers. */
2235 /* If grace period done, leave loop. */
2236 if (!READ_ONCE(rnp->qsmask) &&
2237 !rcu_preempt_blocked_readers_cgp(rnp))
2238 break;
2239 /* If time for quiescent-state forcing, do it. */
2240 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2241 (gf & RCU_GP_FLAG_FQS)) {
2242 trace_rcu_grace_period(rsp->name,
2243 READ_ONCE(rsp->gpnum),
2244 TPS("fqsstart"));
2245 rcu_gp_fqs(rsp, first_gp_fqs);
2246 first_gp_fqs = false;
2247 trace_rcu_grace_period(rsp->name,
2248 READ_ONCE(rsp->gpnum),
2249 TPS("fqsend"));
2250 cond_resched_rcu_qs();
2251 WRITE_ONCE(rsp->gp_activity, jiffies);
2252 ret = 0; /* Force full wait till next FQS. */
2253 j = jiffies_till_next_fqs;
2254 if (j > HZ) {
2255 j = HZ;
2256 jiffies_till_next_fqs = HZ;
2257 } else if (j < 1) {
2258 j = 1;
2259 jiffies_till_next_fqs = 1;
2260 }
2261 } else {
2262 /* Deal with stray signal. */
2263 cond_resched_rcu_qs();
2264 WRITE_ONCE(rsp->gp_activity, jiffies);
2265 WARN_ON(signal_pending(current));
2266 trace_rcu_grace_period(rsp->name,
2267 READ_ONCE(rsp->gpnum),
2268 TPS("fqswaitsig"));
2269 ret = 1; /* Keep old FQS timing. */
2270 j = jiffies;
2271 if (time_after(jiffies, rsp->jiffies_force_qs))
2272 j = 1;
2273 else
2274 j = rsp->jiffies_force_qs - j;
2275 }
2276 }
2277
2278 /* Handle grace-period end. */
2279 rsp->gp_state = RCU_GP_CLEANUP;
2280 rcu_gp_cleanup(rsp);
2281 rsp->gp_state = RCU_GP_CLEANED;
2282 }
2283}
2284
2285/*
2286 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2287 * in preparation for detecting the next grace period. The caller must hold
2288 * the root node's ->lock and hard irqs must be disabled.
2289 *
2290 * Note that it is legal for a dying CPU (which is marked as offline) to
2291 * invoke this function. This can happen when the dying CPU reports its
2292 * quiescent state.
2293 *
2294 * Returns true if the grace-period kthread must be awakened.
2295 */
2296static bool
2297rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
2298 struct rcu_data *rdp)
2299{
2300 raw_lockdep_assert_held_rcu_node(rnp);
2301 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
2302 /*
2303 * Either we have not yet spawned the grace-period
2304 * task, this CPU does not need another grace period,
2305 * or a grace period is already in progress.
2306 * Either way, don't start a new grace period.
2307 */
2308 return false;
2309 }
2310 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2311 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum),
2312 TPS("newreq"));
2313
2314 /*
2315 * We can't do wakeups while holding the rnp->lock, as that
2316 * could cause possible deadlocks with the rq->lock. Defer
2317 * the wakeup to our caller.
2318 */
2319 return true;
2320}
2321
2322/*
2323 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2324 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2325 * is invoked indirectly from rcu_advance_cbs(), which would result in
2326 * endless recursion -- or would do so if it wasn't for the self-deadlock
2327 * that is encountered beforehand.
2328 *
2329 * Returns true if the grace-period kthread needs to be awakened.
2330 */
2331static bool rcu_start_gp(struct rcu_state *rsp)
2332{
2333 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
2334 struct rcu_node *rnp = rcu_get_root(rsp);
2335 bool ret = false;
2336
2337 /*
2338 * If there is no grace period in progress right now, any
2339 * callbacks we have up to this point will be satisfied by the
2340 * next grace period. Also, advancing the callbacks reduces the
2341 * probability of false positives from cpu_needs_another_gp()
2342 * resulting in pointless grace periods. So, advance callbacks
2343 * then start the grace period!
2344 */
2345 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
2346 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
2347 return ret;
2348}
2349
2350/*
2351 * Report a full set of quiescent states to the specified rcu_state data
2352 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2353 * kthread if another grace period is required. Whether we wake
2354 * the grace-period kthread or it awakens itself for the next round
2355 * of quiescent-state forcing, that kthread will clean up after the
2356 * just-completed grace period. Note that the caller must hold rnp->lock,
2357 * which is released before return.
2358 */
2359static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2360 __releases(rcu_get_root(rsp)->lock)
2361{
2362 raw_lockdep_assert_held_rcu_node(rcu_get_root(rsp));
2363 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2364 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2365 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2366 rcu_gp_kthread_wake(rsp);
2367}
2368
2369/*
2370 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2371 * Allows quiescent states for a group of CPUs to be reported at one go
2372 * to the specified rcu_node structure, though all the CPUs in the group
2373 * must be represented by the same rcu_node structure (which need not be a
2374 * leaf rcu_node structure, though it often will be). The gps parameter
2375 * is the grace-period snapshot, which means that the quiescent states
2376 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2377 * must be held upon entry, and it is released before return.
2378 */
2379static void
2380rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2381 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2382 __releases(rnp->lock)
2383{
2384 unsigned long oldmask = 0;
2385 struct rcu_node *rnp_c;
2386
2387 raw_lockdep_assert_held_rcu_node(rnp);
2388
2389 /* Walk up the rcu_node hierarchy. */
2390 for (;;) {
2391 if (!(rnp->qsmask & mask) || rnp->gpnum != gps) {
2392
2393 /*
2394 * Our bit has already been cleared, or the
2395 * relevant grace period is already over, so done.
2396 */
2397 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2398 return;
2399 }
2400 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2401 WARN_ON_ONCE(rnp->level != rcu_num_lvls - 1 &&
2402 rcu_preempt_blocked_readers_cgp(rnp));
2403 rnp->qsmask &= ~mask;
2404 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
2405 mask, rnp->qsmask, rnp->level,
2406 rnp->grplo, rnp->grphi,
2407 !!rnp->gp_tasks);
2408 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2409
2410 /* Other bits still set at this level, so done. */
2411 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2412 return;
2413 }
2414 mask = rnp->grpmask;
2415 if (rnp->parent == NULL) {
2416
2417 /* No more levels. Exit loop holding root lock. */
2418
2419 break;
2420 }
2421 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2422 rnp_c = rnp;
2423 rnp = rnp->parent;
2424 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2425 oldmask = rnp_c->qsmask;
2426 }
2427
2428 /*
2429 * Get here if we are the last CPU to pass through a quiescent
2430 * state for this grace period. Invoke rcu_report_qs_rsp()
2431 * to clean up and start the next grace period if one is needed.
2432 */
2433 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2434}
2435
2436/*
2437 * Record a quiescent state for all tasks that were previously queued
2438 * on the specified rcu_node structure and that were blocking the current
2439 * RCU grace period. The caller must hold the specified rnp->lock with
2440 * irqs disabled, and this lock is released upon return, but irqs remain
2441 * disabled.
2442 */
2443static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2444 struct rcu_node *rnp, unsigned long flags)
2445 __releases(rnp->lock)
2446{
2447 unsigned long gps;
2448 unsigned long mask;
2449 struct rcu_node *rnp_p;
2450
2451 raw_lockdep_assert_held_rcu_node(rnp);
2452 if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
2453 rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2454 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2455 return; /* Still need more quiescent states! */
2456 }
2457
2458 rnp_p = rnp->parent;
2459 if (rnp_p == NULL) {
2460 /*
2461 * Only one rcu_node structure in the tree, so don't
2462 * try to report up to its nonexistent parent!
2463 */
2464 rcu_report_qs_rsp(rsp, flags);
2465 return;
2466 }
2467
2468 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2469 gps = rnp->gpnum;
2470 mask = rnp->grpmask;
2471 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2472 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2473 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2474}
2475
2476/*
2477 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2478 * structure. This must be called from the specified CPU.
2479 */
2480static void
2481rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2482{
2483 unsigned long flags;
2484 unsigned long mask;
2485 bool needwake;
2486 struct rcu_node *rnp;
2487
2488 rnp = rdp->mynode;
2489 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2490 if (rdp->cpu_no_qs.b.norm || rdp->gpnum != rnp->gpnum ||
2491 rnp->completed == rnp->gpnum || rdp->gpwrap) {
2492
2493 /*
2494 * The grace period in which this quiescent state was
2495 * recorded has ended, so don't report it upwards.
