xref: /openbmc/linux/kernel/rcu/tree.c (revision b755c25f)
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 
78 static 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 };
84 static 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. */
98 static bool dump_tree;
99 module_param(dump_tree, bool, 0444);
100 /* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */
101 static bool use_softirq = !IS_ENABLED(CONFIG_PREEMPT_RT);
102 #ifndef CONFIG_PREEMPT_RT
103 module_param(use_softirq, bool, 0444);
104 #endif
105 /* Control rcu_node-tree auto-balancing at boot time. */
106 static bool rcu_fanout_exact;
107 module_param(rcu_fanout_exact, bool, 0444);
108 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
109 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
110 module_param(rcu_fanout_leaf, int, 0444);
111 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
112 /* Number of rcu_nodes at specified level. */
113 int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
114 int 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  */
128 int rcu_scheduler_active __read_mostly;
129 EXPORT_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  */
143 static int rcu_scheduler_fully_active __read_mostly;
144 
145 static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
146 			      unsigned long gps, unsigned long flags);
147 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
148 static void invoke_rcu_core(void);
149 static void rcu_report_exp_rdp(struct rcu_data *rdp);
150 static void sync_sched_exp_online_cleanup(int cpu);
151 static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp);
152 static bool rcu_rdp_is_offloaded(struct rcu_data *rdp);
153 static bool rcu_rdp_cpu_online(struct rcu_data *rdp);
154 static bool rcu_init_invoked(void);
155 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
156 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
157 
158 /*
159  * rcuc/rcub/rcuop kthread realtime priority. The "rcuop"
160  * real-time priority(enabling/disabling) is controlled by
161  * the extra CONFIG_RCU_NOCB_CPU_CB_BOOST configuration.
162  */
163 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
164 module_param(kthread_prio, int, 0444);
165 
166 /* Delay in jiffies for grace-period initialization delays, debug only. */
167 
168 static int gp_preinit_delay;
169 module_param(gp_preinit_delay, int, 0444);
170 static int gp_init_delay;
171 module_param(gp_init_delay, int, 0444);
172 static int gp_cleanup_delay;
173 module_param(gp_cleanup_delay, int, 0444);
174 
175 // Add delay to rcu_read_unlock() for strict grace periods.
176 static int rcu_unlock_delay;
177 #ifdef CONFIG_RCU_STRICT_GRACE_PERIOD
178 module_param(rcu_unlock_delay, int, 0444);
179 #endif
180 
181 /*
182  * This rcu parameter is runtime-read-only. It reflects
183  * a minimum allowed number of objects which can be cached
184  * per-CPU. Object size is equal to one page. This value
185  * can be changed at boot time.
186  */
187 static int rcu_min_cached_objs = 5;
188 module_param(rcu_min_cached_objs, int, 0444);
189 
190 // A page shrinker can ask for pages to be freed to make them
191 // available for other parts of the system. This usually happens
192 // under low memory conditions, and in that case we should also
193 // defer page-cache filling for a short time period.
194 //
195 // The default value is 5 seconds, which is long enough to reduce
196 // interference with the shrinker while it asks other systems to
197 // drain their caches.
198 static int rcu_delay_page_cache_fill_msec = 5000;
199 module_param(rcu_delay_page_cache_fill_msec, int, 0444);
200 
201 /* Retrieve RCU kthreads priority for rcutorture */
202 int rcu_get_gp_kthreads_prio(void)
203 {
204 	return kthread_prio;
205 }
206 EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio);
207 
208 /*
209  * Number of grace periods between delays, normalized by the duration of
210  * the delay.  The longer the delay, the more the grace periods between
211  * each delay.  The reason for this normalization is that it means that,
212  * for non-zero delays, the overall slowdown of grace periods is constant
213  * regardless of the duration of the delay.  This arrangement balances
214  * the need for long delays to increase some race probabilities with the
215  * need for fast grace periods to increase other race probabilities.
216  */
217 #define PER_RCU_NODE_PERIOD 3	/* Number of grace periods between delays for debugging. */
218 
219 /*
220  * Return true if an RCU grace period is in progress.  The READ_ONCE()s
221  * permit this function to be invoked without holding the root rcu_node
222  * structure's ->lock, but of course results can be subject to change.
223  */
224 static int rcu_gp_in_progress(void)
225 {
226 	return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq));
227 }
228 
229 /*
230  * Return the number of callbacks queued on the specified CPU.
231  * Handles both the nocbs and normal cases.
232  */
233 static long rcu_get_n_cbs_cpu(int cpu)
234 {
235 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
236 
237 	if (rcu_segcblist_is_enabled(&rdp->cblist))
238 		return rcu_segcblist_n_cbs(&rdp->cblist);
239 	return 0;
240 }
241 
242 void rcu_softirq_qs(void)
243 {
244 	rcu_qs();
245 	rcu_preempt_deferred_qs(current);
246 	rcu_tasks_qs(current, false);
247 }
248 
249 /*
250  * Reset the current CPU's ->dynticks counter to indicate that the
251  * newly onlined CPU is no longer in an extended quiescent state.
252  * This will either leave the counter unchanged, or increment it
253  * to the next non-quiescent value.
254  *
255  * The non-atomic test/increment sequence works because the upper bits
256  * of the ->dynticks counter are manipulated only by the corresponding CPU,
257  * or when the corresponding CPU is offline.
258  */
259 static void rcu_dynticks_eqs_online(void)
260 {
261 	if (ct_dynticks() & RCU_DYNTICKS_IDX)
262 		return;
263 	ct_state_inc(RCU_DYNTICKS_IDX);
264 }
265 
266 /*
267  * Snapshot the ->dynticks counter with full ordering so as to allow
268  * stable comparison of this counter with past and future snapshots.
269  */
270 static int rcu_dynticks_snap(int cpu)
271 {
272 	smp_mb();  // Fundamental RCU ordering guarantee.
273 	return ct_dynticks_cpu_acquire(cpu);
274 }
275 
276 /*
277  * Return true if the snapshot returned from rcu_dynticks_snap()
278  * indicates that RCU is in an extended quiescent state.
279  */
280 static bool rcu_dynticks_in_eqs(int snap)
281 {
282 	return !(snap & RCU_DYNTICKS_IDX);
283 }
284 
285 /*
286  * Return true if the CPU corresponding to the specified rcu_data
287  * structure has spent some time in an extended quiescent state since
288  * rcu_dynticks_snap() returned the specified snapshot.
289  */
290 static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap)
291 {
292 	return snap != rcu_dynticks_snap(rdp->cpu);
293 }
294 
295 /*
296  * Return true if the referenced integer is zero while the specified
297  * CPU remains within a single extended quiescent state.
298  */
299 bool rcu_dynticks_zero_in_eqs(int cpu, int *vp)
300 {
301 	int snap;
302 
303 	// If not quiescent, force back to earlier extended quiescent state.
304 	snap = ct_dynticks_cpu(cpu) & ~RCU_DYNTICKS_IDX;
305 	smp_rmb(); // Order ->dynticks and *vp reads.
306 	if (READ_ONCE(*vp))
307 		return false;  // Non-zero, so report failure;
308 	smp_rmb(); // Order *vp read and ->dynticks re-read.
309 
310 	// If still in the same extended quiescent state, we are good!
311 	return snap == ct_dynticks_cpu(cpu);
312 }
313 
314 /*
315  * Let the RCU core know that this CPU has gone through the scheduler,
316  * which is a quiescent state.  This is called when the need for a
317  * quiescent state is urgent, so we burn an atomic operation and full
318  * memory barriers to let the RCU core know about it, regardless of what
319  * this CPU might (or might not) do in the near future.
320  *
321  * We inform the RCU core by emulating a zero-duration dyntick-idle period.
322  *
323  * The caller must have disabled interrupts and must not be idle.
324  */
325 notrace void rcu_momentary_dyntick_idle(void)
326 {
327 	int seq;
328 
329 	raw_cpu_write(rcu_data.rcu_need_heavy_qs, false);
330 	seq = ct_state_inc(2 * RCU_DYNTICKS_IDX);
331 	/* It is illegal to call this from idle state. */
332 	WARN_ON_ONCE(!(seq & RCU_DYNTICKS_IDX));
333 	rcu_preempt_deferred_qs(current);
334 }
335 EXPORT_SYMBOL_GPL(rcu_momentary_dyntick_idle);
336 
337 /**
338  * rcu_is_cpu_rrupt_from_idle - see if 'interrupted' from idle
339  *
340  * If the current CPU is idle and running at a first-level (not nested)
341  * interrupt, or directly, from idle, return true.
342  *
343  * The caller must have at least disabled IRQs.
344  */
345 static int rcu_is_cpu_rrupt_from_idle(void)
346 {
347 	long nesting;
348 
349 	/*
350 	 * Usually called from the tick; but also used from smp_function_call()
351 	 * for expedited grace periods. This latter can result in running from
352 	 * the idle task, instead of an actual IPI.
353 	 */
354 	lockdep_assert_irqs_disabled();
355 
356 	/* Check for counter underflows */
357 	RCU_LOCKDEP_WARN(ct_dynticks_nesting() < 0,
358 			 "RCU dynticks_nesting counter underflow!");
359 	RCU_LOCKDEP_WARN(ct_dynticks_nmi_nesting() <= 0,
360 			 "RCU dynticks_nmi_nesting counter underflow/zero!");
361 
362 	/* Are we at first interrupt nesting level? */
363 	nesting = ct_dynticks_nmi_nesting();
364 	if (nesting > 1)
365 		return false;
366 
367 	/*
368 	 * If we're not in an interrupt, we must be in the idle task!
369 	 */
370 	WARN_ON_ONCE(!nesting && !is_idle_task(current));
371 
372 	/* Does CPU appear to be idle from an RCU standpoint? */
373 	return ct_dynticks_nesting() == 0;
374 }
375 
376 #define DEFAULT_RCU_BLIMIT (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 1000 : 10)
377 				// Maximum callbacks per rcu_do_batch ...
378 #define DEFAULT_MAX_RCU_BLIMIT 10000 // ... even during callback flood.
379 static long blimit = DEFAULT_RCU_BLIMIT;
380 #define DEFAULT_RCU_QHIMARK 10000 // If this many pending, ignore blimit.
381 static long qhimark = DEFAULT_RCU_QHIMARK;
382 #define DEFAULT_RCU_QLOMARK 100   // Once only this many pending, use blimit.
383 static long qlowmark = DEFAULT_RCU_QLOMARK;
384 #define DEFAULT_RCU_QOVLD_MULT 2
385 #define DEFAULT_RCU_QOVLD (DEFAULT_RCU_QOVLD_MULT * DEFAULT_RCU_QHIMARK)
386 static long qovld = DEFAULT_RCU_QOVLD; // If this many pending, hammer QS.
387 static long qovld_calc = -1;	  // No pre-initialization lock acquisitions!
388 
389 module_param(blimit, long, 0444);
390 module_param(qhimark, long, 0444);
391 module_param(qlowmark, long, 0444);
392 module_param(qovld, long, 0444);
393 
394 static ulong jiffies_till_first_fqs = IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 0 : ULONG_MAX;
395 static ulong jiffies_till_next_fqs = ULONG_MAX;
396 static bool rcu_kick_kthreads;
397 static int rcu_divisor = 7;
398 module_param(rcu_divisor, int, 0644);
399 
400 /* Force an exit from rcu_do_batch() after 3 milliseconds. */
401 static long rcu_resched_ns = 3 * NSEC_PER_MSEC;
402 module_param(rcu_resched_ns, long, 0644);
403 
404 /*
405  * How long the grace period must be before we start recruiting
406  * quiescent-state help from rcu_note_context_switch().
407  */
408 static ulong jiffies_till_sched_qs = ULONG_MAX;
409 module_param(jiffies_till_sched_qs, ulong, 0444);
410 static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */
411 module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
412 
413 /*
414  * Make sure that we give the grace-period kthread time to detect any
415  * idle CPUs before taking active measures to force quiescent states.
416  * However, don't go below 100 milliseconds, adjusted upwards for really
417  * large systems.
418  */
419 static void adjust_jiffies_till_sched_qs(void)
420 {
421 	unsigned long j;
422 
423 	/* If jiffies_till_sched_qs was specified, respect the request. */
424 	if (jiffies_till_sched_qs != ULONG_MAX) {
425 		WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
426 		return;
427 	}
428 	/* Otherwise, set to third fqs scan, but bound below on large system. */
429 	j = READ_ONCE(jiffies_till_first_fqs) +
430 		      2 * READ_ONCE(jiffies_till_next_fqs);
431 	if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
432 		j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
433 	pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
434 	WRITE_ONCE(jiffies_to_sched_qs, j);
435 }
436 
437 static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
438 {
439 	ulong j;
440 	int ret = kstrtoul(val, 0, &j);
441 
442 	if (!ret) {
443 		WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
444 		adjust_jiffies_till_sched_qs();
445 	}
446 	return ret;
447 }
448 
449 static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
450 {
451 	ulong j;
452 	int ret = kstrtoul(val, 0, &j);
453 
454 	if (!ret) {
455 		WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
456 		adjust_jiffies_till_sched_qs();
457 	}
458 	return ret;
459 }
460 
461 static const struct kernel_param_ops first_fqs_jiffies_ops = {
462 	.set = param_set_first_fqs_jiffies,
463 	.get = param_get_ulong,
464 };
465 
466 static const struct kernel_param_ops next_fqs_jiffies_ops = {
467 	.set = param_set_next_fqs_jiffies,
468 	.get = param_get_ulong,
469 };
470 
471 module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
472 module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
473 module_param(rcu_kick_kthreads, bool, 0644);
474 
475 static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
476 static int rcu_pending(int user);
477 
478 /*
479  * Return the number of RCU GPs completed thus far for debug & stats.
480  */
481 unsigned long rcu_get_gp_seq(void)
482 {
483 	return READ_ONCE(rcu_state.gp_seq);
484 }
485 EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
486 
487 /*
488  * Return the number of RCU expedited batches completed thus far for
489  * debug & stats.  Odd numbers mean that a batch is in progress, even
490  * numbers mean idle.  The value returned will thus be roughly double
491  * the cumulative batches since boot.
492  */
493 unsigned long rcu_exp_batches_completed(void)
494 {
495 	return rcu_state.expedited_sequence;
496 }
497 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
498 
499 /*
500  * Return the root node of the rcu_state structure.
501  */
502 static struct rcu_node *rcu_get_root(void)
503 {
504 	return &rcu_state.node[0];
505 }
506 
507 /*
508  * Send along grace-period-related data for rcutorture diagnostics.
509  */
510 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
511 			    unsigned long *gp_seq)
512 {
513 	switch (test_type) {
514 	case RCU_FLAVOR:
515 		*flags = READ_ONCE(rcu_state.gp_flags);
516 		*gp_seq = rcu_seq_current(&rcu_state.gp_seq);
517 		break;
518 	default:
519 		break;
520 	}
521 }
522 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
523 
524 #if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK))
525 /*
526  * An empty function that will trigger a reschedule on
527  * IRQ tail once IRQs get re-enabled on userspace/guest resume.
528  */
529 static void late_wakeup_func(struct irq_work *work)
530 {
531 }
532 
533 static DEFINE_PER_CPU(struct irq_work, late_wakeup_work) =
534 	IRQ_WORK_INIT(late_wakeup_func);
535 
536 /*
537  * If either:
538  *
539  * 1) the task is about to enter in guest mode and $ARCH doesn't support KVM generic work
540  * 2) the task is about to enter in user mode and $ARCH doesn't support generic entry.
541  *
542  * In these cases the late RCU wake ups aren't supported in the resched loops and our
543  * last resort is to fire a local irq_work that will trigger a reschedule once IRQs
544  * get re-enabled again.
545  */
546 noinstr void rcu_irq_work_resched(void)
547 {
548 	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
549 
550 	if (IS_ENABLED(CONFIG_GENERIC_ENTRY) && !(current->flags & PF_VCPU))
551 		return;
552 
553 	if (IS_ENABLED(CONFIG_KVM_XFER_TO_GUEST_WORK) && (current->flags & PF_VCPU))
554 		return;
555 
556 	instrumentation_begin();
557 	if (do_nocb_deferred_wakeup(rdp) && need_resched()) {
558 		irq_work_queue(this_cpu_ptr(&late_wakeup_work));
559 	}
560 	instrumentation_end();
561 }
562 #endif /* #if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK)) */
563 
564 #ifdef CONFIG_PROVE_RCU
565 /**
566  * rcu_irq_exit_check_preempt - Validate that scheduling is possible
567  */
568 void rcu_irq_exit_check_preempt(void)
569 {
570 	lockdep_assert_irqs_disabled();
571 
572 	RCU_LOCKDEP_WARN(ct_dynticks_nesting() <= 0,
573 			 "RCU dynticks_nesting counter underflow/zero!");
574 	RCU_LOCKDEP_WARN(ct_dynticks_nmi_nesting() !=
575 			 DYNTICK_IRQ_NONIDLE,
576 			 "Bad RCU  dynticks_nmi_nesting counter\n");
577 	RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
578 			 "RCU in extended quiescent state!");
579 }
580 #endif /* #ifdef CONFIG_PROVE_RCU */
581 
582 #ifdef CONFIG_NO_HZ_FULL
583 /**
584  * __rcu_irq_enter_check_tick - Enable scheduler tick on CPU if RCU needs it.
585  *
586  * The scheduler tick is not normally enabled when CPUs enter the kernel
587  * from nohz_full userspace execution.  After all, nohz_full userspace
588  * execution is an RCU quiescent state and the time executing in the kernel
589  * is quite short.  Except of course when it isn't.  And it is not hard to
590  * cause a large system to spend tens of seconds or even minutes looping
591  * in the kernel, which can cause a number of problems, include RCU CPU
592  * stall warnings.
593  *
594  * Therefore, if a nohz_full CPU fails to report a quiescent state
595  * in a timely manner, the RCU grace-period kthread sets that CPU's
596  * ->rcu_urgent_qs flag with the expectation that the next interrupt or
597  * exception will invoke this function, which will turn on the scheduler
598  * tick, which will enable RCU to detect that CPU's quiescent states,
599  * for example, due to cond_resched() calls in CONFIG_PREEMPT=n kernels.
600  * The tick will be disabled once a quiescent state is reported for
601  * this CPU.
602  *
603  * Of course, in carefully tuned systems, there might never be an
604  * interrupt or exception.  In that case, the RCU grace-period kthread
605  * will eventually cause one to happen.  However, in less carefully
606  * controlled environments, this function allows RCU to get what it
607  * needs without creating otherwise useless interruptions.
608  */
609 void __rcu_irq_enter_check_tick(void)
610 {
611 	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
612 
613 	// If we're here from NMI there's nothing to do.
614 	if (in_nmi())
615 		return;
616 
617 	RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
618 			 "Illegal rcu_irq_enter_check_tick() from extended quiescent state");
619 
620 	if (!tick_nohz_full_cpu(rdp->cpu) ||
621 	    !READ_ONCE(rdp->rcu_urgent_qs) ||
622 	    READ_ONCE(rdp->rcu_forced_tick)) {
623 		// RCU doesn't need nohz_full help from this CPU, or it is
624 		// already getting that help.
625 		return;
626 	}
627 
628 	// We get here only when not in an extended quiescent state and
629 	// from interrupts (as opposed to NMIs).  Therefore, (1) RCU is
630 	// already watching and (2) The fact that we are in an interrupt
631 	// handler and that the rcu_node lock is an irq-disabled lock
632 	// prevents self-deadlock.  So we can safely recheck under the lock.
633 	// Note that the nohz_full state currently cannot change.
634 	raw_spin_lock_rcu_node(rdp->mynode);
635 	if (rdp->rcu_urgent_qs && !rdp->rcu_forced_tick) {
636 		// A nohz_full CPU is in the kernel and RCU needs a
637 		// quiescent state.  Turn on the tick!
638 		WRITE_ONCE(rdp->rcu_forced_tick, true);
639 		tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
640 	}
641 	raw_spin_unlock_rcu_node(rdp->mynode);
642 }
643 NOKPROBE_SYMBOL(__rcu_irq_enter_check_tick);
644 #endif /* CONFIG_NO_HZ_FULL */
645 
646 /*
647  * Check to see if any future non-offloaded RCU-related work will need
648  * to be done by the current CPU, even if none need be done immediately,
649  * returning 1 if so.  This function is part of the RCU implementation;
650  * it is -not- an exported member of the RCU API.  This is used by
651  * the idle-entry code to figure out whether it is safe to disable the
652  * scheduler-clock interrupt.
653  *
654  * Just check whether or not this CPU has non-offloaded RCU callbacks
655  * queued.
656  */
657 int rcu_needs_cpu(void)
658 {
659 	return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) &&
660 		!rcu_rdp_is_offloaded(this_cpu_ptr(&rcu_data));
661 }
662 
663 /*
664  * If any sort of urgency was applied to the current CPU (for example,
665  * the scheduler-clock interrupt was enabled on a nohz_full CPU) in order
666  * to get to a quiescent state, disable it.
667  */
668 static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp)
669 {
670 	raw_lockdep_assert_held_rcu_node(rdp->mynode);
671 	WRITE_ONCE(rdp->rcu_urgent_qs, false);
672 	WRITE_ONCE(rdp->rcu_need_heavy_qs, false);
673 	if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) {
674 		tick_dep_clear_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
675 		WRITE_ONCE(rdp->rcu_forced_tick, false);
676 	}
677 }
678 
679 /**
680  * rcu_is_watching - see if RCU thinks that the current CPU is not idle
681  *
682  * Return true if RCU is watching the running CPU, which means that this
683  * CPU can safely enter RCU read-side critical sections.  In other words,
684  * if the current CPU is not in its idle loop or is in an interrupt or
685  * NMI handler, return true.
686  *
687  * Make notrace because it can be called by the internal functions of
688  * ftrace, and making this notrace removes unnecessary recursion calls.
689  */
690 notrace bool rcu_is_watching(void)
691 {
692 	bool ret;
693 
694 	preempt_disable_notrace();
695 	ret = !rcu_dynticks_curr_cpu_in_eqs();
696 	preempt_enable_notrace();
697 	return ret;
698 }
699 EXPORT_SYMBOL_GPL(rcu_is_watching);
700 
701 /*
702  * If a holdout task is actually running, request an urgent quiescent
703  * state from its CPU.  This is unsynchronized, so migrations can cause
704  * the request to go to the wrong CPU.  Which is OK, all that will happen
705  * is that the CPU's next context switch will be a bit slower and next
706  * time around this task will generate another request.
