Linux Audio

Check our new training course

Loading...
Note: File does not exist in v4.6.
   1// SPDX-License-Identifier: GPL-2.0+
   2/*
   3 * Sleepable Read-Copy Update mechanism for mutual exclusion.
   4 *
   5 * Copyright (C) IBM Corporation, 2006
   6 * Copyright (C) Fujitsu, 2012
   7 *
   8 * Authors: Paul McKenney <paulmck@linux.ibm.com>
   9 *	   Lai Jiangshan <laijs@cn.fujitsu.com>
  10 *
  11 * For detailed explanation of Read-Copy Update mechanism see -
  12 *		Documentation/RCU/ *.txt
  13 *
  14 */
  15
  16#define pr_fmt(fmt) "rcu: " fmt
  17
  18#include <linux/export.h>
  19#include <linux/mutex.h>
  20#include <linux/percpu.h>
  21#include <linux/preempt.h>
  22#include <linux/rcupdate_wait.h>
  23#include <linux/sched.h>
  24#include <linux/smp.h>
  25#include <linux/delay.h>
  26#include <linux/module.h>
  27#include <linux/slab.h>
  28#include <linux/srcu.h>
  29
  30#include "rcu.h"
  31#include "rcu_segcblist.h"
  32
  33/* Holdoff in nanoseconds for auto-expediting. */
  34#define DEFAULT_SRCU_EXP_HOLDOFF (25 * 1000)
  35static ulong exp_holdoff = DEFAULT_SRCU_EXP_HOLDOFF;
  36module_param(exp_holdoff, ulong, 0444);
  37
  38/* Overflow-check frequency.  N bits roughly says every 2**N grace periods. */
  39static ulong counter_wrap_check = (ULONG_MAX >> 2);
  40module_param(counter_wrap_check, ulong, 0444);
  41
  42/*
  43 * Control conversion to SRCU_SIZE_BIG:
  44 *    0: Don't convert at all.
  45 *    1: Convert at init_srcu_struct() time.
  46 *    2: Convert when rcutorture invokes srcu_torture_stats_print().
  47 *    3: Decide at boot time based on system shape (default).
  48 * 0x1x: Convert when excessive contention encountered.
  49 */
  50#define SRCU_SIZING_NONE	0
  51#define SRCU_SIZING_INIT	1
  52#define SRCU_SIZING_TORTURE	2
  53#define SRCU_SIZING_AUTO	3
  54#define SRCU_SIZING_CONTEND	0x10
  55#define SRCU_SIZING_IS(x) ((convert_to_big & ~SRCU_SIZING_CONTEND) == x)
  56#define SRCU_SIZING_IS_NONE() (SRCU_SIZING_IS(SRCU_SIZING_NONE))
  57#define SRCU_SIZING_IS_INIT() (SRCU_SIZING_IS(SRCU_SIZING_INIT))
  58#define SRCU_SIZING_IS_TORTURE() (SRCU_SIZING_IS(SRCU_SIZING_TORTURE))
  59#define SRCU_SIZING_IS_CONTEND() (convert_to_big & SRCU_SIZING_CONTEND)
  60static int convert_to_big = SRCU_SIZING_AUTO;
  61module_param(convert_to_big, int, 0444);
  62
  63/* Number of CPUs to trigger init_srcu_struct()-time transition to big. */
  64static int big_cpu_lim __read_mostly = 128;
  65module_param(big_cpu_lim, int, 0444);
  66
  67/* Contention events per jiffy to initiate transition to big. */
  68static int small_contention_lim __read_mostly = 100;
  69module_param(small_contention_lim, int, 0444);
  70
  71/* Early-boot callback-management, so early that no lock is required! */
  72static LIST_HEAD(srcu_boot_list);
  73static bool __read_mostly srcu_init_done;
  74
  75static void srcu_invoke_callbacks(struct work_struct *work);
  76static void srcu_reschedule(struct srcu_struct *ssp, unsigned long delay);
  77static void process_srcu(struct work_struct *work);
  78static void srcu_delay_timer(struct timer_list *t);
  79
  80/* Wrappers for lock acquisition and release, see raw_spin_lock_rcu_node(). */
  81#define spin_lock_rcu_node(p)							\
  82do {										\
  83	spin_lock(&ACCESS_PRIVATE(p, lock));					\
  84	smp_mb__after_unlock_lock();						\
  85} while (0)
  86
  87#define spin_unlock_rcu_node(p) spin_unlock(&ACCESS_PRIVATE(p, lock))
  88
  89#define spin_lock_irq_rcu_node(p)						\
  90do {										\
  91	spin_lock_irq(&ACCESS_PRIVATE(p, lock));				\
  92	smp_mb__after_unlock_lock();						\
  93} while (0)
  94
  95#define spin_unlock_irq_rcu_node(p)						\
  96	spin_unlock_irq(&ACCESS_PRIVATE(p, lock))
  97
  98#define spin_lock_irqsave_rcu_node(p, flags)					\
  99do {										\
 100	spin_lock_irqsave(&ACCESS_PRIVATE(p, lock), flags);			\
 101	smp_mb__after_unlock_lock();						\
 102} while (0)
 103
 104#define spin_trylock_irqsave_rcu_node(p, flags)					\
 105({										\
 106	bool ___locked = spin_trylock_irqsave(&ACCESS_PRIVATE(p, lock), flags); \
 107										\
 108	if (___locked)								\
 109		smp_mb__after_unlock_lock();					\
 110	___locked;								\
 111})
 112
 113#define spin_unlock_irqrestore_rcu_node(p, flags)				\
 114	spin_unlock_irqrestore(&ACCESS_PRIVATE(p, lock), flags)			\
 115
 116/*
 117 * Initialize SRCU per-CPU data.  Note that statically allocated
 118 * srcu_struct structures might already have srcu_read_lock() and
 119 * srcu_read_unlock() running against them.  So if the is_static parameter
 120 * is set, don't initialize ->srcu_lock_count[] and ->srcu_unlock_count[].
 121 */
 122static void init_srcu_struct_data(struct srcu_struct *ssp)
 123{
 124	int cpu;
 125	struct srcu_data *sdp;
 126
 127	/*
 128	 * Initialize the per-CPU srcu_data array, which feeds into the
 129	 * leaves of the srcu_node tree.
 130	 */
 131	WARN_ON_ONCE(ARRAY_SIZE(sdp->srcu_lock_count) !=
 132		     ARRAY_SIZE(sdp->srcu_unlock_count));
 133	for_each_possible_cpu(cpu) {
 134		sdp = per_cpu_ptr(ssp->sda, cpu);
 135		spin_lock_init(&ACCESS_PRIVATE(sdp, lock));
 136		rcu_segcblist_init(&sdp->srcu_cblist);
 137		sdp->srcu_cblist_invoking = false;
 138		sdp->srcu_gp_seq_needed = ssp->srcu_sup->srcu_gp_seq;
 139		sdp->srcu_gp_seq_needed_exp = ssp->srcu_sup->srcu_gp_seq;
 140		sdp->mynode = NULL;
 141		sdp->cpu = cpu;
 142		INIT_WORK(&sdp->work, srcu_invoke_callbacks);
 143		timer_setup(&sdp->delay_work, srcu_delay_timer, 0);
 144		sdp->ssp = ssp;
 145	}
 146}
 147
 148/* Invalid seq state, used during snp node initialization */
 149#define SRCU_SNP_INIT_SEQ		0x2
 150
 151/*
 152 * Check whether sequence number corresponding to snp node,
 153 * is invalid.
 154 */
 155static inline bool srcu_invl_snp_seq(unsigned long s)
 156{
 157	return s == SRCU_SNP_INIT_SEQ;
 158}
 159
 160/*
 161 * Allocated and initialize SRCU combining tree.  Returns @true if
 162 * allocation succeeded and @false otherwise.
 163 */
 164static bool init_srcu_struct_nodes(struct srcu_struct *ssp, gfp_t gfp_flags)
 165{
 166	int cpu;
 167	int i;
 168	int level = 0;
 169	int levelspread[RCU_NUM_LVLS];
 170	struct srcu_data *sdp;
 171	struct srcu_node *snp;
 172	struct srcu_node *snp_first;
 173
 174	/* Initialize geometry if it has not already been initialized. */
 175	rcu_init_geometry();
 176	ssp->srcu_sup->node = kcalloc(rcu_num_nodes, sizeof(*ssp->srcu_sup->node), gfp_flags);
 177	if (!ssp->srcu_sup->node)
 178		return false;
 179
 180	/* Work out the overall tree geometry. */
 181	ssp->srcu_sup->level[0] = &ssp->srcu_sup->node[0];
 182	for (i = 1; i < rcu_num_lvls; i++)
 183		ssp->srcu_sup->level[i] = ssp->srcu_sup->level[i - 1] + num_rcu_lvl[i - 1];
 184	rcu_init_levelspread(levelspread, num_rcu_lvl);
 185
 186	/* Each pass through this loop initializes one srcu_node structure. */
 187	srcu_for_each_node_breadth_first(ssp, snp) {
 188		spin_lock_init(&ACCESS_PRIVATE(snp, lock));
 189		WARN_ON_ONCE(ARRAY_SIZE(snp->srcu_have_cbs) !=
 190			     ARRAY_SIZE(snp->srcu_data_have_cbs));
 191		for (i = 0; i < ARRAY_SIZE(snp->srcu_have_cbs); i++) {
 192			snp->srcu_have_cbs[i] = SRCU_SNP_INIT_SEQ;
 193			snp->srcu_data_have_cbs[i] = 0;
 194		}
 195		snp->srcu_gp_seq_needed_exp = SRCU_SNP_INIT_SEQ;
 196		snp->grplo = -1;
 197		snp->grphi = -1;
 198		if (snp == &ssp->srcu_sup->node[0]) {
 199			/* Root node, special case. */
 200			snp->srcu_parent = NULL;
 201			continue;
 202		}
 203
 204		/* Non-root node. */
 205		if (snp == ssp->srcu_sup->level[level + 1])
 206			level++;
 207		snp->srcu_parent = ssp->srcu_sup->level[level - 1] +
 208				   (snp - ssp->srcu_sup->level[level]) /
 209				   levelspread[level - 1];
 210	}
 211
 212	/*
 213	 * Initialize the per-CPU srcu_data array, which feeds into the
 214	 * leaves of the srcu_node tree.
 215	 */
 216	level = rcu_num_lvls - 1;
 217	snp_first = ssp->srcu_sup->level[level];
 218	for_each_possible_cpu(cpu) {
 219		sdp = per_cpu_ptr(ssp->sda, cpu);
 220		sdp->mynode = &snp_first[cpu / levelspread[level]];
 221		for (snp = sdp->mynode; snp != NULL; snp = snp->srcu_parent) {
 222			if (snp->grplo < 0)
 223				snp->grplo = cpu;
 224			snp->grphi = cpu;
 225		}
 226		sdp->grpmask = 1UL << (cpu - sdp->mynode->grplo);
 227	}
 228	smp_store_release(&ssp->srcu_sup->srcu_size_state, SRCU_SIZE_WAIT_BARRIER);
 229	return true;
 230}
 231
 232/*
 233 * Initialize non-compile-time initialized fields, including the
 234 * associated srcu_node and srcu_data structures.  The is_static parameter
 235 * tells us that ->sda has already been wired up to srcu_data.