2496 * We will instead need a new quiescent state that lies
2497 * within the current grace period.
2498 */
2499 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2500 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
2501 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2502 return;
2503 }
2504 mask = rdp->grpmask;
2505 if ((rnp->qsmask & mask) == 0) {
2506 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2507 } else {
2508 rdp->core_needs_qs = false;
2509
2510 /*
2511 * This GP can't end until cpu checks in, so all of our
2512 * callbacks can be processed during the next GP.
2513 */
2514 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2515
2516 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2517 /* ^^^ Released rnp->lock */
2518 if (needwake)
2519 rcu_gp_kthread_wake(rsp);
2520 }
2521}
2522
2523/*
2524 * Check to see if there is a new grace period of which this CPU
2525 * is not yet aware, and if so, set up local rcu_data state for it.
2526 * Otherwise, see if this CPU has just passed through its first
2527 * quiescent state for this grace period, and record that fact if so.
2528 */
2529static void
2530rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2531{
2532 /* Check for grace-period ends and beginnings. */
2533 note_gp_changes(rsp, rdp);
2534
2535 /*
2536 * Does this CPU still need to do its part for current grace period?
2537 * If no, return and let the other CPUs do their part as well.
2538 */
2539 if (!rdp->core_needs_qs)
2540 return;
2541
2542 /*
2543 * Was there a quiescent state since the beginning of the grace
2544 * period? If no, then exit and wait for the next call.
2545 */
2546 if (rdp->cpu_no_qs.b.norm)
2547 return;
2548
2549 /*
2550 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2551 * judge of that).
2552 */
2553 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2554}
2555
2556/*
2557 * Trace the fact that this CPU is going offline.
2558 */
2559static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2560{
2561 RCU_TRACE(unsigned long mask;)
2562 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda);)
2563 RCU_TRACE(struct rcu_node *rnp = rdp->mynode;)
2564
2565 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2566 return;
2567
2568 RCU_TRACE(mask = rdp->grpmask;)
2569 trace_rcu_grace_period(rsp->name,
2570 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2571 TPS("cpuofl"));
2572}
2573
2574/*
2575 * All CPUs for the specified rcu_node structure have gone offline,
2576 * and all tasks that were preempted within an RCU read-side critical
2577 * section while running on one of those CPUs have since exited their RCU
2578 * read-side critical section. Some other CPU is reporting this fact with
2579 * the specified rcu_node structure's ->lock held and interrupts disabled.
2580 * This function therefore goes up the tree of rcu_node structures,
2581 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2582 * the leaf rcu_node structure's ->qsmaskinit field has already been
2583 * updated
2584 *
2585 * This function does check that the specified rcu_node structure has
2586 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2587 * prematurely. That said, invoking it after the fact will cost you
2588 * a needless lock acquisition. So once it has done its work, don't
2589 * invoke it again.
2590 */
2591static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2592{
2593 long mask;
2594 struct rcu_node *rnp = rnp_leaf;
2595
2596 raw_lockdep_assert_held_rcu_node(rnp);
2597 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2598 rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
2599 return;
2600 for (;;) {
2601 mask = rnp->grpmask;
2602 rnp = rnp->parent;
2603 if (!rnp)
2604 break;
2605 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2606 rnp->qsmaskinit &= ~mask;
2607 rnp->qsmask &= ~mask;
2608 if (rnp->qsmaskinit) {
2609 raw_spin_unlock_rcu_node(rnp);
2610 /* irqs remain disabled. */
2611 return;
2612 }
2613 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2614 }
2615}
2616
2617/*
2618 * The CPU has been completely removed, and some other CPU is reporting
2619 * this fact from process context. Do the remainder of the cleanup.
2620 * There can only be one CPU hotplug operation at a time, so no need for
2621 * explicit locking.
2622 */
2623static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2624{
2625 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2626 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2627
2628 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2629 return;
2630
2631 /* Adjust any no-longer-needed kthreads. */
2632 rcu_boost_kthread_setaffinity(rnp, -1);
2633}
2634
2635/*
2636 * Invoke any RCU callbacks that have made it to the end of their grace
2637 * period. Thottle as specified by rdp->blimit.
2638 */
2639static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2640{
2641 unsigned long flags;
2642 struct rcu_head *rhp;
2643 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2644 long bl, count;
2645
2646 /* If no callbacks are ready, just return. */
2647 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2648 trace_rcu_batch_start(rsp->name,
2649 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2650 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2651 trace_rcu_batch_end(rsp->name, 0,
2652 !rcu_segcblist_empty(&rdp->cblist),
2653 need_resched(), is_idle_task(current),
2654 rcu_is_callbacks_kthread());
2655 return;
2656 }
2657
2658 /*
2659 * Extract the list of ready callbacks, disabling to prevent
2660 * races with call_rcu() from interrupt handlers. Leave the
2661 * callback counts, as rcu_barrier() needs to be conservative.
2662 */
2663 local_irq_save(flags);
2664 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2665 bl = rdp->blimit;
2666 trace_rcu_batch_start(rsp->name, rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2667 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2668 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2669 local_irq_restore(flags);
2670
2671 /* Invoke callbacks. */
2672 rhp = rcu_cblist_dequeue(&rcl);
2673 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2674 debug_rcu_head_unqueue(rhp);
2675 if (__rcu_reclaim(rsp->name, rhp))
2676 rcu_cblist_dequeued_lazy(&rcl);
2677 /*
2678 * Stop only if limit reached and CPU has something to do.
2679 * Note: The rcl structure counts down from zero.
2680 */
2681 if (-rcl.len >= bl &&
2682 (need_resched() ||
2683 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2684 break;
2685 }
2686
2687 local_irq_save(flags);
2688 count = -rcl.len;
2689 trace_rcu_batch_end(rsp->name, count, !!rcl.head, need_resched(),
2690 is_idle_task(current), rcu_is_callbacks_kthread());
2691
2692 /* Update counts and requeue any remaining callbacks. */
2693 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2694 smp_mb(); /* List handling before counting for rcu_barrier(). */
2695 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2696
2697 /* Reinstate batch limit if we have worked down the excess. */
2698 count = rcu_segcblist_n_cbs(&rdp->cblist);
2699 if (rdp->blimit == LONG_MAX && count <= qlowmark)
2700 rdp->blimit = blimit;
2701
2702 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2703 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2704 rdp->qlen_last_fqs_check = 0;
2705 rdp->n_force_qs_snap = rsp->n_force_qs;
2706 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2707 rdp->qlen_last_fqs_check = count;
2708
2709 /*
2710 * The following usually indicates a double call_rcu(). To track
2711 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2712 */
2713 WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) != (count == 0));
2714
2715 local_irq_restore(flags);
2716
2717 /* Re-invoke RCU core processing if there are callbacks remaining. */
2718 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2719 invoke_rcu_core();
2720}
2721
2722/*
2723 * Check to see if this CPU is in a non-context-switch quiescent state
2724 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2725 * Also schedule RCU core processing.
2726 *
2727 * This function must be called from hardirq context. It is normally
2728 * invoked from the scheduling-clock interrupt.
2729 */
2730void rcu_check_callbacks(int user)
2731{
2732 trace_rcu_utilization(TPS("Start scheduler-tick"));
2733 increment_cpu_stall_ticks();
2734 if (user || rcu_is_cpu_rrupt_from_idle()) {
2735
2736 /*
2737 * Get here if this CPU took its interrupt from user
2738 * mode or from the idle loop, and if this is not a
2739 * nested interrupt. In this case, the CPU is in
2740 * a quiescent state, so note it.
2741 *
2742 * No memory barrier is required here because both
2743 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2744 * variables that other CPUs neither access nor modify,
2745 * at least not while the corresponding CPU is online.
2746 */
2747
2748 rcu_sched_qs();
2749 rcu_bh_qs();
2750
2751 } else if (!in_softirq()) {
2752
2753 /*
2754 * Get here if this CPU did not take its interrupt from
2755 * softirq, in other words, if it is not interrupting
2756 * a rcu_bh read-side critical section. This is an _bh
2757 * critical section, so note it.