707  */
708 void rcu_request_urgent_qs_task(struct task_struct *t)
709 {
710 	int cpu;
711 
712 	barrier();
713 	cpu = task_cpu(t);
714 	if (!task_curr(t))
715 		return; /* This task is not running on that CPU. */
716 	smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
717 }
718 
719 /*
720  * When trying to report a quiescent state on behalf of some other CPU,
721  * it is our responsibility to check for and handle potential overflow
722  * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
723  * After all, the CPU might be in deep idle state, and thus executing no
724  * code whatsoever.
725  */
726 static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
727 {
728 	raw_lockdep_assert_held_rcu_node(rnp);
729 	if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
730 			 rnp->gp_seq))
731 		WRITE_ONCE(rdp->gpwrap, true);
732 	if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
733 		rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
734 }
735 
736 /*
737  * Snapshot the specified CPU's dynticks counter so that we can later
738  * credit them with an implicit quiescent state.  Return 1 if this CPU
739  * is in dynticks idle mode, which is an extended quiescent state.
740  */
741 static int dyntick_save_progress_counter(struct rcu_data *rdp)
742 {
743 	rdp->dynticks_snap = rcu_dynticks_snap(rdp->cpu);
744 	if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
745 		trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
746 		rcu_gpnum_ovf(rdp->mynode, rdp);
747 		return 1;
748 	}
749 	return 0;
750 }
751 
752 /*
753  * Return true if the specified CPU has passed through a quiescent
754  * state by virtue of being in or having passed through an dynticks
755  * idle state since the last call to dyntick_save_progress_counter()
756  * for this same CPU, or by virtue of having been offline.
757  */
758 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
759 {
760 	unsigned long jtsq;
761 	struct rcu_node *rnp = rdp->mynode;
762 
763 	/*
764 	 * If the CPU passed through or entered a dynticks idle phase with
765 	 * no active irq/NMI handlers, then we can safely pretend that the CPU
766 	 * already acknowledged the request to pass through a quiescent
767 	 * state.  Either way, that CPU cannot possibly be in an RCU
768 	 * read-side critical section that started before the beginning
769 	 * of the current RCU grace period.
770 	 */
771 	if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
772 		trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
773 		rcu_gpnum_ovf(rnp, rdp);
774 		return 1;
775 	}
776 
777 	/*
778 	 * Complain if a CPU that is considered to be offline from RCU's
779 	 * perspective has not yet reported a quiescent state.  After all,
780 	 * the offline CPU should have reported a quiescent state during
781 	 * the CPU-offline process, or, failing that, by rcu_gp_init()
782 	 * if it ran concurrently with either the CPU going offline or the
783 	 * last task on a leaf rcu_node structure exiting its RCU read-side
784 	 * critical section while all CPUs corresponding to that structure
785 	 * are offline.  This added warning detects bugs in any of these
786 	 * code paths.
787 	 *
788 	 * The rcu_node structure's ->lock is held here, which excludes
789 	 * the relevant portions the CPU-hotplug code, the grace-period
790 	 * initialization code, and the rcu_read_unlock() code paths.
791 	 *
792 	 * For more detail, please refer to the "Hotplug CPU" section
793 	 * of RCU's Requirements documentation.
794 	 */
795 	if (WARN_ON_ONCE(!rcu_rdp_cpu_online(rdp))) {
796 		struct rcu_node *rnp1;
797 
798 		pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
799 			__func__, rnp->grplo, rnp->grphi, rnp->level,
800 			(long)rnp->gp_seq, (long)rnp->completedqs);
801 		for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
802 			pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
803 				__func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
804 		pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
805 			__func__, rdp->cpu, ".o"[rcu_rdp_cpu_online(rdp)],
806 			(long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
807 			(long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
808 		return 1; /* Break things loose after complaining. */
809 	}
810 
811 	/*
812 	 * A CPU running for an extended time within the kernel can
813 	 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
814 	 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
815 	 * both .rcu_need_heavy_qs and .rcu_urgent_qs.  Note that the
816 	 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
817 	 * variable are safe because the assignments are repeated if this
818 	 * CPU failed to pass through a quiescent state.  This code
819 	 * also checks .jiffies_resched in case jiffies_to_sched_qs
820 	 * is set way high.
821 	 */
822 	jtsq = READ_ONCE(jiffies_to_sched_qs);
823 	if (!READ_ONCE(rdp->rcu_need_heavy_qs) &&
824 	    (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
825 	     time_after(jiffies, rcu_state.jiffies_resched) ||
826 	     rcu_state.cbovld)) {
827 		WRITE_ONCE(rdp->rcu_need_heavy_qs, true);
828 		/* Store rcu_need_heavy_qs before rcu_urgent_qs. */
829 		smp_store_release(&rdp->rcu_urgent_qs, true);
830 	} else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
831 		WRITE_ONCE(rdp->rcu_urgent_qs, true);
832 	}
833 
834 	/*
835 	 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
836 	 * The above code handles this, but only for straight cond_resched().
837 	 * And some in-kernel loops check need_resched() before calling
838 	 * cond_resched(), which defeats the above code for CPUs that are
839 	 * running in-kernel with scheduling-clock interrupts disabled.
840 	 * So hit them over the head with the resched_cpu() hammer!
841 	 */
842 	if (tick_nohz_full_cpu(rdp->cpu) &&
843 	    (time_after(jiffies, READ_ONCE(rdp->last_fqs_resched) + jtsq * 3) ||
844 	     rcu_state.cbovld)) {
845 		WRITE_ONCE(rdp->rcu_urgent_qs, true);
846 		resched_cpu(rdp->cpu);
847 		WRITE_ONCE(rdp->last_fqs_resched, jiffies);
848 	}
849 
850 	/*
851 	 * If more than halfway to RCU CPU stall-warning time, invoke
852 	 * resched_cpu() more frequently to try to loosen things up a bit.
853 	 * Also check to see if the CPU is getting hammered with interrupts,
854 	 * but only once per grace period, just to keep the IPIs down to
855 	 * a dull roar.
856 	 */
857 	if (time_after(jiffies, rcu_state.jiffies_resched)) {
858 		if (time_after(jiffies,
859 			       READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
860 			resched_cpu(rdp->cpu);
861 			WRITE_ONCE(rdp->last_fqs_resched, jiffies);
862 		}
863 		if (IS_ENABLED(CONFIG_IRQ_WORK) &&
864 		    !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
865 		    (rnp->ffmask & rdp->grpmask)) {
866 			rdp->rcu_iw_pending = true;
867 			rdp->rcu_iw_gp_seq = rnp->gp_seq;
868 			irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
869 		}
870 
871 		if (rcu_cpu_stall_cputime && rdp->snap_record.gp_seq != rdp->gp_seq) {
872 			int cpu = rdp->cpu;
873 			struct rcu_snap_record *rsrp;
874 			struct kernel_cpustat *kcsp;
875 
876 			kcsp = &kcpustat_cpu(cpu);
877 
878 			rsrp = &rdp->snap_record;
879 			rsrp->cputime_irq     = kcpustat_field(kcsp, CPUTIME_IRQ, cpu);
880 			rsrp->cputime_softirq = kcpustat_field(kcsp, CPUTIME_SOFTIRQ, cpu);
881 			rsrp->cputime_system  = kcpustat_field(kcsp, CPUTIME_SYSTEM, cpu);
882 			rsrp->nr_hardirqs = kstat_cpu_irqs_sum(rdp->cpu);
883 			rsrp->nr_softirqs = kstat_cpu_softirqs_sum(rdp->cpu);
884 			rsrp->nr_csw = nr_context_switches_cpu(rdp->cpu);
885 			rsrp->jiffies = jiffies;
886 			rsrp->gp_seq = rdp->gp_seq;
887 		}
888 	}
889 
890 	return 0;
891 }
892 
893 /* Trace-event wrapper function for trace_rcu_future_grace_period.  */
894 static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
895 			      unsigned long gp_seq_req, const char *s)
896 {
897 	trace_rcu_future_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
898 				      gp_seq_req, rnp->level,
899 				      rnp->grplo, rnp->grphi, s);
900 }
901 
902 /*
903  * rcu_start_this_gp - Request the start of a particular grace period
904  * @rnp_start: The leaf node of the CPU from which to start.
905  * @rdp: The rcu_data corresponding to the CPU from which to start.
906  * @gp_seq_req: The gp_seq of the grace period to start.
907  *
908  * Start the specified grace period, as needed to handle newly arrived
909  * callbacks.  The required future grace periods are recorded in each
910  * rcu_node structure's ->gp_seq_needed field.  Returns true if there
911  * is reason to awaken the grace-period kthread.
912  *
913  * The caller must hold the specified rcu_node structure's ->lock, which
914  * is why the caller is responsible for waking the grace-period kthread.
915  *
916  * Returns true if the GP thread needs to be awakened else false.
917  */
918 static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
919 			      unsigned long gp_seq_req)
920 {
921 	bool ret = false;
922 	struct rcu_node *rnp;
923 
924 	/*
925 	 * Use funnel locking to either acquire the root rcu_node
926 	 * structure's lock or bail out if the need for this grace period
927 	 * has already been recorded -- or if that grace period has in
928 	 * fact already started.  If there is already a grace period in
929 	 * progress in a non-leaf node, no recording is needed because the
930 	 * end of the grace period will scan the leaf rcu_node structures.
931 	 * Note that rnp_start->lock must not be released.
932 	 */
933 	raw_lockdep_assert_held_rcu_node(rnp_start);
934 	trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
935 	for (rnp = rnp_start; 1; rnp = rnp->parent) {
936 		if (rnp != rnp_start)
937 			raw_spin_lock_rcu_node(rnp);
938 		if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
939 		    rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
940 		    (rnp != rnp_start &&
941 		     rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
942 			trace_rcu_this_gp(rnp, rdp, gp_seq_req,
943 					  TPS("Prestarted"));
944 			goto unlock_out;
945 		}
946 		WRITE_ONCE(rnp->gp_seq_needed, gp_seq_req);
947 		if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
948 			/*
949 			 * We just marked the leaf or internal node, and a
950 			 * grace period is in progress, which means that
951 			 * rcu_gp_cleanup() will see the marking.  Bail to
952 			 * reduce contention.
953 			 */
954 			trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
955 					  TPS("Startedleaf"));
956 			goto unlock_out;
957 		}
958 		if (rnp != rnp_start && rnp->parent != NULL)
959 			raw_spin_unlock_rcu_node(rnp);
960 		if (!rnp->parent)
961 			break;  /* At root, and perhaps also leaf. */
962 	}
963 
964 	/* If GP already in progress, just leave, otherwise start one. */
965 	if (rcu_gp_in_progress()) {
966 		trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
967 		goto unlock_out;
968 	}
969 	trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
970 	WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
971 	WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
972 	if (!READ_ONCE(rcu_state.gp_kthread)) {
973 		trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
974 		goto unlock_out;
975 	}
976 	trace_rcu_grace_period(rcu_state.name, data_race(rcu_state.gp_seq), TPS("newreq"));
977 	ret = true;  /* Caller must wake GP kthread. */
978 unlock_out:
979 	/* Push furthest requested GP to leaf node and rcu_data structure. */
980 	if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
981 		WRITE_ONCE(rnp_start->gp_seq_needed, rnp->gp_seq_needed);
982 		WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
983 	}
984 	if (rnp != rnp_start)
985 		raw_spin_unlock_rcu_node(rnp);
986 	return ret;
987 }
988 
989 /*
990  * Clean up any old requests for the just-ended grace period.  Also return
991  * whether any additional grace periods have been requested.
992  */
993 static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
994 {
995 	bool needmore;
996 	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
997 
998 	needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
999 	if (!needmore)
1000 		rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1001 	trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1002 			  needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1003 	return needmore;
1004 }
1005 
1006 /*
1007  * Awaken the grace-period kthread.  Don't do a self-awaken (unless in an
1008  * interrupt or softirq handler, in which case we just might immediately
1009  * sleep upon return, resulting in a grace-period hang), and don't bother
1010  * awakening when there is nothing for the grace-period kthread to do
1011  * (as in several CPUs raced to awaken, we lost), and finally don't try
1012  * to awaken a kthread that has not yet been created.  If all those checks
1013  * are passed, track some debug information and awaken.
1014  *
1015  * So why do the self-wakeup when in an interrupt or softirq handler
1016  * in the grace-period kthread's context?  Because the kthread might have
1017  * been interrupted just as it was going to sleep, and just after the final
1018  * pre-sleep check of the awaken condition.  In this case, a wakeup really
1019  * is required, and is therefore supplied.
1020  */
1021 static void rcu_gp_kthread_wake(void)
1022 {
1023 	struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
1024 
1025 	if ((current == t && !in_hardirq() && !in_serving_softirq()) ||
1026 	    !READ_ONCE(rcu_state.gp_flags) || !t)
1027 		return;
1028 	WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
1029 	WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
1030 	swake_up_one(&rcu_state.gp_wq);
1031 }
1032 
1033 /*
1034  * If there is room, assign a ->gp_seq number to any callbacks on this
1035  * CPU that have not already been assigned.  Also accelerate any callbacks
1036  * that were previously assigned a ->gp_seq number that has since proven
1037  * to be too conservative, which can happen if callbacks get assigned a
1038  * ->gp_seq number while RCU is idle, but with reference to a non-root
1039  * rcu_node structure.  This function is idempotent, so it does not hurt
1040  * to call it repeatedly.  Returns an flag saying that we should awaken
1041  * the RCU grace-period kthread.
1042  *
1043  * The caller must hold rnp->lock with interrupts disabled.
1044  */
1045 static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1046 {
1047 	unsigned long gp_seq_req;
1048 	bool ret = false;
1049 
1050 	rcu_lockdep_assert_cblist_protected(rdp);
1051 	raw_lockdep_assert_held_rcu_node(rnp);
1052 
1053 	/* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1054 	if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1055 		return false;
1056 
1057 	trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPreAcc"));
1058 
1059 	/*
1060 	 * Callbacks are often registered with incomplete grace-period
1061 	 * information.  Something about the fact that getting exact
1062 	 * information requires acquiring a global lock...  RCU therefore
1063 	 * makes a conservative estimate of the grace period number at which
1064 	 * a given callback will become ready to invoke.	The following
1065 	 * code checks this estimate and improves it when possible, thus
1066 	 * accelerating callback invocation to an earlier grace-period
1067 	 * number.
1068 	 */
1069 	gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1070 	if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1071 		ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1072 
1073 	/* Trace depending on how much we were able to accelerate. */
1074 	if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1075 		trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccWaitCB"));
1076 	else
1077 		trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccReadyCB"));
1078 
1079 	trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPostAcc"));
1080 
1081 	return ret;
1082 }
1083 
1084 /*
1085  * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1086  * rcu_node structure's ->lock be held.  It consults the cached value
1087  * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1088  * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1089  * while holding the leaf rcu_node structure's ->lock.
1090  */
1091 static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1092 					struct rcu_data *rdp)
1093 {
1094 	unsigned long c;
1095 	bool needwake;
1096 
1097 	rcu_lockdep_assert_cblist_protected(rdp);
1098 	c = rcu_seq_snap(&rcu_state.gp_seq);
1099 	if (!READ_ONCE(rdp->gpwrap) && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1100 		/* Old request still live, so mark recent callbacks. */
1101 		(void)rcu_segcblist_accelerate(&rdp->cblist, c);
1102 		return;
1103 	}
1104 	raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1105 	needwake = rcu_accelerate_cbs(rnp, rdp);
1106 	raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1107 	if (needwake)
1108 		rcu_gp_kthread_wake();
1109 }
1110 
1111 /*
1112  * Move any callbacks whose grace period has completed to the
1113  * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1114  * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1115  * sublist.  This function is idempotent, so it does not hurt to
1116  * invoke it repeatedly.  As long as it is not invoked -too- often...
1117  * Returns true if the RCU grace-period kthread needs to be awakened.
1118  *
1119  * The caller must hold rnp->lock with interrupts disabled.
1120  */
1121 static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1122 {
1123 	rcu_lockdep_assert_cblist_protected(rdp);
1124 	raw_lockdep_assert_held_rcu_node(rnp);
1125 
1126 	/* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1127 	if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1128 		return false;
1129 
1130 	/*
1131 	 * Find all callbacks whose ->gp_seq numbers indicate that they
1132 	 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1133 	 */
1134 	rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1135 
1136 	/* Classify any remaining callbacks. */
1137 	return rcu_accelerate_cbs(rnp, rdp);
1138 }
1139 
1140 /*
1141  * Move and classify callbacks, but only if doing so won't require
1142  * that the RCU grace-period kthread be awakened.
1143  */
1144 static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
1145 						  struct rcu_data *rdp)
1146 {
1147 	rcu_lockdep_assert_cblist_protected(rdp);
1148 	if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) || !raw_spin_trylock_rcu_node(rnp))
1149 		return;
1150 	// The grace period cannot end while we hold the rcu_node lock.
1151 	if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))
1152 		WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
1153 	raw_spin_unlock_rcu_node(rnp);
1154 }
1155 
1156 /*
1157  * In CONFIG_RCU_STRICT_GRACE_PERIOD=y kernels, attempt to generate a
1158  * quiescent state.  This is intended to be invoked when the CPU notices
1159  * a new grace period.
1160  */
1161 static void rcu_strict_gp_check_qs(void)
1162 {
1163 	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
1164 		rcu_read_lock();
1165 		rcu_read_unlock();
1166 	}
1167 }
1168 
1169 /*
1170  * Update CPU-local rcu_data state to record the beginnings and ends of
1171  * grace periods.  The caller must hold the ->lock of the leaf rcu_node
1172  * structure corresponding to the current CPU, and must have irqs disabled.
1173  * Returns true if the grace-period kthread needs to be awakened.
1174  */
1175 static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
1176 {
1177 	bool ret = false;
1178 	bool need_qs;
1179 	const bool offloaded = rcu_rdp_is_offloaded(rdp);
1180 
1181 	raw_lockdep_assert_held_rcu_node(rnp);
1182 
1183 	if (rdp->gp_seq == rnp->gp_seq)
1184 		return false; /* Nothing to do. */
1185 
1186 	/* Handle the ends of any preceding grace periods first. */
1187 	if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1188 	    unlikely(READ_ONCE(rdp->gpwrap))) {
1189 		if (!offloaded)
1190 			ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
1191 		rdp->core_needs_qs = false;
1192 		trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1193 	} else {
1194 		if (!offloaded)
1195 			ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
1196 		if (rdp->core_needs_qs)
1197 			rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask);
1198 	}
1199 
1200 	/* Now handle the beginnings of any new-to-this-CPU grace periods. */
1201 	if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1202 	    unlikely(READ_ONCE(rdp->gpwrap))) {
1203 		/*
1204 		 * If the current grace period is waiting for this CPU,
1205 		 * set up to detect a quiescent state, otherwise don't
1206 		 * go looking for one.
1207 		 */
1208 		trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1209 		need_qs = !!(rnp->qsmask & rdp->grpmask);
1210 		rdp->cpu_no_qs.b.norm = need_qs;
1211 		rdp->core_needs_qs = need_qs;
1212 		zero_cpu_stall_ticks(rdp);
1213 	}
1214 	rdp->gp_seq = rnp->gp_seq;  /* Remember new grace-period state. */
1215 	if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
1216 		WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1217 	if (IS_ENABLED(CONFIG_PROVE_RCU) && READ_ONCE(rdp->gpwrap))
1218 		WRITE_ONCE(rdp->last_sched_clock, jiffies);
1219 	WRITE_ONCE(rdp->gpwrap, false);
1220 	rcu_gpnum_ovf(rnp, rdp);
1221 	return ret;
1222 }
1223 
1224 static void note_gp_changes(struct rcu_data *rdp)
1225 {
1226 	unsigned long flags;
1227 	bool needwake;
1228 	struct rcu_node *rnp;
1229 
1230 	local_irq_save(flags);
1231 	rnp = rdp->mynode;
1232 	if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1233 	     !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1234 	    !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1235 		local_irq_restore(flags);
1236 		return;
1237 	}
1238 	needwake = __note_gp_changes(rnp, rdp);
1239 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1240 	rcu_strict_gp_check_qs();
1241 	if (needwake)
1242 		rcu_gp_kthread_wake();
1243 }
1244 
1245 static atomic_t *rcu_gp_slow_suppress;
1246 
1247 /* Register a counter to suppress debugging grace-period delays. */
1248 void rcu_gp_slow_register(atomic_t *rgssp)
1249 {
1250 	WARN_ON_ONCE(rcu_gp_slow_suppress);
1251 
1252 	WRITE_ONCE(rcu_gp_slow_suppress, rgssp);
1253 }
1254 EXPORT_SYMBOL_GPL(rcu_gp_slow_register);
1255 
1256 /* Unregister a counter, with NULL for not caring which. */
1257 void rcu_gp_slow_unregister(atomic_t *rgssp)
1258 {
1259 	WARN_ON_ONCE(rgssp && rgssp != rcu_gp_slow_suppress);
1260 
1261 	WRITE_ONCE(rcu_gp_slow_suppress, NULL);
1262 }
1263 EXPORT_SYMBOL_GPL(rcu_gp_slow_unregister);
1264 
1265 static bool rcu_gp_slow_is_suppressed(void)
1266 {
1267 	atomic_t *rgssp = READ_ONCE(rcu_gp_slow_suppress);
1268 
1269 	return rgssp && atomic_read(rgssp);
1270 }
1271 
1272 static void rcu_gp_slow(int delay)
1273 {
1274 	if (!rcu_gp_slow_is_suppressed() && delay > 0 &&
1275 	    !(rcu_seq_ctr(rcu_state.gp_seq) % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1276 		schedule_timeout_idle(delay);
1277 }
1278 
1279 static unsigned long sleep_duration;
1280 
1281 /* Allow rcutorture to stall the grace-period kthread. */
1282 void rcu_gp_set_torture_wait(int duration)
1283 {
1284 	if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST) && duration > 0)
1285 		WRITE_ONCE(sleep_duration, duration);
1286 }
1287 EXPORT_SYMBOL_GPL(rcu_gp_set_torture_wait);
1288 
1289 /* Actually implement the aforementioned wait. */
1290 static void rcu_gp_torture_wait(void)
1291 {
1292 	unsigned long duration;
1293 
1294 	if (!IS_ENABLED(CONFIG_RCU_TORTURE_TEST))
1295 		return;
1296 	duration = xchg(&sleep_duration, 0UL);
1297 	if (duration > 0) {
1298 		pr_alert("%s: Waiting %lu jiffies\n", __func__, duration);
1299 		schedule_timeout_idle(duration);
1300 		pr_alert("%s: Wait complete\n", __func__);
1301 	}
1302 }
1303 
1304 /*
1305  * Handler for on_each_cpu() to invoke the target CPU's RCU core
1306  * processing.
1307  */
1308 static void rcu_strict_gp_boundary(void *unused)
1309 {
1310 	invoke_rcu_core();
1311 }
1312 
1313 // Make the polled API aware of the beginning of a grace period.