 236 */
 237static int init_srcu_struct_fields(struct srcu_struct *ssp, bool is_static)
 238{
 239	if (!is_static)
 240		ssp->srcu_sup = kzalloc(sizeof(*ssp->srcu_sup), GFP_KERNEL);
 241	if (!ssp->srcu_sup)
 242		return -ENOMEM;
 243	if (!is_static)
 244		spin_lock_init(&ACCESS_PRIVATE(ssp->srcu_sup, lock));
 245	ssp->srcu_sup->srcu_size_state = SRCU_SIZE_SMALL;
 246	ssp->srcu_sup->node = NULL;
 247	mutex_init(&ssp->srcu_sup->srcu_cb_mutex);
 248	mutex_init(&ssp->srcu_sup->srcu_gp_mutex);
 249	ssp->srcu_idx = 0;
 250	ssp->srcu_sup->srcu_gp_seq = 0;
 251	ssp->srcu_sup->srcu_barrier_seq = 0;
 252	mutex_init(&ssp->srcu_sup->srcu_barrier_mutex);
 253	atomic_set(&ssp->srcu_sup->srcu_barrier_cpu_cnt, 0);
 254	INIT_DELAYED_WORK(&ssp->srcu_sup->work, process_srcu);
 255	ssp->srcu_sup->sda_is_static = is_static;
 256	if (!is_static)
 257		ssp->sda = alloc_percpu(struct srcu_data);
 258	if (!ssp->sda)
 259		goto err_free_sup;
 260	init_srcu_struct_data(ssp);
 261	ssp->srcu_sup->srcu_gp_seq_needed_exp = 0;
 262	ssp->srcu_sup->srcu_last_gp_end = ktime_get_mono_fast_ns();
 263	if (READ_ONCE(ssp->srcu_sup->srcu_size_state) == SRCU_SIZE_SMALL && SRCU_SIZING_IS_INIT()) {
 264		if (!init_srcu_struct_nodes(ssp, GFP_ATOMIC))
 265			goto err_free_sda;
 266		WRITE_ONCE(ssp->srcu_sup->srcu_size_state, SRCU_SIZE_BIG);
 267	}
 268	ssp->srcu_sup->srcu_ssp = ssp;
 269	smp_store_release(&ssp->srcu_sup->srcu_gp_seq_needed, 0); /* Init done. */
 270	return 0;
 271
 272err_free_sda:
 273	if (!is_static) {
 274		free_percpu(ssp->sda);
 275		ssp->sda = NULL;
 276	}
 277err_free_sup:
 278	if (!is_static) {
 279		kfree(ssp->srcu_sup);
 280		ssp->srcu_sup = NULL;
 281	}
 282	return -ENOMEM;
 283}
 284
 285#ifdef CONFIG_DEBUG_LOCK_ALLOC
 286
 287int __init_srcu_struct(struct srcu_struct *ssp, const char *name,
 288		       struct lock_class_key *key)
 289{
 290	/* Don't re-initialize a lock while it is held. */
 291	debug_check_no_locks_freed((void *)ssp, sizeof(*ssp));
 292	lockdep_init_map(&ssp->dep_map, name, key, 0);
 293	return init_srcu_struct_fields(ssp, false);
 294}
 295EXPORT_SYMBOL_GPL(__init_srcu_struct);
 296
 297#else /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
 298
 299/**
 300 * init_srcu_struct - initialize a sleep-RCU structure
 301 * @ssp: structure to initialize.
 302 *
 303 * Must invoke this on a given srcu_struct before passing that srcu_struct
 304 * to any other function.  Each srcu_struct represents a separate domain
 305 * of SRCU protection.
 306 */
 307int init_srcu_struct(struct srcu_struct *ssp)
 308{
 309	return init_srcu_struct_fields(ssp, false);
 310}
 311EXPORT_SYMBOL_GPL(init_srcu_struct);
 312
 313#endif /* #else #ifdef CONFIG_DEBUG_LOCK_ALLOC */
 314
 315/*
 316 * Initiate a transition to SRCU_SIZE_BIG with lock held.
 317 */
 318static void __srcu_transition_to_big(struct srcu_struct *ssp)
 319{
 320	lockdep_assert_held(&ACCESS_PRIVATE(ssp->srcu_sup, lock));
 321	smp_store_release(&ssp->srcu_sup->srcu_size_state, SRCU_SIZE_ALLOC);
 322}
 323
 324/*
 325 * Initiate an idempotent transition to SRCU_SIZE_BIG.
 326 */
 327static void srcu_transition_to_big(struct srcu_struct *ssp)
 328{
 329	unsigned long flags;
 330
 331	/* Double-checked locking on ->srcu_size-state. */
 332	if (smp_load_acquire(&ssp->srcu_sup->srcu_size_state) != SRCU_SIZE_SMALL)
 333		return;
 334	spin_lock_irqsave_rcu_node(ssp->srcu_sup, flags);
 335	if (smp_load_acquire(&ssp->srcu_sup->srcu_size_state) != SRCU_SIZE_SMALL) {
 336		spin_unlock_irqrestore_rcu_node(ssp->srcu_sup, flags);
 337		return;
 338	}
 339	__srcu_transition_to_big(ssp);
 340	spin_unlock_irqrestore_rcu_node(ssp->srcu_sup, flags);
 341}
 342
 343/*
 344 * Check to see if the just-encountered contention event justifies
 345 * a transition to SRCU_SIZE_BIG.
 346 */
 347static void spin_lock_irqsave_check_contention(struct srcu_struct *ssp)
 348{
 349	unsigned long j;
 350
 351	if (!SRCU_SIZING_IS_CONTEND() || ssp->srcu_sup->srcu_size_state)
 352		return;
 353	j = jiffies;
 354	if (ssp->srcu_sup->srcu_size_jiffies != j) {
 355		ssp->srcu_sup->srcu_size_jiffies = j;
 356		ssp->srcu_sup->srcu_n_lock_retries = 0;
 357	}
 358	if (++ssp->srcu_sup->srcu_n_lock_retries <= small_contention_lim)
 359		return;
 360	__srcu_transition_to_big(ssp);
 361}
 362
 363/*
 364 * Acquire the specified srcu_data structure's ->lock, but check for
 365 * excessive contention, which results in initiation of a transition
 366 * to SRCU_SIZE_BIG.  But only if the srcutree.convert_to_big module
 367 * parameter permits this.
 368 */
 369static void spin_lock_irqsave_sdp_contention(struct srcu_data *sdp, unsigned long *flags)
 370{
 371	struct srcu_struct *ssp = sdp->ssp;
 372
 373	if (spin_trylock_irqsave_rcu_node(sdp, *flags))
 374		return;
 375	spin_lock_irqsave_rcu_node(ssp->srcu_sup, *flags);
 376	spin_lock_irqsave_check_contention(ssp);
 377	spin_unlock_irqrestore_rcu_node(ssp->srcu_sup, *flags);
 378	spin_lock_irqsave_rcu_node(sdp, *flags);
 379}
 380
 381/*
 382 * Acquire the specified srcu_struct structure's ->lock, but check for
 383 * excessive contention, which results in initiation of a transition
 384 * to SRCU_SIZE_BIG.  But only if the srcutree.convert_to_big module
 385 * parameter permits this.
 386 */
 387static void spin_lock_irqsave_ssp_contention(struct srcu_struct *ssp, unsigned long *flags)
 388{
 389	if (spin_trylock_irqsave_rcu_node(ssp->srcu_sup, *flags))
 390		return;
 391	spin_lock_irqsave_rcu_node(ssp->srcu_sup, *flags);
 392	spin_lock_irqsave_check_contention(ssp);
 393}
 394
 395/*
 396 * First-use initialization of statically allocated srcu_struct
 397 * structure.  Wiring up the combining tree is more than can be
 398 * done with compile-time initialization, so this check is added
 399 * to each update-side SRCU primitive.  Use ssp->lock, which -is-
 400 * compile-time initialized, to resolve races involving multiple
 401 * CPUs trying to garner first-use privileges.
 402 */
 403static void check_init_srcu_struct(struct srcu_struct *ssp)
 404{
 405	unsigned long flags;
 406
 407	/* The smp_load_acquire() pairs with the smp_store_release(). */
 408	if (!rcu_seq_state(smp_load_acquire(&ssp->srcu_sup->srcu_gp_seq_needed))) /*^^^*/
 409		return; /* Already initialized. */
 410	spin_lock_irqsave_rcu_node(ssp->srcu_sup, flags);
 411	if (!rcu_seq_state(ssp->srcu_sup->srcu_gp_seq_needed)) {
 412		spin_unlock_irqrestore_rcu_node(ssp->srcu_sup, flags);
 413		return;
 414	}
 415	init_srcu_struct_fields(ssp, true);
 416	spin_unlock_irqrestore_rcu_node(ssp->srcu_sup, flags);
 417}
 418
 419/*
 420 * Returns approximate total of the readers' ->srcu_lock_count[] values
 421 * for the rank of per-CPU counters specified by idx.
 422 */
 423static unsigned long srcu_readers_lock_idx(struct srcu_struct *ssp, int idx)
 424{
 425	int cpu;
 426	unsigned long sum = 0;
 427
 428	for_each_possible_cpu(cpu) {
 429		struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
 430
 431		sum += atomic_long_read(&cpuc->srcu_lock_count[idx]);
 432	}
 433	return sum;
 434}
 435
 436/*
 437 * Returns approximate total of the readers' ->srcu_unlock_count[] values
 438 * for the rank of per-CPU counters specified by idx.
 439 */
 440static unsigned long srcu_readers_unlock_idx(struct srcu_struct *ssp, int idx)
 441{
 442	int cpu;
 443	unsigned long mask = 0;
 444	unsigned long sum = 0;
 445
 446	for_each_possible_cpu(cpu) {
 447		struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
 448
 449		sum += atomic_long_read(&cpuc->srcu_unlock_count[idx]);
 450		if (IS_ENABLED(CONFIG_PROVE_RCU))
 451			mask = mask | READ_ONCE(cpuc->srcu_nmi_safety);
 452	}
 453	WARN_ONCE(IS_ENABLED(CONFIG_PROVE_RCU) && (mask & (mask >> 1)),
 454		  "Mixed NMI-safe readers for srcu_struct at %ps.\n", ssp);
 455	return sum;
 456}
 457
 458/*
 459 * Return true if the number of pre-existing readers is determined to
 460 * be zero.
 461 */
 462static bool srcu_readers_active_idx_check(struct srcu_struct *ssp, int idx)
 463{
 464	unsigned long unlocks;
 465
 466	unlocks = srcu_readers_unlock_idx(ssp, idx);
 467
 468	/*
 469	 * Make sure that a lock is always counted if the corresponding
 470	 * unlock is counted. Needs to be a smp_mb() as the read side may
 471	 * contain a read from a variable that is written to before the
 472	 * synchronize_srcu() in the write side. In this case smp_mb()s
 473	 * A and B act like the store buffering pattern.
 474	 *
 475	 * This smp_mb() also pairs with smp_mb() C to prevent accesses
 476	 * after the synchronize_srcu() from being executed before the
 477	 * grace period ends.
 478	 */
 479	smp_mb(); /* A */
 480
 481	/*
 482	 * If the locks are the same as the unlocks, then there must have
 483	 * been no readers on this index at some point in this function.
 484	 * But there might be more readers, as a task might have read
 485	 * the current ->srcu_idx but not yet have incremented its CPU's
 486	 * ->srcu_lock_count[idx] counter.  In fact, it is possible
 487	 * that most of the tasks have been preempted between fetching
 488	 * ->srcu_idx and incrementing ->srcu_lock_count[idx].  And there
 489	 * could be almost (ULONG_MAX / sizeof(struct task_struct)) tasks
 490	 * in a system whose address space was fully populated with memory.
 491	 * Call this quantity Nt.
 492	 *
 493	 * So suppose that the updater is preempted at this point in the
 494	 * code for a long time.  That now-preempted updater has already
 495	 * flipped ->srcu_idx (possibly during the preceding grace period),
 496	 * done an smp_mb() (again, possibly during the preceding grace
 497	 * period), and summed up the ->srcu_unlock_count[idx] counters.
 498	 * How many times can a given one of the aforementioned Nt tasks
 499	 * increment the old ->srcu_idx value's ->srcu_lock_count[idx]
 500	 * counter, in the absence of nesting?
 501	 *
 502	 * It can clearly do so once, given that it has already fetched
 503	 * the old value of ->srcu_idx and is just about to use that value
 504	 * to index its increment of ->srcu_lock_count[idx].  But as soon as
 505	 * it leaves that SRCU read-side critical section, it will increment
 506	 * ->srcu_unlock_count[idx], which must follow the updater's above
 507	 * read from that same value.  Thus, as soon the reading task does
 508	 * an smp_mb() and a later fetch from ->srcu_idx, that task will be
 509	 * guaranteed to get the new index.  Except that the increment of
 510	 * ->srcu_unlock_count[idx] in __srcu_read_unlock() is after the
 511	 * smp_mb(), and the fetch from ->srcu_idx in __srcu_read_lock()
 512	 * is before the smp_mb().  Thus, that task might not see the new
 513	 * value of ->srcu_idx until the -second- __srcu_read_lock(),
 514	 * which in turn means that this task might well increment
 515	 * ->srcu_lock_count[idx] for the old value of ->srcu_idx twice,
 516	 * not just once.