2758 */
2759
2760 rcu_bh_qs();
2761 }
2762 rcu_preempt_check_callbacks();
2763 if (rcu_pending())
2764 invoke_rcu_core();
2765 if (user)
2766 rcu_note_voluntary_context_switch(current);
2767 trace_rcu_utilization(TPS("End scheduler-tick"));
2768}
2769
2770/*
2771 * Scan the leaf rcu_node structures, processing dyntick state for any that
2772 * have not yet encountered a quiescent state, using the function specified.
2773 * Also initiate boosting for any threads blocked on the root rcu_node.
2774 *
2775 * The caller must have suppressed start of new grace periods.
2776 */
2777static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp))
2778{
2779 int cpu;
2780 unsigned long flags;
2781 unsigned long mask;
2782 struct rcu_node *rnp;
2783
2784 rcu_for_each_leaf_node(rsp, rnp) {
2785 cond_resched_rcu_qs();
2786 mask = 0;
2787 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2788 if (rnp->qsmask == 0) {
2789 if (rcu_state_p == &rcu_sched_state ||
2790 rsp != rcu_state_p ||
2791 rcu_preempt_blocked_readers_cgp(rnp)) {
2792 /*
2793 * No point in scanning bits because they
2794 * are all zero. But we might need to
2795 * priority-boost blocked readers.
2796 */
2797 rcu_initiate_boost(rnp, flags);
2798 /* rcu_initiate_boost() releases rnp->lock */
2799 continue;
2800 }
2801 if (rnp->parent &&
2802 (rnp->parent->qsmask & rnp->grpmask)) {
2803 /*
2804 * Race between grace-period
2805 * initialization and task exiting RCU
2806 * read-side critical section: Report.
2807 */
2808 rcu_report_unblock_qs_rnp(rsp, rnp, flags);
2809 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2810 continue;
2811 }
2812 }
2813 for_each_leaf_node_possible_cpu(rnp, cpu) {
2814 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
2815 if ((rnp->qsmask & bit) != 0) {
2816 if (f(per_cpu_ptr(rsp->rda, cpu)))
2817 mask |= bit;
2818 }
2819 }
2820 if (mask != 0) {
2821 /* Idle/offline CPUs, report (releases rnp->lock. */
2822 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2823 } else {
2824 /* Nothing to do here, so just drop the lock. */
2825 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2826 }
2827 }
2828}
2829
2830/*
2831 * Force quiescent states on reluctant CPUs, and also detect which
2832 * CPUs are in dyntick-idle mode.
2833 */
2834static void force_quiescent_state(struct rcu_state *rsp)
2835{
2836 unsigned long flags;
2837 bool ret;
2838 struct rcu_node *rnp;
2839 struct rcu_node *rnp_old = NULL;
2840
2841 /* Funnel through hierarchy to reduce memory contention. */
2842 rnp = __this_cpu_read(rsp->rda->mynode);
2843 for (; rnp != NULL; rnp = rnp->parent) {
2844 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2845 !raw_spin_trylock(&rnp->fqslock);
2846 if (rnp_old != NULL)
2847 raw_spin_unlock(&rnp_old->fqslock);
2848 if (ret)
2849 return;
2850 rnp_old = rnp;
2851 }
2852 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2853
2854 /* Reached the root of the rcu_node tree, acquire lock. */
2855 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2856 raw_spin_unlock(&rnp_old->fqslock);
2857 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2858 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2859 return; /* Someone beat us to it. */
2860 }
2861 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2862 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2863 rcu_gp_kthread_wake(rsp);
2864}
2865
2866/*
2867 * This does the RCU core processing work for the specified rcu_state
2868 * and rcu_data structures. This may be called only from the CPU to
2869 * whom the rdp belongs.
2870 */
2871static void
2872__rcu_process_callbacks(struct rcu_state *rsp)
2873{
2874 unsigned long flags;
2875 bool needwake;
2876 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2877
2878 WARN_ON_ONCE(!rdp->beenonline);
2879
2880 /* Update RCU state based on any recent quiescent states. */
2881 rcu_check_quiescent_state(rsp, rdp);
2882
2883 /* Does this CPU require a not-yet-started grace period? */
2884 local_irq_save(flags);
2885 if (cpu_needs_another_gp(rsp, rdp)) {
2886 raw_spin_lock_rcu_node(rcu_get_root(rsp)); /* irqs disabled. */
2887 needwake = rcu_start_gp(rsp);
2888 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2889 if (needwake)
2890 rcu_gp_kthread_wake(rsp);
2891 } else {
2892 local_irq_restore(flags);
2893 }
2894
2895 /* If there are callbacks ready, invoke them. */
2896 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2897 invoke_rcu_callbacks(rsp, rdp);
2898
2899 /* Do any needed deferred wakeups of rcuo kthreads. */
2900 do_nocb_deferred_wakeup(rdp);
2901}
2902
2903/*
2904 * Do RCU core processing for the current CPU.
2905 */
2906static __latent_entropy void rcu_process_callbacks(struct softirq_action *unused)
2907{
2908 struct rcu_state *rsp;
2909
2910 if (cpu_is_offline(smp_processor_id()))
2911 return;
2912 trace_rcu_utilization(TPS("Start RCU core"));
2913 for_each_rcu_flavor(rsp)
2914 __rcu_process_callbacks(rsp);
2915 trace_rcu_utilization(TPS("End RCU core"));
2916}
2917
2918/*
2919 * Schedule RCU callback invocation. If the specified type of RCU
2920 * does not support RCU priority boosting, just do a direct call,
2921 * otherwise wake up the per-CPU kernel kthread. Note that because we
2922 * are running on the current CPU with softirqs disabled, the
2923 * rcu_cpu_kthread_task cannot disappear out from under us.
2924 */
2925static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2926{
2927 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
2928 return;
2929 if (likely(!rsp->boost)) {
2930 rcu_do_batch(rsp, rdp);
2931 return;
2932 }
2933 invoke_rcu_callbacks_kthread();
2934}
2935
2936static void invoke_rcu_core(void)
2937{
2938 if (cpu_online(smp_processor_id()))
2939 raise_softirq(RCU_SOFTIRQ);
2940}
2941
2942/*
2943 * Handle any core-RCU processing required by a call_rcu() invocation.
2944 */
2945static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2946 struct rcu_head *head, unsigned long flags)
2947{
2948 bool needwake;
2949
2950 /*
2951 * If called from an extended quiescent state, invoke the RCU
2952 * core in order to force a re-evaluation of RCU's idleness.
2953 */
2954 if (!rcu_is_watching())
2955 invoke_rcu_core();
2956
2957 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2958 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2959 return;
2960
2961 /*
2962 * Force the grace period if too many callbacks or too long waiting.
2963 * Enforce hysteresis, and don't invoke force_quiescent_state()
2964 * if some other CPU has recently done so. Also, don't bother
2965 * invoking force_quiescent_state() if the newly enqueued callback
2966 * is the only one waiting for a grace period to complete.
2967 */
2968 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2969 rdp->qlen_last_fqs_check + qhimark)) {
2970
2971 /* Are we ignoring a completed grace period? */
2972 note_gp_changes(rsp, rdp);
2973
2974 /* Start a new grace period if one not already started. */
2975 if (!rcu_gp_in_progress(rsp)) {
2976 struct rcu_node *rnp_root = rcu_get_root(rsp);
2977
2978 raw_spin_lock_rcu_node(rnp_root);
2979 needwake = rcu_start_gp(rsp);
2980 raw_spin_unlock_rcu_node(rnp_root);
2981 if (needwake)
2982 rcu_gp_kthread_wake(rsp);
2983 } else {
2984 /* Give the grace period a kick. */
2985 rdp->blimit = LONG_MAX;
2986 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2987 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2988 force_quiescent_state(rsp);
2989 rdp->n_force_qs_snap = rsp->n_force_qs;
2990 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2991 }
2992 }
2993}
2994
2995/*
2996 * RCU callback function to leak a callback.
2997 */
2998static void rcu_leak_callback(struct rcu_head *rhp)
2999{
3000}
3001
3002/*
3003 * Helper function for call_rcu() and friends. The cpu argument will
3004 * normally be -1, indicating "currently running CPU". It may specify
3005 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3006 * is expected to specify a CPU.
