Loading...
Note: File does not exist in v3.5.6.
1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
4 *
5 * Copyright IBM Corporation, 2008
6 *
7 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
8 * Manfred Spraul <manfred@colorfullife.com>
9 * Paul E. McKenney <paulmck@linux.ibm.com>
10 *
11 * Based on the original work by Paul McKenney <paulmck@linux.ibm.com>
12 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
13 *
14 * For detailed explanation of Read-Copy Update mechanism see -
15 * Documentation/RCU
16 */
17
18#define pr_fmt(fmt) "rcu: " fmt
19
20#include <linux/types.h>
21#include <linux/kernel.h>
22#include <linux/init.h>
23#include <linux/spinlock.h>
24#include <linux/smp.h>
25#include <linux/rcupdate_wait.h>
26#include <linux/interrupt.h>
27#include <linux/sched.h>
28#include <linux/sched/debug.h>
29#include <linux/nmi.h>
30#include <linux/atomic.h>
31#include <linux/bitops.h>
32#include <linux/export.h>
33#include <linux/completion.h>
34#include <linux/moduleparam.h>
35#include <linux/panic.h>
36#include <linux/panic_notifier.h>
37#include <linux/percpu.h>
38#include <linux/notifier.h>
39#include <linux/cpu.h>
40#include <linux/mutex.h>
41#include <linux/time.h>
42#include <linux/kernel_stat.h>
43#include <linux/wait.h>
44#include <linux/kthread.h>
45#include <uapi/linux/sched/types.h>
46#include <linux/prefetch.h>
47#include <linux/delay.h>
48#include <linux/random.h>
49#include <linux/trace_events.h>
50#include <linux/suspend.h>
51#include <linux/ftrace.h>
52#include <linux/tick.h>
53#include <linux/sysrq.h>
54#include <linux/kprobes.h>
55#include <linux/gfp.h>
56#include <linux/oom.h>
57#include <linux/smpboot.h>
58#include <linux/jiffies.h>
59#include <linux/slab.h>
60#include <linux/sched/isolation.h>
61#include <linux/sched/clock.h>
62#include <linux/vmalloc.h>
63#include <linux/mm.h>
64#include <linux/kasan.h>
65#include <linux/context_tracking.h>
66#include "../time/tick-internal.h"
67
68#include "tree.h"
69#include "rcu.h"
70
71#ifdef MODULE_PARAM_PREFIX
72#undef MODULE_PARAM_PREFIX
73#endif
74#define MODULE_PARAM_PREFIX "rcutree."
75
76/* Data structures. */
77
78static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = {
79 .gpwrap = true,
80#ifdef CONFIG_RCU_NOCB_CPU
81 .cblist.flags = SEGCBLIST_RCU_CORE,
82#endif
83};
84static struct rcu_state rcu_state = {
85 .level = { &rcu_state.node[0] },
86 .gp_state = RCU_GP_IDLE,
87 .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT,
88 .barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex),
89 .barrier_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.barrier_lock),
90 .name = RCU_NAME,
91 .abbr = RCU_ABBR,
92 .exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
93 .exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
94 .ofl_lock = __ARCH_SPIN_LOCK_UNLOCKED,
95};
96
97/* Dump rcu_node combining tree at boot to verify correct setup. */
98static bool dump_tree;
99module_param(dump_tree, bool, 0444);
100/* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */
101static bool use_softirq = !IS_ENABLED(CONFIG_PREEMPT_RT);
102#ifndef CONFIG_PREEMPT_RT
103module_param(use_softirq, bool, 0444);
104#endif
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);
153static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp);
154static bool rcu_rdp_is_offloaded(struct rcu_data *rdp);
155
156/*
157 * rcuc/rcub/rcuop kthread realtime priority. The "rcuop"
158 * real-time priority(enabling/disabling) is controlled by
159 * the extra CONFIG_RCU_NOCB_CPU_CB_BOOST configuration.
160 */
161static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
162module_param(kthread_prio, int, 0444);
163
164/* Delay in jiffies for grace-period initialization delays, debug only. */
165
166static int gp_preinit_delay;
167module_param(gp_preinit_delay, int, 0444);
168static int gp_init_delay;
169module_param(gp_init_delay, int, 0444);
170static int gp_cleanup_delay;
171module_param(gp_cleanup_delay, int, 0444);
172
173// Add delay to rcu_read_unlock() for strict grace periods.
174static int rcu_unlock_delay;
175#ifdef CONFIG_RCU_STRICT_GRACE_PERIOD
176module_param(rcu_unlock_delay, int, 0444);
177#endif
178
179/*
180 * This rcu parameter is runtime-read-only. It reflects
181 * a minimum allowed number of objects which can be cached
182 * per-CPU. Object size is equal to one page. This value
183 * can be changed at boot time.
184 */
185static int rcu_min_cached_objs = 5;
186module_param(rcu_min_cached_objs, int, 0444);
187
188// A page shrinker can ask for pages to be freed to make them
189// available for other parts of the system. This usually happens
190// under low memory conditions, and in that case we should also
191// defer page-cache filling for a short time period.
192//
193// The default value is 5 seconds, which is long enough to reduce
194// interference with the shrinker while it asks other systems to
195// drain their caches.
196static int rcu_delay_page_cache_fill_msec = 5000;
197module_param(rcu_delay_page_cache_fill_msec, int, 0444);
198
199/* Retrieve RCU kthreads priority for rcutorture */
200int rcu_get_gp_kthreads_prio(void)
201{
202 return kthread_prio;
203}
204EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio);
205
206/*
207 * Number of grace periods between delays, normalized by the duration of
208 * the delay. The longer the delay, the more the grace periods between
209 * each delay. The reason for this normalization is that it means that,
210 * for non-zero delays, the overall slowdown of grace periods is constant
211 * regardless of the duration of the delay. This arrangement balances
212 * the need for long delays to increase some race probabilities with the
213 * need for fast grace periods to increase other race probabilities.
214 */
215#define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays for debugging. */
216
217/*
218 * Compute the mask of online CPUs for the specified rcu_node structure.
219 * This will not be stable unless the rcu_node structure's ->lock is
220 * held, but the bit corresponding to the current CPU will be stable
221 * in most contexts.
222 */
223static unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
224{
225 return READ_ONCE(rnp->qsmaskinitnext);
226}
227
228/*
229 * Is the CPU corresponding to the specified rcu_data structure online
230 * from RCU's perspective? This perspective is given by that structure's
231 * ->qsmaskinitnext field rather than by the global cpu_online_mask.
232 */
233static bool rcu_rdp_cpu_online(struct rcu_data *rdp)
234{
235 return !!(rdp->grpmask & rcu_rnp_online_cpus(rdp->mynode));
236}
237
238/*
239 * Return true if an RCU grace period is in progress. The READ_ONCE()s
240 * permit this function to be invoked without holding the root rcu_node
241 * structure's ->lock, but of course results can be subject to change.
242 */
243static int rcu_gp_in_progress(void)
244{
245 return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq));
246}
247
248/*
249 * Return the number of callbacks queued on the specified CPU.
250 * Handles both the nocbs and normal cases.
251 */
252static long rcu_get_n_cbs_cpu(int cpu)
253{
254 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
255
256 if (rcu_segcblist_is_enabled(&rdp->cblist))
257 return rcu_segcblist_n_cbs(&rdp->cblist);
258 return 0;
259}
260
261void rcu_softirq_qs(void)
262{
263 rcu_qs();
264 rcu_preempt_deferred_qs(current);
265 rcu_tasks_qs(current, false);
266}
267
268/*
269 * Reset the current CPU's ->dynticks counter to indicate that the
270 * newly onlined CPU is no longer in an extended quiescent state.
271 * This will either leave the counter unchanged, or increment it
272 * to the next non-quiescent value.
273 *
274 * The non-atomic test/increment sequence works because the upper bits
275 * of the ->dynticks counter are manipulated only by the corresponding CPU,
276 * or when the corresponding CPU is offline.
277 */
278static void rcu_dynticks_eqs_online(void)
279{
280 if (ct_dynticks() & RCU_DYNTICKS_IDX)
281 return;
282 ct_state_inc(RCU_DYNTICKS_IDX);
283}
284
285/*
286 * Snapshot the ->dynticks counter with full ordering so as to allow
287 * stable comparison of this counter with past and future snapshots.
288 */
289static int rcu_dynticks_snap(int cpu)
290{
291 smp_mb(); // Fundamental RCU ordering guarantee.
292 return ct_dynticks_cpu_acquire(cpu);
293}
294
295/*
296 * Return true if the snapshot returned from rcu_dynticks_snap()
297 * indicates that RCU is in an extended quiescent state.
298 */
299static bool rcu_dynticks_in_eqs(int snap)
300{
301 return !(snap & RCU_DYNTICKS_IDX);
302}
303
304/*
305 * Return true if the CPU corresponding to the specified rcu_data
306 * structure has spent some time in an extended quiescent state since
307 * rcu_dynticks_snap() returned the specified snapshot.
308 */
309static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap)
310{
311 return snap != rcu_dynticks_snap(rdp->cpu);
312}
313
314/*
315 * Return true if the referenced integer is zero while the specified
316 * CPU remains within a single extended quiescent state.
317 */
318bool rcu_dynticks_zero_in_eqs(int cpu, int *vp)
319{
320 int snap;
321
322 // If not quiescent, force back to earlier extended quiescent state.
323 snap = ct_dynticks_cpu(cpu) & ~RCU_DYNTICKS_IDX;
324 smp_rmb(); // Order ->dynticks and *vp reads.
325 if (READ_ONCE(*vp))
326 return false; // Non-zero, so report failure;
327 smp_rmb(); // Order *vp read and ->dynticks re-read.
328
329 // If still in the same extended quiescent state, we are good!
330 return snap == ct_dynticks_cpu(cpu);
331}
332
333/*
334 * Let the RCU core know that this CPU has gone through the scheduler,
335 * which is a quiescent state. This is called when the need for a
336 * quiescent state is urgent, so we burn an atomic operation and full
337 * memory barriers to let the RCU core know about it, regardless of what
338 * this CPU might (or might not) do in the near future.
339 *
340 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
341 *
342 * The caller must have disabled interrupts and must not be idle.
343 */
344notrace void rcu_momentary_dyntick_idle(void)
345{
346 int seq;
347
348 raw_cpu_write(rcu_data.rcu_need_heavy_qs, false);
349 seq = ct_state_inc(2 * RCU_DYNTICKS_IDX);
350 /* It is illegal to call this from idle state. */
351 WARN_ON_ONCE(!(seq & RCU_DYNTICKS_IDX));
352 rcu_preempt_deferred_qs(current);
353}
354EXPORT_SYMBOL_GPL(rcu_momentary_dyntick_idle);
355
356/**
357 * rcu_is_cpu_rrupt_from_idle - see if 'interrupted' from idle
358 *
359 * If the current CPU is idle and running at a first-level (not nested)
360 * interrupt, or directly, from idle, return true.
361 *
362 * The caller must have at least disabled IRQs.
363 */
364static int rcu_is_cpu_rrupt_from_idle(void)
365{
366 long nesting;
367
368 /*
369 * Usually called from the tick; but also used from smp_function_call()
370 * for expedited grace periods. This latter can result in running from
371 * the idle task, instead of an actual IPI.
372 */
373 lockdep_assert_irqs_disabled();
374
375 /* Check for counter underflows */
376 RCU_LOCKDEP_WARN(ct_dynticks_nesting() < 0,
377 "RCU dynticks_nesting counter underflow!");
378 RCU_LOCKDEP_WARN(ct_dynticks_nmi_nesting() <= 0,
379 "RCU dynticks_nmi_nesting counter underflow/zero!");
380
381 /* Are we at first interrupt nesting level? */
382 nesting = ct_dynticks_nmi_nesting();
383 if (nesting > 1)
384 return false;
385
386 /*
387 * If we're not in an interrupt, we must be in the idle task!
388 */
389 WARN_ON_ONCE(!nesting && !is_idle_task(current));
390
391 /* Does CPU appear to be idle from an RCU standpoint? */
392 return ct_dynticks_nesting() == 0;
393}
394
395#define DEFAULT_RCU_BLIMIT (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 1000 : 10)
396 // Maximum callbacks per rcu_do_batch ...
397#define DEFAULT_MAX_RCU_BLIMIT 10000 // ... even during callback flood.
398static long blimit = DEFAULT_RCU_BLIMIT;
399#define DEFAULT_RCU_QHIMARK 10000 // If this many pending, ignore blimit.
400static long qhimark = DEFAULT_RCU_QHIMARK;
401#define DEFAULT_RCU_QLOMARK 100 // Once only this many pending, use blimit.
402static long qlowmark = DEFAULT_RCU_QLOMARK;
403#define DEFAULT_RCU_QOVLD_MULT 2
404#define DEFAULT_RCU_QOVLD (DEFAULT_RCU_QOVLD_MULT * DEFAULT_RCU_QHIMARK)
405static long qovld = DEFAULT_RCU_QOVLD; // If this many pending, hammer QS.
406static long qovld_calc = -1; // No pre-initialization lock acquisitions!
407
408module_param(blimit, long, 0444);
409module_param(qhimark, long, 0444);
410module_param(qlowmark, long, 0444);
411module_param(qovld, long, 0444);
412
413static ulong jiffies_till_first_fqs = IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 0 : ULONG_MAX;
414static ulong jiffies_till_next_fqs = ULONG_MAX;
415static bool rcu_kick_kthreads;
416static int rcu_divisor = 7;
417module_param(rcu_divisor, int, 0644);
418
419/* Force an exit from rcu_do_batch() after 3 milliseconds. */
420static long rcu_resched_ns = 3 * NSEC_PER_MSEC;
421module_param(rcu_resched_ns, long, 0644);
422
423/*
424 * How long the grace period must be before we start recruiting
425 * quiescent-state help from rcu_note_context_switch().
426 */
427static ulong jiffies_till_sched_qs = ULONG_MAX;
428module_param(jiffies_till_sched_qs, ulong, 0444);
429static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */
430module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
431
432/*
433 * Make sure that we give the grace-period kthread time to detect any
434 * idle CPUs before taking active measures to force quiescent states.
435 * However, don't go below 100 milliseconds, adjusted upwards for really
436 * large systems.
437 */
438static void adjust_jiffies_till_sched_qs(void)
439{
440 unsigned long j;
441
442 /* If jiffies_till_sched_qs was specified, respect the request. */
443 if (jiffies_till_sched_qs != ULONG_MAX) {
444 WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
445 return;
446 }
447 /* Otherwise, set to third fqs scan, but bound below on large system. */
448 j = READ_ONCE(jiffies_till_first_fqs) +
449 2 * READ_ONCE(jiffies_till_next_fqs);
450 if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
451 j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
452 pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
453 WRITE_ONCE(jiffies_to_sched_qs, j);
454}
455
456static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
457{
458 ulong j;
459 int ret = kstrtoul(val, 0, &j);
460
461 if (!ret) {
462 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
463 adjust_jiffies_till_sched_qs();
464 }
465 return ret;
466}
467
468static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
469{
470 ulong j;
471 int ret = kstrtoul(val, 0, &j);
472
473 if (!ret) {
474 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
475 adjust_jiffies_till_sched_qs();
476 }
477 return ret;
478}
479
480static const struct kernel_param_ops first_fqs_jiffies_ops = {
481 .set = param_set_first_fqs_jiffies,
482 .get = param_get_ulong,
483};
484
485static const struct kernel_param_ops next_fqs_jiffies_ops = {
486 .set = param_set_next_fqs_jiffies,
487 .get = param_get_ulong,
488};
489
490module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
491module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
492module_param(rcu_kick_kthreads, bool, 0644);
493
494static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
495static int rcu_pending(int user);
496
497/*
498 * Return the number of RCU GPs completed thus far for debug & stats.
499 */
500unsigned long rcu_get_gp_seq(void)
501{
502 return READ_ONCE(rcu_state.gp_seq);
503}
504EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
505
506/*
507 * Return the number of RCU expedited batches completed thus far for
508 * debug & stats. Odd numbers mean that a batch is in progress, even
509 * numbers mean idle. The value returned will thus be roughly double
510 * the cumulative batches since boot.
511 */
512unsigned long rcu_exp_batches_completed(void)
513{
514 return rcu_state.expedited_sequence;
515}
516EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
517
518/*
519 * Return the root node of the rcu_state structure.
520 */
521static struct rcu_node *rcu_get_root(void)
522{
523 return &rcu_state.node[0];
524}
525
526/*
527 * Send along grace-period-related data for rcutorture diagnostics.
528 */
529void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
530 unsigned long *gp_seq)
531{
532 switch (test_type) {
533 case RCU_FLAVOR:
534 *flags = READ_ONCE(rcu_state.gp_flags);
535 *gp_seq = rcu_seq_current(&rcu_state.gp_seq);
536 break;
537 default:
538 break;
539 }
540}
541EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
542
543#if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK))
544/*
545 * An empty function that will trigger a reschedule on
546 * IRQ tail once IRQs get re-enabled on userspace/guest resume.
547 */
548static void late_wakeup_func(struct irq_work *work)
549{
550}
551
552static DEFINE_PER_CPU(struct irq_work, late_wakeup_work) =
553 IRQ_WORK_INIT(late_wakeup_func);
554
555/*
556 * If either:
557 *
558 * 1) the task is about to enter in guest mode and $ARCH doesn't support KVM generic work
559 * 2) the task is about to enter in user mode and $ARCH doesn't support generic entry.
560 *
561 * In these cases the late RCU wake ups aren't supported in the resched loops and our
562 * last resort is to fire a local irq_work that will trigger a reschedule once IRQs
563 * get re-enabled again.
564 */
565noinstr void rcu_irq_work_resched(void)
566{
567 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
568
569 if (IS_ENABLED(CONFIG_GENERIC_ENTRY) && !(current->flags & PF_VCPU))
570 return;
571
572 if (IS_ENABLED(CONFIG_KVM_XFER_TO_GUEST_WORK) && (current->flags & PF_VCPU))
573 return;
574
575 instrumentation_begin();
576 if (do_nocb_deferred_wakeup(rdp) && need_resched()) {
577 irq_work_queue(this_cpu_ptr(&late_wakeup_work));
578 }
579 instrumentation_end();
580}
581#endif /* #if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK)) */
582
583#ifdef CONFIG_PROVE_RCU
584/**
585 * rcu_irq_exit_check_preempt - Validate that scheduling is possible
586 */
587void rcu_irq_exit_check_preempt(void)
588{
589 lockdep_assert_irqs_disabled();
590
591 RCU_LOCKDEP_WARN(ct_dynticks_nesting() <= 0,
592 "RCU dynticks_nesting counter underflow/zero!");
593 RCU_LOCKDEP_WARN(ct_dynticks_nmi_nesting() !=
594 DYNTICK_IRQ_NONIDLE,
595 "Bad RCU dynticks_nmi_nesting counter\n");
596 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
597 "RCU in extended quiescent state!");
598}
599#endif /* #ifdef CONFIG_PROVE_RCU */
600
601#ifdef CONFIG_NO_HZ_FULL
602/**
603 * __rcu_irq_enter_check_tick - Enable scheduler tick on CPU if RCU needs it.
604 *
605 * The scheduler tick is not normally enabled when CPUs enter the kernel
606 * from nohz_full userspace execution. After all, nohz_full userspace
607 * execution is an RCU quiescent state and the time executing in the kernel
608 * is quite short. Except of course when it isn't. And it is not hard to
609 * cause a large system to spend tens of seconds or even minutes looping
610 * in the kernel, which can cause a number of problems, include RCU CPU
611 * stall warnings.
612 *
613 * Therefore, if a nohz_full CPU fails to report a quiescent state
614 * in a timely manner, the RCU grace-period kthread sets that CPU's
615 * ->rcu_urgent_qs flag with the expectation that the next interrupt or
616 * exception will invoke this function, which will turn on the scheduler
617 * tick, which will enable RCU to detect that CPU's quiescent states,
618 * for example, due to cond_resched() calls in CONFIG_PREEMPT=n kernels.
619 * The tick will be disabled once a quiescent state is reported for
620 * this CPU.
621 *
622 * Of course, in carefully tuned systems, there might never be an
623 * interrupt or exception. In that case, the RCU grace-period kthread
624 * will eventually cause one to happen. However, in less carefully
625 * controlled environments, this function allows RCU to get what it
626 * needs without creating otherwise useless interruptions.
627 */
628void __rcu_irq_enter_check_tick(void)
629{
630 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
631
632 // If we're here from NMI there's nothing to do.
633 if (in_nmi())
634 return;
635
636 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
637 "Illegal rcu_irq_enter_check_tick() from extended quiescent state");
638
639 if (!tick_nohz_full_cpu(rdp->cpu) ||
640 !READ_ONCE(rdp->rcu_urgent_qs) ||
641 READ_ONCE(rdp->rcu_forced_tick)) {
642 // RCU doesn't need nohz_full help from this CPU, or it is
643 // already getting that help.
644 return;
645 }
646
647 // We get here only when not in an extended quiescent state and
648 // from interrupts (as opposed to NMIs). Therefore, (1) RCU is
649 // already watching and (2) The fact that we are in an interrupt
650 // handler and that the rcu_node lock is an irq-disabled lock
651 // prevents self-deadlock. So we can safely recheck under the lock.
652 // Note that the nohz_full state currently cannot change.
653 raw_spin_lock_rcu_node(rdp->mynode);
654 if (rdp->rcu_urgent_qs && !rdp->rcu_forced_tick) {
655 // A nohz_full CPU is in the kernel and RCU needs a
656 // quiescent state. Turn on the tick!
657 WRITE_ONCE(rdp->rcu_forced_tick, true);
658 tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
659 }
660 raw_spin_unlock_rcu_node(rdp->mynode);
661}
662#endif /* CONFIG_NO_HZ_FULL */
663
664/*
665 * Check to see if any future non-offloaded RCU-related work will need
666 * to be done by the current CPU, even if none need be done immediately,
667 * returning 1 if so. This function is part of the RCU implementation;
668 * it is -not- an exported member of the RCU API. This is used by
669 * the idle-entry code to figure out whether it is safe to disable the
670 * scheduler-clock interrupt.
