tasks.h

   1/* SPDX-License-Identifier: GPL-2.0+ */
   2/*
   3 * Task-based RCU implementations.
   4 *
   5 * Copyright (C) 2020 Paul E. McKenney
   6 */
   7
   8#ifdef CONFIG_TASKS_RCU_GENERIC
   9#include "rcu_segcblist.h"
  10
  11////////////////////////////////////////////////////////////////////////
  12//
  13// Generic data structures.
  14
  15struct rcu_tasks;
  16typedef void (*rcu_tasks_gp_func_t)(struct rcu_tasks *rtp);
  17typedef void (*pregp_func_t)(struct list_head *hop);
  18typedef void (*pertask_func_t)(struct task_struct *t, struct list_head *hop);
  19typedef void (*postscan_func_t)(struct list_head *hop);
  20typedef void (*holdouts_func_t)(struct list_head *hop, bool ndrpt, bool *frptp);
  21typedef void (*postgp_func_t)(struct rcu_tasks *rtp);
  22
  23/**
  24 * struct rcu_tasks_percpu - Per-CPU component of definition for a Tasks-RCU-like mechanism.
  25 * @cblist: Callback list.
  26 * @lock: Lock protecting per-CPU callback list.
  27 * @rtp_jiffies: Jiffies counter value for statistics.
  28 * @lazy_timer: Timer to unlazify callbacks.
  29 * @urgent_gp: Number of additional non-lazy grace periods.
  30 * @rtp_n_lock_retries: Rough lock-contention statistic.
  31 * @rtp_work: Work queue for invoking callbacks.
  32 * @rtp_irq_work: IRQ work queue for deferred wakeups.
  33 * @barrier_q_head: RCU callback for barrier operation.
  34 * @rtp_blkd_tasks: List of tasks blocked as readers.
  35 * @rtp_exit_list: List of tasks in the latter portion of do_exit().
  36 * @cpu: CPU number corresponding to this entry.
  37 * @index: Index of this CPU in rtpcp_array of the rcu_tasks structure.
  38 * @rtpp: Pointer to the rcu_tasks structure.
  39 */
  40struct rcu_tasks_percpu {
  41	struct rcu_segcblist cblist;
  42	raw_spinlock_t __private lock;
  43	unsigned long rtp_jiffies;
  44	unsigned long rtp_n_lock_retries;
  45	struct timer_list lazy_timer;
  46	unsigned int urgent_gp;
  47	struct work_struct rtp_work;
  48	struct irq_work rtp_irq_work;
  49	struct rcu_head barrier_q_head;
  50	struct list_head rtp_blkd_tasks;
  51	struct list_head rtp_exit_list;
  52	int cpu;
  53	int index;
  54	struct rcu_tasks *rtpp;
  55};
  56
  57/**
  58 * struct rcu_tasks - Definition for a Tasks-RCU-like mechanism.
  59 * @cbs_wait: RCU wait allowing a new callback to get kthread's attention.
  60 * @cbs_gbl_lock: Lock protecting callback list.
  61 * @tasks_gp_mutex: Mutex protecting grace period, needed during mid-boot dead zone.
  62 * @gp_func: This flavor's grace-period-wait function.
  63 * @gp_state: Grace period's most recent state transition (debugging).
  64 * @gp_sleep: Per-grace-period sleep to prevent CPU-bound looping.
  65 * @init_fract: Initial backoff sleep interval.
  66 * @gp_jiffies: Time of last @gp_state transition.
  67 * @gp_start: Most recent grace-period start in jiffies.
  68 * @tasks_gp_seq: Number of grace periods completed since boot in upper bits.
  69 * @n_ipis: Number of IPIs sent to encourage grace periods to end.
  70 * @n_ipis_fails: Number of IPI-send failures.
  71 * @kthread_ptr: This flavor's grace-period/callback-invocation kthread.
  72 * @lazy_jiffies: Number of jiffies to allow callbacks to be lazy.
  73 * @pregp_func: This flavor's pre-grace-period function (optional).
  74 * @pertask_func: This flavor's per-task scan function (optional).
  75 * @postscan_func: This flavor's post-task scan function (optional).
  76 * @holdouts_func: This flavor's holdout-list scan function (optional).
  77 * @postgp_func: This flavor's post-grace-period function (optional).
  78 * @call_func: This flavor's call_rcu()-equivalent function.
  79 * @wait_state: Task state for synchronous grace-period waits (default TASK_UNINTERRUPTIBLE).
  80 * @rtpcpu: This flavor's rcu_tasks_percpu structure.
  81 * @rtpcp_array: Array of pointers to rcu_tasks_percpu structure of CPUs in cpu_possible_mask.
  82 * @percpu_enqueue_shift: Shift down CPU ID this much when enqueuing callbacks.
  83 * @percpu_enqueue_lim: Number of per-CPU callback queues in use for enqueuing.
  84 * @percpu_dequeue_lim: Number of per-CPU callback queues in use for dequeuing.
  85 * @percpu_dequeue_gpseq: RCU grace-period number to propagate enqueue limit to dequeuers.
  86 * @barrier_q_mutex: Serialize barrier operations.
  87 * @barrier_q_count: Number of queues being waited on.
  88 * @barrier_q_completion: Barrier wait/wakeup mechanism.
  89 * @barrier_q_seq: Sequence number for barrier operations.
  90 * @barrier_q_start: Most recent barrier start in jiffies.
  91 * @name: This flavor's textual name.
  92 * @kname: This flavor's kthread name.
  93 */
  94struct rcu_tasks {
  95	struct rcuwait cbs_wait;
  96	raw_spinlock_t cbs_gbl_lock;
  97	struct mutex tasks_gp_mutex;
  98	int gp_state;
  99	int gp_sleep;
 100	int init_fract;
 101	unsigned long gp_jiffies;
 102	unsigned long gp_start;
 103	unsigned long tasks_gp_seq;
 104	unsigned long n_ipis;
 105	unsigned long n_ipis_fails;
 106	struct task_struct *kthread_ptr;
 107	unsigned long lazy_jiffies;
 108	rcu_tasks_gp_func_t gp_func;
 109	pregp_func_t pregp_func;
 110	pertask_func_t pertask_func;
 111	postscan_func_t postscan_func;
 112	holdouts_func_t holdouts_func;
 113	postgp_func_t postgp_func;
 114	call_rcu_func_t call_func;
 115	unsigned int wait_state;
 116	struct rcu_tasks_percpu __percpu *rtpcpu;
 117	struct rcu_tasks_percpu **rtpcp_array;
 118	int percpu_enqueue_shift;
 119	int percpu_enqueue_lim;
 120	int percpu_dequeue_lim;
 121	unsigned long percpu_dequeue_gpseq;
 122	struct mutex barrier_q_mutex;
 123	atomic_t barrier_q_count;
 124	struct completion barrier_q_completion;
 125	unsigned long barrier_q_seq;
 126	unsigned long barrier_q_start;
 127	char *name;
 128	char *kname;
 129};
 130
 131static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp);
 132
 133#define DEFINE_RCU_TASKS(rt_name, gp, call, n)						\
 134static DEFINE_PER_CPU(struct rcu_tasks_percpu, rt_name ## __percpu) = {			\
 135	.lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name ## __percpu.cbs_pcpu_lock),		\
 136	.rtp_irq_work = IRQ_WORK_INIT_HARD(call_rcu_tasks_iw_wakeup),			\
 137};											\
 138static struct rcu_tasks rt_name =							\
 139{											\
 140	.cbs_wait = __RCUWAIT_INITIALIZER(rt_name.wait),				\
 141	.cbs_gbl_lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name.cbs_gbl_lock),			\
 142	.tasks_gp_mutex = __MUTEX_INITIALIZER(rt_name.tasks_gp_mutex),			\
 143	.gp_func = gp,									\
 144	.call_func = call,								\
 145	.wait_state = TASK_UNINTERRUPTIBLE,						\
 146	.rtpcpu = &rt_name ## __percpu,							\
 147	.lazy_jiffies = DIV_ROUND_UP(HZ, 4),						\
 148	.name = n,									\
 149	.percpu_enqueue_shift = order_base_2(CONFIG_NR_CPUS),				\
 150	.percpu_enqueue_lim = 1,							\
 151	.percpu_dequeue_lim = 1,							\
 152	.barrier_q_mutex = __MUTEX_INITIALIZER(rt_name.barrier_q_mutex),		\
 153	.barrier_q_seq = (0UL - 50UL) << RCU_SEQ_CTR_SHIFT,				\
 154	.kname = #rt_name,								\
 155}
 156
 157#ifdef CONFIG_TASKS_RCU
 158
 159/* Report delay of scan exiting tasklist in rcu_tasks_postscan(). */
 160static void tasks_rcu_exit_srcu_stall(struct timer_list *unused);
 161static DEFINE_TIMER(tasks_rcu_exit_srcu_stall_timer, tasks_rcu_exit_srcu_stall);
 162#endif
 163
 164/* Avoid IPIing CPUs early in the grace period. */
 165#define RCU_TASK_IPI_DELAY (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) ? HZ / 2 : 0)
 166static int rcu_task_ipi_delay __read_mostly = RCU_TASK_IPI_DELAY;
 167module_param(rcu_task_ipi_delay, int, 0644);
 168
 169/* Control stall timeouts.  Disable with <= 0, otherwise jiffies till stall. */
 170#define RCU_TASK_BOOT_STALL_TIMEOUT (HZ * 30)
 171#define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
 172static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
 173module_param(rcu_task_stall_timeout, int, 0644);
 174#define RCU_TASK_STALL_INFO (HZ * 10)
 175static int rcu_task_stall_info __read_mostly = RCU_TASK_STALL_INFO;
 176module_param(rcu_task_stall_info, int, 0644);
 177static int rcu_task_stall_info_mult __read_mostly = 3;
 178module_param(rcu_task_stall_info_mult, int, 0444);
 179
 180static int rcu_task_enqueue_lim __read_mostly = -1;
 181module_param(rcu_task_enqueue_lim, int, 0444);
 182
 183static bool rcu_task_cb_adjust;
 184static int rcu_task_contend_lim __read_mostly = 100;
 185module_param(rcu_task_contend_lim, int, 0444);
 186static int rcu_task_collapse_lim __read_mostly = 10;
 187module_param(rcu_task_collapse_lim, int, 0444);
 188static int rcu_task_lazy_lim __read_mostly = 32;
 189module_param(rcu_task_lazy_lim, int, 0444);
 190
 191static int rcu_task_cpu_ids;
 192
 193/* RCU tasks grace-period state for debugging. */
 194#define RTGS_INIT		 0
 195#define RTGS_WAIT_WAIT_CBS	 1
 196#define RTGS_WAIT_GP		 2
 197#define RTGS_PRE_WAIT_GP	 3
 198#define RTGS_SCAN_TASKLIST	 4
 199#define RTGS_POST_SCAN_TASKLIST	 5
 200#define RTGS_WAIT_SCAN_HOLDOUTS	 6
 201#define RTGS_SCAN_HOLDOUTS	 7
 202#define RTGS_POST_GP		 8
 203#define RTGS_WAIT_READERS	 9
 204#define RTGS_INVOKE_CBS		10
 205#define RTGS_WAIT_CBS		11
 206#ifndef CONFIG_TINY_RCU
 207static const char * const rcu_tasks_gp_state_names[] = {
 208	"RTGS_INIT",
 209	"RTGS_WAIT_WAIT_CBS",
 210	"RTGS_WAIT_GP",
 211	"RTGS_PRE_WAIT_GP",
 212	"RTGS_SCAN_TASKLIST",
 213	"RTGS_POST_SCAN_TASKLIST",
 214	"RTGS_WAIT_SCAN_HOLDOUTS",
 215	"RTGS_SCAN_HOLDOUTS",
 216	"RTGS_POST_GP",
 217	"RTGS_WAIT_READERS",
 218	"RTGS_INVOKE_CBS",
 219	"RTGS_WAIT_CBS",
 220};
 221#endif /* #ifndef CONFIG_TINY_RCU */
 222
 223////////////////////////////////////////////////////////////////////////
 224//
 225// Generic code.