 517	 *
 518	 * However, it is important to note that a given smp_mb() takes
 519	 * effect not just for the task executing it, but also for any
 520	 * later task running on that same CPU.
 521	 *
 522	 * That is, there can be almost Nt + Nc further increments of
 523	 * ->srcu_lock_count[idx] for the old index, where Nc is the number
 524	 * of CPUs.  But this is OK because the size of the task_struct
 525	 * structure limits the value of Nt and current systems limit Nc
 526	 * to a few thousand.
 527	 *
 528	 * OK, but what about nesting?  This does impose a limit on
 529	 * nesting of half of the size of the task_struct structure
 530	 * (measured in bytes), which should be sufficient.  A late 2022
 531	 * TREE01 rcutorture run reported this size to be no less than
 532	 * 9408 bytes, allowing up to 4704 levels of nesting, which is
 533	 * comfortably beyond excessive.  Especially on 64-bit systems,
 534	 * which are unlikely to be configured with an address space fully
 535	 * populated with memory, at least not anytime soon.
 536	 */
 537	return srcu_readers_lock_idx(ssp, idx) == unlocks;
 538}
 539
 540/**
 541 * srcu_readers_active - returns true if there are readers. and false
 542 *                       otherwise
 543 * @ssp: which srcu_struct to count active readers (holding srcu_read_lock).
 544 *
 545 * Note that this is not an atomic primitive, and can therefore suffer
 546 * severe errors when invoked on an active srcu_struct.  That said, it
 547 * can be useful as an error check at cleanup time.
 548 */
 549static bool srcu_readers_active(struct srcu_struct *ssp)
 550{
 551	int cpu;
 552	unsigned long sum = 0;
 553
 554	for_each_possible_cpu(cpu) {
 555		struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
 556
 557		sum += atomic_long_read(&cpuc->srcu_lock_count[0]);
 558		sum += atomic_long_read(&cpuc->srcu_lock_count[1]);
 559		sum -= atomic_long_read(&cpuc->srcu_unlock_count[0]);
 560		sum -= atomic_long_read(&cpuc->srcu_unlock_count[1]);
 561	}
 562	return sum;
 563}
 564
 565/*
 566 * We use an adaptive strategy for synchronize_srcu() and especially for
 567 * synchronize_srcu_expedited().  We spin for a fixed time period
 568 * (defined below, boot time configurable) to allow SRCU readers to exit
 569 * their read-side critical sections.  If there are still some readers
 570 * after one jiffy, we repeatedly block for one jiffy time periods.
 571 * The blocking time is increased as the grace-period age increases,
 572 * with max blocking time capped at 10 jiffies.
 573 */
 574#define SRCU_DEFAULT_RETRY_CHECK_DELAY		5
 575
 576static ulong srcu_retry_check_delay = SRCU_DEFAULT_RETRY_CHECK_DELAY;
 577module_param(srcu_retry_check_delay, ulong, 0444);
 578
 579#define SRCU_INTERVAL		1		// Base delay if no expedited GPs pending.
 580#define SRCU_MAX_INTERVAL	10		// Maximum incremental delay from slow readers.
 581
 582#define SRCU_DEFAULT_MAX_NODELAY_PHASE_LO	3UL	// Lowmark on default per-GP-phase
 583							// no-delay instances.
 584#define SRCU_DEFAULT_MAX_NODELAY_PHASE_HI	1000UL	// Highmark on default per-GP-phase
 585							// no-delay instances.
 586
 587#define SRCU_UL_CLAMP_LO(val, low)	((val) > (low) ? (val) : (low))
 588#define SRCU_UL_CLAMP_HI(val, high)	((val) < (high) ? (val) : (high))
 589#define SRCU_UL_CLAMP(val, low, high)	SRCU_UL_CLAMP_HI(SRCU_UL_CLAMP_LO((val), (low)), (high))
 590// per-GP-phase no-delay instances adjusted to allow non-sleeping poll upto
 591// one jiffies time duration. Mult by 2 is done to factor in the srcu_get_delay()
 592// called from process_srcu().
 593#define SRCU_DEFAULT_MAX_NODELAY_PHASE_ADJUSTED	\
 594	(2UL * USEC_PER_SEC / HZ / SRCU_DEFAULT_RETRY_CHECK_DELAY)
 595
 596// Maximum per-GP-phase consecutive no-delay instances.
 597#define SRCU_DEFAULT_MAX_NODELAY_PHASE	\
 598	SRCU_UL_CLAMP(SRCU_DEFAULT_MAX_NODELAY_PHASE_ADJUSTED,	\
 599		      SRCU_DEFAULT_MAX_NODELAY_PHASE_LO,	\
 600		      SRCU_DEFAULT_MAX_NODELAY_PHASE_HI)
 601
 602static ulong srcu_max_nodelay_phase = SRCU_DEFAULT_MAX_NODELAY_PHASE;
 603module_param(srcu_max_nodelay_phase, ulong, 0444);
 604
 605// Maximum consecutive no-delay instances.
 606#define SRCU_DEFAULT_MAX_NODELAY	(SRCU_DEFAULT_MAX_NODELAY_PHASE > 100 ?	\
 607					 SRCU_DEFAULT_MAX_NODELAY_PHASE : 100)
 608
 609static ulong srcu_max_nodelay = SRCU_DEFAULT_MAX_NODELAY;
 610module_param(srcu_max_nodelay, ulong, 0444);
 611
 612/*
 613 * Return grace-period delay, zero if there are expedited grace
 614 * periods pending, SRCU_INTERVAL otherwise.
 615 */
 616static unsigned long srcu_get_delay(struct srcu_struct *ssp)
 617{
 618	unsigned long gpstart;
 619	unsigned long j;
 620	unsigned long jbase = SRCU_INTERVAL;
 621	struct srcu_usage *sup = ssp->srcu_sup;
 622
 623	if (ULONG_CMP_LT(READ_ONCE(sup->srcu_gp_seq), READ_ONCE(sup->srcu_gp_seq_needed_exp)))
 624		jbase = 0;
 625	if (rcu_seq_state(READ_ONCE(sup->srcu_gp_seq))) {
 626		j = jiffies - 1;
 627		gpstart = READ_ONCE(sup->srcu_gp_start);
 628		if (time_after(j, gpstart))
 629			jbase += j - gpstart;
 630		if (!jbase) {
 631			WRITE_ONCE(sup->srcu_n_exp_nodelay, READ_ONCE(sup->srcu_n_exp_nodelay) + 1);
 632			if (READ_ONCE(sup->srcu_n_exp_nodelay) > srcu_max_nodelay_phase)
 633				jbase = 1;
 634		}
 635	}
 636	return jbase > SRCU_MAX_INTERVAL ? SRCU_MAX_INTERVAL : jbase;
 637}
 638
 639/**
 640 * cleanup_srcu_struct - deconstruct a sleep-RCU structure
 641 * @ssp: structure to clean up.
 642 *
 643 * Must invoke this after you are finished using a given srcu_struct that
 644 * was initialized via init_srcu_struct(), else you leak memory.
 645 */
 646void cleanup_srcu_struct(struct srcu_struct *ssp)
 647{
 648	int cpu;
 649	struct srcu_usage *sup = ssp->srcu_sup;
 650
 651	if (WARN_ON(!srcu_get_delay(ssp)))
 652		return; /* Just leak it! */
 653	if (WARN_ON(srcu_readers_active(ssp)))
 654		return; /* Just leak it! */
 655	flush_delayed_work(&sup->work);
 656	for_each_possible_cpu(cpu) {
 657		struct srcu_data *sdp = per_cpu_ptr(ssp->sda, cpu);
 658
 659		del_timer_sync(&sdp->delay_work);
 660		flush_work(&sdp->work);
 661		if (WARN_ON(rcu_segcblist_n_cbs(&sdp->srcu_cblist)))
 662			return; /* Forgot srcu_barrier(), so just leak it! */
 663	}
 664	if (WARN_ON(rcu_seq_state(READ_ONCE(sup->srcu_gp_seq)) != SRCU_STATE_IDLE) ||
 665	    WARN_ON(rcu_seq_current(&sup->srcu_gp_seq) != sup->srcu_gp_seq_needed) ||
 666	    WARN_ON(srcu_readers_active(ssp))) {
 667		pr_info("%s: Active srcu_struct %p read state: %d gp state: %lu/%lu\n",
 668			__func__, ssp, rcu_seq_state(READ_ONCE(sup->srcu_gp_seq)),
 669			rcu_seq_current(&sup->srcu_gp_seq), sup->srcu_gp_seq_needed);
 670		return; /* Caller forgot to stop doing call_srcu()? */
 671	}
 672	kfree(sup->node);
 673	sup->node = NULL;
 674	sup->srcu_size_state = SRCU_SIZE_SMALL;
 675	if (!sup->sda_is_static) {
 676		free_percpu(ssp->sda);
 677		ssp->sda = NULL;
 678		kfree(sup);
 679		ssp->srcu_sup = NULL;
 680	}
 681}
 682EXPORT_SYMBOL_GPL(cleanup_srcu_struct);
 683
 684#ifdef CONFIG_PROVE_RCU
 685/*
 686 * Check for consistent NMI safety.
 687 */
 688void srcu_check_nmi_safety(struct srcu_struct *ssp, bool nmi_safe)
 689{
 690	int nmi_safe_mask = 1 << nmi_safe;
 691	int old_nmi_safe_mask;
 692	struct srcu_data *sdp;
 693
 694	/* NMI-unsafe use in NMI is a bad sign */
 695	WARN_ON_ONCE(!nmi_safe && in_nmi());
 696	sdp = raw_cpu_ptr(ssp->sda);
 697	old_nmi_safe_mask = READ_ONCE(sdp->srcu_nmi_safety);
 698	if (!old_nmi_safe_mask) {
 699		WRITE_ONCE(sdp->srcu_nmi_safety, nmi_safe_mask);
 700		return;
 701	}
 702	WARN_ONCE(old_nmi_safe_mask != nmi_safe_mask, "CPU %d old state %d new state %d\n", sdp->cpu, old_nmi_safe_mask, nmi_safe_mask);
 703}
 704EXPORT_SYMBOL_GPL(srcu_check_nmi_safety);
 705#endif /* CONFIG_PROVE_RCU */
 706
 707/*
 708 * Counts the new reader in the appropriate per-CPU element of the
 709 * srcu_struct.
 710 * Returns an index that must be passed to the matching srcu_read_unlock().
 711 */
 712int __srcu_read_lock(struct srcu_struct *ssp)
 713{
 714	int idx;
 715
 716	idx = READ_ONCE(ssp->srcu_idx) & 0x1;
 717	this_cpu_inc(ssp->sda->srcu_lock_count[idx].counter);
 718	smp_mb(); /* B */  /* Avoid leaking the critical section. */
 719	return idx;
 720}
 721EXPORT_SYMBOL_GPL(__srcu_read_lock);
 722
 723/*
 724 * Removes the count for the old reader from the appropriate per-CPU
 725 * element of the srcu_struct.  Note that this may well be a different
 726 * CPU than that which was incremented by the corresponding srcu_read_lock().
 727 */
 728void __srcu_read_unlock(struct srcu_struct *ssp, int idx)
 729{
 730	smp_mb(); /* C */  /* Avoid leaking the critical section. */
 731	this_cpu_inc(ssp->sda->srcu_unlock_count[idx].counter);
 732}
 733EXPORT_SYMBOL_GPL(__srcu_read_unlock);
 734
 735#ifdef CONFIG_NEED_SRCU_NMI_SAFE
 736
 737/*
 738 * Counts the new reader in the appropriate per-CPU element of the
 739 * srcu_struct, but in an NMI-safe manner using RMW atomics.
 740 * Returns an index that must be passed to the matching srcu_read_unlock().