671 *
672 * Just check whether or not this CPU has non-offloaded RCU callbacks
673 * queued.
674 */
675int rcu_needs_cpu(void)
676{
677 return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) &&
678 !rcu_rdp_is_offloaded(this_cpu_ptr(&rcu_data));
679}
680
681/*
682 * If any sort of urgency was applied to the current CPU (for example,
683 * the scheduler-clock interrupt was enabled on a nohz_full CPU) in order
684 * to get to a quiescent state, disable it.
685 */
686static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp)
687{
688 raw_lockdep_assert_held_rcu_node(rdp->mynode);
689 WRITE_ONCE(rdp->rcu_urgent_qs, false);
690 WRITE_ONCE(rdp->rcu_need_heavy_qs, false);
691 if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) {
692 tick_dep_clear_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
693 WRITE_ONCE(rdp->rcu_forced_tick, false);
694 }
695}
696
697/**
698 * rcu_is_watching - see if RCU thinks that the current CPU is not idle
699 *
700 * Return true if RCU is watching the running CPU, which means that this
701 * CPU can safely enter RCU read-side critical sections. In other words,
702 * if the current CPU is not in its idle loop or is in an interrupt or
703 * NMI handler, return true.
704 *
705 * Make notrace because it can be called by the internal functions of
706 * ftrace, and making this notrace removes unnecessary recursion calls.
707 */
708notrace bool rcu_is_watching(void)
709{
710 bool ret;
711
712 preempt_disable_notrace();
713 ret = !rcu_dynticks_curr_cpu_in_eqs();
714 preempt_enable_notrace();
715 return ret;
716}
717EXPORT_SYMBOL_GPL(rcu_is_watching);
718
719/*
720 * If a holdout task is actually running, request an urgent quiescent
721 * state from its CPU. This is unsynchronized, so migrations can cause
722 * the request to go to the wrong CPU. Which is OK, all that will happen
723 * is that the CPU's next context switch will be a bit slower and next
724 * time around this task will generate another request.
725 */
726void rcu_request_urgent_qs_task(struct task_struct *t)
727{
728 int cpu;
729
730 barrier();
731 cpu = task_cpu(t);
732 if (!task_curr(t))
733 return; /* This task is not running on that CPU. */
734 smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
735}
736
737#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
738
739/*
740 * Is the current CPU online as far as RCU is concerned?
741 *
742 * Disable preemption to avoid false positives that could otherwise
743 * happen due to the current CPU number being sampled, this task being
744 * preempted, its old CPU being taken offline, resuming on some other CPU,
745 * then determining that its old CPU is now offline.
746 *
747 * Disable checking if in an NMI handler because we cannot safely
748 * report errors from NMI handlers anyway. In addition, it is OK to use
749 * RCU on an offline processor during initial boot, hence the check for
750 * rcu_scheduler_fully_active.
751 */
752bool rcu_lockdep_current_cpu_online(void)
753{
754 struct rcu_data *rdp;
755 bool ret = false;
756
757 if (in_nmi() || !rcu_scheduler_fully_active)
758 return true;
759 preempt_disable_notrace();
760 rdp = this_cpu_ptr(&rcu_data);
761 /*
762 * Strictly, we care here about the case where the current CPU is
763 * in rcu_cpu_starting() and thus has an excuse for rdp->grpmask
764 * not being up to date. So arch_spin_is_locked() might have a
765 * false positive if it's held by some *other* CPU, but that's
766 * OK because that just means a false *negative* on the warning.
767 */
768 if (rcu_rdp_cpu_online(rdp) || arch_spin_is_locked(&rcu_state.ofl_lock))
769 ret = true;
770 preempt_enable_notrace();
771 return ret;
772}
773EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
774
775#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
776
777/*
778 * When trying to report a quiescent state on behalf of some other CPU,
779 * it is our responsibility to check for and handle potential overflow
780 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
781 * After all, the CPU might be in deep idle state, and thus executing no
782 * code whatsoever.
783 */
784static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
785{
786 raw_lockdep_assert_held_rcu_node(rnp);
787 if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
788 rnp->gp_seq))
789 WRITE_ONCE(rdp->gpwrap, true);
790 if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
791 rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
792}
793
794/*
795 * Snapshot the specified CPU's dynticks counter so that we can later
796 * credit them with an implicit quiescent state. Return 1 if this CPU
797 * is in dynticks idle mode, which is an extended quiescent state.
798 */
799static int dyntick_save_progress_counter(struct rcu_data *rdp)
800{
801 rdp->dynticks_snap = rcu_dynticks_snap(rdp->cpu);
802 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
803 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
804 rcu_gpnum_ovf(rdp->mynode, rdp);
805 return 1;
806 }
807 return 0;
808}
809
810/*
811 * Return true if the specified CPU has passed through a quiescent
812 * state by virtue of being in or having passed through an dynticks
813 * idle state since the last call to dyntick_save_progress_counter()
814 * for this same CPU, or by virtue of having been offline.
815 */
816static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
817{
818 unsigned long jtsq;
819 struct rcu_node *rnp = rdp->mynode;
820
821 /*
822 * If the CPU passed through or entered a dynticks idle phase with
823 * no active irq/NMI handlers, then we can safely pretend that the CPU
824 * already acknowledged the request to pass through a quiescent
825 * state. Either way, that CPU cannot possibly be in an RCU
826 * read-side critical section that started before the beginning
827 * of the current RCU grace period.
828 */
829 if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
830 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
831 rcu_gpnum_ovf(rnp, rdp);
832 return 1;
833 }
834
835 /*
836 * Complain if a CPU that is considered to be offline from RCU's
837 * perspective has not yet reported a quiescent state. After all,
838 * the offline CPU should have reported a quiescent state during
839 * the CPU-offline process, or, failing that, by rcu_gp_init()
840 * if it ran concurrently with either the CPU going offline or the
841 * last task on a leaf rcu_node structure exiting its RCU read-side
842 * critical section while all CPUs corresponding to that structure
843 * are offline. This added warning detects bugs in any of these
844 * code paths.
845 *
846 * The rcu_node structure's ->lock is held here, which excludes
847 * the relevant portions the CPU-hotplug code, the grace-period
848 * initialization code, and the rcu_read_unlock() code paths.
849 *
850 * For more detail, please refer to the "Hotplug CPU" section
851 * of RCU's Requirements documentation.
852 */
853 if (WARN_ON_ONCE(!rcu_rdp_cpu_online(rdp))) {
854 struct rcu_node *rnp1;
855
856 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
857 __func__, rnp->grplo, rnp->grphi, rnp->level,
858 (long)rnp->gp_seq, (long)rnp->completedqs);
859 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
860 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
861 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
862 pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
863 __func__, rdp->cpu, ".o"[rcu_rdp_cpu_online(rdp)],
864 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
865 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
866 return 1; /* Break things loose after complaining. */
867 }
868
869 /*
870 * A CPU running for an extended time within the kernel can
871 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
872 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
873 * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the
874 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
875 * variable are safe because the assignments are repeated if this
876 * CPU failed to pass through a quiescent state. This code
877 * also checks .jiffies_resched in case jiffies_to_sched_qs
878 * is set way high.
879 */
880 jtsq = READ_ONCE(jiffies_to_sched_qs);
881 if (!READ_ONCE(rdp->rcu_need_heavy_qs) &&
882 (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
883 time_after(jiffies, rcu_state.jiffies_resched) ||
884 rcu_state.cbovld)) {
885 WRITE_ONCE(rdp->rcu_need_heavy_qs, true);
886 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
887 smp_store_release(&rdp->rcu_urgent_qs, true);
888 } else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
889 WRITE_ONCE(rdp->rcu_urgent_qs, true);
890 }
891
892 /*
893 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
894 * The above code handles this, but only for straight cond_resched().
895 * And some in-kernel loops check need_resched() before calling
896 * cond_resched(), which defeats the above code for CPUs that are
897 * running in-kernel with scheduling-clock interrupts disabled.
898 * So hit them over the head with the resched_cpu() hammer!
899 */
900 if (tick_nohz_full_cpu(rdp->cpu) &&
901 (time_after(jiffies, READ_ONCE(rdp->last_fqs_resched) + jtsq * 3) ||
902 rcu_state.cbovld)) {
903 WRITE_ONCE(rdp->rcu_urgent_qs, true);
904 resched_cpu(rdp->cpu);
905 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
906 }
907
908 /*
909 * If more than halfway to RCU CPU stall-warning time, invoke
910 * resched_cpu() more frequently to try to loosen things up a bit.
911 * Also check to see if the CPU is getting hammered with interrupts,
912 * but only once per grace period, just to keep the IPIs down to
913 * a dull roar.
914 */
915 if (time_after(jiffies, rcu_state.jiffies_resched)) {
916 if (time_after(jiffies,
917 READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
918 resched_cpu(rdp->cpu);
919 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
920 }
921 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
922 !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
923 (rnp->ffmask & rdp->grpmask)) {
924 rdp->rcu_iw_pending = true;
925 rdp->rcu_iw_gp_seq = rnp->gp_seq;
926 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
927 }
928 }
929
930 return 0;
931}
932
933/* Trace-event wrapper function for trace_rcu_future_grace_period. */
934static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
935 unsigned long gp_seq_req, const char *s)
936{
937 trace_rcu_future_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
938 gp_seq_req, rnp->level,
939 rnp->grplo, rnp->grphi, s);
940}
941
942/*
943 * rcu_start_this_gp - Request the start of a particular grace period
944 * @rnp_start: The leaf node of the CPU from which to start.
945 * @rdp: The rcu_data corresponding to the CPU from which to start.
946 * @gp_seq_req: The gp_seq of the grace period to start.
947 *
948 * Start the specified grace period, as needed to handle newly arrived
949 * callbacks. The required future grace periods are recorded in each
950 * rcu_node structure's ->gp_seq_needed field. Returns true if there
951 * is reason to awaken the grace-period kthread.
952 *
953 * The caller must hold the specified rcu_node structure's ->lock, which
954 * is why the caller is responsible for waking the grace-period kthread.
955 *
956 * Returns true if the GP thread needs to be awakened else false.
957 */
958static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
959 unsigned long gp_seq_req)
960{
961 bool ret = false;
962 struct rcu_node *rnp;
963
964 /*
965 * Use funnel locking to either acquire the root rcu_node
966 * structure's lock or bail out if the need for this grace period
967 * has already been recorded -- or if that grace period has in
968 * fact already started. If there is already a grace period in
969 * progress in a non-leaf node, no recording is needed because the
970 * end of the grace period will scan the leaf rcu_node structures.
971 * Note that rnp_start->lock must not be released.
972 */
973 raw_lockdep_assert_held_rcu_node(rnp_start);
974 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
975 for (rnp = rnp_start; 1; rnp = rnp->parent) {
976 if (rnp != rnp_start)
977 raw_spin_lock_rcu_node(rnp);
978 if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
979 rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
980 (rnp != rnp_start &&
981 rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
982 trace_rcu_this_gp(rnp, rdp, gp_seq_req,
983 TPS("Prestarted"));
984 goto unlock_out;
985 }
986 WRITE_ONCE(rnp->gp_seq_needed, gp_seq_req);
987 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
988 /*
989 * We just marked the leaf or internal node, and a
990 * grace period is in progress, which means that
991 * rcu_gp_cleanup() will see the marking. Bail to
992 * reduce contention.
993 */
994 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
995 TPS("Startedleaf"));
996 goto unlock_out;
997 }
998 if (rnp != rnp_start && rnp->parent != NULL)
999 raw_spin_unlock_rcu_node(rnp);
1000 if (!rnp->parent)
1001 break; /* At root, and perhaps also leaf. */
1002 }
1003
1004 /* If GP already in progress, just leave, otherwise start one. */
1005 if (rcu_gp_in_progress()) {
1006 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
1007 goto unlock_out;
1008 }
1009 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
1010 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
1011 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
1012 if (!READ_ONCE(rcu_state.gp_kthread)) {
1013 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
1014 goto unlock_out;
1015 }
1016 trace_rcu_grace_period(rcu_state.name, data_race(rcu_state.gp_seq), TPS("newreq"));
1017 ret = true; /* Caller must wake GP kthread. */
1018unlock_out:
1019 /* Push furthest requested GP to leaf node and rcu_data structure. */
1020 if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
1021 WRITE_ONCE(rnp_start->gp_seq_needed, rnp->gp_seq_needed);
1022 WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1023 }
1024 if (rnp != rnp_start)
1025 raw_spin_unlock_rcu_node(rnp);
1026 return ret;
1027}
1028
1029/*
1030 * Clean up any old requests for the just-ended grace period. Also return
1031 * whether any additional grace periods have been requested.
1032 */
1033static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
1034{
1035 bool needmore;
1036 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1037
1038 needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1039 if (!needmore)
1040 rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1041 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1042 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1043 return needmore;
1044}
1045
1046/*
1047 * Awaken the grace-period kthread. Don't do a self-awaken (unless in an
1048 * interrupt or softirq handler, in which case we just might immediately
1049 * sleep upon return, resulting in a grace-period hang), and don't bother
1050 * awakening when there is nothing for the grace-period kthread to do
1051 * (as in several CPUs raced to awaken, we lost), and finally don't try
1052 * to awaken a kthread that has not yet been created. If all those checks
1053 * are passed, track some debug information and awaken.
1054 *
1055 * So why do the self-wakeup when in an interrupt or softirq handler
1056 * in the grace-period kthread's context? Because the kthread might have
1057 * been interrupted just as it was going to sleep, and just after the final
1058 * pre-sleep check of the awaken condition. In this case, a wakeup really
1059 * is required, and is therefore supplied.
1060 */
1061static void rcu_gp_kthread_wake(void)
1062{
1063 struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
1064
1065 if ((current == t && !in_hardirq() && !in_serving_softirq()) ||
1066 !READ_ONCE(rcu_state.gp_flags) || !t)
1067 return;
1068 WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
1069 WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
1070 swake_up_one(&rcu_state.gp_wq);
1071}
1072
1073/*
1074 * If there is room, assign a ->gp_seq number to any callbacks on this
1075 * CPU that have not already been assigned. Also accelerate any callbacks
1076 * that were previously assigned a ->gp_seq number that has since proven
1077 * to be too conservative, which can happen if callbacks get assigned a
1078 * ->gp_seq number while RCU is idle, but with reference to a non-root
1079 * rcu_node structure. This function is idempotent, so it does not hurt
1080 * to call it repeatedly. Returns an flag saying that we should awaken
1081 * the RCU grace-period kthread.
1082 *
1083 * The caller must hold rnp->lock with interrupts disabled.
1084 */
1085static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1086{
1087 unsigned long gp_seq_req;
1088 bool ret = false;
1089
1090 rcu_lockdep_assert_cblist_protected(rdp);
1091 raw_lockdep_assert_held_rcu_node(rnp);
1092
1093 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1094 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1095 return false;
1096
1097 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPreAcc"));
1098
1099 /*
1100 * Callbacks are often registered with incomplete grace-period
1101 * information. Something about the fact that getting exact
1102 * information requires acquiring a global lock... RCU therefore
1103 * makes a conservative estimate of the grace period number at which
1104 * a given callback will become ready to invoke. The following
1105 * code checks this estimate and improves it when possible, thus
1106 * accelerating callback invocation to an earlier grace-period
1107 * number.
1108 */
1109 gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1110 if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1111 ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1112
1113 /* Trace depending on how much we were able to accelerate. */
1114 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1115 trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccWaitCB"));
1116 else
1117 trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccReadyCB"));
1118
1119 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPostAcc"));
1120
1121 return ret;
1122}
1123
1124/*
1125 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1126 * rcu_node structure's ->lock be held. It consults the cached value
1127 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1128 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1129 * while holding the leaf rcu_node structure's ->lock.
1130 */
1131static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1132 struct rcu_data *rdp)
1133{
1134 unsigned long c;
1135 bool needwake;
1136
1137 rcu_lockdep_assert_cblist_protected(rdp);
1138 c = rcu_seq_snap(&rcu_state.gp_seq);
1139 if (!READ_ONCE(rdp->gpwrap) && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1140 /* Old request still live, so mark recent callbacks. */
1141 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1142 return;
1143 }
1144 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1145 needwake = rcu_accelerate_cbs(rnp, rdp);
1146 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1147 if (needwake)
1148 rcu_gp_kthread_wake();
1149}
1150
1151/*
1152 * Move any callbacks whose grace period has completed to the
1153 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1154 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1155 * sublist. This function is idempotent, so it does not hurt to
1156 * invoke it repeatedly. As long as it is not invoked -too- often...
1157 * Returns true if the RCU grace-period kthread needs to be awakened.
1158 *
1159 * The caller must hold rnp->lock with interrupts disabled.
1160 */
1161static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1162{
1163 rcu_lockdep_assert_cblist_protected(rdp);
1164 raw_lockdep_assert_held_rcu_node(rnp);
1165
1166 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1167 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1168 return false;
1169
1170 /*
1171 * Find all callbacks whose ->gp_seq numbers indicate that they
1172 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1173 */
1174 rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1175
1176 /* Classify any remaining callbacks. */
1177 return rcu_accelerate_cbs(rnp, rdp);
1178}
1179
1180/*
1181 * Move and classify callbacks, but only if doing so won't require
1182 * that the RCU grace-period kthread be awakened.
1183 */
1184static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
1185 struct rcu_data *rdp)
1186{
1187 rcu_lockdep_assert_cblist_protected(rdp);
1188 if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) || !raw_spin_trylock_rcu_node(rnp))
1189 return;
1190 // The grace period cannot end while we hold the rcu_node lock.
1191 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))
1192 WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
1193 raw_spin_unlock_rcu_node(rnp);
1194}
1195
1196/*
1197 * In CONFIG_RCU_STRICT_GRACE_PERIOD=y kernels, attempt to generate a
1198 * quiescent state. This is intended to be invoked when the CPU notices
1199 * a new grace period.
1200 */
1201static void rcu_strict_gp_check_qs(void)
1202{
1203 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
1204 rcu_read_lock();
1205 rcu_read_unlock();
1206 }
1207}
1208
1209/*
1210 * Update CPU-local rcu_data state to record the beginnings and ends of
1211 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1212 * structure corresponding to the current CPU, and must have irqs disabled.
1213 * Returns true if the grace-period kthread needs to be awakened.
1214 */
1215static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
1216{
1217 bool ret = false;
1218 bool need_qs;
1219 const bool offloaded = rcu_rdp_is_offloaded(rdp);
1220
1221 raw_lockdep_assert_held_rcu_node(rnp);
1222
1223 if (rdp->gp_seq == rnp->gp_seq)
1224 return false; /* Nothing to do. */
1225
1226 /* Handle the ends of any preceding grace periods first. */
1227 if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1228 unlikely(READ_ONCE(rdp->gpwrap))) {
1229 if (!offloaded)
1230 ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
1231 rdp->core_needs_qs = false;
1232 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1233 } else {
1234 if (!offloaded)
1235 ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
1236 if (rdp->core_needs_qs)
1237 rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask);
1238 }
1239
1240 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1241 if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1242 unlikely(READ_ONCE(rdp->gpwrap))) {
1243 /*
1244 * If the current grace period is waiting for this CPU,
1245 * set up to detect a quiescent state, otherwise don't
1246 * go looking for one.
1247 */
1248 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1249 need_qs = !!(rnp->qsmask & rdp->grpmask);
1250 rdp->cpu_no_qs.b.norm = need_qs;
1251 rdp->core_needs_qs = need_qs;
1252 zero_cpu_stall_ticks(rdp);
1253 }
1254 rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
1255 if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
1256 WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1257 if (IS_ENABLED(CONFIG_PROVE_RCU) && READ_ONCE(rdp->gpwrap))
1258 WRITE_ONCE(rdp->last_sched_clock, jiffies);
1259 WRITE_ONCE(rdp->gpwrap, false);
1260 rcu_gpnum_ovf(rnp, rdp);
1261 return ret;
1262}
1263
1264static void note_gp_changes(struct rcu_data *rdp)
1265{
1266 unsigned long flags;
1267 bool needwake;
1268 struct rcu_node *rnp;
1269
1270 local_irq_save(flags);
1271 rnp = rdp->mynode;
1272 if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1273 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1274 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1275 local_irq_restore(flags);
1276 return;
1277 }
1278 needwake = __note_gp_changes(rnp, rdp);
1279 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1280 rcu_strict_gp_check_qs();
1281 if (needwake)
1282 rcu_gp_kthread_wake();
1283}
1284
1285static atomic_t *rcu_gp_slow_suppress;
1286
1287/* Register a counter to suppress debugging grace-period delays. */
1288void rcu_gp_slow_register(atomic_t *rgssp)
1289{
1290 WARN_ON_ONCE(rcu_gp_slow_suppress);
1291
1292 WRITE_ONCE(rcu_gp_slow_suppress, rgssp);
1293}
1294EXPORT_SYMBOL_GPL(rcu_gp_slow_register);
1295
1296/* Unregister a counter, with NULL for not caring which. */
1297void rcu_gp_slow_unregister(atomic_t *rgssp)
1298{
1299 WARN_ON_ONCE(rgssp && rgssp != rcu_gp_slow_suppress);
1300
1301 WRITE_ONCE(rcu_gp_slow_suppress, NULL);
1302}
1303EXPORT_SYMBOL_GPL(rcu_gp_slow_unregister);
1304
1305static bool rcu_gp_slow_is_suppressed(void)
1306{
1307 atomic_t *rgssp = READ_ONCE(rcu_gp_slow_suppress);
1308
1309 return rgssp && atomic_read(rgssp);
1310}
1311
1312static void rcu_gp_slow(int delay)
1313{
1314 if (!rcu_gp_slow_is_suppressed() && delay > 0 &&
1315 !(rcu_seq_ctr(rcu_state.gp_seq) % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1316 schedule_timeout_idle(delay);
1317}
1318
1319static unsigned long sleep_duration;
1320
1321/* Allow rcutorture to stall the grace-period kthread. */
1322void rcu_gp_set_torture_wait(int duration)
1323{
1324 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST) && duration > 0)
1325 WRITE_ONCE(sleep_duration, duration);
1326}
1327EXPORT_SYMBOL_GPL(rcu_gp_set_torture_wait);
1328
1329/* Actually implement the aforementioned wait. */
1330static void rcu_gp_torture_wait(void)
1331{
1332 unsigned long duration;
1333
1334 if (!IS_ENABLED(CONFIG_RCU_TORTURE_TEST))
1335 return;
1336 duration = xchg(&sleep_duration, 0UL);
1337 if (duration > 0) {
1338 pr_alert("%s: Waiting %lu jiffies\n", __func__, duration);
1339 schedule_timeout_idle(duration);
1340 pr_alert("%s: Wait complete\n", __func__);
1341 }
1342}
1343
1344/*
1345 * Handler for on_each_cpu() to invoke the target CPU's RCU core
1346 * processing.