 226
 227static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp);
 228
 229/* Record grace-period phase and time. */
 230static void set_tasks_gp_state(struct rcu_tasks *rtp, int newstate)
 231{
 232	rtp->gp_state = newstate;
 233	rtp->gp_jiffies = jiffies;
 234}
 235
 236#ifndef CONFIG_TINY_RCU
 237/* Return state name. */
 238static const char *tasks_gp_state_getname(struct rcu_tasks *rtp)
 239{
 240	int i = data_race(rtp->gp_state); // Let KCSAN detect update races
 241	int j = READ_ONCE(i); // Prevent the compiler from reading twice
 242
 243	if (j >= ARRAY_SIZE(rcu_tasks_gp_state_names))
 244		return "???";
 245	return rcu_tasks_gp_state_names[j];
 246}
 247#endif /* #ifndef CONFIG_TINY_RCU */
 248
 249// Initialize per-CPU callback lists for the specified flavor of
 250// Tasks RCU.  Do not enqueue callbacks before this function is invoked.
 251static void cblist_init_generic(struct rcu_tasks *rtp)
 252{
 253	int cpu;
 254	int lim;
 255	int shift;
 256	int maxcpu;
 257	int index = 0;
 258
 259	if (rcu_task_enqueue_lim < 0) {
 260		rcu_task_enqueue_lim = 1;
 261		rcu_task_cb_adjust = true;
 262	} else if (rcu_task_enqueue_lim == 0) {
 263		rcu_task_enqueue_lim = 1;
 264	}
 265	lim = rcu_task_enqueue_lim;
 266
 267	rtp->rtpcp_array = kcalloc(num_possible_cpus(), sizeof(struct rcu_tasks_percpu *), GFP_KERNEL);
 268	BUG_ON(!rtp->rtpcp_array);
 269
 270	for_each_possible_cpu(cpu) {
 271		struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
 272
 273		WARN_ON_ONCE(!rtpcp);
 274		if (cpu)
 275			raw_spin_lock_init(&ACCESS_PRIVATE(rtpcp, lock));
 276		if (rcu_segcblist_empty(&rtpcp->cblist))
 277			rcu_segcblist_init(&rtpcp->cblist);
 278		INIT_WORK(&rtpcp->rtp_work, rcu_tasks_invoke_cbs_wq);
 279		rtpcp->cpu = cpu;
 280		rtpcp->rtpp = rtp;
 281		rtpcp->index = index;
 282		rtp->rtpcp_array[index] = rtpcp;
 283		index++;
 284		if (!rtpcp->rtp_blkd_tasks.next)
 285			INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
 286		if (!rtpcp->rtp_exit_list.next)
 287			INIT_LIST_HEAD(&rtpcp->rtp_exit_list);
 288		rtpcp->barrier_q_head.next = &rtpcp->barrier_q_head;
 289		maxcpu = cpu;
 290	}
 291
 292	rcu_task_cpu_ids = maxcpu + 1;
 293	if (lim > rcu_task_cpu_ids)
 294		lim = rcu_task_cpu_ids;
 295	shift = ilog2(rcu_task_cpu_ids / lim);
 296	if (((rcu_task_cpu_ids - 1) >> shift) >= lim)
 297		shift++;
 298	WRITE_ONCE(rtp->percpu_enqueue_shift, shift);
 299	WRITE_ONCE(rtp->percpu_dequeue_lim, lim);
 300	smp_store_release(&rtp->percpu_enqueue_lim, lim);
 301
 302	pr_info("%s: Setting shift to %d and lim to %d rcu_task_cb_adjust=%d rcu_task_cpu_ids=%d.\n",
 303			rtp->name, data_race(rtp->percpu_enqueue_shift), data_race(rtp->percpu_enqueue_lim),
 304			rcu_task_cb_adjust, rcu_task_cpu_ids);
 305}
 306
 307// Compute wakeup time for lazy callback timer.
 308static unsigned long rcu_tasks_lazy_time(struct rcu_tasks *rtp)
 309{
 310	return jiffies + rtp->lazy_jiffies;
 311}
 312
 313// Timer handler that unlazifies lazy callbacks.
 314static void call_rcu_tasks_generic_timer(struct timer_list *tlp)
 315{
 316	unsigned long flags;
 317	bool needwake = false;
 318	struct rcu_tasks *rtp;
 319	struct rcu_tasks_percpu *rtpcp = timer_container_of(rtpcp, tlp,
 320						            lazy_timer);
 321
 322	rtp = rtpcp->rtpp;
 323	raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
 324	if (!rcu_segcblist_empty(&rtpcp->cblist) && rtp->lazy_jiffies) {
 325		if (!rtpcp->urgent_gp)
 326			rtpcp->urgent_gp = 1;
 327		needwake = true;
 328		mod_timer(&rtpcp->lazy_timer, rcu_tasks_lazy_time(rtp));
 329	}
 330	raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
 331	if (needwake)
 332		rcuwait_wake_up(&rtp->cbs_wait);
 333}
 334
 335// IRQ-work handler that does deferred wakeup for call_rcu_tasks_generic().
 336static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp)
 337{
 338	struct rcu_tasks *rtp;
 339	struct rcu_tasks_percpu *rtpcp = container_of(iwp, struct rcu_tasks_percpu, rtp_irq_work);
 340
 341	rtp = rtpcp->rtpp;
 342	rcuwait_wake_up(&rtp->cbs_wait);
 343}
 344
 345// Enqueue a callback for the specified flavor of Tasks RCU.
 346static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
 347				   struct rcu_tasks *rtp)
 348{
 349	int chosen_cpu;
 350	unsigned long flags;
 351	bool havekthread = smp_load_acquire(&rtp->kthread_ptr);
 352	int ideal_cpu;
 353	unsigned long j;
 354	bool needadjust = false;
 355	bool needwake;
 356	struct rcu_tasks_percpu *rtpcp;
 357
 358	rhp->next = NULL;
 359	rhp->func = func;
 360	local_irq_save(flags);
 361	rcu_read_lock();
 362	ideal_cpu = smp_processor_id() >> READ_ONCE(rtp->percpu_enqueue_shift);
 363	chosen_cpu = cpumask_next(ideal_cpu - 1, cpu_possible_mask);
 364	WARN_ON_ONCE(chosen_cpu >= rcu_task_cpu_ids);
 365	rtpcp = per_cpu_ptr(rtp->rtpcpu, chosen_cpu);
 366	if (!raw_spin_trylock_rcu_node(rtpcp)) { // irqs already disabled.
 367		raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
 368		j = jiffies;
 369		if (rtpcp->rtp_jiffies != j) {
 370			rtpcp->rtp_jiffies = j;
 371			rtpcp->rtp_n_lock_retries = 0;
 372		}
 373		if (rcu_task_cb_adjust && ++rtpcp->rtp_n_lock_retries > rcu_task_contend_lim &&
 374		    READ_ONCE(rtp->percpu_enqueue_lim) != rcu_task_cpu_ids)
 375			needadjust = true;  // Defer adjustment to avoid deadlock.
 376	}
 377	// Queuing callbacks before initialization not yet supported.
 378	if (WARN_ON_ONCE(!rcu_segcblist_is_enabled(&rtpcp->cblist)))
 379		rcu_segcblist_init(&rtpcp->cblist);
 380	needwake = (func == wakeme_after_rcu) ||
 381		   (rcu_segcblist_n_cbs(&rtpcp->cblist) == rcu_task_lazy_lim);
 382	if (havekthread && !needwake && !timer_pending(&rtpcp->lazy_timer)) {
 383		if (rtp->lazy_jiffies)
 384			mod_timer(&rtpcp->lazy_timer, rcu_tasks_lazy_time(rtp));
 385		else
 386			needwake = rcu_segcblist_empty(&rtpcp->cblist);
 387	}
 388	if (needwake)
 389		rtpcp->urgent_gp = 3;
 390	rcu_segcblist_enqueue(&rtpcp->cblist, rhp);
 391	raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
 392	if (unlikely(needadjust)) {
 393		raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
 394		if (rtp->percpu_enqueue_lim != rcu_task_cpu_ids) {
 395			WRITE_ONCE(rtp->percpu_enqueue_shift, 0);
 396			WRITE_ONCE(rtp->percpu_dequeue_lim, rcu_task_cpu_ids);
 397			smp_store_release(&rtp->percpu_enqueue_lim, rcu_task_cpu_ids);
 398			pr_info("Switching %s to per-CPU callback queuing.\n", rtp->name);
 399		}
 400		raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
 401	}
 402	rcu_read_unlock();
 403	/* We can't create the thread unless interrupts are enabled. */
 404	if (needwake && READ_ONCE(rtp->kthread_ptr))
 405		irq_work_queue(&rtpcp->rtp_irq_work);
 406}
 407
 408// RCU callback function for rcu_barrier_tasks_generic().
 409static void rcu_barrier_tasks_generic_cb(struct rcu_head *rhp)
 410{
 411	struct rcu_tasks *rtp;
 412	struct rcu_tasks_percpu *rtpcp;
 413
 414	rhp->next = rhp; // Mark the callback as having been invoked.
 415	rtpcp = container_of(rhp, struct rcu_tasks_percpu, barrier_q_head);
 416	rtp = rtpcp->rtpp;
 417	if (atomic_dec_and_test(&rtp->barrier_q_count))
 418		complete(&rtp->barrier_q_completion);
 419}
 420
 421// Wait for all in-flight callbacks for the specified RCU Tasks flavor.
 422// Operates in a manner similar to rcu_barrier().
 423static void __maybe_unused rcu_barrier_tasks_generic(struct rcu_tasks *rtp)
 424{
 425	int cpu;
 426	unsigned long flags;
 427	struct rcu_tasks_percpu *rtpcp;
 428	unsigned long s = rcu_seq_snap(&rtp->barrier_q_seq);
 429
 430	mutex_lock(&rtp->barrier_q_mutex);
 431	if (rcu_seq_done(&rtp->barrier_q_seq, s)) {
 432		smp_mb();
 433		mutex_unlock(&rtp->barrier_q_mutex);
 434		return;
 435	}
 436	rtp->barrier_q_start = jiffies;
 437	rcu_seq_start(&rtp->barrier_q_seq);
 438	init_completion(&rtp->barrier_q_completion);
 439	atomic_set(&rtp->barrier_q_count, 2);
 440	for_each_possible_cpu(cpu) {
 441		if (cpu >= smp_load_acquire(&rtp->percpu_dequeue_lim))
 442			break;
 443		rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
 444		rtpcp->barrier_q_head.func = rcu_barrier_tasks_generic_cb;
 445		raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
 446		if (rcu_segcblist_entrain(&rtpcp->cblist, &rtpcp->barrier_q_head))
 447			atomic_inc(&rtp->barrier_q_count);
 448		raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
 449	}
 450	if (atomic_sub_and_test(2, &rtp->barrier_q_count))
 451		complete(&rtp->barrier_q_completion);
 452	wait_for_completion(&rtp->barrier_q_completion);
 453	rcu_seq_end(&rtp->barrier_q_seq);
 454	mutex_unlock(&rtp->barrier_q_mutex);
 455}
 456
 457// Advance callbacks and indicate whether either a grace period or
 458// callback invocation is needed.
 459static int rcu_tasks_need_gpcb(struct rcu_tasks *rtp)
 460{
 461	int cpu;
 462	int dequeue_limit;
 463	unsigned long flags;
 464	bool gpdone = poll_state_synchronize_rcu(rtp->percpu_dequeue_gpseq);
 465	long n;
 466	long ncbs = 0;
 467	long ncbsnz = 0;
 468	int needgpcb = 0;
 469
 470	dequeue_limit = smp_load_acquire(&rtp->percpu_dequeue_lim);
 471	for (cpu = 0; cpu < dequeue_limit; cpu++) {
 472		if (!cpu_possible(cpu))
 473			continue;
 474		struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
 475
 476		/* Advance and accelerate any new callbacks. */
 477		if (!rcu_segcblist_n_cbs(&rtpcp->cblist))
 478			continue;
 479		raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
 480		// Should we shrink down to a single callback queue?
 481		n = rcu_segcblist_n_cbs(&rtpcp->cblist);
 482		if (n) {
 483			ncbs += n;
 484			if (cpu > 0)
 485				ncbsnz += n;
 486		}
 487		rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
 488		(void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
 489		if (rtpcp->urgent_gp > 0 && rcu_segcblist_pend_cbs(&rtpcp->cblist)) {
 490			if (rtp->lazy_jiffies)
 491				rtpcp->urgent_gp--;
 492			needgpcb |= 0x3;
 493		} else if (rcu_segcblist_empty(&rtpcp->cblist)) {
 494			rtpcp->urgent_gp = 0;
 495		}
 496		if (rcu_segcblist_ready_cbs(&rtpcp->cblist))
 497			needgpcb |= 0x1;
 498		raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
 499	}
 500
 501	// Shrink down to a single callback queue if appropriate.