1314 static void rcu_poll_gp_seq_start(unsigned long *snap)
1315 {
1316 	struct rcu_node *rnp = rcu_get_root();
1317 
1318 	if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1319 		raw_lockdep_assert_held_rcu_node(rnp);
1320 
1321 	// If RCU was idle, note beginning of GP.
1322 	if (!rcu_seq_state(rcu_state.gp_seq_polled))
1323 		rcu_seq_start(&rcu_state.gp_seq_polled);
1324 
1325 	// Either way, record current state.
1326 	*snap = rcu_state.gp_seq_polled;
1327 }
1328 
1329 // Make the polled API aware of the end of a grace period.
1330 static void rcu_poll_gp_seq_end(unsigned long *snap)
1331 {
1332 	struct rcu_node *rnp = rcu_get_root();
1333 
1334 	if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1335 		raw_lockdep_assert_held_rcu_node(rnp);
1336 
1337 	// If the previously noted GP is still in effect, record the
1338 	// end of that GP.  Either way, zero counter to avoid counter-wrap
1339 	// problems.
1340 	if (*snap && *snap == rcu_state.gp_seq_polled) {
1341 		rcu_seq_end(&rcu_state.gp_seq_polled);
1342 		rcu_state.gp_seq_polled_snap = 0;
1343 		rcu_state.gp_seq_polled_exp_snap = 0;
1344 	} else {
1345 		*snap = 0;
1346 	}
1347 }
1348 
1349 // Make the polled API aware of the beginning of a grace period, but
1350 // where caller does not hold the root rcu_node structure's lock.
1351 static void rcu_poll_gp_seq_start_unlocked(unsigned long *snap)
1352 {
1353 	unsigned long flags;
1354 	struct rcu_node *rnp = rcu_get_root();
1355 
1356 	if (rcu_init_invoked()) {
1357 		if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1358 			lockdep_assert_irqs_enabled();
1359 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
1360 	}
1361 	rcu_poll_gp_seq_start(snap);
1362 	if (rcu_init_invoked())
1363 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1364 }
1365 
1366 // Make the polled API aware of the end of a grace period, but where
1367 // caller does not hold the root rcu_node structure's lock.
1368 static void rcu_poll_gp_seq_end_unlocked(unsigned long *snap)
1369 {
1370 	unsigned long flags;
1371 	struct rcu_node *rnp = rcu_get_root();
1372 
1373 	if (rcu_init_invoked()) {
1374 		if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1375 			lockdep_assert_irqs_enabled();
1376 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
1377 	}
1378 	rcu_poll_gp_seq_end(snap);
1379 	if (rcu_init_invoked())
1380 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1381 }
1382 
1383 /*
1384  * Initialize a new grace period.  Return false if no grace period required.
1385  */
1386 static noinline_for_stack bool rcu_gp_init(void)
1387 {
1388 	unsigned long flags;
1389 	unsigned long oldmask;
1390 	unsigned long mask;
1391 	struct rcu_data *rdp;
1392 	struct rcu_node *rnp = rcu_get_root();
1393 
1394 	WRITE_ONCE(rcu_state.gp_activity, jiffies);
1395 	raw_spin_lock_irq_rcu_node(rnp);
1396 	if (!READ_ONCE(rcu_state.gp_flags)) {
1397 		/* Spurious wakeup, tell caller to go back to sleep.  */
1398 		raw_spin_unlock_irq_rcu_node(rnp);
1399 		return false;
1400 	}
1401 	WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1402 
1403 	if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1404 		/*
1405 		 * Grace period already in progress, don't start another.
1406 		 * Not supposed to be able to happen.
1407 		 */
1408 		raw_spin_unlock_irq_rcu_node(rnp);
1409 		return false;
1410 	}
1411 
1412 	/* Advance to a new grace period and initialize state. */
1413 	record_gp_stall_check_time();
1414 	/* Record GP times before starting GP, hence rcu_seq_start(). */
1415 	rcu_seq_start(&rcu_state.gp_seq);
1416 	ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
1417 	trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
1418 	rcu_poll_gp_seq_start(&rcu_state.gp_seq_polled_snap);
1419 	raw_spin_unlock_irq_rcu_node(rnp);
1420 
1421 	/*
1422 	 * Apply per-leaf buffered online and offline operations to
1423 	 * the rcu_node tree. Note that this new grace period need not
1424 	 * wait for subsequent online CPUs, and that RCU hooks in the CPU
1425 	 * offlining path, when combined with checks in this function,
1426 	 * will handle CPUs that are currently going offline or that will
1427 	 * go offline later.  Please also refer to "Hotplug CPU" section
1428 	 * of RCU's Requirements documentation.
1429 	 */
1430 	WRITE_ONCE(rcu_state.gp_state, RCU_GP_ONOFF);
1431 	/* Exclude CPU hotplug operations. */
1432 	rcu_for_each_leaf_node(rnp) {
1433 		local_irq_save(flags);
1434 		arch_spin_lock(&rcu_state.ofl_lock);
1435 		raw_spin_lock_rcu_node(rnp);
1436 		if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1437 		    !rnp->wait_blkd_tasks) {
1438 			/* Nothing to do on this leaf rcu_node structure. */
1439 			raw_spin_unlock_rcu_node(rnp);
1440 			arch_spin_unlock(&rcu_state.ofl_lock);
1441 			local_irq_restore(flags);
1442 			continue;
1443 		}
1444 
1445 		/* Record old state, apply changes to ->qsmaskinit field. */
1446 		oldmask = rnp->qsmaskinit;
1447 		rnp->qsmaskinit = rnp->qsmaskinitnext;
1448 
1449 		/* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1450 		if (!oldmask != !rnp->qsmaskinit) {
1451 			if (!oldmask) { /* First online CPU for rcu_node. */
1452 				if (!rnp->wait_blkd_tasks) /* Ever offline? */
1453 					rcu_init_new_rnp(rnp);
1454 			} else if (rcu_preempt_has_tasks(rnp)) {
1455 				rnp->wait_blkd_tasks = true; /* blocked tasks */
1456 			} else { /* Last offline CPU and can propagate. */
1457 				rcu_cleanup_dead_rnp(rnp);
1458 			}
1459 		}
1460 
1461 		/*
1462 		 * If all waited-on tasks from prior grace period are
1463 		 * done, and if all this rcu_node structure's CPUs are
1464 		 * still offline, propagate up the rcu_node tree and
1465 		 * clear ->wait_blkd_tasks.  Otherwise, if one of this
1466 		 * rcu_node structure's CPUs has since come back online,
1467 		 * simply clear ->wait_blkd_tasks.
1468 		 */
1469 		if (rnp->wait_blkd_tasks &&
1470 		    (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1471 			rnp->wait_blkd_tasks = false;
1472 			if (!rnp->qsmaskinit)
1473 				rcu_cleanup_dead_rnp(rnp);
1474 		}
1475 
1476 		raw_spin_unlock_rcu_node(rnp);
1477 		arch_spin_unlock(&rcu_state.ofl_lock);
1478 		local_irq_restore(flags);
1479 	}
1480 	rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1481 
1482 	/*
1483 	 * Set the quiescent-state-needed bits in all the rcu_node
1484 	 * structures for all currently online CPUs in breadth-first
1485 	 * order, starting from the root rcu_node structure, relying on the
1486 	 * layout of the tree within the rcu_state.node[] array.  Note that
1487 	 * other CPUs will access only the leaves of the hierarchy, thus
1488 	 * seeing that no grace period is in progress, at least until the
1489 	 * corresponding leaf node has been initialized.
1490 	 *
1491 	 * The grace period cannot complete until the initialization
1492 	 * process finishes, because this kthread handles both.
1493 	 */
1494 	WRITE_ONCE(rcu_state.gp_state, RCU_GP_INIT);
1495 	rcu_for_each_node_breadth_first(rnp) {
1496 		rcu_gp_slow(gp_init_delay);
1497 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
1498 		rdp = this_cpu_ptr(&rcu_data);
1499 		rcu_preempt_check_blocked_tasks(rnp);
1500 		rnp->qsmask = rnp->qsmaskinit;
1501 		WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
1502 		if (rnp == rdp->mynode)
1503 			(void)__note_gp_changes(rnp, rdp);
1504 		rcu_preempt_boost_start_gp(rnp);
1505 		trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1506 					    rnp->level, rnp->grplo,
1507 					    rnp->grphi, rnp->qsmask);
1508 		/* Quiescent states for tasks on any now-offline CPUs. */
1509 		mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1510 		rnp->rcu_gp_init_mask = mask;
1511 		if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1512 			rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1513 		else
1514 			raw_spin_unlock_irq_rcu_node(rnp);
1515 		cond_resched_tasks_rcu_qs();
1516 		WRITE_ONCE(rcu_state.gp_activity, jiffies);
1517 	}
1518 
1519 	// If strict, make all CPUs aware of new grace period.
1520 	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
1521 		on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
1522 
1523 	return true;
1524 }
1525 
1526 /*
1527  * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1528  * time.
1529  */
1530 static bool rcu_gp_fqs_check_wake(int *gfp)
1531 {
1532 	struct rcu_node *rnp = rcu_get_root();
1533 
1534 	// If under overload conditions, force an immediate FQS scan.
1535 	if (*gfp & RCU_GP_FLAG_OVLD)
1536 		return true;
1537 
1538 	// Someone like call_rcu() requested a force-quiescent-state scan.
1539 	*gfp = READ_ONCE(rcu_state.gp_flags);
1540 	if (*gfp & RCU_GP_FLAG_FQS)
1541 		return true;
1542 
1543 	// The current grace period has completed.
1544 	if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1545 		return true;
1546 
1547 	return false;
1548 }
1549 
1550 /*
1551  * Do one round of quiescent-state forcing.
1552  */
1553 static void rcu_gp_fqs(bool first_time)
1554 {
1555 	struct rcu_node *rnp = rcu_get_root();
1556 
1557 	WRITE_ONCE(rcu_state.gp_activity, jiffies);
1558 	WRITE_ONCE(rcu_state.n_force_qs, rcu_state.n_force_qs + 1);
1559 	if (first_time) {
1560 		/* Collect dyntick-idle snapshots. */
1561 		force_qs_rnp(dyntick_save_progress_counter);
1562 	} else {
1563 		/* Handle dyntick-idle and offline CPUs. */
1564 		force_qs_rnp(rcu_implicit_dynticks_qs);
1565 	}
1566 	/* Clear flag to prevent immediate re-entry. */
1567 	if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
1568 		raw_spin_lock_irq_rcu_node(rnp);
1569 		WRITE_ONCE(rcu_state.gp_flags,
1570 			   READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
1571 		raw_spin_unlock_irq_rcu_node(rnp);
1572 	}
1573 }
1574 
1575 /*
1576  * Loop doing repeated quiescent-state forcing until the grace period ends.
1577  */
1578 static noinline_for_stack void rcu_gp_fqs_loop(void)
1579 {
1580 	bool first_gp_fqs = true;
1581 	int gf = 0;
1582 	unsigned long j;
1583 	int ret;
1584 	struct rcu_node *rnp = rcu_get_root();
1585 
1586 	j = READ_ONCE(jiffies_till_first_fqs);
1587 	if (rcu_state.cbovld)
1588 		gf = RCU_GP_FLAG_OVLD;
1589 	ret = 0;
1590 	for (;;) {
1591 		if (rcu_state.cbovld) {
1592 			j = (j + 2) / 3;
1593 			if (j <= 0)
1594 				j = 1;
1595 		}
1596 		if (!ret || time_before(jiffies + j, rcu_state.jiffies_force_qs)) {
1597 			WRITE_ONCE(rcu_state.jiffies_force_qs, jiffies + j);
1598 			/*
1599 			 * jiffies_force_qs before RCU_GP_WAIT_FQS state
1600 			 * update; required for stall checks.
1601 			 */
1602 			smp_wmb();
1603 			WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
1604 				   jiffies + (j ? 3 * j : 2));
1605 		}
1606 		trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1607 				       TPS("fqswait"));
1608 		WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_FQS);
1609 		(void)swait_event_idle_timeout_exclusive(rcu_state.gp_wq,
1610 				 rcu_gp_fqs_check_wake(&gf), j);
1611 		rcu_gp_torture_wait();
1612 		WRITE_ONCE(rcu_state.gp_state, RCU_GP_DOING_FQS);
1613 		/* Locking provides needed memory barriers. */
1614 		/*
1615 		 * Exit the loop if the root rcu_node structure indicates that the grace period
1616 		 * has ended, leave the loop.  The rcu_preempt_blocked_readers_cgp(rnp) check
1617 		 * is required only for single-node rcu_node trees because readers blocking
1618 		 * the current grace period are queued only on leaf rcu_node structures.
1619 		 * For multi-node trees, checking the root node's ->qsmask suffices, because a
1620 		 * given root node's ->qsmask bit is cleared only when all CPUs and tasks from
1621 		 * the corresponding leaf nodes have passed through their quiescent state.
1622 		 */
1623 		if (!READ_ONCE(rnp->qsmask) &&
1624 		    !rcu_preempt_blocked_readers_cgp(rnp))
1625 			break;
1626 		/* If time for quiescent-state forcing, do it. */
1627 		if (!time_after(rcu_state.jiffies_force_qs, jiffies) ||
1628 		    (gf & (RCU_GP_FLAG_FQS | RCU_GP_FLAG_OVLD))) {
1629 			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1630 					       TPS("fqsstart"));
1631 			rcu_gp_fqs(first_gp_fqs);
1632 			gf = 0;
1633 			if (first_gp_fqs) {
1634 				first_gp_fqs = false;
1635 				gf = rcu_state.cbovld ? RCU_GP_FLAG_OVLD : 0;
1636 			}
1637 			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1638 					       TPS("fqsend"));
1639 			cond_resched_tasks_rcu_qs();
1640 			WRITE_ONCE(rcu_state.gp_activity, jiffies);
1641 			ret = 0; /* Force full wait till next FQS. */
1642 			j = READ_ONCE(jiffies_till_next_fqs);
1643 		} else {
1644 			/* Deal with stray signal. */
1645 			cond_resched_tasks_rcu_qs();
1646 			WRITE_ONCE(rcu_state.gp_activity, jiffies);
1647 			WARN_ON(signal_pending(current));
1648 			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1649 					       TPS("fqswaitsig"));
1650 			ret = 1; /* Keep old FQS timing. */
1651 			j = jiffies;
1652 			if (time_after(jiffies, rcu_state.jiffies_force_qs))
1653 				j = 1;
1654 			else
1655 				j = rcu_state.jiffies_force_qs - j;
1656 			gf = 0;
1657 		}
1658 	}
1659 }
1660 
1661 /*
1662  * Clean up after the old grace period.
1663  */
1664 static noinline void rcu_gp_cleanup(void)
1665 {
1666 	int cpu;
1667 	bool needgp = false;
1668 	unsigned long gp_duration;
1669 	unsigned long new_gp_seq;
1670 	bool offloaded;
1671 	struct rcu_data *rdp;
1672 	struct rcu_node *rnp = rcu_get_root();
1673 	struct swait_queue_head *sq;
1674 
1675 	WRITE_ONCE(rcu_state.gp_activity, jiffies);
1676 	raw_spin_lock_irq_rcu_node(rnp);
1677 	rcu_state.gp_end = jiffies;
1678 	gp_duration = rcu_state.gp_end - rcu_state.gp_start;
1679 	if (gp_duration > rcu_state.gp_max)
1680 		rcu_state.gp_max = gp_duration;
1681 
1682 	/*
1683 	 * We know the grace period is complete, but to everyone else
1684 	 * it appears to still be ongoing.  But it is also the case
1685 	 * that to everyone else it looks like there is nothing that
1686 	 * they can do to advance the grace period.  It is therefore
1687 	 * safe for us to drop the lock in order to mark the grace
1688 	 * period as completed in all of the rcu_node structures.
1689 	 */
1690 	rcu_poll_gp_seq_end(&rcu_state.gp_seq_polled_snap);
1691 	raw_spin_unlock_irq_rcu_node(rnp);
1692 
1693 	/*
1694 	 * Propagate new ->gp_seq value to rcu_node structures so that
1695 	 * other CPUs don't have to wait until the start of the next grace
1696 	 * period to process their callbacks.  This also avoids some nasty
1697 	 * RCU grace-period initialization races by forcing the end of
1698 	 * the current grace period to be completely recorded in all of
1699 	 * the rcu_node structures before the beginning of the next grace
1700 	 * period is recorded in any of the rcu_node structures.
1701 	 */
1702 	new_gp_seq = rcu_state.gp_seq;
1703 	rcu_seq_end(&new_gp_seq);
1704 	rcu_for_each_node_breadth_first(rnp) {
1705 		raw_spin_lock_irq_rcu_node(rnp);
1706 		if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
1707 			dump_blkd_tasks(rnp, 10);
1708 		WARN_ON_ONCE(rnp->qsmask);
1709 		WRITE_ONCE(rnp->gp_seq, new_gp_seq);
1710 		if (!rnp->parent)
1711 			smp_mb(); // Order against failing poll_state_synchronize_rcu_full().
1712 		rdp = this_cpu_ptr(&rcu_data);
1713 		if (rnp == rdp->mynode)
1714 			needgp = __note_gp_changes(rnp, rdp) || needgp;
1715 		/* smp_mb() provided by prior unlock-lock pair. */
1716 		needgp = rcu_future_gp_cleanup(rnp) || needgp;
1717 		// Reset overload indication for CPUs no longer overloaded
1718 		if (rcu_is_leaf_node(rnp))
1719 			for_each_leaf_node_cpu_mask(rnp, cpu, rnp->cbovldmask) {
1720 				rdp = per_cpu_ptr(&rcu_data, cpu);
1721 				check_cb_ovld_locked(rdp, rnp);
1722 			}
1723 		sq = rcu_nocb_gp_get(rnp);
1724 		raw_spin_unlock_irq_rcu_node(rnp);
1725 		rcu_nocb_gp_cleanup(sq);
1726 		cond_resched_tasks_rcu_qs();
1727 		WRITE_ONCE(rcu_state.gp_activity, jiffies);
1728 		rcu_gp_slow(gp_cleanup_delay);
1729 	}
1730 	rnp = rcu_get_root();
1731 	raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
1732 
1733 	/* Declare grace period done, trace first to use old GP number. */
1734 	trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
1735 	rcu_seq_end(&rcu_state.gp_seq);
1736 	ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
1737 	WRITE_ONCE(rcu_state.gp_state, RCU_GP_IDLE);
1738 	/* Check for GP requests since above loop. */
1739 	rdp = this_cpu_ptr(&rcu_data);
1740 	if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
1741 		trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
1742 				  TPS("CleanupMore"));
1743 		needgp = true;
1744 	}
1745 	/* Advance CBs to reduce false positives below. */
1746 	offloaded = rcu_rdp_is_offloaded(rdp);
1747 	if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
1748 
1749 		// We get here if a grace period was needed (“needgp”)
1750 		// and the above call to rcu_accelerate_cbs() did not set
1751 		// the RCU_GP_FLAG_INIT bit in ->gp_state (which records
1752 		// the need for another grace period).  The purpose
1753 		// of the “offloaded” check is to avoid invoking
1754 		// rcu_accelerate_cbs() on an offloaded CPU because we do not
1755 		// hold the ->nocb_lock needed to safely access an offloaded
1756 		// ->cblist.  We do not want to acquire that lock because
1757 		// it can be heavily contended during callback floods.
1758 
1759 		WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
1760 		WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
1761 		trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("newreq"));
1762 	} else {
1763 
1764 		// We get here either if there is no need for an
1765 		// additional grace period or if rcu_accelerate_cbs() has
1766 		// already set the RCU_GP_FLAG_INIT bit in ->gp_flags. 
1767 		// So all we need to do is to clear all of the other
1768 		// ->gp_flags bits.
1769 
1770 		WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags & RCU_GP_FLAG_INIT);
1771 	}
1772 	raw_spin_unlock_irq_rcu_node(rnp);
1773 
1774 	// If strict, make all CPUs aware of the end of the old grace period.
1775 	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
1776 		on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
1777 }
1778 
1779 /*
1780  * Body of kthread that handles grace periods.
1781  */
1782 static int __noreturn rcu_gp_kthread(void *unused)
1783 {
1784 	rcu_bind_gp_kthread();
1785 	for (;;) {
1786 
1787 		/* Handle grace-period start. */
1788 		for (;;) {
1789 			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1790 					       TPS("reqwait"));
1791 			WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_GPS);
1792 			swait_event_idle_exclusive(rcu_state.gp_wq,
1793 					 READ_ONCE(rcu_state.gp_flags) &
1794 					 RCU_GP_FLAG_INIT);
1795 			rcu_gp_torture_wait();
1796 			WRITE_ONCE(rcu_state.gp_state, RCU_GP_DONE_GPS);
1797 			/* Locking provides needed memory barrier. */
1798 			if (rcu_gp_init())
1799 				break;
1800 			cond_resched_tasks_rcu_qs();
1801 			WRITE_ONCE(rcu_state.gp_activity, jiffies);
1802 			WARN_ON(signal_pending(current));
1803 			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1804 					       TPS("reqwaitsig"));
1805 		}
1806 
1807 		/* Handle quiescent-state forcing. */
1808 		rcu_gp_fqs_loop();
1809 
1810 		/* Handle grace-period end. */
1811 		WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANUP);
1812 		rcu_gp_cleanup();
1813 		WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANED);
1814 	}
1815 }
1816 
1817 /*
1818  * Report a full set of quiescent states to the rcu_state data structure.
1819  * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
1820  * another grace period is required.  Whether we wake the grace-period
1821  * kthread or it awakens itself for the next round of quiescent-state
1822  * forcing, that kthread will clean up after the just-completed grace
1823  * period.  Note that the caller must hold rnp->lock, which is released
1824  * before return.
1825  */
1826 static void rcu_report_qs_rsp(unsigned long flags)
1827 	__releases(rcu_get_root()->lock)
1828 {
1829 	raw_lockdep_assert_held_rcu_node(rcu_get_root());
1830 	WARN_ON_ONCE(!rcu_gp_in_progress());
1831 	WRITE_ONCE(rcu_state.gp_flags,
1832 		   READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
1833 	raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
1834 	rcu_gp_kthread_wake();
1835 }
1836 
1837 /*
1838  * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1839  * Allows quiescent states for a group of CPUs to be reported at one go
1840  * to the specified rcu_node structure, though all the CPUs in the group
1841  * must be represented by the same rcu_node structure (which need not be a
1842  * leaf rcu_node structure, though it often will be).  The gps parameter
1843  * is the grace-period snapshot, which means that the quiescent states
1844  * are valid only if rnp->gp_seq is equal to gps.  That structure's lock
1845  * must be held upon entry, and it is released before return.