3007 */
3008static void
3009__call_rcu(struct rcu_head *head, rcu_callback_t func,
3010 struct rcu_state *rsp, int cpu, bool lazy)
3011{
3012 unsigned long flags;
3013 struct rcu_data *rdp;
3014
3015 /* Misaligned rcu_head! */
3016 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
3017
3018 if (debug_rcu_head_queue(head)) {
3019 /*
3020 * Probable double call_rcu(), so leak the callback.
3021 * Use rcu:rcu_callback trace event to find the previous
3022 * time callback was passed to __call_rcu().
3023 */
3024 WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pF()!!!\n",
3025 head, head->func);
3026 WRITE_ONCE(head->func, rcu_leak_callback);
3027 return;
3028 }
3029 head->func = func;
3030 head->next = NULL;
3031 local_irq_save(flags);
3032 rdp = this_cpu_ptr(rsp->rda);
3033
3034 /* Add the callback to our list. */
3035 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist)) || cpu != -1) {
3036 int offline;
3037
3038 if (cpu != -1)
3039 rdp = per_cpu_ptr(rsp->rda, cpu);
3040 if (likely(rdp->mynode)) {
3041 /* Post-boot, so this should be for a no-CBs CPU. */
3042 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
3043 WARN_ON_ONCE(offline);
3044 /* Offline CPU, _call_rcu() illegal, leak callback. */
3045 local_irq_restore(flags);
3046 return;
3047 }
3048 /*
3049 * Very early boot, before rcu_init(). Initialize if needed
3050 * and then drop through to queue the callback.
3051 */
3052 BUG_ON(cpu != -1);
3053 WARN_ON_ONCE(!rcu_is_watching());
3054 if (rcu_segcblist_empty(&rdp->cblist))
3055 rcu_segcblist_init(&rdp->cblist);
3056 }
3057 rcu_segcblist_enqueue(&rdp->cblist, head, lazy);
3058 if (!lazy)
3059 rcu_idle_count_callbacks_posted();
3060
3061 if (__is_kfree_rcu_offset((unsigned long)func))
3062 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
3063 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
3064 rcu_segcblist_n_cbs(&rdp->cblist));
3065 else
3066 trace_rcu_callback(rsp->name, head,
3067 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
3068 rcu_segcblist_n_cbs(&rdp->cblist));
3069
3070 /* Go handle any RCU core processing required. */
3071 __call_rcu_core(rsp, rdp, head, flags);
3072 local_irq_restore(flags);
3073}
3074
3075/**
3076 * call_rcu_sched() - Queue an RCU for invocation after sched grace period.
3077 * @head: structure to be used for queueing the RCU updates.
3078 * @func: actual callback function to be invoked after the grace period
3079 *
3080 * The callback function will be invoked some time after a full grace
3081 * period elapses, in other words after all currently executing RCU
3082 * read-side critical sections have completed. call_rcu_sched() assumes
3083 * that the read-side critical sections end on enabling of preemption
3084 * or on voluntary preemption.
3085 * RCU read-side critical sections are delimited by:
3086 *
3087 * - rcu_read_lock_sched() and rcu_read_unlock_sched(), OR
3088 * - anything that disables preemption.
3089 *
3090 * These may be nested.
3091 *
3092 * See the description of call_rcu() for more detailed information on
3093 * memory ordering guarantees.
3094 */
3095void call_rcu_sched(struct rcu_head *head, rcu_callback_t func)
3096{
3097 __call_rcu(head, func, &rcu_sched_state, -1, 0);
3098}
3099EXPORT_SYMBOL_GPL(call_rcu_sched);
3100
3101/**
3102 * call_rcu_bh() - Queue an RCU for invocation after a quicker grace period.
3103 * @head: structure to be used for queueing the RCU updates.
3104 * @func: actual callback function to be invoked after the grace period
3105 *
3106 * The callback function will be invoked some time after a full grace
3107 * period elapses, in other words after all currently executing RCU
3108 * read-side critical sections have completed. call_rcu_bh() assumes
3109 * that the read-side critical sections end on completion of a softirq
3110 * handler. This means that read-side critical sections in process
3111 * context must not be interrupted by softirqs. This interface is to be
3112 * used when most of the read-side critical sections are in softirq context.
3113 * RCU read-side critical sections are delimited by:
3114 *
3115 * - rcu_read_lock() and rcu_read_unlock(), if in interrupt context, OR
3116 * - rcu_read_lock_bh() and rcu_read_unlock_bh(), if in process context.
3117 *
3118 * These may be nested.
3119 *
3120 * See the description of call_rcu() for more detailed information on
3121 * memory ordering guarantees.
3122 */
3123void call_rcu_bh(struct rcu_head *head, rcu_callback_t func)
3124{
3125 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3126}
3127EXPORT_SYMBOL_GPL(call_rcu_bh);
3128
3129/*
3130 * Queue an RCU callback for lazy invocation after a grace period.
3131 * This will likely be later named something like "call_rcu_lazy()",
3132 * but this change will require some way of tagging the lazy RCU
3133 * callbacks in the list of pending callbacks. Until then, this
3134 * function may only be called from __kfree_rcu().
3135 */
3136void kfree_call_rcu(struct rcu_head *head,
3137 rcu_callback_t func)
3138{
3139 __call_rcu(head, func, rcu_state_p, -1, 1);
3140}
3141EXPORT_SYMBOL_GPL(kfree_call_rcu);
3142
3143/*
3144 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3145 * any blocking grace-period wait automatically implies a grace period
3146 * if there is only one CPU online at any point time during execution
3147 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3148 * occasionally incorrectly indicate that there are multiple CPUs online
3149 * when there was in fact only one the whole time, as this just adds
3150 * some overhead: RCU still operates correctly.
3151 */
3152static inline int rcu_blocking_is_gp(void)
3153{
3154 int ret;
3155
3156 might_sleep(); /* Check for RCU read-side critical section. */
3157 preempt_disable();
3158 ret = num_online_cpus() <= 1;
3159 preempt_enable();
3160 return ret;
3161}
3162
3163/**
3164 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3165 *
3166 * Control will return to the caller some time after a full rcu-sched
3167 * grace period has elapsed, in other words after all currently executing
3168 * rcu-sched read-side critical sections have completed. These read-side
3169 * critical sections are delimited by rcu_read_lock_sched() and
3170 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3171 * local_irq_disable(), and so on may be used in place of
3172 * rcu_read_lock_sched().
3173 *
3174 * This means that all preempt_disable code sequences, including NMI and
3175 * non-threaded hardware-interrupt handlers, in progress on entry will
3176 * have completed before this primitive returns. However, this does not
3177 * guarantee that softirq handlers will have completed, since in some
3178 * kernels, these handlers can run in process context, and can block.
3179 *
3180 * Note that this guarantee implies further memory-ordering guarantees.
3181 * On systems with more than one CPU, when synchronize_sched() returns,
3182 * each CPU is guaranteed to have executed a full memory barrier since the
3183 * end of its last RCU-sched read-side critical section whose beginning
3184 * preceded the call to synchronize_sched(). In addition, each CPU having
3185 * an RCU read-side critical section that extends beyond the return from
3186 * synchronize_sched() is guaranteed to have executed a full memory barrier
3187 * after the beginning of synchronize_sched() and before the beginning of
3188 * that RCU read-side critical section. Note that these guarantees include
3189 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3190 * that are executing in the kernel.
3191 *
3192 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3193 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3194 * to have executed a full memory barrier during the execution of
3195 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3196 * again only if the system has more than one CPU).
3197 */
3198void synchronize_sched(void)
3199{
3200 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3201 lock_is_held(&rcu_lock_map) ||
3202 lock_is_held(&rcu_sched_lock_map),
3203 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3204 if (rcu_blocking_is_gp())
3205 return;
3206 if (rcu_gp_is_expedited())
3207 synchronize_sched_expedited();
3208 else
3209 wait_rcu_gp(call_rcu_sched);
3210}
3211EXPORT_SYMBOL_GPL(synchronize_sched);
3212
3213/**
3214 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3215 *
3216 * Control will return to the caller some time after a full rcu_bh grace
3217 * period has elapsed, in other words after all currently executing rcu_bh
3218 * read-side critical sections have completed. RCU read-side critical
3219 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3220 * and may be nested.
3221 *
3222 * See the description of synchronize_sched() for more detailed information
3223 * on memory ordering guarantees.