 502	// This is done in two stages: (1) If there are no more than
 503	// rcu_task_collapse_lim callbacks on CPU 0 and none on any other
 504	// CPU, limit enqueueing to CPU 0.  (2) After an RCU grace period,
 505	// if there has not been an increase in callbacks, limit dequeuing
 506	// to CPU 0.  Note the matching RCU read-side critical section in
 507	// call_rcu_tasks_generic().
 508	if (rcu_task_cb_adjust && ncbs <= rcu_task_collapse_lim) {
 509		raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
 510		if (rtp->percpu_enqueue_lim > 1) {
 511			WRITE_ONCE(rtp->percpu_enqueue_shift, order_base_2(rcu_task_cpu_ids));
 512			smp_store_release(&rtp->percpu_enqueue_lim, 1);
 513			rtp->percpu_dequeue_gpseq = get_state_synchronize_rcu();
 514			gpdone = false;
 515			pr_info("Starting switch %s to CPU-0 callback queuing.\n", rtp->name);
 516		}
 517		raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
 518	}
 519	if (rcu_task_cb_adjust && !ncbsnz && gpdone) {
 520		raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
 521		if (rtp->percpu_enqueue_lim < rtp->percpu_dequeue_lim) {
 522			WRITE_ONCE(rtp->percpu_dequeue_lim, 1);
 523			pr_info("Completing switch %s to CPU-0 callback queuing.\n", rtp->name);
 524		}
 525		if (rtp->percpu_dequeue_lim == 1) {
 526			for (cpu = rtp->percpu_dequeue_lim; cpu < rcu_task_cpu_ids; cpu++) {
 527				if (!cpu_possible(cpu))
 528					continue;
 529				struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
 530
 531				WARN_ON_ONCE(rcu_segcblist_n_cbs(&rtpcp->cblist));
 532			}
 533		}
 534		raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
 535	}
 536
 537	return needgpcb;
 538}
 539
 540// Advance callbacks and invoke any that are ready.
 541static void rcu_tasks_invoke_cbs(struct rcu_tasks *rtp, struct rcu_tasks_percpu *rtpcp)
 542{
 543	int cpuwq;
 544	unsigned long flags;
 545	int len;
 546	int index;
 547	struct rcu_head *rhp;
 548	struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
 549	struct rcu_tasks_percpu *rtpcp_next;
 550
 551	index = rtpcp->index * 2 + 1;
 552	if (index < num_possible_cpus()) {
 553		rtpcp_next = rtp->rtpcp_array[index];
 554		if (rtpcp_next->cpu < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
 555			cpuwq = rcu_cpu_beenfullyonline(rtpcp_next->cpu) ? rtpcp_next->cpu : WORK_CPU_UNBOUND;
 556			queue_work_on(cpuwq, system_percpu_wq, &rtpcp_next->rtp_work);
 557			index++;
 558			if (index < num_possible_cpus()) {
 559				rtpcp_next = rtp->rtpcp_array[index];
 560				if (rtpcp_next->cpu < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
 561					cpuwq = rcu_cpu_beenfullyonline(rtpcp_next->cpu) ? rtpcp_next->cpu : WORK_CPU_UNBOUND;
 562					queue_work_on(cpuwq, system_percpu_wq, &rtpcp_next->rtp_work);
 563				}
 564			}
 565		}
 566	}
 567
 568	if (rcu_segcblist_empty(&rtpcp->cblist))
 569		return;
 570	raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
 571	rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
 572	rcu_segcblist_extract_done_cbs(&rtpcp->cblist, &rcl);
 573	raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
 574	len = rcl.len;
 575	for (rhp = rcu_cblist_dequeue(&rcl); rhp; rhp = rcu_cblist_dequeue(&rcl)) {
 576		debug_rcu_head_callback(rhp);
 577		local_bh_disable();
 578		rhp->func(rhp);
 579		local_bh_enable();
 580		cond_resched();
 581	}
 582	raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
 583	rcu_segcblist_add_len(&rtpcp->cblist, -len);
 584	(void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
 585	raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
 586}
 587
 588// Workqueue flood to advance callbacks and invoke any that are ready.
 589static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp)
 590{
 591	struct rcu_tasks *rtp;
 592	struct rcu_tasks_percpu *rtpcp = container_of(wp, struct rcu_tasks_percpu, rtp_work);
 593
 594	rtp = rtpcp->rtpp;
 595	rcu_tasks_invoke_cbs(rtp, rtpcp);
 596}
 597
 598// Wait for one grace period.
 599static void rcu_tasks_one_gp(struct rcu_tasks *rtp, bool midboot)
 600{
 601	int needgpcb;
 602
 603	mutex_lock(&rtp->tasks_gp_mutex);
 604
 605	// If there were none, wait a bit and start over.
 606	if (unlikely(midboot)) {
 607		needgpcb = 0x2;
 608	} else {
 609		mutex_unlock(&rtp->tasks_gp_mutex);
 610		set_tasks_gp_state(rtp, RTGS_WAIT_CBS);
 611		rcuwait_wait_event(&rtp->cbs_wait,
 612				   (needgpcb = rcu_tasks_need_gpcb(rtp)),
 613				   TASK_IDLE);
 614		mutex_lock(&rtp->tasks_gp_mutex);
 615	}
 616
 617	if (needgpcb & 0x2) {
 618		// Wait for one grace period.
 619		set_tasks_gp_state(rtp, RTGS_WAIT_GP);
 620		rtp->gp_start = jiffies;
 621		rcu_seq_start(&rtp->tasks_gp_seq);
 622		rtp->gp_func(rtp);
 623		rcu_seq_end(&rtp->tasks_gp_seq);
 624	}
 625
 626	// Invoke callbacks.
 627	set_tasks_gp_state(rtp, RTGS_INVOKE_CBS);
 628	rcu_tasks_invoke_cbs(rtp, per_cpu_ptr(rtp->rtpcpu, 0));
 629	mutex_unlock(&rtp->tasks_gp_mutex);
 630}
 631
 632// RCU-tasks kthread that detects grace periods and invokes callbacks.
 633static int __noreturn rcu_tasks_kthread(void *arg)
 634{
 635	int cpu;
 636	struct rcu_tasks *rtp = arg;
 637
 638	for_each_possible_cpu(cpu) {
 639		struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
 640
 641		timer_setup(&rtpcp->lazy_timer, call_rcu_tasks_generic_timer, 0);
 642		rtpcp->urgent_gp = 1;
 643	}
 644
 645	/* Run on housekeeping CPUs by default.  Sysadm can move if desired. */
 646	housekeeping_affine(current, HK_TYPE_RCU);
 647	smp_store_release(&rtp->kthread_ptr, current); // Let GPs start!
 648
 649	/*
 650	 * Each pass through the following loop makes one check for
 651	 * newly arrived callbacks, and, if there are some, waits for
 652	 * one RCU-tasks grace period and then invokes the callbacks.
 653	 * This loop is terminated by the system going down.  ;-)
 654	 */
 655	for (;;) {
 656		// Wait for one grace period and invoke any callbacks
 657		// that are ready.
 658		rcu_tasks_one_gp(rtp, false);
 659
 660		// Paranoid sleep to keep this from entering a tight loop.
 661		schedule_timeout_idle(rtp->gp_sleep);
 662	}
 663}
 664
 665// Wait for a grace period for the specified flavor of Tasks RCU.
 666static void synchronize_rcu_tasks_generic(struct rcu_tasks *rtp)
 667{
 668	/* Complain if the scheduler has not started.  */
 669	if (WARN_ONCE(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
 670			 "synchronize_%s() called too soon", rtp->name))
 671		return;
 672
 673	// If the grace-period kthread is running, use it.
 674	if (READ_ONCE(rtp->kthread_ptr)) {
 675		wait_rcu_gp_state(rtp->wait_state, rtp->call_func);
 676		return;
 677	}
 678	rcu_tasks_one_gp(rtp, true);
 679}
 680
 681/* Spawn RCU-tasks grace-period kthread. */
 682static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp)
 683{
 684	struct task_struct *t;
 685
 686	t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname);
 687	if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name))
 688		return;
 689	smp_mb(); /* Ensure others see full kthread. */
 690}
 691
 692#ifndef CONFIG_TINY_RCU
 693
 694/*
 695 * Print any non-default Tasks RCU settings.
 696 */
 697static void __init rcu_tasks_bootup_oddness(void)
 698{
 699#if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
 700	int rtsimc;
 701
 702	if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
 703		pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
 704	rtsimc = clamp(rcu_task_stall_info_mult, 1, 10);
 705	if (rtsimc != rcu_task_stall_info_mult) {
 706		pr_info("\tTasks-RCU CPU stall info multiplier clamped to %d (rcu_task_stall_info_mult).\n", rtsimc);
 707		rcu_task_stall_info_mult = rtsimc;
 708	}
 709#endif /* #ifdef CONFIG_TASKS_RCU */
 710#ifdef CONFIG_TASKS_RCU
 711	pr_info("\tTrampoline variant of Tasks RCU enabled.\n");
 712#endif /* #ifdef CONFIG_TASKS_RCU */
 713#ifdef CONFIG_TASKS_RUDE_RCU
 714	pr_info("\tRude variant of Tasks RCU enabled.\n");
 715#endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
 716#ifdef CONFIG_TASKS_TRACE_RCU
 717	pr_info("\tTracing variant of Tasks RCU enabled.\n");
 718#endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
 719}
 720
 721
 722/* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */
 723static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s)
 724{
 725	int cpu;
 726	bool havecbs = false;
 727	bool haveurgent = false;
 728	bool haveurgentcbs = false;
 729
 730	for_each_possible_cpu(cpu) {
 731		struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
 732
 733		if (!data_race(rcu_segcblist_empty(&rtpcp->cblist)))
 734			havecbs = true;
 735		if (data_race(rtpcp->urgent_gp))
 736			haveurgent = true;
 737		if (!data_race(rcu_segcblist_empty(&rtpcp->cblist)) && data_race(rtpcp->urgent_gp))
 738			haveurgentcbs = true;
 739		if (havecbs && haveurgent && haveurgentcbs)
 740			break;
 741	}
 742	pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c%c%c l:%lu %s\n",
 743		rtp->kname,
 744		tasks_gp_state_getname(rtp), data_race(rtp->gp_state),
 745		jiffies - data_race(rtp->gp_jiffies),
 746		data_race(rcu_seq_current(&rtp->tasks_gp_seq)),
 747		data_race(rtp->n_ipis_fails), data_race(rtp->n_ipis),
 748		".k"[!!data_race(rtp->kthread_ptr)],
 749		".C"[havecbs],
 750		".u"[haveurgent],
 751		".U"[haveurgentcbs],
 752		rtp->lazy_jiffies,
 753		s);
 754}
 755
 756/* Dump out more rcutorture-relevant state common to all RCU-tasks flavors. */
 757static void rcu_tasks_torture_stats_print_generic(struct rcu_tasks *rtp, char *tt,
 758						  char *tf, char *tst)
 759{
 760	cpumask_var_t cm;
 761	int cpu;
 762	bool gotcb = false;
 763	unsigned long j = jiffies;
 764
 765	pr_alert("%s%s Tasks%s RCU g%ld gp_start %lu gp_jiffies %lu gp_state %d (%s).\n",
 766		 tt, tf, tst, data_race(rtp->tasks_gp_seq),
 767		 j - data_race(rtp->gp_start), j - data_race(rtp->gp_jiffies),
 768		 data_race(rtp->gp_state), tasks_gp_state_getname(rtp));
 769	pr_alert("\tEnqueue shift %d limit %d Dequeue limit %d gpseq %lu.\n",
 770		 data_race(rtp->percpu_enqueue_shift),
 771		 data_race(rtp->percpu_enqueue_lim),
 772		 data_race(rtp->percpu_dequeue_lim),
 773		 data_race(rtp->percpu_dequeue_gpseq));
 774	(void)zalloc_cpumask_var(&cm, GFP_KERNEL);
 775	pr_alert("\tCallback counts:");
 776	for_each_possible_cpu(cpu) {
 777		long n;
 778		struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
 779
 780		if (cpumask_available(cm) && !rcu_barrier_cb_is_done(&rtpcp->barrier_q_head))
 781			cpumask_set_cpu(cpu, cm);
 782		n = rcu_segcblist_n_cbs(&rtpcp->cblist);
 783		if (!n)
 784			continue;
 785		pr_cont(" %d:%ld", cpu, n);
 786		gotcb = true;
 787	}
 788	if (gotcb)
 789		pr_cont(".\n");
 790	else
 791		pr_cont(" (none).\n");
 792	pr_alert("\tBarrier seq %lu start %lu count %d holdout CPUs ",
 793		 data_race(rtp->barrier_q_seq), j - data_race(rtp->barrier_q_start),
 794		 atomic_read(&rtp->barrier_q_count));
 795	if (cpumask_available(cm) && !cpumask_empty(cm))
 796		pr_cont(" %*pbl.\n", cpumask_pr_args(cm));
 797	else
 798		pr_cont("(none).\n");
 799	free_cpumask_var(cm);
 800}
 801
 802#endif // #ifndef CONFIG_TINY_RCU
 803
 804static void exit_tasks_rcu_finish_trace(struct task_struct *t);
 805
 806#if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
 807
 808////////////////////////////////////////////////////////////////////////
 809//
 810// Shared code between task-list-scanning variants of Tasks RCU.