1846  *
1847  * As a special case, if mask is zero, the bit-already-cleared check is
1848  * disabled.  This allows propagating quiescent state due to resumed tasks
1849  * during grace-period initialization.
1850  */
1851 static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
1852 			      unsigned long gps, unsigned long flags)
1853 	__releases(rnp->lock)
1854 {
1855 	unsigned long oldmask = 0;
1856 	struct rcu_node *rnp_c;
1857 
1858 	raw_lockdep_assert_held_rcu_node(rnp);
1859 
1860 	/* Walk up the rcu_node hierarchy. */
1861 	for (;;) {
1862 		if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
1863 
1864 			/*
1865 			 * Our bit has already been cleared, or the
1866 			 * relevant grace period is already over, so done.
1867 			 */
1868 			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1869 			return;
1870 		}
1871 		WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
1872 		WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
1873 			     rcu_preempt_blocked_readers_cgp(rnp));
1874 		WRITE_ONCE(rnp->qsmask, rnp->qsmask & ~mask);
1875 		trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
1876 						 mask, rnp->qsmask, rnp->level,
1877 						 rnp->grplo, rnp->grphi,
1878 						 !!rnp->gp_tasks);
1879 		if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1880 
1881 			/* Other bits still set at this level, so done. */
1882 			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1883 			return;
1884 		}
1885 		rnp->completedqs = rnp->gp_seq;
1886 		mask = rnp->grpmask;
1887 		if (rnp->parent == NULL) {
1888 
1889 			/* No more levels.  Exit loop holding root lock. */
1890 
1891 			break;
1892 		}
1893 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1894 		rnp_c = rnp;
1895 		rnp = rnp->parent;
1896 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
1897 		oldmask = READ_ONCE(rnp_c->qsmask);
1898 	}
1899 
1900 	/*
1901 	 * Get here if we are the last CPU to pass through a quiescent
1902 	 * state for this grace period.  Invoke rcu_report_qs_rsp()
1903 	 * to clean up and start the next grace period if one is needed.
1904 	 */
1905 	rcu_report_qs_rsp(flags); /* releases rnp->lock. */
1906 }
1907 
1908 /*
1909  * Record a quiescent state for all tasks that were previously queued
1910  * on the specified rcu_node structure and that were blocking the current
1911  * RCU grace period.  The caller must hold the corresponding rnp->lock with
1912  * irqs disabled, and this lock is released upon return, but irqs remain
1913  * disabled.
1914  */
1915 static void __maybe_unused
1916 rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
1917 	__releases(rnp->lock)
1918 {
1919 	unsigned long gps;
1920 	unsigned long mask;
1921 	struct rcu_node *rnp_p;
1922 
1923 	raw_lockdep_assert_held_rcu_node(rnp);
1924 	if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RCU)) ||
1925 	    WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
1926 	    rnp->qsmask != 0) {
1927 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1928 		return;  /* Still need more quiescent states! */
1929 	}
1930 
1931 	rnp->completedqs = rnp->gp_seq;
1932 	rnp_p = rnp->parent;
1933 	if (rnp_p == NULL) {
1934 		/*
1935 		 * Only one rcu_node structure in the tree, so don't
1936 		 * try to report up to its nonexistent parent!
1937 		 */
1938 		rcu_report_qs_rsp(flags);
1939 		return;
1940 	}
1941 
1942 	/* Report up the rest of the hierarchy, tracking current ->gp_seq. */
1943 	gps = rnp->gp_seq;
1944 	mask = rnp->grpmask;
1945 	raw_spin_unlock_rcu_node(rnp);	/* irqs remain disabled. */
1946 	raw_spin_lock_rcu_node(rnp_p);	/* irqs already disabled. */
1947 	rcu_report_qs_rnp(mask, rnp_p, gps, flags);
1948 }
1949 
1950 /*
1951  * Record a quiescent state for the specified CPU to that CPU's rcu_data
1952  * structure.  This must be called from the specified CPU.
1953  */
1954 static void
1955 rcu_report_qs_rdp(struct rcu_data *rdp)
1956 {
1957 	unsigned long flags;
1958 	unsigned long mask;
1959 	bool needacc = false;
1960 	struct rcu_node *rnp;
1961 
1962 	WARN_ON_ONCE(rdp->cpu != smp_processor_id());
1963 	rnp = rdp->mynode;
1964 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
1965 	if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
1966 	    rdp->gpwrap) {
1967 
1968 		/*
1969 		 * The grace period in which this quiescent state was
1970 		 * recorded has ended, so don't report it upwards.
1971 		 * We will instead need a new quiescent state that lies
1972 		 * within the current grace period.
1973 		 */
1974 		rdp->cpu_no_qs.b.norm = true;	/* need qs for new gp. */
1975 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1976 		return;
1977 	}
1978 	mask = rdp->grpmask;
1979 	rdp->core_needs_qs = false;
1980 	if ((rnp->qsmask & mask) == 0) {
1981 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1982 	} else {
1983 		/*
1984 		 * This GP can't end until cpu checks in, so all of our
1985 		 * callbacks can be processed during the next GP.
1986 		 *
1987 		 * NOCB kthreads have their own way to deal with that...
1988 		 */
1989 		if (!rcu_rdp_is_offloaded(rdp)) {
1990 			/*
1991 			 * The current GP has not yet ended, so it
1992 			 * should not be possible for rcu_accelerate_cbs()
1993 			 * to return true.  So complain, but don't awaken.
1994 			 */
1995 			WARN_ON_ONCE(rcu_accelerate_cbs(rnp, rdp));
1996 		} else if (!rcu_segcblist_completely_offloaded(&rdp->cblist)) {
1997 			/*
1998 			 * ...but NOCB kthreads may miss or delay callbacks acceleration
1999 			 * if in the middle of a (de-)offloading process.
2000 			 */
2001 			needacc = true;
2002 		}
2003 
2004 		rcu_disable_urgency_upon_qs(rdp);
2005 		rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2006 		/* ^^^ Released rnp->lock */
2007 
2008 		if (needacc) {
2009 			rcu_nocb_lock_irqsave(rdp, flags);
2010 			rcu_accelerate_cbs_unlocked(rnp, rdp);
2011 			rcu_nocb_unlock_irqrestore(rdp, flags);
2012 		}
2013 	}
2014 }
2015 
2016 /*
2017  * Check to see if there is a new grace period of which this CPU
2018  * is not yet aware, and if so, set up local rcu_data state for it.
2019  * Otherwise, see if this CPU has just passed through its first
2020  * quiescent state for this grace period, and record that fact if so.
2021  */
2022 static void
2023 rcu_check_quiescent_state(struct rcu_data *rdp)
2024 {
2025 	/* Check for grace-period ends and beginnings. */
2026 	note_gp_changes(rdp);
2027 
2028 	/*
2029 	 * Does this CPU still need to do its part for current grace period?
2030 	 * If no, return and let the other CPUs do their part as well.
2031 	 */
2032 	if (!rdp->core_needs_qs)
2033 		return;
2034 
2035 	/*
2036 	 * Was there a quiescent state since the beginning of the grace
2037 	 * period? If no, then exit and wait for the next call.
2038 	 */
2039 	if (rdp->cpu_no_qs.b.norm)
2040 		return;
2041 
2042 	/*
2043 	 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2044 	 * judge of that).
2045 	 */
2046 	rcu_report_qs_rdp(rdp);
2047 }
2048 
2049 /* Return true if callback-invocation time limit exceeded. */
2050 static bool rcu_do_batch_check_time(long count, long tlimit,
2051 				    bool jlimit_check, unsigned long jlimit)
2052 {
2053 	// Invoke local_clock() only once per 32 consecutive callbacks.
2054 	return unlikely(tlimit) &&
2055 	       (!likely(count & 31) ||
2056 		(IS_ENABLED(CONFIG_RCU_DOUBLE_CHECK_CB_TIME) &&
2057 		 jlimit_check && time_after(jiffies, jlimit))) &&
2058 	       local_clock() >= tlimit;
2059 }
2060 
2061 /*
2062  * Invoke any RCU callbacks that have made it to the end of their grace
2063  * period.  Throttle as specified by rdp->blimit.
2064  */
2065 static void rcu_do_batch(struct rcu_data *rdp)
2066 {
2067 	long bl;
2068 	long count = 0;
2069 	int div;
2070 	bool __maybe_unused empty;
2071 	unsigned long flags;
2072 	unsigned long jlimit;
2073 	bool jlimit_check = false;
2074 	long pending;
2075 	struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2076 	struct rcu_head *rhp;
2077 	long tlimit = 0;
2078 
2079 	/* If no callbacks are ready, just return. */
2080 	if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2081 		trace_rcu_batch_start(rcu_state.name,
2082 				      rcu_segcblist_n_cbs(&rdp->cblist), 0);
2083 		trace_rcu_batch_end(rcu_state.name, 0,
2084 				    !rcu_segcblist_empty(&rdp->cblist),
2085 				    need_resched(), is_idle_task(current),
2086 				    rcu_is_callbacks_kthread(rdp));
2087 		return;
2088 	}
2089 
2090 	/*
2091 	 * Extract the list of ready callbacks, disabling IRQs to prevent
2092 	 * races with call_rcu() from interrupt handlers.  Leave the
2093 	 * callback counts, as rcu_barrier() needs to be conservative.
2094 	 */
2095 	rcu_nocb_lock_irqsave(rdp, flags);
2096 	WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2097 	pending = rcu_segcblist_get_seglen(&rdp->cblist, RCU_DONE_TAIL);
2098 	div = READ_ONCE(rcu_divisor);
2099 	div = div < 0 ? 7 : div > sizeof(long) * 8 - 2 ? sizeof(long) * 8 - 2 : div;
2100 	bl = max(rdp->blimit, pending >> div);
2101 	if ((in_serving_softirq() || rdp->rcu_cpu_kthread_status == RCU_KTHREAD_RUNNING) &&
2102 	    (IS_ENABLED(CONFIG_RCU_DOUBLE_CHECK_CB_TIME) || unlikely(bl > 100))) {
2103 		const long npj = NSEC_PER_SEC / HZ;
2104 		long rrn = READ_ONCE(rcu_resched_ns);
2105 
2106 		rrn = rrn < NSEC_PER_MSEC ? NSEC_PER_MSEC : rrn > NSEC_PER_SEC ? NSEC_PER_SEC : rrn;
2107 		tlimit = local_clock() + rrn;
2108 		jlimit = jiffies + (rrn + npj + 1) / npj;
2109 		jlimit_check = true;
2110 	}
2111 	trace_rcu_batch_start(rcu_state.name,
2112 			      rcu_segcblist_n_cbs(&rdp->cblist), bl);
2113 	rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2114 	if (rcu_rdp_is_offloaded(rdp))
2115 		rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2116 
2117 	trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbDequeued"));
2118 	rcu_nocb_unlock_irqrestore(rdp, flags);
2119 
2120 	/* Invoke callbacks. */
2121 	tick_dep_set_task(current, TICK_DEP_BIT_RCU);
2122 	rhp = rcu_cblist_dequeue(&rcl);
2123 
2124 	for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2125 		rcu_callback_t f;
2126 
2127 		count++;
2128 		debug_rcu_head_unqueue(rhp);
2129 
2130 		rcu_lock_acquire(&rcu_callback_map);
2131 		trace_rcu_invoke_callback(rcu_state.name, rhp);
2132 
2133 		f = rhp->func;
2134 		WRITE_ONCE(rhp->func, (rcu_callback_t)0L);
2135 		f(rhp);
2136 
2137 		rcu_lock_release(&rcu_callback_map);
2138 
2139 		/*
2140 		 * Stop only if limit reached and CPU has something to do.
2141 		 */
2142 		if (in_serving_softirq()) {
2143 			if (count >= bl && (need_resched() || !is_idle_task(current)))
2144 				break;
2145 			/*
2146 			 * Make sure we don't spend too much time here and deprive other
2147 			 * softirq vectors of CPU cycles.
2148 			 */
2149 			if (rcu_do_batch_check_time(count, tlimit, jlimit_check, jlimit))
2150 				break;
2151 		} else {
2152 			// In rcuc/rcuoc context, so no worries about
2153 			// depriving other softirq vectors of CPU cycles.
2154 			local_bh_enable();
2155 			lockdep_assert_irqs_enabled();
2156 			cond_resched_tasks_rcu_qs();
2157 			lockdep_assert_irqs_enabled();
2158 			local_bh_disable();
2159 			// But rcuc kthreads can delay quiescent-state
2160 			// reporting, so check time limits for them.
2161 			if (rdp->rcu_cpu_kthread_status == RCU_KTHREAD_RUNNING &&
2162 			    rcu_do_batch_check_time(count, tlimit, jlimit_check, jlimit)) {
2163 				rdp->rcu_cpu_has_work = 1;
2164 				break;
2165 			}
2166 		}
2167 	}
2168 
2169 	rcu_nocb_lock_irqsave(rdp, flags);
2170 	rdp->n_cbs_invoked += count;
2171 	trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2172 			    is_idle_task(current), rcu_is_callbacks_kthread(rdp));
2173 
2174 	/* Update counts and requeue any remaining callbacks. */
2175 	rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2176 	rcu_segcblist_add_len(&rdp->cblist, -count);
2177 
2178 	/* Reinstate batch limit if we have worked down the excess. */
2179 	count = rcu_segcblist_n_cbs(&rdp->cblist);
2180 	if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
2181 		rdp->blimit = blimit;
2182 
2183 	/* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2184 	if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2185 		rdp->qlen_last_fqs_check = 0;
2186 		rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
2187 	} else if (count < rdp->qlen_last_fqs_check - qhimark)
2188 		rdp->qlen_last_fqs_check = count;
2189 
2190 	/*
2191 	 * The following usually indicates a double call_rcu().  To track
2192 	 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2193 	 */
2194 	empty = rcu_segcblist_empty(&rdp->cblist);
2195 	WARN_ON_ONCE(count == 0 && !empty);
2196 	WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2197 		     count != 0 && empty);
2198 	WARN_ON_ONCE(count == 0 && rcu_segcblist_n_segment_cbs(&rdp->cblist) != 0);
2199 	WARN_ON_ONCE(!empty && rcu_segcblist_n_segment_cbs(&rdp->cblist) == 0);
2200 
2201 	rcu_nocb_unlock_irqrestore(rdp, flags);
2202 
2203 	tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
2204 }
2205 
2206 /*
2207  * This function is invoked from each scheduling-clock interrupt,
2208  * and checks to see if this CPU is in a non-context-switch quiescent
2209  * state, for example, user mode or idle loop.  It also schedules RCU
2210  * core processing.  If the current grace period has gone on too long,
2211  * it will ask the scheduler to manufacture a context switch for the sole
2212  * purpose of providing the needed quiescent state.
2213  */
2214 void rcu_sched_clock_irq(int user)
2215 {
2216 	unsigned long j;
2217 
2218 	if (IS_ENABLED(CONFIG_PROVE_RCU)) {
2219 		j = jiffies;
2220 		WARN_ON_ONCE(time_before(j, __this_cpu_read(rcu_data.last_sched_clock)));
2221 		__this_cpu_write(rcu_data.last_sched_clock, j);
2222 	}
2223 	trace_rcu_utilization(TPS("Start scheduler-tick"));
2224 	lockdep_assert_irqs_disabled();
2225 	raw_cpu_inc(rcu_data.ticks_this_gp);
2226 	/* The load-acquire pairs with the store-release setting to true. */
2227 	if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
2228 		/* Idle and userspace execution already are quiescent states. */
2229 		if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2230 			set_tsk_need_resched(current);
2231 			set_preempt_need_resched();
2232 		}
2233 		__this_cpu_write(rcu_data.rcu_urgent_qs, false);
2234 	}
2235 	rcu_flavor_sched_clock_irq(user);
2236 	if (rcu_pending(user))
2237 		invoke_rcu_core();
2238 	if (user || rcu_is_cpu_rrupt_from_idle())
2239 		rcu_note_voluntary_context_switch(current);
2240 	lockdep_assert_irqs_disabled();
2241 
2242 	trace_rcu_utilization(TPS("End scheduler-tick"));
2243 }
2244 
2245 /*
2246  * Scan the leaf rcu_node structures.  For each structure on which all
2247  * CPUs have reported a quiescent state and on which there are tasks
2248  * blocking the current grace period, initiate RCU priority boosting.
2249  * Otherwise, invoke the specified function to check dyntick state for
2250  * each CPU that has not yet reported a quiescent state.
2251  */
2252 static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
2253 {
2254 	int cpu;
2255 	unsigned long flags;
2256 	unsigned long mask;
2257 	struct rcu_data *rdp;
2258 	struct rcu_node *rnp;
2259 
2260 	rcu_state.cbovld = rcu_state.cbovldnext;
2261 	rcu_state.cbovldnext = false;
2262 	rcu_for_each_leaf_node(rnp) {
2263 		cond_resched_tasks_rcu_qs();
2264 		mask = 0;
2265 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
2266 		rcu_state.cbovldnext |= !!rnp->cbovldmask;
2267 		if (rnp->qsmask == 0) {
2268 			if (rcu_preempt_blocked_readers_cgp(rnp)) {
2269 				/*
2270 				 * No point in scanning bits because they
2271 				 * are all zero.  But we might need to
2272 				 * priority-boost blocked readers.
2273 				 */
2274 				rcu_initiate_boost(rnp, flags);
2275 				/* rcu_initiate_boost() releases rnp->lock */
2276 				continue;
2277 			}
2278 			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2279 			continue;
2280 		}
2281 		for_each_leaf_node_cpu_mask(rnp, cpu, rnp->qsmask) {
2282 			rdp = per_cpu_ptr(&rcu_data, cpu);
2283 			if (f(rdp)) {
2284 				mask |= rdp->grpmask;
2285 				rcu_disable_urgency_upon_qs(rdp);
2286 			}
2287 		}
2288 		if (mask != 0) {
2289 			/* Idle/offline CPUs, report (releases rnp->lock). */
2290 			rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2291 		} else {
2292 			/* Nothing to do here, so just drop the lock. */
2293 			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2294 		}
2295 	}
2296 }
2297 
2298 /*
2299  * Force quiescent states on reluctant CPUs, and also detect which
2300  * CPUs are in dyntick-idle mode.
2301  */
2302 void rcu_force_quiescent_state(void)
2303 {
2304 	unsigned long flags;
2305 	bool ret;
2306 	struct rcu_node *rnp;
2307 	struct rcu_node *rnp_old = NULL;
2308 
2309 	/* Funnel through hierarchy to reduce memory contention. */
2310 	rnp = raw_cpu_read(rcu_data.mynode);
2311 	for (; rnp != NULL; rnp = rnp->parent) {
2312 		ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2313 		       !raw_spin_trylock(&rnp->fqslock);
2314 		if (rnp_old != NULL)
2315 			raw_spin_unlock(&rnp_old->fqslock);
2316 		if (ret)
2317 			return;
2318 		rnp_old = rnp;
2319 	}
2320 	/* rnp_old == rcu_get_root(), rnp == NULL. */
2321 
2322 	/* Reached the root of the rcu_node tree, acquire lock. */
2323 	raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2324 	raw_spin_unlock(&rnp_old->fqslock);
2325 	if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2326 		raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2327 		return;  /* Someone beat us to it. */
2328 	}
2329 	WRITE_ONCE(rcu_state.gp_flags,
2330 		   READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2331 	raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2332 	rcu_gp_kthread_wake();
2333 }
2334 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
2335 
2336 // Workqueue handler for an RCU reader for kernels enforcing struct RCU
2337 // grace periods.
2338 static void strict_work_handler(struct work_struct *work)
2339 {
2340 	rcu_read_lock();
2341 	rcu_read_unlock();
2342 }
2343 
2344 /* Perform RCU core processing work for the current CPU.  */
2345 static __latent_entropy void rcu_core(void)
2346 {
2347 	unsigned long flags;
2348 	struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2349 	struct rcu_node *rnp = rdp->mynode;
2350 	/*
2351 	 * On RT rcu_core() can be preempted when IRQs aren't disabled.
2352 	 * Therefore this function can race with concurrent NOCB (de-)offloading
2353 	 * on this CPU and the below condition must be considered volatile.
2354 	 * However if we race with:
2355 	 *
2356 	 * _ Offloading:   In the worst case we accelerate or process callbacks
2357 	 *                 concurrently with NOCB kthreads. We are guaranteed to
2358 	 *                 call rcu_nocb_lock() if that happens.
2359 	 *
2360 	 * _ Deoffloading: In the worst case we miss callbacks acceleration or
2361 	 *                 processing. This is fine because the early stage
2362 	 *                 of deoffloading invokes rcu_core() after setting
2363 	 *                 SEGCBLIST_RCU_CORE. So we guarantee that we'll process
2364 	 *                 what could have been dismissed without the need to wait
2365 	 *                 for the next rcu_pending() check in the next jiffy.
2366 	 */
2367 	const bool do_batch = !rcu_segcblist_completely_offloaded(&rdp->cblist);
2368 
2369 	if (cpu_is_offline(smp_processor_id()))
2370 		return;
2371 	trace_rcu_utilization(TPS("Start RCU core"));
2372 	WARN_ON_ONCE(!rdp->beenonline);
2373 
2374 	/* Report any deferred quiescent states if preemption enabled. */
2375 	if (IS_ENABLED(CONFIG_PREEMPT_COUNT) && (!(preempt_count() & PREEMPT_MASK))) {
2376 		rcu_preempt_deferred_qs(current);
2377 	} else if (rcu_preempt_need_deferred_qs(current)) {
2378 		set_tsk_need_resched(current);
2379 		set_preempt_need_resched();
2380 	}
2381 
2382 	/* Update RCU state based on any recent quiescent states. */
2383 	rcu_check_quiescent_state(rdp);
2384 
2385 	/* No grace period and unregistered callbacks? */
2386 	if (!rcu_gp_in_progress() &&
2387 	    rcu_segcblist_is_enabled(&rdp->cblist) && do_batch) {
2388 		rcu_nocb_lock_irqsave(rdp, flags);
2389 		if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2390 			rcu_accelerate_cbs_unlocked(rnp, rdp);
2391 		rcu_nocb_unlock_irqrestore(rdp, flags);
2392 	}
2393 
2394 	rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
2395 
2396 	/* If there are callbacks ready, invoke them. */
2397 	if (do_batch && rcu_segcblist_ready_cbs(&rdp->cblist) &&
2398 	    likely(READ_ONCE(rcu_scheduler_fully_active))) {
2399 		rcu_do_batch(rdp);
2400 		/* Re-invoke RCU core processing if there are callbacks remaining. */
2401 		if (rcu_segcblist_ready_cbs(&rdp->cblist))
2402 			invoke_rcu_core();
2403 	}
2404 
2405 	/* Do any needed deferred wakeups of rcuo kthreads. */
2406 	do_nocb_deferred_wakeup(rdp);
2407 	trace_rcu_utilization(TPS("End RCU core"));
2408 
2409 	// If strict GPs, schedule an RCU reader in a clean environment.
