Loading...
1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * Read-Copy Update mechanism for mutual exclusion
4 *
5 * Copyright IBM Corporation, 2008
6 *
7 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
8 * Manfred Spraul <manfred@colorfullife.com>
9 * Paul E. McKenney <paulmck@linux.ibm.com> Hierarchical version
10 *
11 * Based on the original work by Paul McKenney <paulmck@linux.ibm.com>
12 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
13 *
14 * For detailed explanation of Read-Copy Update mechanism see -
15 * Documentation/RCU
16 */
17
18#define pr_fmt(fmt) "rcu: " fmt
19
20#include <linux/types.h>
21#include <linux/kernel.h>
22#include <linux/init.h>
23#include <linux/spinlock.h>
24#include <linux/smp.h>
25#include <linux/rcupdate_wait.h>
26#include <linux/interrupt.h>
27#include <linux/sched.h>
28#include <linux/sched/debug.h>
29#include <linux/nmi.h>
30#include <linux/atomic.h>
31#include <linux/bitops.h>
32#include <linux/export.h>
33#include <linux/completion.h>
34#include <linux/moduleparam.h>
35#include <linux/percpu.h>
36#include <linux/notifier.h>
37#include <linux/cpu.h>
38#include <linux/mutex.h>
39#include <linux/time.h>
40#include <linux/kernel_stat.h>
41#include <linux/wait.h>
42#include <linux/kthread.h>
43#include <uapi/linux/sched/types.h>
44#include <linux/prefetch.h>
45#include <linux/delay.h>
46#include <linux/stop_machine.h>
47#include <linux/random.h>
48#include <linux/trace_events.h>
49#include <linux/suspend.h>
50#include <linux/ftrace.h>
51#include <linux/tick.h>
52#include <linux/sysrq.h>
53#include <linux/kprobes.h>
54#include <linux/gfp.h>
55#include <linux/oom.h>
56#include <linux/smpboot.h>
57#include <linux/jiffies.h>
58#include <linux/sched/isolation.h>
59#include <linux/sched/clock.h>
60#include "../time/tick-internal.h"
61
62#include "tree.h"
63#include "rcu.h"
64
65#ifdef MODULE_PARAM_PREFIX
66#undef MODULE_PARAM_PREFIX
67#endif
68#define MODULE_PARAM_PREFIX "rcutree."
69
70/* Data structures. */
71
72/*
73 * Steal a bit from the bottom of ->dynticks for idle entry/exit
74 * control. Initially this is for TLB flushing.
75 */
76#define RCU_DYNTICK_CTRL_MASK 0x1
77#define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
78#ifndef rcu_eqs_special_exit
79#define rcu_eqs_special_exit() do { } while (0)
80#endif
81
82static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = {
83 .dynticks_nesting = 1,
84 .dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE,
85 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
86};
87struct rcu_state rcu_state = {
88 .level = { &rcu_state.node[0] },
89 .gp_state = RCU_GP_IDLE,
90 .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT,
91 .barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex),
92 .name = RCU_NAME,
93 .abbr = RCU_ABBR,
94 .exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
95 .exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
96 .ofl_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.ofl_lock),
97};
98
99/* Dump rcu_node combining tree at boot to verify correct setup. */
100static bool dump_tree;
101module_param(dump_tree, bool, 0444);
102/* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */
103static bool use_softirq = 1;
104module_param(use_softirq, bool, 0444);
105/* Control rcu_node-tree auto-balancing at boot time. */
106static bool rcu_fanout_exact;
107module_param(rcu_fanout_exact, bool, 0444);
108/* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
109static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
110module_param(rcu_fanout_leaf, int, 0444);
111int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
112/* Number of rcu_nodes at specified level. */
113int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
114int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
115
116/*
117 * The rcu_scheduler_active variable is initialized to the value
118 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
119 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
120 * RCU can assume that there is but one task, allowing RCU to (for example)
121 * optimize synchronize_rcu() to a simple barrier(). When this variable
122 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
123 * to detect real grace periods. This variable is also used to suppress
124 * boot-time false positives from lockdep-RCU error checking. Finally, it
125 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
126 * is fully initialized, including all of its kthreads having been spawned.
127 */
128int rcu_scheduler_active __read_mostly;
129EXPORT_SYMBOL_GPL(rcu_scheduler_active);
130
131/*
132 * The rcu_scheduler_fully_active variable transitions from zero to one
133 * during the early_initcall() processing, which is after the scheduler
134 * is capable of creating new tasks. So RCU processing (for example,
135 * creating tasks for RCU priority boosting) must be delayed until after
136 * rcu_scheduler_fully_active transitions from zero to one. We also
137 * currently delay invocation of any RCU callbacks until after this point.
138 *
139 * It might later prove better for people registering RCU callbacks during
140 * early boot to take responsibility for these callbacks, but one step at
141 * a time.
142 */
143static int rcu_scheduler_fully_active __read_mostly;
144
145static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
146 unsigned long gps, unsigned long flags);
147static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
148static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
149static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
150static void invoke_rcu_core(void);
151static void rcu_report_exp_rdp(struct rcu_data *rdp);
152static void sync_sched_exp_online_cleanup(int cpu);
153
154/* rcuc/rcub kthread realtime priority */
155static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
156module_param(kthread_prio, int, 0444);
157
158/* Delay in jiffies for grace-period initialization delays, debug only. */
159
160static int gp_preinit_delay;
161module_param(gp_preinit_delay, int, 0444);
162static int gp_init_delay;
163module_param(gp_init_delay, int, 0444);
164static int gp_cleanup_delay;
165module_param(gp_cleanup_delay, int, 0444);
166
167/* Retrieve RCU kthreads priority for rcutorture */
168int rcu_get_gp_kthreads_prio(void)
169{
170 return kthread_prio;
171}
172EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio);
173
174/*
175 * Number of grace periods between delays, normalized by the duration of
176 * the delay. The longer the delay, the more the grace periods between
177 * each delay. The reason for this normalization is that it means that,
178 * for non-zero delays, the overall slowdown of grace periods is constant
179 * regardless of the duration of the delay. This arrangement balances
180 * the need for long delays to increase some race probabilities with the
181 * need for fast grace periods to increase other race probabilities.
182 */
183#define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
184
185/*
186 * Compute the mask of online CPUs for the specified rcu_node structure.
187 * This will not be stable unless the rcu_node structure's ->lock is
188 * held, but the bit corresponding to the current CPU will be stable
189 * in most contexts.
190 */
191unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
192{
193 return READ_ONCE(rnp->qsmaskinitnext);
194}
195
196/*
197 * Return true if an RCU grace period is in progress. The READ_ONCE()s
198 * permit this function to be invoked without holding the root rcu_node
199 * structure's ->lock, but of course results can be subject to change.
200 */
201static int rcu_gp_in_progress(void)
202{
203 return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq));
204}
205
206/*
207 * Return the number of callbacks queued on the specified CPU.
208 * Handles both the nocbs and normal cases.
209 */
210static long rcu_get_n_cbs_cpu(int cpu)
211{
212 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
213
214 if (rcu_segcblist_is_enabled(&rdp->cblist))
215 return rcu_segcblist_n_cbs(&rdp->cblist);
216 return 0;
217}
218
219void rcu_softirq_qs(void)
220{
221 rcu_qs();
222 rcu_preempt_deferred_qs(current);
223}
224
225/*
226 * Record entry into an extended quiescent state. This is only to be
227 * called when not already in an extended quiescent state.
228 */
229static void rcu_dynticks_eqs_enter(void)
230{
231 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
232 int seq;
233
234 /*
235 * CPUs seeing atomic_add_return() must see prior RCU read-side
236 * critical sections, and we also must force ordering with the
237 * next idle sojourn.
238 */
239 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
240 /* Better be in an extended quiescent state! */
241 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
242 (seq & RCU_DYNTICK_CTRL_CTR));
243 /* Better not have special action (TLB flush) pending! */
244 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
245 (seq & RCU_DYNTICK_CTRL_MASK));
246}
247
248/*
249 * Record exit from an extended quiescent state. This is only to be
250 * called from an extended quiescent state.
251 */
252static void rcu_dynticks_eqs_exit(void)
253{
254 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
255 int seq;
256
257 /*
258 * CPUs seeing atomic_add_return() must see prior idle sojourns,
259 * and we also must force ordering with the next RCU read-side
260 * critical section.
261 */
262 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
263 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
264 !(seq & RCU_DYNTICK_CTRL_CTR));
265 if (seq & RCU_DYNTICK_CTRL_MASK) {
266 atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdp->dynticks);
267 smp_mb__after_atomic(); /* _exit after clearing mask. */
268 /* Prefer duplicate flushes to losing a flush. */
269 rcu_eqs_special_exit();
270 }
271}
272
273/*
274 * Reset the current CPU's ->dynticks counter to indicate that the
275 * newly onlined CPU is no longer in an extended quiescent state.
276 * This will either leave the counter unchanged, or increment it
277 * to the next non-quiescent value.
278 *
279 * The non-atomic test/increment sequence works because the upper bits
280 * of the ->dynticks counter are manipulated only by the corresponding CPU,
281 * or when the corresponding CPU is offline.
282 */
283static void rcu_dynticks_eqs_online(void)
284{
285 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
286
287 if (atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR)
288 return;
289 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
290}
291
292/*
293 * Is the current CPU in an extended quiescent state?
294 *
295 * No ordering, as we are sampling CPU-local information.
296 */
297bool rcu_dynticks_curr_cpu_in_eqs(void)
298{
299 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
300
301 return !(atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR);
302}
303
304/*
305 * Snapshot the ->dynticks counter with full ordering so as to allow
306 * stable comparison of this counter with past and future snapshots.
307 */
308int rcu_dynticks_snap(struct rcu_data *rdp)
309{
310 int snap = atomic_add_return(0, &rdp->dynticks);
311
312 return snap & ~RCU_DYNTICK_CTRL_MASK;
313}
314
315/*
316 * Return true if the snapshot returned from rcu_dynticks_snap()
317 * indicates that RCU is in an extended quiescent state.
318 */
319static bool rcu_dynticks_in_eqs(int snap)
320{
321 return !(snap & RCU_DYNTICK_CTRL_CTR);
322}
323
324/*
325 * Return true if the CPU corresponding to the specified rcu_data
326 * structure has spent some time in an extended quiescent state since
327 * rcu_dynticks_snap() returned the specified snapshot.
328 */
329static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap)
330{
331 return snap != rcu_dynticks_snap(rdp);
332}
333
334/*
335 * Set the special (bottom) bit of the specified CPU so that it
336 * will take special action (such as flushing its TLB) on the
337 * next exit from an extended quiescent state. Returns true if
338 * the bit was successfully set, or false if the CPU was not in
339 * an extended quiescent state.
340 */
341bool rcu_eqs_special_set(int cpu)
342{
343 int old;
344 int new;
345 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
346
347 do {
348 old = atomic_read(&rdp->dynticks);
349 if (old & RCU_DYNTICK_CTRL_CTR)
350 return false;
351 new = old | RCU_DYNTICK_CTRL_MASK;
352 } while (atomic_cmpxchg(&rdp->dynticks, old, new) != old);
353 return true;
354}
355
356/*
357 * Let the RCU core know that this CPU has gone through the scheduler,
358 * which is a quiescent state. This is called when the need for a
359 * quiescent state is urgent, so we burn an atomic operation and full
360 * memory barriers to let the RCU core know about it, regardless of what
361 * this CPU might (or might not) do in the near future.
362 *
363 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
364 *
365 * The caller must have disabled interrupts and must not be idle.
366 */
367static void __maybe_unused rcu_momentary_dyntick_idle(void)
368{
369 int special;
370
371 raw_cpu_write(rcu_data.rcu_need_heavy_qs, false);
372 special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
373 &this_cpu_ptr(&rcu_data)->dynticks);
374 /* It is illegal to call this from idle state. */
375 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
376 rcu_preempt_deferred_qs(current);
377}
378
379/**
380 * rcu_is_cpu_rrupt_from_idle - see if interrupted from idle
381 *
382 * If the current CPU is idle and running at a first-level (not nested)
383 * interrupt from idle, return true. The caller must have at least
384 * disabled preemption.
385 */
386static int rcu_is_cpu_rrupt_from_idle(void)
387{
388 /* Called only from within the scheduling-clock interrupt */
389 lockdep_assert_in_irq();
390
391 /* Check for counter underflows */
392 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) < 0,
393 "RCU dynticks_nesting counter underflow!");
394 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) <= 0,
395 "RCU dynticks_nmi_nesting counter underflow/zero!");
396
397 /* Are we at first interrupt nesting level? */
398 if (__this_cpu_read(rcu_data.dynticks_nmi_nesting) != 1)
399 return false;
400
401 /* Does CPU appear to be idle from an RCU standpoint? */
402 return __this_cpu_read(rcu_data.dynticks_nesting) == 0;
403}
404
405#define DEFAULT_RCU_BLIMIT 10 /* Maximum callbacks per rcu_do_batch ... */
406#define DEFAULT_MAX_RCU_BLIMIT 10000 /* ... even during callback flood. */
407static long blimit = DEFAULT_RCU_BLIMIT;
408#define DEFAULT_RCU_QHIMARK 10000 /* If this many pending, ignore blimit. */
409static long qhimark = DEFAULT_RCU_QHIMARK;
410#define DEFAULT_RCU_QLOMARK 100 /* Once only this many pending, use blimit. */
411static long qlowmark = DEFAULT_RCU_QLOMARK;
412
413module_param(blimit, long, 0444);
414module_param(qhimark, long, 0444);
415module_param(qlowmark, long, 0444);
416
417static ulong jiffies_till_first_fqs = ULONG_MAX;
418static ulong jiffies_till_next_fqs = ULONG_MAX;
419static bool rcu_kick_kthreads;
420static int rcu_divisor = 7;
421module_param(rcu_divisor, int, 0644);
422
423/* Force an exit from rcu_do_batch() after 3 milliseconds. */
424static long rcu_resched_ns = 3 * NSEC_PER_MSEC;
425module_param(rcu_resched_ns, long, 0644);
426
427/*
428 * How long the grace period must be before we start recruiting
429 * quiescent-state help from rcu_note_context_switch().
430 */
431static ulong jiffies_till_sched_qs = ULONG_MAX;
432module_param(jiffies_till_sched_qs, ulong, 0444);
433static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */
434module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
435
436/*
437 * Make sure that we give the grace-period kthread time to detect any
438 * idle CPUs before taking active measures to force quiescent states.
439 * However, don't go below 100 milliseconds, adjusted upwards for really
440 * large systems.
441 */
442static void adjust_jiffies_till_sched_qs(void)
443{
444 unsigned long j;
445
446 /* If jiffies_till_sched_qs was specified, respect the request. */
447 if (jiffies_till_sched_qs != ULONG_MAX) {
448 WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
449 return;
450 }
451 /* Otherwise, set to third fqs scan, but bound below on large system. */
452 j = READ_ONCE(jiffies_till_first_fqs) +
453 2 * READ_ONCE(jiffies_till_next_fqs);
454 if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
455 j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
456 pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
457 WRITE_ONCE(jiffies_to_sched_qs, j);
458}
459
460static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
461{
462 ulong j;
463 int ret = kstrtoul(val, 0, &j);
464
465 if (!ret) {
466 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
467 adjust_jiffies_till_sched_qs();
468 }
469 return ret;
470}
471
472static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
473{
474 ulong j;
475 int ret = kstrtoul(val, 0, &j);
476
477 if (!ret) {
478 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
479 adjust_jiffies_till_sched_qs();
480 }
481 return ret;
482}
483
484static struct kernel_param_ops first_fqs_jiffies_ops = {
485 .set = param_set_first_fqs_jiffies,
486 .get = param_get_ulong,
487};
488
489static struct kernel_param_ops next_fqs_jiffies_ops = {
490 .set = param_set_next_fqs_jiffies,
491 .get = param_get_ulong,
492};
493
494module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
495module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
496module_param(rcu_kick_kthreads, bool, 0644);
497
498static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
499static int rcu_pending(void);
500
501/*
502 * Return the number of RCU GPs completed thus far for debug & stats.
503 */
504unsigned long rcu_get_gp_seq(void)
505{
506 return READ_ONCE(rcu_state.gp_seq);
507}
508EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
509
510/*
511 * Return the number of RCU expedited batches completed thus far for
512 * debug & stats. Odd numbers mean that a batch is in progress, even
513 * numbers mean idle. The value returned will thus be roughly double
514 * the cumulative batches since boot.
515 */
516unsigned long rcu_exp_batches_completed(void)
517{
518 return rcu_state.expedited_sequence;
519}
520EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
521
522/*
523 * Return the root node of the rcu_state structure.
524 */
525static struct rcu_node *rcu_get_root(void)
526{
527 return &rcu_state.node[0];
528}
529
530/*
531 * Convert a ->gp_state value to a character string.
532 */
533static const char *gp_state_getname(short gs)
534{
535 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
536 return "???";
537 return gp_state_names[gs];
538}
539
540/*
541 * Send along grace-period-related data for rcutorture diagnostics.
542 */
543void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
544 unsigned long *gp_seq)
545{
546 switch (test_type) {
547 case RCU_FLAVOR:
548 *flags = READ_ONCE(rcu_state.gp_flags);
549 *gp_seq = rcu_seq_current(&rcu_state.gp_seq);
550 break;
551 default:
552 break;
553 }
554}
555EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
556
557/*
558 * Enter an RCU extended quiescent state, which can be either the
559 * idle loop or adaptive-tickless usermode execution.
560 *
561 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
562 * the possibility of usermode upcalls having messed up our count
563 * of interrupt nesting level during the prior busy period.
564 */
565static void rcu_eqs_enter(bool user)
566{
567 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
568
569 WARN_ON_ONCE(rdp->dynticks_nmi_nesting != DYNTICK_IRQ_NONIDLE);
570 WRITE_ONCE(rdp->dynticks_nmi_nesting, 0);
571 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
572 rdp->dynticks_nesting == 0);
573 if (rdp->dynticks_nesting != 1) {
574 rdp->dynticks_nesting--;
575 return;
576 }
577
578 lockdep_assert_irqs_disabled();
579 trace_rcu_dyntick(TPS("Start"), rdp->dynticks_nesting, 0, rdp->dynticks);
580 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
581 rdp = this_cpu_ptr(&rcu_data);
582 do_nocb_deferred_wakeup(rdp);
583 rcu_prepare_for_idle();
584 rcu_preempt_deferred_qs(current);
585 WRITE_ONCE(rdp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
586 rcu_dynticks_eqs_enter();
587 rcu_dynticks_task_enter();
588}
589
590/**
591 * rcu_idle_enter - inform RCU that current CPU is entering idle
592 *
593 * Enter idle mode, in other words, -leave- the mode in which RCU
594 * read-side critical sections can occur. (Though RCU read-side
595 * critical sections can occur in irq handlers in idle, a possibility
596 * handled by irq_enter() and irq_exit().)
597 *
598 * If you add or remove a call to rcu_idle_enter(), be sure to test with
599 * CONFIG_RCU_EQS_DEBUG=y.
600 */
601void rcu_idle_enter(void)
602{
603 lockdep_assert_irqs_disabled();
604 rcu_eqs_enter(false);
605}
606
607#ifdef CONFIG_NO_HZ_FULL
608/**
609 * rcu_user_enter - inform RCU that we are resuming userspace.
610 *
611 * Enter RCU idle mode right before resuming userspace. No use of RCU
612 * is permitted between this call and rcu_user_exit(). This way the
613 * CPU doesn't need to maintain the tick for RCU maintenance purposes
614 * when the CPU runs in userspace.
615 *
616 * If you add or remove a call to rcu_user_enter(), be sure to test with
617 * CONFIG_RCU_EQS_DEBUG=y.
618 */
619void rcu_user_enter(void)
620{
621 lockdep_assert_irqs_disabled();
622 rcu_eqs_enter(true);
623}
624#endif /* CONFIG_NO_HZ_FULL */
625
626/*
627 * If we are returning from the outermost NMI handler that interrupted an
628 * RCU-idle period, update rdp->dynticks and rdp->dynticks_nmi_nesting
629 * to let the RCU grace-period handling know that the CPU is back to
630 * being RCU-idle.
631 *
632 * If you add or remove a call to rcu_nmi_exit_common(), be sure to test
633 * with CONFIG_RCU_EQS_DEBUG=y.
634 */
635static __always_inline void rcu_nmi_exit_common(bool irq)
636{
637 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
638
639 /*
640 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
641 * (We are exiting an NMI handler, so RCU better be paying attention
642 * to us!)
643 */
644 WARN_ON_ONCE(rdp->dynticks_nmi_nesting <= 0);
645 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
646
647 /*
648 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
649 * leave it in non-RCU-idle state.
650 */
651 if (rdp->dynticks_nmi_nesting != 1) {
652 trace_rcu_dyntick(TPS("--="), rdp->dynticks_nmi_nesting, rdp->dynticks_nmi_nesting - 2, rdp->dynticks);
653 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* No store tearing. */
654 rdp->dynticks_nmi_nesting - 2);
655 return;
656 }
657
658 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
659 trace_rcu_dyntick(TPS("Startirq"), rdp->dynticks_nmi_nesting, 0, rdp->dynticks);
660 WRITE_ONCE(rdp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
661
662 if (irq)
663 rcu_prepare_for_idle();
664
665 rcu_dynticks_eqs_enter();
666
667 if (irq)
668 rcu_dynticks_task_enter();
669}
670
671/**
672 * rcu_nmi_exit - inform RCU of exit from NMI context
673 *
674 * If you add or remove a call to rcu_nmi_exit(), be sure to test
675 * with CONFIG_RCU_EQS_DEBUG=y.
676 */
677void rcu_nmi_exit(void)
678{
679 rcu_nmi_exit_common(false);
680}
681
682/**
683 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
684 *
685 * Exit from an interrupt handler, which might possibly result in entering
686 * idle mode, in other words, leaving the mode in which read-side critical
687 * sections can occur. The caller must have disabled interrupts.
688 *
689 * This code assumes that the idle loop never does anything that might
690 * result in unbalanced calls to irq_enter() and irq_exit(). If your
691 * architecture's idle loop violates this assumption, RCU will give you what
692 * you deserve, good and hard. But very infrequently and irreproducibly.
693 *
694 * Use things like work queues to work around this limitation.
695 *
696 * You have been warned.
697 *
698 * If you add or remove a call to rcu_irq_exit(), be sure to test with
699 * CONFIG_RCU_EQS_DEBUG=y.
700 */
701void rcu_irq_exit(void)
702{
703 lockdep_assert_irqs_disabled();
704 rcu_nmi_exit_common(true);
705}
706
707/*
708 * Wrapper for rcu_irq_exit() where interrupts are enabled.
709 *
710 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
711 * with CONFIG_RCU_EQS_DEBUG=y.
712 */
713void rcu_irq_exit_irqson(void)
714{
715 unsigned long flags;
716
717 local_irq_save(flags);
718 rcu_irq_exit();
719 local_irq_restore(flags);
720}
721
722/*
723 * Exit an RCU extended quiescent state, which can be either the
724 * idle loop or adaptive-tickless usermode execution.
725 *
726 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
727 * allow for the possibility of usermode upcalls messing up our count of
728 * interrupt nesting level during the busy period that is just now starting.
729 */
730static void rcu_eqs_exit(bool user)
731{
732 struct rcu_data *rdp;
733 long oldval;
734
735 lockdep_assert_irqs_disabled();
736 rdp = this_cpu_ptr(&rcu_data);
737 oldval = rdp->dynticks_nesting;
738 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
739 if (oldval) {
740 rdp->dynticks_nesting++;
741 return;
742 }
743 rcu_dynticks_task_exit();
744 rcu_dynticks_eqs_exit();
745 rcu_cleanup_after_idle();
746 trace_rcu_dyntick(TPS("End"), rdp->dynticks_nesting, 1, rdp->dynticks);
747 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
748 WRITE_ONCE(rdp->dynticks_nesting, 1);
749 WARN_ON_ONCE(rdp->dynticks_nmi_nesting);
750 WRITE_ONCE(rdp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
751}
752
753/**
754 * rcu_idle_exit - inform RCU that current CPU is leaving idle
755 *
756 * Exit idle mode, in other words, -enter- the mode in which RCU
757 * read-side critical sections can occur.
758 *
759 * If you add or remove a call to rcu_idle_exit(), be sure to test with
760 * CONFIG_RCU_EQS_DEBUG=y.
761 */
762void rcu_idle_exit(void)
763{
764 unsigned long flags;
765
766 local_irq_save(flags);
767 rcu_eqs_exit(false);
768 local_irq_restore(flags);
769}
770
771#ifdef CONFIG_NO_HZ_FULL
772/**
773 * rcu_user_exit - inform RCU that we are exiting userspace.
774 *
775 * Exit RCU idle mode while entering the kernel because it can
776 * run a RCU read side critical section anytime.
777 *
778 * If you add or remove a call to rcu_user_exit(), be sure to test with
779 * CONFIG_RCU_EQS_DEBUG=y.
780 */
781void rcu_user_exit(void)
782{
783 rcu_eqs_exit(1);
784}
785#endif /* CONFIG_NO_HZ_FULL */
786
787/**
788 * rcu_nmi_enter_common - inform RCU of entry to NMI context
789 * @irq: Is this call from rcu_irq_enter?
790 *
791 * If the CPU was idle from RCU's viewpoint, update rdp->dynticks and
792 * rdp->dynticks_nmi_nesting to let the RCU grace-period handling know
793 * that the CPU is active. This implementation permits nested NMIs, as
794 * long as the nesting level does not overflow an int. (You will probably
795 * run out of stack space first.)
796 *
797 * If you add or remove a call to rcu_nmi_enter_common(), be sure to test
798 * with CONFIG_RCU_EQS_DEBUG=y.
799 */
800static __always_inline void rcu_nmi_enter_common(bool irq)
801{
802 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
803 long incby = 2;
804
805 /* Complain about underflow. */
806 WARN_ON_ONCE(rdp->dynticks_nmi_nesting < 0);
807
808 /*
809 * If idle from RCU viewpoint, atomically increment ->dynticks
810 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
811 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
812 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
813 * to be in the outermost NMI handler that interrupted an RCU-idle
814 * period (observation due to Andy Lutomirski).
815 */
816 if (rcu_dynticks_curr_cpu_in_eqs()) {
817
818 if (irq)
819 rcu_dynticks_task_exit();
820
821 rcu_dynticks_eqs_exit();
822
823 if (irq)
824 rcu_cleanup_after_idle();
825
826 incby = 1;
827 }
828 trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
829 rdp->dynticks_nmi_nesting,
830 rdp->dynticks_nmi_nesting + incby, rdp->dynticks);
831 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* Prevent store tearing. */
832 rdp->dynticks_nmi_nesting + incby);
833 barrier();
834}
835
836/**
837 * rcu_nmi_enter - inform RCU of entry to NMI context
838 */
839void rcu_nmi_enter(void)
840{
841 rcu_nmi_enter_common(false);
842}
843NOKPROBE_SYMBOL(rcu_nmi_enter);
844
845/**
846 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
847 *
848 * Enter an interrupt handler, which might possibly result in exiting
849 * idle mode, in other words, entering the mode in which read-side critical
850 * sections can occur. The caller must have disabled interrupts.
851 *
852 * Note that the Linux kernel is fully capable of entering an interrupt
853 * handler that it never exits, for example when doing upcalls to user mode!
854 * This code assumes that the idle loop never does upcalls to user mode.
855 * If your architecture's idle loop does do upcalls to user mode (or does
856 * anything else that results in unbalanced calls to the irq_enter() and
857 * irq_exit() functions), RCU will give you what you deserve, good and hard.
858 * But very infrequently and irreproducibly.
859 *
860 * Use things like work queues to work around this limitation.
861 *
862 * You have been warned.
863 *
864 * If you add or remove a call to rcu_irq_enter(), be sure to test with
865 * CONFIG_RCU_EQS_DEBUG=y.
866 */
867void rcu_irq_enter(void)
868{
869 lockdep_assert_irqs_disabled();
870 rcu_nmi_enter_common(true);
871}
872
873/*
874 * Wrapper for rcu_irq_enter() where interrupts are enabled.
875 *
876 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
877 * with CONFIG_RCU_EQS_DEBUG=y.
878 */
879void rcu_irq_enter_irqson(void)
880{
881 unsigned long flags;
882
883 local_irq_save(flags);
884 rcu_irq_enter();
885 local_irq_restore(flags);
886}
887
888/**
889 * rcu_is_watching - see if RCU thinks that the current CPU is not idle
890 *
891 * Return true if RCU is watching the running CPU, which means that this
892 * CPU can safely enter RCU read-side critical sections. In other words,
893 * if the current CPU is not in its idle loop or is in an interrupt or
894 * NMI handler, return true.
895 */
896bool notrace rcu_is_watching(void)
897{
898 bool ret;
899
900 preempt_disable_notrace();
901 ret = !rcu_dynticks_curr_cpu_in_eqs();
902 preempt_enable_notrace();
903 return ret;
904}
905EXPORT_SYMBOL_GPL(rcu_is_watching);
906
907/*
908 * If a holdout task is actually running, request an urgent quiescent
909 * state from its CPU. This is unsynchronized, so migrations can cause
910 * the request to go to the wrong CPU. Which is OK, all that will happen
911 * is that the CPU's next context switch will be a bit slower and next
912 * time around this task will generate another request.
913 */
914void rcu_request_urgent_qs_task(struct task_struct *t)
915{
916 int cpu;
917
918 barrier();
919 cpu = task_cpu(t);
920 if (!task_curr(t))
921 return; /* This task is not running on that CPU. */
922 smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
923}
924
925#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
926
927/*
928 * Is the current CPU online as far as RCU is concerned?
929 *
930 * Disable preemption to avoid false positives that could otherwise
931 * happen due to the current CPU number being sampled, this task being
932 * preempted, its old CPU being taken offline, resuming on some other CPU,
933 * then determining that its old CPU is now offline.
934 *
935 * Disable checking if in an NMI handler because we cannot safely
936 * report errors from NMI handlers anyway. In addition, it is OK to use
937 * RCU on an offline processor during initial boot, hence the check for
938 * rcu_scheduler_fully_active.
939 */
940bool rcu_lockdep_current_cpu_online(void)
941{
942 struct rcu_data *rdp;
943 struct rcu_node *rnp;
944 bool ret = false;
945
946 if (in_nmi() || !rcu_scheduler_fully_active)
947 return true;
948 preempt_disable();
949 rdp = this_cpu_ptr(&rcu_data);
950 rnp = rdp->mynode;
951 if (rdp->grpmask & rcu_rnp_online_cpus(rnp))
952 ret = true;
953 preempt_enable();
954 return ret;
955}
956EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
957
958#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
959
960/*
961 * We are reporting a quiescent state on behalf of some other CPU, so
962 * it is our responsibility to check for and handle potential overflow
963 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
964 * After all, the CPU might be in deep idle state, and thus executing no
965 * code whatsoever.
966 */
967static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
968{
969 raw_lockdep_assert_held_rcu_node(rnp);
970 if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
971 rnp->gp_seq))
972 WRITE_ONCE(rdp->gpwrap, true);
973 if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
974 rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
975}
976
977/*
978 * Snapshot the specified CPU's dynticks counter so that we can later
979 * credit them with an implicit quiescent state. Return 1 if this CPU
980 * is in dynticks idle mode, which is an extended quiescent state.
981 */
982static int dyntick_save_progress_counter(struct rcu_data *rdp)
983{
984 rdp->dynticks_snap = rcu_dynticks_snap(rdp);
985 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
986 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
987 rcu_gpnum_ovf(rdp->mynode, rdp);
988 return 1;
989 }
990 return 0;
991}
992
993/*
994 * Return true if the specified CPU has passed through a quiescent
995 * state by virtue of being in or having passed through an dynticks
996 * idle state since the last call to dyntick_save_progress_counter()
997 * for this same CPU, or by virtue of having been offline.
998 */
999static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1000{
1001 unsigned long jtsq;
1002 bool *rnhqp;
1003 bool *ruqp;
1004 struct rcu_node *rnp = rdp->mynode;
1005
1006 /*
1007 * If the CPU passed through or entered a dynticks idle phase with
1008 * no active irq/NMI handlers, then we can safely pretend that the CPU
1009 * already acknowledged the request to pass through a quiescent
1010 * state. Either way, that CPU cannot possibly be in an RCU
1011 * read-side critical section that started before the beginning
1012 * of the current RCU grace period.
1013 */
1014 if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
1015 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1016 rcu_gpnum_ovf(rnp, rdp);
1017 return 1;
1018 }
1019
1020 /* If waiting too long on an offline CPU, complain. */
1021 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp)) &&
1022 time_after(jiffies, rcu_state.gp_start + HZ)) {
1023 bool onl;
1024 struct rcu_node *rnp1;
1025
1026 WARN_ON(1); /* Offline CPUs are supposed to report QS! */
1027 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
1028 __func__, rnp->grplo, rnp->grphi, rnp->level,
1029 (long)rnp->gp_seq, (long)rnp->completedqs);
1030 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
1031 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
1032 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
1033 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
1034 pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
1035 __func__, rdp->cpu, ".o"[onl],
1036 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
1037 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
1038 return 1; /* Break things loose after complaining. */
1039 }
1040
1041 /*
1042 * A CPU running for an extended time within the kernel can
1043 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
1044 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
1045 * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the
1046 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
1047 * variable are safe because the assignments are repeated if this
1048 * CPU failed to pass through a quiescent state. This code
1049 * also checks .jiffies_resched in case jiffies_to_sched_qs
1050 * is set way high.
1051 */
1052 jtsq = READ_ONCE(jiffies_to_sched_qs);
1053 ruqp = per_cpu_ptr(&rcu_data.rcu_urgent_qs, rdp->cpu);
1054 rnhqp = &per_cpu(rcu_data.rcu_need_heavy_qs, rdp->cpu);
1055 if (!READ_ONCE(*rnhqp) &&
1056 (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
1057 time_after(jiffies, rcu_state.jiffies_resched))) {
1058 WRITE_ONCE(*rnhqp, true);
1059 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1060 smp_store_release(ruqp, true);
1061 } else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
1062 WRITE_ONCE(*ruqp, true);
1063 }
1064
1065 /*
1066 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
1067 * The above code handles this, but only for straight cond_resched().
1068 * And some in-kernel loops check need_resched() before calling
1069 * cond_resched(), which defeats the above code for CPUs that are
1070 * running in-kernel with scheduling-clock interrupts disabled.
1071 * So hit them over the head with the resched_cpu() hammer!
1072 */
1073 if (tick_nohz_full_cpu(rdp->cpu) &&
1074 time_after(jiffies,
1075 READ_ONCE(rdp->last_fqs_resched) + jtsq * 3)) {
1076 resched_cpu(rdp->cpu);
1077 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1078 }
1079
1080 /*
1081 * If more than halfway to RCU CPU stall-warning time, invoke
1082 * resched_cpu() more frequently to try to loosen things up a bit.
1083 * Also check to see if the CPU is getting hammered with interrupts,
1084 * but only once per grace period, just to keep the IPIs down to
1085 * a dull roar.
1086 */
1087 if (time_after(jiffies, rcu_state.jiffies_resched)) {
1088 if (time_after(jiffies,
1089 READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
1090 resched_cpu(rdp->cpu);
1091 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1092 }
1093 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1094 !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
1095 (rnp->ffmask & rdp->grpmask)) {
1096 init_irq_work(&rdp->rcu_iw, rcu_iw_handler);
1097 rdp->rcu_iw_pending = true;
1098 rdp->rcu_iw_gp_seq = rnp->gp_seq;
1099 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1100 }
1101 }
1102
1103 return 0;
1104}
1105
1106/* Trace-event wrapper function for trace_rcu_future_grace_period. */
1107static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1108 unsigned long gp_seq_req, const char *s)
1109{
1110 trace_rcu_future_grace_period(rcu_state.name, rnp->gp_seq, gp_seq_req,
1111 rnp->level, rnp->grplo, rnp->grphi, s);
1112}
1113
1114/*
1115 * rcu_start_this_gp - Request the start of a particular grace period
1116 * @rnp_start: The leaf node of the CPU from which to start.
1117 * @rdp: The rcu_data corresponding to the CPU from which to start.
1118 * @gp_seq_req: The gp_seq of the grace period to start.
1119 *
1120 * Start the specified grace period, as needed to handle newly arrived
1121 * callbacks. The required future grace periods are recorded in each
1122 * rcu_node structure's ->gp_seq_needed field. Returns true if there
1123 * is reason to awaken the grace-period kthread.
1124 *
1125 * The caller must hold the specified rcu_node structure's ->lock, which
1126 * is why the caller is responsible for waking the grace-period kthread.
1127 *
1128 * Returns true if the GP thread needs to be awakened else false.
1129 */
1130static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
1131 unsigned long gp_seq_req)
1132{
1133 bool ret = false;
1134 struct rcu_node *rnp;
1135
1136 /*
1137 * Use funnel locking to either acquire the root rcu_node
1138 * structure's lock or bail out if the need for this grace period
1139 * has already been recorded -- or if that grace period has in
1140 * fact already started. If there is already a grace period in
1141 * progress in a non-leaf node, no recording is needed because the
1142 * end of the grace period will scan the leaf rcu_node structures.
1143 * Note that rnp_start->lock must not be released.
1144 */
1145 raw_lockdep_assert_held_rcu_node(rnp_start);
1146 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
1147 for (rnp = rnp_start; 1; rnp = rnp->parent) {
1148 if (rnp != rnp_start)
1149 raw_spin_lock_rcu_node(rnp);
1150 if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
1151 rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
1152 (rnp != rnp_start &&
1153 rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
1154 trace_rcu_this_gp(rnp, rdp, gp_seq_req,
1155 TPS("Prestarted"));
1156 goto unlock_out;
1157 }
1158 rnp->gp_seq_needed = gp_seq_req;
1159 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
1160 /*
1161 * We just marked the leaf or internal node, and a
1162 * grace period is in progress, which means that
1163 * rcu_gp_cleanup() will see the marking. Bail to
1164 * reduce contention.
1165 */
1166 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
1167 TPS("Startedleaf"));
1168 goto unlock_out;
1169 }
1170 if (rnp != rnp_start && rnp->parent != NULL)
1171 raw_spin_unlock_rcu_node(rnp);
1172 if (!rnp->parent)
1173 break; /* At root, and perhaps also leaf. */
1174 }
1175
1176 /* If GP already in progress, just leave, otherwise start one. */
1177 if (rcu_gp_in_progress()) {
1178 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
1179 goto unlock_out;
1180 }
1181 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
1182 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
1183 rcu_state.gp_req_activity = jiffies;
1184 if (!rcu_state.gp_kthread) {
1185 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
1186 goto unlock_out;
1187 }
1188 trace_rcu_grace_period(rcu_state.name, READ_ONCE(rcu_state.gp_seq), TPS("newreq"));
1189 ret = true; /* Caller must wake GP kthread. */
1190unlock_out:
1191 /* Push furthest requested GP to leaf node and rcu_data structure. */
1192 if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
1193 rnp_start->gp_seq_needed = rnp->gp_seq_needed;
1194 rdp->gp_seq_needed = rnp->gp_seq_needed;
1195 }
1196 if (rnp != rnp_start)
1197 raw_spin_unlock_rcu_node(rnp);
1198 return ret;
1199}
1200
1201/*
1202 * Clean up any old requests for the just-ended grace period. Also return
1203 * whether any additional grace periods have been requested.