 811
 812/* Wait for one RCU-tasks grace period. */
 813static void rcu_tasks_wait_gp(struct rcu_tasks *rtp)
 814{
 815	struct task_struct *g;
 816	int fract;
 817	LIST_HEAD(holdouts);
 818	unsigned long j;
 819	unsigned long lastinfo;
 820	unsigned long lastreport;
 821	bool reported = false;
 822	int rtsi;
 823	struct task_struct *t;
 824
 825	set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP);
 826	rtp->pregp_func(&holdouts);
 827
 828	/*
 829	 * There were callbacks, so we need to wait for an RCU-tasks
 830	 * grace period.  Start off by scanning the task list for tasks
 831	 * that are not already voluntarily blocked.  Mark these tasks
 832	 * and make a list of them in holdouts.
 833	 */
 834	set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST);
 835	if (rtp->pertask_func) {
 836		rcu_read_lock();
 837		for_each_process_thread(g, t)
 838			rtp->pertask_func(t, &holdouts);
 839		rcu_read_unlock();
 840	}
 841
 842	set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST);
 843	rtp->postscan_func(&holdouts);
 844
 845	/*
 846	 * Each pass through the following loop scans the list of holdout
 847	 * tasks, removing any that are no longer holdouts.  When the list
 848	 * is empty, we are done.
 849	 */
 850	lastreport = jiffies;
 851	lastinfo = lastreport;
 852	rtsi = READ_ONCE(rcu_task_stall_info);
 853
 854	// Start off with initial wait and slowly back off to 1 HZ wait.
 855	fract = rtp->init_fract;
 856
 857	while (!list_empty(&holdouts)) {
 858		ktime_t exp;
 859		bool firstreport;
 860		bool needreport;
 861		int rtst;
 862
 863		// Slowly back off waiting for holdouts
 864		set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS);
 865		if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
 866			schedule_timeout_idle(fract);
 867		} else {
 868			exp = jiffies_to_nsecs(fract);
 869			__set_current_state(TASK_IDLE);
 870			schedule_hrtimeout_range(&exp, jiffies_to_nsecs(HZ / 2), HRTIMER_MODE_REL_HARD);
 871		}
 872
 873		if (fract < HZ)
 874			fract++;
 875
 876		rtst = READ_ONCE(rcu_task_stall_timeout);
 877		needreport = rtst > 0 && time_after(jiffies, lastreport + rtst);
 878		if (needreport) {
 879			lastreport = jiffies;
 880			reported = true;
 881		}
 882		firstreport = true;
 883		WARN_ON(signal_pending(current));
 884		set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS);
 885		rtp->holdouts_func(&holdouts, needreport, &firstreport);
 886
 887		// Print pre-stall informational messages if needed.
 888		j = jiffies;
 889		if (rtsi > 0 && !reported && time_after(j, lastinfo + rtsi)) {
 890			lastinfo = j;
 891			rtsi = rtsi * rcu_task_stall_info_mult;
 892			pr_info("%s: %s grace period number %lu (since boot) is %lu jiffies old.\n",
 893				__func__, rtp->kname, rtp->tasks_gp_seq, j - rtp->gp_start);
 894		}
 895	}
 896
 897	set_tasks_gp_state(rtp, RTGS_POST_GP);
 898	rtp->postgp_func(rtp);
 899}
 900
 901#endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */
 902
 903#ifdef CONFIG_TASKS_RCU
 904
 905////////////////////////////////////////////////////////////////////////
 906//
 907// Simple variant of RCU whose quiescent states are voluntary context
 908// switch, cond_resched_tasks_rcu_qs(), user-space execution, and idle.
 909// As such, grace periods can take one good long time.  There are no
 910// read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
 911// because this implementation is intended to get the system into a safe
 912// state for some of the manipulations involved in tracing and the like.
 913// Finally, this implementation does not support high call_rcu_tasks()
 914// rates from multiple CPUs.  If this is required, per-CPU callback lists
 915// will be needed.
 916//
 917// The implementation uses rcu_tasks_wait_gp(), which relies on function
 918// pointers in the rcu_tasks structure.  The rcu_spawn_tasks_kthread()
 919// function sets these function pointers up so that rcu_tasks_wait_gp()
 920// invokes these functions in this order:
 921//
 922// rcu_tasks_pregp_step():
 923//	Invokes synchronize_rcu() in order to wait for all in-flight
 924//	t->on_rq and t->nvcsw transitions to complete.	This works because
 925//	all such transitions are carried out with interrupts disabled.
 926// rcu_tasks_pertask(), invoked on every non-idle task:
 927//	For every runnable non-idle task other than the current one, use
 928//	get_task_struct() to pin down that task, snapshot that task's
 929//	number of voluntary context switches, and add that task to the
 930//	holdout list.
 931// rcu_tasks_postscan():
 932//	Gather per-CPU lists of tasks in do_exit() to ensure that all
 933//	tasks that were in the process of exiting (and which thus might
 934//	not know to synchronize with this RCU Tasks grace period) have
 935//	completed exiting.  The synchronize_rcu() in rcu_tasks_postgp()
 936//	will take care of any tasks stuck in the non-preemptible region
 937//	of do_exit() following its call to exit_tasks_rcu_finish().
 938// check_all_holdout_tasks(), repeatedly until holdout list is empty:
 939//	Scans the holdout list, attempting to identify a quiescent state
 940//	for each task on the list.  If there is a quiescent state, the
 941//	corresponding task is removed from the holdout list.
 942// rcu_tasks_postgp():
 943//	Invokes synchronize_rcu() in order to ensure that all prior
 944//	t->on_rq and t->nvcsw transitions are seen by all CPUs and tasks
 945//	to have happened before the end of this RCU Tasks grace period.
 946//	Again, this works because all such transitions are carried out
 947//	with interrupts disabled.
 948//
 949// For each exiting task, the exit_tasks_rcu_start() and
 950// exit_tasks_rcu_finish() functions add and remove, respectively, the
 951// current task to a per-CPU list of tasks that rcu_tasks_postscan() must
 952// wait on.  This is necessary because rcu_tasks_postscan() must wait on
 953// tasks that have already been removed from the global list of tasks.
 954//
 955// Pre-grace-period update-side code is ordered before the grace
 956// via the raw_spin_lock.*rcu_node().  Pre-grace-period read-side code
 957// is ordered before the grace period via synchronize_rcu() call in
 958// rcu_tasks_pregp_step() and by the scheduler's locks and interrupt
 959// disabling.
 960
 961/* Pre-grace-period preparation. */
 962static void rcu_tasks_pregp_step(struct list_head *hop)
 963{
 964	/*
 965	 * Wait for all pre-existing t->on_rq and t->nvcsw transitions
 966	 * to complete.  Invoking synchronize_rcu() suffices because all
 967	 * these transitions occur with interrupts disabled.  Without this
 968	 * synchronize_rcu(), a read-side critical section that started
 969	 * before the grace period might be incorrectly seen as having
 970	 * started after the grace period.
 971	 *
 972	 * This synchronize_rcu() also dispenses with the need for a
 973	 * memory barrier on the first store to t->rcu_tasks_holdout,
 974	 * as it forces the store to happen after the beginning of the
 975	 * grace period.
 976	 */
 977	synchronize_rcu();
 978}
 979
 980/* Check for quiescent states since the pregp's synchronize_rcu() */
 981static bool rcu_tasks_is_holdout(struct task_struct *t)
 982{
 983	int cpu;
 984
 985	/* Has the task been seen voluntarily sleeping? */
 986	if (!READ_ONCE(t->on_rq))
 987		return false;
 988
 989	/*
 990	 * t->on_rq && !t->se.sched_delayed *could* be considered sleeping but
 991	 * since it is a spurious state (it will transition into the
 992	 * traditional blocked state or get woken up without outside
 993	 * dependencies), not considering it such should only affect timing.
 994	 *
 995	 * Be conservative for now and not include it.
 996	 */
 997
 998	/*
 999	 * Idle tasks (or idle injection) within the idle loop are RCU-tasks
1000	 * quiescent states. But CPU boot code performed by the idle task
1001	 * isn't a quiescent state.
1002	 */
1003	if (is_idle_task(t))
1004		return false;
1005
1006	cpu = task_cpu(t);
1007
1008	/* Idle tasks on offline CPUs are RCU-tasks quiescent states. */
1009	if (t == idle_task(cpu) && !rcu_cpu_online(cpu))
1010		return false;
1011
1012	return true;
1013}
1014
1015/* Per-task initial processing. */
1016static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop)
1017{
1018	if (t != current && rcu_tasks_is_holdout(t)) {
1019		get_task_struct(t);
1020		t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
1021		WRITE_ONCE(t->rcu_tasks_holdout, true);
1022		list_add(&t->rcu_tasks_holdout_list, hop);
1023	}
1024}
1025
1026void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
1027DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
1028
1029/* Processing between scanning taskslist and draining the holdout list. */
1030static void rcu_tasks_postscan(struct list_head *hop)
1031{
1032	int cpu;
1033	int rtsi = READ_ONCE(rcu_task_stall_info);
1034
1035	if (!IS_ENABLED(CONFIG_TINY_RCU)) {
1036		tasks_rcu_exit_srcu_stall_timer.expires = jiffies + rtsi;
1037		add_timer(&tasks_rcu_exit_srcu_stall_timer);
1038	}
1039
1040	/*
1041	 * Exiting tasks may escape the tasklist scan. Those are vulnerable
1042	 * until their final schedule() with TASK_DEAD state. To cope with
1043	 * this, divide the fragile exit path part in two intersecting
1044	 * read side critical sections:
1045	 *
1046	 * 1) A task_struct list addition before calling exit_notify(),
1047	 *    which may remove the task from the tasklist, with the
1048	 *    removal after the final preempt_disable() call in do_exit().
1049	 *
1050	 * 2) An _RCU_ read side starting with the final preempt_disable()
1051	 *    call in do_exit() and ending with the final call to schedule()
1052	 *    with TASK_DEAD state.
1053	 *
1054	 * This handles the part 1). And postgp will handle part 2) with a
1055	 * call to synchronize_rcu().
1056	 */
1057
1058	for_each_possible_cpu(cpu) {
1059		unsigned long j = jiffies + 1;
1060		struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rcu_tasks.rtpcpu, cpu);
1061		struct task_struct *t;
1062		struct task_struct *t1;
1063		struct list_head tmp;
1064
1065		raw_spin_lock_irq_rcu_node(rtpcp);
1066		list_for_each_entry_safe(t, t1, &rtpcp->rtp_exit_list, rcu_tasks_exit_list) {
1067			if (list_empty(&t->rcu_tasks_holdout_list))
1068				rcu_tasks_pertask(t, hop);
1069
1070			// RT kernels need frequent pauses, otherwise
1071			// pause at least once per pair of jiffies.
1072			if (!IS_ENABLED(CONFIG_PREEMPT_RT) && time_before(jiffies, j))
1073				continue;
1074
1075			// Keep our place in the list while pausing.
1076			// Nothing else traverses this list, so adding a
1077			// bare list_head is OK.