1347 */
1348static void rcu_strict_gp_boundary(void *unused)
1349{
1350 invoke_rcu_core();
1351}
1352
1353// Has rcu_init() been invoked? This is used (for example) to determine
1354// whether spinlocks may be acquired safely.
1355static bool rcu_init_invoked(void)
1356{
1357 return !!rcu_state.n_online_cpus;
1358}
1359
1360// Make the polled API aware of the beginning of a grace period.
1361static void rcu_poll_gp_seq_start(unsigned long *snap)
1362{
1363 struct rcu_node *rnp = rcu_get_root();
1364
1365 if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1366 raw_lockdep_assert_held_rcu_node(rnp);
1367
1368 // If RCU was idle, note beginning of GP.
1369 if (!rcu_seq_state(rcu_state.gp_seq_polled))
1370 rcu_seq_start(&rcu_state.gp_seq_polled);
1371
1372 // Either way, record current state.
1373 *snap = rcu_state.gp_seq_polled;
1374}
1375
1376// Make the polled API aware of the end of a grace period.
1377static void rcu_poll_gp_seq_end(unsigned long *snap)
1378{
1379 struct rcu_node *rnp = rcu_get_root();
1380
1381 if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1382 raw_lockdep_assert_held_rcu_node(rnp);
1383
1384 // If the previously noted GP is still in effect, record the
1385 // end of that GP. Either way, zero counter to avoid counter-wrap
1386 // problems.
1387 if (*snap && *snap == rcu_state.gp_seq_polled) {
1388 rcu_seq_end(&rcu_state.gp_seq_polled);
1389 rcu_state.gp_seq_polled_snap = 0;
1390 rcu_state.gp_seq_polled_exp_snap = 0;
1391 } else {
1392 *snap = 0;
1393 }
1394}
1395
1396// Make the polled API aware of the beginning of a grace period, but
1397// where caller does not hold the root rcu_node structure's lock.
1398static void rcu_poll_gp_seq_start_unlocked(unsigned long *snap)
1399{
1400 unsigned long flags;
1401 struct rcu_node *rnp = rcu_get_root();
1402
1403 if (rcu_init_invoked()) {
1404 if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1405 lockdep_assert_irqs_enabled();
1406 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1407 }
1408 rcu_poll_gp_seq_start(snap);
1409 if (rcu_init_invoked())
1410 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1411}
1412
1413// Make the polled API aware of the end of a grace period, but where
1414// caller does not hold the root rcu_node structure's lock.
1415static void rcu_poll_gp_seq_end_unlocked(unsigned long *snap)
1416{
1417 unsigned long flags;
1418 struct rcu_node *rnp = rcu_get_root();
1419
1420 if (rcu_init_invoked()) {
1421 if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1422 lockdep_assert_irqs_enabled();
1423 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1424 }
1425 rcu_poll_gp_seq_end(snap);
1426 if (rcu_init_invoked())
1427 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1428}
1429
1430/*
1431 * Initialize a new grace period. Return false if no grace period required.
1432 */
1433static noinline_for_stack bool rcu_gp_init(void)
1434{
1435 unsigned long flags;
1436 unsigned long oldmask;
1437 unsigned long mask;
1438 struct rcu_data *rdp;
1439 struct rcu_node *rnp = rcu_get_root();
1440
1441 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1442 raw_spin_lock_irq_rcu_node(rnp);
1443 if (!READ_ONCE(rcu_state.gp_flags)) {
1444 /* Spurious wakeup, tell caller to go back to sleep. */
1445 raw_spin_unlock_irq_rcu_node(rnp);
1446 return false;
1447 }
1448 WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1449
1450 if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1451 /*
1452 * Grace period already in progress, don't start another.
1453 * Not supposed to be able to happen.
1454 */
1455 raw_spin_unlock_irq_rcu_node(rnp);
1456 return false;
1457 }
1458
1459 /* Advance to a new grace period and initialize state. */
1460 record_gp_stall_check_time();
1461 /* Record GP times before starting GP, hence rcu_seq_start(). */
1462 rcu_seq_start(&rcu_state.gp_seq);
1463 ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
1464 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
1465 rcu_poll_gp_seq_start(&rcu_state.gp_seq_polled_snap);
1466 raw_spin_unlock_irq_rcu_node(rnp);
1467
1468 /*
1469 * Apply per-leaf buffered online and offline operations to
1470 * the rcu_node tree. Note that this new grace period need not
1471 * wait for subsequent online CPUs, and that RCU hooks in the CPU
1472 * offlining path, when combined with checks in this function,
1473 * will handle CPUs that are currently going offline or that will
1474 * go offline later. Please also refer to "Hotplug CPU" section
1475 * of RCU's Requirements documentation.
1476 */
1477 WRITE_ONCE(rcu_state.gp_state, RCU_GP_ONOFF);
1478 /* Exclude CPU hotplug operations. */
1479 rcu_for_each_leaf_node(rnp) {
1480 local_irq_save(flags);
1481 arch_spin_lock(&rcu_state.ofl_lock);
1482 raw_spin_lock_rcu_node(rnp);
1483 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1484 !rnp->wait_blkd_tasks) {
1485 /* Nothing to do on this leaf rcu_node structure. */
1486 raw_spin_unlock_rcu_node(rnp);
1487 arch_spin_unlock(&rcu_state.ofl_lock);
1488 local_irq_restore(flags);
1489 continue;
1490 }
1491
1492 /* Record old state, apply changes to ->qsmaskinit field. */
1493 oldmask = rnp->qsmaskinit;
1494 rnp->qsmaskinit = rnp->qsmaskinitnext;
1495
1496 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1497 if (!oldmask != !rnp->qsmaskinit) {
1498 if (!oldmask) { /* First online CPU for rcu_node. */
1499 if (!rnp->wait_blkd_tasks) /* Ever offline? */
1500 rcu_init_new_rnp(rnp);
1501 } else if (rcu_preempt_has_tasks(rnp)) {
1502 rnp->wait_blkd_tasks = true; /* blocked tasks */
1503 } else { /* Last offline CPU and can propagate. */
1504 rcu_cleanup_dead_rnp(rnp);
1505 }
1506 }
1507
1508 /*
1509 * If all waited-on tasks from prior grace period are
1510 * done, and if all this rcu_node structure's CPUs are
1511 * still offline, propagate up the rcu_node tree and
1512 * clear ->wait_blkd_tasks. Otherwise, if one of this
1513 * rcu_node structure's CPUs has since come back online,
1514 * simply clear ->wait_blkd_tasks.
1515 */
1516 if (rnp->wait_blkd_tasks &&
1517 (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1518 rnp->wait_blkd_tasks = false;
1519 if (!rnp->qsmaskinit)
1520 rcu_cleanup_dead_rnp(rnp);
1521 }
1522
1523 raw_spin_unlock_rcu_node(rnp);
1524 arch_spin_unlock(&rcu_state.ofl_lock);
1525 local_irq_restore(flags);
1526 }
1527 rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1528
1529 /*
1530 * Set the quiescent-state-needed bits in all the rcu_node
1531 * structures for all currently online CPUs in breadth-first
1532 * order, starting from the root rcu_node structure, relying on the
1533 * layout of the tree within the rcu_state.node[] array. Note that
1534 * other CPUs will access only the leaves of the hierarchy, thus
1535 * seeing that no grace period is in progress, at least until the
1536 * corresponding leaf node has been initialized.
1537 *
1538 * The grace period cannot complete until the initialization
1539 * process finishes, because this kthread handles both.
1540 */
1541 WRITE_ONCE(rcu_state.gp_state, RCU_GP_INIT);
1542 rcu_for_each_node_breadth_first(rnp) {
1543 rcu_gp_slow(gp_init_delay);
1544 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1545 rdp = this_cpu_ptr(&rcu_data);
1546 rcu_preempt_check_blocked_tasks(rnp);
1547 rnp->qsmask = rnp->qsmaskinit;
1548 WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
1549 if (rnp == rdp->mynode)
1550 (void)__note_gp_changes(rnp, rdp);
1551 rcu_preempt_boost_start_gp(rnp);
1552 trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1553 rnp->level, rnp->grplo,
1554 rnp->grphi, rnp->qsmask);
1555 /* Quiescent states for tasks on any now-offline CPUs. */
1556 mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1557 rnp->rcu_gp_init_mask = mask;
1558 if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1559 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1560 else
1561 raw_spin_unlock_irq_rcu_node(rnp);
1562 cond_resched_tasks_rcu_qs();
1563 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1564 }
1565
1566 // If strict, make all CPUs aware of new grace period.
1567 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
1568 on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
1569
1570 return true;
1571}
1572
1573/*
1574 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1575 * time.
1576 */
1577static bool rcu_gp_fqs_check_wake(int *gfp)
1578{
1579 struct rcu_node *rnp = rcu_get_root();
1580
1581 // If under overload conditions, force an immediate FQS scan.
1582 if (*gfp & RCU_GP_FLAG_OVLD)
1583 return true;
1584
1585 // Someone like call_rcu() requested a force-quiescent-state scan.
1586 *gfp = READ_ONCE(rcu_state.gp_flags);
1587 if (*gfp & RCU_GP_FLAG_FQS)
1588 return true;
1589
1590 // The current grace period has completed.
1591 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1592 return true;
1593
1594 return false;
1595}
1596
1597/*
1598 * Do one round of quiescent-state forcing.
1599 */
1600static void rcu_gp_fqs(bool first_time)
1601{
1602 struct rcu_node *rnp = rcu_get_root();
1603
1604 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1605 WRITE_ONCE(rcu_state.n_force_qs, rcu_state.n_force_qs + 1);
1606 if (first_time) {
1607 /* Collect dyntick-idle snapshots. */
1608 force_qs_rnp(dyntick_save_progress_counter);
1609 } else {
1610 /* Handle dyntick-idle and offline CPUs. */
1611 force_qs_rnp(rcu_implicit_dynticks_qs);
1612 }
1613 /* Clear flag to prevent immediate re-entry. */
1614 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
1615 raw_spin_lock_irq_rcu_node(rnp);
1616 WRITE_ONCE(rcu_state.gp_flags,
1617 READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
1618 raw_spin_unlock_irq_rcu_node(rnp);
1619 }
1620}
1621
1622/*
1623 * Loop doing repeated quiescent-state forcing until the grace period ends.
1624 */
1625static noinline_for_stack void rcu_gp_fqs_loop(void)
1626{
1627 bool first_gp_fqs = true;
1628 int gf = 0;
1629 unsigned long j;
1630 int ret;
1631 struct rcu_node *rnp = rcu_get_root();
1632
1633 j = READ_ONCE(jiffies_till_first_fqs);
1634 if (rcu_state.cbovld)
1635 gf = RCU_GP_FLAG_OVLD;
1636 ret = 0;
1637 for (;;) {
1638 if (rcu_state.cbovld) {
1639 j = (j + 2) / 3;
1640 if (j <= 0)
1641 j = 1;
1642 }
1643 if (!ret || time_before(jiffies + j, rcu_state.jiffies_force_qs)) {
1644 WRITE_ONCE(rcu_state.jiffies_force_qs, jiffies + j);
1645 /*
1646 * jiffies_force_qs before RCU_GP_WAIT_FQS state
1647 * update; required for stall checks.
1648 */
1649 smp_wmb();
1650 WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
1651 jiffies + (j ? 3 * j : 2));
1652 }
1653 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1654 TPS("fqswait"));
1655 WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_FQS);
1656 (void)swait_event_idle_timeout_exclusive(rcu_state.gp_wq,
1657 rcu_gp_fqs_check_wake(&gf), j);
1658 rcu_gp_torture_wait();
1659 WRITE_ONCE(rcu_state.gp_state, RCU_GP_DOING_FQS);
1660 /* Locking provides needed memory barriers. */
1661 /*
1662 * Exit the loop if the root rcu_node structure indicates that the grace period
1663 * has ended, leave the loop. The rcu_preempt_blocked_readers_cgp(rnp) check
1664 * is required only for single-node rcu_node trees because readers blocking
1665 * the current grace period are queued only on leaf rcu_node structures.
1666 * For multi-node trees, checking the root node's ->qsmask suffices, because a
1667 * given root node's ->qsmask bit is cleared only when all CPUs and tasks from
1668 * the corresponding leaf nodes have passed through their quiescent state.
1669 */
1670 if (!READ_ONCE(rnp->qsmask) &&
1671 !rcu_preempt_blocked_readers_cgp(rnp))
1672 break;
1673 /* If time for quiescent-state forcing, do it. */
1674 if (!time_after(rcu_state.jiffies_force_qs, jiffies) ||
1675 (gf & (RCU_GP_FLAG_FQS | RCU_GP_FLAG_OVLD))) {
1676 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1677 TPS("fqsstart"));
1678 rcu_gp_fqs(first_gp_fqs);
1679 gf = 0;
1680 if (first_gp_fqs) {
1681 first_gp_fqs = false;
1682 gf = rcu_state.cbovld ? RCU_GP_FLAG_OVLD : 0;
1683 }
1684 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1685 TPS("fqsend"));
1686 cond_resched_tasks_rcu_qs();
1687 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1688 ret = 0; /* Force full wait till next FQS. */
1689 j = READ_ONCE(jiffies_till_next_fqs);
1690 } else {
1691 /* Deal with stray signal. */
1692 cond_resched_tasks_rcu_qs();
1693 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1694 WARN_ON(signal_pending(current));
1695 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1696 TPS("fqswaitsig"));
1697 ret = 1; /* Keep old FQS timing. */
1698 j = jiffies;
1699 if (time_after(jiffies, rcu_state.jiffies_force_qs))
1700 j = 1;
1701 else
1702 j = rcu_state.jiffies_force_qs - j;
1703 gf = 0;
1704 }
1705 }
1706}
1707
1708/*
1709 * Clean up after the old grace period.
1710 */
1711static noinline void rcu_gp_cleanup(void)
1712{
1713 int cpu;
1714 bool needgp = false;
1715 unsigned long gp_duration;
1716 unsigned long new_gp_seq;
1717 bool offloaded;
1718 struct rcu_data *rdp;
1719 struct rcu_node *rnp = rcu_get_root();
1720 struct swait_queue_head *sq;
1721
1722 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1723 raw_spin_lock_irq_rcu_node(rnp);
1724 rcu_state.gp_end = jiffies;
1725 gp_duration = rcu_state.gp_end - rcu_state.gp_start;
1726 if (gp_duration > rcu_state.gp_max)
1727 rcu_state.gp_max = gp_duration;
1728
1729 /*
1730 * We know the grace period is complete, but to everyone else
1731 * it appears to still be ongoing. But it is also the case
1732 * that to everyone else it looks like there is nothing that
1733 * they can do to advance the grace period. It is therefore
1734 * safe for us to drop the lock in order to mark the grace
1735 * period as completed in all of the rcu_node structures.
1736 */
1737 rcu_poll_gp_seq_end(&rcu_state.gp_seq_polled_snap);
1738 raw_spin_unlock_irq_rcu_node(rnp);
1739
1740 /*
1741 * Propagate new ->gp_seq value to rcu_node structures so that
1742 * other CPUs don't have to wait until the start of the next grace
1743 * period to process their callbacks. This also avoids some nasty
1744 * RCU grace-period initialization races by forcing the end of
1745 * the current grace period to be completely recorded in all of
1746 * the rcu_node structures before the beginning of the next grace
1747 * period is recorded in any of the rcu_node structures.
1748 */
1749 new_gp_seq = rcu_state.gp_seq;
1750 rcu_seq_end(&new_gp_seq);
1751 rcu_for_each_node_breadth_first(rnp) {
1752 raw_spin_lock_irq_rcu_node(rnp);
1753 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
1754 dump_blkd_tasks(rnp, 10);
1755 WARN_ON_ONCE(rnp->qsmask);
1756 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
1757 if (!rnp->parent)
1758 smp_mb(); // Order against failing poll_state_synchronize_rcu_full().
1759 rdp = this_cpu_ptr(&rcu_data);
1760 if (rnp == rdp->mynode)
1761 needgp = __note_gp_changes(rnp, rdp) || needgp;
1762 /* smp_mb() provided by prior unlock-lock pair. */
1763 needgp = rcu_future_gp_cleanup(rnp) || needgp;
1764 // Reset overload indication for CPUs no longer overloaded
1765 if (rcu_is_leaf_node(rnp))
1766 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->cbovldmask) {
1767 rdp = per_cpu_ptr(&rcu_data, cpu);
1768 check_cb_ovld_locked(rdp, rnp);
1769 }
1770 sq = rcu_nocb_gp_get(rnp);
1771 raw_spin_unlock_irq_rcu_node(rnp);
1772 rcu_nocb_gp_cleanup(sq);
1773 cond_resched_tasks_rcu_qs();
1774 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1775 rcu_gp_slow(gp_cleanup_delay);
1776 }
1777 rnp = rcu_get_root();
1778 raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
1779
1780 /* Declare grace period done, trace first to use old GP number. */
1781 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
1782 rcu_seq_end(&rcu_state.gp_seq);
1783 ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
1784 WRITE_ONCE(rcu_state.gp_state, RCU_GP_IDLE);
1785 /* Check for GP requests since above loop. */
1786 rdp = this_cpu_ptr(&rcu_data);
1787 if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
1788 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
1789 TPS("CleanupMore"));
1790 needgp = true;
1791 }
1792 /* Advance CBs to reduce false positives below. */
1793 offloaded = rcu_rdp_is_offloaded(rdp);
1794 if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
1795
1796 // We get here if a grace period was needed (“needgp”)
1797 // and the above call to rcu_accelerate_cbs() did not set
1798 // the RCU_GP_FLAG_INIT bit in ->gp_state (which records
1799 // the need for another grace period). The purpose
1800 // of the “offloaded” check is to avoid invoking
1801 // rcu_accelerate_cbs() on an offloaded CPU because we do not
1802 // hold the ->nocb_lock needed to safely access an offloaded
1803 // ->cblist. We do not want to acquire that lock because
1804 // it can be heavily contended during callback floods.
1805
1806 WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
1807 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
1808 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("newreq"));
1809 } else {
1810
1811 // We get here either if there is no need for an
1812 // additional grace period or if rcu_accelerate_cbs() has
1813 // already set the RCU_GP_FLAG_INIT bit in ->gp_flags.
1814 // So all we need to do is to clear all of the other
1815 // ->gp_flags bits.
1816
1817 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags & RCU_GP_FLAG_INIT);
1818 }
1819 raw_spin_unlock_irq_rcu_node(rnp);
1820
1821 // If strict, make all CPUs aware of the end of the old grace period.
1822 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
1823 on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
1824}
1825
1826/*
1827 * Body of kthread that handles grace periods.
1828 */
1829static int __noreturn rcu_gp_kthread(void *unused)
1830{
1831 rcu_bind_gp_kthread();
1832 for (;;) {
1833
1834 /* Handle grace-period start. */
1835 for (;;) {
1836 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1837 TPS("reqwait"));
1838 WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_GPS);
1839 swait_event_idle_exclusive(rcu_state.gp_wq,
1840 READ_ONCE(rcu_state.gp_flags) &
1841 RCU_GP_FLAG_INIT);
1842 rcu_gp_torture_wait();
1843 WRITE_ONCE(rcu_state.gp_state, RCU_GP_DONE_GPS);
1844 /* Locking provides needed memory barrier. */
1845 if (rcu_gp_init())
1846 break;
1847 cond_resched_tasks_rcu_qs();
1848 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1849 WARN_ON(signal_pending(current));
1850 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1851 TPS("reqwaitsig"));
1852 }
1853
1854 /* Handle quiescent-state forcing. */
1855 rcu_gp_fqs_loop();
1856
1857 /* Handle grace-period end. */
1858 WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANUP);
1859 rcu_gp_cleanup();
1860 WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANED);
1861 }
1862}
1863
1864/*
1865 * Report a full set of quiescent states to the rcu_state data structure.
1866 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
1867 * another grace period is required. Whether we wake the grace-period
1868 * kthread or it awakens itself for the next round of quiescent-state
1869 * forcing, that kthread will clean up after the just-completed grace
1870 * period. Note that the caller must hold rnp->lock, which is released
1871 * before return.
1872 */
1873static void rcu_report_qs_rsp(unsigned long flags)
1874 __releases(rcu_get_root()->lock)
1875{
1876 raw_lockdep_assert_held_rcu_node(rcu_get_root());
1877 WARN_ON_ONCE(!rcu_gp_in_progress());
1878 WRITE_ONCE(rcu_state.gp_flags,
1879 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
1880 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
1881 rcu_gp_kthread_wake();
1882}
1883
1884/*
1885 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1886 * Allows quiescent states for a group of CPUs to be reported at one go
1887 * to the specified rcu_node structure, though all the CPUs in the group
1888 * must be represented by the same rcu_node structure (which need not be a
1889 * leaf rcu_node structure, though it often will be). The gps parameter
1890 * is the grace-period snapshot, which means that the quiescent states
1891 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
1892 * must be held upon entry, and it is released before return.