3224 */
3225void synchronize_rcu_bh(void)
3226{
3227 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3228 lock_is_held(&rcu_lock_map) ||
3229 lock_is_held(&rcu_sched_lock_map),
3230 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3231 if (rcu_blocking_is_gp())
3232 return;
3233 if (rcu_gp_is_expedited())
3234 synchronize_rcu_bh_expedited();
3235 else
3236 wait_rcu_gp(call_rcu_bh);
3237}
3238EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3239
3240/**
3241 * get_state_synchronize_rcu - Snapshot current RCU state
3242 *
3243 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3244 * to determine whether or not a full grace period has elapsed in the
3245 * meantime.
3246 */
3247unsigned long get_state_synchronize_rcu(void)
3248{
3249 /*
3250 * Any prior manipulation of RCU-protected data must happen
3251 * before the load from ->gpnum.
3252 */
3253 smp_mb(); /* ^^^ */
3254
3255 /*
3256 * Make sure this load happens before the purportedly
3257 * time-consuming work between get_state_synchronize_rcu()
3258 * and cond_synchronize_rcu().
3259 */
3260 return smp_load_acquire(&rcu_state_p->gpnum);
3261}
3262EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3263
3264/**
3265 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3266 *
3267 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3268 *
3269 * If a full RCU grace period has elapsed since the earlier call to
3270 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3271 * synchronize_rcu() to wait for a full grace period.
3272 *
3273 * Yes, this function does not take counter wrap into account. But
3274 * counter wrap is harmless. If the counter wraps, we have waited for
3275 * more than 2 billion grace periods (and way more on a 64-bit system!),
3276 * so waiting for one additional grace period should be just fine.
3277 */
3278void cond_synchronize_rcu(unsigned long oldstate)
3279{
3280 unsigned long newstate;
3281
3282 /*
3283 * Ensure that this load happens before any RCU-destructive
3284 * actions the caller might carry out after we return.
3285 */
3286 newstate = smp_load_acquire(&rcu_state_p->completed);
3287 if (ULONG_CMP_GE(oldstate, newstate))
3288 synchronize_rcu();
3289}
3290EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3291
3292/**
3293 * get_state_synchronize_sched - Snapshot current RCU-sched state
3294 *
3295 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3296 * to determine whether or not a full grace period has elapsed in the
3297 * meantime.
3298 */
3299unsigned long get_state_synchronize_sched(void)
3300{
3301 /*
3302 * Any prior manipulation of RCU-protected data must happen
3303 * before the load from ->gpnum.
3304 */
3305 smp_mb(); /* ^^^ */
3306
3307 /*
3308 * Make sure this load happens before the purportedly
3309 * time-consuming work between get_state_synchronize_sched()
3310 * and cond_synchronize_sched().
3311 */
3312 return smp_load_acquire(&rcu_sched_state.gpnum);
3313}
3314EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3315
3316/**
3317 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3318 *
3319 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3320 *
3321 * If a full RCU-sched grace period has elapsed since the earlier call to
3322 * get_state_synchronize_sched(), just return. Otherwise, invoke
3323 * synchronize_sched() to wait for a full grace period.
3324 *
3325 * Yes, this function does not take counter wrap into account. But
3326 * counter wrap is harmless. If the counter wraps, we have waited for
3327 * more than 2 billion grace periods (and way more on a 64-bit system!),
3328 * so waiting for one additional grace period should be just fine.
3329 */
3330void cond_synchronize_sched(unsigned long oldstate)
3331{
3332 unsigned long newstate;
3333
3334 /*
3335 * Ensure that this load happens before any RCU-destructive
3336 * actions the caller might carry out after we return.
3337 */
3338 newstate = smp_load_acquire(&rcu_sched_state.completed);
3339 if (ULONG_CMP_GE(oldstate, newstate))
3340 synchronize_sched();
3341}
3342EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3343
3344/*
3345 * Check to see if there is any immediate RCU-related work to be done
3346 * by the current CPU, for the specified type of RCU, returning 1 if so.
3347 * The checks are in order of increasing expense: checks that can be
3348 * carried out against CPU-local state are performed first. However,
3349 * we must check for CPU stalls first, else we might not get a chance.
3350 */
3351static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3352{
3353 struct rcu_node *rnp = rdp->mynode;
3354
3355 /* Check for CPU stalls, if enabled. */
3356 check_cpu_stall(rsp, rdp);
3357
3358 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3359 if (rcu_nohz_full_cpu(rsp))
3360 return 0;
3361
3362 /* Is the RCU core waiting for a quiescent state from this CPU? */
3363 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm)
3364 return 1;
3365
3366 /* Does this CPU have callbacks ready to invoke? */
3367 if (rcu_segcblist_ready_cbs(&rdp->cblist))
3368 return 1;
3369
3370 /* Has RCU gone idle with this CPU needing another grace period? */
3371 if (cpu_needs_another_gp(rsp, rdp))
3372 return 1;
3373
3374 /* Has another RCU grace period completed? */
3375 if (READ_ONCE(rnp->completed) != rdp->completed) /* outside lock */
3376 return 1;
3377
3378 /* Has a new RCU grace period started? */
3379 if (READ_ONCE(rnp->gpnum) != rdp->gpnum ||
3380 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
3381 return 1;
3382
3383 /* Does this CPU need a deferred NOCB wakeup? */
3384 if (rcu_nocb_need_deferred_wakeup(rdp))
3385 return 1;
3386
3387 /* nothing to do */
3388 return 0;
3389}
3390
3391/*
3392 * Check to see if there is any immediate RCU-related work to be done
3393 * by the current CPU, returning 1 if so. This function is part of the
3394 * RCU implementation; it is -not- an exported member of the RCU API.
3395 */
3396static int rcu_pending(void)
3397{
3398 struct rcu_state *rsp;
3399
3400 for_each_rcu_flavor(rsp)
3401 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3402 return 1;
3403 return 0;
3404}
3405
3406/*
3407 * Return true if the specified CPU has any callback. If all_lazy is
3408 * non-NULL, store an indication of whether all callbacks are lazy.
3409 * (If there are no callbacks, all of them are deemed to be lazy.)
3410 */
3411static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3412{
3413 bool al = true;
3414 bool hc = false;
3415 struct rcu_data *rdp;
3416 struct rcu_state *rsp;
3417
3418 for_each_rcu_flavor(rsp) {
3419 rdp = this_cpu_ptr(rsp->rda);
3420 if (rcu_segcblist_empty(&rdp->cblist))
3421 continue;
3422 hc = true;
3423 if (rcu_segcblist_n_nonlazy_cbs(&rdp->cblist) || !all_lazy) {
3424 al = false;
3425 break;
3426 }
3427 }
3428 if (all_lazy)
3429 *all_lazy = al;
3430 return hc;
3431}
3432
3433/*
3434 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3435 * the compiler is expected to optimize this away.
3436 */
3437static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3438 int cpu, unsigned long done)
3439{
3440 trace_rcu_barrier(rsp->name, s, cpu,
3441 atomic_read(&rsp->barrier_cpu_count), done);
3442}
3443
3444/*
3445 * RCU callback function for _rcu_barrier(). If we are last, wake
3446 * up the task executing _rcu_barrier().
3447 */
3448static void rcu_barrier_callback(struct rcu_head *rhp)
3449{
3450 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3451 struct rcu_state *rsp = rdp->rsp;
3452
3453 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3454 _rcu_barrier_trace(rsp, TPS("LastCB"), -1,
3455 rsp->barrier_sequence);
3456 complete(&rsp->barrier_completion);
3457 } else {
3458 _rcu_barrier_trace(rsp, TPS("CB"), -1, rsp->barrier_sequence);
3459 }
3460}
3461
3462/*
3463 * Called with preemption disabled, and from cross-cpu IRQ context.
3464 */
3465static void rcu_barrier_func(void *type)
3466{
3467 struct rcu_state *rsp = type;
3468 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3469
3470 _rcu_barrier_trace(rsp, TPS("IRQ"), -1, rsp->barrier_sequence);
3471 rdp->barrier_head.func = rcu_barrier_callback;
3472 debug_rcu_head_queue(&rdp->barrier_head);
3473 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head, 0)) {
3474 atomic_inc(&rsp->barrier_cpu_count);
3475 } else {
3476 debug_rcu_head_unqueue(&rdp->barrier_head);
3477 _rcu_barrier_trace(rsp, TPS("IRQNQ"), -1,
3478 rsp->barrier_sequence);
3479 }
3480}
3481
3482/*
3483 * Orchestrate the specified type of RCU barrier, waiting for all
3484 * RCU callbacks of the specified type to complete.