1078			list_add(&tmp, &t->rcu_tasks_exit_list);
1079			raw_spin_unlock_irq_rcu_node(rtpcp);
1080			cond_resched(); // For CONFIG_PREEMPT=n kernels
1081			raw_spin_lock_irq_rcu_node(rtpcp);
1082			t1 = list_entry(tmp.next, struct task_struct, rcu_tasks_exit_list);
1083			list_del(&tmp);
1084			j = jiffies + 1;
1085		}
1086		raw_spin_unlock_irq_rcu_node(rtpcp);
1087	}
1088
1089	if (!IS_ENABLED(CONFIG_TINY_RCU))
1090		timer_delete_sync(&tasks_rcu_exit_srcu_stall_timer);
1091}
1092
1093/* See if tasks are still holding out, complain if so. */
1094static void check_holdout_task(struct task_struct *t,
1095			       bool needreport, bool *firstreport)
1096{
1097	int cpu;
1098
1099	if (!READ_ONCE(t->rcu_tasks_holdout) ||
1100	    t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
1101	    !rcu_tasks_is_holdout(t) ||
1102	    (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
1103	     !is_idle_task(t) && READ_ONCE(t->rcu_tasks_idle_cpu) >= 0)) {
1104		WRITE_ONCE(t->rcu_tasks_holdout, false);
1105		list_del_init(&t->rcu_tasks_holdout_list);
1106		put_task_struct(t);
1107		return;
1108	}
1109	rcu_request_urgent_qs_task(t);
1110	if (!needreport)
1111		return;
1112	if (*firstreport) {
1113		pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
1114		*firstreport = false;
1115	}
1116	cpu = task_cpu(t);
1117	pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
1118		 t, ".I"[is_idle_task(t)],
1119		 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
1120		 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
1121		 data_race(t->rcu_tasks_idle_cpu), cpu);
1122	sched_show_task(t);
1123}
1124
1125/* Scan the holdout lists for tasks no longer holding out. */
1126static void check_all_holdout_tasks(struct list_head *hop,
1127				    bool needreport, bool *firstreport)
1128{
1129	struct task_struct *t, *t1;
1130
1131	list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) {
1132		check_holdout_task(t, needreport, firstreport);
1133		cond_resched();
1134	}
1135}
1136
1137/* Finish off the Tasks-RCU grace period. */
1138static void rcu_tasks_postgp(struct rcu_tasks *rtp)
1139{
1140	/*
1141	 * Because ->on_rq and ->nvcsw are not guaranteed to have a full
1142	 * memory barriers prior to them in the schedule() path, memory
1143	 * reordering on other CPUs could cause their RCU-tasks read-side
1144	 * critical sections to extend past the end of the grace period.
1145	 * However, because these ->nvcsw updates are carried out with
1146	 * interrupts disabled, we can use synchronize_rcu() to force the
1147	 * needed ordering on all such CPUs.
1148	 *
1149	 * This synchronize_rcu() also confines all ->rcu_tasks_holdout
1150	 * accesses to be within the grace period, avoiding the need for
1151	 * memory barriers for ->rcu_tasks_holdout accesses.
1152	 *
1153	 * In addition, this synchronize_rcu() waits for exiting tasks
1154	 * to complete their final preempt_disable() region of execution,
1155	 * enforcing the whole region before tasklist removal until
1156	 * the final schedule() with TASK_DEAD state to be an RCU TASKS
1157	 * read side critical section.
1158	 */
1159	synchronize_rcu();
1160}
1161
1162static void tasks_rcu_exit_srcu_stall(struct timer_list *unused)
1163{
1164#ifndef CONFIG_TINY_RCU
1165	int rtsi;
1166
1167	rtsi = READ_ONCE(rcu_task_stall_info);
1168	pr_info("%s: %s grace period number %lu (since boot) gp_state: %s is %lu jiffies old.\n",
1169		__func__, rcu_tasks.kname, rcu_tasks.tasks_gp_seq,
1170		tasks_gp_state_getname(&rcu_tasks), jiffies - rcu_tasks.gp_jiffies);
1171	pr_info("Please check any exiting tasks stuck between calls to exit_tasks_rcu_start() and exit_tasks_rcu_finish()\n");
1172	tasks_rcu_exit_srcu_stall_timer.expires = jiffies + rtsi;
1173	add_timer(&tasks_rcu_exit_srcu_stall_timer);
1174#endif // #ifndef CONFIG_TINY_RCU
1175}
1176
1177/**
1178 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
1179 * @rhp: structure to be used for queueing the RCU updates.
1180 * @func: actual callback function to be invoked after the grace period
1181 *
1182 * The callback function will be invoked some time after a full grace
1183 * period elapses, in other words after all currently executing RCU
1184 * read-side critical sections have completed. call_rcu_tasks() assumes
1185 * that the read-side critical sections end at a voluntary context
1186 * switch (not a preemption!), cond_resched_tasks_rcu_qs(), entry into idle,
1187 * or transition to usermode execution.  As such, there are no read-side
1188 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
1189 * this primitive is intended to determine that all tasks have passed
1190 * through a safe state, not so much for data-structure synchronization.
1191 *
1192 * See the description of call_rcu() for more detailed information on
1193 * memory ordering guarantees.
1194 */
1195void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
1196{
1197	call_rcu_tasks_generic(rhp, func, &rcu_tasks);
1198}
1199EXPORT_SYMBOL_GPL(call_rcu_tasks);
1200
1201/**
1202 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
1203 *
1204 * Control will return to the caller some time after a full rcu-tasks
1205 * grace period has elapsed, in other words after all currently
1206 * executing rcu-tasks read-side critical sections have elapsed.  These
1207 * read-side critical sections are delimited by calls to schedule(),
1208 * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
1209 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
1210 *
1211 * This is a very specialized primitive, intended only for a few uses in
1212 * tracing and other situations requiring manipulation of function
1213 * preambles and profiling hooks.  The synchronize_rcu_tasks() function
1214 * is not (yet) intended for heavy use from multiple CPUs.
1215 *
1216 * See the description of synchronize_rcu() for more detailed information
1217 * on memory ordering guarantees.
1218 */
1219void synchronize_rcu_tasks(void)
1220{
1221	synchronize_rcu_tasks_generic(&rcu_tasks);
1222}
1223EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
1224
1225/**
1226 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
1227 *
1228 * Although the current implementation is guaranteed to wait, it is not
1229 * obligated to, for example, if there are no pending callbacks.
1230 */
1231void rcu_barrier_tasks(void)
1232{
1233	rcu_barrier_tasks_generic(&rcu_tasks);
1234}
1235EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
1236
1237static int rcu_tasks_lazy_ms = -1;
1238module_param(rcu_tasks_lazy_ms, int, 0444);
1239
1240static int __init rcu_spawn_tasks_kthread(void)
1241{
1242	rcu_tasks.gp_sleep = HZ / 10;
1243	rcu_tasks.init_fract = HZ / 10;
1244	if (rcu_tasks_lazy_ms >= 0)
1245		rcu_tasks.lazy_jiffies = msecs_to_jiffies(rcu_tasks_lazy_ms);
1246	rcu_tasks.pregp_func = rcu_tasks_pregp_step;
1247	rcu_tasks.pertask_func = rcu_tasks_pertask;
1248	rcu_tasks.postscan_func = rcu_tasks_postscan;
1249	rcu_tasks.holdouts_func = check_all_holdout_tasks;
1250	rcu_tasks.postgp_func = rcu_tasks_postgp;
1251	rcu_tasks.wait_state = TASK_IDLE;
1252	rcu_spawn_tasks_kthread_generic(&rcu_tasks);
1253	return 0;
1254}
1255
1256#if !defined(CONFIG_TINY_RCU)
1257void show_rcu_tasks_classic_gp_kthread(void)
1258{
1259	show_rcu_tasks_generic_gp_kthread(&rcu_tasks, "");
1260}
1261EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread);
1262
1263void rcu_tasks_torture_stats_print(char *tt, char *tf)
1264{
1265	rcu_tasks_torture_stats_print_generic(&rcu_tasks, tt, tf, "");
1266}
1267EXPORT_SYMBOL_GPL(rcu_tasks_torture_stats_print);
1268#endif // !defined(CONFIG_TINY_RCU)
1269
1270struct task_struct *get_rcu_tasks_gp_kthread(void)
1271{
1272	return rcu_tasks.kthread_ptr;
1273}
1274EXPORT_SYMBOL_GPL(get_rcu_tasks_gp_kthread);
1275
1276void rcu_tasks_get_gp_data(int *flags, unsigned long *gp_seq)
1277{
1278	*flags = 0;
1279	*gp_seq = rcu_seq_current(&rcu_tasks.tasks_gp_seq);
1280}
1281EXPORT_SYMBOL_GPL(rcu_tasks_get_gp_data);
1282
1283/*
1284 * Protect against tasklist scan blind spot while the task is exiting and
1285 * may be removed from the tasklist.  Do this by adding the task to yet
1286 * another list.
1287 *
1288 * Note that the task will remove itself from this list, so there is no
1289 * need for get_task_struct(), except in the case where rcu_tasks_pertask()
1290 * adds it to the holdout list, in which case rcu_tasks_pertask() supplies
1291 * the needed get_task_struct().
1292 */
1293void exit_tasks_rcu_start(void)
1294{
1295	unsigned long flags;
1296	struct rcu_tasks_percpu *rtpcp;
1297	struct task_struct *t = current;
1298
1299	WARN_ON_ONCE(!list_empty(&t->rcu_tasks_exit_list));
1300	preempt_disable();
1301	rtpcp = this_cpu_ptr(rcu_tasks.rtpcpu);
1302	t->rcu_tasks_exit_cpu = smp_processor_id();
1303	raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1304	WARN_ON_ONCE(!rtpcp->rtp_exit_list.next);
1305	list_add(&t->rcu_tasks_exit_list, &rtpcp->rtp_exit_list);
1306	raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1307	preempt_enable();
1308}
1309
1310/*
1311 * Remove the task from the "yet another list" because do_exit() is now
1312 * non-preemptible, allowing synchronize_rcu() to wait beyond this point.
1313 */
1314void exit_tasks_rcu_finish(void)
1315{
1316	unsigned long flags;
1317	struct rcu_tasks_percpu *rtpcp;
1318	struct task_struct *t = current;
1319
1320	WARN_ON_ONCE(list_empty(&t->rcu_tasks_exit_list));
1321	rtpcp = per_cpu_ptr(rcu_tasks.rtpcpu, t->rcu_tasks_exit_cpu);
1322	raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1323	list_del_init(&t->rcu_tasks_exit_list);
1324	raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1325
1326	exit_tasks_rcu_finish_trace(t);
1327}
1328
1329#else /* #ifdef CONFIG_TASKS_RCU */
1330void exit_tasks_rcu_start(void) { }
1331void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); }
1332#endif /* #else #ifdef CONFIG_TASKS_RCU */
1333
1334#ifdef CONFIG_TASKS_RUDE_RCU
1335
1336////////////////////////////////////////////////////////////////////////
1337//
1338// "Rude" variant of Tasks RCU, inspired by Steve Rostedt's
1339// trick of passing an empty function to schedule_on_each_cpu().
1340// This approach provides batching of concurrent calls to the synchronous
1341// synchronize_rcu_tasks_rude() API.  This invokes schedule_on_each_cpu()
1342// in order to send IPIs far and wide and induces otherwise unnecessary
1343// context switches on all online CPUs, whether idle or not.
1344//
1345// Callback handling is provided by the rcu_tasks_kthread() function.
1346//
1347// Ordering is provided by the scheduler's context-switch code.
1348
1349// Empty function to allow workqueues to force a context switch.
1350static void rcu_tasks_be_rude(struct work_struct *work)
1351{
1352}
1353
1354// Wait for one rude RCU-tasks grace period.
1355static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
1356{
1357	rtp->n_ipis += cpumask_weight(cpu_online_mask);
1358	schedule_on_each_cpu(rcu_tasks_be_rude);
1359}
1360
1361static void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
1362DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude,
1363		 "RCU Tasks Rude");
1364
1365/*
1366 * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
1367 * @rhp: structure to be used for queueing the RCU updates.
1368 * @func: actual callback function to be invoked after the grace period
1369 *
1370 * The callback function will be invoked some time after a full grace
1371 * period elapses, in other words after all currently executing RCU
1372 * read-side critical sections have completed. call_rcu_tasks_rude()
1373 * assumes that the read-side critical sections end at context switch,
1374 * cond_resched_tasks_rcu_qs(), or transition to usermode execution (as
1375 * usermode execution is schedulable). As such, there are no read-side
1376 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
1377 * this primitive is intended to determine that all tasks have passed
1378 * through a safe state, not so much for data-structure synchronization.