1893 *
1894 * As a special case, if mask is zero, the bit-already-cleared check is
1895 * disabled. This allows propagating quiescent state due to resumed tasks
1896 * during grace-period initialization.
1897 */
1898static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
1899 unsigned long gps, unsigned long flags)
1900 __releases(rnp->lock)
1901{
1902 unsigned long oldmask = 0;
1903 struct rcu_node *rnp_c;
1904
1905 raw_lockdep_assert_held_rcu_node(rnp);
1906
1907 /* Walk up the rcu_node hierarchy. */
1908 for (;;) {
1909 if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
1910
1911 /*
1912 * Our bit has already been cleared, or the
1913 * relevant grace period is already over, so done.
1914 */
1915 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1916 return;
1917 }
1918 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
1919 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
1920 rcu_preempt_blocked_readers_cgp(rnp));
1921 WRITE_ONCE(rnp->qsmask, rnp->qsmask & ~mask);
1922 trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
1923 mask, rnp->qsmask, rnp->level,
1924 rnp->grplo, rnp->grphi,
1925 !!rnp->gp_tasks);
1926 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1927
1928 /* Other bits still set at this level, so done. */
1929 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1930 return;
1931 }
1932 rnp->completedqs = rnp->gp_seq;
1933 mask = rnp->grpmask;
1934 if (rnp->parent == NULL) {
1935
1936 /* No more levels. Exit loop holding root lock. */
1937
1938 break;
1939 }
1940 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1941 rnp_c = rnp;
1942 rnp = rnp->parent;
1943 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1944 oldmask = READ_ONCE(rnp_c->qsmask);
1945 }
1946
1947 /*
1948 * Get here if we are the last CPU to pass through a quiescent
1949 * state for this grace period. Invoke rcu_report_qs_rsp()
1950 * to clean up and start the next grace period if one is needed.
1951 */
1952 rcu_report_qs_rsp(flags); /* releases rnp->lock. */
1953}
1954
1955/*
1956 * Record a quiescent state for all tasks that were previously queued
1957 * on the specified rcu_node structure and that were blocking the current
1958 * RCU grace period. The caller must hold the corresponding rnp->lock with
1959 * irqs disabled, and this lock is released upon return, but irqs remain
1960 * disabled.
1961 */
1962static void __maybe_unused
1963rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
1964 __releases(rnp->lock)
1965{
1966 unsigned long gps;
1967 unsigned long mask;
1968 struct rcu_node *rnp_p;
1969
1970 raw_lockdep_assert_held_rcu_node(rnp);
1971 if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RCU)) ||
1972 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
1973 rnp->qsmask != 0) {
1974 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1975 return; /* Still need more quiescent states! */
1976 }
1977
1978 rnp->completedqs = rnp->gp_seq;
1979 rnp_p = rnp->parent;
1980 if (rnp_p == NULL) {
1981 /*
1982 * Only one rcu_node structure in the tree, so don't
1983 * try to report up to its nonexistent parent!
1984 */
1985 rcu_report_qs_rsp(flags);
1986 return;
1987 }
1988
1989 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
1990 gps = rnp->gp_seq;
1991 mask = rnp->grpmask;
1992 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1993 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
1994 rcu_report_qs_rnp(mask, rnp_p, gps, flags);
1995}
1996
1997/*
1998 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1999 * structure. This must be called from the specified CPU.
2000 */
2001static void
2002rcu_report_qs_rdp(struct rcu_data *rdp)
2003{
2004 unsigned long flags;
2005 unsigned long mask;
2006 bool needwake = false;
2007 bool needacc = false;
2008 struct rcu_node *rnp;
2009
2010 WARN_ON_ONCE(rdp->cpu != smp_processor_id());
2011 rnp = rdp->mynode;
2012 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2013 if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
2014 rdp->gpwrap) {
2015
2016 /*
2017 * The grace period in which this quiescent state was
2018 * recorded has ended, so don't report it upwards.
2019 * We will instead need a new quiescent state that lies
2020 * within the current grace period.
2021 */
2022 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2023 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2024 return;
2025 }
2026 mask = rdp->grpmask;
2027 rdp->core_needs_qs = false;
2028 if ((rnp->qsmask & mask) == 0) {
2029 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2030 } else {
2031 /*
2032 * This GP can't end until cpu checks in, so all of our
2033 * callbacks can be processed during the next GP.
2034 *
2035 * NOCB kthreads have their own way to deal with that...
2036 */
2037 if (!rcu_rdp_is_offloaded(rdp)) {
2038 needwake = rcu_accelerate_cbs(rnp, rdp);
2039 } else if (!rcu_segcblist_completely_offloaded(&rdp->cblist)) {
2040 /*
2041 * ...but NOCB kthreads may miss or delay callbacks acceleration
2042 * if in the middle of a (de-)offloading process.
2043 */
2044 needacc = true;
2045 }
2046
2047 rcu_disable_urgency_upon_qs(rdp);
2048 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2049 /* ^^^ Released rnp->lock */
2050 if (needwake)
2051 rcu_gp_kthread_wake();
2052
2053 if (needacc) {
2054 rcu_nocb_lock_irqsave(rdp, flags);
2055 rcu_accelerate_cbs_unlocked(rnp, rdp);
2056 rcu_nocb_unlock_irqrestore(rdp, flags);
2057 }
2058 }
2059}
2060
2061/*
2062 * Check to see if there is a new grace period of which this CPU
2063 * is not yet aware, and if so, set up local rcu_data state for it.
2064 * Otherwise, see if this CPU has just passed through its first
2065 * quiescent state for this grace period, and record that fact if so.
2066 */
2067static void
2068rcu_check_quiescent_state(struct rcu_data *rdp)
2069{
2070 /* Check for grace-period ends and beginnings. */
2071 note_gp_changes(rdp);
2072
2073 /*
2074 * Does this CPU still need to do its part for current grace period?
2075 * If no, return and let the other CPUs do their part as well.
2076 */
2077 if (!rdp->core_needs_qs)
2078 return;
2079
2080 /*
2081 * Was there a quiescent state since the beginning of the grace
2082 * period? If no, then exit and wait for the next call.
2083 */
2084 if (rdp->cpu_no_qs.b.norm)
2085 return;
2086
2087 /*
2088 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2089 * judge of that).
2090 */
2091 rcu_report_qs_rdp(rdp);
2092}
2093
2094/*
2095 * Near the end of the offline process. Trace the fact that this CPU
2096 * is going offline.
2097 */
2098int rcutree_dying_cpu(unsigned int cpu)
2099{
2100 bool blkd;
2101 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2102 struct rcu_node *rnp = rdp->mynode;
2103
2104 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2105 return 0;
2106
2107 blkd = !!(READ_ONCE(rnp->qsmask) & rdp->grpmask);
2108 trace_rcu_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
2109 blkd ? TPS("cpuofl-bgp") : TPS("cpuofl"));
2110 return 0;
2111}
2112
2113/*
2114 * All CPUs for the specified rcu_node structure have gone offline,
2115 * and all tasks that were preempted within an RCU read-side critical
2116 * section while running on one of those CPUs have since exited their RCU
2117 * read-side critical section. Some other CPU is reporting this fact with
2118 * the specified rcu_node structure's ->lock held and interrupts disabled.
2119 * This function therefore goes up the tree of rcu_node structures,
2120 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2121 * the leaf rcu_node structure's ->qsmaskinit field has already been
2122 * updated.
2123 *
2124 * This function does check that the specified rcu_node structure has
2125 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2126 * prematurely. That said, invoking it after the fact will cost you
2127 * a needless lock acquisition. So once it has done its work, don't
2128 * invoke it again.
2129 */
2130static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2131{
2132 long mask;
2133 struct rcu_node *rnp = rnp_leaf;
2134
2135 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2136 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2137 WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
2138 WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
2139 return;
2140 for (;;) {
2141 mask = rnp->grpmask;
2142 rnp = rnp->parent;
2143 if (!rnp)
2144 break;
2145 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2146 rnp->qsmaskinit &= ~mask;
2147 /* Between grace periods, so better already be zero! */
2148 WARN_ON_ONCE(rnp->qsmask);
2149 if (rnp->qsmaskinit) {
2150 raw_spin_unlock_rcu_node(rnp);
2151 /* irqs remain disabled. */
2152 return;
2153 }
2154 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2155 }
2156}
2157
2158/*
2159 * The CPU has been completely removed, and some other CPU is reporting
2160 * this fact from process context. Do the remainder of the cleanup.
2161 * There can only be one CPU hotplug operation at a time, so no need for
2162 * explicit locking.
2163 */
2164int rcutree_dead_cpu(unsigned int cpu)
2165{
2166 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2167 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2168
2169 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2170 return 0;
2171
2172 WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus - 1);
2173 /* Adjust any no-longer-needed kthreads. */
2174 rcu_boost_kthread_setaffinity(rnp, -1);
2175 // Stop-machine done, so allow nohz_full to disable tick.
2176 tick_dep_clear(TICK_DEP_BIT_RCU);
2177 return 0;
2178}
2179
2180/*
2181 * Invoke any RCU callbacks that have made it to the end of their grace
2182 * period. Throttle as specified by rdp->blimit.
2183 */
2184static void rcu_do_batch(struct rcu_data *rdp)
2185{
2186 int div;
2187 bool __maybe_unused empty;
2188 unsigned long flags;
2189 struct rcu_head *rhp;
2190 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2191 long bl, count = 0;
2192 long pending, tlimit = 0;
2193
2194 /* If no callbacks are ready, just return. */
2195 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2196 trace_rcu_batch_start(rcu_state.name,
2197 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2198 trace_rcu_batch_end(rcu_state.name, 0,
2199 !rcu_segcblist_empty(&rdp->cblist),
2200 need_resched(), is_idle_task(current),
2201 rcu_is_callbacks_kthread(rdp));
2202 return;
2203 }
2204
2205 /*
2206 * Extract the list of ready callbacks, disabling IRQs to prevent
2207 * races with call_rcu() from interrupt handlers. Leave the
2208 * callback counts, as rcu_barrier() needs to be conservative.
2209 */
2210 rcu_nocb_lock_irqsave(rdp, flags);
2211 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2212 pending = rcu_segcblist_n_cbs(&rdp->cblist);
2213 div = READ_ONCE(rcu_divisor);
2214 div = div < 0 ? 7 : div > sizeof(long) * 8 - 2 ? sizeof(long) * 8 - 2 : div;
2215 bl = max(rdp->blimit, pending >> div);
2216 if (in_serving_softirq() && unlikely(bl > 100)) {
2217 long rrn = READ_ONCE(rcu_resched_ns);
2218
2219 rrn = rrn < NSEC_PER_MSEC ? NSEC_PER_MSEC : rrn > NSEC_PER_SEC ? NSEC_PER_SEC : rrn;
2220 tlimit = local_clock() + rrn;
2221 }
2222 trace_rcu_batch_start(rcu_state.name,
2223 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2224 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2225 if (rcu_rdp_is_offloaded(rdp))
2226 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2227
2228 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbDequeued"));
2229 rcu_nocb_unlock_irqrestore(rdp, flags);
2230
2231 /* Invoke callbacks. */
2232 tick_dep_set_task(current, TICK_DEP_BIT_RCU);
2233 rhp = rcu_cblist_dequeue(&rcl);
2234
2235 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2236 rcu_callback_t f;
2237
2238 count++;
2239 debug_rcu_head_unqueue(rhp);
2240
2241 rcu_lock_acquire(&rcu_callback_map);
2242 trace_rcu_invoke_callback(rcu_state.name, rhp);
2243
2244 f = rhp->func;
2245 WRITE_ONCE(rhp->func, (rcu_callback_t)0L);
2246 f(rhp);
2247
2248 rcu_lock_release(&rcu_callback_map);
2249
2250 /*
2251 * Stop only if limit reached and CPU has something to do.
2252 */
2253 if (in_serving_softirq()) {
2254 if (count >= bl && (need_resched() || !is_idle_task(current)))
2255 break;
2256 /*
2257 * Make sure we don't spend too much time here and deprive other
2258 * softirq vectors of CPU cycles.
2259 */
2260 if (unlikely(tlimit)) {
2261 /* only call local_clock() every 32 callbacks */
2262 if (likely((count & 31) || local_clock() < tlimit))
2263 continue;
2264 /* Exceeded the time limit, so leave. */
2265 break;
2266 }
2267 } else {
2268 local_bh_enable();
2269 lockdep_assert_irqs_enabled();
2270 cond_resched_tasks_rcu_qs();
2271 lockdep_assert_irqs_enabled();
2272 local_bh_disable();
2273 }
2274 }
2275
2276 rcu_nocb_lock_irqsave(rdp, flags);
2277 rdp->n_cbs_invoked += count;
2278 trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2279 is_idle_task(current), rcu_is_callbacks_kthread(rdp));
2280
2281 /* Update counts and requeue any remaining callbacks. */
2282 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2283 rcu_segcblist_add_len(&rdp->cblist, -count);
2284
2285 /* Reinstate batch limit if we have worked down the excess. */
2286 count = rcu_segcblist_n_cbs(&rdp->cblist);
2287 if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
2288 rdp->blimit = blimit;
2289
2290 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2291 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2292 rdp->qlen_last_fqs_check = 0;
2293 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
2294 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2295 rdp->qlen_last_fqs_check = count;
2296
2297 /*
2298 * The following usually indicates a double call_rcu(). To track
2299 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2300 */
2301 empty = rcu_segcblist_empty(&rdp->cblist);
2302 WARN_ON_ONCE(count == 0 && !empty);
2303 WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2304 count != 0 && empty);
2305 WARN_ON_ONCE(count == 0 && rcu_segcblist_n_segment_cbs(&rdp->cblist) != 0);
2306 WARN_ON_ONCE(!empty && rcu_segcblist_n_segment_cbs(&rdp->cblist) == 0);
2307
2308 rcu_nocb_unlock_irqrestore(rdp, flags);
2309
2310 tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
2311}
2312
2313/*
2314 * This function is invoked from each scheduling-clock interrupt,
2315 * and checks to see if this CPU is in a non-context-switch quiescent
2316 * state, for example, user mode or idle loop. It also schedules RCU
2317 * core processing. If the current grace period has gone on too long,
2318 * it will ask the scheduler to manufacture a context switch for the sole
2319 * purpose of providing the needed quiescent state.
2320 */
2321void rcu_sched_clock_irq(int user)
2322{
2323 unsigned long j;
2324
2325 if (IS_ENABLED(CONFIG_PROVE_RCU)) {
2326 j = jiffies;
2327 WARN_ON_ONCE(time_before(j, __this_cpu_read(rcu_data.last_sched_clock)));
2328 __this_cpu_write(rcu_data.last_sched_clock, j);
2329 }
2330 trace_rcu_utilization(TPS("Start scheduler-tick"));
2331 lockdep_assert_irqs_disabled();
2332 raw_cpu_inc(rcu_data.ticks_this_gp);
2333 /* The load-acquire pairs with the store-release setting to true. */
2334 if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
2335 /* Idle and userspace execution already are quiescent states. */
2336 if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2337 set_tsk_need_resched(current);
2338 set_preempt_need_resched();
2339 }
2340 __this_cpu_write(rcu_data.rcu_urgent_qs, false);
2341 }
2342 rcu_flavor_sched_clock_irq(user);
2343 if (rcu_pending(user))
2344 invoke_rcu_core();
2345 if (user || rcu_is_cpu_rrupt_from_idle())
2346 rcu_note_voluntary_context_switch(current);
2347 lockdep_assert_irqs_disabled();
2348
2349 trace_rcu_utilization(TPS("End scheduler-tick"));
2350}
2351
2352/*
2353 * Scan the leaf rcu_node structures. For each structure on which all
2354 * CPUs have reported a quiescent state and on which there are tasks
2355 * blocking the current grace period, initiate RCU priority boosting.
2356 * Otherwise, invoke the specified function to check dyntick state for
2357 * each CPU that has not yet reported a quiescent state.
2358 */
2359static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
2360{
2361 int cpu;
2362 unsigned long flags;
2363 unsigned long mask;
2364 struct rcu_data *rdp;
2365 struct rcu_node *rnp;
2366
2367 rcu_state.cbovld = rcu_state.cbovldnext;
2368 rcu_state.cbovldnext = false;
2369 rcu_for_each_leaf_node(rnp) {
2370 cond_resched_tasks_rcu_qs();
2371 mask = 0;
2372 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2373 rcu_state.cbovldnext |= !!rnp->cbovldmask;
2374 if (rnp->qsmask == 0) {
2375 if (rcu_preempt_blocked_readers_cgp(rnp)) {
2376 /*
2377 * No point in scanning bits because they
2378 * are all zero. But we might need to
2379 * priority-boost blocked readers.
2380 */
2381 rcu_initiate_boost(rnp, flags);
2382 /* rcu_initiate_boost() releases rnp->lock */
2383 continue;
2384 }
2385 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2386 continue;
2387 }
2388 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->qsmask) {
2389 rdp = per_cpu_ptr(&rcu_data, cpu);
2390 if (f(rdp)) {
2391 mask |= rdp->grpmask;
2392 rcu_disable_urgency_upon_qs(rdp);
2393 }
2394 }
2395 if (mask != 0) {
2396 /* Idle/offline CPUs, report (releases rnp->lock). */
2397 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2398 } else {
2399 /* Nothing to do here, so just drop the lock. */
2400 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2401 }
2402 }
2403}
2404
2405/*
2406 * Force quiescent states on reluctant CPUs, and also detect which
2407 * CPUs are in dyntick-idle mode.
2408 */
2409void rcu_force_quiescent_state(void)
2410{
2411 unsigned long flags;
2412 bool ret;
2413 struct rcu_node *rnp;
2414 struct rcu_node *rnp_old = NULL;
2415
2416 /* Funnel through hierarchy to reduce memory contention. */
2417 rnp = raw_cpu_read(rcu_data.mynode);
2418 for (; rnp != NULL; rnp = rnp->parent) {
2419 ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2420 !raw_spin_trylock(&rnp->fqslock);
2421 if (rnp_old != NULL)
2422 raw_spin_unlock(&rnp_old->fqslock);
2423 if (ret)
2424 return;
2425 rnp_old = rnp;
2426 }
2427 /* rnp_old == rcu_get_root(), rnp == NULL. */
2428
2429 /* Reached the root of the rcu_node tree, acquire lock. */
2430 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2431 raw_spin_unlock(&rnp_old->fqslock);
2432 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2433 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2434 return; /* Someone beat us to it. */
2435 }
2436 WRITE_ONCE(rcu_state.gp_flags,
2437 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2438 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2439 rcu_gp_kthread_wake();
2440}
2441EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
2442
2443// Workqueue handler for an RCU reader for kernels enforcing struct RCU
2444// grace periods.
2445static void strict_work_handler(struct work_struct *work)
2446{
2447 rcu_read_lock();
2448 rcu_read_unlock();
2449}
2450
2451/* Perform RCU core processing work for the current CPU. */
2452static __latent_entropy void rcu_core(void)
2453{
2454 unsigned long flags;
2455 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2456 struct rcu_node *rnp = rdp->mynode;
2457 /*
2458 * On RT rcu_core() can be preempted when IRQs aren't disabled.
2459 * Therefore this function can race with concurrent NOCB (de-)offloading
2460 * on this CPU and the below condition must be considered volatile.
2461 * However if we race with:
2462 *
2463 * _ Offloading: In the worst case we accelerate or process callbacks
2464 * concurrently with NOCB kthreads. We are guaranteed to
2465 * call rcu_nocb_lock() if that happens.
2466 *
2467 * _ Deoffloading: In the worst case we miss callbacks acceleration or
2468 * processing. This is fine because the early stage
2469 * of deoffloading invokes rcu_core() after setting
2470 * SEGCBLIST_RCU_CORE. So we guarantee that we'll process
2471 * what could have been dismissed without the need to wait
2472 * for the next rcu_pending() check in the next jiffy.
2473 */
2474 const bool do_batch = !rcu_segcblist_completely_offloaded(&rdp->cblist);
2475
2476 if (cpu_is_offline(smp_processor_id()))
2477 return;
2478 trace_rcu_utilization(TPS("Start RCU core"));
2479 WARN_ON_ONCE(!rdp->beenonline);
2480
2481 /* Report any deferred quiescent states if preemption enabled. */
2482 if (IS_ENABLED(CONFIG_PREEMPT_COUNT) && (!(preempt_count() & PREEMPT_MASK))) {
2483 rcu_preempt_deferred_qs(current);
2484 } else if (rcu_preempt_need_deferred_qs(current)) {
2485 set_tsk_need_resched(current);
2486 set_preempt_need_resched();
2487 }
2488
2489 /* Update RCU state based on any recent quiescent states. */
2490 rcu_check_quiescent_state(rdp);
2491
2492 /* No grace period and unregistered callbacks? */
2493 if (!rcu_gp_in_progress() &&
2494 rcu_segcblist_is_enabled(&rdp->cblist) && do_batch) {
2495 rcu_nocb_lock_irqsave(rdp, flags);
2496 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2497 rcu_accelerate_cbs_unlocked(rnp, rdp);
2498 rcu_nocb_unlock_irqrestore(rdp, flags);
2499 }
2500
2501 rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
2502
2503 /* If there are callbacks ready, invoke them. */
2504 if (do_batch && rcu_segcblist_ready_cbs(&rdp->cblist) &&
2505 likely(READ_ONCE(rcu_scheduler_fully_active))) {
2506 rcu_do_batch(rdp);
2507 /* Re-invoke RCU core processing if there are callbacks remaining. */
2508 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2509 invoke_rcu_core();
2510 }
2511
2512 /* Do any needed deferred wakeups of rcuo kthreads. */
2513 do_nocb_deferred_wakeup(rdp);
2514 trace_rcu_utilization(TPS("End RCU core"));
2515
2516 // If strict GPs, schedule an RCU reader in a clean environment.