 741 */
 742int __srcu_read_lock_nmisafe(struct srcu_struct *ssp)
 743{
 744	int idx;
 745	struct srcu_data *sdp = raw_cpu_ptr(ssp->sda);
 746
 747	idx = READ_ONCE(ssp->srcu_idx) & 0x1;
 748	atomic_long_inc(&sdp->srcu_lock_count[idx]);
 749	smp_mb__after_atomic(); /* B */  /* Avoid leaking the critical section. */
 750	return idx;
 751}
 752EXPORT_SYMBOL_GPL(__srcu_read_lock_nmisafe);
 753
 754/*
 755 * Removes the count for the old reader from the appropriate per-CPU
 756 * element of the srcu_struct.  Note that this may well be a different
 757 * CPU than that which was incremented by the corresponding srcu_read_lock().
 758 */
 759void __srcu_read_unlock_nmisafe(struct srcu_struct *ssp, int idx)
 760{
 761	struct srcu_data *sdp = raw_cpu_ptr(ssp->sda);
 762
 763	smp_mb__before_atomic(); /* C */  /* Avoid leaking the critical section. */
 764	atomic_long_inc(&sdp->srcu_unlock_count[idx]);
 765}
 766EXPORT_SYMBOL_GPL(__srcu_read_unlock_nmisafe);
 767
 768#endif // CONFIG_NEED_SRCU_NMI_SAFE
 769
 770/*
 771 * Start an SRCU grace period.
 772 */
 773static void srcu_gp_start(struct srcu_struct *ssp)
 774{
 775	int state;
 776
 777	lockdep_assert_held(&ACCESS_PRIVATE(ssp->srcu_sup, lock));
 778	WARN_ON_ONCE(ULONG_CMP_GE(ssp->srcu_sup->srcu_gp_seq, ssp->srcu_sup->srcu_gp_seq_needed));
 779	WRITE_ONCE(ssp->srcu_sup->srcu_gp_start, jiffies);
 780	WRITE_ONCE(ssp->srcu_sup->srcu_n_exp_nodelay, 0);
 781	smp_mb(); /* Order prior store to ->srcu_gp_seq_needed vs. GP start. */
 782	rcu_seq_start(&ssp->srcu_sup->srcu_gp_seq);
 783	state = rcu_seq_state(ssp->srcu_sup->srcu_gp_seq);
 784	WARN_ON_ONCE(state != SRCU_STATE_SCAN1);
 785}
 786
 787
 788static void srcu_delay_timer(struct timer_list *t)
 789{
 790	struct srcu_data *sdp = container_of(t, struct srcu_data, delay_work);
 791
 792	queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work);
 793}
 794
 795static void srcu_queue_delayed_work_on(struct srcu_data *sdp,
 796				       unsigned long delay)
 797{
 798	if (!delay) {
 799		queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work);
 800		return;
 801	}
 802
 803	timer_reduce(&sdp->delay_work, jiffies + delay);
 804}
 805
 806/*
 807 * Schedule callback invocation for the specified srcu_data structure,
 808 * if possible, on the corresponding CPU.
 809 */
 810static void srcu_schedule_cbs_sdp(struct srcu_data *sdp, unsigned long delay)
 811{
 812	srcu_queue_delayed_work_on(sdp, delay);
 813}
 814
 815/*
 816 * Schedule callback invocation for all srcu_data structures associated
 817 * with the specified srcu_node structure that have callbacks for the
 818 * just-completed grace period, the one corresponding to idx.  If possible,
 819 * schedule this invocation on the corresponding CPUs.
 820 */
 821static void srcu_schedule_cbs_snp(struct srcu_struct *ssp, struct srcu_node *snp,
 822				  unsigned long mask, unsigned long delay)
 823{
 824	int cpu;
 825
 826	for (cpu = snp->grplo; cpu <= snp->grphi; cpu++) {
 827		if (!(mask & (1UL << (cpu - snp->grplo))))
 828			continue;
 829		srcu_schedule_cbs_sdp(per_cpu_ptr(ssp->sda, cpu), delay);
 830	}
 831}
 832
 833/*
 834 * Note the end of an SRCU grace period.  Initiates callback invocation
 835 * and starts a new grace period if needed.
 836 *
 837 * The ->srcu_cb_mutex acquisition does not protect any data, but
 838 * instead prevents more than one grace period from starting while we
 839 * are initiating callback invocation.  This allows the ->srcu_have_cbs[]
 840 * array to have a finite number of elements.
 841 */
 842static void srcu_gp_end(struct srcu_struct *ssp)
 843{
 844	unsigned long cbdelay = 1;
 845	bool cbs;
 846	bool last_lvl;
 847	int cpu;
 848	unsigned long flags;
 849	unsigned long gpseq;
 850	int idx;
 851	unsigned long mask;
 852	struct srcu_data *sdp;
 853	unsigned long sgsne;
 854	struct srcu_node *snp;
 855	int ss_state;
 856	struct srcu_usage *sup = ssp->srcu_sup;
 857
 858	/* Prevent more than one additional grace period. */
 859	mutex_lock(&sup->srcu_cb_mutex);
 860
 861	/* End the current grace period. */
 862	spin_lock_irq_rcu_node(sup);
 863	idx = rcu_seq_state(sup->srcu_gp_seq);
 864	WARN_ON_ONCE(idx != SRCU_STATE_SCAN2);
 865	if (ULONG_CMP_LT(READ_ONCE(sup->srcu_gp_seq), READ_ONCE(sup->srcu_gp_seq_needed_exp)))
 866		cbdelay = 0;
 867
 868	WRITE_ONCE(sup->srcu_last_gp_end, ktime_get_mono_fast_ns());
 869	rcu_seq_end(&sup->srcu_gp_seq);
 870	gpseq = rcu_seq_current(&sup->srcu_gp_seq);
 871	if (ULONG_CMP_LT(sup->srcu_gp_seq_needed_exp, gpseq))
 872		WRITE_ONCE(sup->srcu_gp_seq_needed_exp, gpseq);
 873	spin_unlock_irq_rcu_node(sup);
 874	mutex_unlock(&sup->srcu_gp_mutex);
 875	/* A new grace period can start at this point.  But only one. */
 876
 877	/* Initiate callback invocation as needed. */
 878	ss_state = smp_load_acquire(&sup->srcu_size_state);
 879	if (ss_state < SRCU_SIZE_WAIT_BARRIER) {
 880		srcu_schedule_cbs_sdp(per_cpu_ptr(ssp->sda, get_boot_cpu_id()),
 881					cbdelay);
 882	} else {
 883		idx = rcu_seq_ctr(gpseq) % ARRAY_SIZE(snp->srcu_have_cbs);
 884		srcu_for_each_node_breadth_first(ssp, snp) {
 885			spin_lock_irq_rcu_node(snp);
 886			cbs = false;
 887			last_lvl = snp >= sup->level[rcu_num_lvls - 1];
 888			if (last_lvl)
 889				cbs = ss_state < SRCU_SIZE_BIG || snp->srcu_have_cbs[idx] == gpseq;
 890			snp->srcu_have_cbs[idx] = gpseq;
 891			rcu_seq_set_state(&snp->srcu_have_cbs[idx], 1);
 892			sgsne = snp->srcu_gp_seq_needed_exp;
 893			if (srcu_invl_snp_seq(sgsne) || ULONG_CMP_LT(sgsne, gpseq))
 894				WRITE_ONCE(snp->srcu_gp_seq_needed_exp, gpseq);
 895			if (ss_state < SRCU_SIZE_BIG)
 896				mask = ~0;
 897			else
 898				mask = snp->srcu_data_have_cbs[idx];
 899			snp->srcu_data_have_cbs[idx] = 0;
 900			spin_unlock_irq_rcu_node(snp);
 901			if (cbs)
 902				srcu_schedule_cbs_snp(ssp, snp, mask, cbdelay);
 903		}
 904	}
 905
 906	/* Occasionally prevent srcu_data counter wrap. */
 907	if (!(gpseq & counter_wrap_check))
 908		for_each_possible_cpu(cpu) {
 909			sdp = per_cpu_ptr(ssp->sda, cpu);
 910			spin_lock_irqsave_rcu_node(sdp, flags);
 911			if (ULONG_CMP_GE(gpseq, sdp->srcu_gp_seq_needed + 100))
 912				sdp->srcu_gp_seq_needed = gpseq;
 913			if (ULONG_CMP_GE(gpseq, sdp->srcu_gp_seq_needed_exp + 100))
 914				sdp->srcu_gp_seq_needed_exp = gpseq;
 915			spin_unlock_irqrestore_rcu_node(sdp, flags);
 916		}
 917
 918	/* Callback initiation done, allow grace periods after next. */
 919	mutex_unlock(&sup->srcu_cb_mutex);
 920
 921	/* Start a new grace period if needed. */
 922	spin_lock_irq_rcu_node(sup);
 923	gpseq = rcu_seq_current(&sup->srcu_gp_seq);
 924	if (!rcu_seq_state(gpseq) &&
 925	    ULONG_CMP_LT(gpseq, sup->srcu_gp_seq_needed)) {
 926		srcu_gp_start(ssp);
 927		spin_unlock_irq_rcu_node(sup);
 928		srcu_reschedule(ssp, 0);
 929	} else {
 930		spin_unlock_irq_rcu_node(sup);
 931	}
 932
 933	/* Transition to big if needed. */
 934	if (ss_state != SRCU_SIZE_SMALL && ss_state != SRCU_SIZE_BIG) {
 935		if (ss_state == SRCU_SIZE_ALLOC)
 936			init_srcu_struct_nodes(ssp, GFP_KERNEL);
 937		else
 938			smp_store_release(&sup->srcu_size_state, ss_state + 1);
 939	}
 940}
 941
 942/*
 943 * Funnel-locking scheme to scalably mediate many concurrent expedited
 944 * grace-period requests.  This function is invoked for the first known
 945 * expedited request for a grace period that has already been requested,
 946 * but without expediting.  To start a completely new grace period,
 947 * whether expedited or not, use srcu_funnel_gp_start() instead.
 948 */
 949static void srcu_funnel_exp_start(struct srcu_struct *ssp, struct srcu_node *snp,
 950				  unsigned long s)
 951{
 952	unsigned long flags;
 953	unsigned long sgsne;
 954
 955	if (snp)
 956		for (; snp != NULL; snp = snp->srcu_parent) {
 957			sgsne = READ_ONCE(snp->srcu_gp_seq_needed_exp);
 958			if (WARN_ON_ONCE(rcu_seq_done(&ssp->srcu_sup->srcu_gp_seq, s)) ||
 959			    (!srcu_invl_snp_seq(sgsne) && ULONG_CMP_GE(sgsne, s)))
 960				return;
 961			spin_lock_irqsave_rcu_node(snp, flags);
 962			sgsne = snp->srcu_gp_seq_needed_exp;
 963			if (!srcu_invl_snp_seq(sgsne) && ULONG_CMP_GE(sgsne, s)) {
 964				spin_unlock_irqrestore_rcu_node(snp, flags);
 965				return;
 966			}
 967			WRITE_ONCE(snp->srcu_gp_seq_needed_exp, s);
 968			spin_unlock_irqrestore_rcu_node(snp, flags);
 969		}
 970	spin_lock_irqsave_ssp_contention(ssp, &flags);
 971	if (ULONG_CMP_LT(ssp->srcu_sup->srcu_gp_seq_needed_exp, s))
 972		WRITE_ONCE(ssp->srcu_sup->srcu_gp_seq_needed_exp, s);
 973	spin_unlock_irqrestore_rcu_node(ssp->srcu_sup, flags);
 974}
 975
 976/*
 977 * Funnel-locking scheme to scalably mediate many concurrent grace-period
 978 * requests.  The winner has to do the work of actually starting grace
 979 * period s.  Losers must either ensure that their desired grace-period
 980 * number is recorded on at least their leaf srcu_node structure, or they
 981 * must take steps to invoke their own callbacks.
 982 *
 983 * Note that this function also does the work of srcu_funnel_exp_start(),
 984 * in some cases by directly invoking it.
 985 *
 986 * The srcu read lock should be hold around this function. And s is a seq snap
 987 * after holding that lock.