2410 	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
2411 		queue_work_on(rdp->cpu, rcu_gp_wq, &rdp->strict_work);
2412 }
2413 
2414 static void rcu_core_si(struct softirq_action *h)
2415 {
2416 	rcu_core();
2417 }
2418 
2419 static void rcu_wake_cond(struct task_struct *t, int status)
2420 {
2421 	/*
2422 	 * If the thread is yielding, only wake it when this
2423 	 * is invoked from idle
2424 	 */
2425 	if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
2426 		wake_up_process(t);
2427 }
2428 
2429 static void invoke_rcu_core_kthread(void)
2430 {
2431 	struct task_struct *t;
2432 	unsigned long flags;
2433 
2434 	local_irq_save(flags);
2435 	__this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
2436 	t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
2437 	if (t != NULL && t != current)
2438 		rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
2439 	local_irq_restore(flags);
2440 }
2441 
2442 /*
2443  * Wake up this CPU's rcuc kthread to do RCU core processing.
2444  */
2445 static void invoke_rcu_core(void)
2446 {
2447 	if (!cpu_online(smp_processor_id()))
2448 		return;
2449 	if (use_softirq)
2450 		raise_softirq(RCU_SOFTIRQ);
2451 	else
2452 		invoke_rcu_core_kthread();
2453 }
2454 
2455 static void rcu_cpu_kthread_park(unsigned int cpu)
2456 {
2457 	per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
2458 }
2459 
2460 static int rcu_cpu_kthread_should_run(unsigned int cpu)
2461 {
2462 	return __this_cpu_read(rcu_data.rcu_cpu_has_work);
2463 }
2464 
2465 /*
2466  * Per-CPU kernel thread that invokes RCU callbacks.  This replaces
2467  * the RCU softirq used in configurations of RCU that do not support RCU
2468  * priority boosting.
2469  */
2470 static void rcu_cpu_kthread(unsigned int cpu)
2471 {
2472 	unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
2473 	char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
2474 	unsigned long *j = this_cpu_ptr(&rcu_data.rcuc_activity);
2475 	int spincnt;
2476 
2477 	trace_rcu_utilization(TPS("Start CPU kthread@rcu_run"));
2478 	for (spincnt = 0; spincnt < 10; spincnt++) {
2479 		WRITE_ONCE(*j, jiffies);
2480 		local_bh_disable();
2481 		*statusp = RCU_KTHREAD_RUNNING;
2482 		local_irq_disable();
2483 		work = *workp;
2484 		WRITE_ONCE(*workp, 0);
2485 		local_irq_enable();
2486 		if (work)
2487 			rcu_core();
2488 		local_bh_enable();
2489 		if (!READ_ONCE(*workp)) {
2490 			trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2491 			*statusp = RCU_KTHREAD_WAITING;
2492 			return;
2493 		}
2494 	}
2495 	*statusp = RCU_KTHREAD_YIELDING;
2496 	trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2497 	schedule_timeout_idle(2);
2498 	trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2499 	*statusp = RCU_KTHREAD_WAITING;
2500 	WRITE_ONCE(*j, jiffies);
2501 }
2502 
2503 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
2504 	.store			= &rcu_data.rcu_cpu_kthread_task,
2505 	.thread_should_run	= rcu_cpu_kthread_should_run,
2506 	.thread_fn		= rcu_cpu_kthread,
2507 	.thread_comm		= "rcuc/%u",
2508 	.setup			= rcu_cpu_kthread_setup,
2509 	.park			= rcu_cpu_kthread_park,
2510 };
2511 
2512 /*
2513  * Spawn per-CPU RCU core processing kthreads.
2514  */
2515 static int __init rcu_spawn_core_kthreads(void)
2516 {
2517 	int cpu;
2518 
2519 	for_each_possible_cpu(cpu)
2520 		per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
2521 	if (use_softirq)
2522 		return 0;
2523 	WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
2524 		  "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
2525 	return 0;
2526 }
2527 
2528 /*
2529  * Handle any core-RCU processing required by a call_rcu() invocation.
2530  */
2531 static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2532 			    unsigned long flags)
2533 {
2534 	/*
2535 	 * If called from an extended quiescent state, invoke the RCU
2536 	 * core in order to force a re-evaluation of RCU's idleness.
2537 	 */
2538 	if (!rcu_is_watching())
2539 		invoke_rcu_core();
2540 
2541 	/* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2542 	if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2543 		return;
2544 
2545 	/*
2546 	 * Force the grace period if too many callbacks or too long waiting.
2547 	 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2548 	 * if some other CPU has recently done so.  Also, don't bother
2549 	 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2550 	 * is the only one waiting for a grace period to complete.
2551 	 */
2552 	if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2553 		     rdp->qlen_last_fqs_check + qhimark)) {
2554 
2555 		/* Are we ignoring a completed grace period? */
2556 		note_gp_changes(rdp);
2557 
2558 		/* Start a new grace period if one not already started. */
2559 		if (!rcu_gp_in_progress()) {
2560 			rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
2561 		} else {
2562 			/* Give the grace period a kick. */
2563 			rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
2564 			if (READ_ONCE(rcu_state.n_force_qs) == rdp->n_force_qs_snap &&
2565 			    rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2566 				rcu_force_quiescent_state();
2567 			rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
2568 			rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2569 		}
2570 	}
2571 }
2572 
2573 /*
2574  * RCU callback function to leak a callback.
2575  */
2576 static void rcu_leak_callback(struct rcu_head *rhp)
2577 {
2578 }
2579 
2580 /*
2581  * Check and if necessary update the leaf rcu_node structure's
2582  * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2583  * number of queued RCU callbacks.  The caller must hold the leaf rcu_node
2584  * structure's ->lock.
2585  */
2586 static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp)
2587 {
2588 	raw_lockdep_assert_held_rcu_node(rnp);
2589 	if (qovld_calc <= 0)
2590 		return; // Early boot and wildcard value set.
2591 	if (rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc)
2592 		WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask | rdp->grpmask);
2593 	else
2594 		WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask & ~rdp->grpmask);
2595 }
2596 
2597 /*
2598  * Check and if necessary update the leaf rcu_node structure's
2599  * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2600  * number of queued RCU callbacks.  No locks need be held, but the
2601  * caller must have disabled interrupts.
2602  *
2603  * Note that this function ignores the possibility that there are a lot
2604  * of callbacks all of which have already seen the end of their respective
2605  * grace periods.  This omission is due to the need for no-CBs CPUs to
2606  * be holding ->nocb_lock to do this check, which is too heavy for a
2607  * common-case operation.
2608  */
2609 static void check_cb_ovld(struct rcu_data *rdp)
2610 {
2611 	struct rcu_node *const rnp = rdp->mynode;
2612 
2613 	if (qovld_calc <= 0 ||
2614 	    ((rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) ==
2615 	     !!(READ_ONCE(rnp->cbovldmask) & rdp->grpmask)))
2616 		return; // Early boot wildcard value or already set correctly.
2617 	raw_spin_lock_rcu_node(rnp);
2618 	check_cb_ovld_locked(rdp, rnp);
2619 	raw_spin_unlock_rcu_node(rnp);
2620 }
2621 
2622 static void
2623 __call_rcu_common(struct rcu_head *head, rcu_callback_t func, bool lazy_in)
2624 {
2625 	static atomic_t doublefrees;
2626 	unsigned long flags;
2627 	bool lazy;
2628 	struct rcu_data *rdp;
2629 	bool was_alldone;
2630 
2631 	/* Misaligned rcu_head! */
2632 	WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2633 
2634 	if (debug_rcu_head_queue(head)) {
2635 		/*
2636 		 * Probable double call_rcu(), so leak the callback.
2637 		 * Use rcu:rcu_callback trace event to find the previous
2638 		 * time callback was passed to call_rcu().
2639 		 */
2640 		if (atomic_inc_return(&doublefrees) < 4) {
2641 			pr_err("%s(): Double-freed CB %p->%pS()!!!  ", __func__, head, head->func);
2642 			mem_dump_obj(head);
2643 		}
2644 		WRITE_ONCE(head->func, rcu_leak_callback);
2645 		return;
2646 	}
2647 	head->func = func;
2648 	head->next = NULL;
2649 	kasan_record_aux_stack_noalloc(head);
2650 	local_irq_save(flags);
2651 	rdp = this_cpu_ptr(&rcu_data);
2652 	lazy = lazy_in && !rcu_async_should_hurry();
2653 
2654 	/* Add the callback to our list. */
2655 	if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
2656 		// This can trigger due to call_rcu() from offline CPU:
2657 		WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
2658 		WARN_ON_ONCE(!rcu_is_watching());
2659 		// Very early boot, before rcu_init().  Initialize if needed
2660 		// and then drop through to queue the callback.
2661 		if (rcu_segcblist_empty(&rdp->cblist))
2662 			rcu_segcblist_init(&rdp->cblist);
2663 	}
2664 
2665 	check_cb_ovld(rdp);
2666 	if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags, lazy))
2667 		return; // Enqueued onto ->nocb_bypass, so just leave.
2668 	// If no-CBs CPU gets here, rcu_nocb_try_bypass() acquired ->nocb_lock.
2669 	rcu_segcblist_enqueue(&rdp->cblist, head);
2670 	if (__is_kvfree_rcu_offset((unsigned long)func))
2671 		trace_rcu_kvfree_callback(rcu_state.name, head,
2672 					 (unsigned long)func,
2673 					 rcu_segcblist_n_cbs(&rdp->cblist));
2674 	else
2675 		trace_rcu_callback(rcu_state.name, head,
2676 				   rcu_segcblist_n_cbs(&rdp->cblist));
2677 
2678 	trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCBQueued"));
2679 
2680 	/* Go handle any RCU core processing required. */
2681 	if (unlikely(rcu_rdp_is_offloaded(rdp))) {
2682 		__call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */
2683 	} else {
2684 		__call_rcu_core(rdp, head, flags);
2685 		local_irq_restore(flags);
2686 	}
2687 }
2688 
2689 #ifdef CONFIG_RCU_LAZY
2690 /**
2691  * call_rcu_hurry() - Queue RCU callback for invocation after grace period, and
2692  * flush all lazy callbacks (including the new one) to the main ->cblist while
2693  * doing so.
2694  *
2695  * @head: structure to be used for queueing the RCU updates.
2696  * @func: actual callback function to be invoked after the grace period
2697  *
2698  * The callback function will be invoked some time after a full grace
2699  * period elapses, in other words after all pre-existing RCU read-side
2700  * critical sections have completed.
2701  *
2702  * Use this API instead of call_rcu() if you don't want the callback to be
2703  * invoked after very long periods of time, which can happen on systems without
2704  * memory pressure and on systems which are lightly loaded or mostly idle.
2705  * This function will cause callbacks to be invoked sooner than later at the
2706  * expense of extra power. Other than that, this function is identical to, and
2707  * reuses call_rcu()'s logic. Refer to call_rcu() for more details about memory
2708  * ordering and other functionality.
2709  */
2710 void call_rcu_hurry(struct rcu_head *head, rcu_callback_t func)
2711 {
2712 	return __call_rcu_common(head, func, false);
2713 }
2714 EXPORT_SYMBOL_GPL(call_rcu_hurry);
2715 #endif
2716 
2717 /**
2718  * call_rcu() - Queue an RCU callback for invocation after a grace period.
2719  * By default the callbacks are 'lazy' and are kept hidden from the main
2720  * ->cblist to prevent starting of grace periods too soon.
2721  * If you desire grace periods to start very soon, use call_rcu_hurry().
2722  *
2723  * @head: structure to be used for queueing the RCU updates.
2724  * @func: actual callback function to be invoked after the grace period
2725  *
2726  * The callback function will be invoked some time after a full grace
2727  * period elapses, in other words after all pre-existing RCU read-side
2728  * critical sections have completed.  However, the callback function
2729  * might well execute concurrently with RCU read-side critical sections
2730  * that started after call_rcu() was invoked.
2731  *
2732  * RCU read-side critical sections are delimited by rcu_read_lock()
2733  * and rcu_read_unlock(), and may be nested.  In addition, but only in
2734  * v5.0 and later, regions of code across which interrupts, preemption,
2735  * or softirqs have been disabled also serve as RCU read-side critical
2736  * sections.  This includes hardware interrupt handlers, softirq handlers,
2737  * and NMI handlers.
2738  *
2739  * Note that all CPUs must agree that the grace period extended beyond
2740  * all pre-existing RCU read-side critical section.  On systems with more
2741  * than one CPU, this means that when "func()" is invoked, each CPU is
2742  * guaranteed to have executed a full memory barrier since the end of its
2743  * last RCU read-side critical section whose beginning preceded the call
2744  * to call_rcu().  It also means that each CPU executing an RCU read-side
2745  * critical section that continues beyond the start of "func()" must have
2746  * executed a memory barrier after the call_rcu() but before the beginning
2747  * of that RCU read-side critical section.  Note that these guarantees
2748  * include CPUs that are offline, idle, or executing in user mode, as
2749  * well as CPUs that are executing in the kernel.
2750  *
2751  * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
2752  * resulting RCU callback function "func()", then both CPU A and CPU B are
2753  * guaranteed to execute a full memory barrier during the time interval
2754  * between the call to call_rcu() and the invocation of "func()" -- even
2755  * if CPU A and CPU B are the same CPU (but again only if the system has
2756  * more than one CPU).
2757  *
2758  * Implementation of these memory-ordering guarantees is described here:
2759  * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
2760  */
2761 void call_rcu(struct rcu_head *head, rcu_callback_t func)
2762 {
2763 	return __call_rcu_common(head, func, IS_ENABLED(CONFIG_RCU_LAZY));
2764 }
2765 EXPORT_SYMBOL_GPL(call_rcu);
2766 
2767 /* Maximum number of jiffies to wait before draining a batch. */
2768 #define KFREE_DRAIN_JIFFIES (5 * HZ)
2769 #define KFREE_N_BATCHES 2
2770 #define FREE_N_CHANNELS 2
2771 
2772 /**
2773  * struct kvfree_rcu_bulk_data - single block to store kvfree_rcu() pointers
2774  * @list: List node. All blocks are linked between each other
2775  * @gp_snap: Snapshot of RCU state for objects placed to this bulk
2776  * @nr_records: Number of active pointers in the array
2777  * @records: Array of the kvfree_rcu() pointers
2778  */
2779 struct kvfree_rcu_bulk_data {
2780 	struct list_head list;
2781 	struct rcu_gp_oldstate gp_snap;
2782 	unsigned long nr_records;
2783 	void *records[];
2784 };
2785 
2786 /*
2787  * This macro defines how many entries the "records" array
2788  * will contain. It is based on the fact that the size of
2789  * kvfree_rcu_bulk_data structure becomes exactly one page.
2790  */
2791 #define KVFREE_BULK_MAX_ENTR \
2792 	((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *))
2793 
2794 /**
2795  * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
2796  * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
2797  * @head_free: List of kfree_rcu() objects waiting for a grace period
2798  * @head_free_gp_snap: Grace-period snapshot to check for attempted premature frees.
2799  * @bulk_head_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
2800  * @krcp: Pointer to @kfree_rcu_cpu structure
2801  */
2802 
2803 struct kfree_rcu_cpu_work {
2804 	struct rcu_work rcu_work;
2805 	struct rcu_head *head_free;
2806 	struct rcu_gp_oldstate head_free_gp_snap;
2807 	struct list_head bulk_head_free[FREE_N_CHANNELS];
2808 	struct kfree_rcu_cpu *krcp;
2809 };
2810 
2811 /**
2812  * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
2813  * @head: List of kfree_rcu() objects not yet waiting for a grace period
2814  * @head_gp_snap: Snapshot of RCU state for objects placed to "@head"
2815  * @bulk_head: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period
2816  * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
2817  * @lock: Synchronize access to this structure
2818  * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES
2819  * @initialized: The @rcu_work fields have been initialized
2820  * @head_count: Number of objects in rcu_head singular list
2821  * @bulk_count: Number of objects in bulk-list
2822  * @bkvcache:
2823  *	A simple cache list that contains objects for reuse purpose.
2824  *	In order to save some per-cpu space the list is singular.
2825  *	Even though it is lockless an access has to be protected by the
2826  *	per-cpu lock.
2827  * @page_cache_work: A work to refill the cache when it is empty
2828  * @backoff_page_cache_fill: Delay cache refills
2829  * @work_in_progress: Indicates that page_cache_work is running
2830  * @hrtimer: A hrtimer for scheduling a page_cache_work
2831  * @nr_bkv_objs: number of allocated objects at @bkvcache.
2832  *
2833  * This is a per-CPU structure.  The reason that it is not included in
2834  * the rcu_data structure is to permit this code to be extracted from
2835  * the RCU files.  Such extraction could allow further optimization of
2836  * the interactions with the slab allocators.
2837  */
2838 struct kfree_rcu_cpu {
2839 	// Objects queued on a linked list
2840 	// through their rcu_head structures.
2841 	struct rcu_head *head;
2842 	unsigned long head_gp_snap;
2843 	atomic_t head_count;
2844 
2845 	// Objects queued on a bulk-list.
2846 	struct list_head bulk_head[FREE_N_CHANNELS];
2847 	atomic_t bulk_count[FREE_N_CHANNELS];
2848 
2849 	struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
2850 	raw_spinlock_t lock;
2851 	struct delayed_work monitor_work;
2852 	bool initialized;
2853 
2854 	struct delayed_work page_cache_work;
2855 	atomic_t backoff_page_cache_fill;
2856 	atomic_t work_in_progress;
2857 	struct hrtimer hrtimer;
2858 
2859 	struct llist_head bkvcache;
2860 	int nr_bkv_objs;
2861 };
2862 
2863 static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc) = {
2864 	.lock = __RAW_SPIN_LOCK_UNLOCKED(krc.lock),
2865 };
2866 
2867 static __always_inline void
2868 debug_rcu_bhead_unqueue(struct kvfree_rcu_bulk_data *bhead)
2869 {
2870 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
2871 	int i;
2872 
2873 	for (i = 0; i < bhead->nr_records; i++)
2874 		debug_rcu_head_unqueue((struct rcu_head *)(bhead->records[i]));
2875 #endif
2876 }
2877 
2878 static inline struct kfree_rcu_cpu *
2879 krc_this_cpu_lock(unsigned long *flags)
2880 {
2881 	struct kfree_rcu_cpu *krcp;
2882 
2883 	local_irq_save(*flags);	// For safely calling this_cpu_ptr().
2884 	krcp = this_cpu_ptr(&krc);
2885 	raw_spin_lock(&krcp->lock);
2886 
2887 	return krcp;
2888 }
2889 
2890 static inline void
2891 krc_this_cpu_unlock(struct kfree_rcu_cpu *krcp, unsigned long flags)
2892 {
2893 	raw_spin_unlock_irqrestore(&krcp->lock, flags);
2894 }
2895 
2896 static inline struct kvfree_rcu_bulk_data *
2897 get_cached_bnode(struct kfree_rcu_cpu *krcp)
2898 {
2899 	if (!krcp->nr_bkv_objs)
2900 		return NULL;
2901 
2902 	WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs - 1);
2903 	return (struct kvfree_rcu_bulk_data *)
2904 		llist_del_first(&krcp->bkvcache);
2905 }
2906 
2907 static inline bool
2908 put_cached_bnode(struct kfree_rcu_cpu *krcp,
2909 	struct kvfree_rcu_bulk_data *bnode)
2910 {
2911 	// Check the limit.
2912 	if (krcp->nr_bkv_objs >= rcu_min_cached_objs)
2913 		return false;
2914 
2915 	llist_add((struct llist_node *) bnode, &krcp->bkvcache);
2916 	WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs + 1);
2917 	return true;
2918 }
2919 
2920 static int
2921 drain_page_cache(struct kfree_rcu_cpu *krcp)
2922 {
2923 	unsigned long flags;
2924 	struct llist_node *page_list, *pos, *n;
2925 	int freed = 0;
2926 
2927 	if (!rcu_min_cached_objs)
2928 		return 0;
2929 
2930 	raw_spin_lock_irqsave(&krcp->lock, flags);
2931 	page_list = llist_del_all(&krcp->bkvcache);
2932 	WRITE_ONCE(krcp->nr_bkv_objs, 0);
2933 	raw_spin_unlock_irqrestore(&krcp->lock, flags);
2934 
2935 	llist_for_each_safe(pos, n, page_list) {
2936 		free_page((unsigned long)pos);
2937 		freed++;
2938 	}
2939 
2940 	return freed;
2941 }
2942 
2943 static void
2944 kvfree_rcu_bulk(struct kfree_rcu_cpu *krcp,
2945 	struct kvfree_rcu_bulk_data *bnode, int idx)
2946 {
2947 	unsigned long flags;
2948 	int i;
2949 
2950 	if (!WARN_ON_ONCE(!poll_state_synchronize_rcu_full(&bnode->gp_snap))) {
2951 		debug_rcu_bhead_unqueue(bnode);
2952 		rcu_lock_acquire(&rcu_callback_map);
2953 		if (idx == 0) { // kmalloc() / kfree().
2954 			trace_rcu_invoke_kfree_bulk_callback(
2955 				rcu_state.name, bnode->nr_records,
2956 				bnode->records);
2957 
2958 			kfree_bulk(bnode->nr_records, bnode->records);
2959 		} else { // vmalloc() / vfree().
2960 			for (i = 0; i < bnode->nr_records; i++) {
2961 				trace_rcu_invoke_kvfree_callback(
2962 					rcu_state.name, bnode->records[i], 0);
2963 
2964 				vfree(bnode->records[i]);
2965 			}
2966 		}
2967 		rcu_lock_release(&rcu_callback_map);
2968 	}
2969 
2970 	raw_spin_lock_irqsave(&krcp->lock, flags);
2971 	if (put_cached_bnode(krcp, bnode))
2972 		bnode = NULL;
2973 	raw_spin_unlock_irqrestore(&krcp->lock, flags);
2974 
2975 	if (bnode)
2976 		free_page((unsigned long) bnode);
2977 
2978 	cond_resched_tasks_rcu_qs();
2979 }
2980 
2981 static void
2982 kvfree_rcu_list(struct rcu_head *head)
2983 {
2984 	struct rcu_head *next;
2985 
2986 	for (; head; head = next) {
2987 		void *ptr = (void *) head->func;
2988 		unsigned long offset = (void *) head - ptr;
2989 
2990 		next = head->next;
2991 		debug_rcu_head_unqueue((struct rcu_head *)ptr);
2992 		rcu_lock_acquire(&rcu_callback_map);
2993 		trace_rcu_invoke_kvfree_callback(rcu_state.name, head, offset);
2994 
2995 		if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset)))
2996 			kvfree(ptr);
2997 
2998 		rcu_lock_release(&rcu_callback_map);
2999 		cond_resched_tasks_rcu_qs();
3000 	}
3001 }
3002 
3003 /*
3004  * This function is invoked in workqueue context after a grace period.