2517 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
2518 queue_work_on(rdp->cpu, rcu_gp_wq, &rdp->strict_work);
2519}
2520
2521static void rcu_core_si(struct softirq_action *h)
2522{
2523 rcu_core();
2524}
2525
2526static void rcu_wake_cond(struct task_struct *t, int status)
2527{
2528 /*
2529 * If the thread is yielding, only wake it when this
2530 * is invoked from idle
2531 */
2532 if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
2533 wake_up_process(t);
2534}
2535
2536static void invoke_rcu_core_kthread(void)
2537{
2538 struct task_struct *t;
2539 unsigned long flags;
2540
2541 local_irq_save(flags);
2542 __this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
2543 t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
2544 if (t != NULL && t != current)
2545 rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
2546 local_irq_restore(flags);
2547}
2548
2549/*
2550 * Wake up this CPU's rcuc kthread to do RCU core processing.
2551 */
2552static void invoke_rcu_core(void)
2553{
2554 if (!cpu_online(smp_processor_id()))
2555 return;
2556 if (use_softirq)
2557 raise_softirq(RCU_SOFTIRQ);
2558 else
2559 invoke_rcu_core_kthread();
2560}
2561
2562static void rcu_cpu_kthread_park(unsigned int cpu)
2563{
2564 per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
2565}
2566
2567static int rcu_cpu_kthread_should_run(unsigned int cpu)
2568{
2569 return __this_cpu_read(rcu_data.rcu_cpu_has_work);
2570}
2571
2572/*
2573 * Per-CPU kernel thread that invokes RCU callbacks. This replaces
2574 * the RCU softirq used in configurations of RCU that do not support RCU
2575 * priority boosting.
2576 */
2577static void rcu_cpu_kthread(unsigned int cpu)
2578{
2579 unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
2580 char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
2581 unsigned long *j = this_cpu_ptr(&rcu_data.rcuc_activity);
2582 int spincnt;
2583
2584 trace_rcu_utilization(TPS("Start CPU kthread@rcu_run"));
2585 for (spincnt = 0; spincnt < 10; spincnt++) {
2586 WRITE_ONCE(*j, jiffies);
2587 local_bh_disable();
2588 *statusp = RCU_KTHREAD_RUNNING;
2589 local_irq_disable();
2590 work = *workp;
2591 *workp = 0;
2592 local_irq_enable();
2593 if (work)
2594 rcu_core();
2595 local_bh_enable();
2596 if (*workp == 0) {
2597 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2598 *statusp = RCU_KTHREAD_WAITING;
2599 return;
2600 }
2601 }
2602 *statusp = RCU_KTHREAD_YIELDING;
2603 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2604 schedule_timeout_idle(2);
2605 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2606 *statusp = RCU_KTHREAD_WAITING;
2607 WRITE_ONCE(*j, jiffies);
2608}
2609
2610static struct smp_hotplug_thread rcu_cpu_thread_spec = {
2611 .store = &rcu_data.rcu_cpu_kthread_task,
2612 .thread_should_run = rcu_cpu_kthread_should_run,
2613 .thread_fn = rcu_cpu_kthread,
2614 .thread_comm = "rcuc/%u",
2615 .setup = rcu_cpu_kthread_setup,
2616 .park = rcu_cpu_kthread_park,
2617};
2618
2619/*
2620 * Spawn per-CPU RCU core processing kthreads.
2621 */
2622static int __init rcu_spawn_core_kthreads(void)
2623{
2624 int cpu;
2625
2626 for_each_possible_cpu(cpu)
2627 per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
2628 if (use_softirq)
2629 return 0;
2630 WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
2631 "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
2632 return 0;
2633}
2634
2635/*
2636 * Handle any core-RCU processing required by a call_rcu() invocation.
2637 */
2638static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2639 unsigned long flags)
2640{
2641 /*
2642 * If called from an extended quiescent state, invoke the RCU
2643 * core in order to force a re-evaluation of RCU's idleness.
2644 */
2645 if (!rcu_is_watching())
2646 invoke_rcu_core();
2647
2648 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2649 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2650 return;
2651
2652 /*
2653 * Force the grace period if too many callbacks or too long waiting.
2654 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2655 * if some other CPU has recently done so. Also, don't bother
2656 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2657 * is the only one waiting for a grace period to complete.
2658 */
2659 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2660 rdp->qlen_last_fqs_check + qhimark)) {
2661
2662 /* Are we ignoring a completed grace period? */
2663 note_gp_changes(rdp);
2664
2665 /* Start a new grace period if one not already started. */
2666 if (!rcu_gp_in_progress()) {
2667 rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
2668 } else {
2669 /* Give the grace period a kick. */
2670 rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
2671 if (READ_ONCE(rcu_state.n_force_qs) == rdp->n_force_qs_snap &&
2672 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2673 rcu_force_quiescent_state();
2674 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
2675 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2676 }
2677 }
2678}
2679
2680/*
2681 * RCU callback function to leak a callback.
2682 */
2683static void rcu_leak_callback(struct rcu_head *rhp)
2684{
2685}
2686
2687/*
2688 * Check and if necessary update the leaf rcu_node structure's
2689 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2690 * number of queued RCU callbacks. The caller must hold the leaf rcu_node
2691 * structure's ->lock.
2692 */
2693static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp)
2694{
2695 raw_lockdep_assert_held_rcu_node(rnp);
2696 if (qovld_calc <= 0)
2697 return; // Early boot and wildcard value set.
2698 if (rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc)
2699 WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask | rdp->grpmask);
2700 else
2701 WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask & ~rdp->grpmask);
2702}
2703
2704/*
2705 * Check and if necessary update the leaf rcu_node structure's
2706 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2707 * number of queued RCU callbacks. No locks need be held, but the
2708 * caller must have disabled interrupts.
2709 *
2710 * Note that this function ignores the possibility that there are a lot
2711 * of callbacks all of which have already seen the end of their respective
2712 * grace periods. This omission is due to the need for no-CBs CPUs to
2713 * be holding ->nocb_lock to do this check, which is too heavy for a
2714 * common-case operation.
2715 */
2716static void check_cb_ovld(struct rcu_data *rdp)
2717{
2718 struct rcu_node *const rnp = rdp->mynode;
2719
2720 if (qovld_calc <= 0 ||
2721 ((rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) ==
2722 !!(READ_ONCE(rnp->cbovldmask) & rdp->grpmask)))
2723 return; // Early boot wildcard value or already set correctly.
2724 raw_spin_lock_rcu_node(rnp);
2725 check_cb_ovld_locked(rdp, rnp);
2726 raw_spin_unlock_rcu_node(rnp);
2727}
2728
2729static void
2730__call_rcu_common(struct rcu_head *head, rcu_callback_t func, bool lazy)
2731{
2732 static atomic_t doublefrees;
2733 unsigned long flags;
2734 struct rcu_data *rdp;
2735 bool was_alldone;
2736
2737 /* Misaligned rcu_head! */
2738 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2739
2740 if (debug_rcu_head_queue(head)) {
2741 /*
2742 * Probable double call_rcu(), so leak the callback.
2743 * Use rcu:rcu_callback trace event to find the previous
2744 * time callback was passed to call_rcu().
2745 */
2746 if (atomic_inc_return(&doublefrees) < 4) {
2747 pr_err("%s(): Double-freed CB %p->%pS()!!! ", __func__, head, head->func);
2748 mem_dump_obj(head);
2749 }
2750 WRITE_ONCE(head->func, rcu_leak_callback);
2751 return;
2752 }
2753 head->func = func;
2754 head->next = NULL;
2755 kasan_record_aux_stack_noalloc(head);
2756 local_irq_save(flags);
2757 rdp = this_cpu_ptr(&rcu_data);
2758
2759 /* Add the callback to our list. */
2760 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
2761 // This can trigger due to call_rcu() from offline CPU:
2762 WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
2763 WARN_ON_ONCE(!rcu_is_watching());
2764 // Very early boot, before rcu_init(). Initialize if needed
2765 // and then drop through to queue the callback.
2766 if (rcu_segcblist_empty(&rdp->cblist))
2767 rcu_segcblist_init(&rdp->cblist);
2768 }
2769
2770 check_cb_ovld(rdp);
2771 if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags, lazy))
2772 return; // Enqueued onto ->nocb_bypass, so just leave.
2773 // If no-CBs CPU gets here, rcu_nocb_try_bypass() acquired ->nocb_lock.
2774 rcu_segcblist_enqueue(&rdp->cblist, head);
2775 if (__is_kvfree_rcu_offset((unsigned long)func))
2776 trace_rcu_kvfree_callback(rcu_state.name, head,
2777 (unsigned long)func,
2778 rcu_segcblist_n_cbs(&rdp->cblist));
2779 else
2780 trace_rcu_callback(rcu_state.name, head,
2781 rcu_segcblist_n_cbs(&rdp->cblist));
2782
2783 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCBQueued"));
2784
2785 /* Go handle any RCU core processing required. */
2786 if (unlikely(rcu_rdp_is_offloaded(rdp))) {
2787 __call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */
2788 } else {
2789 __call_rcu_core(rdp, head, flags);
2790 local_irq_restore(flags);
2791 }
2792}
2793
2794#ifdef CONFIG_RCU_LAZY
2795/**
2796 * call_rcu_hurry() - Queue RCU callback for invocation after grace period, and
2797 * flush all lazy callbacks (including the new one) to the main ->cblist while
2798 * doing so.
2799 *
2800 * @head: structure to be used for queueing the RCU updates.
2801 * @func: actual callback function to be invoked after the grace period
2802 *
2803 * The callback function will be invoked some time after a full grace
2804 * period elapses, in other words after all pre-existing RCU read-side
2805 * critical sections have completed.
2806 *
2807 * Use this API instead of call_rcu() if you don't want the callback to be
2808 * invoked after very long periods of time, which can happen on systems without
2809 * memory pressure and on systems which are lightly loaded or mostly idle.
2810 * This function will cause callbacks to be invoked sooner than later at the
2811 * expense of extra power. Other than that, this function is identical to, and
2812 * reuses call_rcu()'s logic. Refer to call_rcu() for more details about memory
2813 * ordering and other functionality.
2814 */
2815void call_rcu_hurry(struct rcu_head *head, rcu_callback_t func)
2816{
2817 return __call_rcu_common(head, func, false);
2818}
2819EXPORT_SYMBOL_GPL(call_rcu_hurry);
2820#endif
2821
2822/**
2823 * call_rcu() - Queue an RCU callback for invocation after a grace period.
2824 * By default the callbacks are 'lazy' and are kept hidden from the main
2825 * ->cblist to prevent starting of grace periods too soon.
2826 * If you desire grace periods to start very soon, use call_rcu_hurry().
2827 *
2828 * @head: structure to be used for queueing the RCU updates.
2829 * @func: actual callback function to be invoked after the grace period
2830 *
2831 * The callback function will be invoked some time after a full grace
2832 * period elapses, in other words after all pre-existing RCU read-side
2833 * critical sections have completed. However, the callback function
2834 * might well execute concurrently with RCU read-side critical sections
2835 * that started after call_rcu() was invoked.
2836 *
2837 * RCU read-side critical sections are delimited by rcu_read_lock()
2838 * and rcu_read_unlock(), and may be nested. In addition, but only in
2839 * v5.0 and later, regions of code across which interrupts, preemption,
2840 * or softirqs have been disabled also serve as RCU read-side critical
2841 * sections. This includes hardware interrupt handlers, softirq handlers,
2842 * and NMI handlers.
2843 *
2844 * Note that all CPUs must agree that the grace period extended beyond
2845 * all pre-existing RCU read-side critical section. On systems with more
2846 * than one CPU, this means that when "func()" is invoked, each CPU is
2847 * guaranteed to have executed a full memory barrier since the end of its
2848 * last RCU read-side critical section whose beginning preceded the call
2849 * to call_rcu(). It also means that each CPU executing an RCU read-side
2850 * critical section that continues beyond the start of "func()" must have
2851 * executed a memory barrier after the call_rcu() but before the beginning
2852 * of that RCU read-side critical section. Note that these guarantees
2853 * include CPUs that are offline, idle, or executing in user mode, as
2854 * well as CPUs that are executing in the kernel.
2855 *
2856 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
2857 * resulting RCU callback function "func()", then both CPU A and CPU B are
2858 * guaranteed to execute a full memory barrier during the time interval
2859 * between the call to call_rcu() and the invocation of "func()" -- even
2860 * if CPU A and CPU B are the same CPU (but again only if the system has
2861 * more than one CPU).
2862 *
2863 * Implementation of these memory-ordering guarantees is described here:
2864 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
2865 */
2866void call_rcu(struct rcu_head *head, rcu_callback_t func)
2867{
2868 return __call_rcu_common(head, func, IS_ENABLED(CONFIG_RCU_LAZY));
2869}
2870EXPORT_SYMBOL_GPL(call_rcu);
2871
2872/* Maximum number of jiffies to wait before draining a batch. */
2873#define KFREE_DRAIN_JIFFIES (5 * HZ)
2874#define KFREE_N_BATCHES 2
2875#define FREE_N_CHANNELS 2
2876
2877/**
2878 * struct kvfree_rcu_bulk_data - single block to store kvfree_rcu() pointers
2879 * @nr_records: Number of active pointers in the array
2880 * @next: Next bulk object in the block chain
2881 * @records: Array of the kvfree_rcu() pointers
2882 */
2883struct kvfree_rcu_bulk_data {
2884 unsigned long nr_records;
2885 struct kvfree_rcu_bulk_data *next;
2886 void *records[];
2887};
2888
2889/*
2890 * This macro defines how many entries the "records" array
2891 * will contain. It is based on the fact that the size of
2892 * kvfree_rcu_bulk_data structure becomes exactly one page.
2893 */
2894#define KVFREE_BULK_MAX_ENTR \
2895 ((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *))
2896
2897/**
2898 * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
2899 * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
2900 * @head_free: List of kfree_rcu() objects waiting for a grace period
2901 * @bkvhead_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
2902 * @krcp: Pointer to @kfree_rcu_cpu structure
2903 */
2904
2905struct kfree_rcu_cpu_work {
2906 struct rcu_work rcu_work;
2907 struct rcu_head *head_free;
2908 struct kvfree_rcu_bulk_data *bkvhead_free[FREE_N_CHANNELS];
2909 struct kfree_rcu_cpu *krcp;
2910};
2911
2912/**
2913 * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
2914 * @head: List of kfree_rcu() objects not yet waiting for a grace period
2915 * @bkvhead: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period
2916 * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
2917 * @lock: Synchronize access to this structure
2918 * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES
2919 * @initialized: The @rcu_work fields have been initialized
2920 * @count: Number of objects for which GP not started
2921 * @bkvcache:
2922 * A simple cache list that contains objects for reuse purpose.
2923 * In order to save some per-cpu space the list is singular.
2924 * Even though it is lockless an access has to be protected by the
2925 * per-cpu lock.
2926 * @page_cache_work: A work to refill the cache when it is empty
2927 * @backoff_page_cache_fill: Delay cache refills
2928 * @work_in_progress: Indicates that page_cache_work is running
2929 * @hrtimer: A hrtimer for scheduling a page_cache_work
2930 * @nr_bkv_objs: number of allocated objects at @bkvcache.
2931 *
2932 * This is a per-CPU structure. The reason that it is not included in
2933 * the rcu_data structure is to permit this code to be extracted from
2934 * the RCU files. Such extraction could allow further optimization of
2935 * the interactions with the slab allocators.
2936 */
2937struct kfree_rcu_cpu {
2938 struct rcu_head *head;
2939 struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS];
2940 struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
2941 raw_spinlock_t lock;
2942 struct delayed_work monitor_work;
2943 bool initialized;
2944 int count;
2945
2946 struct delayed_work page_cache_work;
2947 atomic_t backoff_page_cache_fill;
2948 atomic_t work_in_progress;
2949 struct hrtimer hrtimer;
2950
2951 struct llist_head bkvcache;
2952 int nr_bkv_objs;
2953};
2954
2955static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc) = {
2956 .lock = __RAW_SPIN_LOCK_UNLOCKED(krc.lock),
2957};
2958
2959static __always_inline void
2960debug_rcu_bhead_unqueue(struct kvfree_rcu_bulk_data *bhead)
2961{
2962#ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
2963 int i;
2964
2965 for (i = 0; i < bhead->nr_records; i++)
2966 debug_rcu_head_unqueue((struct rcu_head *)(bhead->records[i]));
2967#endif
2968}
2969
2970static inline struct kfree_rcu_cpu *
2971krc_this_cpu_lock(unsigned long *flags)
2972{
2973 struct kfree_rcu_cpu *krcp;
2974
2975 local_irq_save(*flags); // For safely calling this_cpu_ptr().
2976 krcp = this_cpu_ptr(&krc);
2977 raw_spin_lock(&krcp->lock);
2978
2979 return krcp;
2980}
2981
2982static inline void
2983krc_this_cpu_unlock(struct kfree_rcu_cpu *krcp, unsigned long flags)
2984{
2985 raw_spin_unlock_irqrestore(&krcp->lock, flags);
2986}
2987
2988static inline struct kvfree_rcu_bulk_data *
2989get_cached_bnode(struct kfree_rcu_cpu *krcp)
2990{
2991 if (!krcp->nr_bkv_objs)
2992 return NULL;
2993
2994 WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs - 1);
2995 return (struct kvfree_rcu_bulk_data *)
2996 llist_del_first(&krcp->bkvcache);
2997}
2998
2999static inline bool
3000put_cached_bnode(struct kfree_rcu_cpu *krcp,
3001 struct kvfree_rcu_bulk_data *bnode)
3002{
3003 // Check the limit.
3004 if (krcp->nr_bkv_objs >= rcu_min_cached_objs)
3005 return false;
3006
3007 llist_add((struct llist_node *) bnode, &krcp->bkvcache);
3008 WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs + 1);
3009 return true;
3010}
3011
3012static int
3013drain_page_cache(struct kfree_rcu_cpu *krcp)
3014{
3015 unsigned long flags;
3016 struct llist_node *page_list, *pos, *n;
3017 int freed = 0;
3018
3019 raw_spin_lock_irqsave(&krcp->lock, flags);
3020 page_list = llist_del_all(&krcp->bkvcache);
3021 WRITE_ONCE(krcp->nr_bkv_objs, 0);
3022 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3023
3024 llist_for_each_safe(pos, n, page_list) {
3025 free_page((unsigned long)pos);
3026 freed++;
3027 }
3028
3029 return freed;
3030}
3031
3032/*
3033 * This function is invoked in workqueue context after a grace period.
3034 * It frees all the objects queued on ->bkvhead_free or ->head_free.
3035 */
3036static void kfree_rcu_work(struct work_struct *work)
3037{
3038 unsigned long flags;
3039 struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS], *bnext;
3040 struct rcu_head *head, *next;
3041 struct kfree_rcu_cpu *krcp;
3042 struct kfree_rcu_cpu_work *krwp;
3043 int i, j;
3044
3045 krwp = container_of(to_rcu_work(work),
3046 struct kfree_rcu_cpu_work, rcu_work);
3047 krcp = krwp->krcp;
3048
3049 raw_spin_lock_irqsave(&krcp->lock, flags);
3050 // Channels 1 and 2.
3051 for (i = 0; i < FREE_N_CHANNELS; i++) {
3052 bkvhead[i] = krwp->bkvhead_free[i];
3053 krwp->bkvhead_free[i] = NULL;
3054 }
3055
3056 // Channel 3.
3057 head = krwp->head_free;
3058 krwp->head_free = NULL;
3059 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3060
3061 // Handle the first two channels.
3062 for (i = 0; i < FREE_N_CHANNELS; i++) {
3063 for (; bkvhead[i]; bkvhead[i] = bnext) {
3064 bnext = bkvhead[i]->next;
3065 debug_rcu_bhead_unqueue(bkvhead[i]);
3066
3067 rcu_lock_acquire(&rcu_callback_map);
3068 if (i == 0) { // kmalloc() / kfree().
3069 trace_rcu_invoke_kfree_bulk_callback(
3070 rcu_state.name, bkvhead[i]->nr_records,
3071 bkvhead[i]->records);
3072
3073 kfree_bulk(bkvhead[i]->nr_records,
3074 bkvhead[i]->records);
3075 } else { // vmalloc() / vfree().
3076 for (j = 0; j < bkvhead[i]->nr_records; j++) {
3077 trace_rcu_invoke_kvfree_callback(
3078 rcu_state.name,
3079 bkvhead[i]->records[j], 0);
3080
3081 vfree(bkvhead[i]->records[j]);
3082 }
3083 }
3084 rcu_lock_release(&rcu_callback_map);
3085
3086 raw_spin_lock_irqsave(&krcp->lock, flags);
3087 if (put_cached_bnode(krcp, bkvhead[i]))
3088 bkvhead[i] = NULL;
3089 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3090
3091 if (bkvhead[i])
3092 free_page((unsigned long) bkvhead[i]);
3093
3094 cond_resched_tasks_rcu_qs();
3095 }
3096 }
3097
3098 /*
3099 * This is used when the "bulk" path can not be used for the
3100 * double-argument of kvfree_rcu(). This happens when the
3101 * page-cache is empty, which means that objects are instead
3102 * queued on a linked list through their rcu_head structures.
3103 * This list is named "Channel 3".
3104 */
3105 for (; head; head = next) {
3106 unsigned long offset = (unsigned long)head->func;
3107 void *ptr = (void *)head - offset;
3108
3109 next = head->next;
3110 debug_rcu_head_unqueue((struct rcu_head *)ptr);
3111 rcu_lock_acquire(&rcu_callback_map);
3112 trace_rcu_invoke_kvfree_callback(rcu_state.name, head, offset);
3113
3114 if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset)))
3115 kvfree(ptr);
3116
3117 rcu_lock_release(&rcu_callback_map);
3118 cond_resched_tasks_rcu_qs();
3119 }
3120}
3121
3122static bool
3123need_offload_krc(struct kfree_rcu_cpu *krcp)
3124{
3125 int i;
3126
3127 for (i = 0; i < FREE_N_CHANNELS; i++)
3128 if (krcp->bkvhead[i])
3129 return true;
3130
3131 return !!krcp->head;
3132}
3133
3134static void
3135schedule_delayed_monitor_work(struct kfree_rcu_cpu *krcp)
3136{
3137 long delay, delay_left;
3138
3139 delay = READ_ONCE(krcp->count) >= KVFREE_BULK_MAX_ENTR ? 1:KFREE_DRAIN_JIFFIES;
3140 if (delayed_work_pending(&krcp->monitor_work)) {
3141 delay_left = krcp->monitor_work.timer.expires - jiffies;
3142 if (delay < delay_left)
3143 mod_delayed_work(system_wq, &krcp->monitor_work, delay);
3144 return;
3145 }
3146 queue_delayed_work(system_wq, &krcp->monitor_work, delay);
3147}
3148
3149/*
3150 * This function is invoked after the KFREE_DRAIN_JIFFIES timeout.