 988 */
 989static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
 990				 unsigned long s, bool do_norm)
 991{
 992	unsigned long flags;
 993	int idx = rcu_seq_ctr(s) % ARRAY_SIZE(sdp->mynode->srcu_have_cbs);
 994	unsigned long sgsne;
 995	struct srcu_node *snp;
 996	struct srcu_node *snp_leaf;
 997	unsigned long snp_seq;
 998	struct srcu_usage *sup = ssp->srcu_sup;
 999
1000	/* Ensure that snp node tree is fully initialized before traversing it */
1001	if (smp_load_acquire(&sup->srcu_size_state) < SRCU_SIZE_WAIT_BARRIER)
1002		snp_leaf = NULL;
1003	else
1004		snp_leaf = sdp->mynode;
1005
1006	if (snp_leaf)
1007		/* Each pass through the loop does one level of the srcu_node tree. */
1008		for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
1009			if (WARN_ON_ONCE(rcu_seq_done(&sup->srcu_gp_seq, s)) && snp != snp_leaf)
1010				return; /* GP already done and CBs recorded. */
1011			spin_lock_irqsave_rcu_node(snp, flags);
1012			snp_seq = snp->srcu_have_cbs[idx];
1013			if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
1014				if (snp == snp_leaf && snp_seq == s)
1015					snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
1016				spin_unlock_irqrestore_rcu_node(snp, flags);
1017				if (snp == snp_leaf && snp_seq != s) {
1018					srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
1019					return;
1020				}
1021				if (!do_norm)
1022					srcu_funnel_exp_start(ssp, snp, s);
1023				return;
1024			}
1025			snp->srcu_have_cbs[idx] = s;
1026			if (snp == snp_leaf)
1027				snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
1028			sgsne = snp->srcu_gp_seq_needed_exp;
1029			if (!do_norm && (srcu_invl_snp_seq(sgsne) || ULONG_CMP_LT(sgsne, s)))
1030				WRITE_ONCE(snp->srcu_gp_seq_needed_exp, s);
1031			spin_unlock_irqrestore_rcu_node(snp, flags);
1032		}
1033
1034	/* Top of tree, must ensure the grace period will be started. */
1035	spin_lock_irqsave_ssp_contention(ssp, &flags);
1036	if (ULONG_CMP_LT(sup->srcu_gp_seq_needed, s)) {
1037		/*
1038		 * Record need for grace period s.  Pair with load
1039		 * acquire setting up for initialization.
1040		 */
1041		smp_store_release(&sup->srcu_gp_seq_needed, s); /*^^^*/
1042	}
1043	if (!do_norm && ULONG_CMP_LT(sup->srcu_gp_seq_needed_exp, s))
1044		WRITE_ONCE(sup->srcu_gp_seq_needed_exp, s);
1045
1046	/* If grace period not already in progress, start it. */
1047	if (!WARN_ON_ONCE(rcu_seq_done(&sup->srcu_gp_seq, s)) &&
1048	    rcu_seq_state(sup->srcu_gp_seq) == SRCU_STATE_IDLE) {
1049		WARN_ON_ONCE(ULONG_CMP_GE(sup->srcu_gp_seq, sup->srcu_gp_seq_needed));
1050		srcu_gp_start(ssp);
1051
1052		// And how can that list_add() in the "else" clause
1053		// possibly be safe for concurrent execution?  Well,
1054		// it isn't.  And it does not have to be.  After all, it
1055		// can only be executed during early boot when there is only
1056		// the one boot CPU running with interrupts still disabled.
1057		if (likely(srcu_init_done))
1058			queue_delayed_work(rcu_gp_wq, &sup->work,
1059					   !!srcu_get_delay(ssp));
1060		else if (list_empty(&sup->work.work.entry))
1061			list_add(&sup->work.work.entry, &srcu_boot_list);
1062	}
1063	spin_unlock_irqrestore_rcu_node(sup, flags);
1064}
1065
1066/*
1067 * Wait until all readers counted by array index idx complete, but
1068 * loop an additional time if there is an expedited grace period pending.
1069 * The caller must ensure that ->srcu_idx is not changed while checking.
1070 */
1071static bool try_check_zero(struct srcu_struct *ssp, int idx, int trycount)
1072{
1073	unsigned long curdelay;
1074
1075	curdelay = !srcu_get_delay(ssp);
1076
1077	for (;;) {
1078		if (srcu_readers_active_idx_check(ssp, idx))
1079			return true;
1080		if ((--trycount + curdelay) <= 0)
1081			return false;
1082		udelay(srcu_retry_check_delay);
1083	}
1084}
1085
1086/*
1087 * Increment the ->srcu_idx counter so that future SRCU readers will
1088 * use the other rank of the ->srcu_(un)lock_count[] arrays.  This allows
1089 * us to wait for pre-existing readers in a starvation-free manner.
1090 */
1091static void srcu_flip(struct srcu_struct *ssp)
1092{
1093	/*
1094	 * Because the flip of ->srcu_idx is executed only if the
1095	 * preceding call to srcu_readers_active_idx_check() found that
1096	 * the ->srcu_unlock_count[] and ->srcu_lock_count[] sums matched
1097	 * and because that summing uses atomic_long_read(), there is
1098	 * ordering due to a control dependency between that summing and
1099	 * the WRITE_ONCE() in this call to srcu_flip().  This ordering
1100	 * ensures that if this updater saw a given reader's increment from
1101	 * __srcu_read_lock(), that reader was using a value of ->srcu_idx
1102	 * from before the previous call to srcu_flip(), which should be
1103	 * quite rare.  This ordering thus helps forward progress because
1104	 * the grace period could otherwise be delayed by additional
1105	 * calls to __srcu_read_lock() using that old (soon to be new)
1106	 * value of ->srcu_idx.
1107	 *
1108	 * This sum-equality check and ordering also ensures that if
1109	 * a given call to __srcu_read_lock() uses the new value of
1110	 * ->srcu_idx, this updater's earlier scans cannot have seen
1111	 * that reader's increments, which is all to the good, because
1112	 * this grace period need not wait on that reader.  After all,
1113	 * if those earlier scans had seen that reader, there would have
1114	 * been a sum mismatch and this code would not be reached.
1115	 *
1116	 * This means that the following smp_mb() is redundant, but
1117	 * it stays until either (1) Compilers learn about this sort of
1118	 * control dependency or (2) Some production workload running on
1119	 * a production system is unduly delayed by this slowpath smp_mb().
1120	 */
1121	smp_mb(); /* E */  /* Pairs with B and C. */
1122
1123	WRITE_ONCE(ssp->srcu_idx, ssp->srcu_idx + 1); // Flip the counter.
1124
1125	/*
1126	 * Ensure that if the updater misses an __srcu_read_unlock()
1127	 * increment, that task's __srcu_read_lock() following its next
1128	 * __srcu_read_lock() or __srcu_read_unlock() will see the above
1129	 * counter update.  Note that both this memory barrier and the
1130	 * one in srcu_readers_active_idx_check() provide the guarantee
1131	 * for __srcu_read_lock().
1132	 */
1133	smp_mb(); /* D */  /* Pairs with C. */
1134}
1135
1136/*
1137 * If SRCU is likely idle, return true, otherwise return false.
1138 *
1139 * Note that it is OK for several current from-idle requests for a new
1140 * grace period from idle to specify expediting because they will all end
1141 * up requesting the same grace period anyhow.  So no loss.
1142 *
1143 * Note also that if any CPU (including the current one) is still invoking
1144 * callbacks, this function will nevertheless say "idle".  This is not
1145 * ideal, but the overhead of checking all CPUs' callback lists is even
1146 * less ideal, especially on large systems.  Furthermore, the wakeup
1147 * can happen before the callback is fully removed, so we have no choice
1148 * but to accept this type of error.
1149 *
1150 * This function is also subject to counter-wrap errors, but let's face
1151 * it, if this function was preempted for enough time for the counters
1152 * to wrap, it really doesn't matter whether or not we expedite the grace
1153 * period.  The extra overhead of a needlessly expedited grace period is
1154 * negligible when amortized over that time period, and the extra latency
1155 * of a needlessly non-expedited grace period is similarly negligible.
1156 */
1157static bool srcu_might_be_idle(struct srcu_struct *ssp)
1158{
1159	unsigned long curseq;
1160	unsigned long flags;
1161	struct srcu_data *sdp;
1162	unsigned long t;
1163	unsigned long tlast;
1164
1165	check_init_srcu_struct(ssp);
1166	/* If the local srcu_data structure has callbacks, not idle.  */
1167	sdp = raw_cpu_ptr(ssp->sda);
1168	spin_lock_irqsave_rcu_node(sdp, flags);
1169	if (rcu_segcblist_pend_cbs(&sdp->srcu_cblist)) {
1170		spin_unlock_irqrestore_rcu_node(sdp, flags);
1171		return false; /* Callbacks already present, so not idle. */
1172	}
1173	spin_unlock_irqrestore_rcu_node(sdp, flags);
1174
1175	/*
1176	 * No local callbacks, so probabilistically probe global state.
1177	 * Exact information would require acquiring locks, which would
1178	 * kill scalability, hence the probabilistic nature of the probe.
1179	 */
1180
1181	/* First, see if enough time has passed since the last GP. */
1182	t = ktime_get_mono_fast_ns();
1183	tlast = READ_ONCE(ssp->srcu_sup->srcu_last_gp_end);
1184	if (exp_holdoff == 0 ||
1185	    time_in_range_open(t, tlast, tlast + exp_holdoff))
1186		return false; /* Too soon after last GP. */
1187
1188	/* Next, check for probable idleness. */
1189	curseq = rcu_seq_current(&ssp->srcu_sup->srcu_gp_seq);
1190	smp_mb(); /* Order ->srcu_gp_seq with ->srcu_gp_seq_needed. */
1191	if (ULONG_CMP_LT(curseq, READ_ONCE(ssp->srcu_sup->srcu_gp_seq_needed)))
1192		return false; /* Grace period in progress, so not idle. */
1193	smp_mb(); /* Order ->srcu_gp_seq with prior access. */
1194	if (curseq != rcu_seq_current(&ssp->srcu_sup->srcu_gp_seq))
1195		return false; /* GP # changed, so not idle. */
1196	return true; /* With reasonable probability, idle! */
1197}
1198
1199/*
1200 * SRCU callback function to leak a callback.
1201 */
1202static void srcu_leak_callback(struct rcu_head *rhp)
1203{
1204}
1205
1206/*
1207 * Start an SRCU grace period, and also queue the callback if non-NULL.
1208 */
1209static unsigned long srcu_gp_start_if_needed(struct srcu_struct *ssp,
1210					     struct rcu_head *rhp, bool do_norm)
1211{
1212	unsigned long flags;
1213	int idx;
1214	bool needexp = false;
1215	bool needgp = false;
1216	unsigned long s;
1217	struct srcu_data *sdp;
1218	struct srcu_node *sdp_mynode;
1219	int ss_state;
1220
1221	check_init_srcu_struct(ssp);
1222	/*
1223	 * While starting a new grace period, make sure we are in an
1224	 * SRCU read-side critical section so that the grace-period
1225	 * sequence number cannot wrap around in the meantime.
1226	 */
1227	idx = __srcu_read_lock_nmisafe(ssp);
1228	ss_state = smp_load_acquire(&ssp->srcu_sup->srcu_size_state);
1229	if (ss_state < SRCU_SIZE_WAIT_CALL)
1230		sdp = per_cpu_ptr(ssp->sda, get_boot_cpu_id());
1231	else
1232		sdp = raw_cpu_ptr(ssp->sda);
1233	spin_lock_irqsave_sdp_contention(sdp, &flags);
1234	if (rhp)
1235		rcu_segcblist_enqueue(&sdp->srcu_cblist, rhp);
1236	/*
1237	 * The snapshot for acceleration must be taken _before_ the read of the
1238	 * current gp sequence used for advancing, otherwise advancing may fail
1239	 * and acceleration may then fail too.
1240	 *
1241	 * This could happen if:
1242	 *
1243	 *  1) The RCU_WAIT_TAIL segment has callbacks (gp_num = X + 4) and the
1244	 *     RCU_NEXT_READY_TAIL also has callbacks (gp_num = X + 8).
1245	 *
1246	 *  2) The grace period for RCU_WAIT_TAIL is seen as started but not
1247	 *     completed so rcu_seq_current() returns X + SRCU_STATE_SCAN1.