3485 */
3486static void _rcu_barrier(struct rcu_state *rsp)
3487{
3488 int cpu;
3489 struct rcu_data *rdp;
3490 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
3491
3492 _rcu_barrier_trace(rsp, TPS("Begin"), -1, s);
3493
3494 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3495 mutex_lock(&rsp->barrier_mutex);
3496
3497 /* Did someone else do our work for us? */
3498 if (rcu_seq_done(&rsp->barrier_sequence, s)) {
3499 _rcu_barrier_trace(rsp, TPS("EarlyExit"), -1,
3500 rsp->barrier_sequence);
3501 smp_mb(); /* caller's subsequent code after above check. */
3502 mutex_unlock(&rsp->barrier_mutex);
3503 return;
3504 }
3505
3506 /* Mark the start of the barrier operation. */
3507 rcu_seq_start(&rsp->barrier_sequence);
3508 _rcu_barrier_trace(rsp, TPS("Inc1"), -1, rsp->barrier_sequence);
3509
3510 /*
3511 * Initialize the count to one rather than to zero in order to
3512 * avoid a too-soon return to zero in case of a short grace period
3513 * (or preemption of this task). Exclude CPU-hotplug operations
3514 * to ensure that no offline CPU has callbacks queued.
3515 */
3516 init_completion(&rsp->barrier_completion);
3517 atomic_set(&rsp->barrier_cpu_count, 1);
3518 get_online_cpus();
3519
3520 /*
3521 * Force each CPU with callbacks to register a new callback.
3522 * When that callback is invoked, we will know that all of the
3523 * corresponding CPU's preceding callbacks have been invoked.
3524 */
3525 for_each_possible_cpu(cpu) {
3526 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3527 continue;
3528 rdp = per_cpu_ptr(rsp->rda, cpu);
3529 if (rcu_is_nocb_cpu(cpu)) {
3530 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
3531 _rcu_barrier_trace(rsp, TPS("OfflineNoCB"), cpu,
3532 rsp->barrier_sequence);
3533 } else {
3534 _rcu_barrier_trace(rsp, TPS("OnlineNoCB"), cpu,
3535 rsp->barrier_sequence);
3536 smp_mb__before_atomic();
3537 atomic_inc(&rsp->barrier_cpu_count);
3538 __call_rcu(&rdp->barrier_head,
3539 rcu_barrier_callback, rsp, cpu, 0);
3540 }
3541 } else if (rcu_segcblist_n_cbs(&rdp->cblist)) {
3542 _rcu_barrier_trace(rsp, TPS("OnlineQ"), cpu,
3543 rsp->barrier_sequence);
3544 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3545 } else {
3546 _rcu_barrier_trace(rsp, TPS("OnlineNQ"), cpu,
3547 rsp->barrier_sequence);
3548 }
3549 }
3550 put_online_cpus();
3551
3552 /*
3553 * Now that we have an rcu_barrier_callback() callback on each
3554 * CPU, and thus each counted, remove the initial count.
3555 */
3556 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3557 complete(&rsp->barrier_completion);
3558
3559 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3560 wait_for_completion(&rsp->barrier_completion);
3561
3562 /* Mark the end of the barrier operation. */
3563 _rcu_barrier_trace(rsp, TPS("Inc2"), -1, rsp->barrier_sequence);
3564 rcu_seq_end(&rsp->barrier_sequence);
3565
3566 /* Other rcu_barrier() invocations can now safely proceed. */
3567 mutex_unlock(&rsp->barrier_mutex);
3568}
3569
3570/**
3571 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3572 */
3573void rcu_barrier_bh(void)
3574{
3575 _rcu_barrier(&rcu_bh_state);
3576}
3577EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3578
3579/**
3580 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3581 */
3582void rcu_barrier_sched(void)
3583{
3584 _rcu_barrier(&rcu_sched_state);
3585}
3586EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3587
3588/*
3589 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3590 * first CPU in a given leaf rcu_node structure coming online. The caller
3591 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3592 * disabled.
3593 */
3594static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3595{
3596 long mask;
3597 struct rcu_node *rnp = rnp_leaf;
3598
3599 raw_lockdep_assert_held_rcu_node(rnp);
3600 for (;;) {
3601 mask = rnp->grpmask;
3602 rnp = rnp->parent;
3603 if (rnp == NULL)
3604 return;
3605 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3606 rnp->qsmaskinit |= mask;
3607 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3608 }
3609}
3610
3611/*
3612 * Do boot-time initialization of a CPU's per-CPU RCU data.
3613 */
3614static void __init
3615rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3616{
3617 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3618
3619 /* Set up local state, ensuring consistent view of global state. */
3620 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3621 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3622 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != 1);
3623 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp->dynticks)));
3624 rdp->cpu = cpu;
3625 rdp->rsp = rsp;
3626 rcu_boot_init_nocb_percpu_data(rdp);
3627}
3628
3629/*
3630 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3631 * offline event can be happening at a given time. Note also that we
3632 * can accept some slop in the rsp->completed access due to the fact
3633 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3634 */
3635static void
3636rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3637{
3638 unsigned long flags;
3639 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3640 struct rcu_node *rnp = rcu_get_root(rsp);
3641
3642 /* Set up local state, ensuring consistent view of global state. */
3643 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3644 rdp->qlen_last_fqs_check = 0;
3645 rdp->n_force_qs_snap = rsp->n_force_qs;
3646 rdp->blimit = blimit;
3647 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3648 !init_nocb_callback_list(rdp))
3649 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
3650 rdp->dynticks->dynticks_nesting = 1; /* CPU not up, no tearing. */
3651 rcu_dynticks_eqs_online();
3652 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3653
3654 /*
3655 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3656 * propagation up the rcu_node tree will happen at the beginning
3657 * of the next grace period.
3658 */
3659 rnp = rdp->mynode;
3660 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3661 rdp->beenonline = true; /* We have now been online. */
3662 rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */
3663 rdp->completed = rnp->completed;
3664 rdp->cpu_no_qs.b.norm = true;
3665 rdp->rcu_qs_ctr_snap = per_cpu(rcu_dynticks.rcu_qs_ctr, cpu);
3666 rdp->core_needs_qs = false;
3667 rdp->rcu_iw_pending = false;
3668 rdp->rcu_iw_gpnum = rnp->gpnum - 1;
3669 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3670 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3671}
3672
3673/*
3674 * Invoked early in the CPU-online process, when pretty much all
3675 * services are available. The incoming CPU is not present.
3676 */
3677int rcutree_prepare_cpu(unsigned int cpu)
3678{
3679 struct rcu_state *rsp;
3680
3681 for_each_rcu_flavor(rsp)
3682 rcu_init_percpu_data(cpu, rsp);
3683
3684 rcu_prepare_kthreads(cpu);
3685 rcu_spawn_all_nocb_kthreads(cpu);
3686
3687 return 0;
3688}
3689
3690/*
3691 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3692 */
3693static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3694{
3695 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3696
3697 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3698}
3699
3700/*
3701 * Near the end of the CPU-online process. Pretty much all services
3702 * enabled, and the CPU is now very much alive.
3703 */
3704int rcutree_online_cpu(unsigned int cpu)
3705{
3706 unsigned long flags;
3707 struct rcu_data *rdp;
3708 struct rcu_node *rnp;
3709 struct rcu_state *rsp;
3710
3711 for_each_rcu_flavor(rsp) {
3712 rdp = per_cpu_ptr(rsp->rda, cpu);
3713 rnp = rdp->mynode;
3714 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3715 rnp->ffmask |= rdp->grpmask;
3716 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3717 }
3718 if (IS_ENABLED(CONFIG_TREE_SRCU))
3719 srcu_online_cpu(cpu);
3720 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
3721 return 0; /* Too early in boot for scheduler work. */
3722 sync_sched_exp_online_cleanup(cpu);
3723 rcutree_affinity_setting(cpu, -1);
3724 return 0;
3725}
3726
3727/*
3728 * Near the beginning of the process. The CPU is still very much alive
3729 * with pretty much all services enabled.