1204 */
1205static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
1206{
1207 bool needmore;
1208 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1209
1210 needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1211 if (!needmore)
1212 rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1213 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1214 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1215 return needmore;
1216}
1217
1218/*
1219 * Awaken the grace-period kthread. Don't do a self-awaken (unless in
1220 * an interrupt or softirq handler), and don't bother awakening when there
1221 * is nothing for the grace-period kthread to do (as in several CPUs raced
1222 * to awaken, and we lost), and finally don't try to awaken a kthread that
1223 * has not yet been created. If all those checks are passed, track some
1224 * debug information and awaken.
1225 *
1226 * So why do the self-wakeup when in an interrupt or softirq handler
1227 * in the grace-period kthread's context? Because the kthread might have
1228 * been interrupted just as it was going to sleep, and just after the final
1229 * pre-sleep check of the awaken condition. In this case, a wakeup really
1230 * is required, and is therefore supplied.
1231 */
1232static void rcu_gp_kthread_wake(void)
1233{
1234 if ((current == rcu_state.gp_kthread &&
1235 !in_irq() && !in_serving_softirq()) ||
1236 !READ_ONCE(rcu_state.gp_flags) ||
1237 !rcu_state.gp_kthread)
1238 return;
1239 WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
1240 WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
1241 swake_up_one(&rcu_state.gp_wq);
1242}
1243
1244/*
1245 * If there is room, assign a ->gp_seq number to any callbacks on this
1246 * CPU that have not already been assigned. Also accelerate any callbacks
1247 * that were previously assigned a ->gp_seq number that has since proven
1248 * to be too conservative, which can happen if callbacks get assigned a
1249 * ->gp_seq number while RCU is idle, but with reference to a non-root
1250 * rcu_node structure. This function is idempotent, so it does not hurt
1251 * to call it repeatedly. Returns an flag saying that we should awaken
1252 * the RCU grace-period kthread.
1253 *
1254 * The caller must hold rnp->lock with interrupts disabled.
1255 */
1256static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1257{
1258 unsigned long gp_seq_req;
1259 bool ret = false;
1260
1261 rcu_lockdep_assert_cblist_protected(rdp);
1262 raw_lockdep_assert_held_rcu_node(rnp);
1263
1264 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1265 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1266 return false;
1267
1268 /*
1269 * Callbacks are often registered with incomplete grace-period
1270 * information. Something about the fact that getting exact
1271 * information requires acquiring a global lock... RCU therefore
1272 * makes a conservative estimate of the grace period number at which
1273 * a given callback will become ready to invoke. The following
1274 * code checks this estimate and improves it when possible, thus
1275 * accelerating callback invocation to an earlier grace-period
1276 * number.
1277 */
1278 gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1279 if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1280 ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1281
1282 /* Trace depending on how much we were able to accelerate. */
1283 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1284 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccWaitCB"));
1285 else
1286 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccReadyCB"));
1287 return ret;
1288}
1289
1290/*
1291 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1292 * rcu_node structure's ->lock be held. It consults the cached value
1293 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1294 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1295 * while holding the leaf rcu_node structure's ->lock.
1296 */
1297static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1298 struct rcu_data *rdp)
1299{
1300 unsigned long c;
1301 bool needwake;
1302
1303 rcu_lockdep_assert_cblist_protected(rdp);
1304 c = rcu_seq_snap(&rcu_state.gp_seq);
1305 if (!rdp->gpwrap && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1306 /* Old request still live, so mark recent callbacks. */
1307 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1308 return;
1309 }
1310 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1311 needwake = rcu_accelerate_cbs(rnp, rdp);
1312 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1313 if (needwake)
1314 rcu_gp_kthread_wake();
1315}
1316
1317/*
1318 * Move any callbacks whose grace period has completed to the
1319 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1320 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1321 * sublist. This function is idempotent, so it does not hurt to
1322 * invoke it repeatedly. As long as it is not invoked -too- often...
1323 * Returns true if the RCU grace-period kthread needs to be awakened.
1324 *
1325 * The caller must hold rnp->lock with interrupts disabled.
1326 */
1327static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1328{
1329 rcu_lockdep_assert_cblist_protected(rdp);
1330 raw_lockdep_assert_held_rcu_node(rnp);
1331
1332 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1333 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1334 return false;
1335
1336 /*
1337 * Find all callbacks whose ->gp_seq numbers indicate that they
1338 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1339 */
1340 rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1341
1342 /* Classify any remaining callbacks. */
1343 return rcu_accelerate_cbs(rnp, rdp);
1344}
1345
1346/*
1347 * Move and classify callbacks, but only if doing so won't require
1348 * that the RCU grace-period kthread be awakened.
1349 */
1350static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
1351 struct rcu_data *rdp)
1352{
1353 rcu_lockdep_assert_cblist_protected(rdp);
1354 if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) ||
1355 !raw_spin_trylock_rcu_node(rnp))
1356 return;
1357 WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
1358 raw_spin_unlock_rcu_node(rnp);
1359}
1360
1361/*
1362 * Update CPU-local rcu_data state to record the beginnings and ends of
1363 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1364 * structure corresponding to the current CPU, and must have irqs disabled.
1365 * Returns true if the grace-period kthread needs to be awakened.
1366 */
1367static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
1368{
1369 bool ret = false;
1370 bool need_gp;
1371 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
1372 rcu_segcblist_is_offloaded(&rdp->cblist);
1373
1374 raw_lockdep_assert_held_rcu_node(rnp);
1375
1376 if (rdp->gp_seq == rnp->gp_seq)
1377 return false; /* Nothing to do. */
1378
1379 /* Handle the ends of any preceding grace periods first. */
1380 if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1381 unlikely(READ_ONCE(rdp->gpwrap))) {
1382 if (!offloaded)
1383 ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
1384 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1385 } else {
1386 if (!offloaded)
1387 ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
1388 }
1389
1390 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1391 if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1392 unlikely(READ_ONCE(rdp->gpwrap))) {
1393 /*
1394 * If the current grace period is waiting for this CPU,
1395 * set up to detect a quiescent state, otherwise don't
1396 * go looking for one.
1397 */
1398 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1399 need_gp = !!(rnp->qsmask & rdp->grpmask);
1400 rdp->cpu_no_qs.b.norm = need_gp;
1401 rdp->core_needs_qs = need_gp;
1402 zero_cpu_stall_ticks(rdp);
1403 }
1404 rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
1405 if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
1406 rdp->gp_seq_needed = rnp->gp_seq_needed;
1407 WRITE_ONCE(rdp->gpwrap, false);
1408 rcu_gpnum_ovf(rnp, rdp);
1409 return ret;
1410}
1411
1412static void note_gp_changes(struct rcu_data *rdp)
1413{
1414 unsigned long flags;
1415 bool needwake;
1416 struct rcu_node *rnp;
1417
1418 local_irq_save(flags);
1419 rnp = rdp->mynode;
1420 if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1421 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1422 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1423 local_irq_restore(flags);
1424 return;
1425 }
1426 needwake = __note_gp_changes(rnp, rdp);
1427 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1428 if (needwake)
1429 rcu_gp_kthread_wake();
1430}
1431
1432static void rcu_gp_slow(int delay)
1433{
1434 if (delay > 0 &&
1435 !(rcu_seq_ctr(rcu_state.gp_seq) %
1436 (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1437 schedule_timeout_uninterruptible(delay);
1438}
1439
1440/*
1441 * Initialize a new grace period. Return false if no grace period required.
1442 */
1443static bool rcu_gp_init(void)
1444{
1445 unsigned long flags;
1446 unsigned long oldmask;
1447 unsigned long mask;
1448 struct rcu_data *rdp;
1449 struct rcu_node *rnp = rcu_get_root();
1450
1451 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1452 raw_spin_lock_irq_rcu_node(rnp);
1453 if (!READ_ONCE(rcu_state.gp_flags)) {
1454 /* Spurious wakeup, tell caller to go back to sleep. */
1455 raw_spin_unlock_irq_rcu_node(rnp);
1456 return false;
1457 }
1458 WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1459
1460 if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1461 /*
1462 * Grace period already in progress, don't start another.
1463 * Not supposed to be able to happen.
1464 */
1465 raw_spin_unlock_irq_rcu_node(rnp);
1466 return false;
1467 }
1468
1469 /* Advance to a new grace period and initialize state. */
1470 record_gp_stall_check_time();
1471 /* Record GP times before starting GP, hence rcu_seq_start(). */
1472 rcu_seq_start(&rcu_state.gp_seq);
1473 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
1474 raw_spin_unlock_irq_rcu_node(rnp);
1475
1476 /*
1477 * Apply per-leaf buffered online and offline operations to the
1478 * rcu_node tree. Note that this new grace period need not wait
1479 * for subsequent online CPUs, and that quiescent-state forcing
1480 * will handle subsequent offline CPUs.
1481 */
1482 rcu_state.gp_state = RCU_GP_ONOFF;
1483 rcu_for_each_leaf_node(rnp) {
1484 raw_spin_lock(&rcu_state.ofl_lock);
1485 raw_spin_lock_irq_rcu_node(rnp);
1486 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1487 !rnp->wait_blkd_tasks) {
1488 /* Nothing to do on this leaf rcu_node structure. */
1489 raw_spin_unlock_irq_rcu_node(rnp);
1490 raw_spin_unlock(&rcu_state.ofl_lock);
1491 continue;
1492 }
1493
1494 /* Record old state, apply changes to ->qsmaskinit field. */
1495 oldmask = rnp->qsmaskinit;
1496 rnp->qsmaskinit = rnp->qsmaskinitnext;
1497
1498 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1499 if (!oldmask != !rnp->qsmaskinit) {
1500 if (!oldmask) { /* First online CPU for rcu_node. */
1501 if (!rnp->wait_blkd_tasks) /* Ever offline? */
1502 rcu_init_new_rnp(rnp);
1503 } else if (rcu_preempt_has_tasks(rnp)) {
1504 rnp->wait_blkd_tasks = true; /* blocked tasks */
1505 } else { /* Last offline CPU and can propagate. */
1506 rcu_cleanup_dead_rnp(rnp);
1507 }
1508 }
1509
1510 /*
1511 * If all waited-on tasks from prior grace period are
1512 * done, and if all this rcu_node structure's CPUs are
1513 * still offline, propagate up the rcu_node tree and
1514 * clear ->wait_blkd_tasks. Otherwise, if one of this
1515 * rcu_node structure's CPUs has since come back online,
1516 * simply clear ->wait_blkd_tasks.
1517 */
1518 if (rnp->wait_blkd_tasks &&
1519 (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1520 rnp->wait_blkd_tasks = false;
1521 if (!rnp->qsmaskinit)
1522 rcu_cleanup_dead_rnp(rnp);
1523 }
1524
1525 raw_spin_unlock_irq_rcu_node(rnp);
1526 raw_spin_unlock(&rcu_state.ofl_lock);
1527 }
1528 rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1529
1530 /*
1531 * Set the quiescent-state-needed bits in all the rcu_node
1532 * structures for all currently online CPUs in breadth-first
1533 * order, starting from the root rcu_node structure, relying on the
1534 * layout of the tree within the rcu_state.node[] array. Note that
1535 * other CPUs will access only the leaves of the hierarchy, thus
1536 * seeing that no grace period is in progress, at least until the
1537 * corresponding leaf node has been initialized.
1538 *
1539 * The grace period cannot complete until the initialization
1540 * process finishes, because this kthread handles both.
1541 */
1542 rcu_state.gp_state = RCU_GP_INIT;
1543 rcu_for_each_node_breadth_first(rnp) {
1544 rcu_gp_slow(gp_init_delay);
1545 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1546 rdp = this_cpu_ptr(&rcu_data);
1547 rcu_preempt_check_blocked_tasks(rnp);
1548 rnp->qsmask = rnp->qsmaskinit;
1549 WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
1550 if (rnp == rdp->mynode)
1551 (void)__note_gp_changes(rnp, rdp);
1552 rcu_preempt_boost_start_gp(rnp);
1553 trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1554 rnp->level, rnp->grplo,
1555 rnp->grphi, rnp->qsmask);
1556 /* Quiescent states for tasks on any now-offline CPUs. */
1557 mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1558 rnp->rcu_gp_init_mask = mask;
1559 if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1560 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1561 else
1562 raw_spin_unlock_irq_rcu_node(rnp);
1563 cond_resched_tasks_rcu_qs();
1564 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1565 }
1566
1567 return true;
1568}
1569
1570/*
1571 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1572 * time.
1573 */
1574static bool rcu_gp_fqs_check_wake(int *gfp)
1575{
1576 struct rcu_node *rnp = rcu_get_root();
1577
1578 /* Someone like call_rcu() requested a force-quiescent-state scan. */
1579 *gfp = READ_ONCE(rcu_state.gp_flags);
1580 if (*gfp & RCU_GP_FLAG_FQS)
1581 return true;
1582
1583 /* The current grace period has completed. */
1584 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1585 return true;
1586
1587 return false;
1588}
1589
1590/*
1591 * Do one round of quiescent-state forcing.
1592 */
1593static void rcu_gp_fqs(bool first_time)
1594{
1595 struct rcu_node *rnp = rcu_get_root();
1596
1597 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1598 rcu_state.n_force_qs++;
1599 if (first_time) {
1600 /* Collect dyntick-idle snapshots. */
1601 force_qs_rnp(dyntick_save_progress_counter);
1602 } else {
1603 /* Handle dyntick-idle and offline CPUs. */
1604 force_qs_rnp(rcu_implicit_dynticks_qs);
1605 }
1606 /* Clear flag to prevent immediate re-entry. */
1607 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
1608 raw_spin_lock_irq_rcu_node(rnp);
1609 WRITE_ONCE(rcu_state.gp_flags,
1610 READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
1611 raw_spin_unlock_irq_rcu_node(rnp);
1612 }
1613}
1614
1615/*
1616 * Loop doing repeated quiescent-state forcing until the grace period ends.
1617 */
1618static void rcu_gp_fqs_loop(void)
1619{
1620 bool first_gp_fqs;
1621 int gf;
1622 unsigned long j;
1623 int ret;
1624 struct rcu_node *rnp = rcu_get_root();
1625
1626 first_gp_fqs = true;
1627 j = READ_ONCE(jiffies_till_first_fqs);
1628 ret = 0;
1629 for (;;) {
1630 if (!ret) {
1631 rcu_state.jiffies_force_qs = jiffies + j;
1632 WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
1633 jiffies + (j ? 3 * j : 2));
1634 }
1635 trace_rcu_grace_period(rcu_state.name,
1636 READ_ONCE(rcu_state.gp_seq),
1637 TPS("fqswait"));
1638 rcu_state.gp_state = RCU_GP_WAIT_FQS;
1639 ret = swait_event_idle_timeout_exclusive(
1640 rcu_state.gp_wq, rcu_gp_fqs_check_wake(&gf), j);
1641 rcu_state.gp_state = RCU_GP_DOING_FQS;
1642 /* Locking provides needed memory barriers. */
1643 /* If grace period done, leave loop. */
1644 if (!READ_ONCE(rnp->qsmask) &&
1645 !rcu_preempt_blocked_readers_cgp(rnp))
1646 break;
1647 /* If time for quiescent-state forcing, do it. */
1648 if (ULONG_CMP_GE(jiffies, rcu_state.jiffies_force_qs) ||
1649 (gf & RCU_GP_FLAG_FQS)) {
1650 trace_rcu_grace_period(rcu_state.name,
1651 READ_ONCE(rcu_state.gp_seq),
1652 TPS("fqsstart"));
1653 rcu_gp_fqs(first_gp_fqs);
1654 first_gp_fqs = false;
1655 trace_rcu_grace_period(rcu_state.name,
1656 READ_ONCE(rcu_state.gp_seq),
1657 TPS("fqsend"));
1658 cond_resched_tasks_rcu_qs();
1659 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1660 ret = 0; /* Force full wait till next FQS. */
1661 j = READ_ONCE(jiffies_till_next_fqs);
1662 } else {
1663 /* Deal with stray signal. */
1664 cond_resched_tasks_rcu_qs();
1665 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1666 WARN_ON(signal_pending(current));
1667 trace_rcu_grace_period(rcu_state.name,
1668 READ_ONCE(rcu_state.gp_seq),
1669 TPS("fqswaitsig"));
1670 ret = 1; /* Keep old FQS timing. */
1671 j = jiffies;
1672 if (time_after(jiffies, rcu_state.jiffies_force_qs))
1673 j = 1;
1674 else
1675 j = rcu_state.jiffies_force_qs - j;
1676 }
1677 }
1678}
1679
1680/*
1681 * Clean up after the old grace period.
1682 */
1683static void rcu_gp_cleanup(void)
1684{
1685 unsigned long gp_duration;
1686 bool needgp = false;
1687 unsigned long new_gp_seq;
1688 bool offloaded;
1689 struct rcu_data *rdp;
1690 struct rcu_node *rnp = rcu_get_root();
1691 struct swait_queue_head *sq;
1692
1693 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1694 raw_spin_lock_irq_rcu_node(rnp);
1695 rcu_state.gp_end = jiffies;
1696 gp_duration = rcu_state.gp_end - rcu_state.gp_start;
1697 if (gp_duration > rcu_state.gp_max)
1698 rcu_state.gp_max = gp_duration;
1699
1700 /*
1701 * We know the grace period is complete, but to everyone else
1702 * it appears to still be ongoing. But it is also the case
1703 * that to everyone else it looks like there is nothing that
1704 * they can do to advance the grace period. It is therefore
1705 * safe for us to drop the lock in order to mark the grace
1706 * period as completed in all of the rcu_node structures.
1707 */
1708 raw_spin_unlock_irq_rcu_node(rnp);
1709
1710 /*
1711 * Propagate new ->gp_seq value to rcu_node structures so that
1712 * other CPUs don't have to wait until the start of the next grace
1713 * period to process their callbacks. This also avoids some nasty
1714 * RCU grace-period initialization races by forcing the end of
1715 * the current grace period to be completely recorded in all of
1716 * the rcu_node structures before the beginning of the next grace
1717 * period is recorded in any of the rcu_node structures.
1718 */
1719 new_gp_seq = rcu_state.gp_seq;
1720 rcu_seq_end(&new_gp_seq);
1721 rcu_for_each_node_breadth_first(rnp) {
1722 raw_spin_lock_irq_rcu_node(rnp);
1723 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
1724 dump_blkd_tasks(rnp, 10);
1725 WARN_ON_ONCE(rnp->qsmask);
1726 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
1727 rdp = this_cpu_ptr(&rcu_data);
1728 if (rnp == rdp->mynode)
1729 needgp = __note_gp_changes(rnp, rdp) || needgp;
1730 /* smp_mb() provided by prior unlock-lock pair. */
1731 needgp = rcu_future_gp_cleanup(rnp) || needgp;
1732 sq = rcu_nocb_gp_get(rnp);
1733 raw_spin_unlock_irq_rcu_node(rnp);
1734 rcu_nocb_gp_cleanup(sq);
1735 cond_resched_tasks_rcu_qs();
1736 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1737 rcu_gp_slow(gp_cleanup_delay);
1738 }
1739 rnp = rcu_get_root();
1740 raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
1741
1742 /* Declare grace period done, trace first to use old GP number. */
1743 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
1744 rcu_seq_end(&rcu_state.gp_seq);
1745 rcu_state.gp_state = RCU_GP_IDLE;
1746 /* Check for GP requests since above loop. */
1747 rdp = this_cpu_ptr(&rcu_data);
1748 if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
1749 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
1750 TPS("CleanupMore"));
1751 needgp = true;
1752 }
1753 /* Advance CBs to reduce false positives below. */
1754 offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
1755 rcu_segcblist_is_offloaded(&rdp->cblist);
1756 if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
1757 WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
1758 rcu_state.gp_req_activity = jiffies;
1759 trace_rcu_grace_period(rcu_state.name,
1760 READ_ONCE(rcu_state.gp_seq),
1761 TPS("newreq"));
1762 } else {
1763 WRITE_ONCE(rcu_state.gp_flags,
1764 rcu_state.gp_flags & RCU_GP_FLAG_INIT);
1765 }
1766 raw_spin_unlock_irq_rcu_node(rnp);
1767}
1768
1769/*
1770 * Body of kthread that handles grace periods.
1771 */
1772static int __noreturn rcu_gp_kthread(void *unused)
1773{
1774 rcu_bind_gp_kthread();
1775 for (;;) {
1776
1777 /* Handle grace-period start. */
1778 for (;;) {
1779 trace_rcu_grace_period(rcu_state.name,
1780 READ_ONCE(rcu_state.gp_seq),
1781 TPS("reqwait"));
1782 rcu_state.gp_state = RCU_GP_WAIT_GPS;
1783 swait_event_idle_exclusive(rcu_state.gp_wq,
1784 READ_ONCE(rcu_state.gp_flags) &
1785 RCU_GP_FLAG_INIT);
1786 rcu_state.gp_state = RCU_GP_DONE_GPS;
1787 /* Locking provides needed memory barrier. */
1788 if (rcu_gp_init())
1789 break;
1790 cond_resched_tasks_rcu_qs();
1791 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1792 WARN_ON(signal_pending(current));
1793 trace_rcu_grace_period(rcu_state.name,
1794 READ_ONCE(rcu_state.gp_seq),
1795 TPS("reqwaitsig"));
1796 }
1797
1798 /* Handle quiescent-state forcing. */
1799 rcu_gp_fqs_loop();
1800
1801 /* Handle grace-period end. */
1802 rcu_state.gp_state = RCU_GP_CLEANUP;
1803 rcu_gp_cleanup();
1804 rcu_state.gp_state = RCU_GP_CLEANED;
1805 }
1806}
1807
1808/*
1809 * Report a full set of quiescent states to the rcu_state data structure.
1810 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
1811 * another grace period is required. Whether we wake the grace-period
1812 * kthread or it awakens itself for the next round of quiescent-state
1813 * forcing, that kthread will clean up after the just-completed grace
1814 * period. Note that the caller must hold rnp->lock, which is released
1815 * before return.
1816 */
1817static void rcu_report_qs_rsp(unsigned long flags)
1818 __releases(rcu_get_root()->lock)
1819{
1820 raw_lockdep_assert_held_rcu_node(rcu_get_root());
1821 WARN_ON_ONCE(!rcu_gp_in_progress());
1822 WRITE_ONCE(rcu_state.gp_flags,
1823 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
1824 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
1825 rcu_gp_kthread_wake();
1826}
1827
1828/*
1829 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1830 * Allows quiescent states for a group of CPUs to be reported at one go
1831 * to the specified rcu_node structure, though all the CPUs in the group
1832 * must be represented by the same rcu_node structure (which need not be a
1833 * leaf rcu_node structure, though it often will be). The gps parameter
1834 * is the grace-period snapshot, which means that the quiescent states
1835 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
1836 * must be held upon entry, and it is released before return.
1837 *
1838 * As a special case, if mask is zero, the bit-already-cleared check is
1839 * disabled. This allows propagating quiescent state due to resumed tasks
1840 * during grace-period initialization.
1841 */
1842static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
1843 unsigned long gps, unsigned long flags)
1844 __releases(rnp->lock)
1845{
1846 unsigned long oldmask = 0;
1847 struct rcu_node *rnp_c;
1848
1849 raw_lockdep_assert_held_rcu_node(rnp);
1850
1851 /* Walk up the rcu_node hierarchy. */
1852 for (;;) {
1853 if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
1854
1855 /*
1856 * Our bit has already been cleared, or the
1857 * relevant grace period is already over, so done.
1858 */
1859 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1860 return;
1861 }
1862 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
1863 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
1864 rcu_preempt_blocked_readers_cgp(rnp));
1865 rnp->qsmask &= ~mask;
1866 trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
1867 mask, rnp->qsmask, rnp->level,
1868 rnp->grplo, rnp->grphi,
1869 !!rnp->gp_tasks);
1870 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1871
1872 /* Other bits still set at this level, so done. */
1873 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1874 return;
1875 }
1876 rnp->completedqs = rnp->gp_seq;
1877 mask = rnp->grpmask;
1878 if (rnp->parent == NULL) {
1879
1880 /* No more levels. Exit loop holding root lock. */
1881
1882 break;
1883 }
1884 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1885 rnp_c = rnp;
1886 rnp = rnp->parent;
1887 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1888 oldmask = rnp_c->qsmask;
1889 }
1890
1891 /*
1892 * Get here if we are the last CPU to pass through a quiescent
1893 * state for this grace period. Invoke rcu_report_qs_rsp()
1894 * to clean up and start the next grace period if one is needed.
1895 */
1896 rcu_report_qs_rsp(flags); /* releases rnp->lock. */
1897}
1898
1899/*
1900 * Record a quiescent state for all tasks that were previously queued
1901 * on the specified rcu_node structure and that were blocking the current
1902 * RCU grace period. The caller must hold the corresponding rnp->lock with
1903 * irqs disabled, and this lock is released upon return, but irqs remain
1904 * disabled.
1905 */
1906static void __maybe_unused
1907rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
1908 __releases(rnp->lock)
1909{
1910 unsigned long gps;
1911 unsigned long mask;
1912 struct rcu_node *rnp_p;
1913
1914 raw_lockdep_assert_held_rcu_node(rnp);
1915 if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPTION)) ||
1916 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
1917 rnp->qsmask != 0) {
1918 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1919 return; /* Still need more quiescent states! */
1920 }
1921
1922 rnp->completedqs = rnp->gp_seq;
1923 rnp_p = rnp->parent;
1924 if (rnp_p == NULL) {
1925 /*
1926 * Only one rcu_node structure in the tree, so don't
1927 * try to report up to its nonexistent parent!
1928 */
1929 rcu_report_qs_rsp(flags);
1930 return;
1931 }
1932
1933 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
1934 gps = rnp->gp_seq;
1935 mask = rnp->grpmask;
1936 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1937 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
1938 rcu_report_qs_rnp(mask, rnp_p, gps, flags);
1939}
1940
1941/*
1942 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1943 * structure. This must be called from the specified CPU.
1944 */
1945static void
1946rcu_report_qs_rdp(int cpu, struct rcu_data *rdp)
1947{
1948 unsigned long flags;
1949 unsigned long mask;
1950 bool needwake = false;
1951 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
1952 rcu_segcblist_is_offloaded(&rdp->cblist);
1953 struct rcu_node *rnp;
1954
1955 rnp = rdp->mynode;
1956 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1957 if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
1958 rdp->gpwrap) {
1959
1960 /*
1961 * The grace period in which this quiescent state was
1962 * recorded has ended, so don't report it upwards.
1963 * We will instead need a new quiescent state that lies
1964 * within the current grace period.
1965 */
1966 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
1967 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1968 return;
1969 }
1970 mask = rdp->grpmask;
1971 rdp->core_needs_qs = false;
1972 if ((rnp->qsmask & mask) == 0) {
1973 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1974 } else {
1975 /*
1976 * This GP can't end until cpu checks in, so all of our
1977 * callbacks can be processed during the next GP.
1978 */
1979 if (!offloaded)
1980 needwake = rcu_accelerate_cbs(rnp, rdp);
1981
1982 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1983 /* ^^^ Released rnp->lock */
1984 if (needwake)
1985 rcu_gp_kthread_wake();
1986 }
1987}
1988
1989/*
1990 * Check to see if there is a new grace period of which this CPU
1991 * is not yet aware, and if so, set up local rcu_data state for it.
1992 * Otherwise, see if this CPU has just passed through its first
1993 * quiescent state for this grace period, and record that fact if so.
1994 */
1995static void
1996rcu_check_quiescent_state(struct rcu_data *rdp)
1997{
1998 /* Check for grace-period ends and beginnings. */
1999 note_gp_changes(rdp);
2000
2001 /*
2002 * Does this CPU still need to do its part for current grace period?
2003 * If no, return and let the other CPUs do their part as well.
2004 */
2005 if (!rdp->core_needs_qs)
2006 return;
2007
2008 /*
2009 * Was there a quiescent state since the beginning of the grace
2010 * period? If no, then exit and wait for the next call.
2011 */
2012 if (rdp->cpu_no_qs.b.norm)
2013 return;
2014
2015 /*
2016 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2017 * judge of that).
2018 */
2019 rcu_report_qs_rdp(rdp->cpu, rdp);
2020}
2021
2022/*
2023 * Near the end of the offline process. Trace the fact that this CPU
2024 * is going offline.
2025 */
2026int rcutree_dying_cpu(unsigned int cpu)
2027{
2028 bool blkd;
2029 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
2030 struct rcu_node *rnp = rdp->mynode;
2031
2032 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2033 return 0;
2034
2035 blkd = !!(rnp->qsmask & rdp->grpmask);
2036 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq,
2037 blkd ? TPS("cpuofl") : TPS("cpuofl-bgp"));
2038 return 0;
2039}
2040
2041/*
2042 * All CPUs for the specified rcu_node structure have gone offline,
2043 * and all tasks that were preempted within an RCU read-side critical
2044 * section while running on one of those CPUs have since exited their RCU
2045 * read-side critical section. Some other CPU is reporting this fact with
2046 * the specified rcu_node structure's ->lock held and interrupts disabled.
2047 * This function therefore goes up the tree of rcu_node structures,
2048 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2049 * the leaf rcu_node structure's ->qsmaskinit field has already been
2050 * updated.
2051 *
2052 * This function does check that the specified rcu_node structure has
2053 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2054 * prematurely. That said, invoking it after the fact will cost you
2055 * a needless lock acquisition. So once it has done its work, don't
2056 * invoke it again.
2057 */
2058static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2059{
2060 long mask;
2061 struct rcu_node *rnp = rnp_leaf;
2062
2063 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2064 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2065 WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
2066 WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
2067 return;
2068 for (;;) {
2069 mask = rnp->grpmask;
2070 rnp = rnp->parent;
2071 if (!rnp)
2072 break;
2073 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2074 rnp->qsmaskinit &= ~mask;
2075 /* Between grace periods, so better already be zero! */
2076 WARN_ON_ONCE(rnp->qsmask);
2077 if (rnp->qsmaskinit) {
2078 raw_spin_unlock_rcu_node(rnp);
2079 /* irqs remain disabled. */
2080 return;
2081 }
2082 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2083 }
2084}
2085
2086/*
2087 * The CPU has been completely removed, and some other CPU is reporting
2088 * this fact from process context. Do the remainder of the cleanup.
2089 * There can only be one CPU hotplug operation at a time, so no need for
2090 * explicit locking.
2091 */
2092int rcutree_dead_cpu(unsigned int cpu)
2093{
2094 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2095 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2096
2097 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2098 return 0;
2099
2100 /* Adjust any no-longer-needed kthreads. */
2101 rcu_boost_kthread_setaffinity(rnp, -1);
2102 /* Do any needed no-CB deferred wakeups from this CPU. */
2103 do_nocb_deferred_wakeup(per_cpu_ptr(&rcu_data, cpu));
2104 return 0;
2105}
2106
2107/*
2108 * Invoke any RCU callbacks that have made it to the end of their grace
2109 * period. Thottle as specified by rdp->blimit.
2110 */
2111static void rcu_do_batch(struct rcu_data *rdp)
2112{
2113 unsigned long flags;
2114 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2115 rcu_segcblist_is_offloaded(&rdp->cblist);
2116 struct rcu_head *rhp;
2117 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2118 long bl, count;
2119 long pending, tlimit = 0;
2120
2121 /* If no callbacks are ready, just return. */
2122 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2123 trace_rcu_batch_start(rcu_state.name,
2124 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2125 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2126 trace_rcu_batch_end(rcu_state.name, 0,
2127 !rcu_segcblist_empty(&rdp->cblist),
2128 need_resched(), is_idle_task(current),
2129 rcu_is_callbacks_kthread());
2130 return;
2131 }
2132
2133 /*
2134 * Extract the list of ready callbacks, disabling to prevent
2135 * races with call_rcu() from interrupt handlers. Leave the
2136 * callback counts, as rcu_barrier() needs to be conservative.
2137 */
2138 local_irq_save(flags);
2139 rcu_nocb_lock(rdp);
2140 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2141 pending = rcu_segcblist_n_cbs(&rdp->cblist);
2142 bl = max(rdp->blimit, pending >> rcu_divisor);
2143 if (unlikely(bl > 100))
2144 tlimit = local_clock() + rcu_resched_ns;
2145 trace_rcu_batch_start(rcu_state.name,
2146 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2147 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2148 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2149 if (offloaded)
2150 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2151 rcu_nocb_unlock_irqrestore(rdp, flags);
2152
2153 /* Invoke callbacks. */
2154 rhp = rcu_cblist_dequeue(&rcl);
2155 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2156 debug_rcu_head_unqueue(rhp);
2157 if (__rcu_reclaim(rcu_state.name, rhp))
2158 rcu_cblist_dequeued_lazy(&rcl);
2159 /*
2160 * Stop only if limit reached and CPU has something to do.
2161 * Note: The rcl structure counts down from zero.
2162 */
2163 if (-rcl.len >= bl && !offloaded &&
2164 (need_resched() ||
2165 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2166 break;
2167 if (unlikely(tlimit)) {
2168 /* only call local_clock() every 32 callbacks */
2169 if (likely((-rcl.len & 31) || local_clock() < tlimit))
2170 continue;
2171 /* Exceeded the time limit, so leave. */
2172 break;
2173 }
2174 if (offloaded) {
2175 WARN_ON_ONCE(in_serving_softirq());
2176 local_bh_enable();
2177 lockdep_assert_irqs_enabled();
2178 cond_resched_tasks_rcu_qs();
2179 lockdep_assert_irqs_enabled();
2180 local_bh_disable();
2181 }
2182 }
2183
2184 local_irq_save(flags);
2185 rcu_nocb_lock(rdp);
2186 count = -rcl.len;
2187 trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2188 is_idle_task(current), rcu_is_callbacks_kthread());
2189
2190 /* Update counts and requeue any remaining callbacks. */
2191 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2192 smp_mb(); /* List handling before counting for rcu_barrier(). */
2193 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2194
2195 /* Reinstate batch limit if we have worked down the excess. */
2196 count = rcu_segcblist_n_cbs(&rdp->cblist);
2197 if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
2198 rdp->blimit = blimit;
2199
2200 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2201 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2202 rdp->qlen_last_fqs_check = 0;
2203 rdp->n_force_qs_snap = rcu_state.n_force_qs;
2204 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2205 rdp->qlen_last_fqs_check = count;
2206
2207 /*
2208 * The following usually indicates a double call_rcu(). To track
2209 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2210 */
2211 WARN_ON_ONCE(count == 0 && !rcu_segcblist_empty(&rdp->cblist));
2212 WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2213 count != 0 && rcu_segcblist_empty(&rdp->cblist));
2214
2215 rcu_nocb_unlock_irqrestore(rdp, flags);
2216
2217 /* Re-invoke RCU core processing if there are callbacks remaining. */
2218 if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist))
2219 invoke_rcu_core();
2220}
2221
2222/*
2223 * This function is invoked from each scheduling-clock interrupt,
2224 * and checks to see if this CPU is in a non-context-switch quiescent
2225 * state, for example, user mode or idle loop. It also schedules RCU
2226 * core processing. If the current grace period has gone on too long,
2227 * it will ask the scheduler to manufacture a context switch for the sole
2228 * purpose of providing a providing the needed quiescent state.
2229 */
2230void rcu_sched_clock_irq(int user)
2231{
2232 trace_rcu_utilization(TPS("Start scheduler-tick"));
2233 raw_cpu_inc(rcu_data.ticks_this_gp);
2234 /* The load-acquire pairs with the store-release setting to true. */
2235 if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
2236 /* Idle and userspace execution already are quiescent states. */
2237 if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2238 set_tsk_need_resched(current);
2239 set_preempt_need_resched();
2240 }
2241 __this_cpu_write(rcu_data.rcu_urgent_qs, false);
2242 }
2243 rcu_flavor_sched_clock_irq(user);
2244 if (rcu_pending())
2245 invoke_rcu_core();
2246
2247 trace_rcu_utilization(TPS("End scheduler-tick"));
2248}
2249
2250/*
2251 * Scan the leaf rcu_node structures. For each structure on which all
2252 * CPUs have reported a quiescent state and on which there are tasks
2253 * blocking the current grace period, initiate RCU priority boosting.
2254 * Otherwise, invoke the specified function to check dyntick state for
2255 * each CPU that has not yet reported a quiescent state.
2256 */
2257static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
2258{
2259 int cpu;
2260 unsigned long flags;
2261 unsigned long mask;
2262 struct rcu_node *rnp;
2263
2264 rcu_for_each_leaf_node(rnp) {
2265 cond_resched_tasks_rcu_qs();
2266 mask = 0;
2267 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2268 if (rnp->qsmask == 0) {
2269 if (!IS_ENABLED(CONFIG_PREEMPTION) ||
2270 rcu_preempt_blocked_readers_cgp(rnp)) {
2271 /*
2272 * No point in scanning bits because they
2273 * are all zero. But we might need to
2274 * priority-boost blocked readers.
2275 */
2276 rcu_initiate_boost(rnp, flags);
2277 /* rcu_initiate_boost() releases rnp->lock */
2278 continue;
2279 }
2280 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2281 continue;
2282 }
2283 for_each_leaf_node_possible_cpu(rnp, cpu) {
2284 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
2285 if ((rnp->qsmask & bit) != 0) {
2286 if (f(per_cpu_ptr(&rcu_data, cpu)))
2287 mask |= bit;
2288 }
2289 }
2290 if (mask != 0) {
2291 /* Idle/offline CPUs, report (releases rnp->lock). */
2292 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2293 } else {
2294 /* Nothing to do here, so just drop the lock. */
2295 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2296 }
2297 }
2298}
2299
2300/*
2301 * Force quiescent states on reluctant CPUs, and also detect which
2302 * CPUs are in dyntick-idle mode.