1248	 *
1249	 *  3) This value is passed to rcu_segcblist_advance() which can't move
1250	 *     any segment forward and fails.
1251	 *
1252	 *  4) srcu_gp_start_if_needed() still proceeds with callback acceleration.
1253	 *     But then the call to rcu_seq_snap() observes the grace period for the
1254	 *     RCU_WAIT_TAIL segment as completed and the subsequent one for the
1255	 *     RCU_NEXT_READY_TAIL segment as started (ie: X + 4 + SRCU_STATE_SCAN1)
1256	 *     so it returns a snapshot of the next grace period, which is X + 12.
1257	 *
1258	 *  5) The value of X + 12 is passed to rcu_segcblist_accelerate() but the
1259	 *     freshly enqueued callback in RCU_NEXT_TAIL can't move to
1260	 *     RCU_NEXT_READY_TAIL which already has callbacks for a previous grace
1261	 *     period (gp_num = X + 8). So acceleration fails.
1262	 */
1263	s = rcu_seq_snap(&ssp->srcu_sup->srcu_gp_seq);
1264	if (rhp) {
1265		rcu_segcblist_advance(&sdp->srcu_cblist,
1266				      rcu_seq_current(&ssp->srcu_sup->srcu_gp_seq));
1267		WARN_ON_ONCE(!rcu_segcblist_accelerate(&sdp->srcu_cblist, s));
1268	}
1269	if (ULONG_CMP_LT(sdp->srcu_gp_seq_needed, s)) {
1270		sdp->srcu_gp_seq_needed = s;
1271		needgp = true;
1272	}
1273	if (!do_norm && ULONG_CMP_LT(sdp->srcu_gp_seq_needed_exp, s)) {
1274		sdp->srcu_gp_seq_needed_exp = s;
1275		needexp = true;
1276	}
1277	spin_unlock_irqrestore_rcu_node(sdp, flags);
1278
1279	/* Ensure that snp node tree is fully initialized before traversing it */
1280	if (ss_state < SRCU_SIZE_WAIT_BARRIER)
1281		sdp_mynode = NULL;
1282	else
1283		sdp_mynode = sdp->mynode;
1284
1285	if (needgp)
1286		srcu_funnel_gp_start(ssp, sdp, s, do_norm);
1287	else if (needexp)
1288		srcu_funnel_exp_start(ssp, sdp_mynode, s);
1289	__srcu_read_unlock_nmisafe(ssp, idx);
1290	return s;
1291}
1292
1293/*
1294 * Enqueue an SRCU callback on the srcu_data structure associated with
1295 * the current CPU and the specified srcu_struct structure, initiating
1296 * grace-period processing if it is not already running.
1297 *
1298 * Note that all CPUs must agree that the grace period extended beyond
1299 * all pre-existing SRCU read-side critical section.  On systems with
1300 * more than one CPU, this means that when "func()" is invoked, each CPU
1301 * is guaranteed to have executed a full memory barrier since the end of
1302 * its last corresponding SRCU read-side critical section whose beginning
1303 * preceded the call to call_srcu().  It also means that each CPU executing
1304 * an SRCU read-side critical section that continues beyond the start of
1305 * "func()" must have executed a memory barrier after the call_srcu()
1306 * but before the beginning of that SRCU read-side critical section.
1307 * Note that these guarantees include CPUs that are offline, idle, or
1308 * executing in user mode, as well as CPUs that are executing in the kernel.
1309 *
1310 * Furthermore, if CPU A invoked call_srcu() and CPU B invoked the
1311 * resulting SRCU callback function "func()", then both CPU A and CPU
1312 * B are guaranteed to execute a full memory barrier during the time
1313 * interval between the call to call_srcu() and the invocation of "func()".
1314 * This guarantee applies even if CPU A and CPU B are the same CPU (but
1315 * again only if the system has more than one CPU).
1316 *
1317 * Of course, these guarantees apply only for invocations of call_srcu(),
1318 * srcu_read_lock(), and srcu_read_unlock() that are all passed the same
1319 * srcu_struct structure.
1320 */
1321static void __call_srcu(struct srcu_struct *ssp, struct rcu_head *rhp,
1322			rcu_callback_t func, bool do_norm)
1323{
1324	if (debug_rcu_head_queue(rhp)) {
1325		/* Probable double call_srcu(), so leak the callback. */
1326		WRITE_ONCE(rhp->func, srcu_leak_callback);
1327		WARN_ONCE(1, "call_srcu(): Leaked duplicate callback\n");
1328		return;
1329	}
1330	rhp->func = func;
1331	(void)srcu_gp_start_if_needed(ssp, rhp, do_norm);
1332}
1333
1334/**
1335 * call_srcu() - Queue a callback for invocation after an SRCU grace period
1336 * @ssp: srcu_struct in queue the callback
1337 * @rhp: structure to be used for queueing the SRCU callback.
1338 * @func: function to be invoked after the SRCU grace period
1339 *
1340 * The callback function will be invoked some time after a full SRCU
1341 * grace period elapses, in other words after all pre-existing SRCU
1342 * read-side critical sections have completed.  However, the callback
1343 * function might well execute concurrently with other SRCU read-side
1344 * critical sections that started after call_srcu() was invoked.  SRCU
1345 * read-side critical sections are delimited by srcu_read_lock() and
1346 * srcu_read_unlock(), and may be nested.
1347 *
1348 * The callback will be invoked from process context, but must nevertheless
1349 * be fast and must not block.
1350 */
1351void call_srcu(struct srcu_struct *ssp, struct rcu_head *rhp,
1352	       rcu_callback_t func)
1353{
1354	__call_srcu(ssp, rhp, func, true);
1355}
1356EXPORT_SYMBOL_GPL(call_srcu);
1357
1358/*
1359 * Helper function for synchronize_srcu() and synchronize_srcu_expedited().
1360 */
1361static void __synchronize_srcu(struct srcu_struct *ssp, bool do_norm)
1362{
1363	struct rcu_synchronize rcu;
1364
1365	srcu_lock_sync(&ssp->dep_map);
1366
1367	RCU_LOCKDEP_WARN(lockdep_is_held(ssp) ||
1368			 lock_is_held(&rcu_bh_lock_map) ||
1369			 lock_is_held(&rcu_lock_map) ||
1370			 lock_is_held(&rcu_sched_lock_map),
1371			 "Illegal synchronize_srcu() in same-type SRCU (or in RCU) read-side critical section");
1372
1373	if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
1374		return;
1375	might_sleep();
1376	check_init_srcu_struct(ssp);
1377	init_completion(&rcu.completion);
1378	init_rcu_head_on_stack(&rcu.head);
1379	__call_srcu(ssp, &rcu.head, wakeme_after_rcu, do_norm);
1380	wait_for_completion(&rcu.completion);
1381	destroy_rcu_head_on_stack(&rcu.head);
1382
1383	/*
1384	 * Make sure that later code is ordered after the SRCU grace
1385	 * period.  This pairs with the spin_lock_irq_rcu_node()
1386	 * in srcu_invoke_callbacks().  Unlike Tree RCU, this is needed
1387	 * because the current CPU might have been totally uninvolved with
1388	 * (and thus unordered against) that grace period.
1389	 */
1390	smp_mb();
1391}
1392
1393/**
1394 * synchronize_srcu_expedited - Brute-force SRCU grace period
1395 * @ssp: srcu_struct with which to synchronize.
1396 *
1397 * Wait for an SRCU grace period to elapse, but be more aggressive about
1398 * spinning rather than blocking when waiting.
1399 *
1400 * Note that synchronize_srcu_expedited() has the same deadlock and
1401 * memory-ordering properties as does synchronize_srcu().
1402 */
1403void synchronize_srcu_expedited(struct srcu_struct *ssp)
1404{
1405	__synchronize_srcu(ssp, rcu_gp_is_normal());
1406}
1407EXPORT_SYMBOL_GPL(synchronize_srcu_expedited);
1408
1409/**
1410 * synchronize_srcu - wait for prior SRCU read-side critical-section completion
1411 * @ssp: srcu_struct with which to synchronize.
1412 *
1413 * Wait for the count to drain to zero of both indexes. To avoid the
1414 * possible starvation of synchronize_srcu(), it waits for the count of
1415 * the index=((->srcu_idx & 1) ^ 1) to drain to zero at first,
1416 * and then flip the srcu_idx and wait for the count of the other index.
1417 *
1418 * Can block; must be called from process context.
1419 *
1420 * Note that it is illegal to call synchronize_srcu() from the corresponding
1421 * SRCU read-side critical section; doing so will result in deadlock.
1422 * However, it is perfectly legal to call synchronize_srcu() on one
1423 * srcu_struct from some other srcu_struct's read-side critical section,
1424 * as long as the resulting graph of srcu_structs is acyclic.
1425 *
1426 * There are memory-ordering constraints implied by synchronize_srcu().
1427 * On systems with more than one CPU, when synchronize_srcu() returns,
1428 * each CPU is guaranteed to have executed a full memory barrier since
1429 * the end of its last corresponding SRCU read-side critical section
1430 * whose beginning preceded the call to synchronize_srcu().  In addition,
1431 * each CPU having an SRCU read-side critical section that extends beyond
1432 * the return from synchronize_srcu() is guaranteed to have executed a
1433 * full memory barrier after the beginning of synchronize_srcu() and before
1434 * the beginning of that SRCU read-side critical section.  Note that these
1435 * guarantees include CPUs that are offline, idle, or executing in user mode,
1436 * as well as CPUs that are executing in the kernel.
1437 *
1438 * Furthermore, if CPU A invoked synchronize_srcu(), which returned
1439 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
1440 * to have executed a full memory barrier during the execution of
1441 * synchronize_srcu().  This guarantee applies even if CPU A and CPU B
1442 * are the same CPU, but again only if the system has more than one CPU.
1443 *
1444 * Of course, these memory-ordering guarantees apply only when
1445 * synchronize_srcu(), srcu_read_lock(), and srcu_read_unlock() are
1446 * passed the same srcu_struct structure.
1447 *
1448 * Implementation of these memory-ordering guarantees is similar to
1449 * that of synchronize_rcu().
1450 *
1451 * If SRCU is likely idle, expedite the first request.  This semantic
1452 * was provided by Classic SRCU, and is relied upon by its users, so TREE
1453 * SRCU must also provide it.  Note that detecting idleness is heuristic
1454 * and subject to both false positives and negatives.
1455 */
1456void synchronize_srcu(struct srcu_struct *ssp)
1457{
1458	if (srcu_might_be_idle(ssp) || rcu_gp_is_expedited())
1459		synchronize_srcu_expedited(ssp);
1460	else
1461		__synchronize_srcu(ssp, true);
1462}
1463EXPORT_SYMBOL_GPL(synchronize_srcu);
1464
1465/**
1466 * get_state_synchronize_srcu - Provide an end-of-grace-period cookie
1467 * @ssp: srcu_struct to provide cookie for.
1468 *
1469 * This function returns a cookie that can be passed to
1470 * poll_state_synchronize_srcu(), which will return true if a full grace
1471 * period has elapsed in the meantime.  It is the caller's responsibility
1472 * to make sure that grace period happens, for example, by invoking
1473 * call_srcu() after return from get_state_synchronize_srcu().
1474 */
1475unsigned long get_state_synchronize_srcu(struct srcu_struct *ssp)
1476{
1477	// Any prior manipulation of SRCU-protected data must happen
1478	// before the load from ->srcu_gp_seq.
1479	smp_mb();
1480	return rcu_seq_snap(&ssp->srcu_sup->srcu_gp_seq);
1481}
1482EXPORT_SYMBOL_GPL(get_state_synchronize_srcu);
1483
1484/**
1485 * start_poll_synchronize_srcu - Provide cookie and start grace period
1486 * @ssp: srcu_struct to provide cookie for.
1487 *
1488 * This function returns a cookie that can be passed to
1489 * poll_state_synchronize_srcu(), which will return true if a full grace
1490 * period has elapsed in the meantime.  Unlike get_state_synchronize_srcu(),
1491 * this function also ensures that any needed SRCU grace period will be
1492 * started.  This convenience does come at a cost in terms of CPU overhead.