3151 */
3152static void kfree_rcu_monitor(struct work_struct *work)
3153{
3154 struct kfree_rcu_cpu *krcp = container_of(work,
3155 struct kfree_rcu_cpu, monitor_work.work);
3156 unsigned long flags;
3157 int i, j;
3158
3159 raw_spin_lock_irqsave(&krcp->lock, flags);
3160
3161 // Attempt to start a new batch.
3162 for (i = 0; i < KFREE_N_BATCHES; i++) {
3163 struct kfree_rcu_cpu_work *krwp = &(krcp->krw_arr[i]);
3164
3165 // Try to detach bkvhead or head and attach it over any
3166 // available corresponding free channel. It can be that
3167 // a previous RCU batch is in progress, it means that
3168 // immediately to queue another one is not possible so
3169 // in that case the monitor work is rearmed.
3170 if ((krcp->bkvhead[0] && !krwp->bkvhead_free[0]) ||
3171 (krcp->bkvhead[1] && !krwp->bkvhead_free[1]) ||
3172 (krcp->head && !krwp->head_free)) {
3173 // Channel 1 corresponds to the SLAB-pointer bulk path.
3174 // Channel 2 corresponds to vmalloc-pointer bulk path.
3175 for (j = 0; j < FREE_N_CHANNELS; j++) {
3176 if (!krwp->bkvhead_free[j]) {
3177 krwp->bkvhead_free[j] = krcp->bkvhead[j];
3178 krcp->bkvhead[j] = NULL;
3179 }
3180 }
3181
3182 // Channel 3 corresponds to both SLAB and vmalloc
3183 // objects queued on the linked list.
3184 if (!krwp->head_free) {
3185 krwp->head_free = krcp->head;
3186 krcp->head = NULL;
3187 }
3188
3189 WRITE_ONCE(krcp->count, 0);
3190
3191 // One work is per one batch, so there are three
3192 // "free channels", the batch can handle. It can
3193 // be that the work is in the pending state when
3194 // channels have been detached following by each
3195 // other.
3196 queue_rcu_work(system_wq, &krwp->rcu_work);
3197 }
3198 }
3199
3200 // If there is nothing to detach, it means that our job is
3201 // successfully done here. In case of having at least one
3202 // of the channels that is still busy we should rearm the
3203 // work to repeat an attempt. Because previous batches are
3204 // still in progress.
3205 if (need_offload_krc(krcp))
3206 schedule_delayed_monitor_work(krcp);
3207
3208 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3209}
3210
3211static enum hrtimer_restart
3212schedule_page_work_fn(struct hrtimer *t)
3213{
3214 struct kfree_rcu_cpu *krcp =
3215 container_of(t, struct kfree_rcu_cpu, hrtimer);
3216
3217 queue_delayed_work(system_highpri_wq, &krcp->page_cache_work, 0);
3218 return HRTIMER_NORESTART;
3219}
3220
3221static void fill_page_cache_func(struct work_struct *work)
3222{
3223 struct kvfree_rcu_bulk_data *bnode;
3224 struct kfree_rcu_cpu *krcp =
3225 container_of(work, struct kfree_rcu_cpu,
3226 page_cache_work.work);
3227 unsigned long flags;
3228 int nr_pages;
3229 bool pushed;
3230 int i;
3231
3232 nr_pages = atomic_read(&krcp->backoff_page_cache_fill) ?
3233 1 : rcu_min_cached_objs;
3234
3235 for (i = 0; i < nr_pages; i++) {
3236 bnode = (struct kvfree_rcu_bulk_data *)
3237 __get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3238
3239 if (!bnode)
3240 break;
3241
3242 raw_spin_lock_irqsave(&krcp->lock, flags);
3243 pushed = put_cached_bnode(krcp, bnode);
3244 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3245
3246 if (!pushed) {
3247 free_page((unsigned long) bnode);
3248 break;
3249 }
3250 }
3251
3252 atomic_set(&krcp->work_in_progress, 0);
3253 atomic_set(&krcp->backoff_page_cache_fill, 0);
3254}
3255
3256static void
3257run_page_cache_worker(struct kfree_rcu_cpu *krcp)
3258{
3259 if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
3260 !atomic_xchg(&krcp->work_in_progress, 1)) {
3261 if (atomic_read(&krcp->backoff_page_cache_fill)) {
3262 queue_delayed_work(system_wq,
3263 &krcp->page_cache_work,
3264 msecs_to_jiffies(rcu_delay_page_cache_fill_msec));
3265 } else {
3266 hrtimer_init(&krcp->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3267 krcp->hrtimer.function = schedule_page_work_fn;
3268 hrtimer_start(&krcp->hrtimer, 0, HRTIMER_MODE_REL);
3269 }
3270 }
3271}
3272
3273// Record ptr in a page managed by krcp, with the pre-krc_this_cpu_lock()
3274// state specified by flags. If can_alloc is true, the caller must
3275// be schedulable and not be holding any locks or mutexes that might be
3276// acquired by the memory allocator or anything that it might invoke.
3277// Returns true if ptr was successfully recorded, else the caller must
3278// use a fallback.
3279static inline bool
3280add_ptr_to_bulk_krc_lock(struct kfree_rcu_cpu **krcp,
3281 unsigned long *flags, void *ptr, bool can_alloc)
3282{
3283 struct kvfree_rcu_bulk_data *bnode;
3284 int idx;
3285
3286 *krcp = krc_this_cpu_lock(flags);
3287 if (unlikely(!(*krcp)->initialized))
3288 return false;
3289
3290 idx = !!is_vmalloc_addr(ptr);
3291
3292 /* Check if a new block is required. */
3293 if (!(*krcp)->bkvhead[idx] ||
3294 (*krcp)->bkvhead[idx]->nr_records == KVFREE_BULK_MAX_ENTR) {
3295 bnode = get_cached_bnode(*krcp);
3296 if (!bnode && can_alloc) {
3297 krc_this_cpu_unlock(*krcp, *flags);
3298
3299 // __GFP_NORETRY - allows a light-weight direct reclaim
3300 // what is OK from minimizing of fallback hitting point of
3301 // view. Apart of that it forbids any OOM invoking what is
3302 // also beneficial since we are about to release memory soon.
3303 //
3304 // __GFP_NOMEMALLOC - prevents from consuming of all the
3305 // memory reserves. Please note we have a fallback path.
3306 //
3307 // __GFP_NOWARN - it is supposed that an allocation can
3308 // be failed under low memory or high memory pressure
3309 // scenarios.
3310 bnode = (struct kvfree_rcu_bulk_data *)
3311 __get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3312 *krcp = krc_this_cpu_lock(flags);
3313 }
3314
3315 if (!bnode)
3316 return false;
3317
3318 /* Initialize the new block. */
3319 bnode->nr_records = 0;
3320 bnode->next = (*krcp)->bkvhead[idx];
3321
3322 /* Attach it to the head. */
3323 (*krcp)->bkvhead[idx] = bnode;
3324 }
3325
3326 /* Finally insert. */
3327 (*krcp)->bkvhead[idx]->records
3328 [(*krcp)->bkvhead[idx]->nr_records++] = ptr;
3329
3330 return true;
3331}
3332
3333/*
3334 * Queue a request for lazy invocation of the appropriate free routine
3335 * after a grace period. Please note that three paths are maintained,
3336 * two for the common case using arrays of pointers and a third one that
3337 * is used only when the main paths cannot be used, for example, due to
3338 * memory pressure.
3339 *
3340 * Each kvfree_call_rcu() request is added to a batch. The batch will be drained
3341 * every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch will
3342 * be free'd in workqueue context. This allows us to: batch requests together to
3343 * reduce the number of grace periods during heavy kfree_rcu()/kvfree_rcu() load.
3344 */
3345void kvfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
3346{
3347 unsigned long flags;
3348 struct kfree_rcu_cpu *krcp;
3349 bool success;
3350 void *ptr;
3351
3352 if (head) {
3353 ptr = (void *) head - (unsigned long) func;
3354 } else {
3355 /*
3356 * Please note there is a limitation for the head-less
3357 * variant, that is why there is a clear rule for such
3358 * objects: it can be used from might_sleep() context
3359 * only. For other places please embed an rcu_head to
3360 * your data.
3361 */
3362 might_sleep();
3363 ptr = (unsigned long *) func;
3364 }
3365
3366 // Queue the object but don't yet schedule the batch.
3367 if (debug_rcu_head_queue(ptr)) {
3368 // Probable double kfree_rcu(), just leak.
3369 WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n",
3370 __func__, head);
3371
3372 // Mark as success and leave.
3373 return;
3374 }
3375
3376 kasan_record_aux_stack_noalloc(ptr);
3377 success = add_ptr_to_bulk_krc_lock(&krcp, &flags, ptr, !head);
3378 if (!success) {
3379 run_page_cache_worker(krcp);
3380
3381 if (head == NULL)
3382 // Inline if kvfree_rcu(one_arg) call.
3383 goto unlock_return;
3384
3385 head->func = func;
3386 head->next = krcp->head;
3387 krcp->head = head;
3388 success = true;
3389 }
3390
3391 WRITE_ONCE(krcp->count, krcp->count + 1);
3392
3393 // Set timer to drain after KFREE_DRAIN_JIFFIES.
3394 if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING)
3395 schedule_delayed_monitor_work(krcp);
3396
3397unlock_return:
3398 krc_this_cpu_unlock(krcp, flags);
3399
3400 /*
3401 * Inline kvfree() after synchronize_rcu(). We can do
3402 * it from might_sleep() context only, so the current
3403 * CPU can pass the QS state.
3404 */
3405 if (!success) {
3406 debug_rcu_head_unqueue((struct rcu_head *) ptr);
3407 synchronize_rcu();
3408 kvfree(ptr);
3409 }
3410}
3411EXPORT_SYMBOL_GPL(kvfree_call_rcu);
3412
3413static unsigned long
3414kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
3415{
3416 int cpu;
3417 unsigned long count = 0;
3418
3419 /* Snapshot count of all CPUs */
3420 for_each_possible_cpu(cpu) {
3421 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3422
3423 count += READ_ONCE(krcp->count);
3424 count += READ_ONCE(krcp->nr_bkv_objs);
3425 atomic_set(&krcp->backoff_page_cache_fill, 1);
3426 }
3427
3428 return count == 0 ? SHRINK_EMPTY : count;
3429}
3430
3431static unsigned long
3432kfree_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
3433{
3434 int cpu, freed = 0;
3435
3436 for_each_possible_cpu(cpu) {
3437 int count;
3438 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3439
3440 count = krcp->count;
3441 count += drain_page_cache(krcp);
3442 kfree_rcu_monitor(&krcp->monitor_work.work);
3443
3444 sc->nr_to_scan -= count;
3445 freed += count;
3446
3447 if (sc->nr_to_scan <= 0)
3448 break;
3449 }
3450
3451 return freed == 0 ? SHRINK_STOP : freed;
3452}
3453
3454static struct shrinker kfree_rcu_shrinker = {
3455 .count_objects = kfree_rcu_shrink_count,
3456 .scan_objects = kfree_rcu_shrink_scan,
3457 .batch = 0,
3458 .seeks = DEFAULT_SEEKS,
3459};
3460
3461void __init kfree_rcu_scheduler_running(void)
3462{
3463 int cpu;
3464 unsigned long flags;
3465
3466 for_each_possible_cpu(cpu) {
3467 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3468
3469 raw_spin_lock_irqsave(&krcp->lock, flags);
3470 if (need_offload_krc(krcp))
3471 schedule_delayed_monitor_work(krcp);
3472 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3473 }
3474}
3475
3476/*
3477 * During early boot, any blocking grace-period wait automatically
3478 * implies a grace period.
3479 *
3480 * Later on, this could in theory be the case for kernels built with
3481 * CONFIG_SMP=y && CONFIG_PREEMPTION=y running on a single CPU, but this
3482 * is not a common case. Furthermore, this optimization would cause
3483 * the rcu_gp_oldstate structure to expand by 50%, so this potential
3484 * grace-period optimization is ignored once the scheduler is running.
3485 */
3486static int rcu_blocking_is_gp(void)
3487{
3488 if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
3489 return false;
3490 might_sleep(); /* Check for RCU read-side critical section. */
3491 return true;
3492}
3493
3494/**
3495 * synchronize_rcu - wait until a grace period has elapsed.
3496 *
3497 * Control will return to the caller some time after a full grace
3498 * period has elapsed, in other words after all currently executing RCU
3499 * read-side critical sections have completed. Note, however, that
3500 * upon return from synchronize_rcu(), the caller might well be executing
3501 * concurrently with new RCU read-side critical sections that began while
3502 * synchronize_rcu() was waiting.
3503 *
3504 * RCU read-side critical sections are delimited by rcu_read_lock()
3505 * and rcu_read_unlock(), and may be nested. In addition, but only in
3506 * v5.0 and later, regions of code across which interrupts, preemption,
3507 * or softirqs have been disabled also serve as RCU read-side critical
3508 * sections. This includes hardware interrupt handlers, softirq handlers,
3509 * and NMI handlers.
3510 *
3511 * Note that this guarantee implies further memory-ordering guarantees.
3512 * On systems with more than one CPU, when synchronize_rcu() returns,
3513 * each CPU is guaranteed to have executed a full memory barrier since
3514 * the end of its last RCU read-side critical section whose beginning
3515 * preceded the call to synchronize_rcu(). In addition, each CPU having
3516 * an RCU read-side critical section that extends beyond the return from
3517 * synchronize_rcu() is guaranteed to have executed a full memory barrier
3518 * after the beginning of synchronize_rcu() and before the beginning of
3519 * that RCU read-side critical section. Note that these guarantees include
3520 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3521 * that are executing in the kernel.
3522 *
3523 * Furthermore, if CPU A invoked synchronize_rcu(), which returned
3524 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3525 * to have executed a full memory barrier during the execution of
3526 * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
3527 * again only if the system has more than one CPU).
3528 *
3529 * Implementation of these memory-ordering guarantees is described here:
3530 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
3531 */
3532void synchronize_rcu(void)
3533{
3534 unsigned long flags;
3535 struct rcu_node *rnp;
3536
3537 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3538 lock_is_held(&rcu_lock_map) ||
3539 lock_is_held(&rcu_sched_lock_map),
3540 "Illegal synchronize_rcu() in RCU read-side critical section");
3541 if (!rcu_blocking_is_gp()) {
3542 if (rcu_gp_is_expedited())
3543 synchronize_rcu_expedited();
3544 else
3545 wait_rcu_gp(call_rcu_hurry);
3546 return;
3547 }
3548
3549 // Context allows vacuous grace periods.
3550 // Note well that this code runs with !PREEMPT && !SMP.
3551 // In addition, all code that advances grace periods runs at
3552 // process level. Therefore, this normal GP overlaps with other
3553 // normal GPs only by being fully nested within them, which allows
3554 // reuse of ->gp_seq_polled_snap.
3555 rcu_poll_gp_seq_start_unlocked(&rcu_state.gp_seq_polled_snap);
3556 rcu_poll_gp_seq_end_unlocked(&rcu_state.gp_seq_polled_snap);
3557
3558 // Update the normal grace-period counters to record
3559 // this grace period, but only those used by the boot CPU.
3560 // The rcu_scheduler_starting() will take care of the rest of
3561 // these counters.
3562 local_irq_save(flags);
3563 WARN_ON_ONCE(num_online_cpus() > 1);
3564 rcu_state.gp_seq += (1 << RCU_SEQ_CTR_SHIFT);
3565 for (rnp = this_cpu_ptr(&rcu_data)->mynode; rnp; rnp = rnp->parent)
3566 rnp->gp_seq_needed = rnp->gp_seq = rcu_state.gp_seq;
3567 local_irq_restore(flags);
3568}
3569EXPORT_SYMBOL_GPL(synchronize_rcu);
3570
3571/**
3572 * get_completed_synchronize_rcu_full - Return a full pre-completed polled state cookie
3573 * @rgosp: Place to put state cookie
3574 *
3575 * Stores into @rgosp a value that will always be treated by functions
3576 * like poll_state_synchronize_rcu_full() as a cookie whose grace period
3577 * has already completed.
3578 */
3579void get_completed_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3580{
3581 rgosp->rgos_norm = RCU_GET_STATE_COMPLETED;
3582 rgosp->rgos_exp = RCU_GET_STATE_COMPLETED;
3583}
3584EXPORT_SYMBOL_GPL(get_completed_synchronize_rcu_full);
3585
3586/**
3587 * get_state_synchronize_rcu - Snapshot current RCU state
3588 *
3589 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3590 * or poll_state_synchronize_rcu() to determine whether or not a full
3591 * grace period has elapsed in the meantime.
3592 */
3593unsigned long get_state_synchronize_rcu(void)
3594{
3595 /*
3596 * Any prior manipulation of RCU-protected data must happen
3597 * before the load from ->gp_seq.
3598 */
3599 smp_mb(); /* ^^^ */
3600 return rcu_seq_snap(&rcu_state.gp_seq_polled);
3601}
3602EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3603
3604/**
3605 * get_state_synchronize_rcu_full - Snapshot RCU state, both normal and expedited
3606 * @rgosp: location to place combined normal/expedited grace-period state
3607 *
3608 * Places the normal and expedited grace-period states in @rgosp. This
3609 * state value can be passed to a later call to cond_synchronize_rcu_full()
3610 * or poll_state_synchronize_rcu_full() to determine whether or not a
3611 * grace period (whether normal or expedited) has elapsed in the meantime.
3612 * The rcu_gp_oldstate structure takes up twice the memory of an unsigned
3613 * long, but is guaranteed to see all grace periods. In contrast, the
3614 * combined state occupies less memory, but can sometimes fail to take
3615 * grace periods into account.
3616 *
3617 * This does not guarantee that the needed grace period will actually
3618 * start.
3619 */
3620void get_state_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3621{
3622 struct rcu_node *rnp = rcu_get_root();
3623
3624 /*
3625 * Any prior manipulation of RCU-protected data must happen
3626 * before the loads from ->gp_seq and ->expedited_sequence.
3627 */
3628 smp_mb(); /* ^^^ */
3629 rgosp->rgos_norm = rcu_seq_snap(&rnp->gp_seq);
3630 rgosp->rgos_exp = rcu_seq_snap(&rcu_state.expedited_sequence);
3631}
3632EXPORT_SYMBOL_GPL(get_state_synchronize_rcu_full);
3633
3634/*
3635 * Helper function for start_poll_synchronize_rcu() and
3636 * start_poll_synchronize_rcu_full().
3637 */
3638static void start_poll_synchronize_rcu_common(void)
3639{
3640 unsigned long flags;
3641 bool needwake;
3642 struct rcu_data *rdp;
3643 struct rcu_node *rnp;
3644
3645 lockdep_assert_irqs_enabled();
3646 local_irq_save(flags);
3647 rdp = this_cpu_ptr(&rcu_data);
3648 rnp = rdp->mynode;
3649 raw_spin_lock_rcu_node(rnp); // irqs already disabled.
3650 // Note it is possible for a grace period to have elapsed between
3651 // the above call to get_state_synchronize_rcu() and the below call
3652 // to rcu_seq_snap. This is OK, the worst that happens is that we
3653 // get a grace period that no one needed. These accesses are ordered
3654 // by smp_mb(), and we are accessing them in the opposite order
3655 // from which they are updated at grace-period start, as required.
3656 needwake = rcu_start_this_gp(rnp, rdp, rcu_seq_snap(&rcu_state.gp_seq));
3657 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3658 if (needwake)
3659 rcu_gp_kthread_wake();
3660}
3661
3662/**
3663 * start_poll_synchronize_rcu - Snapshot and start RCU grace period
3664 *
3665 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3666 * or poll_state_synchronize_rcu() to determine whether or not a full
3667 * grace period has elapsed in the meantime. If the needed grace period
3668 * is not already slated to start, notifies RCU core of the need for that
3669 * grace period.
3670 *
3671 * Interrupts must be enabled for the case where it is necessary to awaken
3672 * the grace-period kthread.
3673 */
3674unsigned long start_poll_synchronize_rcu(void)
3675{
3676 unsigned long gp_seq = get_state_synchronize_rcu();
3677
3678 start_poll_synchronize_rcu_common();
3679 return gp_seq;
3680}
3681EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu);
3682
3683/**
3684 * start_poll_synchronize_rcu_full - Take a full snapshot and start RCU grace period
3685 * @rgosp: value from get_state_synchronize_rcu_full() or start_poll_synchronize_rcu_full()
3686 *
3687 * Places the normal and expedited grace-period states in *@rgos. This
3688 * state value can be passed to a later call to cond_synchronize_rcu_full()
3689 * or poll_state_synchronize_rcu_full() to determine whether or not a
3690 * grace period (whether normal or expedited) has elapsed in the meantime.
3691 * If the needed grace period is not already slated to start, notifies
3692 * RCU core of the need for that grace period.
3693 *
3694 * Interrupts must be enabled for the case where it is necessary to awaken
3695 * the grace-period kthread.
3696 */
3697void start_poll_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3698{
3699 get_state_synchronize_rcu_full(rgosp);
3700
3701 start_poll_synchronize_rcu_common();
3702}
3703EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_full);
3704
3705/**
3706 * poll_state_synchronize_rcu - Has the specified RCU grace period completed?