2303 */
2304void rcu_force_quiescent_state(void)
2305{
2306 unsigned long flags;
2307 bool ret;
2308 struct rcu_node *rnp;
2309 struct rcu_node *rnp_old = NULL;
2310
2311 /* Funnel through hierarchy to reduce memory contention. */
2312 rnp = __this_cpu_read(rcu_data.mynode);
2313 for (; rnp != NULL; rnp = rnp->parent) {
2314 ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2315 !raw_spin_trylock(&rnp->fqslock);
2316 if (rnp_old != NULL)
2317 raw_spin_unlock(&rnp_old->fqslock);
2318 if (ret)
2319 return;
2320 rnp_old = rnp;
2321 }
2322 /* rnp_old == rcu_get_root(), rnp == NULL. */
2323
2324 /* Reached the root of the rcu_node tree, acquire lock. */
2325 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2326 raw_spin_unlock(&rnp_old->fqslock);
2327 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2328 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2329 return; /* Someone beat us to it. */
2330 }
2331 WRITE_ONCE(rcu_state.gp_flags,
2332 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2333 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2334 rcu_gp_kthread_wake();
2335}
2336EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
2337
2338/* Perform RCU core processing work for the current CPU. */
2339static __latent_entropy void rcu_core(void)
2340{
2341 unsigned long flags;
2342 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2343 struct rcu_node *rnp = rdp->mynode;
2344 const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2345 rcu_segcblist_is_offloaded(&rdp->cblist);
2346
2347 if (cpu_is_offline(smp_processor_id()))
2348 return;
2349 trace_rcu_utilization(TPS("Start RCU core"));
2350 WARN_ON_ONCE(!rdp->beenonline);
2351
2352 /* Report any deferred quiescent states if preemption enabled. */
2353 if (!(preempt_count() & PREEMPT_MASK)) {
2354 rcu_preempt_deferred_qs(current);
2355 } else if (rcu_preempt_need_deferred_qs(current)) {
2356 set_tsk_need_resched(current);
2357 set_preempt_need_resched();
2358 }
2359
2360 /* Update RCU state based on any recent quiescent states. */
2361 rcu_check_quiescent_state(rdp);
2362
2363 /* No grace period and unregistered callbacks? */
2364 if (!rcu_gp_in_progress() &&
2365 rcu_segcblist_is_enabled(&rdp->cblist) && !offloaded) {
2366 local_irq_save(flags);
2367 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2368 rcu_accelerate_cbs_unlocked(rnp, rdp);
2369 local_irq_restore(flags);
2370 }
2371
2372 rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
2373
2374 /* If there are callbacks ready, invoke them. */
2375 if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist) &&
2376 likely(READ_ONCE(rcu_scheduler_fully_active)))
2377 rcu_do_batch(rdp);
2378
2379 /* Do any needed deferred wakeups of rcuo kthreads. */
2380 do_nocb_deferred_wakeup(rdp);
2381 trace_rcu_utilization(TPS("End RCU core"));
2382}
2383
2384static void rcu_core_si(struct softirq_action *h)
2385{
2386 rcu_core();
2387}
2388
2389static void rcu_wake_cond(struct task_struct *t, int status)
2390{
2391 /*
2392 * If the thread is yielding, only wake it when this
2393 * is invoked from idle
2394 */
2395 if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
2396 wake_up_process(t);
2397}
2398
2399static void invoke_rcu_core_kthread(void)
2400{
2401 struct task_struct *t;
2402 unsigned long flags;
2403
2404 local_irq_save(flags);
2405 __this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
2406 t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
2407 if (t != NULL && t != current)
2408 rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
2409 local_irq_restore(flags);
2410}
2411
2412/*
2413 * Wake up this CPU's rcuc kthread to do RCU core processing.
2414 */
2415static void invoke_rcu_core(void)
2416{
2417 if (!cpu_online(smp_processor_id()))
2418 return;
2419 if (use_softirq)
2420 raise_softirq(RCU_SOFTIRQ);
2421 else
2422 invoke_rcu_core_kthread();
2423}
2424
2425static void rcu_cpu_kthread_park(unsigned int cpu)
2426{
2427 per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
2428}
2429
2430static int rcu_cpu_kthread_should_run(unsigned int cpu)
2431{
2432 return __this_cpu_read(rcu_data.rcu_cpu_has_work);
2433}
2434
2435/*
2436 * Per-CPU kernel thread that invokes RCU callbacks. This replaces
2437 * the RCU softirq used in configurations of RCU that do not support RCU
2438 * priority boosting.
2439 */
2440static void rcu_cpu_kthread(unsigned int cpu)
2441{
2442 unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
2443 char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
2444 int spincnt;
2445
2446 for (spincnt = 0; spincnt < 10; spincnt++) {
2447 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
2448 local_bh_disable();
2449 *statusp = RCU_KTHREAD_RUNNING;
2450 local_irq_disable();
2451 work = *workp;
2452 *workp = 0;
2453 local_irq_enable();
2454 if (work)
2455 rcu_core();
2456 local_bh_enable();
2457 if (*workp == 0) {
2458 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2459 *statusp = RCU_KTHREAD_WAITING;
2460 return;
2461 }
2462 }
2463 *statusp = RCU_KTHREAD_YIELDING;
2464 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2465 schedule_timeout_interruptible(2);
2466 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2467 *statusp = RCU_KTHREAD_WAITING;
2468}
2469
2470static struct smp_hotplug_thread rcu_cpu_thread_spec = {
2471 .store = &rcu_data.rcu_cpu_kthread_task,
2472 .thread_should_run = rcu_cpu_kthread_should_run,
2473 .thread_fn = rcu_cpu_kthread,
2474 .thread_comm = "rcuc/%u",
2475 .setup = rcu_cpu_kthread_setup,
2476 .park = rcu_cpu_kthread_park,
2477};
2478
2479/*
2480 * Spawn per-CPU RCU core processing kthreads.
2481 */
2482static int __init rcu_spawn_core_kthreads(void)
2483{
2484 int cpu;
2485
2486 for_each_possible_cpu(cpu)
2487 per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
2488 if (!IS_ENABLED(CONFIG_RCU_BOOST) && use_softirq)
2489 return 0;
2490 WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
2491 "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
2492 return 0;
2493}
2494early_initcall(rcu_spawn_core_kthreads);
2495
2496/*
2497 * Handle any core-RCU processing required by a call_rcu() invocation.
2498 */
2499static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2500 unsigned long flags)
2501{
2502 /*
2503 * If called from an extended quiescent state, invoke the RCU
2504 * core in order to force a re-evaluation of RCU's idleness.
2505 */
2506 if (!rcu_is_watching())
2507 invoke_rcu_core();
2508
2509 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2510 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2511 return;
2512
2513 /*
2514 * Force the grace period if too many callbacks or too long waiting.
2515 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2516 * if some other CPU has recently done so. Also, don't bother
2517 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2518 * is the only one waiting for a grace period to complete.
2519 */
2520 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2521 rdp->qlen_last_fqs_check + qhimark)) {
2522
2523 /* Are we ignoring a completed grace period? */
2524 note_gp_changes(rdp);
2525
2526 /* Start a new grace period if one not already started. */
2527 if (!rcu_gp_in_progress()) {
2528 rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
2529 } else {
2530 /* Give the grace period a kick. */
2531 rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
2532 if (rcu_state.n_force_qs == rdp->n_force_qs_snap &&
2533 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2534 rcu_force_quiescent_state();
2535 rdp->n_force_qs_snap = rcu_state.n_force_qs;
2536 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2537 }
2538 }
2539}
2540
2541/*
2542 * RCU callback function to leak a callback.
2543 */
2544static void rcu_leak_callback(struct rcu_head *rhp)
2545{
2546}
2547
2548/*
2549 * Helper function for call_rcu() and friends. The cpu argument will
2550 * normally be -1, indicating "currently running CPU". It may specify
2551 * a CPU only if that CPU is a no-CBs CPU. Currently, only rcu_barrier()
2552 * is expected to specify a CPU.
2553 */
2554static void
2555__call_rcu(struct rcu_head *head, rcu_callback_t func, bool lazy)
2556{
2557 unsigned long flags;
2558 struct rcu_data *rdp;
2559 bool was_alldone;
2560
2561 /* Misaligned rcu_head! */
2562 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2563
2564 if (debug_rcu_head_queue(head)) {
2565 /*
2566 * Probable double call_rcu(), so leak the callback.
2567 * Use rcu:rcu_callback trace event to find the previous
2568 * time callback was passed to __call_rcu().
2569 */
2570 WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pS()!!!\n",
2571 head, head->func);
2572 WRITE_ONCE(head->func, rcu_leak_callback);
2573 return;
2574 }
2575 head->func = func;
2576 head->next = NULL;
2577 local_irq_save(flags);
2578 rdp = this_cpu_ptr(&rcu_data);
2579
2580 /* Add the callback to our list. */
2581 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
2582 // This can trigger due to call_rcu() from offline CPU:
2583 WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
2584 WARN_ON_ONCE(!rcu_is_watching());
2585 // Very early boot, before rcu_init(). Initialize if needed
2586 // and then drop through to queue the callback.
2587 if (rcu_segcblist_empty(&rdp->cblist))
2588 rcu_segcblist_init(&rdp->cblist);
2589 }
2590 if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags))
2591 return; // Enqueued onto ->nocb_bypass, so just leave.
2592 /* If we get here, rcu_nocb_try_bypass() acquired ->nocb_lock. */
2593 rcu_segcblist_enqueue(&rdp->cblist, head, lazy);
2594 if (__is_kfree_rcu_offset((unsigned long)func))
2595 trace_rcu_kfree_callback(rcu_state.name, head,
2596 (unsigned long)func,
2597 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2598 rcu_segcblist_n_cbs(&rdp->cblist));
2599 else
2600 trace_rcu_callback(rcu_state.name, head,
2601 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2602 rcu_segcblist_n_cbs(&rdp->cblist));
2603
2604 /* Go handle any RCU core processing required. */
2605 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2606 unlikely(rcu_segcblist_is_offloaded(&rdp->cblist))) {
2607 __call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */
2608 } else {
2609 __call_rcu_core(rdp, head, flags);
2610 local_irq_restore(flags);
2611 }
2612}
2613
2614/**
2615 * call_rcu() - Queue an RCU callback for invocation after a grace period.
2616 * @head: structure to be used for queueing the RCU updates.
2617 * @func: actual callback function to be invoked after the grace period
2618 *
2619 * The callback function will be invoked some time after a full grace
2620 * period elapses, in other words after all pre-existing RCU read-side
2621 * critical sections have completed. However, the callback function
2622 * might well execute concurrently with RCU read-side critical sections
2623 * that started after call_rcu() was invoked. RCU read-side critical
2624 * sections are delimited by rcu_read_lock() and rcu_read_unlock(), and
2625 * may be nested. In addition, regions of code across which interrupts,
2626 * preemption, or softirqs have been disabled also serve as RCU read-side
2627 * critical sections. This includes hardware interrupt handlers, softirq
2628 * handlers, and NMI handlers.
2629 *
2630 * Note that all CPUs must agree that the grace period extended beyond
2631 * all pre-existing RCU read-side critical section. On systems with more
2632 * than one CPU, this means that when "func()" is invoked, each CPU is
2633 * guaranteed to have executed a full memory barrier since the end of its
2634 * last RCU read-side critical section whose beginning preceded the call
2635 * to call_rcu(). It also means that each CPU executing an RCU read-side
2636 * critical section that continues beyond the start of "func()" must have
2637 * executed a memory barrier after the call_rcu() but before the beginning
2638 * of that RCU read-side critical section. Note that these guarantees
2639 * include CPUs that are offline, idle, or executing in user mode, as
2640 * well as CPUs that are executing in the kernel.
2641 *
2642 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
2643 * resulting RCU callback function "func()", then both CPU A and CPU B are
2644 * guaranteed to execute a full memory barrier during the time interval
2645 * between the call to call_rcu() and the invocation of "func()" -- even
2646 * if CPU A and CPU B are the same CPU (but again only if the system has
2647 * more than one CPU).
2648 */
2649void call_rcu(struct rcu_head *head, rcu_callback_t func)
2650{
2651 __call_rcu(head, func, 0);
2652}
2653EXPORT_SYMBOL_GPL(call_rcu);
2654
2655/*
2656 * Queue an RCU callback for lazy invocation after a grace period.
2657 * This will likely be later named something like "call_rcu_lazy()",
2658 * but this change will require some way of tagging the lazy RCU
2659 * callbacks in the list of pending callbacks. Until then, this
2660 * function may only be called from __kfree_rcu().
2661 */
2662void kfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
2663{
2664 __call_rcu(head, func, 1);
2665}
2666EXPORT_SYMBOL_GPL(kfree_call_rcu);
2667
2668/*
2669 * During early boot, any blocking grace-period wait automatically
2670 * implies a grace period. Later on, this is never the case for PREEMPT.
2671 *
2672 * Howevr, because a context switch is a grace period for !PREEMPT, any
2673 * blocking grace-period wait automatically implies a grace period if
2674 * there is only one CPU online at any point time during execution of
2675 * either synchronize_rcu() or synchronize_rcu_expedited(). It is OK to
2676 * occasionally incorrectly indicate that there are multiple CPUs online
2677 * when there was in fact only one the whole time, as this just adds some
2678 * overhead: RCU still operates correctly.
2679 */
2680static int rcu_blocking_is_gp(void)
2681{
2682 int ret;
2683
2684 if (IS_ENABLED(CONFIG_PREEMPTION))
2685 return rcu_scheduler_active == RCU_SCHEDULER_INACTIVE;
2686 might_sleep(); /* Check for RCU read-side critical section. */
2687 preempt_disable();
2688 ret = num_online_cpus() <= 1;
2689 preempt_enable();
2690 return ret;
2691}
2692
2693/**
2694 * synchronize_rcu - wait until a grace period has elapsed.
2695 *
2696 * Control will return to the caller some time after a full grace
2697 * period has elapsed, in other words after all currently executing RCU
2698 * read-side critical sections have completed. Note, however, that
2699 * upon return from synchronize_rcu(), the caller might well be executing
2700 * concurrently with new RCU read-side critical sections that began while
2701 * synchronize_rcu() was waiting. RCU read-side critical sections are
2702 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
2703 * In addition, regions of code across which interrupts, preemption, or
2704 * softirqs have been disabled also serve as RCU read-side critical
2705 * sections. This includes hardware interrupt handlers, softirq handlers,
2706 * and NMI handlers.
2707 *
2708 * Note that this guarantee implies further memory-ordering guarantees.
2709 * On systems with more than one CPU, when synchronize_rcu() returns,
2710 * each CPU is guaranteed to have executed a full memory barrier since
2711 * the end of its last RCU read-side critical section whose beginning
2712 * preceded the call to synchronize_rcu(). In addition, each CPU having
2713 * an RCU read-side critical section that extends beyond the return from
2714 * synchronize_rcu() is guaranteed to have executed a full memory barrier
2715 * after the beginning of synchronize_rcu() and before the beginning of
2716 * that RCU read-side critical section. Note that these guarantees include
2717 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2718 * that are executing in the kernel.
2719 *
2720 * Furthermore, if CPU A invoked synchronize_rcu(), which returned
2721 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2722 * to have executed a full memory barrier during the execution of
2723 * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
2724 * again only if the system has more than one CPU).
2725 */
2726void synchronize_rcu(void)
2727{
2728 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
2729 lock_is_held(&rcu_lock_map) ||
2730 lock_is_held(&rcu_sched_lock_map),
2731 "Illegal synchronize_rcu() in RCU read-side critical section");
2732 if (rcu_blocking_is_gp())
2733 return;
2734 if (rcu_gp_is_expedited())
2735 synchronize_rcu_expedited();
2736 else
2737 wait_rcu_gp(call_rcu);
2738}
2739EXPORT_SYMBOL_GPL(synchronize_rcu);
2740
2741/**
2742 * get_state_synchronize_rcu - Snapshot current RCU state
2743 *
2744 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
2745 * to determine whether or not a full grace period has elapsed in the
2746 * meantime.
2747 */
2748unsigned long get_state_synchronize_rcu(void)
2749{
2750 /*
2751 * Any prior manipulation of RCU-protected data must happen
2752 * before the load from ->gp_seq.
2753 */
2754 smp_mb(); /* ^^^ */
2755 return rcu_seq_snap(&rcu_state.gp_seq);
2756}
2757EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
2758
2759/**
2760 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
2761 *
2762 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
2763 *
2764 * If a full RCU grace period has elapsed since the earlier call to
2765 * get_state_synchronize_rcu(), just return. Otherwise, invoke
2766 * synchronize_rcu() to wait for a full grace period.
2767 *
2768 * Yes, this function does not take counter wrap into account. But
2769 * counter wrap is harmless. If the counter wraps, we have waited for
2770 * more than 2 billion grace periods (and way more on a 64-bit system!),
2771 * so waiting for one additional grace period should be just fine.
2772 */
2773void cond_synchronize_rcu(unsigned long oldstate)
2774{
2775 if (!rcu_seq_done(&rcu_state.gp_seq, oldstate))
2776 synchronize_rcu();
2777 else
2778 smp_mb(); /* Ensure GP ends before subsequent accesses. */
2779}
2780EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
2781
2782/*
2783 * Check to see if there is any immediate RCU-related work to be done by
2784 * the current CPU, returning 1 if so and zero otherwise. The checks are
2785 * in order of increasing expense: checks that can be carried out against
2786 * CPU-local state are performed first. However, we must check for CPU
2787 * stalls first, else we might not get a chance.
2788 */
2789static int rcu_pending(void)
2790{
2791 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
2792 struct rcu_node *rnp = rdp->mynode;
2793
2794 /* Check for CPU stalls, if enabled. */
2795 check_cpu_stall(rdp);
2796
2797 /* Does this CPU need a deferred NOCB wakeup? */
2798 if (rcu_nocb_need_deferred_wakeup(rdp))
2799 return 1;
2800
2801 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
2802 if (rcu_nohz_full_cpu())
2803 return 0;
2804
2805 /* Is the RCU core waiting for a quiescent state from this CPU? */
2806 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm)
2807 return 1;
2808
2809 /* Does this CPU have callbacks ready to invoke? */
2810 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2811 return 1;
2812
2813 /* Has RCU gone idle with this CPU needing another grace period? */
2814 if (!rcu_gp_in_progress() &&
2815 rcu_segcblist_is_enabled(&rdp->cblist) &&
2816 (!IS_ENABLED(CONFIG_RCU_NOCB_CPU) ||
2817 !rcu_segcblist_is_offloaded(&rdp->cblist)) &&
2818 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2819 return 1;
2820
2821 /* Have RCU grace period completed or started? */
2822 if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
2823 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
2824 return 1;
2825
2826 /* nothing to do */
2827 return 0;
2828}
2829
2830/*
2831 * Helper function for rcu_barrier() tracing. If tracing is disabled,
2832 * the compiler is expected to optimize this away.
2833 */
2834static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
2835{
2836 trace_rcu_barrier(rcu_state.name, s, cpu,
2837 atomic_read(&rcu_state.barrier_cpu_count), done);
2838}
2839
2840/*
2841 * RCU callback function for rcu_barrier(). If we are last, wake
2842 * up the task executing rcu_barrier().
2843 */
2844static void rcu_barrier_callback(struct rcu_head *rhp)
2845{
2846 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
2847 rcu_barrier_trace(TPS("LastCB"), -1,
2848 rcu_state.barrier_sequence);
2849 complete(&rcu_state.barrier_completion);
2850 } else {
2851 rcu_barrier_trace(TPS("CB"), -1, rcu_state.barrier_sequence);
2852 }
2853}
2854
2855/*
2856 * Called with preemption disabled, and from cross-cpu IRQ context.
2857 */
2858static void rcu_barrier_func(void *unused)
2859{
2860 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2861
2862 rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
2863 rdp->barrier_head.func = rcu_barrier_callback;
2864 debug_rcu_head_queue(&rdp->barrier_head);
2865 rcu_nocb_lock(rdp);
2866 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
2867 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head, 0)) {
2868 atomic_inc(&rcu_state.barrier_cpu_count);
2869 } else {
2870 debug_rcu_head_unqueue(&rdp->barrier_head);
2871 rcu_barrier_trace(TPS("IRQNQ"), -1,
2872 rcu_state.barrier_sequence);
2873 }
2874 rcu_nocb_unlock(rdp);
2875}
2876
2877/**
2878 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
2879 *
2880 * Note that this primitive does not necessarily wait for an RCU grace period
2881 * to complete. For example, if there are no RCU callbacks queued anywhere
2882 * in the system, then rcu_barrier() is within its rights to return
2883 * immediately, without waiting for anything, much less an RCU grace period.
2884 */
2885void rcu_barrier(void)
2886{
2887 int cpu;
2888 struct rcu_data *rdp;
2889 unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
2890
2891 rcu_barrier_trace(TPS("Begin"), -1, s);
2892
2893 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2894 mutex_lock(&rcu_state.barrier_mutex);
2895
2896 /* Did someone else do our work for us? */
2897 if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
2898 rcu_barrier_trace(TPS("EarlyExit"), -1,
2899 rcu_state.barrier_sequence);
2900 smp_mb(); /* caller's subsequent code after above check. */
2901 mutex_unlock(&rcu_state.barrier_mutex);
2902 return;
2903 }
2904
2905 /* Mark the start of the barrier operation. */
2906 rcu_seq_start(&rcu_state.barrier_sequence);
2907 rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
2908
2909 /*
2910 * Initialize the count to one rather than to zero in order to
2911 * avoid a too-soon return to zero in case of a short grace period
2912 * (or preemption of this task). Exclude CPU-hotplug operations
2913 * to ensure that no offline CPU has callbacks queued.
2914 */
2915 init_completion(&rcu_state.barrier_completion);
2916 atomic_set(&rcu_state.barrier_cpu_count, 1);
2917 get_online_cpus();
2918
2919 /*
2920 * Force each CPU with callbacks to register a new callback.
2921 * When that callback is invoked, we will know that all of the
2922 * corresponding CPU's preceding callbacks have been invoked.
2923 */
2924 for_each_possible_cpu(cpu) {
2925 rdp = per_cpu_ptr(&rcu_data, cpu);
2926 if (!cpu_online(cpu) &&
2927 !rcu_segcblist_is_offloaded(&rdp->cblist))
2928 continue;
2929 if (rcu_segcblist_n_cbs(&rdp->cblist)) {
2930 rcu_barrier_trace(TPS("OnlineQ"), cpu,
2931 rcu_state.barrier_sequence);
2932 smp_call_function_single(cpu, rcu_barrier_func, NULL, 1);
2933 } else {
2934 rcu_barrier_trace(TPS("OnlineNQ"), cpu,
2935 rcu_state.barrier_sequence);
2936 }
2937 }
2938 put_online_cpus();
2939
2940 /*
2941 * Now that we have an rcu_barrier_callback() callback on each
2942 * CPU, and thus each counted, remove the initial count.
2943 */
2944 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count))
2945 complete(&rcu_state.barrier_completion);
2946
2947 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2948 wait_for_completion(&rcu_state.barrier_completion);
2949
2950 /* Mark the end of the barrier operation. */
2951 rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
2952 rcu_seq_end(&rcu_state.barrier_sequence);
2953
2954 /* Other rcu_barrier() invocations can now safely proceed. */
2955 mutex_unlock(&rcu_state.barrier_mutex);
2956}
2957EXPORT_SYMBOL_GPL(rcu_barrier);
2958
2959/*
2960 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
2961 * first CPU in a given leaf rcu_node structure coming online. The caller
2962 * must hold the corresponding leaf rcu_node ->lock with interrrupts
2963 * disabled.
2964 */
2965static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
2966{
2967 long mask;
2968 long oldmask;
2969 struct rcu_node *rnp = rnp_leaf;
2970
2971 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2972 WARN_ON_ONCE(rnp->wait_blkd_tasks);
2973 for (;;) {
2974 mask = rnp->grpmask;
2975 rnp = rnp->parent;
2976 if (rnp == NULL)
2977 return;
2978 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
2979 oldmask = rnp->qsmaskinit;
2980 rnp->qsmaskinit |= mask;
2981 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
2982 if (oldmask)
2983 return;
2984 }
2985}
2986
2987/*
2988 * Do boot-time initialization of a CPU's per-CPU RCU data.
2989 */
2990static void __init
2991rcu_boot_init_percpu_data(int cpu)
2992{
2993 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2994
2995 /* Set up local state, ensuring consistent view of global state. */
2996 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
2997 WARN_ON_ONCE(rdp->dynticks_nesting != 1);
2998 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp)));
2999 rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
3000 rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
3001 rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
3002 rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
3003 rdp->cpu = cpu;
3004 rcu_boot_init_nocb_percpu_data(rdp);
3005}
3006
3007/*
3008 * Invoked early in the CPU-online process, when pretty much all services
3009 * are available. The incoming CPU is not present.
3010 *
3011 * Initializes a CPU's per-CPU RCU data. Note that only one online or
3012 * offline event can be happening at a given time. Note also that we can
3013 * accept some slop in the rsp->gp_seq access due to the fact that this
3014 * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
3015 * And any offloaded callbacks are being numbered elsewhere.
3016 */
3017int rcutree_prepare_cpu(unsigned int cpu)
3018{
3019 unsigned long flags;
3020 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3021 struct rcu_node *rnp = rcu_get_root();
3022
3023 /* Set up local state, ensuring consistent view of global state. */
3024 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3025 rdp->qlen_last_fqs_check = 0;
3026 rdp->n_force_qs_snap = rcu_state.n_force_qs;
3027 rdp->blimit = blimit;
3028 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3029 !rcu_segcblist_is_offloaded(&rdp->cblist))
3030 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
3031 rdp->dynticks_nesting = 1; /* CPU not up, no tearing. */
3032 rcu_dynticks_eqs_online();
3033 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3034
3035 /*
3036 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3037 * propagation up the rcu_node tree will happen at the beginning
3038 * of the next grace period.
3039 */
3040 rnp = rdp->mynode;
3041 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3042 rdp->beenonline = true; /* We have now been online. */
3043 rdp->gp_seq = rnp->gp_seq;
3044 rdp->gp_seq_needed = rnp->gp_seq;
3045 rdp->cpu_no_qs.b.norm = true;
3046 rdp->core_needs_qs = false;
3047 rdp->rcu_iw_pending = false;
3048 rdp->rcu_iw_gp_seq = rnp->gp_seq - 1;
3049 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
3050 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3051 rcu_prepare_kthreads(cpu);
3052 rcu_spawn_cpu_nocb_kthread(cpu);
3053
3054 return 0;
3055}
3056
3057/*
3058 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3059 */
3060static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3061{
3062 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3063
3064 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3065}
3066
3067/*
3068 * Near the end of the CPU-online process. Pretty much all services
3069 * enabled, and the CPU is now very much alive.
3070 */
3071int rcutree_online_cpu(unsigned int cpu)
3072{
3073 unsigned long flags;
3074 struct rcu_data *rdp;
3075 struct rcu_node *rnp;
3076
3077 rdp = per_cpu_ptr(&rcu_data, cpu);
3078 rnp = rdp->mynode;
3079 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3080 rnp->ffmask |= rdp->grpmask;
3081 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3082 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
3083 return 0; /* Too early in boot for scheduler work. */
3084 sync_sched_exp_online_cleanup(cpu);
3085 rcutree_affinity_setting(cpu, -1);
3086 return 0;
3087}
3088
3089/*
3090 * Near the beginning of the process. The CPU is still very much alive
3091 * with pretty much all services enabled.
3092 */
3093int rcutree_offline_cpu(unsigned int cpu)
3094{
3095 unsigned long flags;
3096 struct rcu_data *rdp;
3097 struct rcu_node *rnp;
3098
3099 rdp = per_cpu_ptr(&rcu_data, cpu);
3100 rnp = rdp->mynode;
3101 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3102 rnp->ffmask &= ~rdp->grpmask;
3103 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3104
3105 rcutree_affinity_setting(cpu, cpu);
3106 return 0;
3107}
3108
3109static DEFINE_PER_CPU(int, rcu_cpu_started);
3110
3111/*
3112 * Mark the specified CPU as being online so that subsequent grace periods
3113 * (both expedited and normal) will wait on it. Note that this means that
3114 * incoming CPUs are not allowed to use RCU read-side critical sections
3115 * until this function is called. Failing to observe this restriction
3116 * will result in lockdep splats.
3117 *
3118 * Note that this function is special in that it is invoked directly
3119 * from the incoming CPU rather than from the cpuhp_step mechanism.
3120 * This is because this function must be invoked at a precise location.
3121 */
3122void rcu_cpu_starting(unsigned int cpu)
3123{
3124 unsigned long flags;
3125 unsigned long mask;
3126 int nbits;
3127 unsigned long oldmask;
3128 struct rcu_data *rdp;
3129 struct rcu_node *rnp;
3130
3131 if (per_cpu(rcu_cpu_started, cpu))
3132 return;
3133
3134 per_cpu(rcu_cpu_started, cpu) = 1;
3135
3136 rdp = per_cpu_ptr(&rcu_data, cpu);
3137 rnp = rdp->mynode;
3138 mask = rdp->grpmask;
3139 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3140 rnp->qsmaskinitnext |= mask;
3141 oldmask = rnp->expmaskinitnext;
3142 rnp->expmaskinitnext |= mask;
3143 oldmask ^= rnp->expmaskinitnext;
3144 nbits = bitmap_weight(&oldmask, BITS_PER_LONG);
3145 /* Allow lockless access for expedited grace periods. */
3146 smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + nbits); /* ^^^ */
3147 rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
3148 rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
3149 rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
3150 if (rnp->qsmask & mask) { /* RCU waiting on incoming CPU? */
3151 /* Report QS -after- changing ->qsmaskinitnext! */
3152 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
3153 } else {
3154 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3155 }
3156 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
3157}
3158
3159#ifdef CONFIG_HOTPLUG_CPU
3160/*
3161 * The outgoing function has no further need of RCU, so remove it from
3162 * the rcu_node tree's ->qsmaskinitnext bit masks.
3163 *
3164 * Note that this function is special in that it is invoked directly
3165 * from the outgoing CPU rather than from the cpuhp_step mechanism.
3166 * This is because this function must be invoked at a precise location.
3167 */
3168void rcu_report_dead(unsigned int cpu)
3169{
3170 unsigned long flags;
3171 unsigned long mask;
3172 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3173 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3174
3175 /* QS for any half-done expedited grace period. */
3176 preempt_disable();
3177 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
3178 preempt_enable();
3179 rcu_preempt_deferred_qs(current);
3180
3181 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3182 mask = rdp->grpmask;
3183 raw_spin_lock(&rcu_state.ofl_lock);
3184 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3185 rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
3186 rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
3187 if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
3188 /* Report quiescent state -before- changing ->qsmaskinitnext! */
3189 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
3190 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3191 }
3192 rnp->qsmaskinitnext &= ~mask;
3193 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3194 raw_spin_unlock(&rcu_state.ofl_lock);
3195
3196 per_cpu(rcu_cpu_started, cpu) = 0;
3197}
3198
3199/*
3200 * The outgoing CPU has just passed through the dying-idle state, and we
3201 * are being invoked from the CPU that was IPIed to continue the offline
3202 * operation. Migrate the outgoing CPU's callbacks to the current CPU.
3203 */
3204void rcutree_migrate_callbacks(int cpu)
3205{
3206 unsigned long flags;
3207 struct rcu_data *my_rdp;
3208 struct rcu_node *my_rnp;
3209 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3210 bool needwake;
3211
3212 if (rcu_segcblist_is_offloaded(&rdp->cblist) ||
3213 rcu_segcblist_empty(&rdp->cblist))
3214 return; /* No callbacks to migrate. */
3215
3216 local_irq_save(flags);
3217 my_rdp = this_cpu_ptr(&rcu_data);
3218 my_rnp = my_rdp->mynode;
3219 rcu_nocb_lock(my_rdp); /* irqs already disabled. */
3220 WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies));
3221 raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
3222 /* Leverage recent GPs and set GP for new callbacks. */
3223 needwake = rcu_advance_cbs(my_rnp, rdp) ||
3224 rcu_advance_cbs(my_rnp, my_rdp);
3225 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
3226 needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
3227 rcu_segcblist_disable(&rdp->cblist);
3228 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
3229 !rcu_segcblist_n_cbs(&my_rdp->cblist));
3230 if (rcu_segcblist_is_offloaded(&my_rdp->cblist)) {
3231 raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
3232 __call_rcu_nocb_wake(my_rdp, true, flags);
3233 } else {
3234 rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
3235 raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags);
3236 }
3237 if (needwake)
3238 rcu_gp_kthread_wake();
3239 lockdep_assert_irqs_enabled();
3240 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
3241 !rcu_segcblist_empty(&rdp->cblist),
3242 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
3243 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
3244 rcu_segcblist_first_cb(&rdp->cblist));
3245}
3246#endif
3247
3248/*
3249 * On non-huge systems, use expedited RCU grace periods to make suspend
3250 * and hibernation run faster.
3251 */
3252static int rcu_pm_notify(struct notifier_block *self,
3253 unsigned long action, void *hcpu)
3254{
3255 switch (action) {
3256 case PM_HIBERNATION_PREPARE:
3257 case PM_SUSPEND_PREPARE:
3258 rcu_expedite_gp();
3259 break;
3260 case PM_POST_HIBERNATION:
3261 case PM_POST_SUSPEND:
3262 rcu_unexpedite_gp();
3263 break;
3264 default:
3265 break;
3266 }
3267 return NOTIFY_OK;
3268}
3269
3270/*
3271 * Spawn the kthreads that handle RCU's grace periods.
3272 */
3273static int __init rcu_spawn_gp_kthread(void)
3274{
3275 unsigned long flags;
3276 int kthread_prio_in = kthread_prio;
3277 struct rcu_node *rnp;
3278 struct sched_param sp;
3279 struct task_struct *t;
3280
3281 /* Force priority into range. */
3282 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
3283 && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
3284 kthread_prio = 2;
3285 else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
3286 kthread_prio = 1;
3287 else if (kthread_prio < 0)
3288 kthread_prio = 0;
3289 else if (kthread_prio > 99)
3290 kthread_prio = 99;
3291
3292 if (kthread_prio != kthread_prio_in)
3293 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3294 kthread_prio, kthread_prio_in);
3295
3296 rcu_scheduler_fully_active = 1;
3297 t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
3298 if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
3299 return 0;
3300 if (kthread_prio) {
3301 sp.sched_priority = kthread_prio;
3302 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3303 }
3304 rnp = rcu_get_root();
3305 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3306 rcu_state.gp_kthread = t;
3307 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3308 wake_up_process(t);
3309 rcu_spawn_nocb_kthreads();
3310 rcu_spawn_boost_kthreads();
3311 return 0;
3312}
3313early_initcall(rcu_spawn_gp_kthread);
3314
3315/*
3316 * This function is invoked towards the end of the scheduler's
3317 * initialization process. Before this is called, the idle task might
3318 * contain synchronous grace-period primitives (during which time, this idle
3319 * task is booting the system, and such primitives are no-ops). After this
3320 * function is called, any synchronous grace-period primitives are run as
3321 * expedited, with the requesting task driving the grace period forward.
3322 * A later core_initcall() rcu_set_runtime_mode() will switch to full
3323 * runtime RCU functionality.
3324 */
3325void rcu_scheduler_starting(void)
3326{
3327 WARN_ON(num_online_cpus() != 1);
3328 WARN_ON(nr_context_switches() > 0);
3329 rcu_test_sync_prims();
3330 rcu_scheduler_active = RCU_SCHEDULER_INIT;
3331 rcu_test_sync_prims();
3332}
3333
3334/*
3335 * Helper function for rcu_init() that initializes the rcu_state structure.
3336 */
3337static void __init rcu_init_one(void)
3338{
3339 static const char * const buf[] = RCU_NODE_NAME_INIT;
3340 static const char * const fqs[] = RCU_FQS_NAME_INIT;
3341 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
3342 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
3343
3344 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
3345 int cpustride = 1;
3346 int i;
3347 int j;
3348 struct rcu_node *rnp;
3349
3350 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3351
3352 /* Silence gcc 4.8 false positive about array index out of range. */
3353 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
3354 panic("rcu_init_one: rcu_num_lvls out of range");
3355
3356 /* Initialize the level-tracking arrays. */
3357
3358 for (i = 1; i < rcu_num_lvls; i++)
3359 rcu_state.level[i] =
3360 rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
3361 rcu_init_levelspread(levelspread, num_rcu_lvl);
3362
3363 /* Initialize the elements themselves, starting from the leaves. */
3364
3365 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3366 cpustride *= levelspread[i];
3367 rnp = rcu_state.level[i];
3368 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
3369 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
3370 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
3371 &rcu_node_class[i], buf[i]);
3372 raw_spin_lock_init(&rnp->fqslock);
3373 lockdep_set_class_and_name(&rnp->fqslock,
3374 &rcu_fqs_class[i], fqs[i]);
3375 rnp->gp_seq = rcu_state.gp_seq;
3376 rnp->gp_seq_needed = rcu_state.gp_seq;
3377 rnp->completedqs = rcu_state.gp_seq;
3378 rnp->qsmask = 0;
3379 rnp->qsmaskinit = 0;
3380 rnp->grplo = j * cpustride;
3381 rnp->grphi = (j + 1) * cpustride - 1;
3382 if (rnp->grphi >= nr_cpu_ids)
3383 rnp->grphi = nr_cpu_ids - 1;
3384 if (i == 0) {
3385 rnp->grpnum = 0;
3386 rnp->grpmask = 0;
3387 rnp->parent = NULL;
3388 } else {
3389 rnp->grpnum = j % levelspread[i - 1];
3390 rnp->grpmask = BIT(rnp->grpnum);
3391 rnp->parent = rcu_state.level[i - 1] +
3392 j / levelspread[i - 1];
3393 }
3394 rnp->level = i;
3395 INIT_LIST_HEAD(&rnp->blkd_tasks);
3396 rcu_init_one_nocb(rnp);
3397 init_waitqueue_head(&rnp->exp_wq[0]);
3398 init_waitqueue_head(&rnp->exp_wq[1]);
3399 init_waitqueue_head(&rnp->exp_wq[2]);
3400 init_waitqueue_head(&rnp->exp_wq[3]);
3401 spin_lock_init(&rnp->exp_lock);
3402 }
3403 }
3404
3405 init_swait_queue_head(&rcu_state.gp_wq);
3406 init_swait_queue_head(&rcu_state.expedited_wq);
3407 rnp = rcu_first_leaf_node();
3408 for_each_possible_cpu(i) {
3409 while (i > rnp->grphi)
3410 rnp++;
3411 per_cpu_ptr(&rcu_data, i)->mynode = rnp;
3412 rcu_boot_init_percpu_data(i);
3413 }
3414}
3415
3416/*
3417 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3418 * replace the definitions in tree.h because those are needed to size
3419 * the ->node array in the rcu_state structure.
3420 */
3421static void __init rcu_init_geometry(void)
3422{
3423 ulong d;
3424 int i;
3425 int rcu_capacity[RCU_NUM_LVLS];
3426
3427 /*
3428 * Initialize any unspecified boot parameters.
3429 * The default values of jiffies_till_first_fqs and
3430 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3431 * value, which is a function of HZ, then adding one for each
3432 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3433 */
3434 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3435 if (jiffies_till_first_fqs == ULONG_MAX)
3436 jiffies_till_first_fqs = d;
3437 if (jiffies_till_next_fqs == ULONG_MAX)
3438 jiffies_till_next_fqs = d;
3439 adjust_jiffies_till_sched_qs();
3440
3441 /* If the compile-time values are accurate, just leave. */
3442 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
3443 nr_cpu_ids == NR_CPUS)
3444 return;
3445 pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
3446 rcu_fanout_leaf, nr_cpu_ids);
3447
3448 /*
3449 * The boot-time rcu_fanout_leaf parameter must be at least two
3450 * and cannot exceed the number of bits in the rcu_node masks.
3451 * Complain and fall back to the compile-time values if this
3452 * limit is exceeded.
3453 */
3454 if (rcu_fanout_leaf < 2 ||
3455 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
3456 rcu_fanout_leaf = RCU_FANOUT_LEAF;
3457 WARN_ON(1);
3458 return;
3459 }
3460
3461 /*
3462 * Compute number of nodes that can be handled an rcu_node tree
3463 * with the given number of levels.
3464 */
3465 rcu_capacity[0] = rcu_fanout_leaf;
3466 for (i = 1; i < RCU_NUM_LVLS; i++)
3467 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
3468
3469 /*
3470 * The tree must be able to accommodate the configured number of CPUs.