1493 */
1494unsigned long start_poll_synchronize_srcu(struct srcu_struct *ssp)
1495{
1496	return srcu_gp_start_if_needed(ssp, NULL, true);
1497}
1498EXPORT_SYMBOL_GPL(start_poll_synchronize_srcu);
1499
1500/**
1501 * poll_state_synchronize_srcu - Has cookie's grace period ended?
1502 * @ssp: srcu_struct to provide cookie for.
1503 * @cookie: Return value from get_state_synchronize_srcu() or start_poll_synchronize_srcu().
1504 *
1505 * This function takes the cookie that was returned from either
1506 * get_state_synchronize_srcu() or start_poll_synchronize_srcu(), and
1507 * returns @true if an SRCU grace period elapsed since the time that the
1508 * cookie was created.
1509 *
1510 * Because cookies are finite in size, wrapping/overflow is possible.
1511 * This is more pronounced on 32-bit systems where cookies are 32 bits,
1512 * where in theory wrapping could happen in about 14 hours assuming
1513 * 25-microsecond expedited SRCU grace periods.  However, a more likely
1514 * overflow lower bound is on the order of 24 days in the case of
1515 * one-millisecond SRCU grace periods.  Of course, wrapping in a 64-bit
1516 * system requires geologic timespans, as in more than seven million years
1517 * even for expedited SRCU grace periods.
1518 *
1519 * Wrapping/overflow is much more of an issue for CONFIG_SMP=n systems
1520 * that also have CONFIG_PREEMPTION=n, which selects Tiny SRCU.  This uses
1521 * a 16-bit cookie, which rcutorture routinely wraps in a matter of a
1522 * few minutes.  If this proves to be a problem, this counter will be
1523 * expanded to the same size as for Tree SRCU.
1524 */
1525bool poll_state_synchronize_srcu(struct srcu_struct *ssp, unsigned long cookie)
1526{
1527	if (!rcu_seq_done(&ssp->srcu_sup->srcu_gp_seq, cookie))
1528		return false;
1529	// Ensure that the end of the SRCU grace period happens before
1530	// any subsequent code that the caller might execute.
1531	smp_mb(); // ^^^
1532	return true;
1533}
1534EXPORT_SYMBOL_GPL(poll_state_synchronize_srcu);
1535
1536/*
1537 * Callback function for srcu_barrier() use.
1538 */
1539static void srcu_barrier_cb(struct rcu_head *rhp)
1540{
1541	struct srcu_data *sdp;
1542	struct srcu_struct *ssp;
1543
1544	sdp = container_of(rhp, struct srcu_data, srcu_barrier_head);
1545	ssp = sdp->ssp;
1546	if (atomic_dec_and_test(&ssp->srcu_sup->srcu_barrier_cpu_cnt))
1547		complete(&ssp->srcu_sup->srcu_barrier_completion);
1548}
1549
1550/*
1551 * Enqueue an srcu_barrier() callback on the specified srcu_data
1552 * structure's ->cblist.  but only if that ->cblist already has at least one
1553 * callback enqueued.  Note that if a CPU already has callbacks enqueue,
1554 * it must have already registered the need for a future grace period,
1555 * so all we need do is enqueue a callback that will use the same grace
1556 * period as the last callback already in the queue.
1557 */
1558static void srcu_barrier_one_cpu(struct srcu_struct *ssp, struct srcu_data *sdp)
1559{
1560	spin_lock_irq_rcu_node(sdp);
1561	atomic_inc(&ssp->srcu_sup->srcu_barrier_cpu_cnt);
1562	sdp->srcu_barrier_head.func = srcu_barrier_cb;
1563	debug_rcu_head_queue(&sdp->srcu_barrier_head);
1564	if (!rcu_segcblist_entrain(&sdp->srcu_cblist,
1565				   &sdp->srcu_barrier_head)) {
1566		debug_rcu_head_unqueue(&sdp->srcu_barrier_head);
1567		atomic_dec(&ssp->srcu_sup->srcu_barrier_cpu_cnt);
1568	}
1569	spin_unlock_irq_rcu_node(sdp);
1570}
1571
1572/**
1573 * srcu_barrier - Wait until all in-flight call_srcu() callbacks complete.
1574 * @ssp: srcu_struct on which to wait for in-flight callbacks.
1575 */
1576void srcu_barrier(struct srcu_struct *ssp)
1577{
1578	int cpu;
1579	int idx;
1580	unsigned long s = rcu_seq_snap(&ssp->srcu_sup->srcu_barrier_seq);
1581
1582	check_init_srcu_struct(ssp);
1583	mutex_lock(&ssp->srcu_sup->srcu_barrier_mutex);
1584	if (rcu_seq_done(&ssp->srcu_sup->srcu_barrier_seq, s)) {
1585		smp_mb(); /* Force ordering following return. */
1586		mutex_unlock(&ssp->srcu_sup->srcu_barrier_mutex);
1587		return; /* Someone else did our work for us. */
1588	}
1589	rcu_seq_start(&ssp->srcu_sup->srcu_barrier_seq);
1590	init_completion(&ssp->srcu_sup->srcu_barrier_completion);
1591
1592	/* Initial count prevents reaching zero until all CBs are posted. */
1593	atomic_set(&ssp->srcu_sup->srcu_barrier_cpu_cnt, 1);
1594
1595	idx = __srcu_read_lock_nmisafe(ssp);
1596	if (smp_load_acquire(&ssp->srcu_sup->srcu_size_state) < SRCU_SIZE_WAIT_BARRIER)
1597		srcu_barrier_one_cpu(ssp, per_cpu_ptr(ssp->sda,	get_boot_cpu_id()));
1598	else
1599		for_each_possible_cpu(cpu)
1600			srcu_barrier_one_cpu(ssp, per_cpu_ptr(ssp->sda, cpu));
1601	__srcu_read_unlock_nmisafe(ssp, idx);
1602
1603	/* Remove the initial count, at which point reaching zero can happen. */
1604	if (atomic_dec_and_test(&ssp->srcu_sup->srcu_barrier_cpu_cnt))
1605		complete(&ssp->srcu_sup->srcu_barrier_completion);
1606	wait_for_completion(&ssp->srcu_sup->srcu_barrier_completion);
1607
1608	rcu_seq_end(&ssp->srcu_sup->srcu_barrier_seq);
1609	mutex_unlock(&ssp->srcu_sup->srcu_barrier_mutex);
1610}
1611EXPORT_SYMBOL_GPL(srcu_barrier);
1612
1613/**
1614 * srcu_batches_completed - return batches completed.
1615 * @ssp: srcu_struct on which to report batch completion.
1616 *
1617 * Report the number of batches, correlated with, but not necessarily
1618 * precisely the same as, the number of grace periods that have elapsed.
1619 */
1620unsigned long srcu_batches_completed(struct srcu_struct *ssp)
1621{
1622	return READ_ONCE(ssp->srcu_idx);
1623}
1624EXPORT_SYMBOL_GPL(srcu_batches_completed);
1625
1626/*
1627 * Core SRCU state machine.  Push state bits of ->srcu_gp_seq
1628 * to SRCU_STATE_SCAN2, and invoke srcu_gp_end() when scan has
1629 * completed in that state.
1630 */
1631static void srcu_advance_state(struct srcu_struct *ssp)
1632{
1633	int idx;
1634
1635	mutex_lock(&ssp->srcu_sup->srcu_gp_mutex);
1636
1637	/*
1638	 * Because readers might be delayed for an extended period after
1639	 * fetching ->srcu_idx for their index, at any point in time there
1640	 * might well be readers using both idx=0 and idx=1.  We therefore
1641	 * need to wait for readers to clear from both index values before
1642	 * invoking a callback.
1643	 *
1644	 * The load-acquire ensures that we see the accesses performed
1645	 * by the prior grace period.
1646	 */
1647	idx = rcu_seq_state(smp_load_acquire(&ssp->srcu_sup->srcu_gp_seq)); /* ^^^ */
1648	if (idx == SRCU_STATE_IDLE) {
1649		spin_lock_irq_rcu_node(ssp->srcu_sup);
1650		if (ULONG_CMP_GE(ssp->srcu_sup->srcu_gp_seq, ssp->srcu_sup->srcu_gp_seq_needed)) {
1651			WARN_ON_ONCE(rcu_seq_state(ssp->srcu_sup->srcu_gp_seq));
1652			spin_unlock_irq_rcu_node(ssp->srcu_sup);
1653			mutex_unlock(&ssp->srcu_sup->srcu_gp_mutex);
1654			return;
1655		}
1656		idx = rcu_seq_state(READ_ONCE(ssp->srcu_sup->srcu_gp_seq));
1657		if (idx == SRCU_STATE_IDLE)
1658			srcu_gp_start(ssp);
1659		spin_unlock_irq_rcu_node(ssp->srcu_sup);
1660		if (idx != SRCU_STATE_IDLE) {
1661			mutex_unlock(&ssp->srcu_sup->srcu_gp_mutex);
1662			return; /* Someone else started the grace period. */
1663		}
1664	}
1665
1666	if (rcu_seq_state(READ_ONCE(ssp->srcu_sup->srcu_gp_seq)) == SRCU_STATE_SCAN1) {
1667		idx = 1 ^ (ssp->srcu_idx & 1);
1668		if (!try_check_zero(ssp, idx, 1)) {
1669			mutex_unlock(&ssp->srcu_sup->srcu_gp_mutex);
1670			return; /* readers present, retry later. */
1671		}
1672		srcu_flip(ssp);
1673		spin_lock_irq_rcu_node(ssp->srcu_sup);
1674		rcu_seq_set_state(&ssp->srcu_sup->srcu_gp_seq, SRCU_STATE_SCAN2);
1675		ssp->srcu_sup->srcu_n_exp_nodelay = 0;
1676		spin_unlock_irq_rcu_node(ssp->srcu_sup);
1677	}
1678
1679	if (rcu_seq_state(READ_ONCE(ssp->srcu_sup->srcu_gp_seq)) == SRCU_STATE_SCAN2) {
1680
1681		/*
1682		 * SRCU read-side critical sections are normally short,
1683		 * so check at least twice in quick succession after a flip.
1684		 */
1685		idx = 1 ^ (ssp->srcu_idx & 1);
1686		if (!try_check_zero(ssp, idx, 2)) {
1687			mutex_unlock(&ssp->srcu_sup->srcu_gp_mutex);
1688			return; /* readers present, retry later. */
1689		}
1690		ssp->srcu_sup->srcu_n_exp_nodelay = 0;
1691		srcu_gp_end(ssp);  /* Releases ->srcu_gp_mutex. */
1692	}
1693}
1694
1695/*
1696 * Invoke a limited number of SRCU callbacks that have passed through
1697 * their grace period.  If there are more to do, SRCU will reschedule
1698 * the workqueue.  Note that needed memory barriers have been executed
1699 * in this task's context by srcu_readers_active_idx_check().
1700 */
1701static void srcu_invoke_callbacks(struct work_struct *work)
1702{
1703	long len;
1704	bool more;
1705	struct rcu_cblist ready_cbs;
1706	struct rcu_head *rhp;
1707	struct srcu_data *sdp;
1708	struct srcu_struct *ssp;
1709
1710	sdp = container_of(work, struct srcu_data, work);
1711
1712	ssp = sdp->ssp;
1713	rcu_cblist_init(&ready_cbs);
1714	spin_lock_irq_rcu_node(sdp);
1715	WARN_ON_ONCE(!rcu_segcblist_segempty(&sdp->srcu_cblist, RCU_NEXT_TAIL));
1716	rcu_segcblist_advance(&sdp->srcu_cblist,
1717			      rcu_seq_current(&ssp->srcu_sup->srcu_gp_seq));
1718	/*
1719	 * Although this function is theoretically re-entrant, concurrent
1720	 * callbacks invocation is disallowed to avoid executing an SRCU barrier
1721	 * too early.