3005  * It frees all the objects queued on ->bulk_head_free or ->head_free.
3006  */
3007 static void kfree_rcu_work(struct work_struct *work)
3008 {
3009 	unsigned long flags;
3010 	struct kvfree_rcu_bulk_data *bnode, *n;
3011 	struct list_head bulk_head[FREE_N_CHANNELS];
3012 	struct rcu_head *head;
3013 	struct kfree_rcu_cpu *krcp;
3014 	struct kfree_rcu_cpu_work *krwp;
3015 	struct rcu_gp_oldstate head_gp_snap;
3016 	int i;
3017 
3018 	krwp = container_of(to_rcu_work(work),
3019 		struct kfree_rcu_cpu_work, rcu_work);
3020 	krcp = krwp->krcp;
3021 
3022 	raw_spin_lock_irqsave(&krcp->lock, flags);
3023 	// Channels 1 and 2.
3024 	for (i = 0; i < FREE_N_CHANNELS; i++)
3025 		list_replace_init(&krwp->bulk_head_free[i], &bulk_head[i]);
3026 
3027 	// Channel 3.
3028 	head = krwp->head_free;
3029 	krwp->head_free = NULL;
3030 	head_gp_snap = krwp->head_free_gp_snap;
3031 	raw_spin_unlock_irqrestore(&krcp->lock, flags);
3032 
3033 	// Handle the first two channels.
3034 	for (i = 0; i < FREE_N_CHANNELS; i++) {
3035 		// Start from the tail page, so a GP is likely passed for it.
3036 		list_for_each_entry_safe(bnode, n, &bulk_head[i], list)
3037 			kvfree_rcu_bulk(krcp, bnode, i);
3038 	}
3039 
3040 	/*
3041 	 * This is used when the "bulk" path can not be used for the
3042 	 * double-argument of kvfree_rcu().  This happens when the
3043 	 * page-cache is empty, which means that objects are instead
3044 	 * queued on a linked list through their rcu_head structures.
3045 	 * This list is named "Channel 3".
3046 	 */
3047 	if (head && !WARN_ON_ONCE(!poll_state_synchronize_rcu_full(&head_gp_snap)))
3048 		kvfree_rcu_list(head);
3049 }
3050 
3051 static bool
3052 need_offload_krc(struct kfree_rcu_cpu *krcp)
3053 {
3054 	int i;
3055 
3056 	for (i = 0; i < FREE_N_CHANNELS; i++)
3057 		if (!list_empty(&krcp->bulk_head[i]))
3058 			return true;
3059 
3060 	return !!READ_ONCE(krcp->head);
3061 }
3062 
3063 static bool
3064 need_wait_for_krwp_work(struct kfree_rcu_cpu_work *krwp)
3065 {
3066 	int i;
3067 
3068 	for (i = 0; i < FREE_N_CHANNELS; i++)
3069 		if (!list_empty(&krwp->bulk_head_free[i]))
3070 			return true;
3071 
3072 	return !!krwp->head_free;
3073 }
3074 
3075 static int krc_count(struct kfree_rcu_cpu *krcp)
3076 {
3077 	int sum = atomic_read(&krcp->head_count);
3078 	int i;
3079 
3080 	for (i = 0; i < FREE_N_CHANNELS; i++)
3081 		sum += atomic_read(&krcp->bulk_count[i]);
3082 
3083 	return sum;
3084 }
3085 
3086 static void
3087 schedule_delayed_monitor_work(struct kfree_rcu_cpu *krcp)
3088 {
3089 	long delay, delay_left;
3090 
3091 	delay = krc_count(krcp) >= KVFREE_BULK_MAX_ENTR ? 1:KFREE_DRAIN_JIFFIES;
3092 	if (delayed_work_pending(&krcp->monitor_work)) {
3093 		delay_left = krcp->monitor_work.timer.expires - jiffies;
3094 		if (delay < delay_left)
3095 			mod_delayed_work(system_wq, &krcp->monitor_work, delay);
3096 		return;
3097 	}
3098 	queue_delayed_work(system_wq, &krcp->monitor_work, delay);
3099 }
3100 
3101 static void
3102 kvfree_rcu_drain_ready(struct kfree_rcu_cpu *krcp)
3103 {
3104 	struct list_head bulk_ready[FREE_N_CHANNELS];
3105 	struct kvfree_rcu_bulk_data *bnode, *n;
3106 	struct rcu_head *head_ready = NULL;
3107 	unsigned long flags;
3108 	int i;
3109 
3110 	raw_spin_lock_irqsave(&krcp->lock, flags);
3111 	for (i = 0; i < FREE_N_CHANNELS; i++) {
3112 		INIT_LIST_HEAD(&bulk_ready[i]);
3113 
3114 		list_for_each_entry_safe_reverse(bnode, n, &krcp->bulk_head[i], list) {
3115 			if (!poll_state_synchronize_rcu_full(&bnode->gp_snap))
3116 				break;
3117 
3118 			atomic_sub(bnode->nr_records, &krcp->bulk_count[i]);
3119 			list_move(&bnode->list, &bulk_ready[i]);
3120 		}
3121 	}
3122 
3123 	if (krcp->head && poll_state_synchronize_rcu(krcp->head_gp_snap)) {
3124 		head_ready = krcp->head;
3125 		atomic_set(&krcp->head_count, 0);
3126 		WRITE_ONCE(krcp->head, NULL);
3127 	}
3128 	raw_spin_unlock_irqrestore(&krcp->lock, flags);
3129 
3130 	for (i = 0; i < FREE_N_CHANNELS; i++) {
3131 		list_for_each_entry_safe(bnode, n, &bulk_ready[i], list)
3132 			kvfree_rcu_bulk(krcp, bnode, i);
3133 	}
3134 
3135 	if (head_ready)
3136 		kvfree_rcu_list(head_ready);
3137 }
3138 
3139 /*
3140  * This function is invoked after the KFREE_DRAIN_JIFFIES timeout.
3141  */
3142 static void kfree_rcu_monitor(struct work_struct *work)
3143 {
3144 	struct kfree_rcu_cpu *krcp = container_of(work,
3145 		struct kfree_rcu_cpu, monitor_work.work);
3146 	unsigned long flags;
3147 	int i, j;
3148 
3149 	// Drain ready for reclaim.
3150 	kvfree_rcu_drain_ready(krcp);
3151 
3152 	raw_spin_lock_irqsave(&krcp->lock, flags);
3153 
3154 	// Attempt to start a new batch.
3155 	for (i = 0; i < KFREE_N_BATCHES; i++) {
3156 		struct kfree_rcu_cpu_work *krwp = &(krcp->krw_arr[i]);
3157 
3158 		// Try to detach bulk_head or head and attach it, only when
3159 		// all channels are free.  Any channel is not free means at krwp
3160 		// there is on-going rcu work to handle krwp's free business.
3161 		if (need_wait_for_krwp_work(krwp))
3162 			continue;
3163 
3164 		// kvfree_rcu_drain_ready() might handle this krcp, if so give up.
3165 		if (need_offload_krc(krcp)) {
3166 			// Channel 1 corresponds to the SLAB-pointer bulk path.
3167 			// Channel 2 corresponds to vmalloc-pointer bulk path.
3168 			for (j = 0; j < FREE_N_CHANNELS; j++) {
3169 				if (list_empty(&krwp->bulk_head_free[j])) {
3170 					atomic_set(&krcp->bulk_count[j], 0);
3171 					list_replace_init(&krcp->bulk_head[j],
3172 						&krwp->bulk_head_free[j]);
3173 				}
3174 			}
3175 
3176 			// Channel 3 corresponds to both SLAB and vmalloc
3177 			// objects queued on the linked list.
3178 			if (!krwp->head_free) {
3179 				krwp->head_free = krcp->head;
3180 				get_state_synchronize_rcu_full(&krwp->head_free_gp_snap);
3181 				atomic_set(&krcp->head_count, 0);
3182 				WRITE_ONCE(krcp->head, NULL);
3183 			}
3184 
3185 			// One work is per one batch, so there are three
3186 			// "free channels", the batch can handle. It can
3187 			// be that the work is in the pending state when
3188 			// channels have been detached following by each
3189 			// other.
3190 			queue_rcu_work(system_wq, &krwp->rcu_work);
3191 		}
3192 	}
3193 
3194 	raw_spin_unlock_irqrestore(&krcp->lock, flags);
3195 
3196 	// If there is nothing to detach, it means that our job is
3197 	// successfully done here. In case of having at least one
3198 	// of the channels that is still busy we should rearm the
3199 	// work to repeat an attempt. Because previous batches are
3200 	// still in progress.
3201 	if (need_offload_krc(krcp))
3202 		schedule_delayed_monitor_work(krcp);
3203 }
3204 
3205 static enum hrtimer_restart
3206 schedule_page_work_fn(struct hrtimer *t)
3207 {
3208 	struct kfree_rcu_cpu *krcp =
3209 		container_of(t, struct kfree_rcu_cpu, hrtimer);
3210 
3211 	queue_delayed_work(system_highpri_wq, &krcp->page_cache_work, 0);
3212 	return HRTIMER_NORESTART;
3213 }
3214 
3215 static void fill_page_cache_func(struct work_struct *work)
3216 {
3217 	struct kvfree_rcu_bulk_data *bnode;
3218 	struct kfree_rcu_cpu *krcp =
3219 		container_of(work, struct kfree_rcu_cpu,
3220 			page_cache_work.work);
3221 	unsigned long flags;
3222 	int nr_pages;
3223 	bool pushed;
3224 	int i;
3225 
3226 	nr_pages = atomic_read(&krcp->backoff_page_cache_fill) ?
3227 		1 : rcu_min_cached_objs;
3228 
3229 	for (i = READ_ONCE(krcp->nr_bkv_objs); i < nr_pages; i++) {
3230 		bnode = (struct kvfree_rcu_bulk_data *)
3231 			__get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3232 
3233 		if (!bnode)
3234 			break;
3235 
3236 		raw_spin_lock_irqsave(&krcp->lock, flags);
3237 		pushed = put_cached_bnode(krcp, bnode);
3238 		raw_spin_unlock_irqrestore(&krcp->lock, flags);
3239 
3240 		if (!pushed) {
3241 			free_page((unsigned long) bnode);
3242 			break;
3243 		}
3244 	}
3245 
3246 	atomic_set(&krcp->work_in_progress, 0);
3247 	atomic_set(&krcp->backoff_page_cache_fill, 0);
3248 }
3249 
3250 static void
3251 run_page_cache_worker(struct kfree_rcu_cpu *krcp)
3252 {
3253 	// If cache disabled, bail out.
3254 	if (!rcu_min_cached_objs)
3255 		return;
3256 
3257 	if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
3258 			!atomic_xchg(&krcp->work_in_progress, 1)) {
3259 		if (atomic_read(&krcp->backoff_page_cache_fill)) {
3260 			queue_delayed_work(system_wq,
3261 				&krcp->page_cache_work,
3262 					msecs_to_jiffies(rcu_delay_page_cache_fill_msec));
3263 		} else {
3264 			hrtimer_init(&krcp->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3265 			krcp->hrtimer.function = schedule_page_work_fn;
3266 			hrtimer_start(&krcp->hrtimer, 0, HRTIMER_MODE_REL);
3267 		}
3268 	}
3269 }
3270 
3271 // Record ptr in a page managed by krcp, with the pre-krc_this_cpu_lock()
3272 // state specified by flags.  If can_alloc is true, the caller must
3273 // be schedulable and not be holding any locks or mutexes that might be
3274 // acquired by the memory allocator or anything that it might invoke.
3275 // Returns true if ptr was successfully recorded, else the caller must
3276 // use a fallback.
3277 static inline bool
3278 add_ptr_to_bulk_krc_lock(struct kfree_rcu_cpu **krcp,
3279 	unsigned long *flags, void *ptr, bool can_alloc)
3280 {
3281 	struct kvfree_rcu_bulk_data *bnode;
3282 	int idx;
3283 
3284 	*krcp = krc_this_cpu_lock(flags);
3285 	if (unlikely(!(*krcp)->initialized))
3286 		return false;
3287 
3288 	idx = !!is_vmalloc_addr(ptr);
3289 	bnode = list_first_entry_or_null(&(*krcp)->bulk_head[idx],
3290 		struct kvfree_rcu_bulk_data, list);
3291 
3292 	/* Check if a new block is required. */
3293 	if (!bnode || bnode->nr_records == KVFREE_BULK_MAX_ENTR) {
3294 		bnode = get_cached_bnode(*krcp);
3295 		if (!bnode && can_alloc) {
3296 			krc_this_cpu_unlock(*krcp, *flags);
3297 
3298 			// __GFP_NORETRY - allows a light-weight direct reclaim
3299 			// what is OK from minimizing of fallback hitting point of
3300 			// view. Apart of that it forbids any OOM invoking what is
3301 			// also beneficial since we are about to release memory soon.
3302 			//
3303 			// __GFP_NOMEMALLOC - prevents from consuming of all the
3304 			// memory reserves. Please note we have a fallback path.
3305 			//
3306 			// __GFP_NOWARN - it is supposed that an allocation can
3307 			// be failed under low memory or high memory pressure
3308 			// scenarios.
3309 			bnode = (struct kvfree_rcu_bulk_data *)
3310 				__get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3311 			raw_spin_lock_irqsave(&(*krcp)->lock, *flags);
3312 		}
3313 
3314 		if (!bnode)
3315 			return false;
3316 
3317 		// Initialize the new block and attach it.
3318 		bnode->nr_records = 0;
3319 		list_add(&bnode->list, &(*krcp)->bulk_head[idx]);
3320 	}
3321 
3322 	// Finally insert and update the GP for this page.
3323 	bnode->records[bnode->nr_records++] = ptr;
3324 	get_state_synchronize_rcu_full(&bnode->gp_snap);
3325 	atomic_inc(&(*krcp)->bulk_count[idx]);
3326 
3327 	return true;
3328 }
3329 
3330 /*
3331  * Queue a request for lazy invocation of the appropriate free routine
3332  * after a grace period.  Please note that three paths are maintained,
3333  * two for the common case using arrays of pointers and a third one that
3334  * is used only when the main paths cannot be used, for example, due to
3335  * memory pressure.
3336  *
3337  * Each kvfree_call_rcu() request is added to a batch. The batch will be drained
3338  * every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch will
3339  * be free'd in workqueue context. This allows us to: batch requests together to
3340  * reduce the number of grace periods during heavy kfree_rcu()/kvfree_rcu() load.
3341  */
3342 void kvfree_call_rcu(struct rcu_head *head, void *ptr)
3343 {
3344 	unsigned long flags;
3345 	struct kfree_rcu_cpu *krcp;
3346 	bool success;
3347 
3348 	/*
3349 	 * Please note there is a limitation for the head-less
3350 	 * variant, that is why there is a clear rule for such
3351 	 * objects: it can be used from might_sleep() context
3352 	 * only. For other places please embed an rcu_head to
3353 	 * your data.
3354 	 */
3355 	if (!head)
3356 		might_sleep();
3357 
3358 	// Queue the object but don't yet schedule the batch.
3359 	if (debug_rcu_head_queue(ptr)) {
3360 		// Probable double kfree_rcu(), just leak.
3361 		WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n",
3362 			  __func__, head);
3363 
3364 		// Mark as success and leave.
3365 		return;
3366 	}
3367 
3368 	kasan_record_aux_stack_noalloc(ptr);
3369 	success = add_ptr_to_bulk_krc_lock(&krcp, &flags, ptr, !head);
3370 	if (!success) {
3371 		run_page_cache_worker(krcp);
3372 
3373 		if (head == NULL)
3374 			// Inline if kvfree_rcu(one_arg) call.
3375 			goto unlock_return;
3376 
3377 		head->func = ptr;
3378 		head->next = krcp->head;
3379 		WRITE_ONCE(krcp->head, head);
3380 		atomic_inc(&krcp->head_count);
3381 
3382 		// Take a snapshot for this krcp.
3383 		krcp->head_gp_snap = get_state_synchronize_rcu();
3384 		success = true;
3385 	}
3386 
3387 	// Set timer to drain after KFREE_DRAIN_JIFFIES.
3388 	if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING)
3389 		schedule_delayed_monitor_work(krcp);
3390 
3391 unlock_return:
3392 	krc_this_cpu_unlock(krcp, flags);
3393 
3394 	/*
3395 	 * Inline kvfree() after synchronize_rcu(). We can do
3396 	 * it from might_sleep() context only, so the current
3397 	 * CPU can pass the QS state.
3398 	 */
3399 	if (!success) {
3400 		debug_rcu_head_unqueue((struct rcu_head *) ptr);
3401 		synchronize_rcu();
3402 		kvfree(ptr);
3403 	}
3404 }
3405 EXPORT_SYMBOL_GPL(kvfree_call_rcu);
3406 
3407 static unsigned long
3408 kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
3409 {
3410 	int cpu;
3411 	unsigned long count = 0;
3412 
3413 	/* Snapshot count of all CPUs */
3414 	for_each_possible_cpu(cpu) {
3415 		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3416 
3417 		count += krc_count(krcp);
3418 		count += READ_ONCE(krcp->nr_bkv_objs);
3419 		atomic_set(&krcp->backoff_page_cache_fill, 1);
3420 	}
3421 
3422 	return count == 0 ? SHRINK_EMPTY : count;
3423 }
3424 
3425 static unsigned long
3426 kfree_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
3427 {
3428 	int cpu, freed = 0;
3429 
3430 	for_each_possible_cpu(cpu) {
3431 		int count;
3432 		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3433 
3434 		count = krc_count(krcp);
3435 		count += drain_page_cache(krcp);
3436 		kfree_rcu_monitor(&krcp->monitor_work.work);
3437 
3438 		sc->nr_to_scan -= count;
3439 		freed += count;
3440 
3441 		if (sc->nr_to_scan <= 0)
3442 			break;
3443 	}
3444 
3445 	return freed == 0 ? SHRINK_STOP : freed;
3446 }
3447 
3448 static struct shrinker kfree_rcu_shrinker = {
3449 	.count_objects = kfree_rcu_shrink_count,
3450 	.scan_objects = kfree_rcu_shrink_scan,
3451 	.batch = 0,
3452 	.seeks = DEFAULT_SEEKS,
3453 };
3454 
3455 void __init kfree_rcu_scheduler_running(void)
3456 {
3457 	int cpu;
3458 
3459 	for_each_possible_cpu(cpu) {
3460 		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3461 
3462 		if (need_offload_krc(krcp))
3463 			schedule_delayed_monitor_work(krcp);
3464 	}
3465 }
3466 
3467 /*
3468  * During early boot, any blocking grace-period wait automatically
3469  * implies a grace period.
3470  *
3471  * Later on, this could in theory be the case for kernels built with
3472  * CONFIG_SMP=y && CONFIG_PREEMPTION=y running on a single CPU, but this
3473  * is not a common case.  Furthermore, this optimization would cause
3474  * the rcu_gp_oldstate structure to expand by 50%, so this potential
3475  * grace-period optimization is ignored once the scheduler is running.
3476  */
3477 static int rcu_blocking_is_gp(void)
3478 {
3479 	if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE) {
3480 		might_sleep();
3481 		return false;
3482 	}
3483 	return true;
3484 }
3485 
3486 /**
3487  * synchronize_rcu - wait until a grace period has elapsed.
3488  *
3489  * Control will return to the caller some time after a full grace
3490  * period has elapsed, in other words after all currently executing RCU
3491  * read-side critical sections have completed.  Note, however, that
3492  * upon return from synchronize_rcu(), the caller might well be executing
3493  * concurrently with new RCU read-side critical sections that began while
3494  * synchronize_rcu() was waiting.
3495  *
3496  * RCU read-side critical sections are delimited by rcu_read_lock()
3497  * and rcu_read_unlock(), and may be nested.  In addition, but only in
3498  * v5.0 and later, regions of code across which interrupts, preemption,
3499  * or softirqs have been disabled also serve as RCU read-side critical
3500  * sections.  This includes hardware interrupt handlers, softirq handlers,
3501  * and NMI handlers.
3502  *
3503  * Note that this guarantee implies further memory-ordering guarantees.
3504  * On systems with more than one CPU, when synchronize_rcu() returns,
3505  * each CPU is guaranteed to have executed a full memory barrier since
3506  * the end of its last RCU read-side critical section whose beginning
3507  * preceded the call to synchronize_rcu().  In addition, each CPU having
3508  * an RCU read-side critical section that extends beyond the return from
3509  * synchronize_rcu() is guaranteed to have executed a full memory barrier
3510  * after the beginning of synchronize_rcu() and before the beginning of
3511  * that RCU read-side critical section.  Note that these guarantees include
3512  * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3513  * that are executing in the kernel.
3514  *
3515  * Furthermore, if CPU A invoked synchronize_rcu(), which returned
3516  * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3517  * to have executed a full memory barrier during the execution of
3518  * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
3519  * again only if the system has more than one CPU).
3520  *
3521  * Implementation of these memory-ordering guarantees is described here:
3522  * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
3523  */
3524 void synchronize_rcu(void)
3525 {
3526 	unsigned long flags;
3527 	struct rcu_node *rnp;
3528 
3529 	RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3530 			 lock_is_held(&rcu_lock_map) ||
3531 			 lock_is_held(&rcu_sched_lock_map),
3532 			 "Illegal synchronize_rcu() in RCU read-side critical section");
3533 	if (!rcu_blocking_is_gp()) {
3534 		if (rcu_gp_is_expedited())
3535 			synchronize_rcu_expedited();
3536 		else
3537 			wait_rcu_gp(call_rcu_hurry);
3538 		return;
3539 	}
3540 
3541 	// Context allows vacuous grace periods.
3542 	// Note well that this code runs with !PREEMPT && !SMP.
3543 	// In addition, all code that advances grace periods runs at
3544 	// process level.  Therefore, this normal GP overlaps with other
3545 	// normal GPs only by being fully nested within them, which allows
3546 	// reuse of ->gp_seq_polled_snap.
3547 	rcu_poll_gp_seq_start_unlocked(&rcu_state.gp_seq_polled_snap);
3548 	rcu_poll_gp_seq_end_unlocked(&rcu_state.gp_seq_polled_snap);
3549 
3550 	// Update the normal grace-period counters to record
3551 	// this grace period, but only those used by the boot CPU.