3471 * If this limit is exceeded, fall back to the compile-time values.
3472 */
3473 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
3474 rcu_fanout_leaf = RCU_FANOUT_LEAF;
3475 WARN_ON(1);
3476 return;
3477 }
3478
3479 /* Calculate the number of levels in the tree. */
3480 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
3481 }
3482 rcu_num_lvls = i + 1;
3483
3484 /* Calculate the number of rcu_nodes at each level of the tree. */
3485 for (i = 0; i < rcu_num_lvls; i++) {
3486 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
3487 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
3488 }
3489
3490 /* Calculate the total number of rcu_node structures. */
3491 rcu_num_nodes = 0;
3492 for (i = 0; i < rcu_num_lvls; i++)
3493 rcu_num_nodes += num_rcu_lvl[i];
3494}
3495
3496/*
3497 * Dump out the structure of the rcu_node combining tree associated
3498 * with the rcu_state structure.
3499 */
3500static void __init rcu_dump_rcu_node_tree(void)
3501{
3502 int level = 0;
3503 struct rcu_node *rnp;
3504
3505 pr_info("rcu_node tree layout dump\n");
3506 pr_info(" ");
3507 rcu_for_each_node_breadth_first(rnp) {
3508 if (rnp->level != level) {
3509 pr_cont("\n");
3510 pr_info(" ");
3511 level = rnp->level;
3512 }
3513 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
3514 }
3515 pr_cont("\n");
3516}
3517
3518struct workqueue_struct *rcu_gp_wq;
3519struct workqueue_struct *rcu_par_gp_wq;
3520
3521void __init rcu_init(void)
3522{
3523 int cpu;
3524
3525 rcu_early_boot_tests();
3526
3527 rcu_bootup_announce();
3528 rcu_init_geometry();
3529 rcu_init_one();
3530 if (dump_tree)
3531 rcu_dump_rcu_node_tree();
3532 if (use_softirq)
3533 open_softirq(RCU_SOFTIRQ, rcu_core_si);
3534
3535 /*
3536 * We don't need protection against CPU-hotplug here because
3537 * this is called early in boot, before either interrupts
3538 * or the scheduler are operational.
3539 */
3540 pm_notifier(rcu_pm_notify, 0);
3541 for_each_online_cpu(cpu) {
3542 rcutree_prepare_cpu(cpu);
3543 rcu_cpu_starting(cpu);
3544 rcutree_online_cpu(cpu);
3545 }
3546
3547 /* Create workqueue for expedited GPs and for Tree SRCU. */
3548 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
3549 WARN_ON(!rcu_gp_wq);
3550 rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
3551 WARN_ON(!rcu_par_gp_wq);
3552 srcu_init();
3553}
3554
3555#include "tree_stall.h"
3556#include "tree_exp.h"
3557#include "tree_plugin.h"
1// 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/kmemleak.h>
35#include <linux/moduleparam.h>
36#include <linux/panic.h>
37#include <linux/panic_notifier.h>
38#include <linux/percpu.h>
39#include <linux/notifier.h>
40#include <linux/cpu.h>
41#include <linux/mutex.h>
42#include <linux/time.h>
43#include <linux/kernel_stat.h>
44#include <linux/wait.h>
45#include <linux/kthread.h>
46#include <uapi/linux/sched/types.h>
47#include <linux/prefetch.h>
48#include <linux/delay.h>
49#include <linux/random.h>
50#include <linux/trace_events.h>
51#include <linux/suspend.h>
52#include <linux/ftrace.h>
53#include <linux/tick.h>
54#include <linux/sysrq.h>
55#include <linux/kprobes.h>
56#include <linux/gfp.h>
57#include <linux/oom.h>
58#include <linux/smpboot.h>
59#include <linux/jiffies.h>
60#include <linux/slab.h>
61#include <linux/sched/isolation.h>
62#include <linux/sched/clock.h>
63#include <linux/vmalloc.h>
64#include <linux/mm.h>
65#include <linux/kasan.h>
66#include <linux/context_tracking.h>
67#include "../time/tick-internal.h"
68
69#include "tree.h"
70#include "rcu.h"
71
72#ifdef MODULE_PARAM_PREFIX
73#undef MODULE_PARAM_PREFIX
74#endif
75#define MODULE_PARAM_PREFIX "rcutree."
76
77/* Data structures. */
78
79static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = {
80 .gpwrap = true,
81#ifdef CONFIG_RCU_NOCB_CPU
82 .cblist.flags = SEGCBLIST_RCU_CORE,
83#endif
84};
85static struct rcu_state rcu_state = {
86 .level = { &rcu_state.node[0] },
87 .gp_state = RCU_GP_IDLE,
88 .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT,
89 .barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex),
90 .barrier_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.barrier_lock),
91 .name = RCU_NAME,
92 .abbr = RCU_ABBR,
93 .exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
94 .exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
95 .ofl_lock = __ARCH_SPIN_LOCK_UNLOCKED,
96};
97
98/* Dump rcu_node combining tree at boot to verify correct setup. */
99static bool dump_tree;
100module_param(dump_tree, bool, 0444);
101/* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */
102static bool use_softirq = !IS_ENABLED(CONFIG_PREEMPT_RT);
103#ifndef CONFIG_PREEMPT_RT
104module_param(use_softirq, bool, 0444);
105#endif
106/* Control rcu_node-tree auto-balancing at boot time. */
107static bool rcu_fanout_exact;
108module_param(rcu_fanout_exact, bool, 0444);
109/* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
110static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
111module_param(rcu_fanout_leaf, int, 0444);
112int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
113/* Number of rcu_nodes at specified level. */
114int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
115int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
116
117/*
118 * The rcu_scheduler_active variable is initialized to the value
119 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
120 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
121 * RCU can assume that there is but one task, allowing RCU to (for example)
122 * optimize synchronize_rcu() to a simple barrier(). When this variable
123 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
124 * to detect real grace periods. This variable is also used to suppress
125 * boot-time false positives from lockdep-RCU error checking. Finally, it
126 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
127 * is fully initialized, including all of its kthreads having been spawned.
128 */
129int rcu_scheduler_active __read_mostly;
130EXPORT_SYMBOL_GPL(rcu_scheduler_active);
131
132/*
133 * The rcu_scheduler_fully_active variable transitions from zero to one
134 * during the early_initcall() processing, which is after the scheduler
135 * is capable of creating new tasks. So RCU processing (for example,
136 * creating tasks for RCU priority boosting) must be delayed until after
137 * rcu_scheduler_fully_active transitions from zero to one. We also
138 * currently delay invocation of any RCU callbacks until after this point.
139 *
140 * It might later prove better for people registering RCU callbacks during
141 * early boot to take responsibility for these callbacks, but one step at
142 * a time.
143 */
144static int rcu_scheduler_fully_active __read_mostly;
145
146static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
147 unsigned long gps, unsigned long flags);
148static struct task_struct *rcu_boost_task(struct rcu_node *rnp);
149static void invoke_rcu_core(void);
150static void rcu_report_exp_rdp(struct rcu_data *rdp);
151static void sync_sched_exp_online_cleanup(int cpu);
152static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp);
153static bool rcu_rdp_is_offloaded(struct rcu_data *rdp);
154static bool rcu_rdp_cpu_online(struct rcu_data *rdp);
155static bool rcu_init_invoked(void);
156static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
157static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
158
159/*
160 * rcuc/rcub/rcuop kthread realtime priority. The "rcuop"
161 * real-time priority(enabling/disabling) is controlled by
162 * the extra CONFIG_RCU_NOCB_CPU_CB_BOOST configuration.
163 */
164static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
165module_param(kthread_prio, int, 0444);
166
167/* Delay in jiffies for grace-period initialization delays, debug only. */
168
169static int gp_preinit_delay;
170module_param(gp_preinit_delay, int, 0444);
171static int gp_init_delay;
172module_param(gp_init_delay, int, 0444);
173static int gp_cleanup_delay;
174module_param(gp_cleanup_delay, int, 0444);
175
176// Add delay to rcu_read_unlock() for strict grace periods.
177static int rcu_unlock_delay;
178#ifdef CONFIG_RCU_STRICT_GRACE_PERIOD
179module_param(rcu_unlock_delay, int, 0444);
180#endif
181
182/*
183 * This rcu parameter is runtime-read-only. It reflects
184 * a minimum allowed number of objects which can be cached
185 * per-CPU. Object size is equal to one page. This value
186 * can be changed at boot time.
187 */
188static int rcu_min_cached_objs = 5;
189module_param(rcu_min_cached_objs, int, 0444);
190
191// A page shrinker can ask for pages to be freed to make them
192// available for other parts of the system. This usually happens
193// under low memory conditions, and in that case we should also
194// defer page-cache filling for a short time period.
195//
196// The default value is 5 seconds, which is long enough to reduce
197// interference with the shrinker while it asks other systems to
198// drain their caches.
199static int rcu_delay_page_cache_fill_msec = 5000;
200module_param(rcu_delay_page_cache_fill_msec, int, 0444);
201
202/* Retrieve RCU kthreads priority for rcutorture */
203int rcu_get_gp_kthreads_prio(void)
204{
205 return kthread_prio;
206}
207EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio);
208
209/*
210 * Number of grace periods between delays, normalized by the duration of
211 * the delay. The longer the delay, the more the grace periods between
212 * each delay. The reason for this normalization is that it means that,
213 * for non-zero delays, the overall slowdown of grace periods is constant
214 * regardless of the duration of the delay. This arrangement balances
215 * the need for long delays to increase some race probabilities with the
216 * need for fast grace periods to increase other race probabilities.
217 */
218#define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays for debugging. */
219
220/*
221 * Return true if an RCU grace period is in progress. The READ_ONCE()s
222 * permit this function to be invoked without holding the root rcu_node
223 * structure's ->lock, but of course results can be subject to change.
224 */
225static int rcu_gp_in_progress(void)
226{
227 return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq));
228}
229
230/*
231 * Return the number of callbacks queued on the specified CPU.
232 * Handles both the nocbs and normal cases.
233 */
234static long rcu_get_n_cbs_cpu(int cpu)
235{
236 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
237
238 if (rcu_segcblist_is_enabled(&rdp->cblist))
239 return rcu_segcblist_n_cbs(&rdp->cblist);
240 return 0;
241}
242
243void rcu_softirq_qs(void)
244{
245 rcu_qs();
246 rcu_preempt_deferred_qs(current);
247 rcu_tasks_qs(current, false);
248}
249
250/*
251 * Reset the current CPU's ->dynticks counter to indicate that the
252 * newly onlined CPU is no longer in an extended quiescent state.
253 * This will either leave the counter unchanged, or increment it
254 * to the next non-quiescent value.
255 *
256 * The non-atomic test/increment sequence works because the upper bits
257 * of the ->dynticks counter are manipulated only by the corresponding CPU,
258 * or when the corresponding CPU is offline.
259 */
260static void rcu_dynticks_eqs_online(void)
261{
262 if (ct_dynticks() & RCU_DYNTICKS_IDX)
263 return;
264 ct_state_inc(RCU_DYNTICKS_IDX);
265}
266
267/*
268 * Snapshot the ->dynticks counter with full ordering so as to allow
269 * stable comparison of this counter with past and future snapshots.
270 */
271static int rcu_dynticks_snap(int cpu)
272{
273 smp_mb(); // Fundamental RCU ordering guarantee.
274 return ct_dynticks_cpu_acquire(cpu);
275}
276
277/*
278 * Return true if the snapshot returned from rcu_dynticks_snap()
279 * indicates that RCU is in an extended quiescent state.
280 */
281static bool rcu_dynticks_in_eqs(int snap)
282{
283 return !(snap & RCU_DYNTICKS_IDX);
284}
285
286/*
287 * Return true if the CPU corresponding to the specified rcu_data
288 * structure has spent some time in an extended quiescent state since
289 * rcu_dynticks_snap() returned the specified snapshot.
290 */
291static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap)
292{
293 return snap != rcu_dynticks_snap(rdp->cpu);
294}
295
296/*
297 * Return true if the referenced integer is zero while the specified
298 * CPU remains within a single extended quiescent state.
299 */
300bool rcu_dynticks_zero_in_eqs(int cpu, int *vp)
301{
302 int snap;
303
304 // If not quiescent, force back to earlier extended quiescent state.
305 snap = ct_dynticks_cpu(cpu) & ~RCU_DYNTICKS_IDX;
306 smp_rmb(); // Order ->dynticks and *vp reads.
307 if (READ_ONCE(*vp))
308 return false; // Non-zero, so report failure;
309 smp_rmb(); // Order *vp read and ->dynticks re-read.
310
311 // If still in the same extended quiescent state, we are good!
312 return snap == ct_dynticks_cpu(cpu);
313}
314
315/*
316 * Let the RCU core know that this CPU has gone through the scheduler,
317 * which is a quiescent state. This is called when the need for a
318 * quiescent state is urgent, so we burn an atomic operation and full
319 * memory barriers to let the RCU core know about it, regardless of what
320 * this CPU might (or might not) do in the near future.
321 *
322 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
323 *
324 * The caller must have disabled interrupts and must not be idle.
325 */
326notrace void rcu_momentary_dyntick_idle(void)
327{
328 int seq;
329
330 raw_cpu_write(rcu_data.rcu_need_heavy_qs, false);
331 seq = ct_state_inc(2 * RCU_DYNTICKS_IDX);
332 /* It is illegal to call this from idle state. */
333 WARN_ON_ONCE(!(seq & RCU_DYNTICKS_IDX));
334 rcu_preempt_deferred_qs(current);
335}
336EXPORT_SYMBOL_GPL(rcu_momentary_dyntick_idle);
337
338/**
339 * rcu_is_cpu_rrupt_from_idle - see if 'interrupted' from idle
340 *
341 * If the current CPU is idle and running at a first-level (not nested)
342 * interrupt, or directly, from idle, return true.
343 *
344 * The caller must have at least disabled IRQs.
345 */
346static int rcu_is_cpu_rrupt_from_idle(void)
347{
348 long nesting;
349
350 /*
351 * Usually called from the tick; but also used from smp_function_call()
352 * for expedited grace periods. This latter can result in running from
353 * the idle task, instead of an actual IPI.
354 */
355 lockdep_assert_irqs_disabled();
356
357 /* Check for counter underflows */
358 RCU_LOCKDEP_WARN(ct_dynticks_nesting() < 0,
359 "RCU dynticks_nesting counter underflow!");
360 RCU_LOCKDEP_WARN(ct_dynticks_nmi_nesting() <= 0,
361 "RCU dynticks_nmi_nesting counter underflow/zero!");
362
363 /* Are we at first interrupt nesting level? */
364 nesting = ct_dynticks_nmi_nesting();
365 if (nesting > 1)
366 return false;
367
368 /*
369 * If we're not in an interrupt, we must be in the idle task!
370 */
371 WARN_ON_ONCE(!nesting && !is_idle_task(current));
372
373 /* Does CPU appear to be idle from an RCU standpoint? */
374 return ct_dynticks_nesting() == 0;
375}
376
377#define DEFAULT_RCU_BLIMIT (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 1000 : 10)
378 // Maximum callbacks per rcu_do_batch ...
379#define DEFAULT_MAX_RCU_BLIMIT 10000 // ... even during callback flood.
380static long blimit = DEFAULT_RCU_BLIMIT;
381#define DEFAULT_RCU_QHIMARK 10000 // If this many pending, ignore blimit.
382static long qhimark = DEFAULT_RCU_QHIMARK;
383#define DEFAULT_RCU_QLOMARK 100 // Once only this many pending, use blimit.
384static long qlowmark = DEFAULT_RCU_QLOMARK;
385#define DEFAULT_RCU_QOVLD_MULT 2
386#define DEFAULT_RCU_QOVLD (DEFAULT_RCU_QOVLD_MULT * DEFAULT_RCU_QHIMARK)
387static long qovld = DEFAULT_RCU_QOVLD; // If this many pending, hammer QS.
388static long qovld_calc = -1; // No pre-initialization lock acquisitions!
389
390module_param(blimit, long, 0444);
391module_param(qhimark, long, 0444);
392module_param(qlowmark, long, 0444);
393module_param(qovld, long, 0444);
394
395static ulong jiffies_till_first_fqs = IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 0 : ULONG_MAX;
396static ulong jiffies_till_next_fqs = ULONG_MAX;
397static bool rcu_kick_kthreads;
398static int rcu_divisor = 7;
399module_param(rcu_divisor, int, 0644);
400
401/* Force an exit from rcu_do_batch() after 3 milliseconds. */
402static long rcu_resched_ns = 3 * NSEC_PER_MSEC;
403module_param(rcu_resched_ns, long, 0644);
404
405/*
406 * How long the grace period must be before we start recruiting
407 * quiescent-state help from rcu_note_context_switch().
408 */
409static ulong jiffies_till_sched_qs = ULONG_MAX;
410module_param(jiffies_till_sched_qs, ulong, 0444);
411static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */
412module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
413
414/*
415 * Make sure that we give the grace-period kthread time to detect any
416 * idle CPUs before taking active measures to force quiescent states.
417 * However, don't go below 100 milliseconds, adjusted upwards for really
418 * large systems.
419 */
420static void adjust_jiffies_till_sched_qs(void)
421{
422 unsigned long j;
423
424 /* If jiffies_till_sched_qs was specified, respect the request. */
425 if (jiffies_till_sched_qs != ULONG_MAX) {
426 WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
427 return;
428 }
429 /* Otherwise, set to third fqs scan, but bound below on large system. */
430 j = READ_ONCE(jiffies_till_first_fqs) +
431 2 * READ_ONCE(jiffies_till_next_fqs);
432 if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
433 j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
434 pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
435 WRITE_ONCE(jiffies_to_sched_qs, j);
436}
437
438static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
439{
440 ulong j;
441 int ret = kstrtoul(val, 0, &j);
442
443 if (!ret) {
444 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
445 adjust_jiffies_till_sched_qs();
446 }
447 return ret;
448}
449
450static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
451{
452 ulong j;
453 int ret = kstrtoul(val, 0, &j);
454
455 if (!ret) {
456 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
457 adjust_jiffies_till_sched_qs();
458 }
459 return ret;
460}
461
462static const struct kernel_param_ops first_fqs_jiffies_ops = {
463 .set = param_set_first_fqs_jiffies,
464 .get = param_get_ulong,
465};
466
467static const struct kernel_param_ops next_fqs_jiffies_ops = {
468 .set = param_set_next_fqs_jiffies,
469 .get = param_get_ulong,
470};
471
472module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
473module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
474module_param(rcu_kick_kthreads, bool, 0644);
475
476static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
477static int rcu_pending(int user);
478
479/*
480 * Return the number of RCU GPs completed thus far for debug & stats.
481 */
482unsigned long rcu_get_gp_seq(void)
483{
484 return READ_ONCE(rcu_state.gp_seq);
485}
486EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
487
488/*
489 * Return the number of RCU expedited batches completed thus far for
490 * debug & stats. Odd numbers mean that a batch is in progress, even
491 * numbers mean idle. The value returned will thus be roughly double
492 * the cumulative batches since boot.
493 */
494unsigned long rcu_exp_batches_completed(void)
495{
496 return rcu_state.expedited_sequence;
497}
498EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
499
500/*
501 * Return the root node of the rcu_state structure.
502 */
503static struct rcu_node *rcu_get_root(void)
504{
505 return &rcu_state.node[0];
506}
507
508/*
509 * Send along grace-period-related data for rcutorture diagnostics.
510 */
511void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
512 unsigned long *gp_seq)
513{
514 switch (test_type) {
515 case RCU_FLAVOR:
516 *flags = READ_ONCE(rcu_state.gp_flags);
517 *gp_seq = rcu_seq_current(&rcu_state.gp_seq);
518 break;
519 default:
520 break;
521 }
522}
523EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
524
525#if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK))
526/*
527 * An empty function that will trigger a reschedule on
528 * IRQ tail once IRQs get re-enabled on userspace/guest resume.
529 */
530static void late_wakeup_func(struct irq_work *work)
531{
532}
533
534static DEFINE_PER_CPU(struct irq_work, late_wakeup_work) =
535 IRQ_WORK_INIT(late_wakeup_func);
536
537/*
538 * If either:
539 *
540 * 1) the task is about to enter in guest mode and $ARCH doesn't support KVM generic work
541 * 2) the task is about to enter in user mode and $ARCH doesn't support generic entry.
542 *
543 * In these cases the late RCU wake ups aren't supported in the resched loops and our
544 * last resort is to fire a local irq_work that will trigger a reschedule once IRQs
545 * get re-enabled again.
546 */
547noinstr void rcu_irq_work_resched(void)
548{
549 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
550
551 if (IS_ENABLED(CONFIG_GENERIC_ENTRY) && !(current->flags & PF_VCPU))
552 return;
553
554 if (IS_ENABLED(CONFIG_KVM_XFER_TO_GUEST_WORK) && (current->flags & PF_VCPU))
555 return;
556
557 instrumentation_begin();
558 if (do_nocb_deferred_wakeup(rdp) && need_resched()) {
559 irq_work_queue(this_cpu_ptr(&late_wakeup_work));
560 }
561 instrumentation_end();
562}
563#endif /* #if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK)) */
564
565#ifdef CONFIG_PROVE_RCU
566/**
567 * rcu_irq_exit_check_preempt - Validate that scheduling is possible
568 */
569void rcu_irq_exit_check_preempt(void)
570{
571 lockdep_assert_irqs_disabled();
572
573 RCU_LOCKDEP_WARN(ct_dynticks_nesting() <= 0,
574 "RCU dynticks_nesting counter underflow/zero!");
575 RCU_LOCKDEP_WARN(ct_dynticks_nmi_nesting() !=
576 DYNTICK_IRQ_NONIDLE,
577 "Bad RCU dynticks_nmi_nesting counter\n");
578 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
579 "RCU in extended quiescent state!");
580}
581#endif /* #ifdef CONFIG_PROVE_RCU */
582
583#ifdef CONFIG_NO_HZ_FULL
584/**
585 * __rcu_irq_enter_check_tick - Enable scheduler tick on CPU if RCU needs it.
586 *
587 * The scheduler tick is not normally enabled when CPUs enter the kernel
588 * from nohz_full userspace execution. After all, nohz_full userspace
589 * execution is an RCU quiescent state and the time executing in the kernel
590 * is quite short. Except of course when it isn't. And it is not hard to
591 * cause a large system to spend tens of seconds or even minutes looping
592 * in the kernel, which can cause a number of problems, include RCU CPU
593 * stall warnings.
594 *
595 * Therefore, if a nohz_full CPU fails to report a quiescent state
596 * in a timely manner, the RCU grace-period kthread sets that CPU's
597 * ->rcu_urgent_qs flag with the expectation that the next interrupt or
598 * exception will invoke this function, which will turn on the scheduler
599 * tick, which will enable RCU to detect that CPU's quiescent states,
600 * for example, due to cond_resched() calls in CONFIG_PREEMPT=n kernels.
601 * The tick will be disabled once a quiescent state is reported for
602 * this CPU.
603 *
604 * Of course, in carefully tuned systems, there might never be an
605 * interrupt or exception. In that case, the RCU grace-period kthread
606 * will eventually cause one to happen. However, in less carefully
607 * controlled environments, this function allows RCU to get what it
608 * needs without creating otherwise useless interruptions.
609 */
610void __rcu_irq_enter_check_tick(void)
611{
612 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
613
614 // If we're here from NMI there's nothing to do.
615 if (in_nmi())
616 return;
617
618 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
619 "Illegal rcu_irq_enter_check_tick() from extended quiescent state");
620
621 if (!tick_nohz_full_cpu(rdp->cpu) ||
622 !READ_ONCE(rdp->rcu_urgent_qs) ||
623 READ_ONCE(rdp->rcu_forced_tick)) {
624 // RCU doesn't need nohz_full help from this CPU, or it is
625 // already getting that help.
626 return;
627 }
628
629 // We get here only when not in an extended quiescent state and
630 // from interrupts (as opposed to NMIs). Therefore, (1) RCU is
631 // already watching and (2) The fact that we are in an interrupt
632 // handler and that the rcu_node lock is an irq-disabled lock
633 // prevents self-deadlock. So we can safely recheck under the lock.
634 // Note that the nohz_full state currently cannot change.
635 raw_spin_lock_rcu_node(rdp->mynode);
636 if (READ_ONCE(rdp->rcu_urgent_qs) && !rdp->rcu_forced_tick) {
637 // A nohz_full CPU is in the kernel and RCU needs a
638 // quiescent state. Turn on the tick!
639 WRITE_ONCE(rdp->rcu_forced_tick, true);
640 tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
641 }
642 raw_spin_unlock_rcu_node(rdp->mynode);
643}
644NOKPROBE_SYMBOL(__rcu_irq_enter_check_tick);
645#endif /* CONFIG_NO_HZ_FULL */
646
647/*
648 * Check to see if any future non-offloaded RCU-related work will need
649 * to be done by the current CPU, even if none need be done immediately,
650 * returning 1 if so. This function is part of the RCU implementation;
651 * it is -not- an exported member of the RCU API. This is used by
652 * the idle-entry code to figure out whether it is safe to disable the
653 * scheduler-clock interrupt.
654 *
655 * Just check whether or not this CPU has non-offloaded RCU callbacks
656 * queued.
657 */
658int rcu_needs_cpu(void)
659{
660 return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) &&
661 !rcu_rdp_is_offloaded(this_cpu_ptr(&rcu_data));
662}
663
664/*
665 * If any sort of urgency was applied to the current CPU (for example,
666 * the scheduler-clock interrupt was enabled on a nohz_full CPU) in order
667 * to get to a quiescent state, disable it.
668 */
669static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp)
670{
671 raw_lockdep_assert_held_rcu_node(rdp->mynode);
672 WRITE_ONCE(rdp->rcu_urgent_qs, false);
673 WRITE_ONCE(rdp->rcu_need_heavy_qs, false);
674 if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) {
675 tick_dep_clear_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
676 WRITE_ONCE(rdp->rcu_forced_tick, false);
677 }
678}
679
680/**
681 * rcu_is_watching - RCU read-side critical sections permitted on current CPU?
682 *
683 * Return @true if RCU is watching the running CPU and @false otherwise.
684 * An @true return means that this CPU can safely enter RCU read-side
685 * critical sections.
686 *
687 * Although calls to rcu_is_watching() from most parts of the kernel
688 * will return @true, there are important exceptions. For example, if the
689 * current CPU is deep within its idle loop, in kernel entry/exit code,
690 * or offline, rcu_is_watching() will return @false.
691 *
692 * Make notrace because it can be called by the internal functions of
693 * ftrace, and making this notrace removes unnecessary recursion calls.
694 */
695notrace bool rcu_is_watching(void)
696{
697 bool ret;
698
699 preempt_disable_notrace();
700 ret = !rcu_dynticks_curr_cpu_in_eqs();
701 preempt_enable_notrace();
702 return ret;
703}
704EXPORT_SYMBOL_GPL(rcu_is_watching);
705
706/*
707 * If a holdout task is actually running, request an urgent quiescent
708 * state from its CPU. This is unsynchronized, so migrations can cause
709 * the request to go to the wrong CPU. Which is OK, all that will happen
710 * is that the CPU's next context switch will be a bit slower and next
711 * time around this task will generate another request.
712 */
713void rcu_request_urgent_qs_task(struct task_struct *t)
714{
715 int cpu;
716
717 barrier();
718 cpu = task_cpu(t);
719 if (!task_curr(t))
720 return; /* This task is not running on that CPU. */
721 smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
722}
723
724/*
725 * When trying to report a quiescent state on behalf of some other CPU,
726 * it is our responsibility to check for and handle potential overflow
727 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
728 * After all, the CPU might be in deep idle state, and thus executing no
729 * code whatsoever.
730 */
731static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
732{
733 raw_lockdep_assert_held_rcu_node(rnp);
734 if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
735 rnp->gp_seq))
736 WRITE_ONCE(rdp->gpwrap, true);
737 if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
738 rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
739}
740
741/*
742 * Snapshot the specified CPU's dynticks counter so that we can later
743 * credit them with an implicit quiescent state. Return 1 if this CPU
744 * is in dynticks idle mode, which is an extended quiescent state.
745 */
746static int dyntick_save_progress_counter(struct rcu_data *rdp)
747{
748 rdp->dynticks_snap = rcu_dynticks_snap(rdp->cpu);
749 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
750 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
751 rcu_gpnum_ovf(rdp->mynode, rdp);
752 return 1;
753 }
754 return 0;
755}
756
757/*
758 * Returns positive if the specified CPU has passed through a quiescent state
759 * by virtue of being in or having passed through an dynticks idle state since
760 * the last call to dyntick_save_progress_counter() for this same CPU, or by
761 * virtue of having been offline.
762 *
763 * Returns negative if the specified CPU needs a force resched.
764 *
765 * Returns zero otherwise.
766 */
767static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
768{
769 unsigned long jtsq;
770 int ret = 0;
771 struct rcu_node *rnp = rdp->mynode;
772
773 /*
774 * If the CPU passed through or entered a dynticks idle phase with
775 * no active irq/NMI handlers, then we can safely pretend that the CPU
776 * already acknowledged the request to pass through a quiescent
777 * state. Either way, that CPU cannot possibly be in an RCU
778 * read-side critical section that started before the beginning
779 * of the current RCU grace period.
780 */
781 if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
782 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
783 rcu_gpnum_ovf(rnp, rdp);
784 return 1;
785 }
786
787 /*
788 * Complain if a CPU that is considered to be offline from RCU's
789 * perspective has not yet reported a quiescent state. After all,
790 * the offline CPU should have reported a quiescent state during
791 * the CPU-offline process, or, failing that, by rcu_gp_init()
792 * if it ran concurrently with either the CPU going offline or the
793 * last task on a leaf rcu_node structure exiting its RCU read-side
794 * critical section while all CPUs corresponding to that structure
795 * are offline. This added warning detects bugs in any of these
796 * code paths.
797 *
798 * The rcu_node structure's ->lock is held here, which excludes
799 * the relevant portions the CPU-hotplug code, the grace-period
800 * initialization code, and the rcu_read_unlock() code paths.
801 *
802 * For more detail, please refer to the "Hotplug CPU" section
803 * of RCU's Requirements documentation.
804 */
805 if (WARN_ON_ONCE(!rcu_rdp_cpu_online(rdp))) {
806 struct rcu_node *rnp1;
807
808 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
809 __func__, rnp->grplo, rnp->grphi, rnp->level,
810 (long)rnp->gp_seq, (long)rnp->completedqs);
811 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
812 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
813 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
814 pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
815 __func__, rdp->cpu, ".o"[rcu_rdp_cpu_online(rdp)],
816 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
817 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
818 return 1; /* Break things loose after complaining. */
819 }
820
821 /*
822 * A CPU running for an extended time within the kernel can
823 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
824 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
825 * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the
826 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
827 * variable are safe because the assignments are repeated if this
828 * CPU failed to pass through a quiescent state. This code
829 * also checks .jiffies_resched in case jiffies_to_sched_qs
830 * is set way high.
831 */
832 jtsq = READ_ONCE(jiffies_to_sched_qs);
833 if (!READ_ONCE(rdp->rcu_need_heavy_qs) &&
834 (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
835 time_after(jiffies, rcu_state.jiffies_resched) ||
836 rcu_state.cbovld)) {
837 WRITE_ONCE(rdp->rcu_need_heavy_qs, true);
838 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
839 smp_store_release(&rdp->rcu_urgent_qs, true);
840 } else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
841 WRITE_ONCE(rdp->rcu_urgent_qs, true);
842 }
843
844 /*
845 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
846 * The above code handles this, but only for straight cond_resched().
847 * And some in-kernel loops check need_resched() before calling
848 * cond_resched(), which defeats the above code for CPUs that are
849 * running in-kernel with scheduling-clock interrupts disabled.
850 * So hit them over the head with the resched_cpu() hammer!
851 */
852 if (tick_nohz_full_cpu(rdp->cpu) &&
853 (time_after(jiffies, READ_ONCE(rdp->last_fqs_resched) + jtsq * 3) ||
854 rcu_state.cbovld)) {
855 WRITE_ONCE(rdp->rcu_urgent_qs, true);
856 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
857 ret = -1;
858 }
859
860 /*
861 * If more than halfway to RCU CPU stall-warning time, invoke
862 * resched_cpu() more frequently to try to loosen things up a bit.
863 * Also check to see if the CPU is getting hammered with interrupts,
864 * but only once per grace period, just to keep the IPIs down to
865 * a dull roar.
866 */
867 if (time_after(jiffies, rcu_state.jiffies_resched)) {
868 if (time_after(jiffies,
869 READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
870 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
871 ret = -1;
872 }
873 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
874 !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
875 (rnp->ffmask & rdp->grpmask)) {
876 rdp->rcu_iw_pending = true;
877 rdp->rcu_iw_gp_seq = rnp->gp_seq;
878 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
879 }
880
881 if (rcu_cpu_stall_cputime && rdp->snap_record.gp_seq != rdp->gp_seq) {
882 int cpu = rdp->cpu;
883 struct rcu_snap_record *rsrp;
884 struct kernel_cpustat *kcsp;
885
886 kcsp = &kcpustat_cpu(cpu);
887
888 rsrp = &rdp->snap_record;
889 rsrp->cputime_irq = kcpustat_field(kcsp, CPUTIME_IRQ, cpu);
890 rsrp->cputime_softirq = kcpustat_field(kcsp, CPUTIME_SOFTIRQ, cpu);
891 rsrp->cputime_system = kcpustat_field(kcsp, CPUTIME_SYSTEM, cpu);
892 rsrp->nr_hardirqs = kstat_cpu_irqs_sum(rdp->cpu);
893 rsrp->nr_softirqs = kstat_cpu_softirqs_sum(rdp->cpu);
894 rsrp->nr_csw = nr_context_switches_cpu(rdp->cpu);
895 rsrp->jiffies = jiffies;
896 rsrp->gp_seq = rdp->gp_seq;
897 }
898 }
899
900 return ret;
901}
902
903/* Trace-event wrapper function for trace_rcu_future_grace_period. */
904static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
905 unsigned long gp_seq_req, const char *s)
906{
907 trace_rcu_future_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
908 gp_seq_req, rnp->level,
909 rnp->grplo, rnp->grphi, s);
910}
911
912/*
913 * rcu_start_this_gp - Request the start of a particular grace period
914 * @rnp_start: The leaf node of the CPU from which to start.
915 * @rdp: The rcu_data corresponding to the CPU from which to start.
916 * @gp_seq_req: The gp_seq of the grace period to start.
917 *
918 * Start the specified grace period, as needed to handle newly arrived
919 * callbacks. The required future grace periods are recorded in each
920 * rcu_node structure's ->gp_seq_needed field. Returns true if there
921 * is reason to awaken the grace-period kthread.
922 *
923 * The caller must hold the specified rcu_node structure's ->lock, which
924 * is why the caller is responsible for waking the grace-period kthread.
925 *
926 * Returns true if the GP thread needs to be awakened else false.
927 */
928static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
929 unsigned long gp_seq_req)
930{
931 bool ret = false;
932 struct rcu_node *rnp;
933
934 /*
935 * Use funnel locking to either acquire the root rcu_node
936 * structure's lock or bail out if the need for this grace period
937 * has already been recorded -- or if that grace period has in
938 * fact already started. If there is already a grace period in
939 * progress in a non-leaf node, no recording is needed because the
940 * end of the grace period will scan the leaf rcu_node structures.
941 * Note that rnp_start->lock must not be released.
942 */
943 raw_lockdep_assert_held_rcu_node(rnp_start);
944 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
945 for (rnp = rnp_start; 1; rnp = rnp->parent) {
946 if (rnp != rnp_start)
947 raw_spin_lock_rcu_node(rnp);
948 if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
949 rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
950 (rnp != rnp_start &&
951 rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
952 trace_rcu_this_gp(rnp, rdp, gp_seq_req,
953 TPS("Prestarted"));
954 goto unlock_out;
955 }
956 WRITE_ONCE(rnp->gp_seq_needed, gp_seq_req);
957 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
958 /*
959 * We just marked the leaf or internal node, and a
960 * grace period is in progress, which means that
961 * rcu_gp_cleanup() will see the marking. Bail to
962 * reduce contention.
963 */
964 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
965 TPS("Startedleaf"));
966 goto unlock_out;
967 }
968 if (rnp != rnp_start && rnp->parent != NULL)
969 raw_spin_unlock_rcu_node(rnp);
970 if (!rnp->parent)
971 break; /* At root, and perhaps also leaf. */
972 }
973
974 /* If GP already in progress, just leave, otherwise start one. */
975 if (rcu_gp_in_progress()) {
976 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
977 goto unlock_out;
978 }
979 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
980 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
981 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
982 if (!READ_ONCE(rcu_state.gp_kthread)) {
983 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
984 goto unlock_out;
985 }
986 trace_rcu_grace_period(rcu_state.name, data_race(rcu_state.gp_seq), TPS("newreq"));
987 ret = true; /* Caller must wake GP kthread. */
988unlock_out:
989 /* Push furthest requested GP to leaf node and rcu_data structure. */
990 if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
991 WRITE_ONCE(rnp_start->gp_seq_needed, rnp->gp_seq_needed);
992 WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
993 }
994 if (rnp != rnp_start)
995 raw_spin_unlock_rcu_node(rnp);
996 return ret;
997}
998
999/*
1000 * Clean up any old requests for the just-ended grace period. Also return
1001 * whether any additional grace periods have been requested.
1002 */
1003static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
1004{
1005 bool needmore;
1006 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1007
1008 needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1009 if (!needmore)
1010 rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1011 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1012 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1013 return needmore;
1014}
1015
1016static void swake_up_one_online_ipi(void *arg)
1017{
1018 struct swait_queue_head *wqh = arg;
1019
1020 swake_up_one(wqh);
1021}
1022
1023static void swake_up_one_online(struct swait_queue_head *wqh)
1024{
1025 int cpu = get_cpu();
1026
1027 /*
1028 * If called from rcutree_report_cpu_starting(), wake up
1029 * is dangerous that late in the CPU-down hotplug process. The
1030 * scheduler might queue an ignored hrtimer. Defer the wake up
1031 * to an online CPU instead.
1032 */
1033 if (unlikely(cpu_is_offline(cpu))) {
1034 int target;
1035
1036 target = cpumask_any_and(housekeeping_cpumask(HK_TYPE_RCU),
1037 cpu_online_mask);
1038
1039 smp_call_function_single(target, swake_up_one_online_ipi,
1040 wqh, 0);
1041 put_cpu();
1042 } else {
1043 put_cpu();
1044 swake_up_one(wqh);
1045 }
1046}
1047
1048/*
1049 * Awaken the grace-period kthread. Don't do a self-awaken (unless in an
1050 * interrupt or softirq handler, in which case we just might immediately
1051 * sleep upon return, resulting in a grace-period hang), and don't bother
1052 * awakening when there is nothing for the grace-period kthread to do
1053 * (as in several CPUs raced to awaken, we lost), and finally don't try
1054 * to awaken a kthread that has not yet been created. If all those checks
1055 * are passed, track some debug information and awaken.
1056 *
1057 * So why do the self-wakeup when in an interrupt or softirq handler
1058 * in the grace-period kthread's context? Because the kthread might have
1059 * been interrupted just as it was going to sleep, and just after the final
1060 * pre-sleep check of the awaken condition. In this case, a wakeup really
1061 * is required, and is therefore supplied.