3707 * @oldstate: value from get_state_synchronize_rcu() or start_poll_synchronize_rcu()
3708 *
3709 * If a full RCU grace period has elapsed since the earlier call from
3710 * which @oldstate was obtained, return @true, otherwise return @false.
3711 * If @false is returned, it is the caller's responsibility to invoke this
3712 * function later on until it does return @true. Alternatively, the caller
3713 * can explicitly wait for a grace period, for example, by passing @oldstate
3714 * to cond_synchronize_rcu() or by directly invoking synchronize_rcu().
3715 *
3716 * Yes, this function does not take counter wrap into account.
3717 * But counter wrap is harmless. If the counter wraps, we have waited for
3718 * more than a billion grace periods (and way more on a 64-bit system!).
3719 * Those needing to keep old state values for very long time periods
3720 * (many hours even on 32-bit systems) should check them occasionally and
3721 * either refresh them or set a flag indicating that the grace period has
3722 * completed. Alternatively, they can use get_completed_synchronize_rcu()
3723 * to get a guaranteed-completed grace-period state.
3724 *
3725 * This function provides the same memory-ordering guarantees that
3726 * would be provided by a synchronize_rcu() that was invoked at the call
3727 * to the function that provided @oldstate, and that returned at the end
3728 * of this function.
3729 */
3730bool poll_state_synchronize_rcu(unsigned long oldstate)
3731{
3732 if (oldstate == RCU_GET_STATE_COMPLETED ||
3733 rcu_seq_done_exact(&rcu_state.gp_seq_polled, oldstate)) {
3734 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3735 return true;
3736 }
3737 return false;
3738}
3739EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu);
3740
3741/**
3742 * poll_state_synchronize_rcu_full - Has the specified RCU grace period completed?
3743 * @rgosp: value from get_state_synchronize_rcu_full() or start_poll_synchronize_rcu_full()
3744 *
3745 * If a full RCU grace period has elapsed since the earlier call from
3746 * which *rgosp was obtained, return @true, otherwise return @false.
3747 * If @false is returned, it is the caller's responsibility to invoke this
3748 * function later on until it does return @true. Alternatively, the caller
3749 * can explicitly wait for a grace period, for example, by passing @rgosp
3750 * to cond_synchronize_rcu() or by directly invoking synchronize_rcu().
3751 *
3752 * Yes, this function does not take counter wrap into account.
3753 * But counter wrap is harmless. If the counter wraps, we have waited
3754 * for more than a billion grace periods (and way more on a 64-bit
3755 * system!). Those needing to keep rcu_gp_oldstate values for very
3756 * long time periods (many hours even on 32-bit systems) should check
3757 * them occasionally and either refresh them or set a flag indicating
3758 * that the grace period has completed. Alternatively, they can use
3759 * get_completed_synchronize_rcu_full() to get a guaranteed-completed
3760 * grace-period state.
3761 *
3762 * This function provides the same memory-ordering guarantees that would
3763 * be provided by a synchronize_rcu() that was invoked at the call to
3764 * the function that provided @rgosp, and that returned at the end of this
3765 * function. And this guarantee requires that the root rcu_node structure's
3766 * ->gp_seq field be checked instead of that of the rcu_state structure.
3767 * The problem is that the just-ending grace-period's callbacks can be
3768 * invoked between the time that the root rcu_node structure's ->gp_seq
3769 * field is updated and the time that the rcu_state structure's ->gp_seq
3770 * field is updated. Therefore, if a single synchronize_rcu() is to
3771 * cause a subsequent poll_state_synchronize_rcu_full() to return @true,
3772 * then the root rcu_node structure is the one that needs to be polled.
3773 */
3774bool poll_state_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3775{
3776 struct rcu_node *rnp = rcu_get_root();
3777
3778 smp_mb(); // Order against root rcu_node structure grace-period cleanup.
3779 if (rgosp->rgos_norm == RCU_GET_STATE_COMPLETED ||
3780 rcu_seq_done_exact(&rnp->gp_seq, rgosp->rgos_norm) ||
3781 rgosp->rgos_exp == RCU_GET_STATE_COMPLETED ||
3782 rcu_seq_done_exact(&rcu_state.expedited_sequence, rgosp->rgos_exp)) {
3783 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3784 return true;
3785 }
3786 return false;
3787}
3788EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu_full);
3789
3790/**
3791 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3792 * @oldstate: value from get_state_synchronize_rcu(), start_poll_synchronize_rcu(), or start_poll_synchronize_rcu_expedited()
3793 *
3794 * If a full RCU grace period has elapsed since the earlier call to
3795 * get_state_synchronize_rcu() or start_poll_synchronize_rcu(), just return.
3796 * Otherwise, invoke synchronize_rcu() to wait for a full grace period.
3797 *
3798 * Yes, this function does not take counter wrap into account.
3799 * But counter wrap is harmless. If the counter wraps, we have waited for
3800 * more than 2 billion grace periods (and way more on a 64-bit system!),
3801 * so waiting for a couple of additional grace periods should be just fine.
3802 *
3803 * This function provides the same memory-ordering guarantees that
3804 * would be provided by a synchronize_rcu() that was invoked at the call
3805 * to the function that provided @oldstate and that returned at the end
3806 * of this function.
3807 */
3808void cond_synchronize_rcu(unsigned long oldstate)
3809{
3810 if (!poll_state_synchronize_rcu(oldstate))
3811 synchronize_rcu();
3812}
3813EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3814
3815/**
3816 * cond_synchronize_rcu_full - Conditionally wait for an RCU grace period
3817 * @rgosp: value from get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(), or start_poll_synchronize_rcu_expedited_full()
3818 *
3819 * If a full RCU grace period has elapsed since the call to
3820 * get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(),
3821 * or start_poll_synchronize_rcu_expedited_full() from which @rgosp was
3822 * obtained, just return. Otherwise, invoke synchronize_rcu() to wait
3823 * for a full grace period.
3824 *
3825 * Yes, this function does not take counter wrap into account.
3826 * But counter wrap is harmless. If the counter wraps, we have waited for
3827 * more than 2 billion grace periods (and way more on a 64-bit system!),
3828 * so waiting for a couple of additional grace periods should be just fine.
3829 *
3830 * This function provides the same memory-ordering guarantees that
3831 * would be provided by a synchronize_rcu() that was invoked at the call
3832 * to the function that provided @rgosp and that returned at the end of
3833 * this function.
3834 */
3835void cond_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3836{
3837 if (!poll_state_synchronize_rcu_full(rgosp))
3838 synchronize_rcu();
3839}
3840EXPORT_SYMBOL_GPL(cond_synchronize_rcu_full);
3841
3842/*
3843 * Check to see if there is any immediate RCU-related work to be done by
3844 * the current CPU, returning 1 if so and zero otherwise. The checks are
3845 * in order of increasing expense: checks that can be carried out against
3846 * CPU-local state are performed first. However, we must check for CPU
3847 * stalls first, else we might not get a chance.
3848 */
3849static int rcu_pending(int user)
3850{
3851 bool gp_in_progress;
3852 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
3853 struct rcu_node *rnp = rdp->mynode;
3854
3855 lockdep_assert_irqs_disabled();
3856
3857 /* Check for CPU stalls, if enabled. */
3858 check_cpu_stall(rdp);
3859
3860 /* Does this CPU need a deferred NOCB wakeup? */
3861 if (rcu_nocb_need_deferred_wakeup(rdp, RCU_NOCB_WAKE))
3862 return 1;
3863
3864 /* Is this a nohz_full CPU in userspace or idle? (Ignore RCU if so.) */
3865 if ((user || rcu_is_cpu_rrupt_from_idle()) && rcu_nohz_full_cpu())
3866 return 0;
3867
3868 /* Is the RCU core waiting for a quiescent state from this CPU? */
3869 gp_in_progress = rcu_gp_in_progress();
3870 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm && gp_in_progress)
3871 return 1;
3872
3873 /* Does this CPU have callbacks ready to invoke? */
3874 if (!rcu_rdp_is_offloaded(rdp) &&
3875 rcu_segcblist_ready_cbs(&rdp->cblist))
3876 return 1;
3877
3878 /* Has RCU gone idle with this CPU needing another grace period? */
3879 if (!gp_in_progress && rcu_segcblist_is_enabled(&rdp->cblist) &&
3880 !rcu_rdp_is_offloaded(rdp) &&
3881 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
3882 return 1;
3883
3884 /* Have RCU grace period completed or started? */
3885 if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
3886 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
3887 return 1;
3888
3889 /* nothing to do */
3890 return 0;
3891}
3892
3893/*
3894 * Helper function for rcu_barrier() tracing. If tracing is disabled,
3895 * the compiler is expected to optimize this away.
3896 */
3897static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
3898{
3899 trace_rcu_barrier(rcu_state.name, s, cpu,
3900 atomic_read(&rcu_state.barrier_cpu_count), done);
3901}
3902
3903/*
3904 * RCU callback function for rcu_barrier(). If we are last, wake
3905 * up the task executing rcu_barrier().
3906 *
3907 * Note that the value of rcu_state.barrier_sequence must be captured
3908 * before the atomic_dec_and_test(). Otherwise, if this CPU is not last,
3909 * other CPUs might count the value down to zero before this CPU gets
3910 * around to invoking rcu_barrier_trace(), which might result in bogus
3911 * data from the next instance of rcu_barrier().
3912 */
3913static void rcu_barrier_callback(struct rcu_head *rhp)
3914{
3915 unsigned long __maybe_unused s = rcu_state.barrier_sequence;
3916
3917 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
3918 rcu_barrier_trace(TPS("LastCB"), -1, s);
3919 complete(&rcu_state.barrier_completion);
3920 } else {
3921 rcu_barrier_trace(TPS("CB"), -1, s);
3922 }
3923}
3924
3925/*
3926 * If needed, entrain an rcu_barrier() callback on rdp->cblist.
3927 */
3928static void rcu_barrier_entrain(struct rcu_data *rdp)
3929{
3930 unsigned long gseq = READ_ONCE(rcu_state.barrier_sequence);
3931 unsigned long lseq = READ_ONCE(rdp->barrier_seq_snap);
3932 bool wake_nocb = false;
3933 bool was_alldone = false;
3934
3935 lockdep_assert_held(&rcu_state.barrier_lock);
3936 if (rcu_seq_state(lseq) || !rcu_seq_state(gseq) || rcu_seq_ctr(lseq) != rcu_seq_ctr(gseq))
3937 return;
3938 rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
3939 rdp->barrier_head.func = rcu_barrier_callback;
3940 debug_rcu_head_queue(&rdp->barrier_head);
3941 rcu_nocb_lock(rdp);
3942 /*
3943 * Flush bypass and wakeup rcuog if we add callbacks to an empty regular
3944 * queue. This way we don't wait for bypass timer that can reach seconds
3945 * if it's fully lazy.
3946 */
3947 was_alldone = rcu_rdp_is_offloaded(rdp) && !rcu_segcblist_pend_cbs(&rdp->cblist);
3948 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies, false));
3949 wake_nocb = was_alldone && rcu_segcblist_pend_cbs(&rdp->cblist);
3950 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head)) {
3951 atomic_inc(&rcu_state.barrier_cpu_count);
3952 } else {
3953 debug_rcu_head_unqueue(&rdp->barrier_head);
3954 rcu_barrier_trace(TPS("IRQNQ"), -1, rcu_state.barrier_sequence);
3955 }
3956 rcu_nocb_unlock(rdp);
3957 if (wake_nocb)
3958 wake_nocb_gp(rdp, false);
3959 smp_store_release(&rdp->barrier_seq_snap, gseq);
3960}
3961
3962/*
3963 * Called with preemption disabled, and from cross-cpu IRQ context.
3964 */
3965static void rcu_barrier_handler(void *cpu_in)
3966{
3967 uintptr_t cpu = (uintptr_t)cpu_in;
3968 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3969
3970 lockdep_assert_irqs_disabled();
3971 WARN_ON_ONCE(cpu != rdp->cpu);
3972 WARN_ON_ONCE(cpu != smp_processor_id());
3973 raw_spin_lock(&rcu_state.barrier_lock);
3974 rcu_barrier_entrain(rdp);
3975 raw_spin_unlock(&rcu_state.barrier_lock);
3976}
3977
3978/**
3979 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
3980 *
3981 * Note that this primitive does not necessarily wait for an RCU grace period
3982 * to complete. For example, if there are no RCU callbacks queued anywhere
3983 * in the system, then rcu_barrier() is within its rights to return
3984 * immediately, without waiting for anything, much less an RCU grace period.
3985 */
3986void rcu_barrier(void)
3987{
3988 uintptr_t cpu;
3989 unsigned long flags;
3990 unsigned long gseq;
3991 struct rcu_data *rdp;
3992 unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
3993
3994 rcu_barrier_trace(TPS("Begin"), -1, s);
3995
3996 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3997 mutex_lock(&rcu_state.barrier_mutex);
3998
3999 /* Did someone else do our work for us? */
4000 if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
4001 rcu_barrier_trace(TPS("EarlyExit"), -1, rcu_state.barrier_sequence);
4002 smp_mb(); /* caller's subsequent code after above check. */
4003 mutex_unlock(&rcu_state.barrier_mutex);
4004 return;
4005 }
4006
4007 /* Mark the start of the barrier operation. */
4008 raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
4009 rcu_seq_start(&rcu_state.barrier_sequence);
4010 gseq = rcu_state.barrier_sequence;
4011 rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
4012
4013 /*
4014 * Initialize the count to two rather than to zero in order
4015 * to avoid a too-soon return to zero in case of an immediate
4016 * invocation of the just-enqueued callback (or preemption of
4017 * this task). Exclude CPU-hotplug operations to ensure that no
4018 * offline non-offloaded CPU has callbacks queued.
4019 */
4020 init_completion(&rcu_state.barrier_completion);
4021 atomic_set(&rcu_state.barrier_cpu_count, 2);
4022 raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4023
4024 /*
4025 * Force each CPU with callbacks to register a new callback.
4026 * When that callback is invoked, we will know that all of the
4027 * corresponding CPU's preceding callbacks have been invoked.
4028 */
4029 for_each_possible_cpu(cpu) {
4030 rdp = per_cpu_ptr(&rcu_data, cpu);
4031retry:
4032 if (smp_load_acquire(&rdp->barrier_seq_snap) == gseq)
4033 continue;
4034 raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
4035 if (!rcu_segcblist_n_cbs(&rdp->cblist)) {
4036 WRITE_ONCE(rdp->barrier_seq_snap, gseq);
4037 raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4038 rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence);
4039 continue;
4040 }
4041 if (!rcu_rdp_cpu_online(rdp)) {
4042 rcu_barrier_entrain(rdp);
4043 WARN_ON_ONCE(READ_ONCE(rdp->barrier_seq_snap) != gseq);
4044 raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4045 rcu_barrier_trace(TPS("OfflineNoCBQ"), cpu, rcu_state.barrier_sequence);
4046 continue;
4047 }
4048 raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4049 if (smp_call_function_single(cpu, rcu_barrier_handler, (void *)cpu, 1)) {
4050 schedule_timeout_uninterruptible(1);
4051 goto retry;
4052 }
4053 WARN_ON_ONCE(READ_ONCE(rdp->barrier_seq_snap) != gseq);
4054 rcu_barrier_trace(TPS("OnlineQ"), cpu, rcu_state.barrier_sequence);
4055 }
4056
4057 /*
4058 * Now that we have an rcu_barrier_callback() callback on each
4059 * CPU, and thus each counted, remove the initial count.
4060 */
4061 if (atomic_sub_and_test(2, &rcu_state.barrier_cpu_count))
4062 complete(&rcu_state.barrier_completion);
4063
4064 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
4065 wait_for_completion(&rcu_state.barrier_completion);
4066
4067 /* Mark the end of the barrier operation. */
4068 rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
4069 rcu_seq_end(&rcu_state.barrier_sequence);
4070 gseq = rcu_state.barrier_sequence;
4071 for_each_possible_cpu(cpu) {
4072 rdp = per_cpu_ptr(&rcu_data, cpu);
4073
4074 WRITE_ONCE(rdp->barrier_seq_snap, gseq);
4075 }
4076
4077 /* Other rcu_barrier() invocations can now safely proceed. */
4078 mutex_unlock(&rcu_state.barrier_mutex);
4079}
4080EXPORT_SYMBOL_GPL(rcu_barrier);
4081
4082/*
4083 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
4084 * first CPU in a given leaf rcu_node structure coming online. The caller
4085 * must hold the corresponding leaf rcu_node ->lock with interrupts
4086 * disabled.
4087 */
4088static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
4089{
4090 long mask;
4091 long oldmask;
4092 struct rcu_node *rnp = rnp_leaf;
4093
4094 raw_lockdep_assert_held_rcu_node(rnp_leaf);
4095 WARN_ON_ONCE(rnp->wait_blkd_tasks);
4096 for (;;) {
4097 mask = rnp->grpmask;
4098 rnp = rnp->parent;
4099 if (rnp == NULL)
4100 return;
4101 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
4102 oldmask = rnp->qsmaskinit;
4103 rnp->qsmaskinit |= mask;
4104 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
4105 if (oldmask)
4106 return;
4107 }
4108}
4109
4110/*
4111 * Do boot-time initialization of a CPU's per-CPU RCU data.
4112 */
4113static void __init
4114rcu_boot_init_percpu_data(int cpu)
4115{
4116 struct context_tracking *ct = this_cpu_ptr(&context_tracking);
4117 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4118
4119 /* Set up local state, ensuring consistent view of global state. */
4120 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
4121 INIT_WORK(&rdp->strict_work, strict_work_handler);
4122 WARN_ON_ONCE(ct->dynticks_nesting != 1);
4123 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(cpu)));
4124 rdp->barrier_seq_snap = rcu_state.barrier_sequence;
4125 rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
4126 rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
4127 rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
4128 rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
4129 rdp->last_sched_clock = jiffies;
4130 rdp->cpu = cpu;
4131 rcu_boot_init_nocb_percpu_data(rdp);
4132}
4133
4134/*
4135 * Invoked early in the CPU-online process, when pretty much all services
4136 * are available. The incoming CPU is not present.
4137 *
4138 * Initializes a CPU's per-CPU RCU data. Note that only one online or
4139 * offline event can be happening at a given time. Note also that we can
4140 * accept some slop in the rsp->gp_seq access due to the fact that this
4141 * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
4142 * And any offloaded callbacks are being numbered elsewhere.
4143 */
4144int rcutree_prepare_cpu(unsigned int cpu)
4145{
4146 unsigned long flags;
4147 struct context_tracking *ct = per_cpu_ptr(&context_tracking, cpu);
4148 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4149 struct rcu_node *rnp = rcu_get_root();
4150
4151 /* Set up local state, ensuring consistent view of global state. */
4152 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4153 rdp->qlen_last_fqs_check = 0;
4154 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
4155 rdp->blimit = blimit;
4156 ct->dynticks_nesting = 1; /* CPU not up, no tearing. */
4157 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
4158
4159 /*
4160 * Only non-NOCB CPUs that didn't have early-boot callbacks need to be
4161 * (re-)initialized.
4162 */
4163 if (!rcu_segcblist_is_enabled(&rdp->cblist))
4164 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
4165
4166 /*
4167 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
4168 * propagation up the rcu_node tree will happen at the beginning
4169 * of the next grace period.
4170 */
4171 rnp = rdp->mynode;
4172 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
4173 rdp->beenonline = true; /* We have now been online. */
4174 rdp->gp_seq = READ_ONCE(rnp->gp_seq);
4175 rdp->gp_seq_needed = rdp->gp_seq;
4176 rdp->cpu_no_qs.b.norm = true;
4177 rdp->core_needs_qs = false;
4178 rdp->rcu_iw_pending = false;
4179 rdp->rcu_iw = IRQ_WORK_INIT_HARD(rcu_iw_handler);
4180 rdp->rcu_iw_gp_seq = rdp->gp_seq - 1;
4181 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
4182 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4183 rcu_spawn_one_boost_kthread(rnp);
4184 rcu_spawn_cpu_nocb_kthread(cpu);
4185 WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus + 1);
4186
4187 return 0;
4188}
4189
4190/*
4191 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
4192 */
4193static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
4194{
4195 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4196
4197 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
4198}
4199
4200/*
4201 * Near the end of the CPU-online process. Pretty much all services
4202 * enabled, and the CPU is now very much alive.
4203 */
4204int rcutree_online_cpu(unsigned int cpu)
4205{
4206 unsigned long flags;
4207 struct rcu_data *rdp;
4208 struct rcu_node *rnp;
4209
4210 rdp = per_cpu_ptr(&rcu_data, cpu);
4211 rnp = rdp->mynode;
4212 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4213 rnp->ffmask |= rdp->grpmask;
4214 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4215 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
4216 return 0; /* Too early in boot for scheduler work. */
4217 sync_sched_exp_online_cleanup(cpu);
4218 rcutree_affinity_setting(cpu, -1);
4219
4220 // Stop-machine done, so allow nohz_full to disable tick.
4221 tick_dep_clear(TICK_DEP_BIT_RCU);
4222 return 0;
4223}
4224
4225/*
4226 * Near the beginning of the process. The CPU is still very much alive
4227 * with pretty much all services enabled.
4228 */
4229int rcutree_offline_cpu(unsigned int cpu)
4230{
4231 unsigned long flags;
4232 struct rcu_data *rdp;
4233 struct rcu_node *rnp;
4234
4235 rdp = per_cpu_ptr(&rcu_data, cpu);
4236 rnp = rdp->mynode;
4237 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4238 rnp->ffmask &= ~rdp->grpmask;
4239 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4240
4241 rcutree_affinity_setting(cpu, cpu);
4242
4243 // nohz_full CPUs need the tick for stop-machine to work quickly
4244 tick_dep_set(TICK_DEP_BIT_RCU);
4245 return 0;
4246}
4247
4248/*
4249 * Mark the specified CPU as being online so that subsequent grace periods
4250 * (both expedited and normal) will wait on it. Note that this means that
4251 * incoming CPUs are not allowed to use RCU read-side critical sections
4252 * until this function is called. Failing to observe this restriction
4253 * will result in lockdep splats.