3730 */
3731int rcutree_offline_cpu(unsigned int cpu)
3732{
3733 unsigned long flags;
3734 struct rcu_data *rdp;
3735 struct rcu_node *rnp;
3736 struct rcu_state *rsp;
3737
3738 for_each_rcu_flavor(rsp) {
3739 rdp = per_cpu_ptr(rsp->rda, cpu);
3740 rnp = rdp->mynode;
3741 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3742 rnp->ffmask &= ~rdp->grpmask;
3743 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3744 }
3745
3746 rcutree_affinity_setting(cpu, cpu);
3747 if (IS_ENABLED(CONFIG_TREE_SRCU))
3748 srcu_offline_cpu(cpu);
3749 return 0;
3750}
3751
3752/*
3753 * Near the end of the offline process. We do only tracing here.
3754 */
3755int rcutree_dying_cpu(unsigned int cpu)
3756{
3757 struct rcu_state *rsp;
3758
3759 for_each_rcu_flavor(rsp)
3760 rcu_cleanup_dying_cpu(rsp);
3761 return 0;
3762}
3763
3764/*
3765 * The outgoing CPU is gone and we are running elsewhere.
3766 */
3767int rcutree_dead_cpu(unsigned int cpu)
3768{
3769 struct rcu_state *rsp;
3770
3771 for_each_rcu_flavor(rsp) {
3772 rcu_cleanup_dead_cpu(cpu, rsp);
3773 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
3774 }
3775 return 0;
3776}
3777
3778/*
3779 * Mark the specified CPU as being online so that subsequent grace periods
3780 * (both expedited and normal) will wait on it. Note that this means that
3781 * incoming CPUs are not allowed to use RCU read-side critical sections
3782 * until this function is called. Failing to observe this restriction
3783 * will result in lockdep splats.
3784 *
3785 * Note that this function is special in that it is invoked directly
3786 * from the incoming CPU rather than from the cpuhp_step mechanism.
3787 * This is because this function must be invoked at a precise location.
3788 */
3789void rcu_cpu_starting(unsigned int cpu)
3790{
3791 unsigned long flags;
3792 unsigned long mask;
3793 int nbits;
3794 unsigned long oldmask;
3795 struct rcu_data *rdp;
3796 struct rcu_node *rnp;
3797 struct rcu_state *rsp;
3798
3799 for_each_rcu_flavor(rsp) {
3800 rdp = per_cpu_ptr(rsp->rda, cpu);
3801 rnp = rdp->mynode;
3802 mask = rdp->grpmask;
3803 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3804 rnp->qsmaskinitnext |= mask;
3805 oldmask = rnp->expmaskinitnext;
3806 rnp->expmaskinitnext |= mask;
3807 oldmask ^= rnp->expmaskinitnext;
3808 nbits = bitmap_weight(&oldmask, BITS_PER_LONG);
3809 /* Allow lockless access for expedited grace periods. */
3810 smp_store_release(&rsp->ncpus, rsp->ncpus + nbits); /* ^^^ */
3811 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3812 }
3813 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
3814}
3815
3816#ifdef CONFIG_HOTPLUG_CPU
3817/*
3818 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
3819 * function. We now remove it from the rcu_node tree's ->qsmaskinit
3820 * bit masks.
3821 */
3822static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
3823{
3824 unsigned long flags;
3825 unsigned long mask;
3826 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3827 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3828
3829 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3830 mask = rdp->grpmask;
3831 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3832 rnp->qsmaskinitnext &= ~mask;
3833 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3834}
3835
3836/*
3837 * The outgoing function has no further need of RCU, so remove it from
3838 * the list of CPUs that RCU must track.
3839 *
3840 * Note that this function is special in that it is invoked directly
3841 * from the outgoing CPU rather than from the cpuhp_step mechanism.
3842 * This is because this function must be invoked at a precise location.
3843 */
3844void rcu_report_dead(unsigned int cpu)
3845{
3846 struct rcu_state *rsp;
3847
3848 /* QS for any half-done expedited RCU-sched GP. */
3849 preempt_disable();
3850 rcu_report_exp_rdp(&rcu_sched_state,
3851 this_cpu_ptr(rcu_sched_state.rda), true);
3852 preempt_enable();
3853 for_each_rcu_flavor(rsp)
3854 rcu_cleanup_dying_idle_cpu(cpu, rsp);
3855}
3856
3857/* Migrate the dead CPU's callbacks to the current CPU. */
3858static void rcu_migrate_callbacks(int cpu, struct rcu_state *rsp)
3859{
3860 unsigned long flags;
3861 struct rcu_data *my_rdp;
3862 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3863 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
3864
3865 if (rcu_is_nocb_cpu(cpu) || rcu_segcblist_empty(&rdp->cblist))
3866 return; /* No callbacks to migrate. */
3867
3868 local_irq_save(flags);
3869 my_rdp = this_cpu_ptr(rsp->rda);
3870 if (rcu_nocb_adopt_orphan_cbs(my_rdp, rdp, flags)) {
3871 local_irq_restore(flags);
3872 return;
3873 }
3874 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
3875 rcu_advance_cbs(rsp, rnp_root, rdp); /* Leverage recent GPs. */
3876 rcu_advance_cbs(rsp, rnp_root, my_rdp); /* Assign GP to pending CBs. */
3877 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
3878 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
3879 !rcu_segcblist_n_cbs(&my_rdp->cblist));
3880 raw_spin_unlock_irqrestore_rcu_node(rnp_root, flags);
3881 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
3882 !rcu_segcblist_empty(&rdp->cblist),
3883 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
3884 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
3885 rcu_segcblist_first_cb(&rdp->cblist));
3886}
3887
3888/*
3889 * The outgoing CPU has just passed through the dying-idle state,
3890 * and we are being invoked from the CPU that was IPIed to continue the
3891 * offline operation. We need to migrate the outgoing CPU's callbacks.
3892 */
3893void rcutree_migrate_callbacks(int cpu)
3894{
3895 struct rcu_state *rsp;
3896
3897 for_each_rcu_flavor(rsp)
3898 rcu_migrate_callbacks(cpu, rsp);
3899}
3900#endif
3901
3902/*
3903 * On non-huge systems, use expedited RCU grace periods to make suspend
3904 * and hibernation run faster.
3905 */
3906static int rcu_pm_notify(struct notifier_block *self,
3907 unsigned long action, void *hcpu)
3908{
3909 switch (action) {
3910 case PM_HIBERNATION_PREPARE:
3911 case PM_SUSPEND_PREPARE:
3912 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3913 rcu_expedite_gp();
3914 break;
3915 case PM_POST_HIBERNATION:
3916 case PM_POST_SUSPEND:
3917 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3918 rcu_unexpedite_gp();
3919 break;
3920 default:
3921 break;
3922 }
3923 return NOTIFY_OK;
3924}
3925
3926/*
3927 * Spawn the kthreads that handle each RCU flavor's grace periods.
3928 */
3929static int __init rcu_spawn_gp_kthread(void)
3930{
3931 unsigned long flags;
3932 int kthread_prio_in = kthread_prio;
3933 struct rcu_node *rnp;
3934 struct rcu_state *rsp;
3935 struct sched_param sp;
3936 struct task_struct *t;
3937
3938 /* Force priority into range. */
3939 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
3940 kthread_prio = 1;
3941 else if (kthread_prio < 0)
3942 kthread_prio = 0;
3943 else if (kthread_prio > 99)
3944 kthread_prio = 99;
3945 if (kthread_prio != kthread_prio_in)
3946 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3947 kthread_prio, kthread_prio_in);
3948
3949 rcu_scheduler_fully_active = 1;
3950 for_each_rcu_flavor(rsp) {
3951 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
3952 BUG_ON(IS_ERR(t));
3953 rnp = rcu_get_root(rsp);
3954 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3955 rsp->gp_kthread = t;
3956 if (kthread_prio) {
3957 sp.sched_priority = kthread_prio;
3958 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3959 }
3960 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3961 wake_up_process(t);
3962 }
3963 rcu_spawn_nocb_kthreads();
3964 rcu_spawn_boost_kthreads();
3965 return 0;
3966}
3967early_initcall(rcu_spawn_gp_kthread);
3968
3969/*
3970 * This function is invoked towards the end of the scheduler's
3971 * initialization process. Before this is called, the idle task might
3972 * contain synchronous grace-period primitives (during which time, this idle
3973 * task is booting the system, and such primitives are no-ops). After this
3974 * function is called, any synchronous grace-period primitives are run as
3975 * expedited, with the requesting task driving the grace period forward.