1062 */
1063static void rcu_gp_kthread_wake(void)
1064{
1065 struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
1066
1067 if ((current == t && !in_hardirq() && !in_serving_softirq()) ||
1068 !READ_ONCE(rcu_state.gp_flags) || !t)
1069 return;
1070 WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
1071 WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
1072 swake_up_one_online(&rcu_state.gp_wq);
1073}
1074
1075/*
1076 * If there is room, assign a ->gp_seq number to any callbacks on this
1077 * CPU that have not already been assigned. Also accelerate any callbacks
1078 * that were previously assigned a ->gp_seq number that has since proven
1079 * to be too conservative, which can happen if callbacks get assigned a
1080 * ->gp_seq number while RCU is idle, but with reference to a non-root
1081 * rcu_node structure. This function is idempotent, so it does not hurt
1082 * to call it repeatedly. Returns an flag saying that we should awaken
1083 * the RCU grace-period kthread.
1084 *
1085 * The caller must hold rnp->lock with interrupts disabled.
1086 */
1087static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1088{
1089 unsigned long gp_seq_req;
1090 bool ret = false;
1091
1092 rcu_lockdep_assert_cblist_protected(rdp);
1093 raw_lockdep_assert_held_rcu_node(rnp);
1094
1095 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1096 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1097 return false;
1098
1099 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPreAcc"));
1100
1101 /*
1102 * Callbacks are often registered with incomplete grace-period
1103 * information. Something about the fact that getting exact
1104 * information requires acquiring a global lock... RCU therefore
1105 * makes a conservative estimate of the grace period number at which
1106 * a given callback will become ready to invoke. The following
1107 * code checks this estimate and improves it when possible, thus
1108 * accelerating callback invocation to an earlier grace-period
1109 * number.
1110 */
1111 gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1112 if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1113 ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1114
1115 /* Trace depending on how much we were able to accelerate. */
1116 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1117 trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccWaitCB"));
1118 else
1119 trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccReadyCB"));
1120
1121 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPostAcc"));
1122
1123 return ret;
1124}
1125
1126/*
1127 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1128 * rcu_node structure's ->lock be held. It consults the cached value
1129 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1130 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1131 * while holding the leaf rcu_node structure's ->lock.
1132 */
1133static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1134 struct rcu_data *rdp)
1135{
1136 unsigned long c;
1137 bool needwake;
1138
1139 rcu_lockdep_assert_cblist_protected(rdp);
1140 c = rcu_seq_snap(&rcu_state.gp_seq);
1141 if (!READ_ONCE(rdp->gpwrap) && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1142 /* Old request still live, so mark recent callbacks. */
1143 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1144 return;
1145 }
1146 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1147 needwake = rcu_accelerate_cbs(rnp, rdp);
1148 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1149 if (needwake)
1150 rcu_gp_kthread_wake();
1151}
1152
1153/*
1154 * Move any callbacks whose grace period has completed to the
1155 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1156 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1157 * sublist. This function is idempotent, so it does not hurt to
1158 * invoke it repeatedly. As long as it is not invoked -too- often...
1159 * Returns true if the RCU grace-period kthread needs to be awakened.
1160 *
1161 * The caller must hold rnp->lock with interrupts disabled.
1162 */
1163static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1164{
1165 rcu_lockdep_assert_cblist_protected(rdp);
1166 raw_lockdep_assert_held_rcu_node(rnp);
1167
1168 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1169 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1170 return false;
1171
1172 /*
1173 * Find all callbacks whose ->gp_seq numbers indicate that they
1174 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1175 */
1176 rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1177
1178 /* Classify any remaining callbacks. */
1179 return rcu_accelerate_cbs(rnp, rdp);
1180}
1181
1182/*
1183 * Move and classify callbacks, but only if doing so won't require
1184 * that the RCU grace-period kthread be awakened.
1185 */
1186static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
1187 struct rcu_data *rdp)
1188{
1189 rcu_lockdep_assert_cblist_protected(rdp);
1190 if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) || !raw_spin_trylock_rcu_node(rnp))
1191 return;
1192 // The grace period cannot end while we hold the rcu_node lock.
1193 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))
1194 WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
1195 raw_spin_unlock_rcu_node(rnp);
1196}
1197
1198/*
1199 * In CONFIG_RCU_STRICT_GRACE_PERIOD=y kernels, attempt to generate a
1200 * quiescent state. This is intended to be invoked when the CPU notices
1201 * a new grace period.
1202 */
1203static void rcu_strict_gp_check_qs(void)
1204{
1205 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
1206 rcu_read_lock();
1207 rcu_read_unlock();
1208 }
1209}
1210
1211/*
1212 * Update CPU-local rcu_data state to record the beginnings and ends of
1213 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1214 * structure corresponding to the current CPU, and must have irqs disabled.
1215 * Returns true if the grace-period kthread needs to be awakened.
1216 */
1217static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
1218{
1219 bool ret = false;
1220 bool need_qs;
1221 const bool offloaded = rcu_rdp_is_offloaded(rdp);
1222
1223 raw_lockdep_assert_held_rcu_node(rnp);
1224
1225 if (rdp->gp_seq == rnp->gp_seq)
1226 return false; /* Nothing to do. */
1227
1228 /* Handle the ends of any preceding grace periods first. */
1229 if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1230 unlikely(READ_ONCE(rdp->gpwrap))) {
1231 if (!offloaded)
1232 ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
1233 rdp->core_needs_qs = false;
1234 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1235 } else {
1236 if (!offloaded)
1237 ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
1238 if (rdp->core_needs_qs)
1239 rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask);
1240 }
1241
1242 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1243 if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1244 unlikely(READ_ONCE(rdp->gpwrap))) {
1245 /*
1246 * If the current grace period is waiting for this CPU,
1247 * set up to detect a quiescent state, otherwise don't
1248 * go looking for one.
1249 */
1250 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1251 need_qs = !!(rnp->qsmask & rdp->grpmask);
1252 rdp->cpu_no_qs.b.norm = need_qs;
1253 rdp->core_needs_qs = need_qs;
1254 zero_cpu_stall_ticks(rdp);
1255 }
1256 rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
1257 if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
1258 WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1259 if (IS_ENABLED(CONFIG_PROVE_RCU) && READ_ONCE(rdp->gpwrap))
1260 WRITE_ONCE(rdp->last_sched_clock, jiffies);
1261 WRITE_ONCE(rdp->gpwrap, false);
1262 rcu_gpnum_ovf(rnp, rdp);
1263 return ret;
1264}
1265
1266static void note_gp_changes(struct rcu_data *rdp)
1267{
1268 unsigned long flags;
1269 bool needwake;
1270 struct rcu_node *rnp;
1271
1272 local_irq_save(flags);
1273 rnp = rdp->mynode;
1274 if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1275 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1276 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1277 local_irq_restore(flags);
1278 return;
1279 }
1280 needwake = __note_gp_changes(rnp, rdp);
1281 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1282 rcu_strict_gp_check_qs();
1283 if (needwake)
1284 rcu_gp_kthread_wake();
1285}
1286
1287static atomic_t *rcu_gp_slow_suppress;
1288
1289/* Register a counter to suppress debugging grace-period delays. */
1290void rcu_gp_slow_register(atomic_t *rgssp)
1291{
1292 WARN_ON_ONCE(rcu_gp_slow_suppress);
1293
1294 WRITE_ONCE(rcu_gp_slow_suppress, rgssp);
1295}
1296EXPORT_SYMBOL_GPL(rcu_gp_slow_register);
1297
1298/* Unregister a counter, with NULL for not caring which. */
1299void rcu_gp_slow_unregister(atomic_t *rgssp)
1300{
1301 WARN_ON_ONCE(rgssp && rgssp != rcu_gp_slow_suppress && rcu_gp_slow_suppress != NULL);
1302
1303 WRITE_ONCE(rcu_gp_slow_suppress, NULL);
1304}
1305EXPORT_SYMBOL_GPL(rcu_gp_slow_unregister);
1306
1307static bool rcu_gp_slow_is_suppressed(void)
1308{
1309 atomic_t *rgssp = READ_ONCE(rcu_gp_slow_suppress);
1310
1311 return rgssp && atomic_read(rgssp);
1312}
1313
1314static void rcu_gp_slow(int delay)
1315{
1316 if (!rcu_gp_slow_is_suppressed() && delay > 0 &&
1317 !(rcu_seq_ctr(rcu_state.gp_seq) % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1318 schedule_timeout_idle(delay);
1319}
1320
1321static unsigned long sleep_duration;
1322
1323/* Allow rcutorture to stall the grace-period kthread. */
1324void rcu_gp_set_torture_wait(int duration)
1325{
1326 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST) && duration > 0)
1327 WRITE_ONCE(sleep_duration, duration);
1328}
1329EXPORT_SYMBOL_GPL(rcu_gp_set_torture_wait);
1330
1331/* Actually implement the aforementioned wait. */
1332static void rcu_gp_torture_wait(void)
1333{
1334 unsigned long duration;
1335
1336 if (!IS_ENABLED(CONFIG_RCU_TORTURE_TEST))
1337 return;
1338 duration = xchg(&sleep_duration, 0UL);
1339 if (duration > 0) {
1340 pr_alert("%s: Waiting %lu jiffies\n", __func__, duration);
1341 schedule_timeout_idle(duration);
1342 pr_alert("%s: Wait complete\n", __func__);
1343 }
1344}
1345
1346/*
1347 * Handler for on_each_cpu() to invoke the target CPU's RCU core
1348 * processing.
1349 */
1350static void rcu_strict_gp_boundary(void *unused)
1351{
1352 invoke_rcu_core();
1353}
1354
1355// Make the polled API aware of the beginning of a grace period.
1356static void rcu_poll_gp_seq_start(unsigned long *snap)
1357{
1358 struct rcu_node *rnp = rcu_get_root();
1359
1360 if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1361 raw_lockdep_assert_held_rcu_node(rnp);
1362
1363 // If RCU was idle, note beginning of GP.
1364 if (!rcu_seq_state(rcu_state.gp_seq_polled))
1365 rcu_seq_start(&rcu_state.gp_seq_polled);
1366
1367 // Either way, record current state.
1368 *snap = rcu_state.gp_seq_polled;
1369}
1370
1371// Make the polled API aware of the end of a grace period.
1372static void rcu_poll_gp_seq_end(unsigned long *snap)
1373{
1374 struct rcu_node *rnp = rcu_get_root();
1375
1376 if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1377 raw_lockdep_assert_held_rcu_node(rnp);
1378
1379 // If the previously noted GP is still in effect, record the
1380 // end of that GP. Either way, zero counter to avoid counter-wrap
1381 // problems.
1382 if (*snap && *snap == rcu_state.gp_seq_polled) {
1383 rcu_seq_end(&rcu_state.gp_seq_polled);
1384 rcu_state.gp_seq_polled_snap = 0;
1385 rcu_state.gp_seq_polled_exp_snap = 0;
1386 } else {
1387 *snap = 0;
1388 }
1389}
1390
1391// Make the polled API aware of the beginning of a grace period, but
1392// where caller does not hold the root rcu_node structure's lock.
1393static void rcu_poll_gp_seq_start_unlocked(unsigned long *snap)
1394{
1395 unsigned long flags;
1396 struct rcu_node *rnp = rcu_get_root();
1397
1398 if (rcu_init_invoked()) {
1399 if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1400 lockdep_assert_irqs_enabled();
1401 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1402 }
1403 rcu_poll_gp_seq_start(snap);
1404 if (rcu_init_invoked())
1405 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1406}
1407
1408// Make the polled API aware of the end of a grace period, but where
1409// caller does not hold the root rcu_node structure's lock.
1410static void rcu_poll_gp_seq_end_unlocked(unsigned long *snap)
1411{
1412 unsigned long flags;
1413 struct rcu_node *rnp = rcu_get_root();
1414
1415 if (rcu_init_invoked()) {
1416 if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1417 lockdep_assert_irqs_enabled();
1418 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1419 }
1420 rcu_poll_gp_seq_end(snap);
1421 if (rcu_init_invoked())
1422 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1423}
1424
1425/*
1426 * Initialize a new grace period. Return false if no grace period required.
1427 */
1428static noinline_for_stack bool rcu_gp_init(void)
1429{
1430 unsigned long flags;
1431 unsigned long oldmask;
1432 unsigned long mask;
1433 struct rcu_data *rdp;
1434 struct rcu_node *rnp = rcu_get_root();
1435
1436 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1437 raw_spin_lock_irq_rcu_node(rnp);
1438 if (!READ_ONCE(rcu_state.gp_flags)) {
1439 /* Spurious wakeup, tell caller to go back to sleep. */
1440 raw_spin_unlock_irq_rcu_node(rnp);
1441 return false;
1442 }
1443 WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1444
1445 if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1446 /*
1447 * Grace period already in progress, don't start another.
1448 * Not supposed to be able to happen.
1449 */
1450 raw_spin_unlock_irq_rcu_node(rnp);
1451 return false;
1452 }
1453
1454 /* Advance to a new grace period and initialize state. */
1455 record_gp_stall_check_time();
1456 /* Record GP times before starting GP, hence rcu_seq_start(). */
1457 rcu_seq_start(&rcu_state.gp_seq);
1458 ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
1459 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
1460 rcu_poll_gp_seq_start(&rcu_state.gp_seq_polled_snap);
1461 raw_spin_unlock_irq_rcu_node(rnp);
1462
1463 /*
1464 * Apply per-leaf buffered online and offline operations to
1465 * the rcu_node tree. Note that this new grace period need not
1466 * wait for subsequent online CPUs, and that RCU hooks in the CPU
1467 * offlining path, when combined with checks in this function,
1468 * will handle CPUs that are currently going offline or that will
1469 * go offline later. Please also refer to "Hotplug CPU" section
1470 * of RCU's Requirements documentation.
1471 */
1472 WRITE_ONCE(rcu_state.gp_state, RCU_GP_ONOFF);
1473 /* Exclude CPU hotplug operations. */
1474 rcu_for_each_leaf_node(rnp) {
1475 local_irq_save(flags);
1476 arch_spin_lock(&rcu_state.ofl_lock);
1477 raw_spin_lock_rcu_node(rnp);
1478 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1479 !rnp->wait_blkd_tasks) {
1480 /* Nothing to do on this leaf rcu_node structure. */
1481 raw_spin_unlock_rcu_node(rnp);
1482 arch_spin_unlock(&rcu_state.ofl_lock);
1483 local_irq_restore(flags);
1484 continue;
1485 }
1486
1487 /* Record old state, apply changes to ->qsmaskinit field. */
1488 oldmask = rnp->qsmaskinit;
1489 rnp->qsmaskinit = rnp->qsmaskinitnext;
1490
1491 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1492 if (!oldmask != !rnp->qsmaskinit) {
1493 if (!oldmask) { /* First online CPU for rcu_node. */
1494 if (!rnp->wait_blkd_tasks) /* Ever offline? */
1495 rcu_init_new_rnp(rnp);
1496 } else if (rcu_preempt_has_tasks(rnp)) {
1497 rnp->wait_blkd_tasks = true; /* blocked tasks */
1498 } else { /* Last offline CPU and can propagate. */
1499 rcu_cleanup_dead_rnp(rnp);
1500 }
1501 }
1502
1503 /*
1504 * If all waited-on tasks from prior grace period are
1505 * done, and if all this rcu_node structure's CPUs are
1506 * still offline, propagate up the rcu_node tree and
1507 * clear ->wait_blkd_tasks. Otherwise, if one of this
1508 * rcu_node structure's CPUs has since come back online,
1509 * simply clear ->wait_blkd_tasks.
1510 */
1511 if (rnp->wait_blkd_tasks &&
1512 (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1513 rnp->wait_blkd_tasks = false;
1514 if (!rnp->qsmaskinit)
1515 rcu_cleanup_dead_rnp(rnp);
1516 }
1517
1518 raw_spin_unlock_rcu_node(rnp);
1519 arch_spin_unlock(&rcu_state.ofl_lock);
1520 local_irq_restore(flags);
1521 }
1522 rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1523
1524 /*
1525 * Set the quiescent-state-needed bits in all the rcu_node
1526 * structures for all currently online CPUs in breadth-first
1527 * order, starting from the root rcu_node structure, relying on the
1528 * layout of the tree within the rcu_state.node[] array. Note that
1529 * other CPUs will access only the leaves of the hierarchy, thus
1530 * seeing that no grace period is in progress, at least until the
1531 * corresponding leaf node has been initialized.
1532 *
1533 * The grace period cannot complete until the initialization
1534 * process finishes, because this kthread handles both.
1535 */
1536 WRITE_ONCE(rcu_state.gp_state, RCU_GP_INIT);
1537 rcu_for_each_node_breadth_first(rnp) {
1538 rcu_gp_slow(gp_init_delay);
1539 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1540 rdp = this_cpu_ptr(&rcu_data);
1541 rcu_preempt_check_blocked_tasks(rnp);
1542 rnp->qsmask = rnp->qsmaskinit;
1543 WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
1544 if (rnp == rdp->mynode)
1545 (void)__note_gp_changes(rnp, rdp);
1546 rcu_preempt_boost_start_gp(rnp);
1547 trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1548 rnp->level, rnp->grplo,
1549 rnp->grphi, rnp->qsmask);
1550 /* Quiescent states for tasks on any now-offline CPUs. */
1551 mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1552 rnp->rcu_gp_init_mask = mask;
1553 if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1554 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1555 else
1556 raw_spin_unlock_irq_rcu_node(rnp);
1557 cond_resched_tasks_rcu_qs();
1558 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1559 }
1560
1561 // If strict, make all CPUs aware of new grace period.
1562 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
1563 on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
1564
1565 return true;
1566}
1567
1568/*
1569 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1570 * time.
1571 */
1572static bool rcu_gp_fqs_check_wake(int *gfp)
1573{
1574 struct rcu_node *rnp = rcu_get_root();
1575
1576 // If under overload conditions, force an immediate FQS scan.
1577 if (*gfp & RCU_GP_FLAG_OVLD)
1578 return true;
1579
1580 // Someone like call_rcu() requested a force-quiescent-state scan.
1581 *gfp = READ_ONCE(rcu_state.gp_flags);
1582 if (*gfp & RCU_GP_FLAG_FQS)
1583 return true;
1584
1585 // The current grace period has completed.
1586 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1587 return true;
1588
1589 return false;
1590}
1591
1592/*
1593 * Do one round of quiescent-state forcing.
1594 */
1595static void rcu_gp_fqs(bool first_time)
1596{
1597 int nr_fqs = READ_ONCE(rcu_state.nr_fqs_jiffies_stall);
1598 struct rcu_node *rnp = rcu_get_root();
1599
1600 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1601 WRITE_ONCE(rcu_state.n_force_qs, rcu_state.n_force_qs + 1);
1602
1603 WARN_ON_ONCE(nr_fqs > 3);
1604 /* Only countdown nr_fqs for stall purposes if jiffies moves. */
1605 if (nr_fqs) {
1606 if (nr_fqs == 1) {
1607 WRITE_ONCE(rcu_state.jiffies_stall,
1608 jiffies + rcu_jiffies_till_stall_check());
1609 }
1610 WRITE_ONCE(rcu_state.nr_fqs_jiffies_stall, --nr_fqs);
1611 }
1612
1613 if (first_time) {
1614 /* Collect dyntick-idle snapshots. */
1615 force_qs_rnp(dyntick_save_progress_counter);
1616 } else {
1617 /* Handle dyntick-idle and offline CPUs. */
1618 force_qs_rnp(rcu_implicit_dynticks_qs);
1619 }
1620 /* Clear flag to prevent immediate re-entry. */
1621 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
1622 raw_spin_lock_irq_rcu_node(rnp);
1623 WRITE_ONCE(rcu_state.gp_flags,
1624 READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
1625 raw_spin_unlock_irq_rcu_node(rnp);
1626 }
1627}
1628
1629/*
1630 * Loop doing repeated quiescent-state forcing until the grace period ends.
1631 */
1632static noinline_for_stack void rcu_gp_fqs_loop(void)
1633{
1634 bool first_gp_fqs = true;
1635 int gf = 0;
1636 unsigned long j;
1637 int ret;
1638 struct rcu_node *rnp = rcu_get_root();
1639
1640 j = READ_ONCE(jiffies_till_first_fqs);
1641 if (rcu_state.cbovld)
1642 gf = RCU_GP_FLAG_OVLD;
1643 ret = 0;
1644 for (;;) {
1645 if (rcu_state.cbovld) {
1646 j = (j + 2) / 3;
1647 if (j <= 0)
1648 j = 1;
1649 }
1650 if (!ret || time_before(jiffies + j, rcu_state.jiffies_force_qs)) {
1651 WRITE_ONCE(rcu_state.jiffies_force_qs, jiffies + j);
1652 /*
1653 * jiffies_force_qs before RCU_GP_WAIT_FQS state
1654 * update; required for stall checks.
1655 */
1656 smp_wmb();
1657 WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
1658 jiffies + (j ? 3 * j : 2));
1659 }
1660 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1661 TPS("fqswait"));
1662 WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_FQS);
1663 (void)swait_event_idle_timeout_exclusive(rcu_state.gp_wq,
1664 rcu_gp_fqs_check_wake(&gf), j);
1665 rcu_gp_torture_wait();
1666 WRITE_ONCE(rcu_state.gp_state, RCU_GP_DOING_FQS);
1667 /* Locking provides needed memory barriers. */
1668 /*
1669 * Exit the loop if the root rcu_node structure indicates that the grace period
1670 * has ended, leave the loop. The rcu_preempt_blocked_readers_cgp(rnp) check
1671 * is required only for single-node rcu_node trees because readers blocking
1672 * the current grace period are queued only on leaf rcu_node structures.
1673 * For multi-node trees, checking the root node's ->qsmask suffices, because a
1674 * given root node's ->qsmask bit is cleared only when all CPUs and tasks from
1675 * the corresponding leaf nodes have passed through their quiescent state.
1676 */
1677 if (!READ_ONCE(rnp->qsmask) &&
1678 !rcu_preempt_blocked_readers_cgp(rnp))
1679 break;
1680 /* If time for quiescent-state forcing, do it. */
1681 if (!time_after(rcu_state.jiffies_force_qs, jiffies) ||
1682 (gf & (RCU_GP_FLAG_FQS | RCU_GP_FLAG_OVLD))) {
1683 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1684 TPS("fqsstart"));
1685 rcu_gp_fqs(first_gp_fqs);
1686 gf = 0;
1687 if (first_gp_fqs) {
1688 first_gp_fqs = false;
1689 gf = rcu_state.cbovld ? RCU_GP_FLAG_OVLD : 0;
1690 }
1691 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1692 TPS("fqsend"));
1693 cond_resched_tasks_rcu_qs();
1694 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1695 ret = 0; /* Force full wait till next FQS. */
1696 j = READ_ONCE(jiffies_till_next_fqs);
1697 } else {
1698 /* Deal with stray signal. */
1699 cond_resched_tasks_rcu_qs();
1700 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1701 WARN_ON(signal_pending(current));
1702 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1703 TPS("fqswaitsig"));
1704 ret = 1; /* Keep old FQS timing. */
1705 j = jiffies;
1706 if (time_after(jiffies, rcu_state.jiffies_force_qs))
1707 j = 1;
1708 else
1709 j = rcu_state.jiffies_force_qs - j;
1710 gf = 0;
1711 }
1712 }
1713}
1714
1715/*
1716 * Clean up after the old grace period.
1717 */
1718static noinline void rcu_gp_cleanup(void)
1719{
1720 int cpu;
1721 bool needgp = false;
1722 unsigned long gp_duration;
1723 unsigned long new_gp_seq;
1724 bool offloaded;
1725 struct rcu_data *rdp;
1726 struct rcu_node *rnp = rcu_get_root();
1727 struct swait_queue_head *sq;
1728
1729 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1730 raw_spin_lock_irq_rcu_node(rnp);
1731 rcu_state.gp_end = jiffies;
1732 gp_duration = rcu_state.gp_end - rcu_state.gp_start;
1733 if (gp_duration > rcu_state.gp_max)
1734 rcu_state.gp_max = gp_duration;
1735
1736 /*
1737 * We know the grace period is complete, but to everyone else
1738 * it appears to still be ongoing. But it is also the case
1739 * that to everyone else it looks like there is nothing that
1740 * they can do to advance the grace period. It is therefore
1741 * safe for us to drop the lock in order to mark the grace
1742 * period as completed in all of the rcu_node structures.
1743 */
1744 rcu_poll_gp_seq_end(&rcu_state.gp_seq_polled_snap);
1745 raw_spin_unlock_irq_rcu_node(rnp);
1746
1747 /*
1748 * Propagate new ->gp_seq value to rcu_node structures so that
1749 * other CPUs don't have to wait until the start of the next grace
1750 * period to process their callbacks. This also avoids some nasty
1751 * RCU grace-period initialization races by forcing the end of
1752 * the current grace period to be completely recorded in all of
1753 * the rcu_node structures before the beginning of the next grace
1754 * period is recorded in any of the rcu_node structures.
1755 */
1756 new_gp_seq = rcu_state.gp_seq;
1757 rcu_seq_end(&new_gp_seq);
1758 rcu_for_each_node_breadth_first(rnp) {
1759 raw_spin_lock_irq_rcu_node(rnp);
1760 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
1761 dump_blkd_tasks(rnp, 10);
1762 WARN_ON_ONCE(rnp->qsmask);
1763 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
1764 if (!rnp->parent)
1765 smp_mb(); // Order against failing poll_state_synchronize_rcu_full().
1766 rdp = this_cpu_ptr(&rcu_data);
1767 if (rnp == rdp->mynode)
1768 needgp = __note_gp_changes(rnp, rdp) || needgp;
1769 /* smp_mb() provided by prior unlock-lock pair. */
1770 needgp = rcu_future_gp_cleanup(rnp) || needgp;
1771 // Reset overload indication for CPUs no longer overloaded
1772 if (rcu_is_leaf_node(rnp))
1773 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->cbovldmask) {
1774 rdp = per_cpu_ptr(&rcu_data, cpu);
1775 check_cb_ovld_locked(rdp, rnp);
1776 }
1777 sq = rcu_nocb_gp_get(rnp);
1778 raw_spin_unlock_irq_rcu_node(rnp);
1779 rcu_nocb_gp_cleanup(sq);
1780 cond_resched_tasks_rcu_qs();
1781 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1782 rcu_gp_slow(gp_cleanup_delay);
1783 }
1784 rnp = rcu_get_root();
1785 raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
1786
1787 /* Declare grace period done, trace first to use old GP number. */
1788 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
1789 rcu_seq_end(&rcu_state.gp_seq);
1790 ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
1791 WRITE_ONCE(rcu_state.gp_state, RCU_GP_IDLE);
1792 /* Check for GP requests since above loop. */
1793 rdp = this_cpu_ptr(&rcu_data);
1794 if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
1795 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
1796 TPS("CleanupMore"));
1797 needgp = true;
1798 }
1799 /* Advance CBs to reduce false positives below. */
1800 offloaded = rcu_rdp_is_offloaded(rdp);
1801 if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
1802
1803 // We get here if a grace period was needed (“needgp”)
1804 // and the above call to rcu_accelerate_cbs() did not set
1805 // the RCU_GP_FLAG_INIT bit in ->gp_state (which records
1806 // the need for another grace period). The purpose
1807 // of the “offloaded” check is to avoid invoking
1808 // rcu_accelerate_cbs() on an offloaded CPU because we do not
1809 // hold the ->nocb_lock needed to safely access an offloaded
1810 // ->cblist. We do not want to acquire that lock because
1811 // it can be heavily contended during callback floods.
1812
1813 WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
1814 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
1815 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("newreq"));
1816 } else {
1817
1818 // We get here either if there is no need for an
1819 // additional grace period or if rcu_accelerate_cbs() has
1820 // already set the RCU_GP_FLAG_INIT bit in ->gp_flags.
1821 // So all we need to do is to clear all of the other
1822 // ->gp_flags bits.
1823
1824 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags & RCU_GP_FLAG_INIT);
1825 }
1826 raw_spin_unlock_irq_rcu_node(rnp);
1827
1828 // If strict, make all CPUs aware of the end of the old grace period.
1829 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
1830 on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
1831}
1832
1833/*
1834 * Body of kthread that handles grace periods.
1835 */
1836static int __noreturn rcu_gp_kthread(void *unused)
1837{
1838 rcu_bind_gp_kthread();
1839 for (;;) {
1840
1841 /* Handle grace-period start. */
1842 for (;;) {
1843 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1844 TPS("reqwait"));
1845 WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_GPS);
1846 swait_event_idle_exclusive(rcu_state.gp_wq,
1847 READ_ONCE(rcu_state.gp_flags) &
1848 RCU_GP_FLAG_INIT);
1849 rcu_gp_torture_wait();
1850 WRITE_ONCE(rcu_state.gp_state, RCU_GP_DONE_GPS);
1851 /* Locking provides needed memory barrier. */
1852 if (rcu_gp_init())
1853 break;
1854 cond_resched_tasks_rcu_qs();
1855 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1856 WARN_ON(signal_pending(current));
1857 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1858 TPS("reqwaitsig"));
1859 }
1860
1861 /* Handle quiescent-state forcing. */
1862 rcu_gp_fqs_loop();
1863
1864 /* Handle grace-period end. */
1865 WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANUP);
1866 rcu_gp_cleanup();
1867 WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANED);
1868 }
1869}
1870
1871/*
1872 * Report a full set of quiescent states to the rcu_state data structure.
1873 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
1874 * another grace period is required. Whether we wake the grace-period
1875 * kthread or it awakens itself for the next round of quiescent-state
1876 * forcing, that kthread will clean up after the just-completed grace
1877 * period. Note that the caller must hold rnp->lock, which is released
1878 * before return.
1879 */
1880static void rcu_report_qs_rsp(unsigned long flags)
1881 __releases(rcu_get_root()->lock)
1882{
1883 raw_lockdep_assert_held_rcu_node(rcu_get_root());
1884 WARN_ON_ONCE(!rcu_gp_in_progress());
1885 WRITE_ONCE(rcu_state.gp_flags,
1886 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
1887 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
1888 rcu_gp_kthread_wake();
1889}
1890
1891/*
1892 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1893 * Allows quiescent states for a group of CPUs to be reported at one go
1894 * to the specified rcu_node structure, though all the CPUs in the group
1895 * must be represented by the same rcu_node structure (which need not be a
1896 * leaf rcu_node structure, though it often will be). The gps parameter
1897 * is the grace-period snapshot, which means that the quiescent states
1898 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
1899 * must be held upon entry, and it is released before return.
1900 *
1901 * As a special case, if mask is zero, the bit-already-cleared check is
1902 * disabled. This allows propagating quiescent state due to resumed tasks
1903 * during grace-period initialization.
1904 */
1905static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
1906 unsigned long gps, unsigned long flags)
1907 __releases(rnp->lock)
1908{
1909 unsigned long oldmask = 0;
1910 struct rcu_node *rnp_c;
1911
1912 raw_lockdep_assert_held_rcu_node(rnp);
1913
1914 /* Walk up the rcu_node hierarchy. */
1915 for (;;) {
1916 if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
1917
1918 /*
1919 * Our bit has already been cleared, or the
1920 * relevant grace period is already over, so done.
1921 */
1922 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1923 return;
1924 }
1925 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
1926 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
1927 rcu_preempt_blocked_readers_cgp(rnp));
1928 WRITE_ONCE(rnp->qsmask, rnp->qsmask & ~mask);
1929 trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
1930 mask, rnp->qsmask, rnp->level,
1931 rnp->grplo, rnp->grphi,
1932 !!rnp->gp_tasks);
1933 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1934
1935 /* Other bits still set at this level, so done. */
1936 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1937 return;
1938 }
1939 rnp->completedqs = rnp->gp_seq;
1940 mask = rnp->grpmask;
1941 if (rnp->parent == NULL) {
1942
1943 /* No more levels. Exit loop holding root lock. */
1944
1945 break;
1946 }
1947 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1948 rnp_c = rnp;
1949 rnp = rnp->parent;
1950 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1951 oldmask = READ_ONCE(rnp_c->qsmask);
1952 }
1953
1954 /*
1955 * Get here if we are the last CPU to pass through a quiescent
1956 * state for this grace period. Invoke rcu_report_qs_rsp()
1957 * to clean up and start the next grace period if one is needed.
1958 */
1959 rcu_report_qs_rsp(flags); /* releases rnp->lock. */
1960}
1961
1962/*
1963 * Record a quiescent state for all tasks that were previously queued
1964 * on the specified rcu_node structure and that were blocking the current
1965 * RCU grace period. The caller must hold the corresponding rnp->lock with
1966 * irqs disabled, and this lock is released upon return, but irqs remain
1967 * disabled.
1968 */
1969static void __maybe_unused
1970rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
1971 __releases(rnp->lock)
1972{
1973 unsigned long gps;
1974 unsigned long mask;
1975 struct rcu_node *rnp_p;
1976
1977 raw_lockdep_assert_held_rcu_node(rnp);
1978 if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RCU)) ||
1979 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
1980 rnp->qsmask != 0) {
1981 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1982 return; /* Still need more quiescent states! */
1983 }
1984
1985 rnp->completedqs = rnp->gp_seq;
1986 rnp_p = rnp->parent;
1987 if (rnp_p == NULL) {
1988 /*
1989 * Only one rcu_node structure in the tree, so don't
1990 * try to report up to its nonexistent parent!
1991 */
1992 rcu_report_qs_rsp(flags);
1993 return;
1994 }
1995
1996 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
1997 gps = rnp->gp_seq;
1998 mask = rnp->grpmask;
1999 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2000 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2001 rcu_report_qs_rnp(mask, rnp_p, gps, flags);
2002}
2003
2004/*
2005 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2006 * structure. This must be called from the specified CPU.
2007 */
2008static void
2009rcu_report_qs_rdp(struct rcu_data *rdp)
2010{
2011 unsigned long flags;
2012 unsigned long mask;
2013 bool needacc = false;
2014 struct rcu_node *rnp;
2015
2016 WARN_ON_ONCE(rdp->cpu != smp_processor_id());
2017 rnp = rdp->mynode;
2018 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2019 if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
2020 rdp->gpwrap) {
2021
2022 /*
2023 * The grace period in which this quiescent state was
2024 * recorded has ended, so don't report it upwards.
2025 * We will instead need a new quiescent state that lies
2026 * within the current grace period.
2027 */
2028 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2029 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2030 return;
2031 }
2032 mask = rdp->grpmask;
2033 rdp->core_needs_qs = false;
2034 if ((rnp->qsmask & mask) == 0) {
2035 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2036 } else {
2037 /*
2038 * This GP can't end until cpu checks in, so all of our
2039 * callbacks can be processed during the next GP.
2040 *
2041 * NOCB kthreads have their own way to deal with that...
2042 */
2043 if (!rcu_rdp_is_offloaded(rdp)) {
2044 /*
2045 * The current GP has not yet ended, so it
2046 * should not be possible for rcu_accelerate_cbs()
2047 * to return true. So complain, but don't awaken.
2048 */
2049 WARN_ON_ONCE(rcu_accelerate_cbs(rnp, rdp));
2050 } else if (!rcu_segcblist_completely_offloaded(&rdp->cblist)) {
2051 /*
2052 * ...but NOCB kthreads may miss or delay callbacks acceleration
2053 * if in the middle of a (de-)offloading process.
2054 */
2055 needacc = true;
2056 }
2057
2058 rcu_disable_urgency_upon_qs(rdp);
2059 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2060 /* ^^^ Released rnp->lock */
2061
2062 if (needacc) {
2063 rcu_nocb_lock_irqsave(rdp, flags);
2064 rcu_accelerate_cbs_unlocked(rnp, rdp);
2065 rcu_nocb_unlock_irqrestore(rdp, flags);
2066 }
2067 }
2068}
2069
2070/*
2071 * Check to see if there is a new grace period of which this CPU
2072 * is not yet aware, and if so, set up local rcu_data state for it.
2073 * Otherwise, see if this CPU has just passed through its first
2074 * quiescent state for this grace period, and record that fact if so.
2075 */
2076static void
2077rcu_check_quiescent_state(struct rcu_data *rdp)
2078{
2079 /* Check for grace-period ends and beginnings. */
2080 note_gp_changes(rdp);
2081
2082 /*
2083 * Does this CPU still need to do its part for current grace period?
2084 * If no, return and let the other CPUs do their part as well.
2085 */
2086 if (!rdp->core_needs_qs)
2087 return;
2088
2089 /*
2090 * Was there a quiescent state since the beginning of the grace
2091 * period? If no, then exit and wait for the next call.
2092 */
2093 if (rdp->cpu_no_qs.b.norm)
2094 return;
2095
2096 /*
2097 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2098 * judge of that).
2099 */
2100 rcu_report_qs_rdp(rdp);
2101}
2102
2103/* Return true if callback-invocation time limit exceeded. */
2104static bool rcu_do_batch_check_time(long count, long tlimit,
2105 bool jlimit_check, unsigned long jlimit)
2106{
2107 // Invoke local_clock() only once per 32 consecutive callbacks.
2108 return unlikely(tlimit) &&
2109 (!likely(count & 31) ||
2110 (IS_ENABLED(CONFIG_RCU_DOUBLE_CHECK_CB_TIME) &&
2111 jlimit_check && time_after(jiffies, jlimit))) &&
2112 local_clock() >= tlimit;
2113}
2114
2115/*
2116 * Invoke any RCU callbacks that have made it to the end of their grace
2117 * period. Throttle as specified by rdp->blimit.
2118 */
2119static void rcu_do_batch(struct rcu_data *rdp)
2120{
2121 long bl;
2122 long count = 0;
2123 int div;
2124 bool __maybe_unused empty;
2125 unsigned long flags;
2126 unsigned long jlimit;
2127 bool jlimit_check = false;
2128 long pending;
2129 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2130 struct rcu_head *rhp;
2131 long tlimit = 0;
2132
2133 /* If no callbacks are ready, just return. */
2134 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2135 trace_rcu_batch_start(rcu_state.name,
2136 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2137 trace_rcu_batch_end(rcu_state.name, 0,
2138 !rcu_segcblist_empty(&rdp->cblist),
2139 need_resched(), is_idle_task(current),
2140 rcu_is_callbacks_kthread(rdp));
2141 return;
2142 }
2143
2144 /*
2145 * Extract the list of ready callbacks, disabling IRQs to prevent
2146 * races with call_rcu() from interrupt handlers. Leave the
2147 * callback counts, as rcu_barrier() needs to be conservative.
2148 *
2149 * Callbacks execution is fully ordered against preceding grace period
2150 * completion (materialized by rnp->gp_seq update) thanks to the
2151 * smp_mb__after_unlock_lock() upon node locking required for callbacks
2152 * advancing. In NOCB mode this ordering is then further relayed through
2153 * the nocb locking that protects both callbacks advancing and extraction.