3552 	// The rcu_scheduler_starting() will take care of the rest of
3553 	// these counters.
3554 	local_irq_save(flags);
3555 	WARN_ON_ONCE(num_online_cpus() > 1);
3556 	rcu_state.gp_seq += (1 << RCU_SEQ_CTR_SHIFT);
3557 	for (rnp = this_cpu_ptr(&rcu_data)->mynode; rnp; rnp = rnp->parent)
3558 		rnp->gp_seq_needed = rnp->gp_seq = rcu_state.gp_seq;
3559 	local_irq_restore(flags);
3560 }
3561 EXPORT_SYMBOL_GPL(synchronize_rcu);
3562 
3563 /**
3564  * get_completed_synchronize_rcu_full - Return a full pre-completed polled state cookie
3565  * @rgosp: Place to put state cookie
3566  *
3567  * Stores into @rgosp a value that will always be treated by functions
3568  * like poll_state_synchronize_rcu_full() as a cookie whose grace period
3569  * has already completed.
3570  */
3571 void get_completed_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3572 {
3573 	rgosp->rgos_norm = RCU_GET_STATE_COMPLETED;
3574 	rgosp->rgos_exp = RCU_GET_STATE_COMPLETED;
3575 }
3576 EXPORT_SYMBOL_GPL(get_completed_synchronize_rcu_full);
3577 
3578 /**
3579  * get_state_synchronize_rcu - Snapshot current RCU state
3580  *
3581  * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3582  * or poll_state_synchronize_rcu() to determine whether or not a full
3583  * grace period has elapsed in the meantime.
3584  */
3585 unsigned long get_state_synchronize_rcu(void)
3586 {
3587 	/*
3588 	 * Any prior manipulation of RCU-protected data must happen
3589 	 * before the load from ->gp_seq.
3590 	 */
3591 	smp_mb();  /* ^^^ */
3592 	return rcu_seq_snap(&rcu_state.gp_seq_polled);
3593 }
3594 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3595 
3596 /**
3597  * get_state_synchronize_rcu_full - Snapshot RCU state, both normal and expedited
3598  * @rgosp: location to place combined normal/expedited grace-period state
3599  *
3600  * Places the normal and expedited grace-period states in @rgosp.  This
3601  * state value can be passed to a later call to cond_synchronize_rcu_full()
3602  * or poll_state_synchronize_rcu_full() to determine whether or not a
3603  * grace period (whether normal or expedited) has elapsed in the meantime.
3604  * The rcu_gp_oldstate structure takes up twice the memory of an unsigned
3605  * long, but is guaranteed to see all grace periods.  In contrast, the
3606  * combined state occupies less memory, but can sometimes fail to take
3607  * grace periods into account.
3608  *
3609  * This does not guarantee that the needed grace period will actually
3610  * start.
3611  */
3612 void get_state_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3613 {
3614 	struct rcu_node *rnp = rcu_get_root();
3615 
3616 	/*
3617 	 * Any prior manipulation of RCU-protected data must happen
3618 	 * before the loads from ->gp_seq and ->expedited_sequence.
3619 	 */
3620 	smp_mb();  /* ^^^ */
3621 	rgosp->rgos_norm = rcu_seq_snap(&rnp->gp_seq);
3622 	rgosp->rgos_exp = rcu_seq_snap(&rcu_state.expedited_sequence);
3623 }
3624 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu_full);
3625 
3626 /*
3627  * Helper function for start_poll_synchronize_rcu() and
3628  * start_poll_synchronize_rcu_full().
3629  */
3630 static void start_poll_synchronize_rcu_common(void)
3631 {
3632 	unsigned long flags;
3633 	bool needwake;
3634 	struct rcu_data *rdp;
3635 	struct rcu_node *rnp;
3636 
3637 	lockdep_assert_irqs_enabled();
3638 	local_irq_save(flags);
3639 	rdp = this_cpu_ptr(&rcu_data);
3640 	rnp = rdp->mynode;
3641 	raw_spin_lock_rcu_node(rnp); // irqs already disabled.
3642 	// Note it is possible for a grace period to have elapsed between
3643 	// the above call to get_state_synchronize_rcu() and the below call
3644 	// to rcu_seq_snap.  This is OK, the worst that happens is that we
3645 	// get a grace period that no one needed.  These accesses are ordered
3646 	// by smp_mb(), and we are accessing them in the opposite order
3647 	// from which they are updated at grace-period start, as required.
3648 	needwake = rcu_start_this_gp(rnp, rdp, rcu_seq_snap(&rcu_state.gp_seq));
3649 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3650 	if (needwake)
3651 		rcu_gp_kthread_wake();
3652 }
3653 
3654 /**
3655  * start_poll_synchronize_rcu - Snapshot and start RCU grace period
3656  *
3657  * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3658  * or poll_state_synchronize_rcu() to determine whether or not a full
3659  * grace period has elapsed in the meantime.  If the needed grace period
3660  * is not already slated to start, notifies RCU core of the need for that
3661  * grace period.
3662  *
3663  * Interrupts must be enabled for the case where it is necessary to awaken
3664  * the grace-period kthread.
3665  */
3666 unsigned long start_poll_synchronize_rcu(void)
3667 {
3668 	unsigned long gp_seq = get_state_synchronize_rcu();
3669 
3670 	start_poll_synchronize_rcu_common();
3671 	return gp_seq;
3672 }
3673 EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu);
3674 
3675 /**
3676  * start_poll_synchronize_rcu_full - Take a full snapshot and start RCU grace period
3677  * @rgosp: value from get_state_synchronize_rcu_full() or start_poll_synchronize_rcu_full()
3678  *
3679  * Places the normal and expedited grace-period states in *@rgos.  This
3680  * state value can be passed to a later call to cond_synchronize_rcu_full()
3681  * or poll_state_synchronize_rcu_full() to determine whether or not a
3682  * grace period (whether normal or expedited) has elapsed in the meantime.
3683  * If the needed grace period is not already slated to start, notifies
3684  * RCU core of the need for that grace period.
3685  *
3686  * Interrupts must be enabled for the case where it is necessary to awaken
3687  * the grace-period kthread.
3688  */
3689 void start_poll_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3690 {
3691 	get_state_synchronize_rcu_full(rgosp);
3692 
3693 	start_poll_synchronize_rcu_common();
3694 }
3695 EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_full);
3696 
3697 /**
3698  * poll_state_synchronize_rcu - Has the specified RCU grace period completed?
3699  * @oldstate: value from get_state_synchronize_rcu() or start_poll_synchronize_rcu()
3700  *
3701  * If a full RCU grace period has elapsed since the earlier call from
3702  * which @oldstate was obtained, return @true, otherwise return @false.
3703  * If @false is returned, it is the caller's responsibility to invoke this
3704  * function later on until it does return @true.  Alternatively, the caller
3705  * can explicitly wait for a grace period, for example, by passing @oldstate
3706  * to either cond_synchronize_rcu() or cond_synchronize_rcu_expedited()
3707  * on the one hand or by directly invoking either synchronize_rcu() or
3708  * synchronize_rcu_expedited() on the other.
3709  *
3710  * Yes, this function does not take counter wrap into account.
3711  * But counter wrap is harmless.  If the counter wraps, we have waited for
3712  * more than a billion grace periods (and way more on a 64-bit system!).
3713  * Those needing to keep old state values for very long time periods
3714  * (many hours even on 32-bit systems) should check them occasionally and
3715  * either refresh them or set a flag indicating that the grace period has
3716  * completed.  Alternatively, they can use get_completed_synchronize_rcu()
3717  * to get a guaranteed-completed grace-period state.
3718  *
3719  * In addition, because oldstate compresses the grace-period state for
3720  * both normal and expedited grace periods into a single unsigned long,
3721  * it can miss a grace period when synchronize_rcu() runs concurrently
3722  * with synchronize_rcu_expedited().  If this is unacceptable, please
3723  * instead use the _full() variant of these polling APIs.
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  */
3730 bool 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 }
3739 EXPORT_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  */
3774 bool 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 }
3788 EXPORT_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  */
3808 void cond_synchronize_rcu(unsigned long oldstate)
3809 {
3810 	if (!poll_state_synchronize_rcu(oldstate))
3811 		synchronize_rcu();
3812 }
3813 EXPORT_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  */
3835 void cond_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3836 {
3837 	if (!poll_state_synchronize_rcu_full(rgosp))
3838 		synchronize_rcu();
3839 }
3840 EXPORT_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  */
3849 static 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  */
3897 static 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  */
3913 static 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  */
3928 static 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  */
3965 static 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  */
3986 void 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);
4031 retry:
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 }
4080 EXPORT_SYMBOL_GPL(rcu_barrier);
4081 
4082 /*
4083  * Compute the mask of online CPUs for the specified rcu_node structure.
4084  * This will not be stable unless the rcu_node structure's ->lock is
4085  * held, but the bit corresponding to the current CPU will be stable
4086  * in most contexts.
4087  */
4088 static unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
4089 {
4090 	return READ_ONCE(rnp->qsmaskinitnext);
4091 }
4092 
4093 /*
4094  * Is the CPU corresponding to the specified rcu_data structure online
4095  * from RCU's perspective?  This perspective is given by that structure's
4096  * ->qsmaskinitnext field rather than by the global cpu_online_mask.
4097  */
4098 static bool rcu_rdp_cpu_online(struct rcu_data *rdp)
4099 {
4100 	return !!(rdp->grpmask & rcu_rnp_online_cpus(rdp->mynode));
4101 }
4102 
4103 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
4104 
4105 /*
4106  * Is the current CPU online as far as RCU is concerned?
4107  *
4108  * Disable preemption to avoid false positives that could otherwise
4109  * happen due to the current CPU number being sampled, this task being
4110  * preempted, its old CPU being taken offline, resuming on some other CPU,
4111  * then determining that its old CPU is now offline.
4112  *
4113  * Disable checking if in an NMI handler because we cannot safely
4114  * report errors from NMI handlers anyway.  In addition, it is OK to use
4115  * RCU on an offline processor during initial boot, hence the check for
4116  * rcu_scheduler_fully_active.
4117  */
4118 bool rcu_lockdep_current_cpu_online(void)
4119 {
4120 	struct rcu_data *rdp;
4121 	bool ret = false;
4122 
4123 	if (in_nmi() || !rcu_scheduler_fully_active)
4124 		return true;
4125 	preempt_disable_notrace();
4126 	rdp = this_cpu_ptr(&rcu_data);
4127 	/*
4128 	 * Strictly, we care here about the case where the current CPU is
4129 	 * in rcu_cpu_starting() and thus has an excuse for rdp->grpmask
4130 	 * not being up to date. So arch_spin_is_locked() might have a
4131 	 * false positive if it's held by some *other* CPU, but that's
4132 	 * OK because that just means a false *negative* on the warning.
4133 	 */
4134 	if (rcu_rdp_cpu_online(rdp) || arch_spin_is_locked(&rcu_state.ofl_lock))
4135 		ret = true;
4136 	preempt_enable_notrace();
4137 	return ret;
4138 }
4139 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
4140 
4141 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
4142 
4143 // Has rcu_init() been invoked?  This is used (for example) to determine
4144 // whether spinlocks may be acquired safely.
4145 static bool rcu_init_invoked(void)
4146 {
4147 	return !!rcu_state.n_online_cpus;
4148 }
4149 
4150 /*
4151  * Near the end of the offline process.  Trace the fact that this CPU
4152  * is going offline.
4153  */
4154 int rcutree_dying_cpu(unsigned int cpu)
4155 {
4156 	bool blkd;
4157 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4158 	struct rcu_node *rnp = rdp->mynode;
4159 
4160 	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
4161 		return 0;
4162 
4163 	blkd = !!(READ_ONCE(rnp->qsmask) & rdp->grpmask);
4164 	trace_rcu_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
4165 			       blkd ? TPS("cpuofl-bgp") : TPS("cpuofl"));
4166 	return 0;
4167 }
4168 
4169 /*
4170  * All CPUs for the specified rcu_node structure have gone offline,
4171  * and all tasks that were preempted within an RCU read-side critical
4172  * section while running on one of those CPUs have since exited their RCU
4173  * read-side critical section.  Some other CPU is reporting this fact with
4174  * the specified rcu_node structure's ->lock held and interrupts disabled.
4175  * This function therefore goes up the tree of rcu_node structures,
4176  * clearing the corresponding bits in the ->qsmaskinit fields.  Note that
4177  * the leaf rcu_node structure's ->qsmaskinit field has already been
4178  * updated.
4179  *
4180  * This function does check that the specified rcu_node structure has
4181  * all CPUs offline and no blocked tasks, so it is OK to invoke it
4182  * prematurely.  That said, invoking it after the fact will cost you
4183  * a needless lock acquisition.  So once it has done its work, don't
4184  * invoke it again.
4185  */
4186 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
4187 {
4188 	long mask;
4189 	struct rcu_node *rnp = rnp_leaf;
4190 
4191 	raw_lockdep_assert_held_rcu_node(rnp_leaf);
4192 	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
4193 	    WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
4194 	    WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
4195 		return;
4196 	for (;;) {
4197 		mask = rnp->grpmask;
4198 		rnp = rnp->parent;
4199 		if (!rnp)
4200 			break;
4201 		raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
4202 		rnp->qsmaskinit &= ~mask;
4203 		/* Between grace periods, so better already be zero! */
4204 		WARN_ON_ONCE(rnp->qsmask);
4205 		if (rnp->qsmaskinit) {
4206 			raw_spin_unlock_rcu_node(rnp);
4207 			/* irqs remain disabled. */
4208 			return;
4209 		}
4210 		raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
4211 	}
4212 }
4213 
4214 /*
4215  * The CPU has been completely removed, and some other CPU is reporting
4216  * this fact from process context.  Do the remainder of the cleanup.
4217  * There can only be one CPU hotplug operation at a time, so no need for
4218  * explicit locking.
4219  */
4220 int rcutree_dead_cpu(unsigned int cpu)
4221 {
4222 	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
4223 		return 0;
4224 
4225 	WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus - 1);
4226 	// Stop-machine done, so allow nohz_full to disable tick.
4227 	tick_dep_clear(TICK_DEP_BIT_RCU);
4228 	return 0;
4229 }
4230 
4231 /*
4232  * Propagate ->qsinitmask bits up the rcu_node tree to account for the
4233  * first CPU in a given leaf rcu_node structure coming online.  The caller
4234  * must hold the corresponding leaf rcu_node ->lock with interrupts
4235  * disabled.
4236  */
4237 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
4238 {
4239 	long mask;
4240 	long oldmask;
4241 	struct rcu_node *rnp = rnp_leaf;
4242 
4243 	raw_lockdep_assert_held_rcu_node(rnp_leaf);
4244 	WARN_ON_ONCE(rnp->wait_blkd_tasks);
4245 	for (;;) {
4246 		mask = rnp->grpmask;
4247 		rnp = rnp->parent;
4248 		if (rnp == NULL)
4249 			return;
4250 		raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
4251 		oldmask = rnp->qsmaskinit;
4252 		rnp->qsmaskinit |= mask;
4253 		raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
4254 		if (oldmask)
4255 			return;
4256 	}
4257 }
4258 
4259 /*
4260  * Do boot-time initialization of a CPU's per-CPU RCU data.
4261  */
4262 static void __init
4263 rcu_boot_init_percpu_data(int cpu)
4264 {
4265 	struct context_tracking *ct = this_cpu_ptr(&context_tracking);
4266 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4267 
4268 	/* Set up local state, ensuring consistent view of global state. */
4269 	rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
4270 	INIT_WORK(&rdp->strict_work, strict_work_handler);
4271 	WARN_ON_ONCE(ct->dynticks_nesting != 1);
4272 	WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(cpu)));
4273 	rdp->barrier_seq_snap = rcu_state.barrier_sequence;
4274 	rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
4275 	rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
4276 	rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
4277 	rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
4278 	rdp->last_sched_clock = jiffies;
4279 	rdp->cpu = cpu;
4280 	rcu_boot_init_nocb_percpu_data(rdp);
4281 }
4282 
4283 /*
4284  * Invoked early in the CPU-online process, when pretty much all services
4285  * are available.  The incoming CPU is not present.
4286  *
4287  * Initializes a CPU's per-CPU RCU data.  Note that only one online or
4288  * offline event can be happening at a given time.  Note also that we can
4289  * accept some slop in the rsp->gp_seq access due to the fact that this
4290  * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
4291  * And any offloaded callbacks are being numbered elsewhere.
4292  */
4293 int rcutree_prepare_cpu(unsigned int cpu)
4294 {
4295 	unsigned long flags;
4296 	struct context_tracking *ct = per_cpu_ptr(&context_tracking, cpu);
4297 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4298 	struct rcu_node *rnp = rcu_get_root();
4299 
4300 	/* Set up local state, ensuring consistent view of global state. */
4301 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
4302 	rdp->qlen_last_fqs_check = 0;
4303 	rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
4304 	rdp->blimit = blimit;
4305 	ct->dynticks_nesting = 1;	/* CPU not up, no tearing. */
4306 	raw_spin_unlock_rcu_node(rnp);		/* irqs remain disabled. */
4307 
4308 	/*
4309 	 * Only non-NOCB CPUs that didn't have early-boot callbacks need to be
4310 	 * (re-)initialized.
4311 	 */
4312 	if (!rcu_segcblist_is_enabled(&rdp->cblist))
4313 		rcu_segcblist_init(&rdp->cblist);  /* Re-enable callbacks. */
4314 
4315 	/*
4316 	 * Add CPU to leaf rcu_node pending-online bitmask.  Any needed
4317 	 * propagation up the rcu_node tree will happen at the beginning
4318 	 * of the next grace period.
4319 	 */
4320 	rnp = rdp->mynode;
4321 	raw_spin_lock_rcu_node(rnp);		/* irqs already disabled. */
4322 	rdp->gp_seq = READ_ONCE(rnp->gp_seq);
4323 	rdp->gp_seq_needed = rdp->gp_seq;
4324 	rdp->cpu_no_qs.b.norm = true;
4325 	rdp->core_needs_qs = false;
4326 	rdp->rcu_iw_pending = false;
4327 	rdp->rcu_iw = IRQ_WORK_INIT_HARD(rcu_iw_handler);
4328 	rdp->rcu_iw_gp_seq = rdp->gp_seq - 1;
4329 	trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
4330 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4331 	rcu_spawn_one_boost_kthread(rnp);
4332 	rcu_spawn_cpu_nocb_kthread(cpu);
4333 	WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus + 1);
4334 
4335 	return 0;
4336 }
4337 
4338 /*
4339  * Update RCU priority boot kthread affinity for CPU-hotplug changes.
4340  */
4341 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
4342 {
4343 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4344 
4345 	rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
4346 }
4347 
4348 /*
4349  * Has the specified (known valid) CPU ever been fully online?
4350  */
4351 bool rcu_cpu_beenfullyonline(int cpu)
4352 {
4353 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4354 
4355 	return smp_load_acquire(&rdp->beenonline);
4356 }
4357 
4358 /*
4359  * Near the end of the CPU-online process.  Pretty much all services
4360  * enabled, and the CPU is now very much alive.
4361  */
4362 int rcutree_online_cpu(unsigned int cpu)
4363 {
4364 	unsigned long flags;
4365 	struct rcu_data *rdp;
4366 	struct rcu_node *rnp;
4367 
4368 	rdp = per_cpu_ptr(&rcu_data, cpu);
4369 	rnp = rdp->mynode;
4370 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
4371 	rnp->ffmask |= rdp->grpmask;
4372 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4373 	if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
4374 		return 0; /* Too early in boot for scheduler work. */
4375 	sync_sched_exp_online_cleanup(cpu);
4376 	rcutree_affinity_setting(cpu, -1);
4377 
4378 	// Stop-machine done, so allow nohz_full to disable tick.
4379 	tick_dep_clear(TICK_DEP_BIT_RCU);
4380 	return 0;
4381 }
4382 
4383 /*
4384  * Near the beginning of the process.  The CPU is still very much alive
4385  * with pretty much all services enabled.
4386  */
4387 int rcutree_offline_cpu(unsigned int cpu)
4388 {
4389 	unsigned long flags;
4390 	struct rcu_data *rdp;
4391 	struct rcu_node *rnp;
4392 
4393 	rdp = per_cpu_ptr(&rcu_data, cpu);
4394 	rnp = rdp->mynode;
4395 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
4396 	rnp->ffmask &= ~rdp->grpmask;
4397 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4398 
4399 	rcutree_affinity_setting(cpu, cpu);
4400 
4401 	// nohz_full CPUs need the tick for stop-machine to work quickly
4402 	tick_dep_set(TICK_DEP_BIT_RCU);
4403 	return 0;
4404 }
4405 
4406 /*
4407  * Mark the specified CPU as being online so that subsequent grace periods
4408  * (both expedited and normal) will wait on it.  Note that this means that
4409  * incoming CPUs are not allowed to use RCU read-side critical sections
4410  * until this function is called.  Failing to observe this restriction
4411  * will result in lockdep splats.
4412  *
4413  * Note that this function is special in that it is invoked directly
4414  * from the incoming CPU rather than from the cpuhp_step mechanism.
4415  * This is because this function must be invoked at a precise location.
4416  * This incoming CPU must not have enabled interrupts yet.
4417  */
4418 void rcu_cpu_starting(unsigned int cpu)
4419 {
4420 	unsigned long mask;
4421 	struct rcu_data *rdp;
4422 	struct rcu_node *rnp;
4423 	bool newcpu;
4424 
4425 	lockdep_assert_irqs_disabled();
4426 	rdp = per_cpu_ptr(&rcu_data, cpu);
4427 	if (rdp->cpu_started)
4428 		return;
4429 	rdp->cpu_started = true;
4430 
4431 	rnp = rdp->mynode;
4432 	mask = rdp->grpmask;
4433 	arch_spin_lock(&rcu_state.ofl_lock);
4434 	rcu_dynticks_eqs_online();
4435 	raw_spin_lock(&rcu_state.barrier_lock);
4436 	raw_spin_lock_rcu_node(rnp);
4437 	WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask);
4438 	raw_spin_unlock(&rcu_state.barrier_lock);
4439 	newcpu = !(rnp->expmaskinitnext & mask);
4440 	rnp->expmaskinitnext |= mask;
4441 	/* Allow lockless access for expedited grace periods. */
4442 	smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + newcpu); /* ^^^ */
4443 	ASSERT_EXCLUSIVE_WRITER(rcu_state.ncpus);
4444 	rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
4445 	rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4446 	rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4447 
4448 	/* An incoming CPU should never be blocking a grace period. */
4449 	if (WARN_ON_ONCE(rnp->qsmask & mask)) { /* RCU waiting on incoming CPU? */
4450 		/* rcu_report_qs_rnp() *really* wants some flags to restore */
4451 		unsigned long flags;
4452 
4453 		local_irq_save(flags);
4454 		rcu_disable_urgency_upon_qs(rdp);
4455 		/* Report QS -after- changing ->qsmaskinitnext! */
4456 		rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4457 	} else {
4458 		raw_spin_unlock_rcu_node(rnp);
4459 	}
4460 	arch_spin_unlock(&rcu_state.ofl_lock);
4461 	smp_store_release(&rdp->beenonline, true);
4462 	smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
4463 }
4464 
4465 /*
4466  * The outgoing function has no further need of RCU, so remove it from
4467  * the rcu_node tree's ->qsmaskinitnext bit masks.