4254 *
4255 * Note that this function is special in that it is invoked directly
4256 * from the incoming CPU rather than from the cpuhp_step mechanism.
4257 * This is because this function must be invoked at a precise location.
4258 */
4259void rcu_cpu_starting(unsigned int cpu)
4260{
4261 unsigned long flags;
4262 unsigned long mask;
4263 struct rcu_data *rdp;
4264 struct rcu_node *rnp;
4265 bool newcpu;
4266
4267 rdp = per_cpu_ptr(&rcu_data, cpu);
4268 if (rdp->cpu_started)
4269 return;
4270 rdp->cpu_started = true;
4271
4272 rnp = rdp->mynode;
4273 mask = rdp->grpmask;
4274 local_irq_save(flags);
4275 arch_spin_lock(&rcu_state.ofl_lock);
4276 rcu_dynticks_eqs_online();
4277 raw_spin_lock(&rcu_state.barrier_lock);
4278 raw_spin_lock_rcu_node(rnp);
4279 WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask);
4280 raw_spin_unlock(&rcu_state.barrier_lock);
4281 newcpu = !(rnp->expmaskinitnext & mask);
4282 rnp->expmaskinitnext |= mask;
4283 /* Allow lockless access for expedited grace periods. */
4284 smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + newcpu); /* ^^^ */
4285 ASSERT_EXCLUSIVE_WRITER(rcu_state.ncpus);
4286 rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
4287 rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4288 rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4289
4290 /* An incoming CPU should never be blocking a grace period. */
4291 if (WARN_ON_ONCE(rnp->qsmask & mask)) { /* RCU waiting on incoming CPU? */
4292 /* rcu_report_qs_rnp() *really* wants some flags to restore */
4293 unsigned long flags2;
4294
4295 local_irq_save(flags2);
4296 rcu_disable_urgency_upon_qs(rdp);
4297 /* Report QS -after- changing ->qsmaskinitnext! */
4298 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags2);
4299 } else {
4300 raw_spin_unlock_rcu_node(rnp);
4301 }
4302 arch_spin_unlock(&rcu_state.ofl_lock);
4303 local_irq_restore(flags);
4304 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
4305}
4306
4307/*
4308 * The outgoing function has no further need of RCU, so remove it from
4309 * the rcu_node tree's ->qsmaskinitnext bit masks.
4310 *
4311 * Note that this function is special in that it is invoked directly
4312 * from the outgoing CPU rather than from the cpuhp_step mechanism.
4313 * This is because this function must be invoked at a precise location.
4314 */
4315void rcu_report_dead(unsigned int cpu)
4316{
4317 unsigned long flags, seq_flags;
4318 unsigned long mask;
4319 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4320 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
4321
4322 // Do any dangling deferred wakeups.
4323 do_nocb_deferred_wakeup(rdp);
4324
4325 rcu_preempt_deferred_qs(current);
4326
4327 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
4328 mask = rdp->grpmask;
4329 local_irq_save(seq_flags);
4330 arch_spin_lock(&rcu_state.ofl_lock);
4331 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
4332 rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4333 rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4334 if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
4335 /* Report quiescent state -before- changing ->qsmaskinitnext! */
4336 rcu_disable_urgency_upon_qs(rdp);
4337 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4338 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4339 }
4340 WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask);
4341 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4342 arch_spin_unlock(&rcu_state.ofl_lock);
4343 local_irq_restore(seq_flags);
4344
4345 rdp->cpu_started = false;
4346}
4347
4348#ifdef CONFIG_HOTPLUG_CPU
4349/*
4350 * The outgoing CPU has just passed through the dying-idle state, and we
4351 * are being invoked from the CPU that was IPIed to continue the offline
4352 * operation. Migrate the outgoing CPU's callbacks to the current CPU.
4353 */
4354void rcutree_migrate_callbacks(int cpu)
4355{
4356 unsigned long flags;
4357 struct rcu_data *my_rdp;
4358 struct rcu_node *my_rnp;
4359 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4360 bool needwake;
4361
4362 if (rcu_rdp_is_offloaded(rdp) ||
4363 rcu_segcblist_empty(&rdp->cblist))
4364 return; /* No callbacks to migrate. */
4365
4366 raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
4367 WARN_ON_ONCE(rcu_rdp_cpu_online(rdp));
4368 rcu_barrier_entrain(rdp);
4369 my_rdp = this_cpu_ptr(&rcu_data);
4370 my_rnp = my_rdp->mynode;
4371 rcu_nocb_lock(my_rdp); /* irqs already disabled. */
4372 WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies, false));
4373 raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
4374 /* Leverage recent GPs and set GP for new callbacks. */
4375 needwake = rcu_advance_cbs(my_rnp, rdp) ||
4376 rcu_advance_cbs(my_rnp, my_rdp);
4377 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
4378 raw_spin_unlock(&rcu_state.barrier_lock); /* irqs remain disabled. */
4379 needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
4380 rcu_segcblist_disable(&rdp->cblist);
4381 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) != !rcu_segcblist_n_cbs(&my_rdp->cblist));
4382 check_cb_ovld_locked(my_rdp, my_rnp);
4383 if (rcu_rdp_is_offloaded(my_rdp)) {
4384 raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
4385 __call_rcu_nocb_wake(my_rdp, true, flags);
4386 } else {
4387 rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
4388 raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags);
4389 }
4390 if (needwake)
4391 rcu_gp_kthread_wake();
4392 lockdep_assert_irqs_enabled();
4393 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
4394 !rcu_segcblist_empty(&rdp->cblist),
4395 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
4396 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
4397 rcu_segcblist_first_cb(&rdp->cblist));
4398}
4399#endif
4400
4401/*
4402 * On non-huge systems, use expedited RCU grace periods to make suspend
4403 * and hibernation run faster.
4404 */
4405static int rcu_pm_notify(struct notifier_block *self,
4406 unsigned long action, void *hcpu)
4407{
4408 switch (action) {
4409 case PM_HIBERNATION_PREPARE:
4410 case PM_SUSPEND_PREPARE:
4411 rcu_expedite_gp();
4412 break;
4413 case PM_POST_HIBERNATION:
4414 case PM_POST_SUSPEND:
4415 rcu_unexpedite_gp();
4416 break;
4417 default:
4418 break;
4419 }
4420 return NOTIFY_OK;
4421}
4422
4423#ifdef CONFIG_RCU_EXP_KTHREAD
4424struct kthread_worker *rcu_exp_gp_kworker;
4425struct kthread_worker *rcu_exp_par_gp_kworker;
4426
4427static void __init rcu_start_exp_gp_kworkers(void)
4428{
4429 const char *par_gp_kworker_name = "rcu_exp_par_gp_kthread_worker";
4430 const char *gp_kworker_name = "rcu_exp_gp_kthread_worker";
4431 struct sched_param param = { .sched_priority = kthread_prio };
4432
4433 rcu_exp_gp_kworker = kthread_create_worker(0, gp_kworker_name);
4434 if (IS_ERR_OR_NULL(rcu_exp_gp_kworker)) {
4435 pr_err("Failed to create %s!\n", gp_kworker_name);
4436 return;
4437 }
4438
4439 rcu_exp_par_gp_kworker = kthread_create_worker(0, par_gp_kworker_name);
4440 if (IS_ERR_OR_NULL(rcu_exp_par_gp_kworker)) {
4441 pr_err("Failed to create %s!\n", par_gp_kworker_name);
4442 kthread_destroy_worker(rcu_exp_gp_kworker);
4443 return;
4444 }
4445
4446 sched_setscheduler_nocheck(rcu_exp_gp_kworker->task, SCHED_FIFO, ¶m);
4447 sched_setscheduler_nocheck(rcu_exp_par_gp_kworker->task, SCHED_FIFO,
4448 ¶m);
4449}
4450
4451static inline void rcu_alloc_par_gp_wq(void)
4452{
4453}
4454#else /* !CONFIG_RCU_EXP_KTHREAD */
4455struct workqueue_struct *rcu_par_gp_wq;
4456
4457static void __init rcu_start_exp_gp_kworkers(void)
4458{
4459}
4460
4461static inline void rcu_alloc_par_gp_wq(void)
4462{
4463 rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
4464 WARN_ON(!rcu_par_gp_wq);
4465}
4466#endif /* CONFIG_RCU_EXP_KTHREAD */
4467
4468/*
4469 * Spawn the kthreads that handle RCU's grace periods.
4470 */
4471static int __init rcu_spawn_gp_kthread(void)
4472{
4473 unsigned long flags;
4474 struct rcu_node *rnp;
4475 struct sched_param sp;
4476 struct task_struct *t;
4477 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
4478
4479 rcu_scheduler_fully_active = 1;
4480 t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
4481 if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
4482 return 0;
4483 if (kthread_prio) {
4484 sp.sched_priority = kthread_prio;
4485 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4486 }
4487 rnp = rcu_get_root();
4488 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4489 WRITE_ONCE(rcu_state.gp_activity, jiffies);
4490 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
4491 // Reset .gp_activity and .gp_req_activity before setting .gp_kthread.
4492 smp_store_release(&rcu_state.gp_kthread, t); /* ^^^ */
4493 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4494 wake_up_process(t);
4495 /* This is a pre-SMP initcall, we expect a single CPU */
4496 WARN_ON(num_online_cpus() > 1);
4497 /*
4498 * Those kthreads couldn't be created on rcu_init() -> rcutree_prepare_cpu()
4499 * due to rcu_scheduler_fully_active.
4500 */
4501 rcu_spawn_cpu_nocb_kthread(smp_processor_id());
4502 rcu_spawn_one_boost_kthread(rdp->mynode);
4503 rcu_spawn_core_kthreads();
4504 /* Create kthread worker for expedited GPs */
4505 rcu_start_exp_gp_kworkers();
4506 return 0;
4507}
4508early_initcall(rcu_spawn_gp_kthread);
4509
4510/*
4511 * This function is invoked towards the end of the scheduler's
4512 * initialization process. Before this is called, the idle task might
4513 * contain synchronous grace-period primitives (during which time, this idle
4514 * task is booting the system, and such primitives are no-ops). After this
4515 * function is called, any synchronous grace-period primitives are run as
4516 * expedited, with the requesting task driving the grace period forward.
4517 * A later core_initcall() rcu_set_runtime_mode() will switch to full
4518 * runtime RCU functionality.
4519 */
4520void rcu_scheduler_starting(void)
4521{
4522 unsigned long flags;
4523 struct rcu_node *rnp;
4524
4525 WARN_ON(num_online_cpus() != 1);
4526 WARN_ON(nr_context_switches() > 0);
4527 rcu_test_sync_prims();
4528
4529 // Fix up the ->gp_seq counters.
4530 local_irq_save(flags);
4531 rcu_for_each_node_breadth_first(rnp)
4532 rnp->gp_seq_needed = rnp->gp_seq = rcu_state.gp_seq;
4533 local_irq_restore(flags);
4534
4535 // Switch out of early boot mode.
4536 rcu_scheduler_active = RCU_SCHEDULER_INIT;
4537 rcu_test_sync_prims();
4538}
4539
4540/*
4541 * Helper function for rcu_init() that initializes the rcu_state structure.
4542 */
4543static void __init rcu_init_one(void)
4544{
4545 static const char * const buf[] = RCU_NODE_NAME_INIT;
4546 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4547 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4548 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4549
4550 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4551 int cpustride = 1;
4552 int i;
4553 int j;
4554 struct rcu_node *rnp;
4555
4556 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4557
4558 /* Silence gcc 4.8 false positive about array index out of range. */
4559 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4560 panic("rcu_init_one: rcu_num_lvls out of range");
4561
4562 /* Initialize the level-tracking arrays. */
4563
4564 for (i = 1; i < rcu_num_lvls; i++)
4565 rcu_state.level[i] =
4566 rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
4567 rcu_init_levelspread(levelspread, num_rcu_lvl);
4568
4569 /* Initialize the elements themselves, starting from the leaves. */
4570
4571 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4572 cpustride *= levelspread[i];
4573 rnp = rcu_state.level[i];
4574 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
4575 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4576 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4577 &rcu_node_class[i], buf[i]);
4578 raw_spin_lock_init(&rnp->fqslock);
4579 lockdep_set_class_and_name(&rnp->fqslock,
4580 &rcu_fqs_class[i], fqs[i]);
4581 rnp->gp_seq = rcu_state.gp_seq;
4582 rnp->gp_seq_needed = rcu_state.gp_seq;
4583 rnp->completedqs = rcu_state.gp_seq;
4584 rnp->qsmask = 0;
4585 rnp->qsmaskinit = 0;
4586 rnp->grplo = j * cpustride;
4587 rnp->grphi = (j + 1) * cpustride - 1;
4588 if (rnp->grphi >= nr_cpu_ids)
4589 rnp->grphi = nr_cpu_ids - 1;
4590 if (i == 0) {
4591 rnp->grpnum = 0;
4592 rnp->grpmask = 0;
4593 rnp->parent = NULL;
4594 } else {
4595 rnp->grpnum = j % levelspread[i - 1];
4596 rnp->grpmask = BIT(rnp->grpnum);
4597 rnp->parent = rcu_state.level[i - 1] +
4598 j / levelspread[i - 1];
4599 }
4600 rnp->level = i;
4601 INIT_LIST_HEAD(&rnp->blkd_tasks);
4602 rcu_init_one_nocb(rnp);
4603 init_waitqueue_head(&rnp->exp_wq[0]);
4604 init_waitqueue_head(&rnp->exp_wq[1]);
4605 init_waitqueue_head(&rnp->exp_wq[2]);
4606 init_waitqueue_head(&rnp->exp_wq[3]);
4607 spin_lock_init(&rnp->exp_lock);
4608 mutex_init(&rnp->boost_kthread_mutex);
4609 raw_spin_lock_init(&rnp->exp_poll_lock);
4610 rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED;
4611 INIT_WORK(&rnp->exp_poll_wq, sync_rcu_do_polled_gp);
4612 }
4613 }
4614
4615 init_swait_queue_head(&rcu_state.gp_wq);
4616 init_swait_queue_head(&rcu_state.expedited_wq);
4617 rnp = rcu_first_leaf_node();
4618 for_each_possible_cpu(i) {
4619 while (i > rnp->grphi)
4620 rnp++;
4621 per_cpu_ptr(&rcu_data, i)->mynode = rnp;
4622 rcu_boot_init_percpu_data(i);
4623 }
4624}
4625
4626/*
4627 * Force priority from the kernel command-line into range.
4628 */
4629static void __init sanitize_kthread_prio(void)
4630{
4631 int kthread_prio_in = kthread_prio;
4632
4633 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
4634 && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
4635 kthread_prio = 2;
4636 else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4637 kthread_prio = 1;
4638 else if (kthread_prio < 0)
4639 kthread_prio = 0;
4640 else if (kthread_prio > 99)
4641 kthread_prio = 99;
4642
4643 if (kthread_prio != kthread_prio_in)
4644 pr_alert("%s: Limited prio to %d from %d\n",
4645 __func__, kthread_prio, kthread_prio_in);
4646}
4647
4648/*
4649 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4650 * replace the definitions in tree.h because those are needed to size
4651 * the ->node array in the rcu_state structure.
4652 */
4653void rcu_init_geometry(void)
4654{
4655 ulong d;
4656 int i;
4657 static unsigned long old_nr_cpu_ids;
4658 int rcu_capacity[RCU_NUM_LVLS];
4659 static bool initialized;
4660
4661 if (initialized) {
4662 /*
4663 * Warn if setup_nr_cpu_ids() had not yet been invoked,
4664 * unless nr_cpus_ids == NR_CPUS, in which case who cares?
4665 */
4666 WARN_ON_ONCE(old_nr_cpu_ids != nr_cpu_ids);
4667 return;
4668 }
4669
4670 old_nr_cpu_ids = nr_cpu_ids;
4671 initialized = true;
4672
4673 /*
4674 * Initialize any unspecified boot parameters.
4675 * The default values of jiffies_till_first_fqs and
4676 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4677 * value, which is a function of HZ, then adding one for each
4678 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4679 */
4680 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4681 if (jiffies_till_first_fqs == ULONG_MAX)
4682 jiffies_till_first_fqs = d;
4683 if (jiffies_till_next_fqs == ULONG_MAX)
4684 jiffies_till_next_fqs = d;
4685 adjust_jiffies_till_sched_qs();
4686
4687 /* If the compile-time values are accurate, just leave. */
4688 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4689 nr_cpu_ids == NR_CPUS)
4690 return;
4691 pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4692 rcu_fanout_leaf, nr_cpu_ids);
4693
4694 /*
4695 * The boot-time rcu_fanout_leaf parameter must be at least two
4696 * and cannot exceed the number of bits in the rcu_node masks.
4697 * Complain and fall back to the compile-time values if this
4698 * limit is exceeded.
4699 */
4700 if (rcu_fanout_leaf < 2 ||
4701 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4702 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4703 WARN_ON(1);
4704 return;
4705 }
4706
4707 /*
4708 * Compute number of nodes that can be handled an rcu_node tree
4709 * with the given number of levels.
4710 */
4711 rcu_capacity[0] = rcu_fanout_leaf;
4712 for (i = 1; i < RCU_NUM_LVLS; i++)
4713 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4714
4715 /*
4716 * The tree must be able to accommodate the configured number of CPUs.
4717 * If this limit is exceeded, fall back to the compile-time values.
4718 */
4719 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4720 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4721 WARN_ON(1);
4722 return;
4723 }
4724
4725 /* Calculate the number of levels in the tree. */
4726 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4727 }
4728 rcu_num_lvls = i + 1;
4729
4730 /* Calculate the number of rcu_nodes at each level of the tree. */
4731 for (i = 0; i < rcu_num_lvls; i++) {
4732 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4733 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4734 }
4735
4736 /* Calculate the total number of rcu_node structures. */
4737 rcu_num_nodes = 0;
4738 for (i = 0; i < rcu_num_lvls; i++)
4739 rcu_num_nodes += num_rcu_lvl[i];
4740}
4741
4742/*
4743 * Dump out the structure of the rcu_node combining tree associated
4744 * with the rcu_state structure.
4745 */
4746static void __init rcu_dump_rcu_node_tree(void)
4747{
4748 int level = 0;
4749 struct rcu_node *rnp;
4750
4751 pr_info("rcu_node tree layout dump\n");
4752 pr_info(" ");
4753 rcu_for_each_node_breadth_first(rnp) {
4754 if (rnp->level != level) {
4755 pr_cont("\n");
4756 pr_info(" ");
4757 level = rnp->level;
4758 }
4759 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4760 }
4761 pr_cont("\n");
4762}
4763
4764struct workqueue_struct *rcu_gp_wq;
4765
4766static void __init kfree_rcu_batch_init(void)
4767{
4768 int cpu;
4769 int i;
4770
4771 /* Clamp it to [0:100] seconds interval. */
4772 if (rcu_delay_page_cache_fill_msec < 0 ||
4773 rcu_delay_page_cache_fill_msec > 100 * MSEC_PER_SEC) {
4774
4775 rcu_delay_page_cache_fill_msec =
4776 clamp(rcu_delay_page_cache_fill_msec, 0,
4777 (int) (100 * MSEC_PER_SEC));
4778
4779 pr_info("Adjusting rcutree.rcu_delay_page_cache_fill_msec to %d ms.\n",
4780 rcu_delay_page_cache_fill_msec);
4781 }
4782
4783 for_each_possible_cpu(cpu) {
4784 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
4785
4786 for (i = 0; i < KFREE_N_BATCHES; i++) {
4787 INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
4788 krcp->krw_arr[i].krcp = krcp;
4789 }
4790
4791 INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
4792 INIT_DELAYED_WORK(&krcp->page_cache_work, fill_page_cache_func);
4793 krcp->initialized = true;
4794 }
4795 if (register_shrinker(&kfree_rcu_shrinker, "rcu-kfree"))
4796 pr_err("Failed to register kfree_rcu() shrinker!\n");
4797}
4798
4799void __init rcu_init(void)
4800{
4801 int cpu = smp_processor_id();
4802
4803 rcu_early_boot_tests();
4804
4805 kfree_rcu_batch_init();
4806 rcu_bootup_announce();
4807 sanitize_kthread_prio();
4808 rcu_init_geometry();
4809 rcu_init_one();
4810 if (dump_tree)
4811 rcu_dump_rcu_node_tree();
4812 if (use_softirq)
4813 open_softirq(RCU_SOFTIRQ, rcu_core_si);
4814
4815 /*
4816 * We don't need protection against CPU-hotplug here because
4817 * this is called early in boot, before either interrupts
4818 * or the scheduler are operational.
4819 */
4820 pm_notifier(rcu_pm_notify, 0);
4821 WARN_ON(num_online_cpus() > 1); // Only one CPU this early in boot.
4822 rcutree_prepare_cpu(cpu);
4823 rcu_cpu_starting(cpu);
4824 rcutree_online_cpu(cpu);
4825
4826 /* Create workqueue for Tree SRCU and for expedited GPs. */
4827 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
4828 WARN_ON(!rcu_gp_wq);
4829 rcu_alloc_par_gp_wq();
4830
4831 /* Fill in default value for rcutree.qovld boot parameter. */
4832 /* -After- the rcu_node ->lock fields are initialized! */
4833 if (qovld < 0)
4834 qovld_calc = DEFAULT_RCU_QOVLD_MULT * qhimark;
4835 else
4836 qovld_calc = qovld;
4837
4838 // Kick-start any polled grace periods that started early.
4839 if (!(per_cpu_ptr(&rcu_data, cpu)->mynode->exp_seq_poll_rq & 0x1))
4840 (void)start_poll_synchronize_rcu_expedited();
4841}
4842
4843#include "tree_stall.h"
4844#include "tree_exp.h"
4845#include "tree_nocb.h"
4846#include "tree_plugin.h"