2154 */
2155 rcu_nocb_lock_irqsave(rdp, flags);
2156 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2157 pending = rcu_segcblist_get_seglen(&rdp->cblist, RCU_DONE_TAIL);
2158 div = READ_ONCE(rcu_divisor);
2159 div = div < 0 ? 7 : div > sizeof(long) * 8 - 2 ? sizeof(long) * 8 - 2 : div;
2160 bl = max(rdp->blimit, pending >> div);
2161 if ((in_serving_softirq() || rdp->rcu_cpu_kthread_status == RCU_KTHREAD_RUNNING) &&
2162 (IS_ENABLED(CONFIG_RCU_DOUBLE_CHECK_CB_TIME) || unlikely(bl > 100))) {
2163 const long npj = NSEC_PER_SEC / HZ;
2164 long rrn = READ_ONCE(rcu_resched_ns);
2165
2166 rrn = rrn < NSEC_PER_MSEC ? NSEC_PER_MSEC : rrn > NSEC_PER_SEC ? NSEC_PER_SEC : rrn;
2167 tlimit = local_clock() + rrn;
2168 jlimit = jiffies + (rrn + npj + 1) / npj;
2169 jlimit_check = true;
2170 }
2171 trace_rcu_batch_start(rcu_state.name,
2172 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2173 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2174 if (rcu_rdp_is_offloaded(rdp))
2175 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2176
2177 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbDequeued"));
2178 rcu_nocb_unlock_irqrestore(rdp, flags);
2179
2180 /* Invoke callbacks. */
2181 tick_dep_set_task(current, TICK_DEP_BIT_RCU);
2182 rhp = rcu_cblist_dequeue(&rcl);
2183
2184 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2185 rcu_callback_t f;
2186
2187 count++;
2188 debug_rcu_head_unqueue(rhp);
2189
2190 rcu_lock_acquire(&rcu_callback_map);
2191 trace_rcu_invoke_callback(rcu_state.name, rhp);
2192
2193 f = rhp->func;
2194 debug_rcu_head_callback(rhp);
2195 WRITE_ONCE(rhp->func, (rcu_callback_t)0L);
2196 f(rhp);
2197
2198 rcu_lock_release(&rcu_callback_map);
2199
2200 /*
2201 * Stop only if limit reached and CPU has something to do.
2202 */
2203 if (in_serving_softirq()) {
2204 if (count >= bl && (need_resched() || !is_idle_task(current)))
2205 break;
2206 /*
2207 * Make sure we don't spend too much time here and deprive other
2208 * softirq vectors of CPU cycles.
2209 */
2210 if (rcu_do_batch_check_time(count, tlimit, jlimit_check, jlimit))
2211 break;
2212 } else {
2213 // In rcuc/rcuoc context, so no worries about
2214 // depriving other softirq vectors of CPU cycles.
2215 local_bh_enable();
2216 lockdep_assert_irqs_enabled();
2217 cond_resched_tasks_rcu_qs();
2218 lockdep_assert_irqs_enabled();
2219 local_bh_disable();
2220 // But rcuc kthreads can delay quiescent-state
2221 // reporting, so check time limits for them.
2222 if (rdp->rcu_cpu_kthread_status == RCU_KTHREAD_RUNNING &&
2223 rcu_do_batch_check_time(count, tlimit, jlimit_check, jlimit)) {
2224 rdp->rcu_cpu_has_work = 1;
2225 break;
2226 }
2227 }
2228 }
2229
2230 rcu_nocb_lock_irqsave(rdp, flags);
2231 rdp->n_cbs_invoked += count;
2232 trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2233 is_idle_task(current), rcu_is_callbacks_kthread(rdp));
2234
2235 /* Update counts and requeue any remaining callbacks. */
2236 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2237 rcu_segcblist_add_len(&rdp->cblist, -count);
2238
2239 /* Reinstate batch limit if we have worked down the excess. */
2240 count = rcu_segcblist_n_cbs(&rdp->cblist);
2241 if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
2242 rdp->blimit = blimit;
2243
2244 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2245 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2246 rdp->qlen_last_fqs_check = 0;
2247 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
2248 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2249 rdp->qlen_last_fqs_check = count;
2250
2251 /*
2252 * The following usually indicates a double call_rcu(). To track
2253 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2254 */
2255 empty = rcu_segcblist_empty(&rdp->cblist);
2256 WARN_ON_ONCE(count == 0 && !empty);
2257 WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2258 count != 0 && empty);
2259 WARN_ON_ONCE(count == 0 && rcu_segcblist_n_segment_cbs(&rdp->cblist) != 0);
2260 WARN_ON_ONCE(!empty && rcu_segcblist_n_segment_cbs(&rdp->cblist) == 0);
2261
2262 rcu_nocb_unlock_irqrestore(rdp, flags);
2263
2264 tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
2265}
2266
2267/*
2268 * This function is invoked from each scheduling-clock interrupt,
2269 * and checks to see if this CPU is in a non-context-switch quiescent
2270 * state, for example, user mode or idle loop. It also schedules RCU
2271 * core processing. If the current grace period has gone on too long,
2272 * it will ask the scheduler to manufacture a context switch for the sole
2273 * purpose of providing the needed quiescent state.
2274 */
2275void rcu_sched_clock_irq(int user)
2276{
2277 unsigned long j;
2278
2279 if (IS_ENABLED(CONFIG_PROVE_RCU)) {
2280 j = jiffies;
2281 WARN_ON_ONCE(time_before(j, __this_cpu_read(rcu_data.last_sched_clock)));
2282 __this_cpu_write(rcu_data.last_sched_clock, j);
2283 }
2284 trace_rcu_utilization(TPS("Start scheduler-tick"));
2285 lockdep_assert_irqs_disabled();
2286 raw_cpu_inc(rcu_data.ticks_this_gp);
2287 /* The load-acquire pairs with the store-release setting to true. */
2288 if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
2289 /* Idle and userspace execution already are quiescent states. */
2290 if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2291 set_tsk_need_resched(current);
2292 set_preempt_need_resched();
2293 }
2294 __this_cpu_write(rcu_data.rcu_urgent_qs, false);
2295 }
2296 rcu_flavor_sched_clock_irq(user);
2297 if (rcu_pending(user))
2298 invoke_rcu_core();
2299 if (user || rcu_is_cpu_rrupt_from_idle())
2300 rcu_note_voluntary_context_switch(current);
2301 lockdep_assert_irqs_disabled();
2302
2303 trace_rcu_utilization(TPS("End scheduler-tick"));
2304}
2305
2306/*
2307 * Scan the leaf rcu_node structures. For each structure on which all
2308 * CPUs have reported a quiescent state and on which there are tasks
2309 * blocking the current grace period, initiate RCU priority boosting.
2310 * Otherwise, invoke the specified function to check dyntick state for
2311 * each CPU that has not yet reported a quiescent state.
2312 */
2313static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
2314{
2315 int cpu;
2316 unsigned long flags;
2317 struct rcu_node *rnp;
2318
2319 rcu_state.cbovld = rcu_state.cbovldnext;
2320 rcu_state.cbovldnext = false;
2321 rcu_for_each_leaf_node(rnp) {
2322 unsigned long mask = 0;
2323 unsigned long rsmask = 0;
2324
2325 cond_resched_tasks_rcu_qs();
2326 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2327 rcu_state.cbovldnext |= !!rnp->cbovldmask;
2328 if (rnp->qsmask == 0) {
2329 if (rcu_preempt_blocked_readers_cgp(rnp)) {
2330 /*
2331 * No point in scanning bits because they
2332 * are all zero. But we might need to
2333 * priority-boost blocked readers.
2334 */
2335 rcu_initiate_boost(rnp, flags);
2336 /* rcu_initiate_boost() releases rnp->lock */
2337 continue;
2338 }
2339 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2340 continue;
2341 }
2342 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->qsmask) {
2343 struct rcu_data *rdp;
2344 int ret;
2345
2346 rdp = per_cpu_ptr(&rcu_data, cpu);
2347 ret = f(rdp);
2348 if (ret > 0) {
2349 mask |= rdp->grpmask;
2350 rcu_disable_urgency_upon_qs(rdp);
2351 }
2352 if (ret < 0)
2353 rsmask |= rdp->grpmask;
2354 }
2355 if (mask != 0) {
2356 /* Idle/offline CPUs, report (releases rnp->lock). */
2357 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2358 } else {
2359 /* Nothing to do here, so just drop the lock. */
2360 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2361 }
2362
2363 for_each_leaf_node_cpu_mask(rnp, cpu, rsmask)
2364 resched_cpu(cpu);
2365 }
2366}
2367
2368/*
2369 * Force quiescent states on reluctant CPUs, and also detect which
2370 * CPUs are in dyntick-idle mode.
2371 */
2372void rcu_force_quiescent_state(void)
2373{
2374 unsigned long flags;
2375 bool ret;
2376 struct rcu_node *rnp;
2377 struct rcu_node *rnp_old = NULL;
2378
2379 if (!rcu_gp_in_progress())
2380 return;
2381 /* Funnel through hierarchy to reduce memory contention. */
2382 rnp = raw_cpu_read(rcu_data.mynode);
2383 for (; rnp != NULL; rnp = rnp->parent) {
2384 ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2385 !raw_spin_trylock(&rnp->fqslock);
2386 if (rnp_old != NULL)
2387 raw_spin_unlock(&rnp_old->fqslock);
2388 if (ret)
2389 return;
2390 rnp_old = rnp;
2391 }
2392 /* rnp_old == rcu_get_root(), rnp == NULL. */
2393
2394 /* Reached the root of the rcu_node tree, acquire lock. */
2395 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2396 raw_spin_unlock(&rnp_old->fqslock);
2397 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2398 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2399 return; /* Someone beat us to it. */
2400 }
2401 WRITE_ONCE(rcu_state.gp_flags,
2402 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2403 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2404 rcu_gp_kthread_wake();
2405}
2406EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
2407
2408// Workqueue handler for an RCU reader for kernels enforcing struct RCU
2409// grace periods.
2410static void strict_work_handler(struct work_struct *work)
2411{
2412 rcu_read_lock();
2413 rcu_read_unlock();
2414}
2415
2416/* Perform RCU core processing work for the current CPU. */
2417static __latent_entropy void rcu_core(void)
2418{
2419 unsigned long flags;
2420 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2421 struct rcu_node *rnp = rdp->mynode;
2422 /*
2423 * On RT rcu_core() can be preempted when IRQs aren't disabled.
2424 * Therefore this function can race with concurrent NOCB (de-)offloading
2425 * on this CPU and the below condition must be considered volatile.
2426 * However if we race with:
2427 *
2428 * _ Offloading: In the worst case we accelerate or process callbacks
2429 * concurrently with NOCB kthreads. We are guaranteed to
2430 * call rcu_nocb_lock() if that happens.
2431 *
2432 * _ Deoffloading: In the worst case we miss callbacks acceleration or
2433 * processing. This is fine because the early stage
2434 * of deoffloading invokes rcu_core() after setting
2435 * SEGCBLIST_RCU_CORE. So we guarantee that we'll process
2436 * what could have been dismissed without the need to wait
2437 * for the next rcu_pending() check in the next jiffy.
2438 */
2439 const bool do_batch = !rcu_segcblist_completely_offloaded(&rdp->cblist);
2440
2441 if (cpu_is_offline(smp_processor_id()))
2442 return;
2443 trace_rcu_utilization(TPS("Start RCU core"));
2444 WARN_ON_ONCE(!rdp->beenonline);
2445
2446 /* Report any deferred quiescent states if preemption enabled. */
2447 if (IS_ENABLED(CONFIG_PREEMPT_COUNT) && (!(preempt_count() & PREEMPT_MASK))) {
2448 rcu_preempt_deferred_qs(current);
2449 } else if (rcu_preempt_need_deferred_qs(current)) {
2450 set_tsk_need_resched(current);
2451 set_preempt_need_resched();
2452 }
2453
2454 /* Update RCU state based on any recent quiescent states. */
2455 rcu_check_quiescent_state(rdp);
2456
2457 /* No grace period and unregistered callbacks? */
2458 if (!rcu_gp_in_progress() &&
2459 rcu_segcblist_is_enabled(&rdp->cblist) && do_batch) {
2460 rcu_nocb_lock_irqsave(rdp, flags);
2461 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2462 rcu_accelerate_cbs_unlocked(rnp, rdp);
2463 rcu_nocb_unlock_irqrestore(rdp, flags);
2464 }
2465
2466 rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
2467
2468 /* If there are callbacks ready, invoke them. */
2469 if (do_batch && rcu_segcblist_ready_cbs(&rdp->cblist) &&
2470 likely(READ_ONCE(rcu_scheduler_fully_active))) {
2471 rcu_do_batch(rdp);
2472 /* Re-invoke RCU core processing if there are callbacks remaining. */
2473 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2474 invoke_rcu_core();
2475 }
2476
2477 /* Do any needed deferred wakeups of rcuo kthreads. */
2478 do_nocb_deferred_wakeup(rdp);
2479 trace_rcu_utilization(TPS("End RCU core"));
2480
2481 // If strict GPs, schedule an RCU reader in a clean environment.
2482 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
2483 queue_work_on(rdp->cpu, rcu_gp_wq, &rdp->strict_work);
2484}
2485
2486static void rcu_core_si(struct softirq_action *h)
2487{
2488 rcu_core();
2489}
2490
2491static void rcu_wake_cond(struct task_struct *t, int status)
2492{
2493 /*
2494 * If the thread is yielding, only wake it when this
2495 * is invoked from idle
2496 */
2497 if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
2498 wake_up_process(t);
2499}
2500
2501static void invoke_rcu_core_kthread(void)
2502{
2503 struct task_struct *t;
2504 unsigned long flags;
2505
2506 local_irq_save(flags);
2507 __this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
2508 t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
2509 if (t != NULL && t != current)
2510 rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
2511 local_irq_restore(flags);
2512}
2513
2514/*
2515 * Wake up this CPU's rcuc kthread to do RCU core processing.
2516 */
2517static void invoke_rcu_core(void)
2518{
2519 if (!cpu_online(smp_processor_id()))
2520 return;
2521 if (use_softirq)
2522 raise_softirq(RCU_SOFTIRQ);
2523 else
2524 invoke_rcu_core_kthread();
2525}
2526
2527static void rcu_cpu_kthread_park(unsigned int cpu)
2528{
2529 per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
2530}
2531
2532static int rcu_cpu_kthread_should_run(unsigned int cpu)
2533{
2534 return __this_cpu_read(rcu_data.rcu_cpu_has_work);
2535}
2536
2537/*
2538 * Per-CPU kernel thread that invokes RCU callbacks. This replaces
2539 * the RCU softirq used in configurations of RCU that do not support RCU
2540 * priority boosting.
2541 */
2542static void rcu_cpu_kthread(unsigned int cpu)
2543{
2544 unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
2545 char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
2546 unsigned long *j = this_cpu_ptr(&rcu_data.rcuc_activity);
2547 int spincnt;
2548
2549 trace_rcu_utilization(TPS("Start CPU kthread@rcu_run"));
2550 for (spincnt = 0; spincnt < 10; spincnt++) {
2551 WRITE_ONCE(*j, jiffies);
2552 local_bh_disable();
2553 *statusp = RCU_KTHREAD_RUNNING;
2554 local_irq_disable();
2555 work = *workp;
2556 WRITE_ONCE(*workp, 0);
2557 local_irq_enable();
2558 if (work)
2559 rcu_core();
2560 local_bh_enable();
2561 if (!READ_ONCE(*workp)) {
2562 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2563 *statusp = RCU_KTHREAD_WAITING;
2564 return;
2565 }
2566 }
2567 *statusp = RCU_KTHREAD_YIELDING;
2568 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2569 schedule_timeout_idle(2);
2570 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2571 *statusp = RCU_KTHREAD_WAITING;
2572 WRITE_ONCE(*j, jiffies);
2573}
2574
2575static struct smp_hotplug_thread rcu_cpu_thread_spec = {
2576 .store = &rcu_data.rcu_cpu_kthread_task,
2577 .thread_should_run = rcu_cpu_kthread_should_run,
2578 .thread_fn = rcu_cpu_kthread,
2579 .thread_comm = "rcuc/%u",
2580 .setup = rcu_cpu_kthread_setup,
2581 .park = rcu_cpu_kthread_park,
2582};
2583
2584/*
2585 * Spawn per-CPU RCU core processing kthreads.
2586 */
2587static int __init rcu_spawn_core_kthreads(void)
2588{
2589 int cpu;
2590
2591 for_each_possible_cpu(cpu)
2592 per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
2593 if (use_softirq)
2594 return 0;
2595 WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
2596 "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
2597 return 0;
2598}
2599
2600static void rcutree_enqueue(struct rcu_data *rdp, struct rcu_head *head, rcu_callback_t func)
2601{
2602 rcu_segcblist_enqueue(&rdp->cblist, head);
2603 if (__is_kvfree_rcu_offset((unsigned long)func))
2604 trace_rcu_kvfree_callback(rcu_state.name, head,
2605 (unsigned long)func,
2606 rcu_segcblist_n_cbs(&rdp->cblist));
2607 else
2608 trace_rcu_callback(rcu_state.name, head,
2609 rcu_segcblist_n_cbs(&rdp->cblist));
2610 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCBQueued"));
2611}
2612
2613/*
2614 * Handle any core-RCU processing required by a call_rcu() invocation.
2615 */
2616static void call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2617 rcu_callback_t func, unsigned long flags)
2618{
2619 rcutree_enqueue(rdp, head, func);
2620 /*
2621 * If called from an extended quiescent state, invoke the RCU
2622 * core in order to force a re-evaluation of RCU's idleness.
2623 */
2624 if (!rcu_is_watching())
2625 invoke_rcu_core();
2626
2627 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2628 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2629 return;
2630
2631 /*
2632 * Force the grace period if too many callbacks or too long waiting.
2633 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2634 * if some other CPU has recently done so. Also, don't bother
2635 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2636 * is the only one waiting for a grace period to complete.
2637 */
2638 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2639 rdp->qlen_last_fqs_check + qhimark)) {
2640
2641 /* Are we ignoring a completed grace period? */
2642 note_gp_changes(rdp);
2643
2644 /* Start a new grace period if one not already started. */
2645 if (!rcu_gp_in_progress()) {
2646 rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
2647 } else {
2648 /* Give the grace period a kick. */
2649 rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
2650 if (READ_ONCE(rcu_state.n_force_qs) == rdp->n_force_qs_snap &&
2651 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2652 rcu_force_quiescent_state();
2653 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
2654 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2655 }
2656 }
2657}
2658
2659/*
2660 * RCU callback function to leak a callback.
2661 */
2662static void rcu_leak_callback(struct rcu_head *rhp)
2663{
2664}
2665
2666/*
2667 * Check and if necessary update the leaf rcu_node structure's
2668 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2669 * number of queued RCU callbacks. The caller must hold the leaf rcu_node
2670 * structure's ->lock.
2671 */
2672static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp)
2673{
2674 raw_lockdep_assert_held_rcu_node(rnp);
2675 if (qovld_calc <= 0)
2676 return; // Early boot and wildcard value set.
2677 if (rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc)
2678 WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask | rdp->grpmask);
2679 else
2680 WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask & ~rdp->grpmask);
2681}
2682
2683/*
2684 * Check and if necessary update the leaf rcu_node structure's
2685 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2686 * number of queued RCU callbacks. No locks need be held, but the
2687 * caller must have disabled interrupts.
2688 *
2689 * Note that this function ignores the possibility that there are a lot
2690 * of callbacks all of which have already seen the end of their respective
2691 * grace periods. This omission is due to the need for no-CBs CPUs to
2692 * be holding ->nocb_lock to do this check, which is too heavy for a
2693 * common-case operation.
2694 */
2695static void check_cb_ovld(struct rcu_data *rdp)
2696{
2697 struct rcu_node *const rnp = rdp->mynode;
2698
2699 if (qovld_calc <= 0 ||
2700 ((rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) ==
2701 !!(READ_ONCE(rnp->cbovldmask) & rdp->grpmask)))
2702 return; // Early boot wildcard value or already set correctly.
2703 raw_spin_lock_rcu_node(rnp);
2704 check_cb_ovld_locked(rdp, rnp);
2705 raw_spin_unlock_rcu_node(rnp);
2706}
2707
2708static void
2709__call_rcu_common(struct rcu_head *head, rcu_callback_t func, bool lazy_in)
2710{
2711 static atomic_t doublefrees;
2712 unsigned long flags;
2713 bool lazy;
2714 struct rcu_data *rdp;
2715
2716 /* Misaligned rcu_head! */
2717 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2718
2719 if (debug_rcu_head_queue(head)) {
2720 /*
2721 * Probable double call_rcu(), so leak the callback.
2722 * Use rcu:rcu_callback trace event to find the previous
2723 * time callback was passed to call_rcu().
2724 */
2725 if (atomic_inc_return(&doublefrees) < 4) {
2726 pr_err("%s(): Double-freed CB %p->%pS()!!! ", __func__, head, head->func);
2727 mem_dump_obj(head);
2728 }
2729 WRITE_ONCE(head->func, rcu_leak_callback);
2730 return;
2731 }
2732 head->func = func;
2733 head->next = NULL;
2734 kasan_record_aux_stack_noalloc(head);
2735 local_irq_save(flags);
2736 rdp = this_cpu_ptr(&rcu_data);
2737 lazy = lazy_in && !rcu_async_should_hurry();
2738
2739 /* Add the callback to our list. */
2740 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
2741 // This can trigger due to call_rcu() from offline CPU:
2742 WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
2743 WARN_ON_ONCE(!rcu_is_watching());
2744 // Very early boot, before rcu_init(). Initialize if needed
2745 // and then drop through to queue the callback.
2746 if (rcu_segcblist_empty(&rdp->cblist))
2747 rcu_segcblist_init(&rdp->cblist);
2748 }
2749
2750 check_cb_ovld(rdp);
2751
2752 if (unlikely(rcu_rdp_is_offloaded(rdp)))
2753 call_rcu_nocb(rdp, head, func, flags, lazy);
2754 else
2755 call_rcu_core(rdp, head, func, flags);
2756 local_irq_restore(flags);
2757}
2758
2759#ifdef CONFIG_RCU_LAZY
2760static bool enable_rcu_lazy __read_mostly = !IS_ENABLED(CONFIG_RCU_LAZY_DEFAULT_OFF);
2761module_param(enable_rcu_lazy, bool, 0444);
2762
2763/**
2764 * call_rcu_hurry() - Queue RCU callback for invocation after grace period, and
2765 * flush all lazy callbacks (including the new one) to the main ->cblist while
2766 * doing so.
2767 *
2768 * @head: structure to be used for queueing the RCU updates.
2769 * @func: actual callback function to be invoked after the grace period
2770 *
2771 * The callback function will be invoked some time after a full grace
2772 * period elapses, in other words after all pre-existing RCU read-side
2773 * critical sections have completed.
2774 *
2775 * Use this API instead of call_rcu() if you don't want the callback to be
2776 * invoked after very long periods of time, which can happen on systems without
2777 * memory pressure and on systems which are lightly loaded or mostly idle.
2778 * This function will cause callbacks to be invoked sooner than later at the
2779 * expense of extra power. Other than that, this function is identical to, and
2780 * reuses call_rcu()'s logic. Refer to call_rcu() for more details about memory
2781 * ordering and other functionality.
2782 */
2783void call_rcu_hurry(struct rcu_head *head, rcu_callback_t func)
2784{
2785 __call_rcu_common(head, func, false);
2786}
2787EXPORT_SYMBOL_GPL(call_rcu_hurry);
2788#else
2789#define enable_rcu_lazy false
2790#endif
2791
2792/**
2793 * call_rcu() - Queue an RCU callback for invocation after a grace period.
2794 * By default the callbacks are 'lazy' and are kept hidden from the main
2795 * ->cblist to prevent starting of grace periods too soon.
2796 * If you desire grace periods to start very soon, use call_rcu_hurry().
2797 *
2798 * @head: structure to be used for queueing the RCU updates.
2799 * @func: actual callback function to be invoked after the grace period
2800 *
2801 * The callback function will be invoked some time after a full grace
2802 * period elapses, in other words after all pre-existing RCU read-side
2803 * critical sections have completed. However, the callback function
2804 * might well execute concurrently with RCU read-side critical sections
2805 * that started after call_rcu() was invoked.
2806 *
2807 * RCU read-side critical sections are delimited by rcu_read_lock()
2808 * and rcu_read_unlock(), and may be nested. In addition, but only in
2809 * v5.0 and later, regions of code across which interrupts, preemption,
2810 * or softirqs have been disabled also serve as RCU read-side critical
2811 * sections. This includes hardware interrupt handlers, softirq handlers,
2812 * and NMI handlers.
2813 *
2814 * Note that all CPUs must agree that the grace period extended beyond
2815 * all pre-existing RCU read-side critical section. On systems with more
2816 * than one CPU, this means that when "func()" is invoked, each CPU is
2817 * guaranteed to have executed a full memory barrier since the end of its
2818 * last RCU read-side critical section whose beginning preceded the call
2819 * to call_rcu(). It also means that each CPU executing an RCU read-side
2820 * critical section that continues beyond the start of "func()" must have
2821 * executed a memory barrier after the call_rcu() but before the beginning
2822 * of that RCU read-side critical section. Note that these guarantees
2823 * include CPUs that are offline, idle, or executing in user mode, as
2824 * well as CPUs that are executing in the kernel.
2825 *
2826 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
2827 * resulting RCU callback function "func()", then both CPU A and CPU B are
2828 * guaranteed to execute a full memory barrier during the time interval
2829 * between the call to call_rcu() and the invocation of "func()" -- even
2830 * if CPU A and CPU B are the same CPU (but again only if the system has
2831 * more than one CPU).
2832 *
2833 * Implementation of these memory-ordering guarantees is described here:
2834 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
2835 */
2836void call_rcu(struct rcu_head *head, rcu_callback_t func)
2837{
2838 __call_rcu_common(head, func, enable_rcu_lazy);
2839}
2840EXPORT_SYMBOL_GPL(call_rcu);
2841
2842/* Maximum number of jiffies to wait before draining a batch. */
2843#define KFREE_DRAIN_JIFFIES (5 * HZ)
2844#define KFREE_N_BATCHES 2
2845#define FREE_N_CHANNELS 2
2846
2847/**
2848 * struct kvfree_rcu_bulk_data - single block to store kvfree_rcu() pointers
2849 * @list: List node. All blocks are linked between each other
2850 * @gp_snap: Snapshot of RCU state for objects placed to this bulk
2851 * @nr_records: Number of active pointers in the array
2852 * @records: Array of the kvfree_rcu() pointers
2853 */
2854struct kvfree_rcu_bulk_data {
2855 struct list_head list;
2856 struct rcu_gp_oldstate gp_snap;
2857 unsigned long nr_records;
2858 void *records[];
2859};
2860
2861/*
2862 * This macro defines how many entries the "records" array
2863 * will contain. It is based on the fact that the size of
2864 * kvfree_rcu_bulk_data structure becomes exactly one page.
2865 */
2866#define KVFREE_BULK_MAX_ENTR \
2867 ((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *))
2868
2869/**
2870 * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
2871 * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
2872 * @head_free: List of kfree_rcu() objects waiting for a grace period
2873 * @head_free_gp_snap: Grace-period snapshot to check for attempted premature frees.
2874 * @bulk_head_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
2875 * @krcp: Pointer to @kfree_rcu_cpu structure
2876 */
2877
2878struct kfree_rcu_cpu_work {
2879 struct rcu_work rcu_work;
2880 struct rcu_head *head_free;
2881 struct rcu_gp_oldstate head_free_gp_snap;
2882 struct list_head bulk_head_free[FREE_N_CHANNELS];
2883 struct kfree_rcu_cpu *krcp;
2884};
2885
2886/**
2887 * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
2888 * @head: List of kfree_rcu() objects not yet waiting for a grace period
2889 * @head_gp_snap: Snapshot of RCU state for objects placed to "@head"
2890 * @bulk_head: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period
2891 * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
2892 * @lock: Synchronize access to this structure
2893 * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES
2894 * @initialized: The @rcu_work fields have been initialized
2895 * @head_count: Number of objects in rcu_head singular list
2896 * @bulk_count: Number of objects in bulk-list
2897 * @bkvcache:
2898 * A simple cache list that contains objects for reuse purpose.
2899 * In order to save some per-cpu space the list is singular.
2900 * Even though it is lockless an access has to be protected by the
2901 * per-cpu lock.
2902 * @page_cache_work: A work to refill the cache when it is empty
2903 * @backoff_page_cache_fill: Delay cache refills
2904 * @work_in_progress: Indicates that page_cache_work is running
2905 * @hrtimer: A hrtimer for scheduling a page_cache_work
2906 * @nr_bkv_objs: number of allocated objects at @bkvcache.
2907 *
2908 * This is a per-CPU structure. The reason that it is not included in
2909 * the rcu_data structure is to permit this code to be extracted from
2910 * the RCU files. Such extraction could allow further optimization of
2911 * the interactions with the slab allocators.
2912 */
2913struct kfree_rcu_cpu {
2914 // Objects queued on a linked list
2915 // through their rcu_head structures.
2916 struct rcu_head *head;
2917 unsigned long head_gp_snap;
2918 atomic_t head_count;
2919
2920 // Objects queued on a bulk-list.
2921 struct list_head bulk_head[FREE_N_CHANNELS];
2922 atomic_t bulk_count[FREE_N_CHANNELS];
2923
2924 struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
2925 raw_spinlock_t lock;
2926 struct delayed_work monitor_work;
2927 bool initialized;
2928
2929 struct delayed_work page_cache_work;
2930 atomic_t backoff_page_cache_fill;
2931 atomic_t work_in_progress;
2932 struct hrtimer hrtimer;
2933
2934 struct llist_head bkvcache;
2935 int nr_bkv_objs;
2936};
2937
2938static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc) = {
2939 .lock = __RAW_SPIN_LOCK_UNLOCKED(krc.lock),
2940};
2941
2942static __always_inline void
2943debug_rcu_bhead_unqueue(struct kvfree_rcu_bulk_data *bhead)
2944{
2945#ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
2946 int i;
2947
2948 for (i = 0; i < bhead->nr_records; i++)
2949 debug_rcu_head_unqueue((struct rcu_head *)(bhead->records[i]));
2950#endif
2951}
2952
2953static inline struct kfree_rcu_cpu *
2954krc_this_cpu_lock(unsigned long *flags)
2955{
2956 struct kfree_rcu_cpu *krcp;
2957
2958 local_irq_save(*flags); // For safely calling this_cpu_ptr().
2959 krcp = this_cpu_ptr(&krc);
2960 raw_spin_lock(&krcp->lock);
2961
2962 return krcp;
2963}
2964
2965static inline void
2966krc_this_cpu_unlock(struct kfree_rcu_cpu *krcp, unsigned long flags)
2967{
2968 raw_spin_unlock_irqrestore(&krcp->lock, flags);
2969}
2970
2971static inline struct kvfree_rcu_bulk_data *
2972get_cached_bnode(struct kfree_rcu_cpu *krcp)
2973{
2974 if (!krcp->nr_bkv_objs)
2975 return NULL;
2976
2977 WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs - 1);
2978 return (struct kvfree_rcu_bulk_data *)
2979 llist_del_first(&krcp->bkvcache);
2980}
2981
2982static inline bool
2983put_cached_bnode(struct kfree_rcu_cpu *krcp,
2984 struct kvfree_rcu_bulk_data *bnode)
2985{
2986 // Check the limit.
2987 if (krcp->nr_bkv_objs >= rcu_min_cached_objs)
2988 return false;
2989
2990 llist_add((struct llist_node *) bnode, &krcp->bkvcache);
2991 WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs + 1);
2992 return true;
2993}
2994
2995static int
2996drain_page_cache(struct kfree_rcu_cpu *krcp)
2997{
2998 unsigned long flags;
2999 struct llist_node *page_list, *pos, *n;
3000 int freed = 0;
3001
3002 if (!rcu_min_cached_objs)
3003 return 0;
3004
3005 raw_spin_lock_irqsave(&krcp->lock, flags);
3006 page_list = llist_del_all(&krcp->bkvcache);
3007 WRITE_ONCE(krcp->nr_bkv_objs, 0);
3008 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3009
3010 llist_for_each_safe(pos, n, page_list) {
3011 free_page((unsigned long)pos);
3012 freed++;
3013 }
3014
3015 return freed;
3016}
3017
3018static void
3019kvfree_rcu_bulk(struct kfree_rcu_cpu *krcp,
3020 struct kvfree_rcu_bulk_data *bnode, int idx)
3021{
3022 unsigned long flags;
3023 int i;
3024
3025 if (!WARN_ON_ONCE(!poll_state_synchronize_rcu_full(&bnode->gp_snap))) {
3026 debug_rcu_bhead_unqueue(bnode);
3027 rcu_lock_acquire(&rcu_callback_map);
3028 if (idx == 0) { // kmalloc() / kfree().
3029 trace_rcu_invoke_kfree_bulk_callback(
3030 rcu_state.name, bnode->nr_records,
3031 bnode->records);
3032
3033 kfree_bulk(bnode->nr_records, bnode->records);
3034 } else { // vmalloc() / vfree().
3035 for (i = 0; i < bnode->nr_records; i++) {
3036 trace_rcu_invoke_kvfree_callback(
3037 rcu_state.name, bnode->records[i], 0);
3038
3039 vfree(bnode->records[i]);
3040 }
3041 }
3042 rcu_lock_release(&rcu_callback_map);
3043 }
3044
3045 raw_spin_lock_irqsave(&krcp->lock, flags);
3046 if (put_cached_bnode(krcp, bnode))
3047 bnode = NULL;
3048 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3049
3050 if (bnode)
3051 free_page((unsigned long) bnode);
3052
3053 cond_resched_tasks_rcu_qs();
3054}
3055
3056static void
3057kvfree_rcu_list(struct rcu_head *head)
3058{
3059 struct rcu_head *next;
3060
3061 for (; head; head = next) {
3062 void *ptr = (void *) head->func;
3063 unsigned long offset = (void *) head - ptr;
3064
3065 next = head->next;
3066 debug_rcu_head_unqueue((struct rcu_head *)ptr);
3067 rcu_lock_acquire(&rcu_callback_map);
3068 trace_rcu_invoke_kvfree_callback(rcu_state.name, head, offset);
3069
3070 if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset)))
3071 kvfree(ptr);
3072
3073 rcu_lock_release(&rcu_callback_map);
3074 cond_resched_tasks_rcu_qs();
3075 }
3076}
3077
3078/*
3079 * This function is invoked in workqueue context after a grace period.
3080 * It frees all the objects queued on ->bulk_head_free or ->head_free.
3081 */
3082static void kfree_rcu_work(struct work_struct *work)
3083{
3084 unsigned long flags;
3085 struct kvfree_rcu_bulk_data *bnode, *n;
3086 struct list_head bulk_head[FREE_N_CHANNELS];
3087 struct rcu_head *head;
3088 struct kfree_rcu_cpu *krcp;
3089 struct kfree_rcu_cpu_work *krwp;
3090 struct rcu_gp_oldstate head_gp_snap;
3091 int i;
3092
3093 krwp = container_of(to_rcu_work(work),
3094 struct kfree_rcu_cpu_work, rcu_work);
3095 krcp = krwp->krcp;
3096
3097 raw_spin_lock_irqsave(&krcp->lock, flags);
3098 // Channels 1 and 2.
3099 for (i = 0; i < FREE_N_CHANNELS; i++)
3100 list_replace_init(&krwp->bulk_head_free[i], &bulk_head[i]);
3101
3102 // Channel 3.
3103 head = krwp->head_free;
3104 krwp->head_free = NULL;
3105 head_gp_snap = krwp->head_free_gp_snap;
3106 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3107
3108 // Handle the first two channels.
3109 for (i = 0; i < FREE_N_CHANNELS; i++) {
3110 // Start from the tail page, so a GP is likely passed for it.
3111 list_for_each_entry_safe(bnode, n, &bulk_head[i], list)
3112 kvfree_rcu_bulk(krcp, bnode, i);
3113 }
3114
3115 /*
3116 * This is used when the "bulk" path can not be used for the
3117 * double-argument of kvfree_rcu(). This happens when the
3118 * page-cache is empty, which means that objects are instead
3119 * queued on a linked list through their rcu_head structures.
3120 * This list is named "Channel 3".
3121 */
3122 if (head && !WARN_ON_ONCE(!poll_state_synchronize_rcu_full(&head_gp_snap)))
3123 kvfree_rcu_list(head);
3124}
3125
3126static bool
3127need_offload_krc(struct kfree_rcu_cpu *krcp)
3128{
3129 int i;
3130
3131 for (i = 0; i < FREE_N_CHANNELS; i++)
3132 if (!list_empty(&krcp->bulk_head[i]))
3133 return true;
3134
3135 return !!READ_ONCE(krcp->head);
3136}
3137
3138static bool
3139need_wait_for_krwp_work(struct kfree_rcu_cpu_work *krwp)
3140{
3141 int i;
3142
3143 for (i = 0; i < FREE_N_CHANNELS; i++)
3144 if (!list_empty(&krwp->bulk_head_free[i]))
3145 return true;
3146
3147 return !!krwp->head_free;
3148}
3149
3150static int krc_count(struct kfree_rcu_cpu *krcp)
3151{
3152 int sum = atomic_read(&krcp->head_count);
3153 int i;
3154
3155 for (i = 0; i < FREE_N_CHANNELS; i++)
3156 sum += atomic_read(&krcp->bulk_count[i]);
3157
3158 return sum;
3159}
3160
3161static void
3162schedule_delayed_monitor_work(struct kfree_rcu_cpu *krcp)
3163{
3164 long delay, delay_left;
3165
3166 delay = krc_count(krcp) >= KVFREE_BULK_MAX_ENTR ? 1:KFREE_DRAIN_JIFFIES;
3167 if (delayed_work_pending(&krcp->monitor_work)) {
3168 delay_left = krcp->monitor_work.timer.expires - jiffies;
3169 if (delay < delay_left)
3170 mod_delayed_work(system_wq, &krcp->monitor_work, delay);
3171 return;
3172 }
3173 queue_delayed_work(system_wq, &krcp->monitor_work, delay);
3174}
3175
3176static void
3177kvfree_rcu_drain_ready(struct kfree_rcu_cpu *krcp)
3178{
3179 struct list_head bulk_ready[FREE_N_CHANNELS];
3180 struct kvfree_rcu_bulk_data *bnode, *n;
3181 struct rcu_head *head_ready = NULL;
3182 unsigned long flags;
3183 int i;
3184
3185 raw_spin_lock_irqsave(&krcp->lock, flags);
3186 for (i = 0; i < FREE_N_CHANNELS; i++) {
3187 INIT_LIST_HEAD(&bulk_ready[i]);
3188
3189 list_for_each_entry_safe_reverse(bnode, n, &krcp->bulk_head[i], list) {
3190 if (!poll_state_synchronize_rcu_full(&bnode->gp_snap))
3191 break;
3192
3193 atomic_sub(bnode->nr_records, &krcp->bulk_count[i]);
3194 list_move(&bnode->list, &bulk_ready[i]);
3195 }
3196 }
3197
3198 if (krcp->head && poll_state_synchronize_rcu(krcp->head_gp_snap)) {
3199 head_ready = krcp->head;
3200 atomic_set(&krcp->head_count, 0);
3201 WRITE_ONCE(krcp->head, NULL);
3202 }
3203 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3204
3205 for (i = 0; i < FREE_N_CHANNELS; i++) {
3206 list_for_each_entry_safe(bnode, n, &bulk_ready[i], list)
3207 kvfree_rcu_bulk(krcp, bnode, i);
3208 }
3209
3210 if (head_ready)
3211 kvfree_rcu_list(head_ready);
3212}
3213
3214/*
3215 * This function is invoked after the KFREE_DRAIN_JIFFIES timeout.