1379 *
1380 * See the description of call_rcu() for more detailed information on
1381 * memory ordering guarantees.
1382 *
1383 * This is no longer exported, and is instead reserved for use by
1384 * synchronize_rcu_tasks_rude().
1385 */
1386static void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
1387{
1388	call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude);
1389}
1390
1391/**
1392 * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
1393 *
1394 * Control will return to the caller some time after a rude rcu-tasks
1395 * grace period has elapsed, in other words after all currently
1396 * executing rcu-tasks read-side critical sections have elapsed.  These
1397 * read-side critical sections are delimited by calls to schedule(),
1398 * cond_resched_tasks_rcu_qs(), userspace execution (which is a schedulable
1399 * context), and (in theory, anyway) cond_resched().
1400 *
1401 * This is a very specialized primitive, intended only for a few uses in
1402 * tracing and other situations requiring manipulation of function preambles
1403 * and profiling hooks.  The synchronize_rcu_tasks_rude() function is not
1404 * (yet) intended for heavy use from multiple CPUs.
1405 *
1406 * See the description of synchronize_rcu() for more detailed information
1407 * on memory ordering guarantees.
1408 */
1409void synchronize_rcu_tasks_rude(void)
1410{
1411	if (!IS_ENABLED(CONFIG_ARCH_WANTS_NO_INSTR) || IS_ENABLED(CONFIG_FORCE_TASKS_RUDE_RCU))
1412		synchronize_rcu_tasks_generic(&rcu_tasks_rude);
1413}
1414EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
1415
1416static int __init rcu_spawn_tasks_rude_kthread(void)
1417{
1418	rcu_tasks_rude.gp_sleep = HZ / 10;
1419	rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude);
1420	return 0;
1421}
1422
1423#if !defined(CONFIG_TINY_RCU)
1424void show_rcu_tasks_rude_gp_kthread(void)
1425{
1426	show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, "");
1427}
1428EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread);
1429
1430void rcu_tasks_rude_torture_stats_print(char *tt, char *tf)
1431{
1432	rcu_tasks_torture_stats_print_generic(&rcu_tasks_rude, tt, tf, "");
1433}
1434EXPORT_SYMBOL_GPL(rcu_tasks_rude_torture_stats_print);
1435#endif // !defined(CONFIG_TINY_RCU)
1436
1437struct task_struct *get_rcu_tasks_rude_gp_kthread(void)
1438{
1439	return rcu_tasks_rude.kthread_ptr;
1440}
1441EXPORT_SYMBOL_GPL(get_rcu_tasks_rude_gp_kthread);
1442
1443void rcu_tasks_rude_get_gp_data(int *flags, unsigned long *gp_seq)
1444{
1445	*flags = 0;
1446	*gp_seq = rcu_seq_current(&rcu_tasks_rude.tasks_gp_seq);
1447}
1448EXPORT_SYMBOL_GPL(rcu_tasks_rude_get_gp_data);
1449
1450#endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
1451
1452////////////////////////////////////////////////////////////////////////
1453//
1454// Tracing variant of Tasks RCU.  This variant is designed to be used
1455// to protect tracing hooks, including those of BPF.  This variant
1456// therefore:
1457//
1458// 1.	Has explicit read-side markers to allow finite grace periods
1459//	in the face of in-kernel loops for PREEMPT=n builds.
1460//
1461// 2.	Protects code in the idle loop, exception entry/exit, and
1462//	CPU-hotplug code paths, similar to the capabilities of SRCU.
1463//
1464// 3.	Avoids expensive read-side instructions, having overhead similar
1465//	to that of Preemptible RCU.
1466//
1467// There are of course downsides.  For example, the grace-period code
1468// can send IPIs to CPUs, even when those CPUs are in the idle loop or
1469// in nohz_full userspace.  If needed, these downsides can be at least
1470// partially remedied.
1471//
1472// Perhaps most important, this variant of RCU does not affect the vanilla
1473// flavors, rcu_preempt and rcu_sched.  The fact that RCU Tasks Trace
1474// readers can operate from idle, offline, and exception entry/exit in no
1475// way allows rcu_preempt and rcu_sched readers to also do so.
1476//
1477// The implementation uses rcu_tasks_wait_gp(), which relies on function
1478// pointers in the rcu_tasks structure.  The rcu_spawn_tasks_trace_kthread()
1479// function sets these function pointers up so that rcu_tasks_wait_gp()
1480// invokes these functions in this order:
1481//
1482// rcu_tasks_trace_pregp_step():
1483//	Disables CPU hotplug, adds all currently executing tasks to the
1484//	holdout list, then checks the state of all tasks that blocked
1485//	or were preempted within their current RCU Tasks Trace read-side
1486//	critical section, adding them to the holdout list if appropriate.
1487//	Finally, this function re-enables CPU hotplug.
1488// The ->pertask_func() pointer is NULL, so there is no per-task processing.
1489// rcu_tasks_trace_postscan():
1490//	Invokes synchronize_rcu() to wait for late-stage exiting tasks
1491//	to finish exiting.
1492// check_all_holdout_tasks_trace(), repeatedly until holdout list is empty:
1493//	Scans the holdout list, attempting to identify a quiescent state
1494//	for each task on the list.  If there is a quiescent state, the
1495//	corresponding task is removed from the holdout list.  Once this
1496//	list is empty, the grace period has completed.
1497// rcu_tasks_trace_postgp():
1498//	Provides the needed full memory barrier and does debug checks.
1499//
1500// The exit_tasks_rcu_finish_trace() synchronizes with exiting tasks.
1501//
1502// Pre-grace-period update-side code is ordered before the grace period
1503// via the ->cbs_lock and barriers in rcu_tasks_kthread().  Pre-grace-period
1504// read-side code is ordered before the grace period by atomic operations
1505// on .b.need_qs flag of each task involved in this process, or by scheduler
1506// context-switch ordering (for locked-down non-running readers).
1507
1508// The lockdep state must be outside of #ifdef to be useful.
1509#ifdef CONFIG_DEBUG_LOCK_ALLOC
1510static struct lock_class_key rcu_lock_trace_key;
1511struct lockdep_map rcu_trace_lock_map =
1512	STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key);
1513EXPORT_SYMBOL_GPL(rcu_trace_lock_map);
1514#endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
1515
1516#ifdef CONFIG_TASKS_TRACE_RCU
1517
1518// Record outstanding IPIs to each CPU.  No point in sending two...
1519static DEFINE_PER_CPU(bool, trc_ipi_to_cpu);
1520
1521// The number of detections of task quiescent state relying on
1522// heavyweight readers executing explicit memory barriers.
1523static unsigned long n_heavy_reader_attempts;
1524static unsigned long n_heavy_reader_updates;
1525static unsigned long n_heavy_reader_ofl_updates;
1526static unsigned long n_trc_holdouts;
1527
1528void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func);
1529DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace,
1530		 "RCU Tasks Trace");
1531
1532/* Load from ->trc_reader_special.b.need_qs with proper ordering. */
1533static u8 rcu_ld_need_qs(struct task_struct *t)
1534{
1535	smp_mb(); // Enforce full grace-period ordering.
1536	return smp_load_acquire(&t->trc_reader_special.b.need_qs);
1537}
1538
1539/* Store to ->trc_reader_special.b.need_qs with proper ordering. */
1540static void rcu_st_need_qs(struct task_struct *t, u8 v)
1541{
1542	smp_store_release(&t->trc_reader_special.b.need_qs, v);
1543	smp_mb(); // Enforce full grace-period ordering.
1544}
1545
1546/*
1547 * Do a cmpxchg() on ->trc_reader_special.b.need_qs, allowing for
1548 * the four-byte operand-size restriction of some platforms.
1549 *
1550 * Returns the old value, which is often ignored.
1551 */
1552u8 rcu_trc_cmpxchg_need_qs(struct task_struct *t, u8 old, u8 new)
1553{
1554	return cmpxchg(&t->trc_reader_special.b.need_qs, old, new);
1555}
1556EXPORT_SYMBOL_GPL(rcu_trc_cmpxchg_need_qs);
1557
1558/*
1559 * If we are the last reader, signal the grace-period kthread.
1560 * Also remove from the per-CPU list of blocked tasks.
1561 */
1562void rcu_read_unlock_trace_special(struct task_struct *t)
1563{
1564	unsigned long flags;
1565	struct rcu_tasks_percpu *rtpcp;
1566	union rcu_special trs;
1567
1568	// Open-coded full-word version of rcu_ld_need_qs().
1569	smp_mb(); // Enforce full grace-period ordering.
1570	trs = smp_load_acquire(&t->trc_reader_special);
1571
1572	if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && t->trc_reader_special.b.need_mb)
1573		smp_mb(); // Pairs with update-side barriers.
1574	// Update .need_qs before ->trc_reader_nesting for irq/NMI handlers.
1575	if (trs.b.need_qs == (TRC_NEED_QS_CHECKED | TRC_NEED_QS)) {
1576		u8 result = rcu_trc_cmpxchg_need_qs(t, TRC_NEED_QS_CHECKED | TRC_NEED_QS,
1577						       TRC_NEED_QS_CHECKED);
1578
1579		WARN_ONCE(result != trs.b.need_qs, "%s: result = %d", __func__, result);
1580	}
1581	if (trs.b.blocked) {
1582		rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, t->trc_blkd_cpu);
1583		raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1584		list_del_init(&t->trc_blkd_node);
1585		WRITE_ONCE(t->trc_reader_special.b.blocked, false);
1586		raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1587	}
1588	WRITE_ONCE(t->trc_reader_nesting, 0);
1589}
1590EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special);
1591
1592/* Add a newly blocked reader task to its CPU's list. */
1593void rcu_tasks_trace_qs_blkd(struct task_struct *t)
1594{
1595	unsigned long flags;
1596	struct rcu_tasks_percpu *rtpcp;
1597
1598	local_irq_save(flags);
1599	rtpcp = this_cpu_ptr(rcu_tasks_trace.rtpcpu);
1600	raw_spin_lock_rcu_node(rtpcp); // irqs already disabled
1601	t->trc_blkd_cpu = smp_processor_id();
1602	if (!rtpcp->rtp_blkd_tasks.next)
1603		INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
1604	list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1605	WRITE_ONCE(t->trc_reader_special.b.blocked, true);
1606	raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1607}
1608EXPORT_SYMBOL_GPL(rcu_tasks_trace_qs_blkd);
1609
1610/* Add a task to the holdout list, if it is not already on the list. */
1611static void trc_add_holdout(struct task_struct *t, struct list_head *bhp)
1612{
1613	if (list_empty(&t->trc_holdout_list)) {
1614		get_task_struct(t);
1615		list_add(&t->trc_holdout_list, bhp);
1616		n_trc_holdouts++;
1617	}
1618}
1619
1620/* Remove a task from the holdout list, if it is in fact present. */
1621static void trc_del_holdout(struct task_struct *t)
1622{
1623	if (!list_empty(&t->trc_holdout_list)) {
1624		list_del_init(&t->trc_holdout_list);
1625		put_task_struct(t);
1626		n_trc_holdouts--;
1627	}
1628}
1629
1630/* IPI handler to check task state. */
1631static void trc_read_check_handler(void *t_in)
1632{
1633	int nesting;
1634	struct task_struct *t = current;
1635	struct task_struct *texp = t_in;
1636
1637	// If the task is no longer running on this CPU, leave.
1638	if (unlikely(texp != t))
1639		goto reset_ipi; // Already on holdout list, so will check later.
1640
1641	// If the task is not in a read-side critical section, and
1642	// if this is the last reader, awaken the grace-period kthread.
1643	nesting = READ_ONCE(t->trc_reader_nesting);
1644	if (likely(!nesting)) {
1645		rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1646		goto reset_ipi;
1647	}
1648	// If we are racing with an rcu_read_unlock_trace(), try again later.
1649	if (unlikely(nesting < 0))
1650		goto reset_ipi;
1651
1652	// Get here if the task is in a read-side critical section.
1653	// Set its state so that it will update state for the grace-period
1654	// kthread upon exit from that critical section.
1655	rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED);
1656
1657reset_ipi:
1658	// Allow future IPIs to be sent on CPU and for task.