1722	 */
1723	if (sdp->srcu_cblist_invoking ||
1724	    !rcu_segcblist_ready_cbs(&sdp->srcu_cblist)) {
1725		spin_unlock_irq_rcu_node(sdp);
1726		return;  /* Someone else on the job or nothing to do. */
1727	}
1728
1729	/* We are on the job!  Extract and invoke ready callbacks. */
1730	sdp->srcu_cblist_invoking = true;
1731	rcu_segcblist_extract_done_cbs(&sdp->srcu_cblist, &ready_cbs);
1732	len = ready_cbs.len;
1733	spin_unlock_irq_rcu_node(sdp);
1734	rhp = rcu_cblist_dequeue(&ready_cbs);
1735	for (; rhp != NULL; rhp = rcu_cblist_dequeue(&ready_cbs)) {
1736		debug_rcu_head_unqueue(rhp);
1737		debug_rcu_head_callback(rhp);
1738		local_bh_disable();
1739		rhp->func(rhp);
1740		local_bh_enable();
1741	}
1742	WARN_ON_ONCE(ready_cbs.len);
1743
1744	/*
1745	 * Update counts, accelerate new callbacks, and if needed,
1746	 * schedule another round of callback invocation.
1747	 */
1748	spin_lock_irq_rcu_node(sdp);
1749	rcu_segcblist_add_len(&sdp->srcu_cblist, -len);
1750	sdp->srcu_cblist_invoking = false;
1751	more = rcu_segcblist_ready_cbs(&sdp->srcu_cblist);
1752	spin_unlock_irq_rcu_node(sdp);
1753	/* An SRCU barrier or callbacks from previous nesting work pending */
1754	if (more)
1755		srcu_schedule_cbs_sdp(sdp, 0);
1756}
1757
1758/*
1759 * Finished one round of SRCU grace period.  Start another if there are
1760 * more SRCU callbacks queued, otherwise put SRCU into not-running state.
1761 */
1762static void srcu_reschedule(struct srcu_struct *ssp, unsigned long delay)
1763{
1764	bool pushgp = true;
1765
1766	spin_lock_irq_rcu_node(ssp->srcu_sup);
1767	if (ULONG_CMP_GE(ssp->srcu_sup->srcu_gp_seq, ssp->srcu_sup->srcu_gp_seq_needed)) {
1768		if (!WARN_ON_ONCE(rcu_seq_state(ssp->srcu_sup->srcu_gp_seq))) {
1769			/* All requests fulfilled, time to go idle. */
1770			pushgp = false;
1771		}
1772	} else if (!rcu_seq_state(ssp->srcu_sup->srcu_gp_seq)) {
1773		/* Outstanding request and no GP.  Start one. */
1774		srcu_gp_start(ssp);
1775	}
1776	spin_unlock_irq_rcu_node(ssp->srcu_sup);
1777
1778	if (pushgp)
1779		queue_delayed_work(rcu_gp_wq, &ssp->srcu_sup->work, delay);
1780}
1781
1782/*
1783 * This is the work-queue function that handles SRCU grace periods.
1784 */
1785static void process_srcu(struct work_struct *work)
1786{
1787	unsigned long curdelay;
1788	unsigned long j;
1789	struct srcu_struct *ssp;
1790	struct srcu_usage *sup;
1791
1792	sup = container_of(work, struct srcu_usage, work.work);
1793	ssp = sup->srcu_ssp;
1794
1795	srcu_advance_state(ssp);
1796	curdelay = srcu_get_delay(ssp);
1797	if (curdelay) {
1798		WRITE_ONCE(sup->reschedule_count, 0);
1799	} else {
1800		j = jiffies;
1801		if (READ_ONCE(sup->reschedule_jiffies) == j) {
1802			WRITE_ONCE(sup->reschedule_count, READ_ONCE(sup->reschedule_count) + 1);
1803			if (READ_ONCE(sup->reschedule_count) > srcu_max_nodelay)
1804				curdelay = 1;
1805		} else {
1806			WRITE_ONCE(sup->reschedule_count, 1);
1807			WRITE_ONCE(sup->reschedule_jiffies, j);
1808		}
1809	}
1810	srcu_reschedule(ssp, curdelay);
1811}
1812
1813void srcutorture_get_gp_data(enum rcutorture_type test_type,
1814			     struct srcu_struct *ssp, int *flags,
1815			     unsigned long *gp_seq)
1816{
1817	if (test_type != SRCU_FLAVOR)
1818		return;
1819	*flags = 0;
1820	*gp_seq = rcu_seq_current(&ssp->srcu_sup->srcu_gp_seq);
1821}
1822EXPORT_SYMBOL_GPL(srcutorture_get_gp_data);
1823
1824static const char * const srcu_size_state_name[] = {
1825	"SRCU_SIZE_SMALL",
1826	"SRCU_SIZE_ALLOC",
1827	"SRCU_SIZE_WAIT_BARRIER",
1828	"SRCU_SIZE_WAIT_CALL",
1829	"SRCU_SIZE_WAIT_CBS1",
1830	"SRCU_SIZE_WAIT_CBS2",
1831	"SRCU_SIZE_WAIT_CBS3",
1832	"SRCU_SIZE_WAIT_CBS4",
1833	"SRCU_SIZE_BIG",
1834	"SRCU_SIZE_???",
1835};
1836
1837void srcu_torture_stats_print(struct srcu_struct *ssp, char *tt, char *tf)
1838{
1839	int cpu;
1840	int idx;
1841	unsigned long s0 = 0, s1 = 0;
1842	int ss_state = READ_ONCE(ssp->srcu_sup->srcu_size_state);
1843	int ss_state_idx = ss_state;
1844
1845	idx = ssp->srcu_idx & 0x1;
1846	if (ss_state < 0 || ss_state >= ARRAY_SIZE(srcu_size_state_name))
1847		ss_state_idx = ARRAY_SIZE(srcu_size_state_name) - 1;
1848	pr_alert("%s%s Tree SRCU g%ld state %d (%s)",
1849		 tt, tf, rcu_seq_current(&ssp->srcu_sup->srcu_gp_seq), ss_state,
1850		 srcu_size_state_name[ss_state_idx]);
1851	if (!ssp->sda) {
1852		// Called after cleanup_srcu_struct(), perhaps.
1853		pr_cont(" No per-CPU srcu_data structures (->sda == NULL).\n");
1854	} else {
1855		pr_cont(" per-CPU(idx=%d):", idx);
1856		for_each_possible_cpu(cpu) {
1857			unsigned long l0, l1;
1858			unsigned long u0, u1;
1859			long c0, c1;
1860			struct srcu_data *sdp;
1861
1862			sdp = per_cpu_ptr(ssp->sda, cpu);
1863			u0 = data_race(atomic_long_read(&sdp->srcu_unlock_count[!idx]));
1864			u1 = data_race(atomic_long_read(&sdp->srcu_unlock_count[idx]));
1865
1866			/*
1867			 * Make sure that a lock is always counted if the corresponding
1868			 * unlock is counted.
1869			 */
1870			smp_rmb();
1871
1872			l0 = data_race(atomic_long_read(&sdp->srcu_lock_count[!idx]));
1873			l1 = data_race(atomic_long_read(&sdp->srcu_lock_count[idx]));
1874
1875			c0 = l0 - u0;
1876			c1 = l1 - u1;
1877			pr_cont(" %d(%ld,%ld %c)",
1878				cpu, c0, c1,
1879				"C."[rcu_segcblist_empty(&sdp->srcu_cblist)]);
1880			s0 += c0;
1881			s1 += c1;
1882		}
1883		pr_cont(" T(%ld,%ld)\n", s0, s1);
1884	}
1885	if (SRCU_SIZING_IS_TORTURE())
1886		srcu_transition_to_big(ssp);
1887}
1888EXPORT_SYMBOL_GPL(srcu_torture_stats_print);
1889
1890static int __init srcu_bootup_announce(void)
1891{
1892	pr_info("Hierarchical SRCU implementation.\n");
1893	if (exp_holdoff != DEFAULT_SRCU_EXP_HOLDOFF)
1894		pr_info("\tNon-default auto-expedite holdoff of %lu ns.\n", exp_holdoff);
1895	if (srcu_retry_check_delay != SRCU_DEFAULT_RETRY_CHECK_DELAY)
1896		pr_info("\tNon-default retry check delay of %lu us.\n", srcu_retry_check_delay);
1897	if (srcu_max_nodelay != SRCU_DEFAULT_MAX_NODELAY)
1898		pr_info("\tNon-default max no-delay of %lu.\n", srcu_max_nodelay);
1899	pr_info("\tMax phase no-delay instances is %lu.\n", srcu_max_nodelay_phase);
1900	return 0;
1901}
1902early_initcall(srcu_bootup_announce);
1903
1904void __init srcu_init(void)
1905{
1906	struct srcu_usage *sup;
1907
1908	/* Decide on srcu_struct-size strategy. */
1909	if (SRCU_SIZING_IS(SRCU_SIZING_AUTO)) {
1910		if (nr_cpu_ids >= big_cpu_lim) {
1911			convert_to_big = SRCU_SIZING_INIT; // Don't bother waiting for contention.
1912			pr_info("%s: Setting srcu_struct sizes to big.\n", __func__);
1913		} else {
1914			convert_to_big = SRCU_SIZING_NONE | SRCU_SIZING_CONTEND;
1915			pr_info("%s: Setting srcu_struct sizes based on contention.\n", __func__);
1916		}
1917	}
1918
1919	/*
1920	 * Once that is set, call_srcu() can follow the normal path and
1921	 * queue delayed work. This must follow RCU workqueues creation
1922	 * and timers initialization.
1923	 */
1924	srcu_init_done = true;
1925	while (!list_empty(&srcu_boot_list)) {
1926		sup = list_first_entry(&srcu_boot_list, struct srcu_usage,
1927				      work.work.entry);
1928		list_del_init(&sup->work.work.entry);
1929		if (SRCU_SIZING_IS(SRCU_SIZING_INIT) &&
1930		    sup->srcu_size_state == SRCU_SIZE_SMALL)
1931			sup->srcu_size_state = SRCU_SIZE_ALLOC;
1932		queue_work(rcu_gp_wq, &sup->work.work);
1933	}
1934}
1935
1936#ifdef CONFIG_MODULES
1937
1938/* Initialize any global-scope srcu_struct structures used by this module. */
1939static int srcu_module_coming(struct module *mod)
1940{
1941	int i;
1942	struct srcu_struct *ssp;
1943	struct srcu_struct **sspp = mod->srcu_struct_ptrs;
1944
1945	for (i = 0; i < mod->num_srcu_structs; i++) {
1946		ssp = *(sspp++);
1947		ssp->sda = alloc_percpu(struct srcu_data);
1948		if (WARN_ON_ONCE(!ssp->sda))
1949			return -ENOMEM;
1950	}
1951	return 0;
1952}
1953
1954/* Clean up any global-scope srcu_struct structures used by this module. */
1955static void srcu_module_going(struct module *mod)
1956{
1957	int i;
1958	struct srcu_struct *ssp;
1959	struct srcu_struct **sspp = mod->srcu_struct_ptrs;
1960
1961	for (i = 0; i < mod->num_srcu_structs; i++) {
1962		ssp = *(sspp++);
1963		if (!rcu_seq_state(smp_load_acquire(&ssp->srcu_sup->srcu_gp_seq_needed)) &&
1964		    !WARN_ON_ONCE(!ssp->srcu_sup->sda_is_static))
1965			cleanup_srcu_struct(ssp);
1966		if (!WARN_ON(srcu_readers_active(ssp)))
1967			free_percpu(ssp->sda);
1968	}
1969}
1970
1971/* Handle one module, either coming or going. */
1972static int srcu_module_notify(struct notifier_block *self,
1973			      unsigned long val, void *data)
1974{
1975	struct module *mod = data;
1976	int ret = 0;
1977
1978	switch (val) {
1979	case MODULE_STATE_COMING:
1980		ret = srcu_module_coming(mod);
1981		break;
1982	case MODULE_STATE_GOING:
1983		srcu_module_going(mod);
1984		break;
1985	default:
1986		break;
1987	}
1988	return ret;
1989}
1990
1991static struct notifier_block srcu_module_nb = {
1992	.notifier_call = srcu_module_notify,
1993	.priority = 0,
1994};
1995
1996static __init int init_srcu_module_notifier(void)
1997{
1998	int ret;
1999
2000	ret = register_module_notifier(&srcu_module_nb);
2001	if (ret)
2002		pr_warn("Failed to register srcu module notifier\n");
2003	return ret;
2004}
2005late_initcall(init_srcu_module_notifier);
2006
2007#endif /* #ifdef CONFIG_MODULES */