3216 */
3217static void kfree_rcu_monitor(struct work_struct *work)
3218{
3219 struct kfree_rcu_cpu *krcp = container_of(work,
3220 struct kfree_rcu_cpu, monitor_work.work);
3221 unsigned long flags;
3222 int i, j;
3223
3224 // Drain ready for reclaim.
3225 kvfree_rcu_drain_ready(krcp);
3226
3227 raw_spin_lock_irqsave(&krcp->lock, flags);
3228
3229 // Attempt to start a new batch.
3230 for (i = 0; i < KFREE_N_BATCHES; i++) {
3231 struct kfree_rcu_cpu_work *krwp = &(krcp->krw_arr[i]);
3232
3233 // Try to detach bulk_head or head and attach it, only when
3234 // all channels are free. Any channel is not free means at krwp
3235 // there is on-going rcu work to handle krwp's free business.
3236 if (need_wait_for_krwp_work(krwp))
3237 continue;
3238
3239 // kvfree_rcu_drain_ready() might handle this krcp, if so give up.
3240 if (need_offload_krc(krcp)) {
3241 // Channel 1 corresponds to the SLAB-pointer bulk path.
3242 // Channel 2 corresponds to vmalloc-pointer bulk path.
3243 for (j = 0; j < FREE_N_CHANNELS; j++) {
3244 if (list_empty(&krwp->bulk_head_free[j])) {
3245 atomic_set(&krcp->bulk_count[j], 0);
3246 list_replace_init(&krcp->bulk_head[j],
3247 &krwp->bulk_head_free[j]);
3248 }
3249 }
3250
3251 // Channel 3 corresponds to both SLAB and vmalloc
3252 // objects queued on the linked list.
3253 if (!krwp->head_free) {
3254 krwp->head_free = krcp->head;
3255 get_state_synchronize_rcu_full(&krwp->head_free_gp_snap);
3256 atomic_set(&krcp->head_count, 0);
3257 WRITE_ONCE(krcp->head, NULL);
3258 }
3259
3260 // One work is per one batch, so there are three
3261 // "free channels", the batch can handle. It can
3262 // be that the work is in the pending state when
3263 // channels have been detached following by each
3264 // other.
3265 queue_rcu_work(system_wq, &krwp->rcu_work);
3266 }
3267 }
3268
3269 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3270
3271 // If there is nothing to detach, it means that our job is
3272 // successfully done here. In case of having at least one
3273 // of the channels that is still busy we should rearm the
3274 // work to repeat an attempt. Because previous batches are
3275 // still in progress.
3276 if (need_offload_krc(krcp))
3277 schedule_delayed_monitor_work(krcp);
3278}
3279
3280static enum hrtimer_restart
3281schedule_page_work_fn(struct hrtimer *t)
3282{
3283 struct kfree_rcu_cpu *krcp =
3284 container_of(t, struct kfree_rcu_cpu, hrtimer);
3285
3286 queue_delayed_work(system_highpri_wq, &krcp->page_cache_work, 0);
3287 return HRTIMER_NORESTART;
3288}
3289
3290static void fill_page_cache_func(struct work_struct *work)
3291{
3292 struct kvfree_rcu_bulk_data *bnode;
3293 struct kfree_rcu_cpu *krcp =
3294 container_of(work, struct kfree_rcu_cpu,
3295 page_cache_work.work);
3296 unsigned long flags;
3297 int nr_pages;
3298 bool pushed;
3299 int i;
3300
3301 nr_pages = atomic_read(&krcp->backoff_page_cache_fill) ?
3302 1 : rcu_min_cached_objs;
3303
3304 for (i = READ_ONCE(krcp->nr_bkv_objs); i < nr_pages; i++) {
3305 bnode = (struct kvfree_rcu_bulk_data *)
3306 __get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3307
3308 if (!bnode)
3309 break;
3310
3311 raw_spin_lock_irqsave(&krcp->lock, flags);
3312 pushed = put_cached_bnode(krcp, bnode);
3313 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3314
3315 if (!pushed) {
3316 free_page((unsigned long) bnode);
3317 break;
3318 }
3319 }
3320
3321 atomic_set(&krcp->work_in_progress, 0);
3322 atomic_set(&krcp->backoff_page_cache_fill, 0);
3323}
3324
3325static void
3326run_page_cache_worker(struct kfree_rcu_cpu *krcp)
3327{
3328 // If cache disabled, bail out.
3329 if (!rcu_min_cached_objs)
3330 return;
3331
3332 if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
3333 !atomic_xchg(&krcp->work_in_progress, 1)) {
3334 if (atomic_read(&krcp->backoff_page_cache_fill)) {
3335 queue_delayed_work(system_wq,
3336 &krcp->page_cache_work,
3337 msecs_to_jiffies(rcu_delay_page_cache_fill_msec));
3338 } else {
3339 hrtimer_init(&krcp->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3340 krcp->hrtimer.function = schedule_page_work_fn;
3341 hrtimer_start(&krcp->hrtimer, 0, HRTIMER_MODE_REL);
3342 }
3343 }
3344}
3345
3346// Record ptr in a page managed by krcp, with the pre-krc_this_cpu_lock()
3347// state specified by flags. If can_alloc is true, the caller must
3348// be schedulable and not be holding any locks or mutexes that might be
3349// acquired by the memory allocator or anything that it might invoke.
3350// Returns true if ptr was successfully recorded, else the caller must
3351// use a fallback.
3352static inline bool
3353add_ptr_to_bulk_krc_lock(struct kfree_rcu_cpu **krcp,
3354 unsigned long *flags, void *ptr, bool can_alloc)
3355{
3356 struct kvfree_rcu_bulk_data *bnode;
3357 int idx;
3358
3359 *krcp = krc_this_cpu_lock(flags);
3360 if (unlikely(!(*krcp)->initialized))
3361 return false;
3362
3363 idx = !!is_vmalloc_addr(ptr);
3364 bnode = list_first_entry_or_null(&(*krcp)->bulk_head[idx],
3365 struct kvfree_rcu_bulk_data, list);
3366
3367 /* Check if a new block is required. */
3368 if (!bnode || bnode->nr_records == KVFREE_BULK_MAX_ENTR) {
3369 bnode = get_cached_bnode(*krcp);
3370 if (!bnode && can_alloc) {
3371 krc_this_cpu_unlock(*krcp, *flags);
3372
3373 // __GFP_NORETRY - allows a light-weight direct reclaim
3374 // what is OK from minimizing of fallback hitting point of
3375 // view. Apart of that it forbids any OOM invoking what is
3376 // also beneficial since we are about to release memory soon.
3377 //
3378 // __GFP_NOMEMALLOC - prevents from consuming of all the
3379 // memory reserves. Please note we have a fallback path.
3380 //
3381 // __GFP_NOWARN - it is supposed that an allocation can
3382 // be failed under low memory or high memory pressure
3383 // scenarios.
3384 bnode = (struct kvfree_rcu_bulk_data *)
3385 __get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3386 raw_spin_lock_irqsave(&(*krcp)->lock, *flags);
3387 }
3388
3389 if (!bnode)
3390 return false;
3391
3392 // Initialize the new block and attach it.
3393 bnode->nr_records = 0;
3394 list_add(&bnode->list, &(*krcp)->bulk_head[idx]);
3395 }
3396
3397 // Finally insert and update the GP for this page.
3398 bnode->records[bnode->nr_records++] = ptr;
3399 get_state_synchronize_rcu_full(&bnode->gp_snap);
3400 atomic_inc(&(*krcp)->bulk_count[idx]);
3401
3402 return true;
3403}
3404
3405/*
3406 * Queue a request for lazy invocation of the appropriate free routine
3407 * after a grace period. Please note that three paths are maintained,
3408 * two for the common case using arrays of pointers and a third one that
3409 * is used only when the main paths cannot be used, for example, due to
3410 * memory pressure.
3411 *
3412 * Each kvfree_call_rcu() request is added to a batch. The batch will be drained
3413 * every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch will
3414 * be free'd in workqueue context. This allows us to: batch requests together to
3415 * reduce the number of grace periods during heavy kfree_rcu()/kvfree_rcu() load.
3416 */
3417void kvfree_call_rcu(struct rcu_head *head, void *ptr)
3418{
3419 unsigned long flags;
3420 struct kfree_rcu_cpu *krcp;
3421 bool success;
3422
3423 /*
3424 * Please note there is a limitation for the head-less
3425 * variant, that is why there is a clear rule for such
3426 * objects: it can be used from might_sleep() context
3427 * only. For other places please embed an rcu_head to
3428 * your data.
3429 */
3430 if (!head)
3431 might_sleep();
3432
3433 // Queue the object but don't yet schedule the batch.
3434 if (debug_rcu_head_queue(ptr)) {
3435 // Probable double kfree_rcu(), just leak.
3436 WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n",
3437 __func__, head);
3438
3439 // Mark as success and leave.
3440 return;
3441 }
3442
3443 kasan_record_aux_stack_noalloc(ptr);
3444 success = add_ptr_to_bulk_krc_lock(&krcp, &flags, ptr, !head);
3445 if (!success) {
3446 run_page_cache_worker(krcp);
3447
3448 if (head == NULL)
3449 // Inline if kvfree_rcu(one_arg) call.
3450 goto unlock_return;
3451
3452 head->func = ptr;
3453 head->next = krcp->head;
3454 WRITE_ONCE(krcp->head, head);
3455 atomic_inc(&krcp->head_count);
3456
3457 // Take a snapshot for this krcp.
3458 krcp->head_gp_snap = get_state_synchronize_rcu();
3459 success = true;
3460 }
3461
3462 /*
3463 * The kvfree_rcu() caller considers the pointer freed at this point
3464 * and likely removes any references to it. Since the actual slab
3465 * freeing (and kmemleak_free()) is deferred, tell kmemleak to ignore
3466 * this object (no scanning or false positives reporting).
3467 */
3468 kmemleak_ignore(ptr);
3469
3470 // Set timer to drain after KFREE_DRAIN_JIFFIES.
3471 if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING)
3472 schedule_delayed_monitor_work(krcp);
3473
3474unlock_return:
3475 krc_this_cpu_unlock(krcp, flags);
3476
3477 /*
3478 * Inline kvfree() after synchronize_rcu(). We can do
3479 * it from might_sleep() context only, so the current
3480 * CPU can pass the QS state.
3481 */
3482 if (!success) {
3483 debug_rcu_head_unqueue((struct rcu_head *) ptr);
3484 synchronize_rcu();
3485 kvfree(ptr);
3486 }
3487}
3488EXPORT_SYMBOL_GPL(kvfree_call_rcu);
3489
3490static unsigned long
3491kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
3492{
3493 int cpu;
3494 unsigned long count = 0;
3495
3496 /* Snapshot count of all CPUs */
3497 for_each_possible_cpu(cpu) {
3498 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3499
3500 count += krc_count(krcp);
3501 count += READ_ONCE(krcp->nr_bkv_objs);
3502 atomic_set(&krcp->backoff_page_cache_fill, 1);
3503 }
3504
3505 return count == 0 ? SHRINK_EMPTY : count;
3506}
3507
3508static unsigned long
3509kfree_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
3510{
3511 int cpu, freed = 0;
3512
3513 for_each_possible_cpu(cpu) {
3514 int count;
3515 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3516
3517 count = krc_count(krcp);
3518 count += drain_page_cache(krcp);
3519 kfree_rcu_monitor(&krcp->monitor_work.work);
3520
3521 sc->nr_to_scan -= count;
3522 freed += count;
3523
3524 if (sc->nr_to_scan <= 0)
3525 break;
3526 }
3527
3528 return freed == 0 ? SHRINK_STOP : freed;
3529}
3530
3531void __init kfree_rcu_scheduler_running(void)
3532{
3533 int cpu;
3534
3535 for_each_possible_cpu(cpu) {
3536 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3537
3538 if (need_offload_krc(krcp))
3539 schedule_delayed_monitor_work(krcp);
3540 }
3541}
3542
3543/*
3544 * During early boot, any blocking grace-period wait automatically
3545 * implies a grace period.
3546 *
3547 * Later on, this could in theory be the case for kernels built with
3548 * CONFIG_SMP=y && CONFIG_PREEMPTION=y running on a single CPU, but this
3549 * is not a common case. Furthermore, this optimization would cause
3550 * the rcu_gp_oldstate structure to expand by 50%, so this potential
3551 * grace-period optimization is ignored once the scheduler is running.
3552 */
3553static int rcu_blocking_is_gp(void)
3554{
3555 if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE) {
3556 might_sleep();
3557 return false;
3558 }
3559 return true;
3560}
3561
3562/**
3563 * synchronize_rcu - wait until a grace period has elapsed.
3564 *
3565 * Control will return to the caller some time after a full grace
3566 * period has elapsed, in other words after all currently executing RCU
3567 * read-side critical sections have completed. Note, however, that
3568 * upon return from synchronize_rcu(), the caller might well be executing
3569 * concurrently with new RCU read-side critical sections that began while
3570 * synchronize_rcu() was waiting.
3571 *
3572 * RCU read-side critical sections are delimited by rcu_read_lock()
3573 * and rcu_read_unlock(), and may be nested. In addition, but only in
3574 * v5.0 and later, regions of code across which interrupts, preemption,
3575 * or softirqs have been disabled also serve as RCU read-side critical
3576 * sections. This includes hardware interrupt handlers, softirq handlers,
3577 * and NMI handlers.
3578 *
3579 * Note that this guarantee implies further memory-ordering guarantees.
3580 * On systems with more than one CPU, when synchronize_rcu() returns,
3581 * each CPU is guaranteed to have executed a full memory barrier since
3582 * the end of its last RCU read-side critical section whose beginning
3583 * preceded the call to synchronize_rcu(). In addition, each CPU having
3584 * an RCU read-side critical section that extends beyond the return from
3585 * synchronize_rcu() is guaranteed to have executed a full memory barrier
3586 * after the beginning of synchronize_rcu() and before the beginning of
3587 * that RCU read-side critical section. Note that these guarantees include
3588 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3589 * that are executing in the kernel.
3590 *
3591 * Furthermore, if CPU A invoked synchronize_rcu(), which returned
3592 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3593 * to have executed a full memory barrier during the execution of
3594 * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
3595 * again only if the system has more than one CPU).
3596 *
3597 * Implementation of these memory-ordering guarantees is described here:
3598 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
3599 */
3600void synchronize_rcu(void)
3601{
3602 unsigned long flags;
3603 struct rcu_node *rnp;
3604
3605 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3606 lock_is_held(&rcu_lock_map) ||
3607 lock_is_held(&rcu_sched_lock_map),
3608 "Illegal synchronize_rcu() in RCU read-side critical section");
3609 if (!rcu_blocking_is_gp()) {
3610 if (rcu_gp_is_expedited())
3611 synchronize_rcu_expedited();
3612 else
3613 wait_rcu_gp(call_rcu_hurry);
3614 return;
3615 }
3616
3617 // Context allows vacuous grace periods.
3618 // Note well that this code runs with !PREEMPT && !SMP.
3619 // In addition, all code that advances grace periods runs at
3620 // process level. Therefore, this normal GP overlaps with other
3621 // normal GPs only by being fully nested within them, which allows
3622 // reuse of ->gp_seq_polled_snap.
3623 rcu_poll_gp_seq_start_unlocked(&rcu_state.gp_seq_polled_snap);
3624 rcu_poll_gp_seq_end_unlocked(&rcu_state.gp_seq_polled_snap);
3625
3626 // Update the normal grace-period counters to record
3627 // this grace period, but only those used by the boot CPU.
3628 // The rcu_scheduler_starting() will take care of the rest of
3629 // these counters.
3630 local_irq_save(flags);
3631 WARN_ON_ONCE(num_online_cpus() > 1);
3632 rcu_state.gp_seq += (1 << RCU_SEQ_CTR_SHIFT);
3633 for (rnp = this_cpu_ptr(&rcu_data)->mynode; rnp; rnp = rnp->parent)
3634 rnp->gp_seq_needed = rnp->gp_seq = rcu_state.gp_seq;
3635 local_irq_restore(flags);
3636}
3637EXPORT_SYMBOL_GPL(synchronize_rcu);
3638
3639/**
3640 * get_completed_synchronize_rcu_full - Return a full pre-completed polled state cookie
3641 * @rgosp: Place to put state cookie
3642 *
3643 * Stores into @rgosp a value that will always be treated by functions
3644 * like poll_state_synchronize_rcu_full() as a cookie whose grace period
3645 * has already completed.
3646 */
3647void get_completed_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3648{
3649 rgosp->rgos_norm = RCU_GET_STATE_COMPLETED;
3650 rgosp->rgos_exp = RCU_GET_STATE_COMPLETED;
3651}
3652EXPORT_SYMBOL_GPL(get_completed_synchronize_rcu_full);
3653
3654/**
3655 * get_state_synchronize_rcu - Snapshot current RCU state
3656 *
3657 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3658 * or poll_state_synchronize_rcu() to determine whether or not a full
3659 * grace period has elapsed in the meantime.
3660 */
3661unsigned long get_state_synchronize_rcu(void)
3662{
3663 /*
3664 * Any prior manipulation of RCU-protected data must happen
3665 * before the load from ->gp_seq.
3666 */
3667 smp_mb(); /* ^^^ */
3668 return rcu_seq_snap(&rcu_state.gp_seq_polled);
3669}
3670EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3671
3672/**
3673 * get_state_synchronize_rcu_full - Snapshot RCU state, both normal and expedited
3674 * @rgosp: location to place combined normal/expedited grace-period state
3675 *
3676 * Places the normal and expedited grace-period states in @rgosp. This
3677 * state value can be passed to a later call to cond_synchronize_rcu_full()
3678 * or poll_state_synchronize_rcu_full() to determine whether or not a
3679 * grace period (whether normal or expedited) has elapsed in the meantime.
3680 * The rcu_gp_oldstate structure takes up twice the memory of an unsigned
3681 * long, but is guaranteed to see all grace periods. In contrast, the
3682 * combined state occupies less memory, but can sometimes fail to take
3683 * grace periods into account.
3684 *
3685 * This does not guarantee that the needed grace period will actually
3686 * start.
3687 */
3688void get_state_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3689{
3690 struct rcu_node *rnp = rcu_get_root();
3691
3692 /*
3693 * Any prior manipulation of RCU-protected data must happen
3694 * before the loads from ->gp_seq and ->expedited_sequence.
3695 */
3696 smp_mb(); /* ^^^ */
3697 rgosp->rgos_norm = rcu_seq_snap(&rnp->gp_seq);
3698 rgosp->rgos_exp = rcu_seq_snap(&rcu_state.expedited_sequence);
3699}
3700EXPORT_SYMBOL_GPL(get_state_synchronize_rcu_full);
3701
3702/*
3703 * Helper function for start_poll_synchronize_rcu() and
3704 * start_poll_synchronize_rcu_full().
3705 */
3706static void start_poll_synchronize_rcu_common(void)
3707{
3708 unsigned long flags;
3709 bool needwake;
3710 struct rcu_data *rdp;
3711 struct rcu_node *rnp;
3712
3713 lockdep_assert_irqs_enabled();
3714 local_irq_save(flags);
3715 rdp = this_cpu_ptr(&rcu_data);
3716 rnp = rdp->mynode;
3717 raw_spin_lock_rcu_node(rnp); // irqs already disabled.
3718 // Note it is possible for a grace period to have elapsed between
3719 // the above call to get_state_synchronize_rcu() and the below call
3720 // to rcu_seq_snap. This is OK, the worst that happens is that we
3721 // get a grace period that no one needed. These accesses are ordered
3722 // by smp_mb(), and we are accessing them in the opposite order
3723 // from which they are updated at grace-period start, as required.
3724 needwake = rcu_start_this_gp(rnp, rdp, rcu_seq_snap(&rcu_state.gp_seq));
3725 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3726 if (needwake)
3727 rcu_gp_kthread_wake();
3728}
3729
3730/**
3731 * start_poll_synchronize_rcu - Snapshot and start RCU grace period
3732 *
3733 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3734 * or poll_state_synchronize_rcu() to determine whether or not a full
3735 * grace period has elapsed in the meantime. If the needed grace period
3736 * is not already slated to start, notifies RCU core of the need for that
3737 * grace period.
3738 *
3739 * Interrupts must be enabled for the case where it is necessary to awaken
3740 * the grace-period kthread.
3741 */
3742unsigned long start_poll_synchronize_rcu(void)
3743{
3744 unsigned long gp_seq = get_state_synchronize_rcu();
3745
3746 start_poll_synchronize_rcu_common();
3747 return gp_seq;
3748}
3749EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu);
3750
3751/**
3752 * start_poll_synchronize_rcu_full - Take a full snapshot and start RCU grace period
3753 * @rgosp: value from get_state_synchronize_rcu_full() or start_poll_synchronize_rcu_full()
3754 *
3755 * Places the normal and expedited grace-period states in *@rgos. This
3756 * state value can be passed to a later call to cond_synchronize_rcu_full()
3757 * or poll_state_synchronize_rcu_full() to determine whether or not a
3758 * grace period (whether normal or expedited) has elapsed in the meantime.
3759 * If the needed grace period is not already slated to start, notifies
3760 * RCU core of the need for that grace period.
3761 *
3762 * Interrupts must be enabled for the case where it is necessary to awaken
3763 * the grace-period kthread.
3764 */
3765void start_poll_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3766{
3767 get_state_synchronize_rcu_full(rgosp);
3768
3769 start_poll_synchronize_rcu_common();
3770}
3771EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_full);
3772
3773/**
3774 * poll_state_synchronize_rcu - Has the specified RCU grace period completed?
3775 * @oldstate: value from get_state_synchronize_rcu() or start_poll_synchronize_rcu()
3776 *
3777 * If a full RCU grace period has elapsed since the earlier call from
3778 * which @oldstate was obtained, return @true, otherwise return @false.
3779 * If @false is returned, it is the caller's responsibility to invoke this
3780 * function later on until it does return @true. Alternatively, the caller
3781 * can explicitly wait for a grace period, for example, by passing @oldstate
3782 * to either cond_synchronize_rcu() or cond_synchronize_rcu_expedited()
3783 * on the one hand or by directly invoking either synchronize_rcu() or
3784 * synchronize_rcu_expedited() on the other.
3785 *
3786 * Yes, this function does not take counter wrap into account.
3787 * But counter wrap is harmless. If the counter wraps, we have waited for
3788 * more than a billion grace periods (and way more on a 64-bit system!).
3789 * Those needing to keep old state values for very long time periods
3790 * (many hours even on 32-bit systems) should check them occasionally and
3791 * either refresh them or set a flag indicating that the grace period has
3792 * completed. Alternatively, they can use get_completed_synchronize_rcu()
3793 * to get a guaranteed-completed grace-period state.
3794 *
3795 * In addition, because oldstate compresses the grace-period state for
3796 * both normal and expedited grace periods into a single unsigned long,
3797 * it can miss a grace period when synchronize_rcu() runs concurrently
3798 * with synchronize_rcu_expedited(). If this is unacceptable, please
3799 * instead use the _full() variant of these polling APIs.
3800 *
3801 * This function provides the same memory-ordering guarantees that
3802 * would be provided by a synchronize_rcu() that was invoked at the call
3803 * to the function that provided @oldstate, and that returned at the end
3804 * of this function.
3805 */
3806bool poll_state_synchronize_rcu(unsigned long oldstate)
3807{
3808 if (oldstate == RCU_GET_STATE_COMPLETED ||
3809 rcu_seq_done_exact(&rcu_state.gp_seq_polled, oldstate)) {
3810 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3811 return true;
3812 }
3813 return false;
3814}
3815EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu);
3816
3817/**
3818 * poll_state_synchronize_rcu_full - Has the specified RCU grace period completed?
3819 * @rgosp: value from get_state_synchronize_rcu_full() or start_poll_synchronize_rcu_full()
3820 *
3821 * If a full RCU grace period has elapsed since the earlier call from
3822 * which *rgosp was obtained, return @true, otherwise return @false.
3823 * If @false is returned, it is the caller's responsibility to invoke this
3824 * function later on until it does return @true. Alternatively, the caller
3825 * can explicitly wait for a grace period, for example, by passing @rgosp
3826 * to cond_synchronize_rcu() or by directly invoking synchronize_rcu().
3827 *
3828 * Yes, this function does not take counter wrap into account.
3829 * But counter wrap is harmless. If the counter wraps, we have waited
3830 * for more than a billion grace periods (and way more on a 64-bit
3831 * system!). Those needing to keep rcu_gp_oldstate values for very
3832 * long time periods (many hours even on 32-bit systems) should check
3833 * them occasionally and either refresh them or set a flag indicating
3834 * that the grace period has completed. Alternatively, they can use
3835 * get_completed_synchronize_rcu_full() to get a guaranteed-completed
3836 * grace-period state.
3837 *
3838 * This function provides the same memory-ordering guarantees that would
3839 * be provided by a synchronize_rcu() that was invoked at the call to
3840 * the function that provided @rgosp, and that returned at the end of this
3841 * function. And this guarantee requires that the root rcu_node structure's
3842 * ->gp_seq field be checked instead of that of the rcu_state structure.
3843 * The problem is that the just-ending grace-period's callbacks can be
3844 * invoked between the time that the root rcu_node structure's ->gp_seq
3845 * field is updated and the time that the rcu_state structure's ->gp_seq
3846 * field is updated. Therefore, if a single synchronize_rcu() is to
3847 * cause a subsequent poll_state_synchronize_rcu_full() to return @true,
3848 * then the root rcu_node structure is the one that needs to be polled.
3849 */
3850bool poll_state_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3851{
3852 struct rcu_node *rnp = rcu_get_root();
3853
3854 smp_mb(); // Order against root rcu_node structure grace-period cleanup.
3855 if (rgosp->rgos_norm == RCU_GET_STATE_COMPLETED ||
3856 rcu_seq_done_exact(&rnp->gp_seq, rgosp->rgos_norm) ||
3857 rgosp->rgos_exp == RCU_GET_STATE_COMPLETED ||
3858 rcu_seq_done_exact(&rcu_state.expedited_sequence, rgosp->rgos_exp)) {
3859 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3860 return true;
3861 }
3862 return false;
3863}
3864EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu_full);
3865
3866/**
3867 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3868 * @oldstate: value from get_state_synchronize_rcu(), start_poll_synchronize_rcu(), or start_poll_synchronize_rcu_expedited()
3869 *
3870 * If a full RCU grace period has elapsed since the earlier call to
3871 * get_state_synchronize_rcu() or start_poll_synchronize_rcu(), just return.
3872 * Otherwise, invoke synchronize_rcu() to wait for a full grace period.
3873 *
3874 * Yes, this function does not take counter wrap into account.
3875 * But counter wrap is harmless. If the counter wraps, we have waited for
3876 * more than 2 billion grace periods (and way more on a 64-bit system!),
3877 * so waiting for a couple of additional grace periods should be just fine.
3878 *
3879 * This function provides the same memory-ordering guarantees that
3880 * would be provided by a synchronize_rcu() that was invoked at the call
3881 * to the function that provided @oldstate and that returned at the end
3882 * of this function.
3883 */
3884void cond_synchronize_rcu(unsigned long oldstate)
3885{
3886 if (!poll_state_synchronize_rcu(oldstate))
3887 synchronize_rcu();
3888}
3889EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3890
3891/**
3892 * cond_synchronize_rcu_full - Conditionally wait for an RCU grace period
3893 * @rgosp: value from get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(), or start_poll_synchronize_rcu_expedited_full()
3894 *
3895 * If a full RCU grace period has elapsed since the call to
3896 * get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(),
3897 * or start_poll_synchronize_rcu_expedited_full() from which @rgosp was
3898 * obtained, just return. Otherwise, invoke synchronize_rcu() to wait
3899 * for a full grace period.
3900 *
3901 * Yes, this function does not take counter wrap into account.
3902 * But counter wrap is harmless. If the counter wraps, we have waited for
3903 * more than 2 billion grace periods (and way more on a 64-bit system!),
3904 * so waiting for a couple of additional grace periods should be just fine.
3905 *
3906 * This function provides the same memory-ordering guarantees that
3907 * would be provided by a synchronize_rcu() that was invoked at the call
3908 * to the function that provided @rgosp and that returned at the end of
3909 * this function.
3910 */
3911void cond_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3912{
3913 if (!poll_state_synchronize_rcu_full(rgosp))
3914 synchronize_rcu();
3915}
3916EXPORT_SYMBOL_GPL(cond_synchronize_rcu_full);
3917
3918/*
3919 * Check to see if there is any immediate RCU-related work to be done by
3920 * the current CPU, returning 1 if so and zero otherwise. The checks are
3921 * in order of increasing expense: checks that can be carried out against
3922 * CPU-local state are performed first. However, we must check for CPU
3923 * stalls first, else we might not get a chance.
3924 */
3925static int rcu_pending(int user)
3926{
3927 bool gp_in_progress;
3928 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
3929 struct rcu_node *rnp = rdp->mynode;
3930
3931 lockdep_assert_irqs_disabled();
3932
3933 /* Check for CPU stalls, if enabled. */
3934 check_cpu_stall(rdp);
3935
3936 /* Does this CPU need a deferred NOCB wakeup? */
3937 if (rcu_nocb_need_deferred_wakeup(rdp, RCU_NOCB_WAKE))
3938 return 1;
3939
3940 /* Is this a nohz_full CPU in userspace or idle? (Ignore RCU if so.) */
3941 if ((user || rcu_is_cpu_rrupt_from_idle()) && rcu_nohz_full_cpu())
3942 return 0;
3943
3944 /* Is the RCU core waiting for a quiescent state from this CPU? */
3945 gp_in_progress = rcu_gp_in_progress();
3946 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm && gp_in_progress)
3947 return 1;
3948
3949 /* Does this CPU have callbacks ready to invoke? */
3950 if (!rcu_rdp_is_offloaded(rdp) &&
3951 rcu_segcblist_ready_cbs(&rdp->cblist))
3952 return 1;
3953
3954 /* Has RCU gone idle with this CPU needing another grace period? */
3955 if (!gp_in_progress && rcu_segcblist_is_enabled(&rdp->cblist) &&
3956 !rcu_rdp_is_offloaded(rdp) &&
3957 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
3958 return 1;
3959
3960 /* Have RCU grace period completed or started? */
3961 if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
3962 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
3963 return 1;
3964
3965 /* nothing to do */
3966 return 0;
3967}
3968
3969/*
3970 * Helper function for rcu_barrier() tracing. If tracing is disabled,
3971 * the compiler is expected to optimize this away.
3972 */
3973static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
3974{
3975 trace_rcu_barrier(rcu_state.name, s, cpu,
3976 atomic_read(&rcu_state.barrier_cpu_count), done);
3977}
3978
3979/*
3980 * RCU callback function for rcu_barrier(). If we are last, wake
3981 * up the task executing rcu_barrier().
3982 *
3983 * Note that the value of rcu_state.barrier_sequence must be captured
3984 * before the atomic_dec_and_test(). Otherwise, if this CPU is not last,
3985 * other CPUs might count the value down to zero before this CPU gets
3986 * around to invoking rcu_barrier_trace(), which might result in bogus
3987 * data from the next instance of rcu_barrier().
3988 */
3989static void rcu_barrier_callback(struct rcu_head *rhp)
3990{
3991 unsigned long __maybe_unused s = rcu_state.barrier_sequence;
3992
3993 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
3994 rcu_barrier_trace(TPS("LastCB"), -1, s);
3995 complete(&rcu_state.barrier_completion);
3996 } else {
3997 rcu_barrier_trace(TPS("CB"), -1, s);
3998 }
3999}
4000
4001/*
4002 * If needed, entrain an rcu_barrier() callback on rdp->cblist.
4003 */
4004static void rcu_barrier_entrain(struct rcu_data *rdp)
4005{
4006 unsigned long gseq = READ_ONCE(rcu_state.barrier_sequence);
4007 unsigned long lseq = READ_ONCE(rdp->barrier_seq_snap);
4008 bool wake_nocb = false;
4009 bool was_alldone = false;
4010
4011 lockdep_assert_held(&rcu_state.barrier_lock);
4012 if (rcu_seq_state(lseq) || !rcu_seq_state(gseq) || rcu_seq_ctr(lseq) != rcu_seq_ctr(gseq))
4013 return;
4014 rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
4015 rdp->barrier_head.func = rcu_barrier_callback;
4016 debug_rcu_head_queue(&rdp->barrier_head);
4017 rcu_nocb_lock(rdp);
4018 /*
4019 * Flush bypass and wakeup rcuog if we add callbacks to an empty regular
4020 * queue. This way we don't wait for bypass timer that can reach seconds
4021 * if it's fully lazy.
4022 */
4023 was_alldone = rcu_rdp_is_offloaded(rdp) && !rcu_segcblist_pend_cbs(&rdp->cblist);
4024 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies, false));
4025 wake_nocb = was_alldone && rcu_segcblist_pend_cbs(&rdp->cblist);
4026 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head)) {
4027 atomic_inc(&rcu_state.barrier_cpu_count);
4028 } else {
4029 debug_rcu_head_unqueue(&rdp->barrier_head);
4030 rcu_barrier_trace(TPS("IRQNQ"), -1, rcu_state.barrier_sequence);
4031 }
4032 rcu_nocb_unlock(rdp);
4033 if (wake_nocb)
4034 wake_nocb_gp(rdp, false);
4035 smp_store_release(&rdp->barrier_seq_snap, gseq);
4036}
4037
4038/*
4039 * Called with preemption disabled, and from cross-cpu IRQ context.
4040 */
4041static void rcu_barrier_handler(void *cpu_in)
4042{
4043 uintptr_t cpu = (uintptr_t)cpu_in;
4044 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4045
4046 lockdep_assert_irqs_disabled();
4047 WARN_ON_ONCE(cpu != rdp->cpu);
4048 WARN_ON_ONCE(cpu != smp_processor_id());
4049 raw_spin_lock(&rcu_state.barrier_lock);
4050 rcu_barrier_entrain(rdp);
4051 raw_spin_unlock(&rcu_state.barrier_lock);
4052}
4053
4054/**
4055 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
4056 *
4057 * Note that this primitive does not necessarily wait for an RCU grace period
4058 * to complete. For example, if there are no RCU callbacks queued anywhere
4059 * in the system, then rcu_barrier() is within its rights to return
4060 * immediately, without waiting for anything, much less an RCU grace period.
4061 */
4062void rcu_barrier(void)
4063{
4064 uintptr_t cpu;
4065 unsigned long flags;
4066 unsigned long gseq;
4067 struct rcu_data *rdp;
4068 unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
4069
4070 rcu_barrier_trace(TPS("Begin"), -1, s);
4071
4072 /* Take mutex to serialize concurrent rcu_barrier() requests. */
4073 mutex_lock(&rcu_state.barrier_mutex);
4074
4075 /* Did someone else do our work for us? */
4076 if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
4077 rcu_barrier_trace(TPS("EarlyExit"), -1, rcu_state.barrier_sequence);
4078 smp_mb(); /* caller's subsequent code after above check. */
4079 mutex_unlock(&rcu_state.barrier_mutex);
4080 return;
4081 }
4082
4083 /* Mark the start of the barrier operation. */
4084 raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
4085 rcu_seq_start(&rcu_state.barrier_sequence);
4086 gseq = rcu_state.barrier_sequence;
4087 rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
4088
4089 /*
4090 * Initialize the count to two rather than to zero in order
4091 * to avoid a too-soon return to zero in case of an immediate
4092 * invocation of the just-enqueued callback (or preemption of
4093 * this task). Exclude CPU-hotplug operations to ensure that no
4094 * offline non-offloaded CPU has callbacks queued.
4095 */
4096 init_completion(&rcu_state.barrier_completion);
4097 atomic_set(&rcu_state.barrier_cpu_count, 2);
4098 raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4099
4100 /*
4101 * Force each CPU with callbacks to register a new callback.
4102 * When that callback is invoked, we will know that all of the
4103 * corresponding CPU's preceding callbacks have been invoked.
4104 */
4105 for_each_possible_cpu(cpu) {
4106 rdp = per_cpu_ptr(&rcu_data, cpu);
4107retry:
4108 if (smp_load_acquire(&rdp->barrier_seq_snap) == gseq)
4109 continue;
4110 raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
4111 if (!rcu_segcblist_n_cbs(&rdp->cblist)) {
4112 WRITE_ONCE(rdp->barrier_seq_snap, gseq);
4113 raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4114 rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence);
4115 continue;
4116 }
4117 if (!rcu_rdp_cpu_online(rdp)) {
4118 rcu_barrier_entrain(rdp);
4119 WARN_ON_ONCE(READ_ONCE(rdp->barrier_seq_snap) != gseq);
4120 raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4121 rcu_barrier_trace(TPS("OfflineNoCBQ"), cpu, rcu_state.barrier_sequence);
4122 continue;
4123 }
4124 raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4125 if (smp_call_function_single(cpu, rcu_barrier_handler, (void *)cpu, 1)) {
4126 schedule_timeout_uninterruptible(1);
4127 goto retry;
4128 }
4129 WARN_ON_ONCE(READ_ONCE(rdp->barrier_seq_snap) != gseq);
4130 rcu_barrier_trace(TPS("OnlineQ"), cpu, rcu_state.barrier_sequence);
4131 }
4132
4133 /*
4134 * Now that we have an rcu_barrier_callback() callback on each
4135 * CPU, and thus each counted, remove the initial count.
4136 */
4137 if (atomic_sub_and_test(2, &rcu_state.barrier_cpu_count))
4138 complete(&rcu_state.barrier_completion);
4139
4140 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
4141 wait_for_completion(&rcu_state.barrier_completion);
4142
4143 /* Mark the end of the barrier operation. */
4144 rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
4145 rcu_seq_end(&rcu_state.barrier_sequence);
4146 gseq = rcu_state.barrier_sequence;
4147 for_each_possible_cpu(cpu) {
4148 rdp = per_cpu_ptr(&rcu_data, cpu);
4149
4150 WRITE_ONCE(rdp->barrier_seq_snap, gseq);
4151 }
4152
4153 /* Other rcu_barrier() invocations can now safely proceed. */
4154 mutex_unlock(&rcu_state.barrier_mutex);
4155}
4156EXPORT_SYMBOL_GPL(rcu_barrier);
4157
4158static unsigned long rcu_barrier_last_throttle;
4159
4160/**
4161 * rcu_barrier_throttled - Do rcu_barrier(), but limit to one per second
4162 *
4163 * This can be thought of as guard rails around rcu_barrier() that
4164 * permits unrestricted userspace use, at least assuming the hardware's
4165 * try_cmpxchg() is robust. There will be at most one call per second to
4166 * rcu_barrier() system-wide from use of this function, which means that
4167 * callers might needlessly wait a second or three.
4168 *
4169 * This is intended for use by test suites to avoid OOM by flushing RCU
4170 * callbacks from the previous test before starting the next. See the
4171 * rcutree.do_rcu_barrier module parameter for more information.