3976 * A later core_initcall() rcu_set_runtime_mode() will switch to full
3977 * runtime RCU functionality.
3978 */
3979void rcu_scheduler_starting(void)
3980{
3981 WARN_ON(num_online_cpus() != 1);
3982 WARN_ON(nr_context_switches() > 0);
3983 rcu_test_sync_prims();
3984 rcu_scheduler_active = RCU_SCHEDULER_INIT;
3985 rcu_test_sync_prims();
3986}
3987
3988/*
3989 * Helper function for rcu_init() that initializes one rcu_state structure.
3990 */
3991static void __init rcu_init_one(struct rcu_state *rsp)
3992{
3993 static const char * const buf[] = RCU_NODE_NAME_INIT;
3994 static const char * const fqs[] = RCU_FQS_NAME_INIT;
3995 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
3996 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
3997
3998 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
3999 int cpustride = 1;
4000 int i;
4001 int j;
4002 struct rcu_node *rnp;
4003
4004 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4005
4006 /* Silence gcc 4.8 false positive about array index out of range. */
4007 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4008 panic("rcu_init_one: rcu_num_lvls out of range");
4009
4010 /* Initialize the level-tracking arrays. */
4011
4012 for (i = 1; i < rcu_num_lvls; i++)
4013 rsp->level[i] = rsp->level[i - 1] + num_rcu_lvl[i - 1];
4014 rcu_init_levelspread(levelspread, num_rcu_lvl);
4015
4016 /* Initialize the elements themselves, starting from the leaves. */
4017
4018 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4019 cpustride *= levelspread[i];
4020 rnp = rsp->level[i];
4021 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
4022 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4023 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4024 &rcu_node_class[i], buf[i]);
4025 raw_spin_lock_init(&rnp->fqslock);
4026 lockdep_set_class_and_name(&rnp->fqslock,
4027 &rcu_fqs_class[i], fqs[i]);
4028 rnp->gpnum = rsp->gpnum;
4029 rnp->completed = rsp->completed;
4030 rnp->qsmask = 0;
4031 rnp->qsmaskinit = 0;
4032 rnp->grplo = j * cpustride;
4033 rnp->grphi = (j + 1) * cpustride - 1;
4034 if (rnp->grphi >= nr_cpu_ids)
4035 rnp->grphi = nr_cpu_ids - 1;
4036 if (i == 0) {
4037 rnp->grpnum = 0;
4038 rnp->grpmask = 0;
4039 rnp->parent = NULL;
4040 } else {
4041 rnp->grpnum = j % levelspread[i - 1];
4042 rnp->grpmask = 1UL << rnp->grpnum;
4043 rnp->parent = rsp->level[i - 1] +
4044 j / levelspread[i - 1];
4045 }
4046 rnp->level = i;
4047 INIT_LIST_HEAD(&rnp->blkd_tasks);
4048 rcu_init_one_nocb(rnp);
4049 init_waitqueue_head(&rnp->exp_wq[0]);
4050 init_waitqueue_head(&rnp->exp_wq[1]);
4051 init_waitqueue_head(&rnp->exp_wq[2]);
4052 init_waitqueue_head(&rnp->exp_wq[3]);
4053 spin_lock_init(&rnp->exp_lock);
4054 }
4055 }
4056
4057 init_swait_queue_head(&rsp->gp_wq);
4058 init_swait_queue_head(&rsp->expedited_wq);
4059 rnp = rsp->level[rcu_num_lvls - 1];
4060 for_each_possible_cpu(i) {
4061 while (i > rnp->grphi)
4062 rnp++;
4063 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
4064 rcu_boot_init_percpu_data(i, rsp);
4065 }
4066 list_add(&rsp->flavors, &rcu_struct_flavors);
4067}
4068
4069/*
4070 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4071 * replace the definitions in tree.h because those are needed to size
4072 * the ->node array in the rcu_state structure.
4073 */
4074static void __init rcu_init_geometry(void)
4075{
4076 ulong d;
4077 int i;
4078 int rcu_capacity[RCU_NUM_LVLS];
4079
4080 /*
4081 * Initialize any unspecified boot parameters.
4082 * The default values of jiffies_till_first_fqs and
4083 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4084 * value, which is a function of HZ, then adding one for each
4085 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4086 */
4087 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4088 if (jiffies_till_first_fqs == ULONG_MAX)
4089 jiffies_till_first_fqs = d;
4090 if (jiffies_till_next_fqs == ULONG_MAX)
4091 jiffies_till_next_fqs = d;
4092
4093 /* If the compile-time values are accurate, just leave. */
4094 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4095 nr_cpu_ids == NR_CPUS)
4096 return;
4097 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4098 rcu_fanout_leaf, nr_cpu_ids);
4099
4100 /*
4101 * The boot-time rcu_fanout_leaf parameter must be at least two
4102 * and cannot exceed the number of bits in the rcu_node masks.
4103 * Complain and fall back to the compile-time values if this
4104 * limit is exceeded.
4105 */
4106 if (rcu_fanout_leaf < 2 ||
4107 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4108 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4109 WARN_ON(1);
4110 return;
4111 }
4112
4113 /*
4114 * Compute number of nodes that can be handled an rcu_node tree
4115 * with the given number of levels.
4116 */
4117 rcu_capacity[0] = rcu_fanout_leaf;
4118 for (i = 1; i < RCU_NUM_LVLS; i++)
4119 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4120
4121 /*
4122 * The tree must be able to accommodate the configured number of CPUs.
4123 * If this limit is exceeded, fall back to the compile-time values.
4124 */
4125 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4126 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4127 WARN_ON(1);
4128 return;
4129 }
4130
4131 /* Calculate the number of levels in the tree. */
4132 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4133 }
4134 rcu_num_lvls = i + 1;
4135
4136 /* Calculate the number of rcu_nodes at each level of the tree. */
4137 for (i = 0; i < rcu_num_lvls; i++) {
4138 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4139 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4140 }
4141
4142 /* Calculate the total number of rcu_node structures. */
4143 rcu_num_nodes = 0;
4144 for (i = 0; i < rcu_num_lvls; i++)
4145 rcu_num_nodes += num_rcu_lvl[i];
4146}
4147
4148/*
4149 * Dump out the structure of the rcu_node combining tree associated
4150 * with the rcu_state structure referenced by rsp.
4151 */
4152static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4153{
4154 int level = 0;
4155 struct rcu_node *rnp;
4156
4157 pr_info("rcu_node tree layout dump\n");
4158 pr_info(" ");
4159 rcu_for_each_node_breadth_first(rsp, rnp) {
4160 if (rnp->level != level) {
4161 pr_cont("\n");
4162 pr_info(" ");
4163 level = rnp->level;
4164 }
4165 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4166 }
4167 pr_cont("\n");
4168}
4169
4170struct workqueue_struct *rcu_gp_wq;
4171
4172void __init rcu_init(void)
4173{
4174 int cpu;
4175
4176 rcu_early_boot_tests();
4177
4178 rcu_bootup_announce();
4179 rcu_init_geometry();
4180 rcu_init_one(&rcu_bh_state);
4181 rcu_init_one(&rcu_sched_state);
4182 if (dump_tree)
4183 rcu_dump_rcu_node_tree(&rcu_sched_state);
4184 __rcu_init_preempt();
4185 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4186
4187 /*
4188 * We don't need protection against CPU-hotplug here because
4189 * this is called early in boot, before either interrupts
4190 * or the scheduler are operational.
4191 */
4192 pm_notifier(rcu_pm_notify, 0);
4193 for_each_online_cpu(cpu) {
4194 rcutree_prepare_cpu(cpu);
4195 rcu_cpu_starting(cpu);
4196 rcutree_online_cpu(cpu);
4197 }
4198
4199 /* Create workqueue for expedited GPs and for Tree SRCU. */
4200 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
4201 WARN_ON(!rcu_gp_wq);
4202}
4203
4204#include "tree_exp.h"
4205#include "tree_plugin.h"