1659	// Also order this IPI handler against any later manipulations of
1660	// the intended task.
1661	smp_store_release(per_cpu_ptr(&trc_ipi_to_cpu, smp_processor_id()), false); // ^^^
1662	smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^
1663}
1664
1665/* Callback function for scheduler to check locked-down task.  */
1666static int trc_inspect_reader(struct task_struct *t, void *bhp_in)
1667{
1668	struct list_head *bhp = bhp_in;
1669	int cpu = task_cpu(t);
1670	int nesting;
1671	bool ofl = cpu_is_offline(cpu);
1672
1673	if (task_curr(t) && !ofl) {
1674		// If no chance of heavyweight readers, do it the hard way.
1675		if (!IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
1676			return -EINVAL;
1677
1678		// If heavyweight readers are enabled on the remote task,
1679		// we can inspect its state despite its currently running.
1680		// However, we cannot safely change its state.
1681		n_heavy_reader_attempts++;
1682		// Check for "running" idle tasks on offline CPUs.
1683		if (!rcu_watching_zero_in_eqs(cpu, &t->trc_reader_nesting))
1684			return -EINVAL; // No quiescent state, do it the hard way.
1685		n_heavy_reader_updates++;
1686		nesting = 0;
1687	} else {
1688		// The task is not running, so C-language access is safe.
1689		nesting = t->trc_reader_nesting;
1690		WARN_ON_ONCE(ofl && task_curr(t) && (t != idle_task(task_cpu(t))));
1691		if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && ofl)
1692			n_heavy_reader_ofl_updates++;
1693	}
1694
1695	// If not exiting a read-side critical section, mark as checked
1696	// so that the grace-period kthread will remove it from the
1697	// holdout list.
1698	if (!nesting) {
1699		rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1700		return 0;  // In QS, so done.
1701	}
1702	if (nesting < 0)
1703		return -EINVAL; // Reader transitioning, try again later.
1704
1705	// The task is in a read-side critical section, so set up its
1706	// state so that it will update state upon exit from that critical
1707	// section.
1708	if (!rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED))
1709		trc_add_holdout(t, bhp);
1710	return 0;
1711}
1712
1713/* Attempt to extract the state for the specified task. */
1714static void trc_wait_for_one_reader(struct task_struct *t,
1715				    struct list_head *bhp)
1716{
1717	int cpu;
1718
1719	// If a previous IPI is still in flight, let it complete.
1720	if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI
1721		return;
1722
1723	// The current task had better be in a quiescent state.
1724	if (t == current) {
1725		rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1726		WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1727		return;
1728	}
1729
1730	// Attempt to nail down the task for inspection.
1731	get_task_struct(t);
1732	if (!task_call_func(t, trc_inspect_reader, bhp)) {
1733		put_task_struct(t);
1734		return;
1735	}
1736	put_task_struct(t);
1737
1738	// If this task is not yet on the holdout list, then we are in
1739	// an RCU read-side critical section.  Otherwise, the invocation of
1740	// trc_add_holdout() that added it to the list did the necessary
1741	// get_task_struct().  Either way, the task cannot be freed out
1742	// from under this code.
1743
1744	// If currently running, send an IPI, either way, add to list.
1745	trc_add_holdout(t, bhp);
1746	if (task_curr(t) &&
1747	    time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) {
1748		// The task is currently running, so try IPIing it.
1749		cpu = task_cpu(t);
1750
1751		// If there is already an IPI outstanding, let it happen.
1752		if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0)
1753			return;
1754
1755		per_cpu(trc_ipi_to_cpu, cpu) = true;
1756		t->trc_ipi_to_cpu = cpu;
1757		rcu_tasks_trace.n_ipis++;
1758		if (smp_call_function_single(cpu, trc_read_check_handler, t, 0)) {
1759			// Just in case there is some other reason for
1760			// failure than the target CPU being offline.
1761			WARN_ONCE(1, "%s():  smp_call_function_single() failed for CPU: %d\n",
1762				  __func__, cpu);
1763			rcu_tasks_trace.n_ipis_fails++;
1764			per_cpu(trc_ipi_to_cpu, cpu) = false;
1765			t->trc_ipi_to_cpu = -1;
1766		}
1767	}
1768}
1769
1770/*
1771 * Initialize for first-round processing for the specified task.
1772 * Return false if task is NULL or already taken care of, true otherwise.
1773 */
1774static bool rcu_tasks_trace_pertask_prep(struct task_struct *t, bool notself)
1775{
1776	// During early boot when there is only the one boot CPU, there
1777	// is no idle task for the other CPUs.	Also, the grace-period
1778	// kthread is always in a quiescent state.  In addition, just return
1779	// if this task is already on the list.
1780	if (unlikely(t == NULL) || (t == current && notself) || !list_empty(&t->trc_holdout_list))
1781		return false;
1782
1783	rcu_st_need_qs(t, 0);
1784	t->trc_ipi_to_cpu = -1;
1785	return true;
1786}
1787
1788/* Do first-round processing for the specified task. */
1789static void rcu_tasks_trace_pertask(struct task_struct *t, struct list_head *hop)
1790{
1791	if (rcu_tasks_trace_pertask_prep(t, true))
1792		trc_wait_for_one_reader(t, hop);
1793}
1794
1795/* Initialize for a new RCU-tasks-trace grace period. */
1796static void rcu_tasks_trace_pregp_step(struct list_head *hop)
1797{
1798	LIST_HEAD(blkd_tasks);
1799	int cpu;
1800	unsigned long flags;
1801	struct rcu_tasks_percpu *rtpcp;
1802	struct task_struct *t;
1803
1804	// There shouldn't be any old IPIs, but...
1805	for_each_possible_cpu(cpu)
1806		WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu));
1807
1808	// Disable CPU hotplug across the CPU scan for the benefit of
1809	// any IPIs that might be needed.  This also waits for all readers
1810	// in CPU-hotplug code paths.
1811	cpus_read_lock();
1812
1813	// These rcu_tasks_trace_pertask_prep() calls are serialized to
1814	// allow safe access to the hop list.
1815	for_each_online_cpu(cpu) {
1816		rcu_read_lock();
1817		// Note that cpu_curr_snapshot() picks up the target
1818		// CPU's current task while its runqueue is locked with
1819		// an smp_mb__after_spinlock().  This ensures that either
1820		// the grace-period kthread will see that task's read-side
1821		// critical section or the task will see the updater's pre-GP
1822		// accesses.  The trailing smp_mb() in cpu_curr_snapshot()
1823		// does not currently play a role other than simplify
1824		// that function's ordering semantics.  If these simplified
1825		// ordering semantics continue to be redundant, that smp_mb()
1826		// might be removed.
1827		t = cpu_curr_snapshot(cpu);
1828		if (rcu_tasks_trace_pertask_prep(t, true))
1829			trc_add_holdout(t, hop);
1830		rcu_read_unlock();
1831		cond_resched_tasks_rcu_qs();
1832	}
1833
1834	// Only after all running tasks have been accounted for is it
1835	// safe to take care of the tasks that have blocked within their
1836	// current RCU tasks trace read-side critical section.
1837	for_each_possible_cpu(cpu) {
1838		rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, cpu);
1839		raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1840		list_splice_init(&rtpcp->rtp_blkd_tasks, &blkd_tasks);
1841		while (!list_empty(&blkd_tasks)) {
1842			rcu_read_lock();
1843			t = list_first_entry(&blkd_tasks, struct task_struct, trc_blkd_node);
1844			list_del_init(&t->trc_blkd_node);
1845			list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1846			raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1847			rcu_tasks_trace_pertask(t, hop);
1848			rcu_read_unlock();
1849			raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1850		}
1851		raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1852		cond_resched_tasks_rcu_qs();
1853	}
1854
1855	// Re-enable CPU hotplug now that the holdout list is populated.
1856	cpus_read_unlock();
1857}
1858
1859/*
1860 * Do intermediate processing between task and holdout scans.
1861 */
1862static void rcu_tasks_trace_postscan(struct list_head *hop)
1863{
1864	// Wait for late-stage exiting tasks to finish exiting.
1865	// These might have passed the call to exit_tasks_rcu_finish().
1866
1867	// If you remove the following line, update rcu_trace_implies_rcu_gp()!!!
1868	synchronize_rcu();
1869	// Any tasks that exit after this point will set
1870	// TRC_NEED_QS_CHECKED in ->trc_reader_special.b.need_qs.
1871}
1872
1873/* Communicate task state back to the RCU tasks trace stall warning request. */
1874struct trc_stall_chk_rdr {
1875	int nesting;
1876	int ipi_to_cpu;
1877	u8 needqs;
1878};
1879
1880static int trc_check_slow_task(struct task_struct *t, void *arg)
1881{
1882	struct trc_stall_chk_rdr *trc_rdrp = arg;
1883
1884	if (task_curr(t) && cpu_online(task_cpu(t)))
1885		return false; // It is running, so decline to inspect it.
1886	trc_rdrp->nesting = READ_ONCE(t->trc_reader_nesting);
1887	trc_rdrp->ipi_to_cpu = READ_ONCE(t->trc_ipi_to_cpu);
1888	trc_rdrp->needqs = rcu_ld_need_qs(t);
1889	return true;
1890}
1891
1892/* Show the state of a task stalling the current RCU tasks trace GP. */
1893static void show_stalled_task_trace(struct task_struct *t, bool *firstreport)
1894{
1895	int cpu;
1896	struct trc_stall_chk_rdr trc_rdr;
1897	bool is_idle_tsk = is_idle_task(t);
1898
1899	if (*firstreport) {
1900		pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n");
1901		*firstreport = false;
1902	}
1903	cpu = task_cpu(t);
1904	if (!task_call_func(t, trc_check_slow_task, &trc_rdr))
1905		pr_alert("P%d: %c%c\n",
1906			 t->pid,
1907			 ".I"[t->trc_ipi_to_cpu >= 0],
1908			 ".i"[is_idle_tsk]);
1909	else
1910		pr_alert("P%d: %c%c%c%c nesting: %d%c%c cpu: %d%s\n",
1911			 t->pid,
1912			 ".I"[trc_rdr.ipi_to_cpu >= 0],
1913			 ".i"[is_idle_tsk],
1914			 ".N"[cpu >= 0 && tick_nohz_full_cpu(cpu)],
1915			 ".B"[!!data_race(t->trc_reader_special.b.blocked)],
1916			 trc_rdr.nesting,
1917			 " !CN"[trc_rdr.needqs & 0x3],
1918			 " ?"[trc_rdr.needqs > 0x3],
1919			 cpu, cpu_online(cpu) ? "" : "(offline)");
1920	sched_show_task(t);
1921}
1922
1923/* List stalled IPIs for RCU tasks trace. */
1924static void show_stalled_ipi_trace(void)
1925{
1926	int cpu;
1927
1928	for_each_possible_cpu(cpu)
1929		if (per_cpu(trc_ipi_to_cpu, cpu))
1930			pr_alert("\tIPI outstanding to CPU %d\n", cpu);
1931}
1932
1933/* Do one scan of the holdout list. */
1934static void check_all_holdout_tasks_trace(struct list_head *hop,
1935					  bool needreport, bool *firstreport)
1936{
1937	struct task_struct *g, *t;
1938
1939	// Disable CPU hotplug across the holdout list scan for IPIs.
1940	cpus_read_lock();
1941
1942	list_for_each_entry_safe(t, g, hop, trc_holdout_list) {
1943		// If safe and needed, try to check the current task.
1944		if (READ_ONCE(t->trc_ipi_to_cpu) == -1 &&
1945		    !(rcu_ld_need_qs(t) & TRC_NEED_QS_CHECKED))
1946			trc_wait_for_one_reader(t, hop);
1947
1948		// If check succeeded, remove this task from the list.
1949		if (smp_load_acquire(&t->trc_ipi_to_cpu) == -1 &&
1950		    rcu_ld_need_qs(t) == TRC_NEED_QS_CHECKED)
1951			trc_del_holdout(t);
1952		else if (needreport)
1953			show_stalled_task_trace(t, firstreport);
1954		cond_resched_tasks_rcu_qs();
1955	}
1956
1957	// Re-enable CPU hotplug now that the holdout list scan has completed.