4172 *
4173 * Why not simply make rcu_barrier() more scalable? That might be
4174 * the eventual endpoint, but let's keep it simple for the time being.
4175 * Note that the module parameter infrastructure serializes calls to a
4176 * given .set() function, but should concurrent .set() invocation ever be
4177 * possible, we are ready!
4178 */
4179static void rcu_barrier_throttled(void)
4180{
4181 unsigned long j = jiffies;
4182 unsigned long old = READ_ONCE(rcu_barrier_last_throttle);
4183 unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
4184
4185 while (time_in_range(j, old, old + HZ / 16) ||
4186 !try_cmpxchg(&rcu_barrier_last_throttle, &old, j)) {
4187 schedule_timeout_idle(HZ / 16);
4188 if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
4189 smp_mb(); /* caller's subsequent code after above check. */
4190 return;
4191 }
4192 j = jiffies;
4193 old = READ_ONCE(rcu_barrier_last_throttle);
4194 }
4195 rcu_barrier();
4196}
4197
4198/*
4199 * Invoke rcu_barrier_throttled() when a rcutree.do_rcu_barrier
4200 * request arrives. We insist on a true value to allow for possible
4201 * future expansion.
4202 */
4203static int param_set_do_rcu_barrier(const char *val, const struct kernel_param *kp)
4204{
4205 bool b;
4206 int ret;
4207
4208 if (rcu_scheduler_active != RCU_SCHEDULER_RUNNING)
4209 return -EAGAIN;
4210 ret = kstrtobool(val, &b);
4211 if (!ret && b) {
4212 atomic_inc((atomic_t *)kp->arg);
4213 rcu_barrier_throttled();
4214 atomic_dec((atomic_t *)kp->arg);
4215 }
4216 return ret;
4217}
4218
4219/*
4220 * Output the number of outstanding rcutree.do_rcu_barrier requests.
4221 */
4222static int param_get_do_rcu_barrier(char *buffer, const struct kernel_param *kp)
4223{
4224 return sprintf(buffer, "%d\n", atomic_read((atomic_t *)kp->arg));
4225}
4226
4227static const struct kernel_param_ops do_rcu_barrier_ops = {
4228 .set = param_set_do_rcu_barrier,
4229 .get = param_get_do_rcu_barrier,
4230};
4231static atomic_t do_rcu_barrier;
4232module_param_cb(do_rcu_barrier, &do_rcu_barrier_ops, &do_rcu_barrier, 0644);
4233
4234/*
4235 * Compute the mask of online CPUs for the specified rcu_node structure.
4236 * This will not be stable unless the rcu_node structure's ->lock is
4237 * held, but the bit corresponding to the current CPU will be stable
4238 * in most contexts.
4239 */
4240static unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
4241{
4242 return READ_ONCE(rnp->qsmaskinitnext);
4243}
4244
4245/*
4246 * Is the CPU corresponding to the specified rcu_data structure online
4247 * from RCU's perspective? This perspective is given by that structure's
4248 * ->qsmaskinitnext field rather than by the global cpu_online_mask.
4249 */
4250static bool rcu_rdp_cpu_online(struct rcu_data *rdp)
4251{
4252 return !!(rdp->grpmask & rcu_rnp_online_cpus(rdp->mynode));
4253}
4254
4255bool rcu_cpu_online(int cpu)
4256{
4257 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4258
4259 return rcu_rdp_cpu_online(rdp);
4260}
4261
4262#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
4263
4264/*
4265 * Is the current CPU online as far as RCU is concerned?
4266 *
4267 * Disable preemption to avoid false positives that could otherwise
4268 * happen due to the current CPU number being sampled, this task being
4269 * preempted, its old CPU being taken offline, resuming on some other CPU,
4270 * then determining that its old CPU is now offline.
4271 *
4272 * Disable checking if in an NMI handler because we cannot safely
4273 * report errors from NMI handlers anyway. In addition, it is OK to use
4274 * RCU on an offline processor during initial boot, hence the check for
4275 * rcu_scheduler_fully_active.
4276 */
4277bool rcu_lockdep_current_cpu_online(void)
4278{
4279 struct rcu_data *rdp;
4280 bool ret = false;
4281
4282 if (in_nmi() || !rcu_scheduler_fully_active)
4283 return true;
4284 preempt_disable_notrace();
4285 rdp = this_cpu_ptr(&rcu_data);
4286 /*
4287 * Strictly, we care here about the case where the current CPU is
4288 * in rcutree_report_cpu_starting() and thus has an excuse for rdp->grpmask
4289 * not being up to date. So arch_spin_is_locked() might have a
4290 * false positive if it's held by some *other* CPU, but that's
4291 * OK because that just means a false *negative* on the warning.
4292 */
4293 if (rcu_rdp_cpu_online(rdp) || arch_spin_is_locked(&rcu_state.ofl_lock))
4294 ret = true;
4295 preempt_enable_notrace();
4296 return ret;
4297}
4298EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
4299
4300#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
4301
4302// Has rcu_init() been invoked? This is used (for example) to determine
4303// whether spinlocks may be acquired safely.
4304static bool rcu_init_invoked(void)
4305{
4306 return !!rcu_state.n_online_cpus;
4307}
4308
4309/*
4310 * All CPUs for the specified rcu_node structure have gone offline,
4311 * and all tasks that were preempted within an RCU read-side critical
4312 * section while running on one of those CPUs have since exited their RCU
4313 * read-side critical section. Some other CPU is reporting this fact with
4314 * the specified rcu_node structure's ->lock held and interrupts disabled.
4315 * This function therefore goes up the tree of rcu_node structures,
4316 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
4317 * the leaf rcu_node structure's ->qsmaskinit field has already been
4318 * updated.
4319 *
4320 * This function does check that the specified rcu_node structure has
4321 * all CPUs offline and no blocked tasks, so it is OK to invoke it
4322 * prematurely. That said, invoking it after the fact will cost you
4323 * a needless lock acquisition. So once it has done its work, don't
4324 * invoke it again.
4325 */
4326static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
4327{
4328 long mask;
4329 struct rcu_node *rnp = rnp_leaf;
4330
4331 raw_lockdep_assert_held_rcu_node(rnp_leaf);
4332 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
4333 WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
4334 WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
4335 return;
4336 for (;;) {
4337 mask = rnp->grpmask;
4338 rnp = rnp->parent;
4339 if (!rnp)
4340 break;
4341 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
4342 rnp->qsmaskinit &= ~mask;
4343 /* Between grace periods, so better already be zero! */
4344 WARN_ON_ONCE(rnp->qsmask);
4345 if (rnp->qsmaskinit) {
4346 raw_spin_unlock_rcu_node(rnp);
4347 /* irqs remain disabled. */
4348 return;
4349 }
4350 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
4351 }
4352}
4353
4354/*
4355 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
4356 * first CPU in a given leaf rcu_node structure coming online. The caller
4357 * must hold the corresponding leaf rcu_node ->lock with interrupts
4358 * disabled.
4359 */
4360static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
4361{
4362 long mask;
4363 long oldmask;
4364 struct rcu_node *rnp = rnp_leaf;
4365
4366 raw_lockdep_assert_held_rcu_node(rnp_leaf);
4367 WARN_ON_ONCE(rnp->wait_blkd_tasks);
4368 for (;;) {
4369 mask = rnp->grpmask;
4370 rnp = rnp->parent;
4371 if (rnp == NULL)
4372 return;
4373 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
4374 oldmask = rnp->qsmaskinit;
4375 rnp->qsmaskinit |= mask;
4376 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
4377 if (oldmask)
4378 return;
4379 }
4380}
4381
4382/*
4383 * Do boot-time initialization of a CPU's per-CPU RCU data.
4384 */
4385static void __init
4386rcu_boot_init_percpu_data(int cpu)
4387{
4388 struct context_tracking *ct = this_cpu_ptr(&context_tracking);
4389 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4390
4391 /* Set up local state, ensuring consistent view of global state. */
4392 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
4393 INIT_WORK(&rdp->strict_work, strict_work_handler);
4394 WARN_ON_ONCE(ct->dynticks_nesting != 1);
4395 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(cpu)));
4396 rdp->barrier_seq_snap = rcu_state.barrier_sequence;
4397 rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
4398 rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
4399 rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
4400 rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
4401 rdp->last_sched_clock = jiffies;
4402 rdp->cpu = cpu;
4403 rcu_boot_init_nocb_percpu_data(rdp);
4404}
4405
4406struct kthread_worker *rcu_exp_gp_kworker;
4407
4408static void rcu_spawn_exp_par_gp_kworker(struct rcu_node *rnp)
4409{
4410 struct kthread_worker *kworker;
4411 const char *name = "rcu_exp_par_gp_kthread_worker/%d";
4412 struct sched_param param = { .sched_priority = kthread_prio };
4413 int rnp_index = rnp - rcu_get_root();
4414
4415 if (rnp->exp_kworker)
4416 return;
4417
4418 kworker = kthread_create_worker(0, name, rnp_index);
4419 if (IS_ERR_OR_NULL(kworker)) {
4420 pr_err("Failed to create par gp kworker on %d/%d\n",
4421 rnp->grplo, rnp->grphi);
4422 return;
4423 }
4424 WRITE_ONCE(rnp->exp_kworker, kworker);
4425
4426 if (IS_ENABLED(CONFIG_RCU_EXP_KTHREAD))
4427 sched_setscheduler_nocheck(kworker->task, SCHED_FIFO, ¶m);
4428}
4429
4430static struct task_struct *rcu_exp_par_gp_task(struct rcu_node *rnp)
4431{
4432 struct kthread_worker *kworker = READ_ONCE(rnp->exp_kworker);
4433
4434 if (!kworker)
4435 return NULL;
4436
4437 return kworker->task;
4438}
4439
4440static void __init rcu_start_exp_gp_kworker(void)
4441{
4442 const char *name = "rcu_exp_gp_kthread_worker";
4443 struct sched_param param = { .sched_priority = kthread_prio };
4444
4445 rcu_exp_gp_kworker = kthread_create_worker(0, name);
4446 if (IS_ERR_OR_NULL(rcu_exp_gp_kworker)) {
4447 pr_err("Failed to create %s!\n", name);
4448 rcu_exp_gp_kworker = NULL;
4449 return;
4450 }
4451
4452 if (IS_ENABLED(CONFIG_RCU_EXP_KTHREAD))
4453 sched_setscheduler_nocheck(rcu_exp_gp_kworker->task, SCHED_FIFO, ¶m);
4454}
4455
4456static void rcu_spawn_rnp_kthreads(struct rcu_node *rnp)
4457{
4458 if (rcu_scheduler_fully_active) {
4459 mutex_lock(&rnp->kthread_mutex);
4460 rcu_spawn_one_boost_kthread(rnp);
4461 rcu_spawn_exp_par_gp_kworker(rnp);
4462 mutex_unlock(&rnp->kthread_mutex);
4463 }
4464}
4465
4466/*
4467 * Invoked early in the CPU-online process, when pretty much all services
4468 * are available. The incoming CPU is not present.
4469 *
4470 * Initializes a CPU's per-CPU RCU data. Note that only one online or
4471 * offline event can be happening at a given time. Note also that we can
4472 * accept some slop in the rsp->gp_seq access due to the fact that this
4473 * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
4474 * And any offloaded callbacks are being numbered elsewhere.
4475 */
4476int rcutree_prepare_cpu(unsigned int cpu)
4477{
4478 unsigned long flags;
4479 struct context_tracking *ct = per_cpu_ptr(&context_tracking, cpu);
4480 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4481 struct rcu_node *rnp = rcu_get_root();
4482
4483 /* Set up local state, ensuring consistent view of global state. */
4484 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4485 rdp->qlen_last_fqs_check = 0;
4486 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
4487 rdp->blimit = blimit;
4488 ct->dynticks_nesting = 1; /* CPU not up, no tearing. */
4489 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
4490
4491 /*
4492 * Only non-NOCB CPUs that didn't have early-boot callbacks need to be
4493 * (re-)initialized.
4494 */
4495 if (!rcu_segcblist_is_enabled(&rdp->cblist))
4496 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
4497
4498 /*
4499 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
4500 * propagation up the rcu_node tree will happen at the beginning
4501 * of the next grace period.
4502 */
4503 rnp = rdp->mynode;
4504 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
4505 rdp->gp_seq = READ_ONCE(rnp->gp_seq);
4506 rdp->gp_seq_needed = rdp->gp_seq;
4507 rdp->cpu_no_qs.b.norm = true;
4508 rdp->core_needs_qs = false;
4509 rdp->rcu_iw_pending = false;
4510 rdp->rcu_iw = IRQ_WORK_INIT_HARD(rcu_iw_handler);
4511 rdp->rcu_iw_gp_seq = rdp->gp_seq - 1;
4512 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
4513 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4514 rcu_spawn_rnp_kthreads(rnp);
4515 rcu_spawn_cpu_nocb_kthread(cpu);
4516 WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus + 1);
4517
4518 return 0;
4519}
4520
4521/*
4522 * Update kthreads affinity during CPU-hotplug changes.
4523 *
4524 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
4525 * served by the rcu_node in question. The CPU hotplug lock is still
4526 * held, so the value of rnp->qsmaskinit will be stable.
4527 *
4528 * We don't include outgoingcpu in the affinity set, use -1 if there is
4529 * no outgoing CPU. If there are no CPUs left in the affinity set,
4530 * this function allows the kthread to execute on any CPU.
4531 *
4532 * Any future concurrent calls are serialized via ->kthread_mutex.
4533 */
4534static void rcutree_affinity_setting(unsigned int cpu, int outgoingcpu)
4535{
4536 cpumask_var_t cm;
4537 unsigned long mask;
4538 struct rcu_data *rdp;
4539 struct rcu_node *rnp;
4540 struct task_struct *task_boost, *task_exp;
4541
4542 rdp = per_cpu_ptr(&rcu_data, cpu);
4543 rnp = rdp->mynode;
4544
4545 task_boost = rcu_boost_task(rnp);
4546 task_exp = rcu_exp_par_gp_task(rnp);
4547
4548 /*
4549 * If CPU is the boot one, those tasks are created later from early
4550 * initcall since kthreadd must be created first.
4551 */
4552 if (!task_boost && !task_exp)
4553 return;
4554
4555 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
4556 return;
4557
4558 mutex_lock(&rnp->kthread_mutex);
4559 mask = rcu_rnp_online_cpus(rnp);
4560 for_each_leaf_node_possible_cpu(rnp, cpu)
4561 if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
4562 cpu != outgoingcpu)
4563 cpumask_set_cpu(cpu, cm);
4564 cpumask_and(cm, cm, housekeeping_cpumask(HK_TYPE_RCU));
4565 if (cpumask_empty(cm)) {
4566 cpumask_copy(cm, housekeeping_cpumask(HK_TYPE_RCU));
4567 if (outgoingcpu >= 0)
4568 cpumask_clear_cpu(outgoingcpu, cm);
4569 }
4570
4571 if (task_exp)
4572 set_cpus_allowed_ptr(task_exp, cm);
4573
4574 if (task_boost)
4575 set_cpus_allowed_ptr(task_boost, cm);
4576
4577 mutex_unlock(&rnp->kthread_mutex);
4578
4579 free_cpumask_var(cm);
4580}
4581
4582/*
4583 * Has the specified (known valid) CPU ever been fully online?
4584 */
4585bool rcu_cpu_beenfullyonline(int cpu)
4586{
4587 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4588
4589 return smp_load_acquire(&rdp->beenonline);
4590}
4591
4592/*
4593 * Near the end of the CPU-online process. Pretty much all services
4594 * enabled, and the CPU is now very much alive.
4595 */
4596int rcutree_online_cpu(unsigned int cpu)
4597{
4598 unsigned long flags;
4599 struct rcu_data *rdp;
4600 struct rcu_node *rnp;
4601
4602 rdp = per_cpu_ptr(&rcu_data, cpu);
4603 rnp = rdp->mynode;
4604 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4605 rnp->ffmask |= rdp->grpmask;
4606 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4607 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
4608 return 0; /* Too early in boot for scheduler work. */
4609 sync_sched_exp_online_cleanup(cpu);
4610 rcutree_affinity_setting(cpu, -1);
4611
4612 // Stop-machine done, so allow nohz_full to disable tick.
4613 tick_dep_clear(TICK_DEP_BIT_RCU);
4614 return 0;
4615}
4616
4617/*
4618 * Mark the specified CPU as being online so that subsequent grace periods
4619 * (both expedited and normal) will wait on it. Note that this means that
4620 * incoming CPUs are not allowed to use RCU read-side critical sections
4621 * until this function is called. Failing to observe this restriction
4622 * will result in lockdep splats.
4623 *
4624 * Note that this function is special in that it is invoked directly
4625 * from the incoming CPU rather than from the cpuhp_step mechanism.
4626 * This is because this function must be invoked at a precise location.
4627 * This incoming CPU must not have enabled interrupts yet.
4628 *
4629 * This mirrors the effects of rcutree_report_cpu_dead().
4630 */
4631void rcutree_report_cpu_starting(unsigned int cpu)
4632{
4633 unsigned long mask;
4634 struct rcu_data *rdp;
4635 struct rcu_node *rnp;
4636 bool newcpu;
4637
4638 lockdep_assert_irqs_disabled();
4639 rdp = per_cpu_ptr(&rcu_data, cpu);
4640 if (rdp->cpu_started)
4641 return;
4642 rdp->cpu_started = true;
4643
4644 rnp = rdp->mynode;
4645 mask = rdp->grpmask;
4646 arch_spin_lock(&rcu_state.ofl_lock);
4647 rcu_dynticks_eqs_online();
4648 raw_spin_lock(&rcu_state.barrier_lock);
4649 raw_spin_lock_rcu_node(rnp);
4650 WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask);
4651 raw_spin_unlock(&rcu_state.barrier_lock);
4652 newcpu = !(rnp->expmaskinitnext & mask);
4653 rnp->expmaskinitnext |= mask;
4654 /* Allow lockless access for expedited grace periods. */
4655 smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + newcpu); /* ^^^ */
4656 ASSERT_EXCLUSIVE_WRITER(rcu_state.ncpus);
4657 rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
4658 rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4659 rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4660
4661 /* An incoming CPU should never be blocking a grace period. */
4662 if (WARN_ON_ONCE(rnp->qsmask & mask)) { /* RCU waiting on incoming CPU? */
4663 /* rcu_report_qs_rnp() *really* wants some flags to restore */
4664 unsigned long flags;
4665
4666 local_irq_save(flags);
4667 rcu_disable_urgency_upon_qs(rdp);
4668 /* Report QS -after- changing ->qsmaskinitnext! */
4669 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4670 } else {
4671 raw_spin_unlock_rcu_node(rnp);
4672 }
4673 arch_spin_unlock(&rcu_state.ofl_lock);
4674 smp_store_release(&rdp->beenonline, true);
4675 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
4676}
4677
4678/*
4679 * The outgoing function has no further need of RCU, so remove it from
4680 * the rcu_node tree's ->qsmaskinitnext bit masks.
4681 *
4682 * Note that this function is special in that it is invoked directly
4683 * from the outgoing CPU rather than from the cpuhp_step mechanism.
4684 * This is because this function must be invoked at a precise location.
4685 *
4686 * This mirrors the effect of rcutree_report_cpu_starting().
4687 */
4688void rcutree_report_cpu_dead(void)
4689{
4690 unsigned long flags;
4691 unsigned long mask;
4692 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
4693 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
4694
4695 /*
4696 * IRQS must be disabled from now on and until the CPU dies, or an interrupt
4697 * may introduce a new READ-side while it is actually off the QS masks.
4698 */
4699 lockdep_assert_irqs_disabled();
4700 // Do any dangling deferred wakeups.
4701 do_nocb_deferred_wakeup(rdp);
4702
4703 rcu_preempt_deferred_qs(current);
4704
4705 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
4706 mask = rdp->grpmask;
4707 arch_spin_lock(&rcu_state.ofl_lock);
4708 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
4709 rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4710 rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4711 if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
4712 /* Report quiescent state -before- changing ->qsmaskinitnext! */
4713 rcu_disable_urgency_upon_qs(rdp);
4714 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4715 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4716 }
4717 WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask);
4718 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4719 arch_spin_unlock(&rcu_state.ofl_lock);
4720 rdp->cpu_started = false;
4721}
4722
4723#ifdef CONFIG_HOTPLUG_CPU
4724/*
4725 * The outgoing CPU has just passed through the dying-idle state, and we
4726 * are being invoked from the CPU that was IPIed to continue the offline
4727 * operation. Migrate the outgoing CPU's callbacks to the current CPU.
4728 */
4729void rcutree_migrate_callbacks(int cpu)
4730{
4731 unsigned long flags;
4732 struct rcu_data *my_rdp;
4733 struct rcu_node *my_rnp;
4734 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4735 bool needwake;
4736
4737 if (rcu_rdp_is_offloaded(rdp) ||
4738 rcu_segcblist_empty(&rdp->cblist))
4739 return; /* No callbacks to migrate. */
4740
4741 raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
4742 WARN_ON_ONCE(rcu_rdp_cpu_online(rdp));
4743 rcu_barrier_entrain(rdp);
4744 my_rdp = this_cpu_ptr(&rcu_data);
4745 my_rnp = my_rdp->mynode;
4746 rcu_nocb_lock(my_rdp); /* irqs already disabled. */
4747 WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies, false));
4748 raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
4749 /* Leverage recent GPs and set GP for new callbacks. */
4750 needwake = rcu_advance_cbs(my_rnp, rdp) ||
4751 rcu_advance_cbs(my_rnp, my_rdp);
4752 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
4753 raw_spin_unlock(&rcu_state.barrier_lock); /* irqs remain disabled. */
4754 needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
4755 rcu_segcblist_disable(&rdp->cblist);
4756 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) != !rcu_segcblist_n_cbs(&my_rdp->cblist));
4757 check_cb_ovld_locked(my_rdp, my_rnp);
4758 if (rcu_rdp_is_offloaded(my_rdp)) {
4759 raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
4760 __call_rcu_nocb_wake(my_rdp, true, flags);
4761 } else {
4762 rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
4763 raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
4764 }
4765 local_irq_restore(flags);
4766 if (needwake)
4767 rcu_gp_kthread_wake();
4768 lockdep_assert_irqs_enabled();
4769 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
4770 !rcu_segcblist_empty(&rdp->cblist),
4771 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
4772 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
4773 rcu_segcblist_first_cb(&rdp->cblist));
4774}
4775
4776/*
4777 * The CPU has been completely removed, and some other CPU is reporting
4778 * this fact from process context. Do the remainder of the cleanup.
4779 * There can only be one CPU hotplug operation at a time, so no need for
4780 * explicit locking.
4781 */
4782int rcutree_dead_cpu(unsigned int cpu)
4783{
4784 WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus - 1);
4785 // Stop-machine done, so allow nohz_full to disable tick.
4786 tick_dep_clear(TICK_DEP_BIT_RCU);
4787 return 0;
4788}
4789
4790/*
4791 * Near the end of the offline process. Trace the fact that this CPU
4792 * is going offline.
4793 */
4794int rcutree_dying_cpu(unsigned int cpu)
4795{
4796 bool blkd;
4797 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4798 struct rcu_node *rnp = rdp->mynode;
4799
4800 blkd = !!(READ_ONCE(rnp->qsmask) & rdp->grpmask);
4801 trace_rcu_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
4802 blkd ? TPS("cpuofl-bgp") : TPS("cpuofl"));
4803 return 0;
4804}
4805
4806/*
4807 * Near the beginning of the process. The CPU is still very much alive
4808 * with pretty much all services enabled.
4809 */
4810int rcutree_offline_cpu(unsigned int cpu)
4811{
4812 unsigned long flags;
4813 struct rcu_data *rdp;
4814 struct rcu_node *rnp;
4815
4816 rdp = per_cpu_ptr(&rcu_data, cpu);
4817 rnp = rdp->mynode;
4818 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4819 rnp->ffmask &= ~rdp->grpmask;
4820 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4821
4822 rcutree_affinity_setting(cpu, cpu);
4823
4824 // nohz_full CPUs need the tick for stop-machine to work quickly
4825 tick_dep_set(TICK_DEP_BIT_RCU);
4826 return 0;
4827}
4828#endif /* #ifdef CONFIG_HOTPLUG_CPU */
4829
4830/*
4831 * On non-huge systems, use expedited RCU grace periods to make suspend
4832 * and hibernation run faster.
4833 */
4834static int rcu_pm_notify(struct notifier_block *self,
4835 unsigned long action, void *hcpu)
4836{
4837 switch (action) {
4838 case PM_HIBERNATION_PREPARE:
4839 case PM_SUSPEND_PREPARE:
4840 rcu_async_hurry();
4841 rcu_expedite_gp();
4842 break;
4843 case PM_POST_HIBERNATION:
4844 case PM_POST_SUSPEND:
4845 rcu_unexpedite_gp();
4846 rcu_async_relax();
4847 break;
4848 default:
4849 break;
4850 }
4851 return NOTIFY_OK;
4852}
4853
4854/*
4855 * Spawn the kthreads that handle RCU's grace periods.
4856 */
4857static int __init rcu_spawn_gp_kthread(void)
4858{
4859 unsigned long flags;
4860 struct rcu_node *rnp;
4861 struct sched_param sp;
4862 struct task_struct *t;
4863 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
4864
4865 rcu_scheduler_fully_active = 1;
4866 t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
4867 if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
4868 return 0;
4869 if (kthread_prio) {
4870 sp.sched_priority = kthread_prio;
4871 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4872 }
4873 rnp = rcu_get_root();
4874 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4875 WRITE_ONCE(rcu_state.gp_activity, jiffies);
4876 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
4877 // Reset .gp_activity and .gp_req_activity before setting .gp_kthread.
4878 smp_store_release(&rcu_state.gp_kthread, t); /* ^^^ */
4879 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4880 wake_up_process(t);
4881 /* This is a pre-SMP initcall, we expect a single CPU */
4882 WARN_ON(num_online_cpus() > 1);
4883 /*
4884 * Those kthreads couldn't be created on rcu_init() -> rcutree_prepare_cpu()
4885 * due to rcu_scheduler_fully_active.
4886 */
4887 rcu_spawn_cpu_nocb_kthread(smp_processor_id());
4888 rcu_spawn_rnp_kthreads(rdp->mynode);
4889 rcu_spawn_core_kthreads();
4890 /* Create kthread worker for expedited GPs */
4891 rcu_start_exp_gp_kworker();
4892 return 0;
4893}
4894early_initcall(rcu_spawn_gp_kthread);
4895
4896/*
4897 * This function is invoked towards the end of the scheduler's
4898 * initialization process. Before this is called, the idle task might
4899 * contain synchronous grace-period primitives (during which time, this idle
4900 * task is booting the system, and such primitives are no-ops). After this
4901 * function is called, any synchronous grace-period primitives are run as
4902 * expedited, with the requesting task driving the grace period forward.
4903 * A later core_initcall() rcu_set_runtime_mode() will switch to full
4904 * runtime RCU functionality.
4905 */
4906void rcu_scheduler_starting(void)
4907{
4908 unsigned long flags;
4909 struct rcu_node *rnp;
4910
4911 WARN_ON(num_online_cpus() != 1);
4912 WARN_ON(nr_context_switches() > 0);
4913 rcu_test_sync_prims();
4914
4915 // Fix up the ->gp_seq counters.
4916 local_irq_save(flags);
4917 rcu_for_each_node_breadth_first(rnp)
4918 rnp->gp_seq_needed = rnp->gp_seq = rcu_state.gp_seq;
4919 local_irq_restore(flags);
4920
4921 // Switch out of early boot mode.
4922 rcu_scheduler_active = RCU_SCHEDULER_INIT;
4923 rcu_test_sync_prims();
4924}
4925
4926/*
4927 * Helper function for rcu_init() that initializes the rcu_state structure.
4928 */
4929static void __init rcu_init_one(void)
4930{
4931 static const char * const buf[] = RCU_NODE_NAME_INIT;
4932 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4933 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4934 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4935
4936 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4937 int cpustride = 1;
4938 int i;
4939 int j;
4940 struct rcu_node *rnp;
4941
4942 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4943
4944 /* Silence gcc 4.8 false positive about array index out of range. */
4945 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4946 panic("rcu_init_one: rcu_num_lvls out of range");
4947
4948 /* Initialize the level-tracking arrays. */
4949
4950 for (i = 1; i < rcu_num_lvls; i++)
4951 rcu_state.level[i] =
4952 rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
4953 rcu_init_levelspread(levelspread, num_rcu_lvl);
4954
4955 /* Initialize the elements themselves, starting from the leaves. */
4956
4957 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4958 cpustride *= levelspread[i];
4959 rnp = rcu_state.level[i];
4960 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
4961 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4962 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4963 &rcu_node_class[i], buf[i]);
4964 raw_spin_lock_init(&rnp->fqslock);
4965 lockdep_set_class_and_name(&rnp->fqslock,
4966 &rcu_fqs_class[i], fqs[i]);
4967 rnp->gp_seq = rcu_state.gp_seq;
4968 rnp->gp_seq_needed = rcu_state.gp_seq;
4969 rnp->completedqs = rcu_state.gp_seq;
4970 rnp->qsmask = 0;
4971 rnp->qsmaskinit = 0;
4972 rnp->grplo = j * cpustride;
4973 rnp->grphi = (j + 1) * cpustride - 1;
4974 if (rnp->grphi >= nr_cpu_ids)
4975 rnp->grphi = nr_cpu_ids - 1;
4976 if (i == 0) {
4977 rnp->grpnum = 0;
4978 rnp->grpmask = 0;
4979 rnp->parent = NULL;
4980 } else {
4981 rnp->grpnum = j % levelspread[i - 1];
4982 rnp->grpmask = BIT(rnp->grpnum);
4983 rnp->parent = rcu_state.level[i - 1] +
4984 j / levelspread[i - 1];
4985 }
4986 rnp->level = i;
4987 INIT_LIST_HEAD(&rnp->blkd_tasks);
4988 rcu_init_one_nocb(rnp);
4989 init_waitqueue_head(&rnp->exp_wq[0]);
4990 init_waitqueue_head(&rnp->exp_wq[1]);
4991 init_waitqueue_head(&rnp->exp_wq[2]);
4992 init_waitqueue_head(&rnp->exp_wq[3]);
4993 spin_lock_init(&rnp->exp_lock);
4994 mutex_init(&rnp->kthread_mutex);
4995 raw_spin_lock_init(&rnp->exp_poll_lock);
4996 rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED;
4997 INIT_WORK(&rnp->exp_poll_wq, sync_rcu_do_polled_gp);
4998 }
4999 }
5000
5001 init_swait_queue_head(&rcu_state.gp_wq);
5002 init_swait_queue_head(&rcu_state.expedited_wq);
5003 rnp = rcu_first_leaf_node();
5004 for_each_possible_cpu(i) {
5005 while (i > rnp->grphi)
5006 rnp++;
5007 per_cpu_ptr(&rcu_data, i)->mynode = rnp;
5008 rcu_boot_init_percpu_data(i);
5009 }
5010}
5011
5012/*
5013 * Force priority from the kernel command-line into range.
5014 */
5015static void __init sanitize_kthread_prio(void)
5016{
5017 int kthread_prio_in = kthread_prio;
5018
5019 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
5020 && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
5021 kthread_prio = 2;
5022 else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
5023 kthread_prio = 1;
5024 else if (kthread_prio < 0)
5025 kthread_prio = 0;
5026 else if (kthread_prio > 99)
5027 kthread_prio = 99;
5028
5029 if (kthread_prio != kthread_prio_in)
5030 pr_alert("%s: Limited prio to %d from %d\n",
5031 __func__, kthread_prio, kthread_prio_in);
5032}
5033
5034/*
5035 * Compute the rcu_node tree geometry from kernel parameters. This cannot
5036 * replace the definitions in tree.h because those are needed to size
5037 * the ->node array in the rcu_state structure.
5038 */
5039void rcu_init_geometry(void)
5040{
5041 ulong d;
5042 int i;
5043 static unsigned long old_nr_cpu_ids;
5044 int rcu_capacity[RCU_NUM_LVLS];
5045 static bool initialized;
5046
5047 if (initialized) {
5048 /*
5049 * Warn if setup_nr_cpu_ids() had not yet been invoked,
5050 * unless nr_cpus_ids == NR_CPUS, in which case who cares?
5051 */
5052 WARN_ON_ONCE(old_nr_cpu_ids != nr_cpu_ids);
5053 return;
5054 }
5055
5056 old_nr_cpu_ids = nr_cpu_ids;
5057 initialized = true;
5058
5059 /*
5060 * Initialize any unspecified boot parameters.
5061 * The default values of jiffies_till_first_fqs and
5062 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
5063 * value, which is a function of HZ, then adding one for each
5064 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
5065 */
5066 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
5067 if (jiffies_till_first_fqs == ULONG_MAX)
5068 jiffies_till_first_fqs = d;
5069 if (jiffies_till_next_fqs == ULONG_MAX)
5070 jiffies_till_next_fqs = d;
5071 adjust_jiffies_till_sched_qs();
5072
5073 /* If the compile-time values are accurate, just leave. */
5074 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
5075 nr_cpu_ids == NR_CPUS)
5076 return;
5077 pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
5078 rcu_fanout_leaf, nr_cpu_ids);
5079
5080 /*
5081 * The boot-time rcu_fanout_leaf parameter must be at least two
5082 * and cannot exceed the number of bits in the rcu_node masks.
5083 * Complain and fall back to the compile-time values if this
5084 * limit is exceeded.
5085 */
5086 if (rcu_fanout_leaf < 2 ||
5087 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
5088 rcu_fanout_leaf = RCU_FANOUT_LEAF;
5089 WARN_ON(1);
5090 return;
5091 }
5092
5093 /*
5094 * Compute number of nodes that can be handled an rcu_node tree
5095 * with the given number of levels.
5096 */
5097 rcu_capacity[0] = rcu_fanout_leaf;
5098 for (i = 1; i < RCU_NUM_LVLS; i++)
5099 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
5100
5101 /*
5102 * The tree must be able to accommodate the configured number of CPUs.
5103 * If this limit is exceeded, fall back to the compile-time values.
5104 */
5105 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
5106 rcu_fanout_leaf = RCU_FANOUT_LEAF;
5107 WARN_ON(1);
5108 return;
5109 }
5110
5111 /* Calculate the number of levels in the tree. */
5112 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
5113 }
5114 rcu_num_lvls = i + 1;
5115
5116 /* Calculate the number of rcu_nodes at each level of the tree. */
5117 for (i = 0; i < rcu_num_lvls; i++) {
5118 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
5119 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
5120 }
5121
5122 /* Calculate the total number of rcu_node structures. */
5123 rcu_num_nodes = 0;
5124 for (i = 0; i < rcu_num_lvls; i++)
5125 rcu_num_nodes += num_rcu_lvl[i];
5126}
5127
5128/*
5129 * Dump out the structure of the rcu_node combining tree associated
5130 * with the rcu_state structure.
5131 */
5132static void __init rcu_dump_rcu_node_tree(void)
5133{
5134 int level = 0;
5135 struct rcu_node *rnp;
5136
5137 pr_info("rcu_node tree layout dump\n");
5138 pr_info(" ");
5139 rcu_for_each_node_breadth_first(rnp) {
5140 if (rnp->level != level) {
5141 pr_cont("\n");
5142 pr_info(" ");
5143 level = rnp->level;
5144 }
5145 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
5146 }
5147 pr_cont("\n");
5148}
5149
5150struct workqueue_struct *rcu_gp_wq;
5151
5152static void __init kfree_rcu_batch_init(void)
5153{
5154 int cpu;
5155 int i, j;
5156 struct shrinker *kfree_rcu_shrinker;
5157
5158 /* Clamp it to [0:100] seconds interval. */
5159 if (rcu_delay_page_cache_fill_msec < 0 ||
5160 rcu_delay_page_cache_fill_msec > 100 * MSEC_PER_SEC) {
5161
5162 rcu_delay_page_cache_fill_msec =
5163 clamp(rcu_delay_page_cache_fill_msec, 0,
5164 (int) (100 * MSEC_PER_SEC));
5165
5166 pr_info("Adjusting rcutree.rcu_delay_page_cache_fill_msec to %d ms.\n",
5167 rcu_delay_page_cache_fill_msec);
5168 }
5169
5170 for_each_possible_cpu(cpu) {
5171 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
5172
5173 for (i = 0; i < KFREE_N_BATCHES; i++) {
5174 INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
5175 krcp->krw_arr[i].krcp = krcp;
5176
5177 for (j = 0; j < FREE_N_CHANNELS; j++)
5178 INIT_LIST_HEAD(&krcp->krw_arr[i].bulk_head_free[j]);
5179 }
5180
5181 for (i = 0; i < FREE_N_CHANNELS; i++)
5182 INIT_LIST_HEAD(&krcp->bulk_head[i]);
5183
5184 INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
5185 INIT_DELAYED_WORK(&krcp->page_cache_work, fill_page_cache_func);
5186 krcp->initialized = true;
5187 }
5188
5189 kfree_rcu_shrinker = shrinker_alloc(0, "rcu-kfree");
5190 if (!kfree_rcu_shrinker) {
5191 pr_err("Failed to allocate kfree_rcu() shrinker!\n");
5192 return;
5193 }
5194
5195 kfree_rcu_shrinker->count_objects = kfree_rcu_shrink_count;
5196 kfree_rcu_shrinker->scan_objects = kfree_rcu_shrink_scan;
5197
5198 shrinker_register(kfree_rcu_shrinker);
5199}
5200
5201void __init rcu_init(void)
5202{
5203 int cpu = smp_processor_id();
5204
5205 rcu_early_boot_tests();
5206
5207 kfree_rcu_batch_init();
5208 rcu_bootup_announce();
5209 sanitize_kthread_prio();
5210 rcu_init_geometry();
5211 rcu_init_one();
5212 if (dump_tree)
5213 rcu_dump_rcu_node_tree();
5214 if (use_softirq)
5215 open_softirq(RCU_SOFTIRQ, rcu_core_si);
5216
5217 /*
5218 * We don't need protection against CPU-hotplug here because
5219 * this is called early in boot, before either interrupts
5220 * or the scheduler are operational.
5221 */
5222 pm_notifier(rcu_pm_notify, 0);
5223 WARN_ON(num_online_cpus() > 1); // Only one CPU this early in boot.
5224 rcutree_prepare_cpu(cpu);
5225 rcutree_report_cpu_starting(cpu);
5226 rcutree_online_cpu(cpu);
5227
5228 /* Create workqueue for Tree SRCU and for expedited GPs. */
5229 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
5230 WARN_ON(!rcu_gp_wq);
5231
5232 /* Fill in default value for rcutree.qovld boot parameter. */
5233 /* -After- the rcu_node ->lock fields are initialized! */
5234 if (qovld < 0)
5235 qovld_calc = DEFAULT_RCU_QOVLD_MULT * qhimark;
5236 else
5237 qovld_calc = qovld;
5238
5239 // Kick-start in case any polled grace periods started early.
5240 (void)start_poll_synchronize_rcu_expedited();
5241
5242 rcu_test_sync_prims();
5243
5244 tasks_cblist_init_generic();
5245}
5246
5247#include "tree_stall.h"
5248#include "tree_exp.h"
5249#include "tree_nocb.h"
5250#include "tree_plugin.h"