4468  *
4469  * Note that this function is special in that it is invoked directly
4470  * from the outgoing CPU rather than from the cpuhp_step mechanism.
4471  * This is because this function must be invoked at a precise location.
4472  */
4473 void rcu_report_dead(unsigned int cpu)
4474 {
4475 	unsigned long flags, seq_flags;
4476 	unsigned long mask;
4477 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4478 	struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
4479 
4480 	// Do any dangling deferred wakeups.
4481 	do_nocb_deferred_wakeup(rdp);
4482 
4483 	rcu_preempt_deferred_qs(current);
4484 
4485 	/* Remove outgoing CPU from mask in the leaf rcu_node structure. */
4486 	mask = rdp->grpmask;
4487 	local_irq_save(seq_flags);
4488 	arch_spin_lock(&rcu_state.ofl_lock);
4489 	raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
4490 	rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4491 	rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4492 	if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
4493 		/* Report quiescent state -before- changing ->qsmaskinitnext! */
4494 		rcu_disable_urgency_upon_qs(rdp);
4495 		rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4496 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
4497 	}
4498 	WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask);
4499 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4500 	arch_spin_unlock(&rcu_state.ofl_lock);
4501 	local_irq_restore(seq_flags);
4502 
4503 	rdp->cpu_started = false;
4504 }
4505 
4506 #ifdef CONFIG_HOTPLUG_CPU
4507 /*
4508  * The outgoing CPU has just passed through the dying-idle state, and we
4509  * are being invoked from the CPU that was IPIed to continue the offline
4510  * operation.  Migrate the outgoing CPU's callbacks to the current CPU.
4511  */
4512 void rcutree_migrate_callbacks(int cpu)
4513 {
4514 	unsigned long flags;
4515 	struct rcu_data *my_rdp;
4516 	struct rcu_node *my_rnp;
4517 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4518 	bool needwake;
4519 
4520 	if (rcu_rdp_is_offloaded(rdp) ||
4521 	    rcu_segcblist_empty(&rdp->cblist))
4522 		return;  /* No callbacks to migrate. */
4523 
4524 	raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
4525 	WARN_ON_ONCE(rcu_rdp_cpu_online(rdp));
4526 	rcu_barrier_entrain(rdp);
4527 	my_rdp = this_cpu_ptr(&rcu_data);
4528 	my_rnp = my_rdp->mynode;
4529 	rcu_nocb_lock(my_rdp); /* irqs already disabled. */
4530 	WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies, false));
4531 	raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
4532 	/* Leverage recent GPs and set GP for new callbacks. */
4533 	needwake = rcu_advance_cbs(my_rnp, rdp) ||
4534 		   rcu_advance_cbs(my_rnp, my_rdp);
4535 	rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
4536 	raw_spin_unlock(&rcu_state.barrier_lock); /* irqs remain disabled. */
4537 	needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
4538 	rcu_segcblist_disable(&rdp->cblist);
4539 	WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) != !rcu_segcblist_n_cbs(&my_rdp->cblist));
4540 	check_cb_ovld_locked(my_rdp, my_rnp);
4541 	if (rcu_rdp_is_offloaded(my_rdp)) {
4542 		raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
4543 		__call_rcu_nocb_wake(my_rdp, true, flags);
4544 	} else {
4545 		rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
4546 		raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags);
4547 	}
4548 	if (needwake)
4549 		rcu_gp_kthread_wake();
4550 	lockdep_assert_irqs_enabled();
4551 	WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
4552 		  !rcu_segcblist_empty(&rdp->cblist),
4553 		  "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
4554 		  cpu, rcu_segcblist_n_cbs(&rdp->cblist),
4555 		  rcu_segcblist_first_cb(&rdp->cblist));
4556 }
4557 #endif
4558 
4559 /*
4560  * On non-huge systems, use expedited RCU grace periods to make suspend
4561  * and hibernation run faster.
4562  */
4563 static int rcu_pm_notify(struct notifier_block *self,
4564 			 unsigned long action, void *hcpu)
4565 {
4566 	switch (action) {
4567 	case PM_HIBERNATION_PREPARE:
4568 	case PM_SUSPEND_PREPARE:
4569 		rcu_async_hurry();
4570 		rcu_expedite_gp();
4571 		break;
4572 	case PM_POST_HIBERNATION:
4573 	case PM_POST_SUSPEND:
4574 		rcu_unexpedite_gp();
4575 		rcu_async_relax();
4576 		break;
4577 	default:
4578 		break;
4579 	}
4580 	return NOTIFY_OK;
4581 }
4582 
4583 #ifdef CONFIG_RCU_EXP_KTHREAD
4584 struct kthread_worker *rcu_exp_gp_kworker;
4585 struct kthread_worker *rcu_exp_par_gp_kworker;
4586 
4587 static void __init rcu_start_exp_gp_kworkers(void)
4588 {
4589 	const char *par_gp_kworker_name = "rcu_exp_par_gp_kthread_worker";
4590 	const char *gp_kworker_name = "rcu_exp_gp_kthread_worker";
4591 	struct sched_param param = { .sched_priority = kthread_prio };
4592 
4593 	rcu_exp_gp_kworker = kthread_create_worker(0, gp_kworker_name);
4594 	if (IS_ERR_OR_NULL(rcu_exp_gp_kworker)) {
4595 		pr_err("Failed to create %s!\n", gp_kworker_name);
4596 		return;
4597 	}
4598 
4599 	rcu_exp_par_gp_kworker = kthread_create_worker(0, par_gp_kworker_name);
4600 	if (IS_ERR_OR_NULL(rcu_exp_par_gp_kworker)) {
4601 		pr_err("Failed to create %s!\n", par_gp_kworker_name);
4602 		kthread_destroy_worker(rcu_exp_gp_kworker);
4603 		return;
4604 	}
4605 
4606 	sched_setscheduler_nocheck(rcu_exp_gp_kworker->task, SCHED_FIFO, &param);
4607 	sched_setscheduler_nocheck(rcu_exp_par_gp_kworker->task, SCHED_FIFO,
4608 				   &param);
4609 }
4610 
4611 static inline void rcu_alloc_par_gp_wq(void)
4612 {
4613 }
4614 #else /* !CONFIG_RCU_EXP_KTHREAD */
4615 struct workqueue_struct *rcu_par_gp_wq;
4616 
4617 static void __init rcu_start_exp_gp_kworkers(void)
4618 {
4619 }
4620 
4621 static inline void rcu_alloc_par_gp_wq(void)
4622 {
4623 	rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
4624 	WARN_ON(!rcu_par_gp_wq);
4625 }
4626 #endif /* CONFIG_RCU_EXP_KTHREAD */
4627 
4628 /*
4629  * Spawn the kthreads that handle RCU's grace periods.
4630  */
4631 static int __init rcu_spawn_gp_kthread(void)
4632 {
4633 	unsigned long flags;
4634 	struct rcu_node *rnp;
4635 	struct sched_param sp;
4636 	struct task_struct *t;
4637 	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
4638 
4639 	rcu_scheduler_fully_active = 1;
4640 	t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
4641 	if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
4642 		return 0;
4643 	if (kthread_prio) {
4644 		sp.sched_priority = kthread_prio;
4645 		sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4646 	}
4647 	rnp = rcu_get_root();
4648 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
4649 	WRITE_ONCE(rcu_state.gp_activity, jiffies);
4650 	WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
4651 	// Reset .gp_activity and .gp_req_activity before setting .gp_kthread.
4652 	smp_store_release(&rcu_state.gp_kthread, t);  /* ^^^ */
4653 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4654 	wake_up_process(t);
4655 	/* This is a pre-SMP initcall, we expect a single CPU */
4656 	WARN_ON(num_online_cpus() > 1);
4657 	/*
4658 	 * Those kthreads couldn't be created on rcu_init() -> rcutree_prepare_cpu()
4659 	 * due to rcu_scheduler_fully_active.
4660 	 */
4661 	rcu_spawn_cpu_nocb_kthread(smp_processor_id());
4662 	rcu_spawn_one_boost_kthread(rdp->mynode);
4663 	rcu_spawn_core_kthreads();
4664 	/* Create kthread worker for expedited GPs */
4665 	rcu_start_exp_gp_kworkers();
4666 	return 0;
4667 }
4668 early_initcall(rcu_spawn_gp_kthread);
4669 
4670 /*
4671  * This function is invoked towards the end of the scheduler's
4672  * initialization process.  Before this is called, the idle task might
4673  * contain synchronous grace-period primitives (during which time, this idle
4674  * task is booting the system, and such primitives are no-ops).  After this
4675  * function is called, any synchronous grace-period primitives are run as
4676  * expedited, with the requesting task driving the grace period forward.
4677  * A later core_initcall() rcu_set_runtime_mode() will switch to full
4678  * runtime RCU functionality.
4679  */
4680 void rcu_scheduler_starting(void)
4681 {
4682 	unsigned long flags;
4683 	struct rcu_node *rnp;
4684 
4685 	WARN_ON(num_online_cpus() != 1);
4686 	WARN_ON(nr_context_switches() > 0);
4687 	rcu_test_sync_prims();
4688 
4689 	// Fix up the ->gp_seq counters.
4690 	local_irq_save(flags);
4691 	rcu_for_each_node_breadth_first(rnp)
4692 		rnp->gp_seq_needed = rnp->gp_seq = rcu_state.gp_seq;
4693 	local_irq_restore(flags);
4694 
4695 	// Switch out of early boot mode.
4696 	rcu_scheduler_active = RCU_SCHEDULER_INIT;
4697 	rcu_test_sync_prims();
4698 }
4699 
4700 /*
4701  * Helper function for rcu_init() that initializes the rcu_state structure.
4702  */
4703 static void __init rcu_init_one(void)
4704 {
4705 	static const char * const buf[] = RCU_NODE_NAME_INIT;
4706 	static const char * const fqs[] = RCU_FQS_NAME_INIT;
4707 	static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4708 	static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4709 
4710 	int levelspread[RCU_NUM_LVLS];		/* kids/node in each level. */
4711 	int cpustride = 1;
4712 	int i;
4713 	int j;
4714 	struct rcu_node *rnp;
4715 
4716 	BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf));  /* Fix buf[] init! */
4717 
4718 	/* Silence gcc 4.8 false positive about array index out of range. */
4719 	if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4720 		panic("rcu_init_one: rcu_num_lvls out of range");
4721 
4722 	/* Initialize the level-tracking arrays. */
4723 
4724 	for (i = 1; i < rcu_num_lvls; i++)
4725 		rcu_state.level[i] =
4726 			rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
4727 	rcu_init_levelspread(levelspread, num_rcu_lvl);
4728 
4729 	/* Initialize the elements themselves, starting from the leaves. */
4730 
4731 	for (i = rcu_num_lvls - 1; i >= 0; i--) {
4732 		cpustride *= levelspread[i];
4733 		rnp = rcu_state.level[i];
4734 		for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
4735 			raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4736 			lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4737 						   &rcu_node_class[i], buf[i]);
4738 			raw_spin_lock_init(&rnp->fqslock);
4739 			lockdep_set_class_and_name(&rnp->fqslock,
4740 						   &rcu_fqs_class[i], fqs[i]);
4741 			rnp->gp_seq = rcu_state.gp_seq;
4742 			rnp->gp_seq_needed = rcu_state.gp_seq;
4743 			rnp->completedqs = rcu_state.gp_seq;
4744 			rnp->qsmask = 0;
4745 			rnp->qsmaskinit = 0;
4746 			rnp->grplo = j * cpustride;
4747 			rnp->grphi = (j + 1) * cpustride - 1;
4748 			if (rnp->grphi >= nr_cpu_ids)
4749 				rnp->grphi = nr_cpu_ids - 1;
4750 			if (i == 0) {
4751 				rnp->grpnum = 0;
4752 				rnp->grpmask = 0;
4753 				rnp->parent = NULL;
4754 			} else {
4755 				rnp->grpnum = j % levelspread[i - 1];
4756 				rnp->grpmask = BIT(rnp->grpnum);
4757 				rnp->parent = rcu_state.level[i - 1] +
4758 					      j / levelspread[i - 1];
4759 			}
4760 			rnp->level = i;
4761 			INIT_LIST_HEAD(&rnp->blkd_tasks);
4762 			rcu_init_one_nocb(rnp);
4763 			init_waitqueue_head(&rnp->exp_wq[0]);
4764 			init_waitqueue_head(&rnp->exp_wq[1]);
4765 			init_waitqueue_head(&rnp->exp_wq[2]);
4766 			init_waitqueue_head(&rnp->exp_wq[3]);
4767 			spin_lock_init(&rnp->exp_lock);
4768 			mutex_init(&rnp->boost_kthread_mutex);
4769 			raw_spin_lock_init(&rnp->exp_poll_lock);
4770 			rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED;
4771 			INIT_WORK(&rnp->exp_poll_wq, sync_rcu_do_polled_gp);
4772 		}
4773 	}
4774 
4775 	init_swait_queue_head(&rcu_state.gp_wq);
4776 	init_swait_queue_head(&rcu_state.expedited_wq);
4777 	rnp = rcu_first_leaf_node();
4778 	for_each_possible_cpu(i) {
4779 		while (i > rnp->grphi)
4780 			rnp++;
4781 		per_cpu_ptr(&rcu_data, i)->mynode = rnp;
4782 		rcu_boot_init_percpu_data(i);
4783 	}
4784 }
4785 
4786 /*
4787  * Force priority from the kernel command-line into range.
4788  */
4789 static void __init sanitize_kthread_prio(void)
4790 {
4791 	int kthread_prio_in = kthread_prio;
4792 
4793 	if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
4794 	    && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
4795 		kthread_prio = 2;
4796 	else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4797 		kthread_prio = 1;
4798 	else if (kthread_prio < 0)
4799 		kthread_prio = 0;
4800 	else if (kthread_prio > 99)
4801 		kthread_prio = 99;
4802 
4803 	if (kthread_prio != kthread_prio_in)
4804 		pr_alert("%s: Limited prio to %d from %d\n",
4805 			 __func__, kthread_prio, kthread_prio_in);
4806 }
4807 
4808 /*
4809  * Compute the rcu_node tree geometry from kernel parameters.  This cannot
4810  * replace the definitions in tree.h because those are needed to size
4811  * the ->node array in the rcu_state structure.
4812  */
4813 void rcu_init_geometry(void)
4814 {
4815 	ulong d;
4816 	int i;
4817 	static unsigned long old_nr_cpu_ids;
4818 	int rcu_capacity[RCU_NUM_LVLS];
4819 	static bool initialized;
4820 
4821 	if (initialized) {
4822 		/*
4823 		 * Warn if setup_nr_cpu_ids() had not yet been invoked,
4824 		 * unless nr_cpus_ids == NR_CPUS, in which case who cares?
4825 		 */
4826 		WARN_ON_ONCE(old_nr_cpu_ids != nr_cpu_ids);
4827 		return;
4828 	}
4829 
4830 	old_nr_cpu_ids = nr_cpu_ids;
4831 	initialized = true;
4832 
4833 	/*
4834 	 * Initialize any unspecified boot parameters.
4835 	 * The default values of jiffies_till_first_fqs and
4836 	 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4837 	 * value, which is a function of HZ, then adding one for each
4838 	 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4839 	 */
4840 	d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4841 	if (jiffies_till_first_fqs == ULONG_MAX)
4842 		jiffies_till_first_fqs = d;
4843 	if (jiffies_till_next_fqs == ULONG_MAX)
4844 		jiffies_till_next_fqs = d;
4845 	adjust_jiffies_till_sched_qs();
4846 
4847 	/* If the compile-time values are accurate, just leave. */
4848 	if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4849 	    nr_cpu_ids == NR_CPUS)
4850 		return;
4851 	pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4852 		rcu_fanout_leaf, nr_cpu_ids);
4853 
4854 	/*
4855 	 * The boot-time rcu_fanout_leaf parameter must be at least two
4856 	 * and cannot exceed the number of bits in the rcu_node masks.
4857 	 * Complain and fall back to the compile-time values if this
4858 	 * limit is exceeded.
4859 	 */
4860 	if (rcu_fanout_leaf < 2 ||
4861 	    rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4862 		rcu_fanout_leaf = RCU_FANOUT_LEAF;
4863 		WARN_ON(1);
4864 		return;
4865 	}
4866 
4867 	/*
4868 	 * Compute number of nodes that can be handled an rcu_node tree
4869 	 * with the given number of levels.
4870 	 */
4871 	rcu_capacity[0] = rcu_fanout_leaf;
4872 	for (i = 1; i < RCU_NUM_LVLS; i++)
4873 		rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4874 
4875 	/*
4876 	 * The tree must be able to accommodate the configured number of CPUs.
4877 	 * If this limit is exceeded, fall back to the compile-time values.
4878 	 */
4879 	if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4880 		rcu_fanout_leaf = RCU_FANOUT_LEAF;
4881 		WARN_ON(1);
4882 		return;
4883 	}
4884 
4885 	/* Calculate the number of levels in the tree. */
4886 	for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4887 	}
4888 	rcu_num_lvls = i + 1;
4889 
4890 	/* Calculate the number of rcu_nodes at each level of the tree. */
4891 	for (i = 0; i < rcu_num_lvls; i++) {
4892 		int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4893 		num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4894 	}
4895 
4896 	/* Calculate the total number of rcu_node structures. */
4897 	rcu_num_nodes = 0;
4898 	for (i = 0; i < rcu_num_lvls; i++)
4899 		rcu_num_nodes += num_rcu_lvl[i];
4900 }
4901 
4902 /*
4903  * Dump out the structure of the rcu_node combining tree associated
4904  * with the rcu_state structure.
4905  */
4906 static void __init rcu_dump_rcu_node_tree(void)
4907 {
4908 	int level = 0;
4909 	struct rcu_node *rnp;
4910 
4911 	pr_info("rcu_node tree layout dump\n");
4912 	pr_info(" ");
4913 	rcu_for_each_node_breadth_first(rnp) {
4914 		if (rnp->level != level) {
4915 			pr_cont("\n");
4916 			pr_info(" ");
4917 			level = rnp->level;
4918 		}
4919 		pr_cont("%d:%d ^%d  ", rnp->grplo, rnp->grphi, rnp->grpnum);
4920 	}
4921 	pr_cont("\n");
4922 }
4923 
4924 struct workqueue_struct *rcu_gp_wq;
4925 
4926 static void __init kfree_rcu_batch_init(void)
4927 {
4928 	int cpu;
4929 	int i, j;
4930 
4931 	/* Clamp it to [0:100] seconds interval. */
4932 	if (rcu_delay_page_cache_fill_msec < 0 ||
4933 		rcu_delay_page_cache_fill_msec > 100 * MSEC_PER_SEC) {
4934 
4935 		rcu_delay_page_cache_fill_msec =
4936 			clamp(rcu_delay_page_cache_fill_msec, 0,
4937 				(int) (100 * MSEC_PER_SEC));
4938 
4939 		pr_info("Adjusting rcutree.rcu_delay_page_cache_fill_msec to %d ms.\n",
4940 			rcu_delay_page_cache_fill_msec);
4941 	}
4942 
4943 	for_each_possible_cpu(cpu) {
4944 		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
4945 
4946 		for (i = 0; i < KFREE_N_BATCHES; i++) {
4947 			INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
4948 			krcp->krw_arr[i].krcp = krcp;
4949 
4950 			for (j = 0; j < FREE_N_CHANNELS; j++)
4951 				INIT_LIST_HEAD(&krcp->krw_arr[i].bulk_head_free[j]);
4952 		}
4953 
4954 		for (i = 0; i < FREE_N_CHANNELS; i++)
4955 			INIT_LIST_HEAD(&krcp->bulk_head[i]);
4956 
4957 		INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
4958 		INIT_DELAYED_WORK(&krcp->page_cache_work, fill_page_cache_func);
4959 		krcp->initialized = true;
4960 	}
4961 	if (register_shrinker(&kfree_rcu_shrinker, "rcu-kfree"))
4962 		pr_err("Failed to register kfree_rcu() shrinker!\n");
4963 }
4964 
4965 void __init rcu_init(void)
4966 {
4967 	int cpu = smp_processor_id();
4968 
4969 	rcu_early_boot_tests();
4970 
4971 	kfree_rcu_batch_init();
4972 	rcu_bootup_announce();
4973 	sanitize_kthread_prio();
4974 	rcu_init_geometry();
4975 	rcu_init_one();
4976 	if (dump_tree)
4977 		rcu_dump_rcu_node_tree();
4978 	if (use_softirq)
4979 		open_softirq(RCU_SOFTIRQ, rcu_core_si);
4980 
4981 	/*
4982 	 * We don't need protection against CPU-hotplug here because
4983 	 * this is called early in boot, before either interrupts
4984 	 * or the scheduler are operational.
4985 	 */
4986 	pm_notifier(rcu_pm_notify, 0);
4987 	WARN_ON(num_online_cpus() > 1); // Only one CPU this early in boot.
4988 	rcutree_prepare_cpu(cpu);
4989 	rcu_cpu_starting(cpu);
4990 	rcutree_online_cpu(cpu);
4991 
4992 	/* Create workqueue for Tree SRCU and for expedited GPs. */
4993 	rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
4994 	WARN_ON(!rcu_gp_wq);
4995 	rcu_alloc_par_gp_wq();
4996 
4997 	/* Fill in default value for rcutree.qovld boot parameter. */
4998 	/* -After- the rcu_node ->lock fields are initialized! */
4999 	if (qovld < 0)
5000 		qovld_calc = DEFAULT_RCU_QOVLD_MULT * qhimark;
5001 	else
5002 		qovld_calc = qovld;
5003 
5004 	// Kick-start in case any polled grace periods started early.
5005 	(void)start_poll_synchronize_rcu_expedited();
5006 
5007 	rcu_test_sync_prims();
5008 }
5009 
5010 #include "tree_stall.h"
5011 #include "tree_exp.h"
5012 #include "tree_nocb.h"
5013 #include "tree_plugin.h"
5014