1958	cpus_read_unlock();
1959
1960	if (needreport) {
1961		if (*firstreport)
1962			pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n");
1963		show_stalled_ipi_trace();
1964	}
1965}
1966
1967static void rcu_tasks_trace_empty_fn(void *unused)
1968{
1969}
1970
1971/* Wait for grace period to complete and provide ordering. */
1972static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp)
1973{
1974	int cpu;
1975
1976	// Wait for any lingering IPI handlers to complete.  Note that
1977	// if a CPU has gone offline or transitioned to userspace in the
1978	// meantime, all IPI handlers should have been drained beforehand.
1979	// Yes, this assumes that CPUs process IPIs in order.  If that ever
1980	// changes, there will need to be a recheck and/or timed wait.
1981	for_each_online_cpu(cpu)
1982		if (WARN_ON_ONCE(smp_load_acquire(per_cpu_ptr(&trc_ipi_to_cpu, cpu))))
1983			smp_call_function_single(cpu, rcu_tasks_trace_empty_fn, NULL, 1);
1984
1985	smp_mb(); // Caller's code must be ordered after wakeup.
1986		  // Pairs with pretty much every ordering primitive.
1987}
1988
1989/* Report any needed quiescent state for this exiting task. */
1990static void exit_tasks_rcu_finish_trace(struct task_struct *t)
1991{
1992	union rcu_special trs = READ_ONCE(t->trc_reader_special);
1993
1994	rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1995	WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1996	if (WARN_ON_ONCE(rcu_ld_need_qs(t) & TRC_NEED_QS || trs.b.blocked))
1997		rcu_read_unlock_trace_special(t);
1998	else
1999		WRITE_ONCE(t->trc_reader_nesting, 0);
2000}
2001
2002/**
2003 * call_rcu_tasks_trace() - Queue a callback trace task-based grace period
2004 * @rhp: structure to be used for queueing the RCU updates.
2005 * @func: actual callback function to be invoked after the grace period
2006 *
2007 * The callback function will be invoked some time after a trace rcu-tasks
2008 * grace period elapses, in other words after all currently executing
2009 * trace rcu-tasks read-side critical sections have completed. These
2010 * read-side critical sections are delimited by calls to rcu_read_lock_trace()
2011 * and rcu_read_unlock_trace().
2012 *
2013 * See the description of call_rcu() for more detailed information on
2014 * memory ordering guarantees.
2015 */
2016void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func)
2017{
2018	call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace);
2019}
2020EXPORT_SYMBOL_GPL(call_rcu_tasks_trace);
2021
2022/**
2023 * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period
2024 *
2025 * Control will return to the caller some time after a trace rcu-tasks
2026 * grace period has elapsed, in other words after all currently executing
2027 * trace rcu-tasks read-side critical sections have elapsed. These read-side
2028 * critical sections are delimited by calls to rcu_read_lock_trace()
2029 * and rcu_read_unlock_trace().
2030 *
2031 * This is a very specialized primitive, intended only for a few uses in
2032 * tracing and other situations requiring manipulation of function preambles
2033 * and profiling hooks.  The synchronize_rcu_tasks_trace() function is not
2034 * (yet) intended for heavy use from multiple CPUs.
2035 *
2036 * See the description of synchronize_rcu() for more detailed information
2037 * on memory ordering guarantees.
2038 */
2039void synchronize_rcu_tasks_trace(void)
2040{
2041	RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section");
2042	synchronize_rcu_tasks_generic(&rcu_tasks_trace);
2043}
2044EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace);
2045
2046/**
2047 * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks.
2048 *
2049 * Although the current implementation is guaranteed to wait, it is not
2050 * obligated to, for example, if there are no pending callbacks.
2051 */
2052void rcu_barrier_tasks_trace(void)
2053{
2054	rcu_barrier_tasks_generic(&rcu_tasks_trace);
2055}
2056EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace);
2057
2058int rcu_tasks_trace_lazy_ms = -1;
2059module_param(rcu_tasks_trace_lazy_ms, int, 0444);
2060
2061static int __init rcu_spawn_tasks_trace_kthread(void)
2062{
2063	if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) {
2064		rcu_tasks_trace.gp_sleep = HZ / 10;
2065		rcu_tasks_trace.init_fract = HZ / 10;
2066	} else {
2067		rcu_tasks_trace.gp_sleep = HZ / 200;
2068		if (rcu_tasks_trace.gp_sleep <= 0)
2069			rcu_tasks_trace.gp_sleep = 1;
2070		rcu_tasks_trace.init_fract = HZ / 200;
2071		if (rcu_tasks_trace.init_fract <= 0)
2072			rcu_tasks_trace.init_fract = 1;
2073	}
2074	if (rcu_tasks_trace_lazy_ms >= 0)
2075		rcu_tasks_trace.lazy_jiffies = msecs_to_jiffies(rcu_tasks_trace_lazy_ms);
2076	rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step;
2077	rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan;
2078	rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace;
2079	rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp;
2080	rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace);
2081	return 0;
2082}
2083
2084#if !defined(CONFIG_TINY_RCU)
2085void show_rcu_tasks_trace_gp_kthread(void)
2086{
2087	char buf[64];
2088
2089	snprintf(buf, sizeof(buf), "N%lu h:%lu/%lu/%lu",
2090		data_race(n_trc_holdouts),
2091		data_race(n_heavy_reader_ofl_updates),
2092		data_race(n_heavy_reader_updates),
2093		data_race(n_heavy_reader_attempts));
2094	show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf);
2095}
2096EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread);
2097
2098void rcu_tasks_trace_torture_stats_print(char *tt, char *tf)
2099{
2100	rcu_tasks_torture_stats_print_generic(&rcu_tasks_trace, tt, tf, "");
2101}
2102EXPORT_SYMBOL_GPL(rcu_tasks_trace_torture_stats_print);
2103#endif // !defined(CONFIG_TINY_RCU)
2104
2105struct task_struct *get_rcu_tasks_trace_gp_kthread(void)
2106{
2107	return rcu_tasks_trace.kthread_ptr;
2108}
2109EXPORT_SYMBOL_GPL(get_rcu_tasks_trace_gp_kthread);
2110
2111void rcu_tasks_trace_get_gp_data(int *flags, unsigned long *gp_seq)
2112{
2113	*flags = 0;
2114	*gp_seq = rcu_seq_current(&rcu_tasks_trace.tasks_gp_seq);
2115}
2116EXPORT_SYMBOL_GPL(rcu_tasks_trace_get_gp_data);
2117
2118#else /* #ifdef CONFIG_TASKS_TRACE_RCU */
2119static void exit_tasks_rcu_finish_trace(struct task_struct *t) { }
2120#endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */
2121
2122#ifndef CONFIG_TINY_RCU
2123void show_rcu_tasks_gp_kthreads(void)
2124{
2125	show_rcu_tasks_classic_gp_kthread();
2126	show_rcu_tasks_rude_gp_kthread();
2127	show_rcu_tasks_trace_gp_kthread();
2128}
2129#endif /* #ifndef CONFIG_TINY_RCU */
2130
2131#ifdef CONFIG_PROVE_RCU
2132struct rcu_tasks_test_desc {
2133	struct rcu_head rh;
2134	const char *name;
2135	bool notrun;
2136	unsigned long runstart;
2137};
2138
2139static struct rcu_tasks_test_desc tests[] = {
2140	{
2141		.name = "call_rcu_tasks()",
2142		/* If not defined, the test is skipped. */
2143		.notrun = IS_ENABLED(CONFIG_TASKS_RCU),
2144	},
2145	{
2146		.name = "call_rcu_tasks_trace()",
2147		/* If not defined, the test is skipped. */
2148		.notrun = IS_ENABLED(CONFIG_TASKS_TRACE_RCU)
2149	}
2150};
2151
2152#if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
2153static void test_rcu_tasks_callback(struct rcu_head *rhp)
2154{
2155	struct rcu_tasks_test_desc *rttd =
2156		container_of(rhp, struct rcu_tasks_test_desc, rh);
2157
2158	pr_info("Callback from %s invoked.\n", rttd->name);
2159
2160	rttd->notrun = false;
2161}
2162#endif // #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
2163
2164static void rcu_tasks_initiate_self_tests(void)
2165{
2166#ifdef CONFIG_TASKS_RCU
2167	pr_info("Running RCU Tasks wait API self tests\n");
2168	tests[0].runstart = jiffies;
2169	synchronize_rcu_tasks();
2170	call_rcu_tasks(&tests[0].rh, test_rcu_tasks_callback);
2171#endif
2172
2173#ifdef CONFIG_TASKS_RUDE_RCU
2174	pr_info("Running RCU Tasks Rude wait API self tests\n");
2175	synchronize_rcu_tasks_rude();
2176#endif
2177
2178#ifdef CONFIG_TASKS_TRACE_RCU
2179	pr_info("Running RCU Tasks Trace wait API self tests\n");
2180	tests[1].runstart = jiffies;
2181	synchronize_rcu_tasks_trace();
2182	call_rcu_tasks_trace(&tests[1].rh, test_rcu_tasks_callback);
2183#endif
2184}
2185
2186/*
2187 * Return:  0 - test passed
2188 *	    1 - test failed, but have not timed out yet
2189 *	   -1 - test failed and timed out
2190 */
2191static int rcu_tasks_verify_self_tests(void)
2192{
2193	int ret = 0;
2194	int i;
2195	unsigned long bst = rcu_task_stall_timeout;
2196
2197	if (bst <= 0 || bst > RCU_TASK_BOOT_STALL_TIMEOUT)
2198		bst = RCU_TASK_BOOT_STALL_TIMEOUT;
2199	for (i = 0; i < ARRAY_SIZE(tests); i++) {
2200		while (tests[i].notrun) {		// still hanging.
2201			if (time_after(jiffies, tests[i].runstart + bst)) {
2202				pr_err("%s has failed boot-time tests.\n", tests[i].name);
2203				ret = -1;
2204				break;
2205			}
2206			ret = 1;
2207			break;
2208		}
2209	}
2210	WARN_ON(ret < 0);
2211
2212	return ret;
2213}
2214
2215/*
2216 * Repeat the rcu_tasks_verify_self_tests() call once every second until the
2217 * test passes or has timed out.
2218 */
2219static struct delayed_work rcu_tasks_verify_work;
2220static void rcu_tasks_verify_work_fn(struct work_struct *work __maybe_unused)
2221{
2222	int ret = rcu_tasks_verify_self_tests();
2223
2224	if (ret <= 0)
2225		return;
2226
2227	/* Test fails but not timed out yet, reschedule another check */
2228	schedule_delayed_work(&rcu_tasks_verify_work, HZ);
2229}
2230
2231static int rcu_tasks_verify_schedule_work(void)
2232{
2233	INIT_DELAYED_WORK(&rcu_tasks_verify_work, rcu_tasks_verify_work_fn);
2234	rcu_tasks_verify_work_fn(NULL);
2235	return 0;
2236}
2237late_initcall(rcu_tasks_verify_schedule_work);
2238#else /* #ifdef CONFIG_PROVE_RCU */
2239static void rcu_tasks_initiate_self_tests(void) { }
2240#endif /* #else #ifdef CONFIG_PROVE_RCU */
2241
2242void __init tasks_cblist_init_generic(void)
2243{
2244	lockdep_assert_irqs_disabled();
2245	WARN_ON(num_online_cpus() > 1);
2246
2247#ifdef CONFIG_TASKS_RCU
2248	cblist_init_generic(&rcu_tasks);
2249#endif
2250
2251#ifdef CONFIG_TASKS_RUDE_RCU
2252	cblist_init_generic(&rcu_tasks_rude);
2253#endif
2254
2255#ifdef CONFIG_TASKS_TRACE_RCU
2256	cblist_init_generic(&rcu_tasks_trace);
2257#endif
2258}
2259
2260static int __init rcu_init_tasks_generic(void)
2261{
2262#ifdef CONFIG_TASKS_RCU
2263	rcu_spawn_tasks_kthread();
2264#endif
2265
2266#ifdef CONFIG_TASKS_RUDE_RCU
2267	rcu_spawn_tasks_rude_kthread();
2268#endif
2269
2270#ifdef CONFIG_TASKS_TRACE_RCU
2271	rcu_spawn_tasks_trace_kthread();
2272#endif
2273
2274	// Run the self-tests.
2275	rcu_tasks_initiate_self_tests();
2276
2277	return 0;
2278}
2279core_initcall(rcu_init_tasks_generic);
2280
2281#else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
2282static inline void rcu_tasks_bootup_oddness(void) {}
2283#endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */