Linux Audio

Check our new training course

Loading...
v6.8
   1// SPDX-License-Identifier: GPL-2.0+
   2/*
   3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
   4 *
   5 * Copyright IBM Corporation, 2008
   6 *
   7 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
   8 *	    Manfred Spraul <manfred@colorfullife.com>
   9 *	    Paul E. McKenney <paulmck@linux.ibm.com>
  10 *
  11 * Based on the original work by Paul McKenney <paulmck@linux.ibm.com>
  12 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
  13 *
  14 * For detailed explanation of Read-Copy Update mechanism see -
  15 *	Documentation/RCU
  16 */
  17
  18#define pr_fmt(fmt) "rcu: " fmt
  19
  20#include <linux/types.h>
  21#include <linux/kernel.h>
  22#include <linux/init.h>
  23#include <linux/spinlock.h>
  24#include <linux/smp.h>
  25#include <linux/rcupdate_wait.h>
  26#include <linux/interrupt.h>
  27#include <linux/sched.h>
  28#include <linux/sched/debug.h>
  29#include <linux/nmi.h>
  30#include <linux/atomic.h>
  31#include <linux/bitops.h>
  32#include <linux/export.h>
  33#include <linux/completion.h>
  34#include <linux/kmemleak.h>
  35#include <linux/moduleparam.h>
  36#include <linux/panic.h>
  37#include <linux/panic_notifier.h>
  38#include <linux/percpu.h>
  39#include <linux/notifier.h>
  40#include <linux/cpu.h>
  41#include <linux/mutex.h>
  42#include <linux/time.h>
  43#include <linux/kernel_stat.h>
  44#include <linux/wait.h>
  45#include <linux/kthread.h>
  46#include <uapi/linux/sched/types.h>
  47#include <linux/prefetch.h>
  48#include <linux/delay.h>
  49#include <linux/random.h>
  50#include <linux/trace_events.h>
  51#include <linux/suspend.h>
  52#include <linux/ftrace.h>
  53#include <linux/tick.h>
  54#include <linux/sysrq.h>
  55#include <linux/kprobes.h>
  56#include <linux/gfp.h>
  57#include <linux/oom.h>
  58#include <linux/smpboot.h>
  59#include <linux/jiffies.h>
  60#include <linux/slab.h>
  61#include <linux/sched/isolation.h>
  62#include <linux/sched/clock.h>
  63#include <linux/vmalloc.h>
  64#include <linux/mm.h>
  65#include <linux/kasan.h>
  66#include <linux/context_tracking.h>
  67#include "../time/tick-internal.h"
  68
  69#include "tree.h"
  70#include "rcu.h"
  71
  72#ifdef MODULE_PARAM_PREFIX
  73#undef MODULE_PARAM_PREFIX
  74#endif
  75#define MODULE_PARAM_PREFIX "rcutree."
  76
  77/* Data structures. */
  78
 
 
 
 
 
 
 
  79static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = {
  80	.gpwrap = true,
 
 
  81#ifdef CONFIG_RCU_NOCB_CPU
  82	.cblist.flags = SEGCBLIST_RCU_CORE,
  83#endif
  84};
  85static struct rcu_state rcu_state = {
  86	.level = { &rcu_state.node[0] },
  87	.gp_state = RCU_GP_IDLE,
  88	.gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT,
  89	.barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex),
  90	.barrier_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.barrier_lock),
  91	.name = RCU_NAME,
  92	.abbr = RCU_ABBR,
  93	.exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
  94	.exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
  95	.ofl_lock = __ARCH_SPIN_LOCK_UNLOCKED,
  96};
  97
  98/* Dump rcu_node combining tree at boot to verify correct setup. */
  99static bool dump_tree;
 100module_param(dump_tree, bool, 0444);
 101/* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */
 102static bool use_softirq = !IS_ENABLED(CONFIG_PREEMPT_RT);
 103#ifndef CONFIG_PREEMPT_RT
 104module_param(use_softirq, bool, 0444);
 105#endif
 106/* Control rcu_node-tree auto-balancing at boot time. */
 107static bool rcu_fanout_exact;
 108module_param(rcu_fanout_exact, bool, 0444);
 109/* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
 110static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
 111module_param(rcu_fanout_leaf, int, 0444);
 112int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
 113/* Number of rcu_nodes at specified level. */
 114int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
 115int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
 116
 117/*
 118 * The rcu_scheduler_active variable is initialized to the value
 119 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
 120 * first task is spawned.  So when this variable is RCU_SCHEDULER_INACTIVE,
 121 * RCU can assume that there is but one task, allowing RCU to (for example)
 122 * optimize synchronize_rcu() to a simple barrier().  When this variable
 123 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
 124 * to detect real grace periods.  This variable is also used to suppress
 125 * boot-time false positives from lockdep-RCU error checking.  Finally, it
 126 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
 127 * is fully initialized, including all of its kthreads having been spawned.
 128 */
 129int rcu_scheduler_active __read_mostly;
 130EXPORT_SYMBOL_GPL(rcu_scheduler_active);
 131
 132/*
 133 * The rcu_scheduler_fully_active variable transitions from zero to one
 134 * during the early_initcall() processing, which is after the scheduler
 135 * is capable of creating new tasks.  So RCU processing (for example,
 136 * creating tasks for RCU priority boosting) must be delayed until after
 137 * rcu_scheduler_fully_active transitions from zero to one.  We also
 138 * currently delay invocation of any RCU callbacks until after this point.
 139 *
 140 * It might later prove better for people registering RCU callbacks during
 141 * early boot to take responsibility for these callbacks, but one step at
 142 * a time.
 143 */
 144static int rcu_scheduler_fully_active __read_mostly;
 145
 146static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
 147			      unsigned long gps, unsigned long flags);
 
 
 148static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
 149static void invoke_rcu_core(void);
 150static void rcu_report_exp_rdp(struct rcu_data *rdp);
 151static void sync_sched_exp_online_cleanup(int cpu);
 152static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp);
 153static bool rcu_rdp_is_offloaded(struct rcu_data *rdp);
 154static bool rcu_rdp_cpu_online(struct rcu_data *rdp);
 155static bool rcu_init_invoked(void);
 156static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
 157static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
 158
 159/*
 160 * rcuc/rcub/rcuop kthread realtime priority. The "rcuop"
 161 * real-time priority(enabling/disabling) is controlled by
 162 * the extra CONFIG_RCU_NOCB_CPU_CB_BOOST configuration.
 163 */
 164static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
 165module_param(kthread_prio, int, 0444);
 166
 167/* Delay in jiffies for grace-period initialization delays, debug only. */
 168
 169static int gp_preinit_delay;
 170module_param(gp_preinit_delay, int, 0444);
 171static int gp_init_delay;
 172module_param(gp_init_delay, int, 0444);
 173static int gp_cleanup_delay;
 174module_param(gp_cleanup_delay, int, 0444);
 175
 176// Add delay to rcu_read_unlock() for strict grace periods.
 177static int rcu_unlock_delay;
 178#ifdef CONFIG_RCU_STRICT_GRACE_PERIOD
 179module_param(rcu_unlock_delay, int, 0444);
 180#endif
 181
 182/*
 183 * This rcu parameter is runtime-read-only. It reflects
 184 * a minimum allowed number of objects which can be cached
 185 * per-CPU. Object size is equal to one page. This value
 186 * can be changed at boot time.
 187 */
 188static int rcu_min_cached_objs = 5;
 189module_param(rcu_min_cached_objs, int, 0444);
 190
 191// A page shrinker can ask for pages to be freed to make them
 192// available for other parts of the system. This usually happens
 193// under low memory conditions, and in that case we should also
 194// defer page-cache filling for a short time period.
 195//
 196// The default value is 5 seconds, which is long enough to reduce
 197// interference with the shrinker while it asks other systems to
 198// drain their caches.
 199static int rcu_delay_page_cache_fill_msec = 5000;
 200module_param(rcu_delay_page_cache_fill_msec, int, 0444);
 201
 202/* Retrieve RCU kthreads priority for rcutorture */
 203int rcu_get_gp_kthreads_prio(void)
 204{
 205	return kthread_prio;
 206}
 207EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio);
 208
 209/*
 210 * Number of grace periods between delays, normalized by the duration of
 211 * the delay.  The longer the delay, the more the grace periods between
 212 * each delay.  The reason for this normalization is that it means that,
 213 * for non-zero delays, the overall slowdown of grace periods is constant
 214 * regardless of the duration of the delay.  This arrangement balances
 215 * the need for long delays to increase some race probabilities with the
 216 * need for fast grace periods to increase other race probabilities.
 217 */
 218#define PER_RCU_NODE_PERIOD 3	/* Number of grace periods between delays for debugging. */
 219
 220/*
 
 
 
 
 
 
 
 
 
 
 
 221 * Return true if an RCU grace period is in progress.  The READ_ONCE()s
 222 * permit this function to be invoked without holding the root rcu_node
 223 * structure's ->lock, but of course results can be subject to change.
 224 */
 225static int rcu_gp_in_progress(void)
 226{
 227	return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq));
 228}
 229
 230/*
 231 * Return the number of callbacks queued on the specified CPU.
 232 * Handles both the nocbs and normal cases.
 233 */
 234static long rcu_get_n_cbs_cpu(int cpu)
 235{
 236	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 237
 238	if (rcu_segcblist_is_enabled(&rdp->cblist))
 239		return rcu_segcblist_n_cbs(&rdp->cblist);
 240	return 0;
 241}
 242
 243void rcu_softirq_qs(void)
 244{
 245	rcu_qs();
 246	rcu_preempt_deferred_qs(current);
 247	rcu_tasks_qs(current, false);
 248}
 249
 250/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 251 * Reset the current CPU's ->dynticks counter to indicate that the
 252 * newly onlined CPU is no longer in an extended quiescent state.
 253 * This will either leave the counter unchanged, or increment it
 254 * to the next non-quiescent value.
 255 *
 256 * The non-atomic test/increment sequence works because the upper bits
 257 * of the ->dynticks counter are manipulated only by the corresponding CPU,
 258 * or when the corresponding CPU is offline.
 259 */
 260static void rcu_dynticks_eqs_online(void)
 261{
 262	if (ct_dynticks() & RCU_DYNTICKS_IDX)
 
 
 263		return;
 264	ct_state_inc(RCU_DYNTICKS_IDX);
 
 
 
 
 
 
 
 
 
 
 
 
 265}
 266
 267/*
 268 * Snapshot the ->dynticks counter with full ordering so as to allow
 269 * stable comparison of this counter with past and future snapshots.
 270 */
 271static int rcu_dynticks_snap(int cpu)
 272{
 273	smp_mb();  // Fundamental RCU ordering guarantee.
 274	return ct_dynticks_cpu_acquire(cpu);
 
 275}
 276
 277/*
 278 * Return true if the snapshot returned from rcu_dynticks_snap()
 279 * indicates that RCU is in an extended quiescent state.
 280 */
 281static bool rcu_dynticks_in_eqs(int snap)
 282{
 283	return !(snap & RCU_DYNTICKS_IDX);
 
 
 
 
 
 
 
 
 284}
 285
 286/*
 287 * Return true if the CPU corresponding to the specified rcu_data
 288 * structure has spent some time in an extended quiescent state since
 289 * rcu_dynticks_snap() returned the specified snapshot.
 290 */
 291static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap)
 292{
 293	return snap != rcu_dynticks_snap(rdp->cpu);
 294}
 295
 296/*
 297 * Return true if the referenced integer is zero while the specified
 298 * CPU remains within a single extended quiescent state.
 299 */
 300bool rcu_dynticks_zero_in_eqs(int cpu, int *vp)
 301{
 
 302	int snap;
 303
 304	// If not quiescent, force back to earlier extended quiescent state.
 305	snap = ct_dynticks_cpu(cpu) & ~RCU_DYNTICKS_IDX;
 
 
 306	smp_rmb(); // Order ->dynticks and *vp reads.
 307	if (READ_ONCE(*vp))
 308		return false;  // Non-zero, so report failure;
 309	smp_rmb(); // Order *vp read and ->dynticks re-read.
 310
 311	// If still in the same extended quiescent state, we are good!
 312	return snap == ct_dynticks_cpu(cpu);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 313}
 314
 315/*
 316 * Let the RCU core know that this CPU has gone through the scheduler,
 317 * which is a quiescent state.  This is called when the need for a
 318 * quiescent state is urgent, so we burn an atomic operation and full
 319 * memory barriers to let the RCU core know about it, regardless of what
 320 * this CPU might (or might not) do in the near future.
 321 *
 322 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
 323 *
 324 * The caller must have disabled interrupts and must not be idle.
 325 */
 326notrace void rcu_momentary_dyntick_idle(void)
 327{
 328	int seq;
 329
 330	raw_cpu_write(rcu_data.rcu_need_heavy_qs, false);
 331	seq = ct_state_inc(2 * RCU_DYNTICKS_IDX);
 
 332	/* It is illegal to call this from idle state. */
 333	WARN_ON_ONCE(!(seq & RCU_DYNTICKS_IDX));
 334	rcu_preempt_deferred_qs(current);
 335}
 336EXPORT_SYMBOL_GPL(rcu_momentary_dyntick_idle);
 337
 338/**
 339 * rcu_is_cpu_rrupt_from_idle - see if 'interrupted' from idle
 340 *
 341 * If the current CPU is idle and running at a first-level (not nested)
 342 * interrupt, or directly, from idle, return true.
 343 *
 344 * The caller must have at least disabled IRQs.
 345 */
 346static int rcu_is_cpu_rrupt_from_idle(void)
 347{
 348	long nesting;
 349
 350	/*
 351	 * Usually called from the tick; but also used from smp_function_call()
 352	 * for expedited grace periods. This latter can result in running from
 353	 * the idle task, instead of an actual IPI.
 354	 */
 355	lockdep_assert_irqs_disabled();
 356
 357	/* Check for counter underflows */
 358	RCU_LOCKDEP_WARN(ct_dynticks_nesting() < 0,
 359			 "RCU dynticks_nesting counter underflow!");
 360	RCU_LOCKDEP_WARN(ct_dynticks_nmi_nesting() <= 0,
 361			 "RCU dynticks_nmi_nesting counter underflow/zero!");
 362
 363	/* Are we at first interrupt nesting level? */
 364	nesting = ct_dynticks_nmi_nesting();
 365	if (nesting > 1)
 366		return false;
 367
 368	/*
 369	 * If we're not in an interrupt, we must be in the idle task!
 370	 */
 371	WARN_ON_ONCE(!nesting && !is_idle_task(current));
 372
 373	/* Does CPU appear to be idle from an RCU standpoint? */
 374	return ct_dynticks_nesting() == 0;
 375}
 376
 377#define DEFAULT_RCU_BLIMIT (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 1000 : 10)
 378				// Maximum callbacks per rcu_do_batch ...
 379#define DEFAULT_MAX_RCU_BLIMIT 10000 // ... even during callback flood.
 380static long blimit = DEFAULT_RCU_BLIMIT;
 381#define DEFAULT_RCU_QHIMARK 10000 // If this many pending, ignore blimit.
 382static long qhimark = DEFAULT_RCU_QHIMARK;
 383#define DEFAULT_RCU_QLOMARK 100   // Once only this many pending, use blimit.
 384static long qlowmark = DEFAULT_RCU_QLOMARK;
 385#define DEFAULT_RCU_QOVLD_MULT 2
 386#define DEFAULT_RCU_QOVLD (DEFAULT_RCU_QOVLD_MULT * DEFAULT_RCU_QHIMARK)
 387static long qovld = DEFAULT_RCU_QOVLD; // If this many pending, hammer QS.
 388static long qovld_calc = -1;	  // No pre-initialization lock acquisitions!
 389
 390module_param(blimit, long, 0444);
 391module_param(qhimark, long, 0444);
 392module_param(qlowmark, long, 0444);
 393module_param(qovld, long, 0444);
 394
 395static ulong jiffies_till_first_fqs = IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 0 : ULONG_MAX;
 396static ulong jiffies_till_next_fqs = ULONG_MAX;
 397static bool rcu_kick_kthreads;
 398static int rcu_divisor = 7;
 399module_param(rcu_divisor, int, 0644);
 400
 401/* Force an exit from rcu_do_batch() after 3 milliseconds. */
 402static long rcu_resched_ns = 3 * NSEC_PER_MSEC;
 403module_param(rcu_resched_ns, long, 0644);
 404
 405/*
 406 * How long the grace period must be before we start recruiting
 407 * quiescent-state help from rcu_note_context_switch().
 408 */
 409static ulong jiffies_till_sched_qs = ULONG_MAX;
 410module_param(jiffies_till_sched_qs, ulong, 0444);
 411static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */
 412module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
 413
 414/*
 415 * Make sure that we give the grace-period kthread time to detect any
 416 * idle CPUs before taking active measures to force quiescent states.
 417 * However, don't go below 100 milliseconds, adjusted upwards for really
 418 * large systems.
 419 */
 420static void adjust_jiffies_till_sched_qs(void)
 421{
 422	unsigned long j;
 423
 424	/* If jiffies_till_sched_qs was specified, respect the request. */
 425	if (jiffies_till_sched_qs != ULONG_MAX) {
 426		WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
 427		return;
 428	}
 429	/* Otherwise, set to third fqs scan, but bound below on large system. */
 430	j = READ_ONCE(jiffies_till_first_fqs) +
 431		      2 * READ_ONCE(jiffies_till_next_fqs);
 432	if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
 433		j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
 434	pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
 435	WRITE_ONCE(jiffies_to_sched_qs, j);
 436}
 437
 438static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
 439{
 440	ulong j;
 441	int ret = kstrtoul(val, 0, &j);
 442
 443	if (!ret) {
 444		WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
 445		adjust_jiffies_till_sched_qs();
 446	}
 447	return ret;
 448}
 449
 450static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
 451{
 452	ulong j;
 453	int ret = kstrtoul(val, 0, &j);
 454
 455	if (!ret) {
 456		WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
 457		adjust_jiffies_till_sched_qs();
 458	}
 459	return ret;
 460}
 461
 462static const struct kernel_param_ops first_fqs_jiffies_ops = {
 463	.set = param_set_first_fqs_jiffies,
 464	.get = param_get_ulong,
 465};
 466
 467static const struct kernel_param_ops next_fqs_jiffies_ops = {
 468	.set = param_set_next_fqs_jiffies,
 469	.get = param_get_ulong,
 470};
 471
 472module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
 473module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
 474module_param(rcu_kick_kthreads, bool, 0644);
 475
 476static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
 477static int rcu_pending(int user);
 478
 479/*
 480 * Return the number of RCU GPs completed thus far for debug & stats.
 481 */
 482unsigned long rcu_get_gp_seq(void)
 483{
 484	return READ_ONCE(rcu_state.gp_seq);
 485}
 486EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
 487
 488/*
 489 * Return the number of RCU expedited batches completed thus far for
 490 * debug & stats.  Odd numbers mean that a batch is in progress, even
 491 * numbers mean idle.  The value returned will thus be roughly double
 492 * the cumulative batches since boot.
 493 */
 494unsigned long rcu_exp_batches_completed(void)
 495{
 496	return rcu_state.expedited_sequence;
 497}
 498EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
 499
 500/*
 501 * Return the root node of the rcu_state structure.
 502 */
 503static struct rcu_node *rcu_get_root(void)
 504{
 505	return &rcu_state.node[0];
 506}
 507
 508/*
 509 * Send along grace-period-related data for rcutorture diagnostics.
 510 */
 511void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
 512			    unsigned long *gp_seq)
 513{
 514	switch (test_type) {
 515	case RCU_FLAVOR:
 516		*flags = READ_ONCE(rcu_state.gp_flags);
 517		*gp_seq = rcu_seq_current(&rcu_state.gp_seq);
 518		break;
 519	default:
 520		break;
 521	}
 522}
 523EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
 524
 525#if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 526/*
 527 * An empty function that will trigger a reschedule on
 528 * IRQ tail once IRQs get re-enabled on userspace/guest resume.
 529 */
 530static void late_wakeup_func(struct irq_work *work)
 531{
 532}
 533
 534static DEFINE_PER_CPU(struct irq_work, late_wakeup_work) =
 535	IRQ_WORK_INIT(late_wakeup_func);
 536
 537/*
 538 * If either:
 539 *
 540 * 1) the task is about to enter in guest mode and $ARCH doesn't support KVM generic work
 541 * 2) the task is about to enter in user mode and $ARCH doesn't support generic entry.
 542 *
 543 * In these cases the late RCU wake ups aren't supported in the resched loops and our
 544 * last resort is to fire a local irq_work that will trigger a reschedule once IRQs
 545 * get re-enabled again.
 546 */
 547noinstr void rcu_irq_work_resched(void)
 548{
 549	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 550
 551	if (IS_ENABLED(CONFIG_GENERIC_ENTRY) && !(current->flags & PF_VCPU))
 552		return;
 553
 554	if (IS_ENABLED(CONFIG_KVM_XFER_TO_GUEST_WORK) && (current->flags & PF_VCPU))
 555		return;
 556
 557	instrumentation_begin();
 558	if (do_nocb_deferred_wakeup(rdp) && need_resched()) {
 559		irq_work_queue(this_cpu_ptr(&late_wakeup_work));
 560	}
 561	instrumentation_end();
 562}
 563#endif /* #if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK)) */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 564
 565#ifdef CONFIG_PROVE_RCU
 566/**
 567 * rcu_irq_exit_check_preempt - Validate that scheduling is possible
 568 */
 569void rcu_irq_exit_check_preempt(void)
 570{
 571	lockdep_assert_irqs_disabled();
 572
 573	RCU_LOCKDEP_WARN(ct_dynticks_nesting() <= 0,
 574			 "RCU dynticks_nesting counter underflow/zero!");
 575	RCU_LOCKDEP_WARN(ct_dynticks_nmi_nesting() !=
 576			 DYNTICK_IRQ_NONIDLE,
 577			 "Bad RCU  dynticks_nmi_nesting counter\n");
 578	RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
 579			 "RCU in extended quiescent state!");
 580}
 581#endif /* #ifdef CONFIG_PROVE_RCU */
 582
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 583#ifdef CONFIG_NO_HZ_FULL
 584/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 585 * __rcu_irq_enter_check_tick - Enable scheduler tick on CPU if RCU needs it.
 586 *
 587 * The scheduler tick is not normally enabled when CPUs enter the kernel
 588 * from nohz_full userspace execution.  After all, nohz_full userspace
 589 * execution is an RCU quiescent state and the time executing in the kernel
 590 * is quite short.  Except of course when it isn't.  And it is not hard to
 591 * cause a large system to spend tens of seconds or even minutes looping
 592 * in the kernel, which can cause a number of problems, include RCU CPU
 593 * stall warnings.
 594 *
 595 * Therefore, if a nohz_full CPU fails to report a quiescent state
 596 * in a timely manner, the RCU grace-period kthread sets that CPU's
 597 * ->rcu_urgent_qs flag with the expectation that the next interrupt or
 598 * exception will invoke this function, which will turn on the scheduler
 599 * tick, which will enable RCU to detect that CPU's quiescent states,
 600 * for example, due to cond_resched() calls in CONFIG_PREEMPT=n kernels.
 601 * The tick will be disabled once a quiescent state is reported for
 602 * this CPU.
 603 *
 604 * Of course, in carefully tuned systems, there might never be an
 605 * interrupt or exception.  In that case, the RCU grace-period kthread
 606 * will eventually cause one to happen.  However, in less carefully
 607 * controlled environments, this function allows RCU to get what it
 608 * needs without creating otherwise useless interruptions.
 609 */
 610void __rcu_irq_enter_check_tick(void)
 611{
 612	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 613
 614	// If we're here from NMI there's nothing to do.
 615	if (in_nmi())
 616		return;
 617
 618	RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
 619			 "Illegal rcu_irq_enter_check_tick() from extended quiescent state");
 620
 621	if (!tick_nohz_full_cpu(rdp->cpu) ||
 622	    !READ_ONCE(rdp->rcu_urgent_qs) ||
 623	    READ_ONCE(rdp->rcu_forced_tick)) {
 624		// RCU doesn't need nohz_full help from this CPU, or it is
 625		// already getting that help.
 626		return;
 627	}
 628
 629	// We get here only when not in an extended quiescent state and
 630	// from interrupts (as opposed to NMIs).  Therefore, (1) RCU is
 631	// already watching and (2) The fact that we are in an interrupt
 632	// handler and that the rcu_node lock is an irq-disabled lock
 633	// prevents self-deadlock.  So we can safely recheck under the lock.
 634	// Note that the nohz_full state currently cannot change.
 635	raw_spin_lock_rcu_node(rdp->mynode);
 636	if (READ_ONCE(rdp->rcu_urgent_qs) && !rdp->rcu_forced_tick) {
 637		// A nohz_full CPU is in the kernel and RCU needs a
 638		// quiescent state.  Turn on the tick!
 639		WRITE_ONCE(rdp->rcu_forced_tick, true);
 640		tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
 641	}
 642	raw_spin_unlock_rcu_node(rdp->mynode);
 643}
 644NOKPROBE_SYMBOL(__rcu_irq_enter_check_tick);
 645#endif /* CONFIG_NO_HZ_FULL */
 646
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 647/*
 648 * Check to see if any future non-offloaded RCU-related work will need
 649 * to be done by the current CPU, even if none need be done immediately,
 650 * returning 1 if so.  This function is part of the RCU implementation;
 651 * it is -not- an exported member of the RCU API.  This is used by
 652 * the idle-entry code to figure out whether it is safe to disable the
 653 * scheduler-clock interrupt.
 654 *
 655 * Just check whether or not this CPU has non-offloaded RCU callbacks
 656 * queued.
 657 */
 658int rcu_needs_cpu(void)
 659{
 660	return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) &&
 661		!rcu_rdp_is_offloaded(this_cpu_ptr(&rcu_data));
 
 
 
 662}
 663
 664/*
 665 * If any sort of urgency was applied to the current CPU (for example,
 666 * the scheduler-clock interrupt was enabled on a nohz_full CPU) in order
 667 * to get to a quiescent state, disable it.
 668 */
 669static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp)
 670{
 671	raw_lockdep_assert_held_rcu_node(rdp->mynode);
 672	WRITE_ONCE(rdp->rcu_urgent_qs, false);
 673	WRITE_ONCE(rdp->rcu_need_heavy_qs, false);
 674	if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) {
 675		tick_dep_clear_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
 676		WRITE_ONCE(rdp->rcu_forced_tick, false);
 677	}
 678}
 679
 680/**
 681 * rcu_is_watching - RCU read-side critical sections permitted on current CPU?
 682 *
 683 * Return @true if RCU is watching the running CPU and @false otherwise.
 684 * An @true return means that this CPU can safely enter RCU read-side
 685 * critical sections.
 686 *
 687 * Although calls to rcu_is_watching() from most parts of the kernel
 688 * will return @true, there are important exceptions.  For example, if the
 689 * current CPU is deep within its idle loop, in kernel entry/exit code,
 690 * or offline, rcu_is_watching() will return @false.
 691 *
 692 * Make notrace because it can be called by the internal functions of
 693 * ftrace, and making this notrace removes unnecessary recursion calls.
 694 */
 695notrace bool rcu_is_watching(void)
 696{
 697	bool ret;
 698
 699	preempt_disable_notrace();
 700	ret = !rcu_dynticks_curr_cpu_in_eqs();
 701	preempt_enable_notrace();
 702	return ret;
 703}
 704EXPORT_SYMBOL_GPL(rcu_is_watching);
 705
 706/*
 707 * If a holdout task is actually running, request an urgent quiescent
 708 * state from its CPU.  This is unsynchronized, so migrations can cause
 709 * the request to go to the wrong CPU.  Which is OK, all that will happen
 710 * is that the CPU's next context switch will be a bit slower and next
 711 * time around this task will generate another request.
 712 */
 713void rcu_request_urgent_qs_task(struct task_struct *t)
 714{
 715	int cpu;
 716
 717	barrier();
 718	cpu = task_cpu(t);
 719	if (!task_curr(t))
 720		return; /* This task is not running on that CPU. */
 721	smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
 722}
 723
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 724/*
 725 * When trying to report a quiescent state on behalf of some other CPU,
 726 * it is our responsibility to check for and handle potential overflow
 727 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
 728 * After all, the CPU might be in deep idle state, and thus executing no
 729 * code whatsoever.
 730 */
 731static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
 732{
 733	raw_lockdep_assert_held_rcu_node(rnp);
 734	if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
 735			 rnp->gp_seq))
 736		WRITE_ONCE(rdp->gpwrap, true);
 737	if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
 738		rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
 739}
 740
 741/*
 742 * Snapshot the specified CPU's dynticks counter so that we can later
 743 * credit them with an implicit quiescent state.  Return 1 if this CPU
 744 * is in dynticks idle mode, which is an extended quiescent state.
 745 */
 746static int dyntick_save_progress_counter(struct rcu_data *rdp)
 747{
 748	rdp->dynticks_snap = rcu_dynticks_snap(rdp->cpu);
 749	if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
 750		trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
 751		rcu_gpnum_ovf(rdp->mynode, rdp);
 752		return 1;
 753	}
 754	return 0;
 755}
 756
 757/*
 758 * Returns positive if the specified CPU has passed through a quiescent state
 759 * by virtue of being in or having passed through an dynticks idle state since
 760 * the last call to dyntick_save_progress_counter() for this same CPU, or by
 761 * virtue of having been offline.
 762 *
 763 * Returns negative if the specified CPU needs a force resched.
 764 *
 765 * Returns zero otherwise.
 766 */
 767static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
 768{
 769	unsigned long jtsq;
 770	int ret = 0;
 
 771	struct rcu_node *rnp = rdp->mynode;
 772
 773	/*
 774	 * If the CPU passed through or entered a dynticks idle phase with
 775	 * no active irq/NMI handlers, then we can safely pretend that the CPU
 776	 * already acknowledged the request to pass through a quiescent
 777	 * state.  Either way, that CPU cannot possibly be in an RCU
 778	 * read-side critical section that started before the beginning
 779	 * of the current RCU grace period.
 780	 */
 781	if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
 782		trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
 783		rcu_gpnum_ovf(rnp, rdp);
 784		return 1;
 785	}
 786
 787	/*
 788	 * Complain if a CPU that is considered to be offline from RCU's
 789	 * perspective has not yet reported a quiescent state.  After all,
 790	 * the offline CPU should have reported a quiescent state during
 791	 * the CPU-offline process, or, failing that, by rcu_gp_init()
 792	 * if it ran concurrently with either the CPU going offline or the
 793	 * last task on a leaf rcu_node structure exiting its RCU read-side
 794	 * critical section while all CPUs corresponding to that structure
 795	 * are offline.  This added warning detects bugs in any of these
 796	 * code paths.
 797	 *
 798	 * The rcu_node structure's ->lock is held here, which excludes
 799	 * the relevant portions the CPU-hotplug code, the grace-period
 800	 * initialization code, and the rcu_read_unlock() code paths.
 801	 *
 802	 * For more detail, please refer to the "Hotplug CPU" section
 803	 * of RCU's Requirements documentation.
 804	 */
 805	if (WARN_ON_ONCE(!rcu_rdp_cpu_online(rdp))) {
 
 806		struct rcu_node *rnp1;
 807
 808		pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
 809			__func__, rnp->grplo, rnp->grphi, rnp->level,
 810			(long)rnp->gp_seq, (long)rnp->completedqs);
 811		for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
 812			pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
 813				__func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
 
 814		pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
 815			__func__, rdp->cpu, ".o"[rcu_rdp_cpu_online(rdp)],
 816			(long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
 817			(long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
 818		return 1; /* Break things loose after complaining. */
 819	}
 820
 821	/*
 822	 * A CPU running for an extended time within the kernel can
 823	 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
 824	 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
 825	 * both .rcu_need_heavy_qs and .rcu_urgent_qs.  Note that the
 826	 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
 827	 * variable are safe because the assignments are repeated if this
 828	 * CPU failed to pass through a quiescent state.  This code
 829	 * also checks .jiffies_resched in case jiffies_to_sched_qs
 830	 * is set way high.
 831	 */
 832	jtsq = READ_ONCE(jiffies_to_sched_qs);
 833	if (!READ_ONCE(rdp->rcu_need_heavy_qs) &&
 
 
 834	    (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
 835	     time_after(jiffies, rcu_state.jiffies_resched) ||
 836	     rcu_state.cbovld)) {
 837		WRITE_ONCE(rdp->rcu_need_heavy_qs, true);
 838		/* Store rcu_need_heavy_qs before rcu_urgent_qs. */
 839		smp_store_release(&rdp->rcu_urgent_qs, true);
 840	} else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
 841		WRITE_ONCE(rdp->rcu_urgent_qs, true);
 842	}
 843
 844	/*
 845	 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
 846	 * The above code handles this, but only for straight cond_resched().
 847	 * And some in-kernel loops check need_resched() before calling
 848	 * cond_resched(), which defeats the above code for CPUs that are
 849	 * running in-kernel with scheduling-clock interrupts disabled.
 850	 * So hit them over the head with the resched_cpu() hammer!
 851	 */
 852	if (tick_nohz_full_cpu(rdp->cpu) &&
 853	    (time_after(jiffies, READ_ONCE(rdp->last_fqs_resched) + jtsq * 3) ||
 854	     rcu_state.cbovld)) {
 855		WRITE_ONCE(rdp->rcu_urgent_qs, true);
 
 856		WRITE_ONCE(rdp->last_fqs_resched, jiffies);
 857		ret = -1;
 858	}
 859
 860	/*
 861	 * If more than halfway to RCU CPU stall-warning time, invoke
 862	 * resched_cpu() more frequently to try to loosen things up a bit.
 863	 * Also check to see if the CPU is getting hammered with interrupts,
 864	 * but only once per grace period, just to keep the IPIs down to
 865	 * a dull roar.
 866	 */
 867	if (time_after(jiffies, rcu_state.jiffies_resched)) {
 868		if (time_after(jiffies,
 869			       READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
 
 870			WRITE_ONCE(rdp->last_fqs_resched, jiffies);
 871			ret = -1;
 872		}
 873		if (IS_ENABLED(CONFIG_IRQ_WORK) &&
 874		    !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
 875		    (rnp->ffmask & rdp->grpmask)) {
 876			rdp->rcu_iw_pending = true;
 877			rdp->rcu_iw_gp_seq = rnp->gp_seq;
 878			irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
 879		}
 880
 881		if (rcu_cpu_stall_cputime && rdp->snap_record.gp_seq != rdp->gp_seq) {
 882			int cpu = rdp->cpu;
 883			struct rcu_snap_record *rsrp;
 884			struct kernel_cpustat *kcsp;
 885
 886			kcsp = &kcpustat_cpu(cpu);
 887
 888			rsrp = &rdp->snap_record;
 889			rsrp->cputime_irq     = kcpustat_field(kcsp, CPUTIME_IRQ, cpu);
 890			rsrp->cputime_softirq = kcpustat_field(kcsp, CPUTIME_SOFTIRQ, cpu);
 891			rsrp->cputime_system  = kcpustat_field(kcsp, CPUTIME_SYSTEM, cpu);
 892			rsrp->nr_hardirqs = kstat_cpu_irqs_sum(rdp->cpu);
 893			rsrp->nr_softirqs = kstat_cpu_softirqs_sum(rdp->cpu);
 894			rsrp->nr_csw = nr_context_switches_cpu(rdp->cpu);
 895			rsrp->jiffies = jiffies;
 896			rsrp->gp_seq = rdp->gp_seq;
 897		}
 898	}
 899
 900	return ret;
 901}
 902
 903/* Trace-event wrapper function for trace_rcu_future_grace_period.  */
 904static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
 905			      unsigned long gp_seq_req, const char *s)
 906{
 907	trace_rcu_future_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
 908				      gp_seq_req, rnp->level,
 909				      rnp->grplo, rnp->grphi, s);
 910}
 911
 912/*
 913 * rcu_start_this_gp - Request the start of a particular grace period
 914 * @rnp_start: The leaf node of the CPU from which to start.
 915 * @rdp: The rcu_data corresponding to the CPU from which to start.
 916 * @gp_seq_req: The gp_seq of the grace period to start.
 917 *
 918 * Start the specified grace period, as needed to handle newly arrived
 919 * callbacks.  The required future grace periods are recorded in each
 920 * rcu_node structure's ->gp_seq_needed field.  Returns true if there
 921 * is reason to awaken the grace-period kthread.
 922 *
 923 * The caller must hold the specified rcu_node structure's ->lock, which
 924 * is why the caller is responsible for waking the grace-period kthread.
 925 *
 926 * Returns true if the GP thread needs to be awakened else false.
 927 */
 928static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
 929			      unsigned long gp_seq_req)
 930{
 931	bool ret = false;
 932	struct rcu_node *rnp;
 933
 934	/*
 935	 * Use funnel locking to either acquire the root rcu_node
 936	 * structure's lock or bail out if the need for this grace period
 937	 * has already been recorded -- or if that grace period has in
 938	 * fact already started.  If there is already a grace period in
 939	 * progress in a non-leaf node, no recording is needed because the
 940	 * end of the grace period will scan the leaf rcu_node structures.
 941	 * Note that rnp_start->lock must not be released.
 942	 */
 943	raw_lockdep_assert_held_rcu_node(rnp_start);
 944	trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
 945	for (rnp = rnp_start; 1; rnp = rnp->parent) {
 946		if (rnp != rnp_start)
 947			raw_spin_lock_rcu_node(rnp);
 948		if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
 949		    rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
 950		    (rnp != rnp_start &&
 951		     rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
 952			trace_rcu_this_gp(rnp, rdp, gp_seq_req,
 953					  TPS("Prestarted"));
 954			goto unlock_out;
 955		}
 956		WRITE_ONCE(rnp->gp_seq_needed, gp_seq_req);
 957		if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
 958			/*
 959			 * We just marked the leaf or internal node, and a
 960			 * grace period is in progress, which means that
 961			 * rcu_gp_cleanup() will see the marking.  Bail to
 962			 * reduce contention.
 963			 */
 964			trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
 965					  TPS("Startedleaf"));
 966			goto unlock_out;
 967		}
 968		if (rnp != rnp_start && rnp->parent != NULL)
 969			raw_spin_unlock_rcu_node(rnp);
 970		if (!rnp->parent)
 971			break;  /* At root, and perhaps also leaf. */
 972	}
 973
 974	/* If GP already in progress, just leave, otherwise start one. */
 975	if (rcu_gp_in_progress()) {
 976		trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
 977		goto unlock_out;
 978	}
 979	trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
 980	WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
 981	WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
 982	if (!READ_ONCE(rcu_state.gp_kthread)) {
 983		trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
 984		goto unlock_out;
 985	}
 986	trace_rcu_grace_period(rcu_state.name, data_race(rcu_state.gp_seq), TPS("newreq"));
 987	ret = true;  /* Caller must wake GP kthread. */
 988unlock_out:
 989	/* Push furthest requested GP to leaf node and rcu_data structure. */
 990	if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
 991		WRITE_ONCE(rnp_start->gp_seq_needed, rnp->gp_seq_needed);
 992		WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
 993	}
 994	if (rnp != rnp_start)
 995		raw_spin_unlock_rcu_node(rnp);
 996	return ret;
 997}
 998
 999/*
1000 * Clean up any old requests for the just-ended grace period.  Also return
1001 * whether any additional grace periods have been requested.
1002 */
1003static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
1004{
1005	bool needmore;
1006	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1007
1008	needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1009	if (!needmore)
1010		rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1011	trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1012			  needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1013	return needmore;
1014}
1015
1016static void swake_up_one_online_ipi(void *arg)
1017{
1018	struct swait_queue_head *wqh = arg;
1019
1020	swake_up_one(wqh);
1021}
1022
1023static void swake_up_one_online(struct swait_queue_head *wqh)
1024{
1025	int cpu = get_cpu();
1026
1027	/*
1028	 * If called from rcutree_report_cpu_starting(), wake up
1029	 * is dangerous that late in the CPU-down hotplug process. The
1030	 * scheduler might queue an ignored hrtimer. Defer the wake up
1031	 * to an online CPU instead.
1032	 */
1033	if (unlikely(cpu_is_offline(cpu))) {
1034		int target;
1035
1036		target = cpumask_any_and(housekeeping_cpumask(HK_TYPE_RCU),
1037					 cpu_online_mask);
1038
1039		smp_call_function_single(target, swake_up_one_online_ipi,
1040					 wqh, 0);
1041		put_cpu();
1042	} else {
1043		put_cpu();
1044		swake_up_one(wqh);
1045	}
1046}
1047
1048/*
1049 * Awaken the grace-period kthread.  Don't do a self-awaken (unless in an
1050 * interrupt or softirq handler, in which case we just might immediately
1051 * sleep upon return, resulting in a grace-period hang), and don't bother
1052 * awakening when there is nothing for the grace-period kthread to do
1053 * (as in several CPUs raced to awaken, we lost), and finally don't try
1054 * to awaken a kthread that has not yet been created.  If all those checks
1055 * are passed, track some debug information and awaken.
1056 *
1057 * So why do the self-wakeup when in an interrupt or softirq handler
1058 * in the grace-period kthread's context?  Because the kthread might have
1059 * been interrupted just as it was going to sleep, and just after the final
1060 * pre-sleep check of the awaken condition.  In this case, a wakeup really
1061 * is required, and is therefore supplied.
1062 */
1063static void rcu_gp_kthread_wake(void)
1064{
1065	struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
1066
1067	if ((current == t && !in_hardirq() && !in_serving_softirq()) ||
1068	    !READ_ONCE(rcu_state.gp_flags) || !t)
1069		return;
1070	WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
1071	WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
1072	swake_up_one_online(&rcu_state.gp_wq);
1073}
1074
1075/*
1076 * If there is room, assign a ->gp_seq number to any callbacks on this
1077 * CPU that have not already been assigned.  Also accelerate any callbacks
1078 * that were previously assigned a ->gp_seq number that has since proven
1079 * to be too conservative, which can happen if callbacks get assigned a
1080 * ->gp_seq number while RCU is idle, but with reference to a non-root
1081 * rcu_node structure.  This function is idempotent, so it does not hurt
1082 * to call it repeatedly.  Returns an flag saying that we should awaken
1083 * the RCU grace-period kthread.
1084 *
1085 * The caller must hold rnp->lock with interrupts disabled.
1086 */
1087static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1088{
1089	unsigned long gp_seq_req;
1090	bool ret = false;
1091
1092	rcu_lockdep_assert_cblist_protected(rdp);
1093	raw_lockdep_assert_held_rcu_node(rnp);
1094
1095	/* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1096	if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1097		return false;
1098
1099	trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPreAcc"));
1100
1101	/*
1102	 * Callbacks are often registered with incomplete grace-period
1103	 * information.  Something about the fact that getting exact
1104	 * information requires acquiring a global lock...  RCU therefore
1105	 * makes a conservative estimate of the grace period number at which
1106	 * a given callback will become ready to invoke.	The following
1107	 * code checks this estimate and improves it when possible, thus
1108	 * accelerating callback invocation to an earlier grace-period
1109	 * number.
1110	 */
1111	gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1112	if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1113		ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1114
1115	/* Trace depending on how much we were able to accelerate. */
1116	if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1117		trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccWaitCB"));
1118	else
1119		trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccReadyCB"));
1120
1121	trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPostAcc"));
1122
1123	return ret;
1124}
1125
1126/*
1127 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1128 * rcu_node structure's ->lock be held.  It consults the cached value
1129 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1130 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1131 * while holding the leaf rcu_node structure's ->lock.
1132 */
1133static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1134					struct rcu_data *rdp)
1135{
1136	unsigned long c;
1137	bool needwake;
1138
1139	rcu_lockdep_assert_cblist_protected(rdp);
1140	c = rcu_seq_snap(&rcu_state.gp_seq);
1141	if (!READ_ONCE(rdp->gpwrap) && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1142		/* Old request still live, so mark recent callbacks. */
1143		(void)rcu_segcblist_accelerate(&rdp->cblist, c);
1144		return;
1145	}
1146	raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1147	needwake = rcu_accelerate_cbs(rnp, rdp);
1148	raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1149	if (needwake)
1150		rcu_gp_kthread_wake();
1151}
1152
1153/*
1154 * Move any callbacks whose grace period has completed to the
1155 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1156 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1157 * sublist.  This function is idempotent, so it does not hurt to
1158 * invoke it repeatedly.  As long as it is not invoked -too- often...
1159 * Returns true if the RCU grace-period kthread needs to be awakened.
1160 *
1161 * The caller must hold rnp->lock with interrupts disabled.
1162 */
1163static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1164{
1165	rcu_lockdep_assert_cblist_protected(rdp);
1166	raw_lockdep_assert_held_rcu_node(rnp);
1167
1168	/* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1169	if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1170		return false;
1171
1172	/*
1173	 * Find all callbacks whose ->gp_seq numbers indicate that they
1174	 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1175	 */
1176	rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1177
1178	/* Classify any remaining callbacks. */
1179	return rcu_accelerate_cbs(rnp, rdp);
1180}
1181
1182/*
1183 * Move and classify callbacks, but only if doing so won't require
1184 * that the RCU grace-period kthread be awakened.
1185 */
1186static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
1187						  struct rcu_data *rdp)
1188{
1189	rcu_lockdep_assert_cblist_protected(rdp);
1190	if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) || !raw_spin_trylock_rcu_node(rnp))
 
1191		return;
1192	// The grace period cannot end while we hold the rcu_node lock.
1193	if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))
1194		WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
1195	raw_spin_unlock_rcu_node(rnp);
1196}
1197
1198/*
1199 * In CONFIG_RCU_STRICT_GRACE_PERIOD=y kernels, attempt to generate a
1200 * quiescent state.  This is intended to be invoked when the CPU notices
1201 * a new grace period.
1202 */
1203static void rcu_strict_gp_check_qs(void)
1204{
1205	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
1206		rcu_read_lock();
1207		rcu_read_unlock();
1208	}
1209}
1210
1211/*
1212 * Update CPU-local rcu_data state to record the beginnings and ends of
1213 * grace periods.  The caller must hold the ->lock of the leaf rcu_node
1214 * structure corresponding to the current CPU, and must have irqs disabled.
1215 * Returns true if the grace-period kthread needs to be awakened.
1216 */
1217static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
1218{
1219	bool ret = false;
1220	bool need_qs;
1221	const bool offloaded = rcu_rdp_is_offloaded(rdp);
1222
1223	raw_lockdep_assert_held_rcu_node(rnp);
1224
1225	if (rdp->gp_seq == rnp->gp_seq)
1226		return false; /* Nothing to do. */
1227
1228	/* Handle the ends of any preceding grace periods first. */
1229	if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1230	    unlikely(READ_ONCE(rdp->gpwrap))) {
1231		if (!offloaded)
1232			ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
1233		rdp->core_needs_qs = false;
1234		trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1235	} else {
1236		if (!offloaded)
1237			ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
1238		if (rdp->core_needs_qs)
1239			rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask);
1240	}
1241
1242	/* Now handle the beginnings of any new-to-this-CPU grace periods. */
1243	if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1244	    unlikely(READ_ONCE(rdp->gpwrap))) {
1245		/*
1246		 * If the current grace period is waiting for this CPU,
1247		 * set up to detect a quiescent state, otherwise don't
1248		 * go looking for one.
1249		 */
1250		trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1251		need_qs = !!(rnp->qsmask & rdp->grpmask);
1252		rdp->cpu_no_qs.b.norm = need_qs;
1253		rdp->core_needs_qs = need_qs;
1254		zero_cpu_stall_ticks(rdp);
1255	}
1256	rdp->gp_seq = rnp->gp_seq;  /* Remember new grace-period state. */
1257	if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
1258		WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1259	if (IS_ENABLED(CONFIG_PROVE_RCU) && READ_ONCE(rdp->gpwrap))
1260		WRITE_ONCE(rdp->last_sched_clock, jiffies);
1261	WRITE_ONCE(rdp->gpwrap, false);
1262	rcu_gpnum_ovf(rnp, rdp);
1263	return ret;
1264}
1265
1266static void note_gp_changes(struct rcu_data *rdp)
1267{
1268	unsigned long flags;
1269	bool needwake;
1270	struct rcu_node *rnp;
1271
1272	local_irq_save(flags);
1273	rnp = rdp->mynode;
1274	if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1275	     !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1276	    !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1277		local_irq_restore(flags);
1278		return;
1279	}
1280	needwake = __note_gp_changes(rnp, rdp);
1281	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1282	rcu_strict_gp_check_qs();
1283	if (needwake)
1284		rcu_gp_kthread_wake();
1285}
1286
1287static atomic_t *rcu_gp_slow_suppress;
1288
1289/* Register a counter to suppress debugging grace-period delays. */
1290void rcu_gp_slow_register(atomic_t *rgssp)
1291{
1292	WARN_ON_ONCE(rcu_gp_slow_suppress);
1293
1294	WRITE_ONCE(rcu_gp_slow_suppress, rgssp);
1295}
1296EXPORT_SYMBOL_GPL(rcu_gp_slow_register);
1297
1298/* Unregister a counter, with NULL for not caring which. */
1299void rcu_gp_slow_unregister(atomic_t *rgssp)
1300{
1301	WARN_ON_ONCE(rgssp && rgssp != rcu_gp_slow_suppress && rcu_gp_slow_suppress != NULL);
1302
1303	WRITE_ONCE(rcu_gp_slow_suppress, NULL);
1304}
1305EXPORT_SYMBOL_GPL(rcu_gp_slow_unregister);
1306
1307static bool rcu_gp_slow_is_suppressed(void)
1308{
1309	atomic_t *rgssp = READ_ONCE(rcu_gp_slow_suppress);
1310
1311	return rgssp && atomic_read(rgssp);
1312}
1313
1314static void rcu_gp_slow(int delay)
1315{
1316	if (!rcu_gp_slow_is_suppressed() && delay > 0 &&
1317	    !(rcu_seq_ctr(rcu_state.gp_seq) % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
 
1318		schedule_timeout_idle(delay);
1319}
1320
1321static unsigned long sleep_duration;
1322
1323/* Allow rcutorture to stall the grace-period kthread. */
1324void rcu_gp_set_torture_wait(int duration)
1325{
1326	if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST) && duration > 0)
1327		WRITE_ONCE(sleep_duration, duration);
1328}
1329EXPORT_SYMBOL_GPL(rcu_gp_set_torture_wait);
1330
1331/* Actually implement the aforementioned wait. */
1332static void rcu_gp_torture_wait(void)
1333{
1334	unsigned long duration;
1335
1336	if (!IS_ENABLED(CONFIG_RCU_TORTURE_TEST))
1337		return;
1338	duration = xchg(&sleep_duration, 0UL);
1339	if (duration > 0) {
1340		pr_alert("%s: Waiting %lu jiffies\n", __func__, duration);
1341		schedule_timeout_idle(duration);
1342		pr_alert("%s: Wait complete\n", __func__);
1343	}
1344}
1345
1346/*
1347 * Handler for on_each_cpu() to invoke the target CPU's RCU core
1348 * processing.
1349 */
1350static void rcu_strict_gp_boundary(void *unused)
1351{
1352	invoke_rcu_core();
1353}
1354
1355// Make the polled API aware of the beginning of a grace period.
1356static void rcu_poll_gp_seq_start(unsigned long *snap)
1357{
1358	struct rcu_node *rnp = rcu_get_root();
1359
1360	if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1361		raw_lockdep_assert_held_rcu_node(rnp);
1362
1363	// If RCU was idle, note beginning of GP.
1364	if (!rcu_seq_state(rcu_state.gp_seq_polled))
1365		rcu_seq_start(&rcu_state.gp_seq_polled);
1366
1367	// Either way, record current state.
1368	*snap = rcu_state.gp_seq_polled;
1369}
1370
1371// Make the polled API aware of the end of a grace period.
1372static void rcu_poll_gp_seq_end(unsigned long *snap)
1373{
1374	struct rcu_node *rnp = rcu_get_root();
1375
1376	if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1377		raw_lockdep_assert_held_rcu_node(rnp);
1378
1379	// If the previously noted GP is still in effect, record the
1380	// end of that GP.  Either way, zero counter to avoid counter-wrap
1381	// problems.
1382	if (*snap && *snap == rcu_state.gp_seq_polled) {
1383		rcu_seq_end(&rcu_state.gp_seq_polled);
1384		rcu_state.gp_seq_polled_snap = 0;
1385		rcu_state.gp_seq_polled_exp_snap = 0;
1386	} else {
1387		*snap = 0;
1388	}
1389}
1390
1391// Make the polled API aware of the beginning of a grace period, but
1392// where caller does not hold the root rcu_node structure's lock.
1393static void rcu_poll_gp_seq_start_unlocked(unsigned long *snap)
1394{
1395	unsigned long flags;
1396	struct rcu_node *rnp = rcu_get_root();
1397
1398	if (rcu_init_invoked()) {
1399		if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1400			lockdep_assert_irqs_enabled();
1401		raw_spin_lock_irqsave_rcu_node(rnp, flags);
1402	}
1403	rcu_poll_gp_seq_start(snap);
1404	if (rcu_init_invoked())
1405		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1406}
1407
1408// Make the polled API aware of the end of a grace period, but where
1409// caller does not hold the root rcu_node structure's lock.
1410static void rcu_poll_gp_seq_end_unlocked(unsigned long *snap)
1411{
1412	unsigned long flags;
1413	struct rcu_node *rnp = rcu_get_root();
1414
1415	if (rcu_init_invoked()) {
1416		if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1417			lockdep_assert_irqs_enabled();
1418		raw_spin_lock_irqsave_rcu_node(rnp, flags);
1419	}
1420	rcu_poll_gp_seq_end(snap);
1421	if (rcu_init_invoked())
1422		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1423}
1424
1425/*
1426 * Initialize a new grace period.  Return false if no grace period required.
1427 */
1428static noinline_for_stack bool rcu_gp_init(void)
1429{
 
1430	unsigned long flags;
1431	unsigned long oldmask;
1432	unsigned long mask;
1433	struct rcu_data *rdp;
1434	struct rcu_node *rnp = rcu_get_root();
1435
1436	WRITE_ONCE(rcu_state.gp_activity, jiffies);
1437	raw_spin_lock_irq_rcu_node(rnp);
1438	if (!READ_ONCE(rcu_state.gp_flags)) {
1439		/* Spurious wakeup, tell caller to go back to sleep.  */
1440		raw_spin_unlock_irq_rcu_node(rnp);
1441		return false;
1442	}
1443	WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1444
1445	if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1446		/*
1447		 * Grace period already in progress, don't start another.
1448		 * Not supposed to be able to happen.
1449		 */
1450		raw_spin_unlock_irq_rcu_node(rnp);
1451		return false;
1452	}
1453
1454	/* Advance to a new grace period and initialize state. */
1455	record_gp_stall_check_time();
1456	/* Record GP times before starting GP, hence rcu_seq_start(). */
1457	rcu_seq_start(&rcu_state.gp_seq);
1458	ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
1459	trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
1460	rcu_poll_gp_seq_start(&rcu_state.gp_seq_polled_snap);
1461	raw_spin_unlock_irq_rcu_node(rnp);
1462
1463	/*
1464	 * Apply per-leaf buffered online and offline operations to
1465	 * the rcu_node tree. Note that this new grace period need not
1466	 * wait for subsequent online CPUs, and that RCU hooks in the CPU
1467	 * offlining path, when combined with checks in this function,
1468	 * will handle CPUs that are currently going offline or that will
1469	 * go offline later.  Please also refer to "Hotplug CPU" section
1470	 * of RCU's Requirements documentation.
1471	 */
1472	WRITE_ONCE(rcu_state.gp_state, RCU_GP_ONOFF);
1473	/* Exclude CPU hotplug operations. */
1474	rcu_for_each_leaf_node(rnp) {
1475		local_irq_save(flags);
1476		arch_spin_lock(&rcu_state.ofl_lock);
1477		raw_spin_lock_rcu_node(rnp);
 
 
 
 
 
1478		if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1479		    !rnp->wait_blkd_tasks) {
1480			/* Nothing to do on this leaf rcu_node structure. */
1481			raw_spin_unlock_rcu_node(rnp);
1482			arch_spin_unlock(&rcu_state.ofl_lock);
1483			local_irq_restore(flags);
1484			continue;
1485		}
1486
1487		/* Record old state, apply changes to ->qsmaskinit field. */
1488		oldmask = rnp->qsmaskinit;
1489		rnp->qsmaskinit = rnp->qsmaskinitnext;
1490
1491		/* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1492		if (!oldmask != !rnp->qsmaskinit) {
1493			if (!oldmask) { /* First online CPU for rcu_node. */
1494				if (!rnp->wait_blkd_tasks) /* Ever offline? */
1495					rcu_init_new_rnp(rnp);
1496			} else if (rcu_preempt_has_tasks(rnp)) {
1497				rnp->wait_blkd_tasks = true; /* blocked tasks */
1498			} else { /* Last offline CPU and can propagate. */
1499				rcu_cleanup_dead_rnp(rnp);
1500			}
1501		}
1502
1503		/*
1504		 * If all waited-on tasks from prior grace period are
1505		 * done, and if all this rcu_node structure's CPUs are
1506		 * still offline, propagate up the rcu_node tree and
1507		 * clear ->wait_blkd_tasks.  Otherwise, if one of this
1508		 * rcu_node structure's CPUs has since come back online,
1509		 * simply clear ->wait_blkd_tasks.
1510		 */
1511		if (rnp->wait_blkd_tasks &&
1512		    (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1513			rnp->wait_blkd_tasks = false;
1514			if (!rnp->qsmaskinit)
1515				rcu_cleanup_dead_rnp(rnp);
1516		}
1517
1518		raw_spin_unlock_rcu_node(rnp);
1519		arch_spin_unlock(&rcu_state.ofl_lock);
1520		local_irq_restore(flags);
1521	}
1522	rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1523
1524	/*
1525	 * Set the quiescent-state-needed bits in all the rcu_node
1526	 * structures for all currently online CPUs in breadth-first
1527	 * order, starting from the root rcu_node structure, relying on the
1528	 * layout of the tree within the rcu_state.node[] array.  Note that
1529	 * other CPUs will access only the leaves of the hierarchy, thus
1530	 * seeing that no grace period is in progress, at least until the
1531	 * corresponding leaf node has been initialized.
1532	 *
1533	 * The grace period cannot complete until the initialization
1534	 * process finishes, because this kthread handles both.
1535	 */
1536	WRITE_ONCE(rcu_state.gp_state, RCU_GP_INIT);
1537	rcu_for_each_node_breadth_first(rnp) {
1538		rcu_gp_slow(gp_init_delay);
1539		raw_spin_lock_irqsave_rcu_node(rnp, flags);
1540		rdp = this_cpu_ptr(&rcu_data);
1541		rcu_preempt_check_blocked_tasks(rnp);
1542		rnp->qsmask = rnp->qsmaskinit;
1543		WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
1544		if (rnp == rdp->mynode)
1545			(void)__note_gp_changes(rnp, rdp);
1546		rcu_preempt_boost_start_gp(rnp);
1547		trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1548					    rnp->level, rnp->grplo,
1549					    rnp->grphi, rnp->qsmask);
1550		/* Quiescent states for tasks on any now-offline CPUs. */
1551		mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1552		rnp->rcu_gp_init_mask = mask;
1553		if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1554			rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1555		else
1556			raw_spin_unlock_irq_rcu_node(rnp);
1557		cond_resched_tasks_rcu_qs();
1558		WRITE_ONCE(rcu_state.gp_activity, jiffies);
1559	}
1560
1561	// If strict, make all CPUs aware of new grace period.
1562	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
1563		on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
1564
1565	return true;
1566}
1567
1568/*
1569 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1570 * time.
1571 */
1572static bool rcu_gp_fqs_check_wake(int *gfp)
1573{
1574	struct rcu_node *rnp = rcu_get_root();
1575
1576	// If under overload conditions, force an immediate FQS scan.
1577	if (*gfp & RCU_GP_FLAG_OVLD)
1578		return true;
1579
1580	// Someone like call_rcu() requested a force-quiescent-state scan.
1581	*gfp = READ_ONCE(rcu_state.gp_flags);
1582	if (*gfp & RCU_GP_FLAG_FQS)
1583		return true;
1584
1585	// The current grace period has completed.
1586	if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1587		return true;
1588
1589	return false;
1590}
1591
1592/*
1593 * Do one round of quiescent-state forcing.
1594 */
1595static void rcu_gp_fqs(bool first_time)
1596{
1597	int nr_fqs = READ_ONCE(rcu_state.nr_fqs_jiffies_stall);
1598	struct rcu_node *rnp = rcu_get_root();
1599
1600	WRITE_ONCE(rcu_state.gp_activity, jiffies);
1601	WRITE_ONCE(rcu_state.n_force_qs, rcu_state.n_force_qs + 1);
1602
1603	WARN_ON_ONCE(nr_fqs > 3);
1604	/* Only countdown nr_fqs for stall purposes if jiffies moves. */
1605	if (nr_fqs) {
1606		if (nr_fqs == 1) {
1607			WRITE_ONCE(rcu_state.jiffies_stall,
1608				   jiffies + rcu_jiffies_till_stall_check());
1609		}
1610		WRITE_ONCE(rcu_state.nr_fqs_jiffies_stall, --nr_fqs);
1611	}
1612
1613	if (first_time) {
1614		/* Collect dyntick-idle snapshots. */
1615		force_qs_rnp(dyntick_save_progress_counter);
1616	} else {
1617		/* Handle dyntick-idle and offline CPUs. */
1618		force_qs_rnp(rcu_implicit_dynticks_qs);
1619	}
1620	/* Clear flag to prevent immediate re-entry. */
1621	if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
1622		raw_spin_lock_irq_rcu_node(rnp);
1623		WRITE_ONCE(rcu_state.gp_flags,
1624			   READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
1625		raw_spin_unlock_irq_rcu_node(rnp);
1626	}
1627}
1628
1629/*
1630 * Loop doing repeated quiescent-state forcing until the grace period ends.
1631 */
1632static noinline_for_stack void rcu_gp_fqs_loop(void)
1633{
1634	bool first_gp_fqs = true;
1635	int gf = 0;
1636	unsigned long j;
1637	int ret;
1638	struct rcu_node *rnp = rcu_get_root();
1639
 
1640	j = READ_ONCE(jiffies_till_first_fqs);
1641	if (rcu_state.cbovld)
1642		gf = RCU_GP_FLAG_OVLD;
1643	ret = 0;
1644	for (;;) {
1645		if (rcu_state.cbovld) {
1646			j = (j + 2) / 3;
1647			if (j <= 0)
1648				j = 1;
1649		}
1650		if (!ret || time_before(jiffies + j, rcu_state.jiffies_force_qs)) {
1651			WRITE_ONCE(rcu_state.jiffies_force_qs, jiffies + j);
1652			/*
1653			 * jiffies_force_qs before RCU_GP_WAIT_FQS state
1654			 * update; required for stall checks.
1655			 */
1656			smp_wmb();
1657			WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
1658				   jiffies + (j ? 3 * j : 2));
1659		}
1660		trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1661				       TPS("fqswait"));
1662		WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_FQS);
1663		(void)swait_event_idle_timeout_exclusive(rcu_state.gp_wq,
1664				 rcu_gp_fqs_check_wake(&gf), j);
1665		rcu_gp_torture_wait();
1666		WRITE_ONCE(rcu_state.gp_state, RCU_GP_DOING_FQS);
1667		/* Locking provides needed memory barriers. */
1668		/*
1669		 * Exit the loop if the root rcu_node structure indicates that the grace period
1670		 * has ended, leave the loop.  The rcu_preempt_blocked_readers_cgp(rnp) check
1671		 * is required only for single-node rcu_node trees because readers blocking
1672		 * the current grace period are queued only on leaf rcu_node structures.
1673		 * For multi-node trees, checking the root node's ->qsmask suffices, because a
1674		 * given root node's ->qsmask bit is cleared only when all CPUs and tasks from
1675		 * the corresponding leaf nodes have passed through their quiescent state.
1676		 */
1677		if (!READ_ONCE(rnp->qsmask) &&
1678		    !rcu_preempt_blocked_readers_cgp(rnp))
1679			break;
1680		/* If time for quiescent-state forcing, do it. */
1681		if (!time_after(rcu_state.jiffies_force_qs, jiffies) ||
1682		    (gf & (RCU_GP_FLAG_FQS | RCU_GP_FLAG_OVLD))) {
1683			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1684					       TPS("fqsstart"));
1685			rcu_gp_fqs(first_gp_fqs);
1686			gf = 0;
1687			if (first_gp_fqs) {
1688				first_gp_fqs = false;
1689				gf = rcu_state.cbovld ? RCU_GP_FLAG_OVLD : 0;
1690			}
1691			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1692					       TPS("fqsend"));
1693			cond_resched_tasks_rcu_qs();
1694			WRITE_ONCE(rcu_state.gp_activity, jiffies);
1695			ret = 0; /* Force full wait till next FQS. */
1696			j = READ_ONCE(jiffies_till_next_fqs);
1697		} else {
1698			/* Deal with stray signal. */
1699			cond_resched_tasks_rcu_qs();
1700			WRITE_ONCE(rcu_state.gp_activity, jiffies);
1701			WARN_ON(signal_pending(current));
1702			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1703					       TPS("fqswaitsig"));
1704			ret = 1; /* Keep old FQS timing. */
1705			j = jiffies;
1706			if (time_after(jiffies, rcu_state.jiffies_force_qs))
1707				j = 1;
1708			else
1709				j = rcu_state.jiffies_force_qs - j;
1710			gf = 0;
1711		}
1712	}
1713}
1714
1715/*
1716 * Clean up after the old grace period.
1717 */
1718static noinline void rcu_gp_cleanup(void)
1719{
1720	int cpu;
1721	bool needgp = false;
1722	unsigned long gp_duration;
1723	unsigned long new_gp_seq;
1724	bool offloaded;
1725	struct rcu_data *rdp;
1726	struct rcu_node *rnp = rcu_get_root();
1727	struct swait_queue_head *sq;
1728
1729	WRITE_ONCE(rcu_state.gp_activity, jiffies);
1730	raw_spin_lock_irq_rcu_node(rnp);
1731	rcu_state.gp_end = jiffies;
1732	gp_duration = rcu_state.gp_end - rcu_state.gp_start;
1733	if (gp_duration > rcu_state.gp_max)
1734		rcu_state.gp_max = gp_duration;
1735
1736	/*
1737	 * We know the grace period is complete, but to everyone else
1738	 * it appears to still be ongoing.  But it is also the case
1739	 * that to everyone else it looks like there is nothing that
1740	 * they can do to advance the grace period.  It is therefore
1741	 * safe for us to drop the lock in order to mark the grace
1742	 * period as completed in all of the rcu_node structures.
1743	 */
1744	rcu_poll_gp_seq_end(&rcu_state.gp_seq_polled_snap);
1745	raw_spin_unlock_irq_rcu_node(rnp);
1746
1747	/*
1748	 * Propagate new ->gp_seq value to rcu_node structures so that
1749	 * other CPUs don't have to wait until the start of the next grace
1750	 * period to process their callbacks.  This also avoids some nasty
1751	 * RCU grace-period initialization races by forcing the end of
1752	 * the current grace period to be completely recorded in all of
1753	 * the rcu_node structures before the beginning of the next grace
1754	 * period is recorded in any of the rcu_node structures.
1755	 */
1756	new_gp_seq = rcu_state.gp_seq;
1757	rcu_seq_end(&new_gp_seq);
1758	rcu_for_each_node_breadth_first(rnp) {
1759		raw_spin_lock_irq_rcu_node(rnp);
1760		if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
1761			dump_blkd_tasks(rnp, 10);
1762		WARN_ON_ONCE(rnp->qsmask);
1763		WRITE_ONCE(rnp->gp_seq, new_gp_seq);
1764		if (!rnp->parent)
1765			smp_mb(); // Order against failing poll_state_synchronize_rcu_full().
1766		rdp = this_cpu_ptr(&rcu_data);
1767		if (rnp == rdp->mynode)
1768			needgp = __note_gp_changes(rnp, rdp) || needgp;
1769		/* smp_mb() provided by prior unlock-lock pair. */
1770		needgp = rcu_future_gp_cleanup(rnp) || needgp;
1771		// Reset overload indication for CPUs no longer overloaded
1772		if (rcu_is_leaf_node(rnp))
1773			for_each_leaf_node_cpu_mask(rnp, cpu, rnp->cbovldmask) {
1774				rdp = per_cpu_ptr(&rcu_data, cpu);
1775				check_cb_ovld_locked(rdp, rnp);
1776			}
1777		sq = rcu_nocb_gp_get(rnp);
1778		raw_spin_unlock_irq_rcu_node(rnp);
1779		rcu_nocb_gp_cleanup(sq);
1780		cond_resched_tasks_rcu_qs();
1781		WRITE_ONCE(rcu_state.gp_activity, jiffies);
1782		rcu_gp_slow(gp_cleanup_delay);
1783	}
1784	rnp = rcu_get_root();
1785	raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
1786
1787	/* Declare grace period done, trace first to use old GP number. */
1788	trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
1789	rcu_seq_end(&rcu_state.gp_seq);
1790	ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
1791	WRITE_ONCE(rcu_state.gp_state, RCU_GP_IDLE);
1792	/* Check for GP requests since above loop. */
1793	rdp = this_cpu_ptr(&rcu_data);
1794	if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
1795		trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
1796				  TPS("CleanupMore"));
1797		needgp = true;
1798	}
1799	/* Advance CBs to reduce false positives below. */
1800	offloaded = rcu_rdp_is_offloaded(rdp);
1801	if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
1802
1803		// We get here if a grace period was needed (“needgp”)
1804		// and the above call to rcu_accelerate_cbs() did not set
1805		// the RCU_GP_FLAG_INIT bit in ->gp_state (which records
1806		// the need for another grace period).  The purpose
1807		// of the “offloaded” check is to avoid invoking
1808		// rcu_accelerate_cbs() on an offloaded CPU because we do not
1809		// hold the ->nocb_lock needed to safely access an offloaded
1810		// ->cblist.  We do not want to acquire that lock because
1811		// it can be heavily contended during callback floods.
1812
1813		WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
1814		WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
1815		trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("newreq"));
 
 
1816	} else {
1817
1818		// We get here either if there is no need for an
1819		// additional grace period or if rcu_accelerate_cbs() has
1820		// already set the RCU_GP_FLAG_INIT bit in ->gp_flags. 
1821		// So all we need to do is to clear all of the other
1822		// ->gp_flags bits.
1823
1824		WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags & RCU_GP_FLAG_INIT);
1825	}
1826	raw_spin_unlock_irq_rcu_node(rnp);
1827
1828	// If strict, make all CPUs aware of the end of the old grace period.
1829	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
1830		on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
1831}
1832
1833/*
1834 * Body of kthread that handles grace periods.
1835 */
1836static int __noreturn rcu_gp_kthread(void *unused)
1837{
1838	rcu_bind_gp_kthread();
1839	for (;;) {
1840
1841		/* Handle grace-period start. */
1842		for (;;) {
1843			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1844					       TPS("reqwait"));
1845			WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_GPS);
1846			swait_event_idle_exclusive(rcu_state.gp_wq,
1847					 READ_ONCE(rcu_state.gp_flags) &
1848					 RCU_GP_FLAG_INIT);
1849			rcu_gp_torture_wait();
1850			WRITE_ONCE(rcu_state.gp_state, RCU_GP_DONE_GPS);
1851			/* Locking provides needed memory barrier. */
1852			if (rcu_gp_init())
1853				break;
1854			cond_resched_tasks_rcu_qs();
1855			WRITE_ONCE(rcu_state.gp_activity, jiffies);
1856			WARN_ON(signal_pending(current));
1857			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1858					       TPS("reqwaitsig"));
1859		}
1860
1861		/* Handle quiescent-state forcing. */
1862		rcu_gp_fqs_loop();
1863
1864		/* Handle grace-period end. */
1865		WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANUP);
1866		rcu_gp_cleanup();
1867		WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANED);
1868	}
1869}
1870
1871/*
1872 * Report a full set of quiescent states to the rcu_state data structure.
1873 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
1874 * another grace period is required.  Whether we wake the grace-period
1875 * kthread or it awakens itself for the next round of quiescent-state
1876 * forcing, that kthread will clean up after the just-completed grace
1877 * period.  Note that the caller must hold rnp->lock, which is released
1878 * before return.
1879 */
1880static void rcu_report_qs_rsp(unsigned long flags)
1881	__releases(rcu_get_root()->lock)
1882{
1883	raw_lockdep_assert_held_rcu_node(rcu_get_root());
1884	WARN_ON_ONCE(!rcu_gp_in_progress());
1885	WRITE_ONCE(rcu_state.gp_flags,
1886		   READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
1887	raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
1888	rcu_gp_kthread_wake();
1889}
1890
1891/*
1892 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1893 * Allows quiescent states for a group of CPUs to be reported at one go
1894 * to the specified rcu_node structure, though all the CPUs in the group
1895 * must be represented by the same rcu_node structure (which need not be a
1896 * leaf rcu_node structure, though it often will be).  The gps parameter
1897 * is the grace-period snapshot, which means that the quiescent states
1898 * are valid only if rnp->gp_seq is equal to gps.  That structure's lock
1899 * must be held upon entry, and it is released before return.
1900 *
1901 * As a special case, if mask is zero, the bit-already-cleared check is
1902 * disabled.  This allows propagating quiescent state due to resumed tasks
1903 * during grace-period initialization.
1904 */
1905static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
1906			      unsigned long gps, unsigned long flags)
1907	__releases(rnp->lock)
1908{
1909	unsigned long oldmask = 0;
1910	struct rcu_node *rnp_c;
1911
1912	raw_lockdep_assert_held_rcu_node(rnp);
1913
1914	/* Walk up the rcu_node hierarchy. */
1915	for (;;) {
1916		if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
1917
1918			/*
1919			 * Our bit has already been cleared, or the
1920			 * relevant grace period is already over, so done.
1921			 */
1922			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1923			return;
1924		}
1925		WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
1926		WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
1927			     rcu_preempt_blocked_readers_cgp(rnp));
1928		WRITE_ONCE(rnp->qsmask, rnp->qsmask & ~mask);
1929		trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
1930						 mask, rnp->qsmask, rnp->level,
1931						 rnp->grplo, rnp->grphi,
1932						 !!rnp->gp_tasks);
1933		if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1934
1935			/* Other bits still set at this level, so done. */
1936			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1937			return;
1938		}
1939		rnp->completedqs = rnp->gp_seq;
1940		mask = rnp->grpmask;
1941		if (rnp->parent == NULL) {
1942
1943			/* No more levels.  Exit loop holding root lock. */
1944
1945			break;
1946		}
1947		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1948		rnp_c = rnp;
1949		rnp = rnp->parent;
1950		raw_spin_lock_irqsave_rcu_node(rnp, flags);
1951		oldmask = READ_ONCE(rnp_c->qsmask);
1952	}
1953
1954	/*
1955	 * Get here if we are the last CPU to pass through a quiescent
1956	 * state for this grace period.  Invoke rcu_report_qs_rsp()
1957	 * to clean up and start the next grace period if one is needed.
1958	 */
1959	rcu_report_qs_rsp(flags); /* releases rnp->lock. */
1960}
1961
1962/*
1963 * Record a quiescent state for all tasks that were previously queued
1964 * on the specified rcu_node structure and that were blocking the current
1965 * RCU grace period.  The caller must hold the corresponding rnp->lock with
1966 * irqs disabled, and this lock is released upon return, but irqs remain
1967 * disabled.
1968 */
1969static void __maybe_unused
1970rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
1971	__releases(rnp->lock)
1972{
1973	unsigned long gps;
1974	unsigned long mask;
1975	struct rcu_node *rnp_p;
1976
1977	raw_lockdep_assert_held_rcu_node(rnp);
1978	if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RCU)) ||
1979	    WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
1980	    rnp->qsmask != 0) {
1981		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1982		return;  /* Still need more quiescent states! */
1983	}
1984
1985	rnp->completedqs = rnp->gp_seq;
1986	rnp_p = rnp->parent;
1987	if (rnp_p == NULL) {
1988		/*
1989		 * Only one rcu_node structure in the tree, so don't
1990		 * try to report up to its nonexistent parent!
1991		 */
1992		rcu_report_qs_rsp(flags);
1993		return;
1994	}
1995
1996	/* Report up the rest of the hierarchy, tracking current ->gp_seq. */
1997	gps = rnp->gp_seq;
1998	mask = rnp->grpmask;
1999	raw_spin_unlock_rcu_node(rnp);	/* irqs remain disabled. */
2000	raw_spin_lock_rcu_node(rnp_p);	/* irqs already disabled. */
2001	rcu_report_qs_rnp(mask, rnp_p, gps, flags);
2002}
2003
2004/*
2005 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2006 * structure.  This must be called from the specified CPU.
2007 */
2008static void
2009rcu_report_qs_rdp(struct rcu_data *rdp)
2010{
2011	unsigned long flags;
2012	unsigned long mask;
2013	bool needacc = false;
 
2014	struct rcu_node *rnp;
2015
2016	WARN_ON_ONCE(rdp->cpu != smp_processor_id());
2017	rnp = rdp->mynode;
2018	raw_spin_lock_irqsave_rcu_node(rnp, flags);
2019	if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
2020	    rdp->gpwrap) {
2021
2022		/*
2023		 * The grace period in which this quiescent state was
2024		 * recorded has ended, so don't report it upwards.
2025		 * We will instead need a new quiescent state that lies
2026		 * within the current grace period.
2027		 */
2028		rdp->cpu_no_qs.b.norm = true;	/* need qs for new gp. */
2029		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2030		return;
2031	}
2032	mask = rdp->grpmask;
2033	rdp->core_needs_qs = false;
2034	if ((rnp->qsmask & mask) == 0) {
2035		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2036	} else {
2037		/*
2038		 * This GP can't end until cpu checks in, so all of our
2039		 * callbacks can be processed during the next GP.
2040		 *
2041		 * NOCB kthreads have their own way to deal with that...
2042		 */
2043		if (!rcu_rdp_is_offloaded(rdp)) {
2044			/*
2045			 * The current GP has not yet ended, so it
2046			 * should not be possible for rcu_accelerate_cbs()
2047			 * to return true.  So complain, but don't awaken.
2048			 */
2049			WARN_ON_ONCE(rcu_accelerate_cbs(rnp, rdp));
2050		} else if (!rcu_segcblist_completely_offloaded(&rdp->cblist)) {
2051			/*
2052			 * ...but NOCB kthreads may miss or delay callbacks acceleration
2053			 * if in the middle of a (de-)offloading process.
2054			 */
2055			needacc = true;
2056		}
2057
2058		rcu_disable_urgency_upon_qs(rdp);
2059		rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2060		/* ^^^ Released rnp->lock */
2061
2062		if (needacc) {
2063			rcu_nocb_lock_irqsave(rdp, flags);
2064			rcu_accelerate_cbs_unlocked(rnp, rdp);
2065			rcu_nocb_unlock_irqrestore(rdp, flags);
2066		}
2067	}
2068}
2069
2070/*
2071 * Check to see if there is a new grace period of which this CPU
2072 * is not yet aware, and if so, set up local rcu_data state for it.
2073 * Otherwise, see if this CPU has just passed through its first
2074 * quiescent state for this grace period, and record that fact if so.
2075 */
2076static void
2077rcu_check_quiescent_state(struct rcu_data *rdp)
2078{
2079	/* Check for grace-period ends and beginnings. */
2080	note_gp_changes(rdp);
2081
2082	/*
2083	 * Does this CPU still need to do its part for current grace period?
2084	 * If no, return and let the other CPUs do their part as well.
2085	 */
2086	if (!rdp->core_needs_qs)
2087		return;
2088
2089	/*
2090	 * Was there a quiescent state since the beginning of the grace
2091	 * period? If no, then exit and wait for the next call.
2092	 */
2093	if (rdp->cpu_no_qs.b.norm)
2094		return;
2095
2096	/*
2097	 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2098	 * judge of that).
2099	 */
2100	rcu_report_qs_rdp(rdp);
2101}
2102
2103/* Return true if callback-invocation time limit exceeded. */
2104static bool rcu_do_batch_check_time(long count, long tlimit,
2105				    bool jlimit_check, unsigned long jlimit)
2106{
2107	// Invoke local_clock() only once per 32 consecutive callbacks.
2108	return unlikely(tlimit) &&
2109	       (!likely(count & 31) ||
2110		(IS_ENABLED(CONFIG_RCU_DOUBLE_CHECK_CB_TIME) &&
2111		 jlimit_check && time_after(jiffies, jlimit))) &&
2112	       local_clock() >= tlimit;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2113}
2114
2115/*
2116 * Invoke any RCU callbacks that have made it to the end of their grace
2117 * period.  Throttle as specified by rdp->blimit.
2118 */
2119static void rcu_do_batch(struct rcu_data *rdp)
2120{
2121	long bl;
2122	long count = 0;
2123	int div;
2124	bool __maybe_unused empty;
2125	unsigned long flags;
2126	unsigned long jlimit;
2127	bool jlimit_check = false;
2128	long pending;
2129	struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2130	struct rcu_head *rhp;
2131	long tlimit = 0;
 
 
2132
2133	/* If no callbacks are ready, just return. */
2134	if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2135		trace_rcu_batch_start(rcu_state.name,
2136				      rcu_segcblist_n_cbs(&rdp->cblist), 0);
2137		trace_rcu_batch_end(rcu_state.name, 0,
2138				    !rcu_segcblist_empty(&rdp->cblist),
2139				    need_resched(), is_idle_task(current),
2140				    rcu_is_callbacks_kthread(rdp));
2141		return;
2142	}
2143
2144	/*
2145	 * Extract the list of ready callbacks, disabling IRQs to prevent
2146	 * races with call_rcu() from interrupt handlers.  Leave the
2147	 * callback counts, as rcu_barrier() needs to be conservative.
2148	 */
2149	rcu_nocb_lock_irqsave(rdp, flags);
 
2150	WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2151	pending = rcu_segcblist_get_seglen(&rdp->cblist, RCU_DONE_TAIL);
2152	div = READ_ONCE(rcu_divisor);
2153	div = div < 0 ? 7 : div > sizeof(long) * 8 - 2 ? sizeof(long) * 8 - 2 : div;
2154	bl = max(rdp->blimit, pending >> div);
2155	if ((in_serving_softirq() || rdp->rcu_cpu_kthread_status == RCU_KTHREAD_RUNNING) &&
2156	    (IS_ENABLED(CONFIG_RCU_DOUBLE_CHECK_CB_TIME) || unlikely(bl > 100))) {
2157		const long npj = NSEC_PER_SEC / HZ;
2158		long rrn = READ_ONCE(rcu_resched_ns);
2159
2160		rrn = rrn < NSEC_PER_MSEC ? NSEC_PER_MSEC : rrn > NSEC_PER_SEC ? NSEC_PER_SEC : rrn;
2161		tlimit = local_clock() + rrn;
2162		jlimit = jiffies + (rrn + npj + 1) / npj;
2163		jlimit_check = true;
2164	}
2165	trace_rcu_batch_start(rcu_state.name,
2166			      rcu_segcblist_n_cbs(&rdp->cblist), bl);
2167	rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2168	if (rcu_rdp_is_offloaded(rdp))
2169		rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2170
2171	trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbDequeued"));
2172	rcu_nocb_unlock_irqrestore(rdp, flags);
2173
2174	/* Invoke callbacks. */
2175	tick_dep_set_task(current, TICK_DEP_BIT_RCU);
2176	rhp = rcu_cblist_dequeue(&rcl);
2177
2178	for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2179		rcu_callback_t f;
2180
2181		count++;
2182		debug_rcu_head_unqueue(rhp);
2183
2184		rcu_lock_acquire(&rcu_callback_map);
2185		trace_rcu_invoke_callback(rcu_state.name, rhp);
2186
2187		f = rhp->func;
2188		debug_rcu_head_callback(rhp);
2189		WRITE_ONCE(rhp->func, (rcu_callback_t)0L);
2190		f(rhp);
2191
2192		rcu_lock_release(&rcu_callback_map);
2193
2194		/*
2195		 * Stop only if limit reached and CPU has something to do.
2196		 */
2197		if (in_serving_softirq()) {
2198			if (count >= bl && (need_resched() || !is_idle_task(current)))
2199				break;
2200			/*
2201			 * Make sure we don't spend too much time here and deprive other
2202			 * softirq vectors of CPU cycles.
2203			 */
2204			if (rcu_do_batch_check_time(count, tlimit, jlimit_check, jlimit))
2205				break;
2206		} else {
2207			// In rcuc/rcuoc context, so no worries about
2208			// depriving other softirq vectors of CPU cycles.
2209			local_bh_enable();
2210			lockdep_assert_irqs_enabled();
2211			cond_resched_tasks_rcu_qs();
2212			lockdep_assert_irqs_enabled();
2213			local_bh_disable();
2214			// But rcuc kthreads can delay quiescent-state
2215			// reporting, so check time limits for them.
2216			if (rdp->rcu_cpu_kthread_status == RCU_KTHREAD_RUNNING &&
2217			    rcu_do_batch_check_time(count, tlimit, jlimit_check, jlimit)) {
2218				rdp->rcu_cpu_has_work = 1;
2219				break;
2220			}
2221		}
2222	}
2223
2224	rcu_nocb_lock_irqsave(rdp, flags);
 
2225	rdp->n_cbs_invoked += count;
2226	trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2227			    is_idle_task(current), rcu_is_callbacks_kthread(rdp));
2228
2229	/* Update counts and requeue any remaining callbacks. */
2230	rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2231	rcu_segcblist_add_len(&rdp->cblist, -count);
2232
2233	/* Reinstate batch limit if we have worked down the excess. */
2234	count = rcu_segcblist_n_cbs(&rdp->cblist);
2235	if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
2236		rdp->blimit = blimit;
2237
2238	/* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2239	if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2240		rdp->qlen_last_fqs_check = 0;
2241		rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
2242	} else if (count < rdp->qlen_last_fqs_check - qhimark)
2243		rdp->qlen_last_fqs_check = count;
2244
2245	/*
2246	 * The following usually indicates a double call_rcu().  To track
2247	 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2248	 */
2249	empty = rcu_segcblist_empty(&rdp->cblist);
2250	WARN_ON_ONCE(count == 0 && !empty);
2251	WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2252		     count != 0 && empty);
2253	WARN_ON_ONCE(count == 0 && rcu_segcblist_n_segment_cbs(&rdp->cblist) != 0);
2254	WARN_ON_ONCE(!empty && rcu_segcblist_n_segment_cbs(&rdp->cblist) == 0);
2255
2256	rcu_nocb_unlock_irqrestore(rdp, flags);
2257
 
 
 
2258	tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
2259}
2260
2261/*
2262 * This function is invoked from each scheduling-clock interrupt,
2263 * and checks to see if this CPU is in a non-context-switch quiescent
2264 * state, for example, user mode or idle loop.  It also schedules RCU
2265 * core processing.  If the current grace period has gone on too long,
2266 * it will ask the scheduler to manufacture a context switch for the sole
2267 * purpose of providing the needed quiescent state.
2268 */
2269void rcu_sched_clock_irq(int user)
2270{
2271	unsigned long j;
2272
2273	if (IS_ENABLED(CONFIG_PROVE_RCU)) {
2274		j = jiffies;
2275		WARN_ON_ONCE(time_before(j, __this_cpu_read(rcu_data.last_sched_clock)));
2276		__this_cpu_write(rcu_data.last_sched_clock, j);
2277	}
2278	trace_rcu_utilization(TPS("Start scheduler-tick"));
2279	lockdep_assert_irqs_disabled();
2280	raw_cpu_inc(rcu_data.ticks_this_gp);
2281	/* The load-acquire pairs with the store-release setting to true. */
2282	if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
2283		/* Idle and userspace execution already are quiescent states. */
2284		if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2285			set_tsk_need_resched(current);
2286			set_preempt_need_resched();
2287		}
2288		__this_cpu_write(rcu_data.rcu_urgent_qs, false);
2289	}
2290	rcu_flavor_sched_clock_irq(user);
2291	if (rcu_pending(user))
2292		invoke_rcu_core();
2293	if (user || rcu_is_cpu_rrupt_from_idle())
2294		rcu_note_voluntary_context_switch(current);
2295	lockdep_assert_irqs_disabled();
2296
2297	trace_rcu_utilization(TPS("End scheduler-tick"));
2298}
2299
2300/*
2301 * Scan the leaf rcu_node structures.  For each structure on which all
2302 * CPUs have reported a quiescent state and on which there are tasks
2303 * blocking the current grace period, initiate RCU priority boosting.
2304 * Otherwise, invoke the specified function to check dyntick state for
2305 * each CPU that has not yet reported a quiescent state.
2306 */
2307static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
2308{
2309	int cpu;
2310	unsigned long flags;
 
 
2311	struct rcu_node *rnp;
2312
2313	rcu_state.cbovld = rcu_state.cbovldnext;
2314	rcu_state.cbovldnext = false;
2315	rcu_for_each_leaf_node(rnp) {
2316		unsigned long mask = 0;
2317		unsigned long rsmask = 0;
2318
2319		cond_resched_tasks_rcu_qs();
 
2320		raw_spin_lock_irqsave_rcu_node(rnp, flags);
2321		rcu_state.cbovldnext |= !!rnp->cbovldmask;
2322		if (rnp->qsmask == 0) {
2323			if (rcu_preempt_blocked_readers_cgp(rnp)) {
2324				/*
2325				 * No point in scanning bits because they
2326				 * are all zero.  But we might need to
2327				 * priority-boost blocked readers.
2328				 */
2329				rcu_initiate_boost(rnp, flags);
2330				/* rcu_initiate_boost() releases rnp->lock */
2331				continue;
2332			}
2333			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2334			continue;
2335		}
2336		for_each_leaf_node_cpu_mask(rnp, cpu, rnp->qsmask) {
2337			struct rcu_data *rdp;
2338			int ret;
2339
2340			rdp = per_cpu_ptr(&rcu_data, cpu);
2341			ret = f(rdp);
2342			if (ret > 0) {
2343				mask |= rdp->grpmask;
2344				rcu_disable_urgency_upon_qs(rdp);
2345			}
2346			if (ret < 0)
2347				rsmask |= rdp->grpmask;
2348		}
2349		if (mask != 0) {
2350			/* Idle/offline CPUs, report (releases rnp->lock). */
2351			rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2352		} else {
2353			/* Nothing to do here, so just drop the lock. */
2354			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2355		}
2356
2357		for_each_leaf_node_cpu_mask(rnp, cpu, rsmask)
2358			resched_cpu(cpu);
2359	}
2360}
2361
2362/*
2363 * Force quiescent states on reluctant CPUs, and also detect which
2364 * CPUs are in dyntick-idle mode.
2365 */
2366void rcu_force_quiescent_state(void)
2367{
2368	unsigned long flags;
2369	bool ret;
2370	struct rcu_node *rnp;
2371	struct rcu_node *rnp_old = NULL;
2372
2373	if (!rcu_gp_in_progress())
2374		return;
2375	/* Funnel through hierarchy to reduce memory contention. */
2376	rnp = raw_cpu_read(rcu_data.mynode);
2377	for (; rnp != NULL; rnp = rnp->parent) {
2378		ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2379		       !raw_spin_trylock(&rnp->fqslock);
2380		if (rnp_old != NULL)
2381			raw_spin_unlock(&rnp_old->fqslock);
2382		if (ret)
2383			return;
2384		rnp_old = rnp;
2385	}
2386	/* rnp_old == rcu_get_root(), rnp == NULL. */
2387
2388	/* Reached the root of the rcu_node tree, acquire lock. */
2389	raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2390	raw_spin_unlock(&rnp_old->fqslock);
2391	if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2392		raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2393		return;  /* Someone beat us to it. */
2394	}
2395	WRITE_ONCE(rcu_state.gp_flags,
2396		   READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2397	raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2398	rcu_gp_kthread_wake();
2399}
2400EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
2401
2402// Workqueue handler for an RCU reader for kernels enforcing struct RCU
2403// grace periods.
2404static void strict_work_handler(struct work_struct *work)
2405{
2406	rcu_read_lock();
2407	rcu_read_unlock();
2408}
2409
2410/* Perform RCU core processing work for the current CPU.  */
2411static __latent_entropy void rcu_core(void)
2412{
2413	unsigned long flags;
2414	struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2415	struct rcu_node *rnp = rdp->mynode;
2416	/*
2417	 * On RT rcu_core() can be preempted when IRQs aren't disabled.
2418	 * Therefore this function can race with concurrent NOCB (de-)offloading
2419	 * on this CPU and the below condition must be considered volatile.
2420	 * However if we race with:
2421	 *
2422	 * _ Offloading:   In the worst case we accelerate or process callbacks
2423	 *                 concurrently with NOCB kthreads. We are guaranteed to
2424	 *                 call rcu_nocb_lock() if that happens.
2425	 *
2426	 * _ Deoffloading: In the worst case we miss callbacks acceleration or
2427	 *                 processing. This is fine because the early stage
2428	 *                 of deoffloading invokes rcu_core() after setting
2429	 *                 SEGCBLIST_RCU_CORE. So we guarantee that we'll process
2430	 *                 what could have been dismissed without the need to wait
2431	 *                 for the next rcu_pending() check in the next jiffy.
2432	 */
2433	const bool do_batch = !rcu_segcblist_completely_offloaded(&rdp->cblist);
2434
2435	if (cpu_is_offline(smp_processor_id()))
2436		return;
2437	trace_rcu_utilization(TPS("Start RCU core"));
2438	WARN_ON_ONCE(!rdp->beenonline);
2439
2440	/* Report any deferred quiescent states if preemption enabled. */
2441	if (IS_ENABLED(CONFIG_PREEMPT_COUNT) && (!(preempt_count() & PREEMPT_MASK))) {
2442		rcu_preempt_deferred_qs(current);
2443	} else if (rcu_preempt_need_deferred_qs(current)) {
2444		set_tsk_need_resched(current);
2445		set_preempt_need_resched();
2446	}
2447
2448	/* Update RCU state based on any recent quiescent states. */
2449	rcu_check_quiescent_state(rdp);
2450
2451	/* No grace period and unregistered callbacks? */
2452	if (!rcu_gp_in_progress() &&
2453	    rcu_segcblist_is_enabled(&rdp->cblist) && do_batch) {
2454		rcu_nocb_lock_irqsave(rdp, flags);
2455		if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2456			rcu_accelerate_cbs_unlocked(rnp, rdp);
2457		rcu_nocb_unlock_irqrestore(rdp, flags);
2458	}
2459
2460	rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
2461
2462	/* If there are callbacks ready, invoke them. */
2463	if (do_batch && rcu_segcblist_ready_cbs(&rdp->cblist) &&
2464	    likely(READ_ONCE(rcu_scheduler_fully_active))) {
2465		rcu_do_batch(rdp);
2466		/* Re-invoke RCU core processing if there are callbacks remaining. */
2467		if (rcu_segcblist_ready_cbs(&rdp->cblist))
2468			invoke_rcu_core();
2469	}
2470
2471	/* Do any needed deferred wakeups of rcuo kthreads. */
2472	do_nocb_deferred_wakeup(rdp);
2473	trace_rcu_utilization(TPS("End RCU core"));
2474
2475	// If strict GPs, schedule an RCU reader in a clean environment.
2476	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
2477		queue_work_on(rdp->cpu, rcu_gp_wq, &rdp->strict_work);
2478}
2479
2480static void rcu_core_si(struct softirq_action *h)
2481{
2482	rcu_core();
2483}
2484
2485static void rcu_wake_cond(struct task_struct *t, int status)
2486{
2487	/*
2488	 * If the thread is yielding, only wake it when this
2489	 * is invoked from idle
2490	 */
2491	if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
2492		wake_up_process(t);
2493}
2494
2495static void invoke_rcu_core_kthread(void)
2496{
2497	struct task_struct *t;
2498	unsigned long flags;
2499
2500	local_irq_save(flags);
2501	__this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
2502	t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
2503	if (t != NULL && t != current)
2504		rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
2505	local_irq_restore(flags);
2506}
2507
2508/*
2509 * Wake up this CPU's rcuc kthread to do RCU core processing.
2510 */
2511static void invoke_rcu_core(void)
2512{
2513	if (!cpu_online(smp_processor_id()))
2514		return;
2515	if (use_softirq)
2516		raise_softirq(RCU_SOFTIRQ);
2517	else
2518		invoke_rcu_core_kthread();
2519}
2520
2521static void rcu_cpu_kthread_park(unsigned int cpu)
2522{
2523	per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
2524}
2525
2526static int rcu_cpu_kthread_should_run(unsigned int cpu)
2527{
2528	return __this_cpu_read(rcu_data.rcu_cpu_has_work);
2529}
2530
2531/*
2532 * Per-CPU kernel thread that invokes RCU callbacks.  This replaces
2533 * the RCU softirq used in configurations of RCU that do not support RCU
2534 * priority boosting.
2535 */
2536static void rcu_cpu_kthread(unsigned int cpu)
2537{
2538	unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
2539	char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
2540	unsigned long *j = this_cpu_ptr(&rcu_data.rcuc_activity);
2541	int spincnt;
2542
2543	trace_rcu_utilization(TPS("Start CPU kthread@rcu_run"));
2544	for (spincnt = 0; spincnt < 10; spincnt++) {
2545		WRITE_ONCE(*j, jiffies);
2546		local_bh_disable();
2547		*statusp = RCU_KTHREAD_RUNNING;
2548		local_irq_disable();
2549		work = *workp;
2550		WRITE_ONCE(*workp, 0);
2551		local_irq_enable();
2552		if (work)
2553			rcu_core();
2554		local_bh_enable();
2555		if (!READ_ONCE(*workp)) {
2556			trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2557			*statusp = RCU_KTHREAD_WAITING;
2558			return;
2559		}
2560	}
2561	*statusp = RCU_KTHREAD_YIELDING;
2562	trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2563	schedule_timeout_idle(2);
2564	trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2565	*statusp = RCU_KTHREAD_WAITING;
2566	WRITE_ONCE(*j, jiffies);
2567}
2568
2569static struct smp_hotplug_thread rcu_cpu_thread_spec = {
2570	.store			= &rcu_data.rcu_cpu_kthread_task,
2571	.thread_should_run	= rcu_cpu_kthread_should_run,
2572	.thread_fn		= rcu_cpu_kthread,
2573	.thread_comm		= "rcuc/%u",
2574	.setup			= rcu_cpu_kthread_setup,
2575	.park			= rcu_cpu_kthread_park,
2576};
2577
2578/*
2579 * Spawn per-CPU RCU core processing kthreads.
2580 */
2581static int __init rcu_spawn_core_kthreads(void)
2582{
2583	int cpu;
2584
2585	for_each_possible_cpu(cpu)
2586		per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
2587	if (use_softirq)
2588		return 0;
2589	WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
2590		  "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
2591	return 0;
2592}
2593
2594/*
2595 * Handle any core-RCU processing required by a call_rcu() invocation.
2596 */
2597static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2598			    unsigned long flags)
2599{
2600	/*
2601	 * If called from an extended quiescent state, invoke the RCU
2602	 * core in order to force a re-evaluation of RCU's idleness.
2603	 */
2604	if (!rcu_is_watching())
2605		invoke_rcu_core();
2606
2607	/* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2608	if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2609		return;
2610
2611	/*
2612	 * Force the grace period if too many callbacks or too long waiting.
2613	 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2614	 * if some other CPU has recently done so.  Also, don't bother
2615	 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2616	 * is the only one waiting for a grace period to complete.
2617	 */
2618	if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2619		     rdp->qlen_last_fqs_check + qhimark)) {
2620
2621		/* Are we ignoring a completed grace period? */
2622		note_gp_changes(rdp);
2623
2624		/* Start a new grace period if one not already started. */
2625		if (!rcu_gp_in_progress()) {
2626			rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
2627		} else {
2628			/* Give the grace period a kick. */
2629			rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
2630			if (READ_ONCE(rcu_state.n_force_qs) == rdp->n_force_qs_snap &&
2631			    rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2632				rcu_force_quiescent_state();
2633			rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
2634			rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2635		}
2636	}
2637}
2638
2639/*
2640 * RCU callback function to leak a callback.
2641 */
2642static void rcu_leak_callback(struct rcu_head *rhp)
2643{
2644}
2645
2646/*
2647 * Check and if necessary update the leaf rcu_node structure's
2648 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2649 * number of queued RCU callbacks.  The caller must hold the leaf rcu_node
2650 * structure's ->lock.
2651 */
2652static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp)
2653{
2654	raw_lockdep_assert_held_rcu_node(rnp);
2655	if (qovld_calc <= 0)
2656		return; // Early boot and wildcard value set.
2657	if (rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc)
2658		WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask | rdp->grpmask);
2659	else
2660		WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask & ~rdp->grpmask);
2661}
2662
2663/*
2664 * Check and if necessary update the leaf rcu_node structure's
2665 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2666 * number of queued RCU callbacks.  No locks need be held, but the
2667 * caller must have disabled interrupts.
2668 *
2669 * Note that this function ignores the possibility that there are a lot
2670 * of callbacks all of which have already seen the end of their respective
2671 * grace periods.  This omission is due to the need for no-CBs CPUs to
2672 * be holding ->nocb_lock to do this check, which is too heavy for a
2673 * common-case operation.
2674 */
2675static void check_cb_ovld(struct rcu_data *rdp)
2676{
2677	struct rcu_node *const rnp = rdp->mynode;
2678
2679	if (qovld_calc <= 0 ||
2680	    ((rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) ==
2681	     !!(READ_ONCE(rnp->cbovldmask) & rdp->grpmask)))
2682		return; // Early boot wildcard value or already set correctly.
2683	raw_spin_lock_rcu_node(rnp);
2684	check_cb_ovld_locked(rdp, rnp);
2685	raw_spin_unlock_rcu_node(rnp);
2686}
2687
 
2688static void
2689__call_rcu_common(struct rcu_head *head, rcu_callback_t func, bool lazy_in)
2690{
2691	static atomic_t doublefrees;
2692	unsigned long flags;
2693	bool lazy;
2694	struct rcu_data *rdp;
2695	bool was_alldone;
2696
2697	/* Misaligned rcu_head! */
2698	WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2699
2700	if (debug_rcu_head_queue(head)) {
2701		/*
2702		 * Probable double call_rcu(), so leak the callback.
2703		 * Use rcu:rcu_callback trace event to find the previous
2704		 * time callback was passed to call_rcu().
2705		 */
2706		if (atomic_inc_return(&doublefrees) < 4) {
2707			pr_err("%s(): Double-freed CB %p->%pS()!!!  ", __func__, head, head->func);
2708			mem_dump_obj(head);
2709		}
2710		WRITE_ONCE(head->func, rcu_leak_callback);
2711		return;
2712	}
2713	head->func = func;
2714	head->next = NULL;
2715	kasan_record_aux_stack_noalloc(head);
2716	local_irq_save(flags);
 
2717	rdp = this_cpu_ptr(&rcu_data);
2718	lazy = lazy_in && !rcu_async_should_hurry();
2719
2720	/* Add the callback to our list. */
2721	if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
2722		// This can trigger due to call_rcu() from offline CPU:
2723		WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
2724		WARN_ON_ONCE(!rcu_is_watching());
2725		// Very early boot, before rcu_init().  Initialize if needed
2726		// and then drop through to queue the callback.
2727		if (rcu_segcblist_empty(&rdp->cblist))
2728			rcu_segcblist_init(&rdp->cblist);
2729	}
2730
2731	check_cb_ovld(rdp);
2732	if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags, lazy))
2733		return; // Enqueued onto ->nocb_bypass, so just leave.
2734	// If no-CBs CPU gets here, rcu_nocb_try_bypass() acquired ->nocb_lock.
2735	rcu_segcblist_enqueue(&rdp->cblist, head);
2736	if (__is_kvfree_rcu_offset((unsigned long)func))
2737		trace_rcu_kvfree_callback(rcu_state.name, head,
2738					 (unsigned long)func,
2739					 rcu_segcblist_n_cbs(&rdp->cblist));
2740	else
2741		trace_rcu_callback(rcu_state.name, head,
2742				   rcu_segcblist_n_cbs(&rdp->cblist));
2743
2744	trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCBQueued"));
2745
2746	/* Go handle any RCU core processing required. */
2747	if (unlikely(rcu_rdp_is_offloaded(rdp))) {
2748		__call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */
2749	} else {
2750		__call_rcu_core(rdp, head, flags);
2751		local_irq_restore(flags);
2752	}
2753}
2754
2755#ifdef CONFIG_RCU_LAZY
2756/**
2757 * call_rcu_hurry() - Queue RCU callback for invocation after grace period, and
2758 * flush all lazy callbacks (including the new one) to the main ->cblist while
2759 * doing so.
2760 *
2761 * @head: structure to be used for queueing the RCU updates.
2762 * @func: actual callback function to be invoked after the grace period
2763 *
2764 * The callback function will be invoked some time after a full grace
2765 * period elapses, in other words after all pre-existing RCU read-side
2766 * critical sections have completed.
2767 *
2768 * Use this API instead of call_rcu() if you don't want the callback to be
2769 * invoked after very long periods of time, which can happen on systems without
2770 * memory pressure and on systems which are lightly loaded or mostly idle.
2771 * This function will cause callbacks to be invoked sooner than later at the
2772 * expense of extra power. Other than that, this function is identical to, and
2773 * reuses call_rcu()'s logic. Refer to call_rcu() for more details about memory
2774 * ordering and other functionality.
2775 */
2776void call_rcu_hurry(struct rcu_head *head, rcu_callback_t func)
2777{
2778	__call_rcu_common(head, func, false);
2779}
2780EXPORT_SYMBOL_GPL(call_rcu_hurry);
2781#endif
2782
2783/**
2784 * call_rcu() - Queue an RCU callback for invocation after a grace period.
2785 * By default the callbacks are 'lazy' and are kept hidden from the main
2786 * ->cblist to prevent starting of grace periods too soon.
2787 * If you desire grace periods to start very soon, use call_rcu_hurry().
2788 *
2789 * @head: structure to be used for queueing the RCU updates.
2790 * @func: actual callback function to be invoked after the grace period
2791 *
2792 * The callback function will be invoked some time after a full grace
2793 * period elapses, in other words after all pre-existing RCU read-side
2794 * critical sections have completed.  However, the callback function
2795 * might well execute concurrently with RCU read-side critical sections
2796 * that started after call_rcu() was invoked.
2797 *
2798 * RCU read-side critical sections are delimited by rcu_read_lock()
2799 * and rcu_read_unlock(), and may be nested.  In addition, but only in
2800 * v5.0 and later, regions of code across which interrupts, preemption,
2801 * or softirqs have been disabled also serve as RCU read-side critical
2802 * sections.  This includes hardware interrupt handlers, softirq handlers,
2803 * and NMI handlers.
2804 *
2805 * Note that all CPUs must agree that the grace period extended beyond
2806 * all pre-existing RCU read-side critical section.  On systems with more
2807 * than one CPU, this means that when "func()" is invoked, each CPU is
2808 * guaranteed to have executed a full memory barrier since the end of its
2809 * last RCU read-side critical section whose beginning preceded the call
2810 * to call_rcu().  It also means that each CPU executing an RCU read-side
2811 * critical section that continues beyond the start of "func()" must have
2812 * executed a memory barrier after the call_rcu() but before the beginning
2813 * of that RCU read-side critical section.  Note that these guarantees
2814 * include CPUs that are offline, idle, or executing in user mode, as
2815 * well as CPUs that are executing in the kernel.
2816 *
2817 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
2818 * resulting RCU callback function "func()", then both CPU A and CPU B are
2819 * guaranteed to execute a full memory barrier during the time interval
2820 * between the call to call_rcu() and the invocation of "func()" -- even
2821 * if CPU A and CPU B are the same CPU (but again only if the system has
2822 * more than one CPU).
2823 *
2824 * Implementation of these memory-ordering guarantees is described here:
2825 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
2826 */
2827void call_rcu(struct rcu_head *head, rcu_callback_t func)
2828{
2829	__call_rcu_common(head, func, IS_ENABLED(CONFIG_RCU_LAZY));
2830}
2831EXPORT_SYMBOL_GPL(call_rcu);
2832
 
2833/* Maximum number of jiffies to wait before draining a batch. */
2834#define KFREE_DRAIN_JIFFIES (5 * HZ)
2835#define KFREE_N_BATCHES 2
2836#define FREE_N_CHANNELS 2
2837
2838/**
2839 * struct kvfree_rcu_bulk_data - single block to store kvfree_rcu() pointers
2840 * @list: List node. All blocks are linked between each other
2841 * @gp_snap: Snapshot of RCU state for objects placed to this bulk
2842 * @nr_records: Number of active pointers in the array
 
2843 * @records: Array of the kvfree_rcu() pointers
2844 */
2845struct kvfree_rcu_bulk_data {
2846	struct list_head list;
2847	struct rcu_gp_oldstate gp_snap;
2848	unsigned long nr_records;
 
2849	void *records[];
2850};
2851
2852/*
2853 * This macro defines how many entries the "records" array
2854 * will contain. It is based on the fact that the size of
2855 * kvfree_rcu_bulk_data structure becomes exactly one page.
2856 */
2857#define KVFREE_BULK_MAX_ENTR \
2858	((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *))
2859
2860/**
2861 * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
2862 * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
2863 * @head_free: List of kfree_rcu() objects waiting for a grace period
2864 * @head_free_gp_snap: Grace-period snapshot to check for attempted premature frees.
2865 * @bulk_head_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
2866 * @krcp: Pointer to @kfree_rcu_cpu structure
2867 */
2868
2869struct kfree_rcu_cpu_work {
2870	struct rcu_work rcu_work;
2871	struct rcu_head *head_free;
2872	struct rcu_gp_oldstate head_free_gp_snap;
2873	struct list_head bulk_head_free[FREE_N_CHANNELS];
2874	struct kfree_rcu_cpu *krcp;
2875};
2876
2877/**
2878 * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
2879 * @head: List of kfree_rcu() objects not yet waiting for a grace period
2880 * @head_gp_snap: Snapshot of RCU state for objects placed to "@head"
2881 * @bulk_head: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period
2882 * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
2883 * @lock: Synchronize access to this structure
2884 * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES
 
2885 * @initialized: The @rcu_work fields have been initialized
2886 * @head_count: Number of objects in rcu_head singular list
2887 * @bulk_count: Number of objects in bulk-list
2888 * @bkvcache:
2889 *	A simple cache list that contains objects for reuse purpose.
2890 *	In order to save some per-cpu space the list is singular.
2891 *	Even though it is lockless an access has to be protected by the
2892 *	per-cpu lock.
2893 * @page_cache_work: A work to refill the cache when it is empty
2894 * @backoff_page_cache_fill: Delay cache refills
2895 * @work_in_progress: Indicates that page_cache_work is running
2896 * @hrtimer: A hrtimer for scheduling a page_cache_work
2897 * @nr_bkv_objs: number of allocated objects at @bkvcache.
2898 *
2899 * This is a per-CPU structure.  The reason that it is not included in
2900 * the rcu_data structure is to permit this code to be extracted from
2901 * the RCU files.  Such extraction could allow further optimization of
2902 * the interactions with the slab allocators.
2903 */
2904struct kfree_rcu_cpu {
2905	// Objects queued on a linked list
2906	// through their rcu_head structures.
2907	struct rcu_head *head;
2908	unsigned long head_gp_snap;
2909	atomic_t head_count;
2910
2911	// Objects queued on a bulk-list.
2912	struct list_head bulk_head[FREE_N_CHANNELS];
2913	atomic_t bulk_count[FREE_N_CHANNELS];
2914
2915	struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
2916	raw_spinlock_t lock;
2917	struct delayed_work monitor_work;
 
2918	bool initialized;
 
2919
2920	struct delayed_work page_cache_work;
2921	atomic_t backoff_page_cache_fill;
2922	atomic_t work_in_progress;
2923	struct hrtimer hrtimer;
2924
2925	struct llist_head bkvcache;
2926	int nr_bkv_objs;
2927};
2928
2929static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc) = {
2930	.lock = __RAW_SPIN_LOCK_UNLOCKED(krc.lock),
2931};
2932
2933static __always_inline void
2934debug_rcu_bhead_unqueue(struct kvfree_rcu_bulk_data *bhead)
2935{
2936#ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
2937	int i;
2938
2939	for (i = 0; i < bhead->nr_records; i++)
2940		debug_rcu_head_unqueue((struct rcu_head *)(bhead->records[i]));
2941#endif
2942}
2943
2944static inline struct kfree_rcu_cpu *
2945krc_this_cpu_lock(unsigned long *flags)
2946{
2947	struct kfree_rcu_cpu *krcp;
2948
2949	local_irq_save(*flags);	// For safely calling this_cpu_ptr().
2950	krcp = this_cpu_ptr(&krc);
2951	raw_spin_lock(&krcp->lock);
2952
2953	return krcp;
2954}
2955
2956static inline void
2957krc_this_cpu_unlock(struct kfree_rcu_cpu *krcp, unsigned long flags)
2958{
2959	raw_spin_unlock_irqrestore(&krcp->lock, flags);
2960}
2961
2962static inline struct kvfree_rcu_bulk_data *
2963get_cached_bnode(struct kfree_rcu_cpu *krcp)
2964{
2965	if (!krcp->nr_bkv_objs)
2966		return NULL;
2967
2968	WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs - 1);
2969	return (struct kvfree_rcu_bulk_data *)
2970		llist_del_first(&krcp->bkvcache);
2971}
2972
2973static inline bool
2974put_cached_bnode(struct kfree_rcu_cpu *krcp,
2975	struct kvfree_rcu_bulk_data *bnode)
2976{
2977	// Check the limit.
2978	if (krcp->nr_bkv_objs >= rcu_min_cached_objs)
2979		return false;
2980
2981	llist_add((struct llist_node *) bnode, &krcp->bkvcache);
2982	WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs + 1);
2983	return true;
2984}
2985
2986static int
2987drain_page_cache(struct kfree_rcu_cpu *krcp)
2988{
2989	unsigned long flags;
2990	struct llist_node *page_list, *pos, *n;
2991	int freed = 0;
2992
2993	if (!rcu_min_cached_objs)
2994		return 0;
2995
2996	raw_spin_lock_irqsave(&krcp->lock, flags);
2997	page_list = llist_del_all(&krcp->bkvcache);
2998	WRITE_ONCE(krcp->nr_bkv_objs, 0);
2999	raw_spin_unlock_irqrestore(&krcp->lock, flags);
3000
3001	llist_for_each_safe(pos, n, page_list) {
3002		free_page((unsigned long)pos);
3003		freed++;
3004	}
3005
3006	return freed;
3007}
3008
3009static void
3010kvfree_rcu_bulk(struct kfree_rcu_cpu *krcp,
3011	struct kvfree_rcu_bulk_data *bnode, int idx)
3012{
3013	unsigned long flags;
3014	int i;
3015
3016	if (!WARN_ON_ONCE(!poll_state_synchronize_rcu_full(&bnode->gp_snap))) {
3017		debug_rcu_bhead_unqueue(bnode);
3018		rcu_lock_acquire(&rcu_callback_map);
3019		if (idx == 0) { // kmalloc() / kfree().
3020			trace_rcu_invoke_kfree_bulk_callback(
3021				rcu_state.name, bnode->nr_records,
3022				bnode->records);
3023
3024			kfree_bulk(bnode->nr_records, bnode->records);
3025		} else { // vmalloc() / vfree().
3026			for (i = 0; i < bnode->nr_records; i++) {
3027				trace_rcu_invoke_kvfree_callback(
3028					rcu_state.name, bnode->records[i], 0);
3029
3030				vfree(bnode->records[i]);
3031			}
3032		}
3033		rcu_lock_release(&rcu_callback_map);
3034	}
3035
3036	raw_spin_lock_irqsave(&krcp->lock, flags);
3037	if (put_cached_bnode(krcp, bnode))
3038		bnode = NULL;
3039	raw_spin_unlock_irqrestore(&krcp->lock, flags);
3040
3041	if (bnode)
3042		free_page((unsigned long) bnode);
3043
3044	cond_resched_tasks_rcu_qs();
3045}
3046
3047static void
3048kvfree_rcu_list(struct rcu_head *head)
3049{
3050	struct rcu_head *next;
3051
3052	for (; head; head = next) {
3053		void *ptr = (void *) head->func;
3054		unsigned long offset = (void *) head - ptr;
3055
3056		next = head->next;
3057		debug_rcu_head_unqueue((struct rcu_head *)ptr);
3058		rcu_lock_acquire(&rcu_callback_map);
3059		trace_rcu_invoke_kvfree_callback(rcu_state.name, head, offset);
3060
3061		if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset)))
3062			kvfree(ptr);
3063
3064		rcu_lock_release(&rcu_callback_map);
3065		cond_resched_tasks_rcu_qs();
3066	}
3067}
3068
3069/*
3070 * This function is invoked in workqueue context after a grace period.
3071 * It frees all the objects queued on ->bulk_head_free or ->head_free.
3072 */
3073static void kfree_rcu_work(struct work_struct *work)
3074{
3075	unsigned long flags;
3076	struct kvfree_rcu_bulk_data *bnode, *n;
3077	struct list_head bulk_head[FREE_N_CHANNELS];
3078	struct rcu_head *head;
3079	struct kfree_rcu_cpu *krcp;
3080	struct kfree_rcu_cpu_work *krwp;
3081	struct rcu_gp_oldstate head_gp_snap;
3082	int i;
3083
3084	krwp = container_of(to_rcu_work(work),
3085		struct kfree_rcu_cpu_work, rcu_work);
3086	krcp = krwp->krcp;
3087
3088	raw_spin_lock_irqsave(&krcp->lock, flags);
3089	// Channels 1 and 2.
3090	for (i = 0; i < FREE_N_CHANNELS; i++)
3091		list_replace_init(&krwp->bulk_head_free[i], &bulk_head[i]);
 
 
3092
3093	// Channel 3.
3094	head = krwp->head_free;
3095	krwp->head_free = NULL;
3096	head_gp_snap = krwp->head_free_gp_snap;
3097	raw_spin_unlock_irqrestore(&krcp->lock, flags);
3098
3099	// Handle the first two channels.
3100	for (i = 0; i < FREE_N_CHANNELS; i++) {
3101		// Start from the tail page, so a GP is likely passed for it.
3102		list_for_each_entry_safe(bnode, n, &bulk_head[i], list)
3103			kvfree_rcu_bulk(krcp, bnode, i);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3104	}
3105
3106	/*
3107	 * This is used when the "bulk" path can not be used for the
3108	 * double-argument of kvfree_rcu().  This happens when the
3109	 * page-cache is empty, which means that objects are instead
3110	 * queued on a linked list through their rcu_head structures.
3111	 * This list is named "Channel 3".
3112	 */
3113	if (head && !WARN_ON_ONCE(!poll_state_synchronize_rcu_full(&head_gp_snap)))
3114		kvfree_rcu_list(head);
3115}
3116
3117static bool
3118need_offload_krc(struct kfree_rcu_cpu *krcp)
3119{
3120	int i;
3121
3122	for (i = 0; i < FREE_N_CHANNELS; i++)
3123		if (!list_empty(&krcp->bulk_head[i]))
3124			return true;
3125
3126	return !!READ_ONCE(krcp->head);
3127}
3128
3129static bool
3130need_wait_for_krwp_work(struct kfree_rcu_cpu_work *krwp)
3131{
3132	int i;
3133
3134	for (i = 0; i < FREE_N_CHANNELS; i++)
3135		if (!list_empty(&krwp->bulk_head_free[i]))
3136			return true;
3137
3138	return !!krwp->head_free;
3139}
3140
3141static int krc_count(struct kfree_rcu_cpu *krcp)
3142{
3143	int sum = atomic_read(&krcp->head_count);
3144	int i;
3145
3146	for (i = 0; i < FREE_N_CHANNELS; i++)
3147		sum += atomic_read(&krcp->bulk_count[i]);
3148
3149	return sum;
3150}
3151
3152static void
3153schedule_delayed_monitor_work(struct kfree_rcu_cpu *krcp)
3154{
3155	long delay, delay_left;
3156
3157	delay = krc_count(krcp) >= KVFREE_BULK_MAX_ENTR ? 1:KFREE_DRAIN_JIFFIES;
3158	if (delayed_work_pending(&krcp->monitor_work)) {
3159		delay_left = krcp->monitor_work.timer.expires - jiffies;
3160		if (delay < delay_left)
3161			mod_delayed_work(system_wq, &krcp->monitor_work, delay);
3162		return;
3163	}
3164	queue_delayed_work(system_wq, &krcp->monitor_work, delay);
3165}
3166
3167static void
3168kvfree_rcu_drain_ready(struct kfree_rcu_cpu *krcp)
3169{
3170	struct list_head bulk_ready[FREE_N_CHANNELS];
3171	struct kvfree_rcu_bulk_data *bnode, *n;
3172	struct rcu_head *head_ready = NULL;
3173	unsigned long flags;
3174	int i;
3175
3176	raw_spin_lock_irqsave(&krcp->lock, flags);
3177	for (i = 0; i < FREE_N_CHANNELS; i++) {
3178		INIT_LIST_HEAD(&bulk_ready[i]);
3179
3180		list_for_each_entry_safe_reverse(bnode, n, &krcp->bulk_head[i], list) {
3181			if (!poll_state_synchronize_rcu_full(&bnode->gp_snap))
3182				break;
3183
3184			atomic_sub(bnode->nr_records, &krcp->bulk_count[i]);
3185			list_move(&bnode->list, &bulk_ready[i]);
3186		}
3187	}
3188
3189	if (krcp->head && poll_state_synchronize_rcu(krcp->head_gp_snap)) {
3190		head_ready = krcp->head;
3191		atomic_set(&krcp->head_count, 0);
3192		WRITE_ONCE(krcp->head, NULL);
3193	}
3194	raw_spin_unlock_irqrestore(&krcp->lock, flags);
3195
3196	for (i = 0; i < FREE_N_CHANNELS; i++) {
3197		list_for_each_entry_safe(bnode, n, &bulk_ready[i], list)
3198			kvfree_rcu_bulk(krcp, bnode, i);
3199	}
3200
3201	if (head_ready)
3202		kvfree_rcu_list(head_ready);
3203}
3204
3205/*
3206 * This function is invoked after the KFREE_DRAIN_JIFFIES timeout.
3207 */
3208static void kfree_rcu_monitor(struct work_struct *work)
3209{
3210	struct kfree_rcu_cpu *krcp = container_of(work,
3211		struct kfree_rcu_cpu, monitor_work.work);
3212	unsigned long flags;
3213	int i, j;
3214
3215	// Drain ready for reclaim.
3216	kvfree_rcu_drain_ready(krcp);
3217
3218	raw_spin_lock_irqsave(&krcp->lock, flags);
3219
3220	// Attempt to start a new batch.
3221	for (i = 0; i < KFREE_N_BATCHES; i++) {
3222		struct kfree_rcu_cpu_work *krwp = &(krcp->krw_arr[i]);
3223
3224		// Try to detach bulk_head or head and attach it, only when
3225		// all channels are free.  Any channel is not free means at krwp
3226		// there is on-going rcu work to handle krwp's free business.
3227		if (need_wait_for_krwp_work(krwp))
3228			continue;
3229
3230		// kvfree_rcu_drain_ready() might handle this krcp, if so give up.
3231		if (need_offload_krc(krcp)) {
3232			// Channel 1 corresponds to the SLAB-pointer bulk path.
3233			// Channel 2 corresponds to vmalloc-pointer bulk path.
3234			for (j = 0; j < FREE_N_CHANNELS; j++) {
3235				if (list_empty(&krwp->bulk_head_free[j])) {
3236					atomic_set(&krcp->bulk_count[j], 0);
3237					list_replace_init(&krcp->bulk_head[j],
3238						&krwp->bulk_head_free[j]);
3239				}
3240			}
3241
3242			// Channel 3 corresponds to both SLAB and vmalloc
3243			// objects queued on the linked list.
3244			if (!krwp->head_free) {
3245				krwp->head_free = krcp->head;
3246				get_state_synchronize_rcu_full(&krwp->head_free_gp_snap);
3247				atomic_set(&krcp->head_count, 0);
3248				WRITE_ONCE(krcp->head, NULL);
3249			}
3250
 
 
3251			// One work is per one batch, so there are three
3252			// "free channels", the batch can handle. It can
3253			// be that the work is in the pending state when
3254			// channels have been detached following by each
3255			// other.
3256			queue_rcu_work(system_wq, &krwp->rcu_work);
3257		}
3258	}
3259
3260	raw_spin_unlock_irqrestore(&krcp->lock, flags);
3261
3262	// If there is nothing to detach, it means that our job is
3263	// successfully done here. In case of having at least one
3264	// of the channels that is still busy we should rearm the
3265	// work to repeat an attempt. Because previous batches are
3266	// still in progress.
3267	if (need_offload_krc(krcp))
3268		schedule_delayed_monitor_work(krcp);
 
 
 
 
3269}
3270
3271static enum hrtimer_restart
3272schedule_page_work_fn(struct hrtimer *t)
3273{
3274	struct kfree_rcu_cpu *krcp =
3275		container_of(t, struct kfree_rcu_cpu, hrtimer);
3276
3277	queue_delayed_work(system_highpri_wq, &krcp->page_cache_work, 0);
3278	return HRTIMER_NORESTART;
3279}
3280
3281static void fill_page_cache_func(struct work_struct *work)
3282{
3283	struct kvfree_rcu_bulk_data *bnode;
3284	struct kfree_rcu_cpu *krcp =
3285		container_of(work, struct kfree_rcu_cpu,
3286			page_cache_work.work);
3287	unsigned long flags;
3288	int nr_pages;
3289	bool pushed;
3290	int i;
3291
3292	nr_pages = atomic_read(&krcp->backoff_page_cache_fill) ?
3293		1 : rcu_min_cached_objs;
3294
3295	for (i = READ_ONCE(krcp->nr_bkv_objs); i < nr_pages; i++) {
3296		bnode = (struct kvfree_rcu_bulk_data *)
3297			__get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3298
3299		if (!bnode)
3300			break;
3301
3302		raw_spin_lock_irqsave(&krcp->lock, flags);
3303		pushed = put_cached_bnode(krcp, bnode);
3304		raw_spin_unlock_irqrestore(&krcp->lock, flags);
3305
3306		if (!pushed) {
3307			free_page((unsigned long) bnode);
3308			break;
 
3309		}
3310	}
3311
3312	atomic_set(&krcp->work_in_progress, 0);
3313	atomic_set(&krcp->backoff_page_cache_fill, 0);
3314}
3315
3316static void
3317run_page_cache_worker(struct kfree_rcu_cpu *krcp)
3318{
3319	// If cache disabled, bail out.
3320	if (!rcu_min_cached_objs)
3321		return;
3322
3323	if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
3324			!atomic_xchg(&krcp->work_in_progress, 1)) {
3325		if (atomic_read(&krcp->backoff_page_cache_fill)) {
3326			queue_delayed_work(system_wq,
3327				&krcp->page_cache_work,
3328					msecs_to_jiffies(rcu_delay_page_cache_fill_msec));
3329		} else {
3330			hrtimer_init(&krcp->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3331			krcp->hrtimer.function = schedule_page_work_fn;
3332			hrtimer_start(&krcp->hrtimer, 0, HRTIMER_MODE_REL);
3333		}
3334	}
3335}
3336
3337// Record ptr in a page managed by krcp, with the pre-krc_this_cpu_lock()
3338// state specified by flags.  If can_alloc is true, the caller must
3339// be schedulable and not be holding any locks or mutexes that might be
3340// acquired by the memory allocator or anything that it might invoke.
3341// Returns true if ptr was successfully recorded, else the caller must
3342// use a fallback.
3343static inline bool
3344add_ptr_to_bulk_krc_lock(struct kfree_rcu_cpu **krcp,
3345	unsigned long *flags, void *ptr, bool can_alloc)
3346{
3347	struct kvfree_rcu_bulk_data *bnode;
3348	int idx;
3349
3350	*krcp = krc_this_cpu_lock(flags);
3351	if (unlikely(!(*krcp)->initialized))
3352		return false;
3353
3354	idx = !!is_vmalloc_addr(ptr);
3355	bnode = list_first_entry_or_null(&(*krcp)->bulk_head[idx],
3356		struct kvfree_rcu_bulk_data, list);
3357
3358	/* Check if a new block is required. */
3359	if (!bnode || bnode->nr_records == KVFREE_BULK_MAX_ENTR) {
 
3360		bnode = get_cached_bnode(*krcp);
3361		if (!bnode && can_alloc) {
3362			krc_this_cpu_unlock(*krcp, *flags);
3363
3364			// __GFP_NORETRY - allows a light-weight direct reclaim
3365			// what is OK from minimizing of fallback hitting point of
3366			// view. Apart of that it forbids any OOM invoking what is
3367			// also beneficial since we are about to release memory soon.
3368			//
3369			// __GFP_NOMEMALLOC - prevents from consuming of all the
3370			// memory reserves. Please note we have a fallback path.
3371			//
3372			// __GFP_NOWARN - it is supposed that an allocation can
3373			// be failed under low memory or high memory pressure
3374			// scenarios.
3375			bnode = (struct kvfree_rcu_bulk_data *)
3376				__get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3377			raw_spin_lock_irqsave(&(*krcp)->lock, *flags);
3378		}
3379
3380		if (!bnode)
3381			return false;
3382
3383		// Initialize the new block and attach it.
3384		bnode->nr_records = 0;
3385		list_add(&bnode->list, &(*krcp)->bulk_head[idx]);
 
 
 
3386	}
3387
3388	// Finally insert and update the GP for this page.
3389	bnode->records[bnode->nr_records++] = ptr;
3390	get_state_synchronize_rcu_full(&bnode->gp_snap);
3391	atomic_inc(&(*krcp)->bulk_count[idx]);
3392
3393	return true;
3394}
3395
3396/*
3397 * Queue a request for lazy invocation of the appropriate free routine
3398 * after a grace period.  Please note that three paths are maintained,
3399 * two for the common case using arrays of pointers and a third one that
3400 * is used only when the main paths cannot be used, for example, due to
3401 * memory pressure.
3402 *
3403 * Each kvfree_call_rcu() request is added to a batch. The batch will be drained
3404 * every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch will
3405 * be free'd in workqueue context. This allows us to: batch requests together to
3406 * reduce the number of grace periods during heavy kfree_rcu()/kvfree_rcu() load.
3407 */
3408void kvfree_call_rcu(struct rcu_head *head, void *ptr)
3409{
3410	unsigned long flags;
3411	struct kfree_rcu_cpu *krcp;
3412	bool success;
 
3413
3414	/*
3415	 * Please note there is a limitation for the head-less
3416	 * variant, that is why there is a clear rule for such
3417	 * objects: it can be used from might_sleep() context
3418	 * only. For other places please embed an rcu_head to
3419	 * your data.
3420	 */
3421	if (!head)
 
 
3422		might_sleep();
 
 
3423
3424	// Queue the object but don't yet schedule the batch.
3425	if (debug_rcu_head_queue(ptr)) {
3426		// Probable double kfree_rcu(), just leak.
3427		WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n",
3428			  __func__, head);
3429
3430		// Mark as success and leave.
3431		return;
3432	}
3433
3434	kasan_record_aux_stack_noalloc(ptr);
3435	success = add_ptr_to_bulk_krc_lock(&krcp, &flags, ptr, !head);
3436	if (!success) {
3437		run_page_cache_worker(krcp);
3438
3439		if (head == NULL)
3440			// Inline if kvfree_rcu(one_arg) call.
3441			goto unlock_return;
3442
3443		head->func = ptr;
3444		head->next = krcp->head;
3445		WRITE_ONCE(krcp->head, head);
3446		atomic_inc(&krcp->head_count);
3447
3448		// Take a snapshot for this krcp.
3449		krcp->head_gp_snap = get_state_synchronize_rcu();
3450		success = true;
3451	}
3452
3453	/*
3454	 * The kvfree_rcu() caller considers the pointer freed at this point
3455	 * and likely removes any references to it. Since the actual slab
3456	 * freeing (and kmemleak_free()) is deferred, tell kmemleak to ignore
3457	 * this object (no scanning or false positives reporting).
3458	 */
3459	kmemleak_ignore(ptr);
3460
3461	// Set timer to drain after KFREE_DRAIN_JIFFIES.
3462	if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING)
3463		schedule_delayed_monitor_work(krcp);
 
 
 
3464
3465unlock_return:
3466	krc_this_cpu_unlock(krcp, flags);
3467
3468	/*
3469	 * Inline kvfree() after synchronize_rcu(). We can do
3470	 * it from might_sleep() context only, so the current
3471	 * CPU can pass the QS state.
3472	 */
3473	if (!success) {
3474		debug_rcu_head_unqueue((struct rcu_head *) ptr);
3475		synchronize_rcu();
3476		kvfree(ptr);
3477	}
3478}
3479EXPORT_SYMBOL_GPL(kvfree_call_rcu);
3480
3481static unsigned long
3482kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
3483{
3484	int cpu;
3485	unsigned long count = 0;
3486
3487	/* Snapshot count of all CPUs */
3488	for_each_possible_cpu(cpu) {
3489		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3490
3491		count += krc_count(krcp);
3492		count += READ_ONCE(krcp->nr_bkv_objs);
3493		atomic_set(&krcp->backoff_page_cache_fill, 1);
3494	}
3495
3496	return count == 0 ? SHRINK_EMPTY : count;
3497}
3498
3499static unsigned long
3500kfree_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
3501{
3502	int cpu, freed = 0;
3503
3504	for_each_possible_cpu(cpu) {
3505		int count;
3506		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3507
3508		count = krc_count(krcp);
3509		count += drain_page_cache(krcp);
3510		kfree_rcu_monitor(&krcp->monitor_work.work);
3511
3512		sc->nr_to_scan -= count;
3513		freed += count;
3514
3515		if (sc->nr_to_scan <= 0)
3516			break;
3517	}
3518
3519	return freed == 0 ? SHRINK_STOP : freed;
3520}
3521
 
 
 
 
 
 
 
3522void __init kfree_rcu_scheduler_running(void)
3523{
3524	int cpu;
 
3525
3526	for_each_possible_cpu(cpu) {
3527		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3528
3529		if (need_offload_krc(krcp))
3530			schedule_delayed_monitor_work(krcp);
 
 
 
 
 
 
 
 
3531	}
3532}
3533
3534/*
3535 * During early boot, any blocking grace-period wait automatically
3536 * implies a grace period.
3537 *
3538 * Later on, this could in theory be the case for kernels built with
3539 * CONFIG_SMP=y && CONFIG_PREEMPTION=y running on a single CPU, but this
3540 * is not a common case.  Furthermore, this optimization would cause
3541 * the rcu_gp_oldstate structure to expand by 50%, so this potential
3542 * grace-period optimization is ignored once the scheduler is running.
 
 
3543 */
3544static int rcu_blocking_is_gp(void)
3545{
3546	if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE) {
3547		might_sleep();
3548		return false;
3549	}
3550	return true;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3551}
3552
3553/**
3554 * synchronize_rcu - wait until a grace period has elapsed.
3555 *
3556 * Control will return to the caller some time after a full grace
3557 * period has elapsed, in other words after all currently executing RCU
3558 * read-side critical sections have completed.  Note, however, that
3559 * upon return from synchronize_rcu(), the caller might well be executing
3560 * concurrently with new RCU read-side critical sections that began while
3561 * synchronize_rcu() was waiting.
3562 *
3563 * RCU read-side critical sections are delimited by rcu_read_lock()
3564 * and rcu_read_unlock(), and may be nested.  In addition, but only in
3565 * v5.0 and later, regions of code across which interrupts, preemption,
3566 * or softirqs have been disabled also serve as RCU read-side critical
3567 * sections.  This includes hardware interrupt handlers, softirq handlers,
3568 * and NMI handlers.
3569 *
3570 * Note that this guarantee implies further memory-ordering guarantees.
3571 * On systems with more than one CPU, when synchronize_rcu() returns,
3572 * each CPU is guaranteed to have executed a full memory barrier since
3573 * the end of its last RCU read-side critical section whose beginning
3574 * preceded the call to synchronize_rcu().  In addition, each CPU having
3575 * an RCU read-side critical section that extends beyond the return from
3576 * synchronize_rcu() is guaranteed to have executed a full memory barrier
3577 * after the beginning of synchronize_rcu() and before the beginning of
3578 * that RCU read-side critical section.  Note that these guarantees include
3579 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3580 * that are executing in the kernel.
3581 *
3582 * Furthermore, if CPU A invoked synchronize_rcu(), which returned
3583 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3584 * to have executed a full memory barrier during the execution of
3585 * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
3586 * again only if the system has more than one CPU).
3587 *
3588 * Implementation of these memory-ordering guarantees is described here:
3589 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
3590 */
3591void synchronize_rcu(void)
3592{
3593	unsigned long flags;
3594	struct rcu_node *rnp;
3595
3596	RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3597			 lock_is_held(&rcu_lock_map) ||
3598			 lock_is_held(&rcu_sched_lock_map),
3599			 "Illegal synchronize_rcu() in RCU read-side critical section");
3600	if (!rcu_blocking_is_gp()) {
3601		if (rcu_gp_is_expedited())
3602			synchronize_rcu_expedited();
3603		else
3604			wait_rcu_gp(call_rcu_hurry);
3605		return;
3606	}
3607
3608	// Context allows vacuous grace periods.
3609	// Note well that this code runs with !PREEMPT && !SMP.
3610	// In addition, all code that advances grace periods runs at
3611	// process level.  Therefore, this normal GP overlaps with other
3612	// normal GPs only by being fully nested within them, which allows
3613	// reuse of ->gp_seq_polled_snap.
3614	rcu_poll_gp_seq_start_unlocked(&rcu_state.gp_seq_polled_snap);
3615	rcu_poll_gp_seq_end_unlocked(&rcu_state.gp_seq_polled_snap);
3616
3617	// Update the normal grace-period counters to record
3618	// this grace period, but only those used by the boot CPU.
3619	// The rcu_scheduler_starting() will take care of the rest of
3620	// these counters.
3621	local_irq_save(flags);
3622	WARN_ON_ONCE(num_online_cpus() > 1);
3623	rcu_state.gp_seq += (1 << RCU_SEQ_CTR_SHIFT);
3624	for (rnp = this_cpu_ptr(&rcu_data)->mynode; rnp; rnp = rnp->parent)
3625		rnp->gp_seq_needed = rnp->gp_seq = rcu_state.gp_seq;
3626	local_irq_restore(flags);
3627}
3628EXPORT_SYMBOL_GPL(synchronize_rcu);
3629
3630/**
3631 * get_completed_synchronize_rcu_full - Return a full pre-completed polled state cookie
3632 * @rgosp: Place to put state cookie
3633 *
3634 * Stores into @rgosp a value that will always be treated by functions
3635 * like poll_state_synchronize_rcu_full() as a cookie whose grace period
3636 * has already completed.
3637 */
3638void get_completed_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3639{
3640	rgosp->rgos_norm = RCU_GET_STATE_COMPLETED;
3641	rgosp->rgos_exp = RCU_GET_STATE_COMPLETED;
3642}
3643EXPORT_SYMBOL_GPL(get_completed_synchronize_rcu_full);
3644
3645/**
3646 * get_state_synchronize_rcu - Snapshot current RCU state
3647 *
3648 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3649 * or poll_state_synchronize_rcu() to determine whether or not a full
3650 * grace period has elapsed in the meantime.
3651 */
3652unsigned long get_state_synchronize_rcu(void)
3653{
3654	/*
3655	 * Any prior manipulation of RCU-protected data must happen
3656	 * before the load from ->gp_seq.
3657	 */
3658	smp_mb();  /* ^^^ */
3659	return rcu_seq_snap(&rcu_state.gp_seq_polled);
3660}
3661EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3662
3663/**
3664 * get_state_synchronize_rcu_full - Snapshot RCU state, both normal and expedited
3665 * @rgosp: location to place combined normal/expedited grace-period state
3666 *
3667 * Places the normal and expedited grace-period states in @rgosp.  This
3668 * state value can be passed to a later call to cond_synchronize_rcu_full()
3669 * or poll_state_synchronize_rcu_full() to determine whether or not a
3670 * grace period (whether normal or expedited) has elapsed in the meantime.
3671 * The rcu_gp_oldstate structure takes up twice the memory of an unsigned
3672 * long, but is guaranteed to see all grace periods.  In contrast, the
3673 * combined state occupies less memory, but can sometimes fail to take
3674 * grace periods into account.
3675 *
3676 * This does not guarantee that the needed grace period will actually
3677 * start.
3678 */
3679void get_state_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3680{
3681	struct rcu_node *rnp = rcu_get_root();
3682
3683	/*
3684	 * Any prior manipulation of RCU-protected data must happen
3685	 * before the loads from ->gp_seq and ->expedited_sequence.
3686	 */
3687	smp_mb();  /* ^^^ */
3688	rgosp->rgos_norm = rcu_seq_snap(&rnp->gp_seq);
3689	rgosp->rgos_exp = rcu_seq_snap(&rcu_state.expedited_sequence);
3690}
3691EXPORT_SYMBOL_GPL(get_state_synchronize_rcu_full);
3692
3693/*
3694 * Helper function for start_poll_synchronize_rcu() and
3695 * start_poll_synchronize_rcu_full().
3696 */
3697static void start_poll_synchronize_rcu_common(void)
3698{
3699	unsigned long flags;
 
3700	bool needwake;
3701	struct rcu_data *rdp;
3702	struct rcu_node *rnp;
3703
3704	lockdep_assert_irqs_enabled();
3705	local_irq_save(flags);
3706	rdp = this_cpu_ptr(&rcu_data);
3707	rnp = rdp->mynode;
3708	raw_spin_lock_rcu_node(rnp); // irqs already disabled.
3709	// Note it is possible for a grace period to have elapsed between
3710	// the above call to get_state_synchronize_rcu() and the below call
3711	// to rcu_seq_snap.  This is OK, the worst that happens is that we
3712	// get a grace period that no one needed.  These accesses are ordered
3713	// by smp_mb(), and we are accessing them in the opposite order
3714	// from which they are updated at grace-period start, as required.
3715	needwake = rcu_start_this_gp(rnp, rdp, rcu_seq_snap(&rcu_state.gp_seq));
3716	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3717	if (needwake)
3718		rcu_gp_kthread_wake();
3719}
3720
3721/**
3722 * start_poll_synchronize_rcu - Snapshot and start RCU grace period
3723 *
3724 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3725 * or poll_state_synchronize_rcu() to determine whether or not a full
3726 * grace period has elapsed in the meantime.  If the needed grace period
3727 * is not already slated to start, notifies RCU core of the need for that
3728 * grace period.
3729 *
3730 * Interrupts must be enabled for the case where it is necessary to awaken
3731 * the grace-period kthread.
3732 */
3733unsigned long start_poll_synchronize_rcu(void)
3734{
3735	unsigned long gp_seq = get_state_synchronize_rcu();
3736
3737	start_poll_synchronize_rcu_common();
3738	return gp_seq;
3739}
3740EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu);
3741
3742/**
3743 * start_poll_synchronize_rcu_full - Take a full snapshot and start RCU grace period
3744 * @rgosp: value from get_state_synchronize_rcu_full() or start_poll_synchronize_rcu_full()
3745 *
3746 * Places the normal and expedited grace-period states in *@rgos.  This
3747 * state value can be passed to a later call to cond_synchronize_rcu_full()
3748 * or poll_state_synchronize_rcu_full() to determine whether or not a
3749 * grace period (whether normal or expedited) has elapsed in the meantime.
3750 * If the needed grace period is not already slated to start, notifies
3751 * RCU core of the need for that grace period.
3752 *
3753 * Interrupts must be enabled for the case where it is necessary to awaken
3754 * the grace-period kthread.
3755 */
3756void start_poll_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3757{
3758	get_state_synchronize_rcu_full(rgosp);
3759
3760	start_poll_synchronize_rcu_common();
3761}
3762EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_full);
3763
3764/**
3765 * poll_state_synchronize_rcu - Has the specified RCU grace period completed?
3766 * @oldstate: value from get_state_synchronize_rcu() or start_poll_synchronize_rcu()
3767 *
3768 * If a full RCU grace period has elapsed since the earlier call from
3769 * which @oldstate was obtained, return @true, otherwise return @false.
3770 * If @false is returned, it is the caller's responsibility to invoke this
3771 * function later on until it does return @true.  Alternatively, the caller
3772 * can explicitly wait for a grace period, for example, by passing @oldstate
3773 * to either cond_synchronize_rcu() or cond_synchronize_rcu_expedited()
3774 * on the one hand or by directly invoking either synchronize_rcu() or
3775 * synchronize_rcu_expedited() on the other.
3776 *
3777 * Yes, this function does not take counter wrap into account.
3778 * But counter wrap is harmless.  If the counter wraps, we have waited for
3779 * more than a billion grace periods (and way more on a 64-bit system!).
3780 * Those needing to keep old state values for very long time periods
3781 * (many hours even on 32-bit systems) should check them occasionally and
3782 * either refresh them or set a flag indicating that the grace period has
3783 * completed.  Alternatively, they can use get_completed_synchronize_rcu()
3784 * to get a guaranteed-completed grace-period state.
3785 *
3786 * In addition, because oldstate compresses the grace-period state for
3787 * both normal and expedited grace periods into a single unsigned long,
3788 * it can miss a grace period when synchronize_rcu() runs concurrently
3789 * with synchronize_rcu_expedited().  If this is unacceptable, please
3790 * instead use the _full() variant of these polling APIs.
3791 *
3792 * This function provides the same memory-ordering guarantees that
3793 * would be provided by a synchronize_rcu() that was invoked at the call
3794 * to the function that provided @oldstate, and that returned at the end
3795 * of this function.
3796 */
3797bool poll_state_synchronize_rcu(unsigned long oldstate)
3798{
3799	if (oldstate == RCU_GET_STATE_COMPLETED ||
3800	    rcu_seq_done_exact(&rcu_state.gp_seq_polled, oldstate)) {
3801		smp_mb(); /* Ensure GP ends before subsequent accesses. */
3802		return true;
3803	}
3804	return false;
3805}
3806EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu);
3807
3808/**
3809 * poll_state_synchronize_rcu_full - Has the specified RCU grace period completed?
3810 * @rgosp: value from get_state_synchronize_rcu_full() or start_poll_synchronize_rcu_full()
3811 *
3812 * If a full RCU grace period has elapsed since the earlier call from
3813 * which *rgosp was obtained, return @true, otherwise return @false.
3814 * If @false is returned, it is the caller's responsibility to invoke this
3815 * function later on until it does return @true.  Alternatively, the caller
3816 * can explicitly wait for a grace period, for example, by passing @rgosp
3817 * to cond_synchronize_rcu() or by directly invoking synchronize_rcu().
3818 *
3819 * Yes, this function does not take counter wrap into account.
3820 * But counter wrap is harmless.  If the counter wraps, we have waited
3821 * for more than a billion grace periods (and way more on a 64-bit
3822 * system!).  Those needing to keep rcu_gp_oldstate values for very
3823 * long time periods (many hours even on 32-bit systems) should check
3824 * them occasionally and either refresh them or set a flag indicating
3825 * that the grace period has completed.  Alternatively, they can use
3826 * get_completed_synchronize_rcu_full() to get a guaranteed-completed
3827 * grace-period state.
3828 *
3829 * This function provides the same memory-ordering guarantees that would
3830 * be provided by a synchronize_rcu() that was invoked at the call to
3831 * the function that provided @rgosp, and that returned at the end of this
3832 * function.  And this guarantee requires that the root rcu_node structure's
3833 * ->gp_seq field be checked instead of that of the rcu_state structure.
3834 * The problem is that the just-ending grace-period's callbacks can be
3835 * invoked between the time that the root rcu_node structure's ->gp_seq
3836 * field is updated and the time that the rcu_state structure's ->gp_seq
3837 * field is updated.  Therefore, if a single synchronize_rcu() is to
3838 * cause a subsequent poll_state_synchronize_rcu_full() to return @true,
3839 * then the root rcu_node structure is the one that needs to be polled.
3840 */
3841bool poll_state_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3842{
3843	struct rcu_node *rnp = rcu_get_root();
3844
3845	smp_mb(); // Order against root rcu_node structure grace-period cleanup.
3846	if (rgosp->rgos_norm == RCU_GET_STATE_COMPLETED ||
3847	    rcu_seq_done_exact(&rnp->gp_seq, rgosp->rgos_norm) ||
3848	    rgosp->rgos_exp == RCU_GET_STATE_COMPLETED ||
3849	    rcu_seq_done_exact(&rcu_state.expedited_sequence, rgosp->rgos_exp)) {
3850		smp_mb(); /* Ensure GP ends before subsequent accesses. */
3851		return true;
3852	}
3853	return false;
3854}
3855EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu_full);
3856
3857/**
3858 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3859 * @oldstate: value from get_state_synchronize_rcu(), start_poll_synchronize_rcu(), or start_poll_synchronize_rcu_expedited()
 
3860 *
3861 * If a full RCU grace period has elapsed since the earlier call to
3862 * get_state_synchronize_rcu() or start_poll_synchronize_rcu(), just return.
3863 * Otherwise, invoke synchronize_rcu() to wait for a full grace period.
3864 *
3865 * Yes, this function does not take counter wrap into account.
3866 * But counter wrap is harmless.  If the counter wraps, we have waited for
3867 * more than 2 billion grace periods (and way more on a 64-bit system!),
3868 * so waiting for a couple of additional grace periods should be just fine.
3869 *
3870 * This function provides the same memory-ordering guarantees that
3871 * would be provided by a synchronize_rcu() that was invoked at the call
3872 * to the function that provided @oldstate and that returned at the end
3873 * of this function.
3874 */
3875void cond_synchronize_rcu(unsigned long oldstate)
3876{
3877	if (!poll_state_synchronize_rcu(oldstate))
3878		synchronize_rcu();
3879}
3880EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3881
3882/**
3883 * cond_synchronize_rcu_full - Conditionally wait for an RCU grace period
3884 * @rgosp: value from get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(), or start_poll_synchronize_rcu_expedited_full()
3885 *
3886 * If a full RCU grace period has elapsed since the call to
3887 * get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(),
3888 * or start_poll_synchronize_rcu_expedited_full() from which @rgosp was
3889 * obtained, just return.  Otherwise, invoke synchronize_rcu() to wait
3890 * for a full grace period.
3891 *
3892 * Yes, this function does not take counter wrap into account.
3893 * But counter wrap is harmless.  If the counter wraps, we have waited for
3894 * more than 2 billion grace periods (and way more on a 64-bit system!),
3895 * so waiting for a couple of additional grace periods should be just fine.
3896 *
3897 * This function provides the same memory-ordering guarantees that
3898 * would be provided by a synchronize_rcu() that was invoked at the call
3899 * to the function that provided @rgosp and that returned at the end of
3900 * this function.
3901 */
3902void cond_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
3903{
3904	if (!poll_state_synchronize_rcu_full(rgosp))
3905		synchronize_rcu();
3906}
3907EXPORT_SYMBOL_GPL(cond_synchronize_rcu_full);
3908
3909/*
3910 * Check to see if there is any immediate RCU-related work to be done by
3911 * the current CPU, returning 1 if so and zero otherwise.  The checks are
3912 * in order of increasing expense: checks that can be carried out against
3913 * CPU-local state are performed first.  However, we must check for CPU
3914 * stalls first, else we might not get a chance.
3915 */
3916static int rcu_pending(int user)
3917{
3918	bool gp_in_progress;
3919	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
3920	struct rcu_node *rnp = rdp->mynode;
3921
3922	lockdep_assert_irqs_disabled();
3923
3924	/* Check for CPU stalls, if enabled. */
3925	check_cpu_stall(rdp);
3926
3927	/* Does this CPU need a deferred NOCB wakeup? */
3928	if (rcu_nocb_need_deferred_wakeup(rdp, RCU_NOCB_WAKE))
3929		return 1;
3930
3931	/* Is this a nohz_full CPU in userspace or idle?  (Ignore RCU if so.) */
3932	if ((user || rcu_is_cpu_rrupt_from_idle()) && rcu_nohz_full_cpu())
3933		return 0;
3934
3935	/* Is the RCU core waiting for a quiescent state from this CPU? */
3936	gp_in_progress = rcu_gp_in_progress();
3937	if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm && gp_in_progress)
3938		return 1;
3939
3940	/* Does this CPU have callbacks ready to invoke? */
3941	if (!rcu_rdp_is_offloaded(rdp) &&
3942	    rcu_segcblist_ready_cbs(&rdp->cblist))
3943		return 1;
3944
3945	/* Has RCU gone idle with this CPU needing another grace period? */
3946	if (!gp_in_progress && rcu_segcblist_is_enabled(&rdp->cblist) &&
3947	    !rcu_rdp_is_offloaded(rdp) &&
3948	    !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
3949		return 1;
3950
3951	/* Have RCU grace period completed or started?  */
3952	if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
3953	    unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
3954		return 1;
3955
3956	/* nothing to do */
3957	return 0;
3958}
3959
3960/*
3961 * Helper function for rcu_barrier() tracing.  If tracing is disabled,
3962 * the compiler is expected to optimize this away.
3963 */
3964static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
3965{
3966	trace_rcu_barrier(rcu_state.name, s, cpu,
3967			  atomic_read(&rcu_state.barrier_cpu_count), done);
3968}
3969
3970/*
3971 * RCU callback function for rcu_barrier().  If we are last, wake
3972 * up the task executing rcu_barrier().
3973 *
3974 * Note that the value of rcu_state.barrier_sequence must be captured
3975 * before the atomic_dec_and_test().  Otherwise, if this CPU is not last,
3976 * other CPUs might count the value down to zero before this CPU gets
3977 * around to invoking rcu_barrier_trace(), which might result in bogus
3978 * data from the next instance of rcu_barrier().
3979 */
3980static void rcu_barrier_callback(struct rcu_head *rhp)
3981{
3982	unsigned long __maybe_unused s = rcu_state.barrier_sequence;
3983
3984	if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
3985		rcu_barrier_trace(TPS("LastCB"), -1, s);
3986		complete(&rcu_state.barrier_completion);
3987	} else {
3988		rcu_barrier_trace(TPS("CB"), -1, s);
3989	}
3990}
3991
3992/*
3993 * If needed, entrain an rcu_barrier() callback on rdp->cblist.
3994 */
3995static void rcu_barrier_entrain(struct rcu_data *rdp)
3996{
3997	unsigned long gseq = READ_ONCE(rcu_state.barrier_sequence);
3998	unsigned long lseq = READ_ONCE(rdp->barrier_seq_snap);
3999	bool wake_nocb = false;
4000	bool was_alldone = false;
4001
4002	lockdep_assert_held(&rcu_state.barrier_lock);
4003	if (rcu_seq_state(lseq) || !rcu_seq_state(gseq) || rcu_seq_ctr(lseq) != rcu_seq_ctr(gseq))
4004		return;
4005	rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
4006	rdp->barrier_head.func = rcu_barrier_callback;
4007	debug_rcu_head_queue(&rdp->barrier_head);
4008	rcu_nocb_lock(rdp);
4009	/*
4010	 * Flush bypass and wakeup rcuog if we add callbacks to an empty regular
4011	 * queue. This way we don't wait for bypass timer that can reach seconds
4012	 * if it's fully lazy.
4013	 */
4014	was_alldone = rcu_rdp_is_offloaded(rdp) && !rcu_segcblist_pend_cbs(&rdp->cblist);
4015	WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies, false));
4016	wake_nocb = was_alldone && rcu_segcblist_pend_cbs(&rdp->cblist);
4017	if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head)) {
4018		atomic_inc(&rcu_state.barrier_cpu_count);
4019	} else {
4020		debug_rcu_head_unqueue(&rdp->barrier_head);
4021		rcu_barrier_trace(TPS("IRQNQ"), -1, rcu_state.barrier_sequence);
 
4022	}
4023	rcu_nocb_unlock(rdp);
4024	if (wake_nocb)
4025		wake_nocb_gp(rdp, false);
4026	smp_store_release(&rdp->barrier_seq_snap, gseq);
4027}
4028
4029/*
4030 * Called with preemption disabled, and from cross-cpu IRQ context.
4031 */
4032static void rcu_barrier_handler(void *cpu_in)
4033{
4034	uintptr_t cpu = (uintptr_t)cpu_in;
4035	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4036
4037	lockdep_assert_irqs_disabled();
4038	WARN_ON_ONCE(cpu != rdp->cpu);
4039	WARN_ON_ONCE(cpu != smp_processor_id());
4040	raw_spin_lock(&rcu_state.barrier_lock);
4041	rcu_barrier_entrain(rdp);
4042	raw_spin_unlock(&rcu_state.barrier_lock);
4043}
4044
4045/**
4046 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
4047 *
4048 * Note that this primitive does not necessarily wait for an RCU grace period
4049 * to complete.  For example, if there are no RCU callbacks queued anywhere
4050 * in the system, then rcu_barrier() is within its rights to return
4051 * immediately, without waiting for anything, much less an RCU grace period.
4052 */
4053void rcu_barrier(void)
4054{
4055	uintptr_t cpu;
4056	unsigned long flags;
4057	unsigned long gseq;
4058	struct rcu_data *rdp;
4059	unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
4060
4061	rcu_barrier_trace(TPS("Begin"), -1, s);
4062
4063	/* Take mutex to serialize concurrent rcu_barrier() requests. */
4064	mutex_lock(&rcu_state.barrier_mutex);
4065
4066	/* Did someone else do our work for us? */
4067	if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
4068		rcu_barrier_trace(TPS("EarlyExit"), -1, rcu_state.barrier_sequence);
 
4069		smp_mb(); /* caller's subsequent code after above check. */
4070		mutex_unlock(&rcu_state.barrier_mutex);
4071		return;
4072	}
4073
4074	/* Mark the start of the barrier operation. */
4075	raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
4076	rcu_seq_start(&rcu_state.barrier_sequence);
4077	gseq = rcu_state.barrier_sequence;
4078	rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
4079
4080	/*
4081	 * Initialize the count to two rather than to zero in order
4082	 * to avoid a too-soon return to zero in case of an immediate
4083	 * invocation of the just-enqueued callback (or preemption of
4084	 * this task).  Exclude CPU-hotplug operations to ensure that no
4085	 * offline non-offloaded CPU has callbacks queued.
4086	 */
4087	init_completion(&rcu_state.barrier_completion);
4088	atomic_set(&rcu_state.barrier_cpu_count, 2);
4089	raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4090
4091	/*
4092	 * Force each CPU with callbacks to register a new callback.
4093	 * When that callback is invoked, we will know that all of the
4094	 * corresponding CPU's preceding callbacks have been invoked.
4095	 */
4096	for_each_possible_cpu(cpu) {
4097		rdp = per_cpu_ptr(&rcu_data, cpu);
4098retry:
4099		if (smp_load_acquire(&rdp->barrier_seq_snap) == gseq)
4100			continue;
4101		raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
4102		if (!rcu_segcblist_n_cbs(&rdp->cblist)) {
4103			WRITE_ONCE(rdp->barrier_seq_snap, gseq);
4104			raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4105			rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence);
4106			continue;
4107		}
4108		if (!rcu_rdp_cpu_online(rdp)) {
4109			rcu_barrier_entrain(rdp);
4110			WARN_ON_ONCE(READ_ONCE(rdp->barrier_seq_snap) != gseq);
4111			raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4112			rcu_barrier_trace(TPS("OfflineNoCBQ"), cpu, rcu_state.barrier_sequence);
4113			continue;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4114		}
4115		raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4116		if (smp_call_function_single(cpu, rcu_barrier_handler, (void *)cpu, 1)) {
4117			schedule_timeout_uninterruptible(1);
4118			goto retry;
4119		}
4120		WARN_ON_ONCE(READ_ONCE(rdp->barrier_seq_snap) != gseq);
4121		rcu_barrier_trace(TPS("OnlineQ"), cpu, rcu_state.barrier_sequence);
4122	}
 
4123
4124	/*
4125	 * Now that we have an rcu_barrier_callback() callback on each
4126	 * CPU, and thus each counted, remove the initial count.
4127	 */
4128	if (atomic_sub_and_test(2, &rcu_state.barrier_cpu_count))
4129		complete(&rcu_state.barrier_completion);
4130
4131	/* Wait for all rcu_barrier_callback() callbacks to be invoked. */
4132	wait_for_completion(&rcu_state.barrier_completion);
4133
4134	/* Mark the end of the barrier operation. */
4135	rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
4136	rcu_seq_end(&rcu_state.barrier_sequence);
4137	gseq = rcu_state.barrier_sequence;
4138	for_each_possible_cpu(cpu) {
4139		rdp = per_cpu_ptr(&rcu_data, cpu);
4140
4141		WRITE_ONCE(rdp->barrier_seq_snap, gseq);
4142	}
4143
4144	/* Other rcu_barrier() invocations can now safely proceed. */
4145	mutex_unlock(&rcu_state.barrier_mutex);
4146}
4147EXPORT_SYMBOL_GPL(rcu_barrier);
4148
4149static unsigned long rcu_barrier_last_throttle;
4150
4151/**
4152 * rcu_barrier_throttled - Do rcu_barrier(), but limit to one per second
4153 *
4154 * This can be thought of as guard rails around rcu_barrier() that
4155 * permits unrestricted userspace use, at least assuming the hardware's
4156 * try_cmpxchg() is robust.  There will be at most one call per second to
4157 * rcu_barrier() system-wide from use of this function, which means that
4158 * callers might needlessly wait a second or three.
4159 *
4160 * This is intended for use by test suites to avoid OOM by flushing RCU
4161 * callbacks from the previous test before starting the next.  See the
4162 * rcutree.do_rcu_barrier module parameter for more information.
4163 *
4164 * Why not simply make rcu_barrier() more scalable?  That might be
4165 * the eventual endpoint, but let's keep it simple for the time being.
4166 * Note that the module parameter infrastructure serializes calls to a
4167 * given .set() function, but should concurrent .set() invocation ever be
4168 * possible, we are ready!
4169 */
4170static void rcu_barrier_throttled(void)
4171{
4172	unsigned long j = jiffies;
4173	unsigned long old = READ_ONCE(rcu_barrier_last_throttle);
4174	unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
4175
4176	while (time_in_range(j, old, old + HZ / 16) ||
4177	       !try_cmpxchg(&rcu_barrier_last_throttle, &old, j)) {
4178		schedule_timeout_idle(HZ / 16);
4179		if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
4180			smp_mb(); /* caller's subsequent code after above check. */
4181			return;
4182		}
4183		j = jiffies;
4184		old = READ_ONCE(rcu_barrier_last_throttle);
4185	}
4186	rcu_barrier();
4187}
4188
4189/*
4190 * Invoke rcu_barrier_throttled() when a rcutree.do_rcu_barrier
4191 * request arrives.  We insist on a true value to allow for possible
4192 * future expansion.
4193 */
4194static int param_set_do_rcu_barrier(const char *val, const struct kernel_param *kp)
4195{
4196	bool b;
4197	int ret;
4198
4199	if (rcu_scheduler_active != RCU_SCHEDULER_RUNNING)
4200		return -EAGAIN;
4201	ret = kstrtobool(val, &b);
4202	if (!ret && b) {
4203		atomic_inc((atomic_t *)kp->arg);
4204		rcu_barrier_throttled();
4205		atomic_dec((atomic_t *)kp->arg);
4206	}
4207	return ret;
4208}
4209
4210/*
4211 * Output the number of outstanding rcutree.do_rcu_barrier requests.
4212 */
4213static int param_get_do_rcu_barrier(char *buffer, const struct kernel_param *kp)
4214{
4215	return sprintf(buffer, "%d\n", atomic_read((atomic_t *)kp->arg));
4216}
4217
4218static const struct kernel_param_ops do_rcu_barrier_ops = {
4219	.set = param_set_do_rcu_barrier,
4220	.get = param_get_do_rcu_barrier,
4221};
4222static atomic_t do_rcu_barrier;
4223module_param_cb(do_rcu_barrier, &do_rcu_barrier_ops, &do_rcu_barrier, 0644);
4224
4225/*
4226 * Compute the mask of online CPUs for the specified rcu_node structure.
4227 * This will not be stable unless the rcu_node structure's ->lock is
4228 * held, but the bit corresponding to the current CPU will be stable
4229 * in most contexts.
4230 */
4231static unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
4232{
4233	return READ_ONCE(rnp->qsmaskinitnext);
4234}
4235
4236/*
4237 * Is the CPU corresponding to the specified rcu_data structure online
4238 * from RCU's perspective?  This perspective is given by that structure's
4239 * ->qsmaskinitnext field rather than by the global cpu_online_mask.
4240 */
4241static bool rcu_rdp_cpu_online(struct rcu_data *rdp)
4242{
4243	return !!(rdp->grpmask & rcu_rnp_online_cpus(rdp->mynode));
4244}
4245
4246bool rcu_cpu_online(int cpu)
4247{
4248	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4249
4250	return rcu_rdp_cpu_online(rdp);
4251}
4252
4253#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
4254
4255/*
4256 * Is the current CPU online as far as RCU is concerned?
4257 *
4258 * Disable preemption to avoid false positives that could otherwise
4259 * happen due to the current CPU number being sampled, this task being
4260 * preempted, its old CPU being taken offline, resuming on some other CPU,
4261 * then determining that its old CPU is now offline.
4262 *
4263 * Disable checking if in an NMI handler because we cannot safely
4264 * report errors from NMI handlers anyway.  In addition, it is OK to use
4265 * RCU on an offline processor during initial boot, hence the check for
4266 * rcu_scheduler_fully_active.
4267 */
4268bool rcu_lockdep_current_cpu_online(void)
4269{
4270	struct rcu_data *rdp;
4271	bool ret = false;
4272
4273	if (in_nmi() || !rcu_scheduler_fully_active)
4274		return true;
4275	preempt_disable_notrace();
4276	rdp = this_cpu_ptr(&rcu_data);
4277	/*
4278	 * Strictly, we care here about the case where the current CPU is
4279	 * in rcutree_report_cpu_starting() and thus has an excuse for rdp->grpmask
4280	 * not being up to date. So arch_spin_is_locked() might have a
4281	 * false positive if it's held by some *other* CPU, but that's
4282	 * OK because that just means a false *negative* on the warning.
4283	 */
4284	if (rcu_rdp_cpu_online(rdp) || arch_spin_is_locked(&rcu_state.ofl_lock))
4285		ret = true;
4286	preempt_enable_notrace();
4287	return ret;
4288}
4289EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
4290
4291#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
4292
4293// Has rcu_init() been invoked?  This is used (for example) to determine
4294// whether spinlocks may be acquired safely.
4295static bool rcu_init_invoked(void)
4296{
4297	return !!rcu_state.n_online_cpus;
4298}
4299
4300/*
4301 * All CPUs for the specified rcu_node structure have gone offline,
4302 * and all tasks that were preempted within an RCU read-side critical
4303 * section while running on one of those CPUs have since exited their RCU
4304 * read-side critical section.  Some other CPU is reporting this fact with
4305 * the specified rcu_node structure's ->lock held and interrupts disabled.
4306 * This function therefore goes up the tree of rcu_node structures,
4307 * clearing the corresponding bits in the ->qsmaskinit fields.  Note that
4308 * the leaf rcu_node structure's ->qsmaskinit field has already been
4309 * updated.
4310 *
4311 * This function does check that the specified rcu_node structure has
4312 * all CPUs offline and no blocked tasks, so it is OK to invoke it
4313 * prematurely.  That said, invoking it after the fact will cost you
4314 * a needless lock acquisition.  So once it has done its work, don't
4315 * invoke it again.
4316 */
4317static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
4318{
4319	long mask;
4320	struct rcu_node *rnp = rnp_leaf;
4321
4322	raw_lockdep_assert_held_rcu_node(rnp_leaf);
4323	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
4324	    WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
4325	    WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
4326		return;
4327	for (;;) {
4328		mask = rnp->grpmask;
4329		rnp = rnp->parent;
4330		if (!rnp)
4331			break;
4332		raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
4333		rnp->qsmaskinit &= ~mask;
4334		/* Between grace periods, so better already be zero! */
4335		WARN_ON_ONCE(rnp->qsmask);
4336		if (rnp->qsmaskinit) {
4337			raw_spin_unlock_rcu_node(rnp);
4338			/* irqs remain disabled. */
4339			return;
4340		}
4341		raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
4342	}
4343}
4344
4345/*
4346 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
4347 * first CPU in a given leaf rcu_node structure coming online.  The caller
4348 * must hold the corresponding leaf rcu_node ->lock with interrupts
4349 * disabled.
4350 */
4351static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
4352{
4353	long mask;
4354	long oldmask;
4355	struct rcu_node *rnp = rnp_leaf;
4356
4357	raw_lockdep_assert_held_rcu_node(rnp_leaf);
4358	WARN_ON_ONCE(rnp->wait_blkd_tasks);
4359	for (;;) {
4360		mask = rnp->grpmask;
4361		rnp = rnp->parent;
4362		if (rnp == NULL)
4363			return;
4364		raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
4365		oldmask = rnp->qsmaskinit;
4366		rnp->qsmaskinit |= mask;
4367		raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
4368		if (oldmask)
4369			return;
4370	}
4371}
4372
4373/*
4374 * Do boot-time initialization of a CPU's per-CPU RCU data.
4375 */
4376static void __init
4377rcu_boot_init_percpu_data(int cpu)
4378{
4379	struct context_tracking *ct = this_cpu_ptr(&context_tracking);
4380	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4381
4382	/* Set up local state, ensuring consistent view of global state. */
4383	rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
4384	INIT_WORK(&rdp->strict_work, strict_work_handler);
4385	WARN_ON_ONCE(ct->dynticks_nesting != 1);
4386	WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(cpu)));
4387	rdp->barrier_seq_snap = rcu_state.barrier_sequence;
4388	rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
4389	rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
4390	rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
4391	rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
4392	rdp->last_sched_clock = jiffies;
4393	rdp->cpu = cpu;
4394	rcu_boot_init_nocb_percpu_data(rdp);
4395}
4396
4397/*
4398 * Invoked early in the CPU-online process, when pretty much all services
4399 * are available.  The incoming CPU is not present.
4400 *
4401 * Initializes a CPU's per-CPU RCU data.  Note that only one online or
4402 * offline event can be happening at a given time.  Note also that we can
4403 * accept some slop in the rsp->gp_seq access due to the fact that this
4404 * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
4405 * And any offloaded callbacks are being numbered elsewhere.
4406 */
4407int rcutree_prepare_cpu(unsigned int cpu)
4408{
4409	unsigned long flags;
4410	struct context_tracking *ct = per_cpu_ptr(&context_tracking, cpu);
4411	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4412	struct rcu_node *rnp = rcu_get_root();
4413
4414	/* Set up local state, ensuring consistent view of global state. */
4415	raw_spin_lock_irqsave_rcu_node(rnp, flags);
4416	rdp->qlen_last_fqs_check = 0;
4417	rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
4418	rdp->blimit = blimit;
4419	ct->dynticks_nesting = 1;	/* CPU not up, no tearing. */
 
4420	raw_spin_unlock_rcu_node(rnp);		/* irqs remain disabled. */
4421
4422	/*
4423	 * Only non-NOCB CPUs that didn't have early-boot callbacks need to be
4424	 * (re-)initialized.
4425	 */
4426	if (!rcu_segcblist_is_enabled(&rdp->cblist))
4427		rcu_segcblist_init(&rdp->cblist);  /* Re-enable callbacks. */
4428
4429	/*
4430	 * Add CPU to leaf rcu_node pending-online bitmask.  Any needed
4431	 * propagation up the rcu_node tree will happen at the beginning
4432	 * of the next grace period.
4433	 */
4434	rnp = rdp->mynode;
4435	raw_spin_lock_rcu_node(rnp);		/* irqs already disabled. */
 
4436	rdp->gp_seq = READ_ONCE(rnp->gp_seq);
4437	rdp->gp_seq_needed = rdp->gp_seq;
4438	rdp->cpu_no_qs.b.norm = true;
4439	rdp->core_needs_qs = false;
4440	rdp->rcu_iw_pending = false;
4441	rdp->rcu_iw = IRQ_WORK_INIT_HARD(rcu_iw_handler);
4442	rdp->rcu_iw_gp_seq = rdp->gp_seq - 1;
4443	trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
4444	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4445	rcu_spawn_one_boost_kthread(rnp);
4446	rcu_spawn_cpu_nocb_kthread(cpu);
4447	WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus + 1);
4448
4449	return 0;
4450}
4451
4452/*
4453 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
4454 */
4455static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
4456{
4457	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4458
4459	rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
4460}
4461
4462/*
4463 * Has the specified (known valid) CPU ever been fully online?
4464 */
4465bool rcu_cpu_beenfullyonline(int cpu)
4466{
4467	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4468
4469	return smp_load_acquire(&rdp->beenonline);
4470}
4471
4472/*
4473 * Near the end of the CPU-online process.  Pretty much all services
4474 * enabled, and the CPU is now very much alive.
4475 */
4476int rcutree_online_cpu(unsigned int cpu)
4477{
4478	unsigned long flags;
4479	struct rcu_data *rdp;
4480	struct rcu_node *rnp;
4481
4482	rdp = per_cpu_ptr(&rcu_data, cpu);
4483	rnp = rdp->mynode;
4484	raw_spin_lock_irqsave_rcu_node(rnp, flags);
4485	rnp->ffmask |= rdp->grpmask;
4486	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4487	if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
4488		return 0; /* Too early in boot for scheduler work. */
4489	sync_sched_exp_online_cleanup(cpu);
4490	rcutree_affinity_setting(cpu, -1);
4491
4492	// Stop-machine done, so allow nohz_full to disable tick.
4493	tick_dep_clear(TICK_DEP_BIT_RCU);
4494	return 0;
4495}
4496
4497/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4498 * Mark the specified CPU as being online so that subsequent grace periods
4499 * (both expedited and normal) will wait on it.  Note that this means that
4500 * incoming CPUs are not allowed to use RCU read-side critical sections
4501 * until this function is called.  Failing to observe this restriction
4502 * will result in lockdep splats.
4503 *
4504 * Note that this function is special in that it is invoked directly
4505 * from the incoming CPU rather than from the cpuhp_step mechanism.
4506 * This is because this function must be invoked at a precise location.
4507 * This incoming CPU must not have enabled interrupts yet.
4508 *
4509 * This mirrors the effects of rcutree_report_cpu_dead().
4510 */
4511void rcutree_report_cpu_starting(unsigned int cpu)
4512{
 
4513	unsigned long mask;
4514	struct rcu_data *rdp;
4515	struct rcu_node *rnp;
4516	bool newcpu;
4517
4518	lockdep_assert_irqs_disabled();
4519	rdp = per_cpu_ptr(&rcu_data, cpu);
4520	if (rdp->cpu_started)
4521		return;
4522	rdp->cpu_started = true;
4523
4524	rnp = rdp->mynode;
4525	mask = rdp->grpmask;
4526	arch_spin_lock(&rcu_state.ofl_lock);
4527	rcu_dynticks_eqs_online();
4528	raw_spin_lock(&rcu_state.barrier_lock);
4529	raw_spin_lock_rcu_node(rnp);
4530	WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask);
4531	raw_spin_unlock(&rcu_state.barrier_lock);
4532	newcpu = !(rnp->expmaskinitnext & mask);
4533	rnp->expmaskinitnext |= mask;
4534	/* Allow lockless access for expedited grace periods. */
4535	smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + newcpu); /* ^^^ */
4536	ASSERT_EXCLUSIVE_WRITER(rcu_state.ncpus);
4537	rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
4538	rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4539	rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4540
4541	/* An incoming CPU should never be blocking a grace period. */
4542	if (WARN_ON_ONCE(rnp->qsmask & mask)) { /* RCU waiting on incoming CPU? */
4543		/* rcu_report_qs_rnp() *really* wants some flags to restore */
4544		unsigned long flags;
4545
4546		local_irq_save(flags);
4547		rcu_disable_urgency_upon_qs(rdp);
4548		/* Report QS -after- changing ->qsmaskinitnext! */
4549		rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4550	} else {
4551		raw_spin_unlock_rcu_node(rnp);
4552	}
4553	arch_spin_unlock(&rcu_state.ofl_lock);
4554	smp_store_release(&rdp->beenonline, true);
 
4555	smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
4556}
4557
4558/*
4559 * The outgoing function has no further need of RCU, so remove it from
4560 * the rcu_node tree's ->qsmaskinitnext bit masks.
4561 *
4562 * Note that this function is special in that it is invoked directly
4563 * from the outgoing CPU rather than from the cpuhp_step mechanism.
4564 * This is because this function must be invoked at a precise location.
4565 *
4566 * This mirrors the effect of rcutree_report_cpu_starting().
4567 */
4568void rcutree_report_cpu_dead(void)
4569{
4570	unsigned long flags;
4571	unsigned long mask;
4572	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
4573	struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
4574
4575	/*
4576	 * IRQS must be disabled from now on and until the CPU dies, or an interrupt
4577	 * may introduce a new READ-side while it is actually off the QS masks.
4578	 */
4579	lockdep_assert_irqs_disabled();
4580	// Do any dangling deferred wakeups.
4581	do_nocb_deferred_wakeup(rdp);
4582
 
 
 
 
4583	rcu_preempt_deferred_qs(current);
4584
4585	/* Remove outgoing CPU from mask in the leaf rcu_node structure. */
4586	mask = rdp->grpmask;
4587	arch_spin_lock(&rcu_state.ofl_lock);
 
 
 
4588	raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
4589	rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4590	rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4591	if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
4592		/* Report quiescent state -before- changing ->qsmaskinitnext! */
4593		rcu_disable_urgency_upon_qs(rdp);
4594		rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4595		raw_spin_lock_irqsave_rcu_node(rnp, flags);
4596	}
4597	WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask);
4598	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4599	arch_spin_unlock(&rcu_state.ofl_lock);
 
 
 
 
4600	rdp->cpu_started = false;
4601}
4602
4603#ifdef CONFIG_HOTPLUG_CPU
4604/*
4605 * The outgoing CPU has just passed through the dying-idle state, and we
4606 * are being invoked from the CPU that was IPIed to continue the offline
4607 * operation.  Migrate the outgoing CPU's callbacks to the current CPU.
4608 */
4609void rcutree_migrate_callbacks(int cpu)
4610{
4611	unsigned long flags;
4612	struct rcu_data *my_rdp;
4613	struct rcu_node *my_rnp;
4614	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4615	bool needwake;
4616
4617	if (rcu_rdp_is_offloaded(rdp) ||
4618	    rcu_segcblist_empty(&rdp->cblist))
4619		return;  /* No callbacks to migrate. */
4620
4621	raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
4622	WARN_ON_ONCE(rcu_rdp_cpu_online(rdp));
4623	rcu_barrier_entrain(rdp);
4624	my_rdp = this_cpu_ptr(&rcu_data);
4625	my_rnp = my_rdp->mynode;
4626	rcu_nocb_lock(my_rdp); /* irqs already disabled. */
4627	WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies, false));
4628	raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
4629	/* Leverage recent GPs and set GP for new callbacks. */
4630	needwake = rcu_advance_cbs(my_rnp, rdp) ||
4631		   rcu_advance_cbs(my_rnp, my_rdp);
4632	rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
4633	raw_spin_unlock(&rcu_state.barrier_lock); /* irqs remain disabled. */
4634	needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
4635	rcu_segcblist_disable(&rdp->cblist);
4636	WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) != !rcu_segcblist_n_cbs(&my_rdp->cblist));
4637	check_cb_ovld_locked(my_rdp, my_rnp);
4638	if (rcu_rdp_is_offloaded(my_rdp)) {
4639		raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
4640		__call_rcu_nocb_wake(my_rdp, true, flags);
4641	} else {
4642		rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
4643		raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags);
4644	}
4645	if (needwake)
4646		rcu_gp_kthread_wake();
4647	lockdep_assert_irqs_enabled();
4648	WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
4649		  !rcu_segcblist_empty(&rdp->cblist),
4650		  "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
4651		  cpu, rcu_segcblist_n_cbs(&rdp->cblist),
4652		  rcu_segcblist_first_cb(&rdp->cblist));
4653}
4654
4655/*
4656 * The CPU has been completely removed, and some other CPU is reporting
4657 * this fact from process context.  Do the remainder of the cleanup.
4658 * There can only be one CPU hotplug operation at a time, so no need for
4659 * explicit locking.
4660 */
4661int rcutree_dead_cpu(unsigned int cpu)
4662{
4663	WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus - 1);
4664	// Stop-machine done, so allow nohz_full to disable tick.
4665	tick_dep_clear(TICK_DEP_BIT_RCU);
4666	return 0;
4667}
4668
4669/*
4670 * Near the end of the offline process.  Trace the fact that this CPU
4671 * is going offline.
4672 */
4673int rcutree_dying_cpu(unsigned int cpu)
4674{
4675	bool blkd;
4676	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4677	struct rcu_node *rnp = rdp->mynode;
4678
4679	blkd = !!(READ_ONCE(rnp->qsmask) & rdp->grpmask);
4680	trace_rcu_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
4681			       blkd ? TPS("cpuofl-bgp") : TPS("cpuofl"));
4682	return 0;
4683}
4684
4685/*
4686 * Near the beginning of the process.  The CPU is still very much alive
4687 * with pretty much all services enabled.
4688 */
4689int rcutree_offline_cpu(unsigned int cpu)
4690{
4691	unsigned long flags;
4692	struct rcu_data *rdp;
4693	struct rcu_node *rnp;
4694
4695	rdp = per_cpu_ptr(&rcu_data, cpu);
4696	rnp = rdp->mynode;
4697	raw_spin_lock_irqsave_rcu_node(rnp, flags);
4698	rnp->ffmask &= ~rdp->grpmask;
4699	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4700
4701	rcutree_affinity_setting(cpu, cpu);
4702
4703	// nohz_full CPUs need the tick for stop-machine to work quickly
4704	tick_dep_set(TICK_DEP_BIT_RCU);
4705	return 0;
4706}
4707#endif /* #ifdef CONFIG_HOTPLUG_CPU */
4708
4709/*
4710 * On non-huge systems, use expedited RCU grace periods to make suspend
4711 * and hibernation run faster.
4712 */
4713static int rcu_pm_notify(struct notifier_block *self,
4714			 unsigned long action, void *hcpu)
4715{
4716	switch (action) {
4717	case PM_HIBERNATION_PREPARE:
4718	case PM_SUSPEND_PREPARE:
4719		rcu_async_hurry();
4720		rcu_expedite_gp();
4721		break;
4722	case PM_POST_HIBERNATION:
4723	case PM_POST_SUSPEND:
4724		rcu_unexpedite_gp();
4725		rcu_async_relax();
4726		break;
4727	default:
4728		break;
4729	}
4730	return NOTIFY_OK;
4731}
4732
4733#ifdef CONFIG_RCU_EXP_KTHREAD
4734struct kthread_worker *rcu_exp_gp_kworker;
4735struct kthread_worker *rcu_exp_par_gp_kworker;
4736
4737static void __init rcu_start_exp_gp_kworkers(void)
4738{
4739	const char *par_gp_kworker_name = "rcu_exp_par_gp_kthread_worker";
4740	const char *gp_kworker_name = "rcu_exp_gp_kthread_worker";
4741	struct sched_param param = { .sched_priority = kthread_prio };
4742
4743	rcu_exp_gp_kworker = kthread_create_worker(0, gp_kworker_name);
4744	if (IS_ERR_OR_NULL(rcu_exp_gp_kworker)) {
4745		pr_err("Failed to create %s!\n", gp_kworker_name);
4746		return;
4747	}
4748
4749	rcu_exp_par_gp_kworker = kthread_create_worker(0, par_gp_kworker_name);
4750	if (IS_ERR_OR_NULL(rcu_exp_par_gp_kworker)) {
4751		pr_err("Failed to create %s!\n", par_gp_kworker_name);
4752		kthread_destroy_worker(rcu_exp_gp_kworker);
4753		return;
4754	}
4755
4756	sched_setscheduler_nocheck(rcu_exp_gp_kworker->task, SCHED_FIFO, &param);
4757	sched_setscheduler_nocheck(rcu_exp_par_gp_kworker->task, SCHED_FIFO,
4758				   &param);
4759}
4760
4761static inline void rcu_alloc_par_gp_wq(void)
4762{
4763}
4764#else /* !CONFIG_RCU_EXP_KTHREAD */
4765struct workqueue_struct *rcu_par_gp_wq;
4766
4767static void __init rcu_start_exp_gp_kworkers(void)
4768{
4769}
4770
4771static inline void rcu_alloc_par_gp_wq(void)
4772{
4773	rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
4774	WARN_ON(!rcu_par_gp_wq);
4775}
4776#endif /* CONFIG_RCU_EXP_KTHREAD */
4777
4778/*
4779 * Spawn the kthreads that handle RCU's grace periods.
4780 */
4781static int __init rcu_spawn_gp_kthread(void)
4782{
4783	unsigned long flags;
 
4784	struct rcu_node *rnp;
4785	struct sched_param sp;
4786	struct task_struct *t;
4787	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4788
4789	rcu_scheduler_fully_active = 1;
4790	t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
4791	if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
4792		return 0;
4793	if (kthread_prio) {
4794		sp.sched_priority = kthread_prio;
4795		sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4796	}
4797	rnp = rcu_get_root();
4798	raw_spin_lock_irqsave_rcu_node(rnp, flags);
4799	WRITE_ONCE(rcu_state.gp_activity, jiffies);
4800	WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
4801	// Reset .gp_activity and .gp_req_activity before setting .gp_kthread.
4802	smp_store_release(&rcu_state.gp_kthread, t);  /* ^^^ */
4803	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4804	wake_up_process(t);
4805	/* This is a pre-SMP initcall, we expect a single CPU */
4806	WARN_ON(num_online_cpus() > 1);
4807	/*
4808	 * Those kthreads couldn't be created on rcu_init() -> rcutree_prepare_cpu()
4809	 * due to rcu_scheduler_fully_active.
4810	 */
4811	rcu_spawn_cpu_nocb_kthread(smp_processor_id());
4812	rcu_spawn_one_boost_kthread(rdp->mynode);
4813	rcu_spawn_core_kthreads();
4814	/* Create kthread worker for expedited GPs */
4815	rcu_start_exp_gp_kworkers();
4816	return 0;
4817}
4818early_initcall(rcu_spawn_gp_kthread);
4819
4820/*
4821 * This function is invoked towards the end of the scheduler's
4822 * initialization process.  Before this is called, the idle task might
4823 * contain synchronous grace-period primitives (during which time, this idle
4824 * task is booting the system, and such primitives are no-ops).  After this
4825 * function is called, any synchronous grace-period primitives are run as
4826 * expedited, with the requesting task driving the grace period forward.
4827 * A later core_initcall() rcu_set_runtime_mode() will switch to full
4828 * runtime RCU functionality.
4829 */
4830void rcu_scheduler_starting(void)
4831{
4832	unsigned long flags;
4833	struct rcu_node *rnp;
4834
4835	WARN_ON(num_online_cpus() != 1);
4836	WARN_ON(nr_context_switches() > 0);
4837	rcu_test_sync_prims();
4838
4839	// Fix up the ->gp_seq counters.
4840	local_irq_save(flags);
4841	rcu_for_each_node_breadth_first(rnp)
4842		rnp->gp_seq_needed = rnp->gp_seq = rcu_state.gp_seq;
4843	local_irq_restore(flags);
4844
4845	// Switch out of early boot mode.
4846	rcu_scheduler_active = RCU_SCHEDULER_INIT;
4847	rcu_test_sync_prims();
4848}
4849
4850/*
4851 * Helper function for rcu_init() that initializes the rcu_state structure.
4852 */
4853static void __init rcu_init_one(void)
4854{
4855	static const char * const buf[] = RCU_NODE_NAME_INIT;
4856	static const char * const fqs[] = RCU_FQS_NAME_INIT;
4857	static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4858	static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4859
4860	int levelspread[RCU_NUM_LVLS];		/* kids/node in each level. */
4861	int cpustride = 1;
4862	int i;
4863	int j;
4864	struct rcu_node *rnp;
4865
4866	BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf));  /* Fix buf[] init! */
4867
4868	/* Silence gcc 4.8 false positive about array index out of range. */
4869	if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4870		panic("rcu_init_one: rcu_num_lvls out of range");
4871
4872	/* Initialize the level-tracking arrays. */
4873
4874	for (i = 1; i < rcu_num_lvls; i++)
4875		rcu_state.level[i] =
4876			rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
4877	rcu_init_levelspread(levelspread, num_rcu_lvl);
4878
4879	/* Initialize the elements themselves, starting from the leaves. */
4880
4881	for (i = rcu_num_lvls - 1; i >= 0; i--) {
4882		cpustride *= levelspread[i];
4883		rnp = rcu_state.level[i];
4884		for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
4885			raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4886			lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4887						   &rcu_node_class[i], buf[i]);
4888			raw_spin_lock_init(&rnp->fqslock);
4889			lockdep_set_class_and_name(&rnp->fqslock,
4890						   &rcu_fqs_class[i], fqs[i]);
4891			rnp->gp_seq = rcu_state.gp_seq;
4892			rnp->gp_seq_needed = rcu_state.gp_seq;
4893			rnp->completedqs = rcu_state.gp_seq;
4894			rnp->qsmask = 0;
4895			rnp->qsmaskinit = 0;
4896			rnp->grplo = j * cpustride;
4897			rnp->grphi = (j + 1) * cpustride - 1;
4898			if (rnp->grphi >= nr_cpu_ids)
4899				rnp->grphi = nr_cpu_ids - 1;
4900			if (i == 0) {
4901				rnp->grpnum = 0;
4902				rnp->grpmask = 0;
4903				rnp->parent = NULL;
4904			} else {
4905				rnp->grpnum = j % levelspread[i - 1];
4906				rnp->grpmask = BIT(rnp->grpnum);
4907				rnp->parent = rcu_state.level[i - 1] +
4908					      j / levelspread[i - 1];
4909			}
4910			rnp->level = i;
4911			INIT_LIST_HEAD(&rnp->blkd_tasks);
4912			rcu_init_one_nocb(rnp);
4913			init_waitqueue_head(&rnp->exp_wq[0]);
4914			init_waitqueue_head(&rnp->exp_wq[1]);
4915			init_waitqueue_head(&rnp->exp_wq[2]);
4916			init_waitqueue_head(&rnp->exp_wq[3]);
4917			spin_lock_init(&rnp->exp_lock);
4918			mutex_init(&rnp->boost_kthread_mutex);
4919			raw_spin_lock_init(&rnp->exp_poll_lock);
4920			rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED;
4921			INIT_WORK(&rnp->exp_poll_wq, sync_rcu_do_polled_gp);
4922		}
4923	}
4924
4925	init_swait_queue_head(&rcu_state.gp_wq);
4926	init_swait_queue_head(&rcu_state.expedited_wq);
4927	rnp = rcu_first_leaf_node();
4928	for_each_possible_cpu(i) {
4929		while (i > rnp->grphi)
4930			rnp++;
4931		per_cpu_ptr(&rcu_data, i)->mynode = rnp;
4932		rcu_boot_init_percpu_data(i);
4933	}
4934}
4935
4936/*
4937 * Force priority from the kernel command-line into range.
4938 */
4939static void __init sanitize_kthread_prio(void)
4940{
4941	int kthread_prio_in = kthread_prio;
4942
4943	if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
4944	    && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
4945		kthread_prio = 2;
4946	else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4947		kthread_prio = 1;
4948	else if (kthread_prio < 0)
4949		kthread_prio = 0;
4950	else if (kthread_prio > 99)
4951		kthread_prio = 99;
4952
4953	if (kthread_prio != kthread_prio_in)
4954		pr_alert("%s: Limited prio to %d from %d\n",
4955			 __func__, kthread_prio, kthread_prio_in);
4956}
4957
4958/*
4959 * Compute the rcu_node tree geometry from kernel parameters.  This cannot
4960 * replace the definitions in tree.h because those are needed to size
4961 * the ->node array in the rcu_state structure.
4962 */
4963void rcu_init_geometry(void)
4964{
4965	ulong d;
4966	int i;
4967	static unsigned long old_nr_cpu_ids;
4968	int rcu_capacity[RCU_NUM_LVLS];
4969	static bool initialized;
4970
4971	if (initialized) {
4972		/*
4973		 * Warn if setup_nr_cpu_ids() had not yet been invoked,
4974		 * unless nr_cpus_ids == NR_CPUS, in which case who cares?
4975		 */
4976		WARN_ON_ONCE(old_nr_cpu_ids != nr_cpu_ids);
4977		return;
4978	}
4979
4980	old_nr_cpu_ids = nr_cpu_ids;
4981	initialized = true;
4982
4983	/*
4984	 * Initialize any unspecified boot parameters.
4985	 * The default values of jiffies_till_first_fqs and
4986	 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4987	 * value, which is a function of HZ, then adding one for each
4988	 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4989	 */
4990	d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4991	if (jiffies_till_first_fqs == ULONG_MAX)
4992		jiffies_till_first_fqs = d;
4993	if (jiffies_till_next_fqs == ULONG_MAX)
4994		jiffies_till_next_fqs = d;
4995	adjust_jiffies_till_sched_qs();
4996
4997	/* If the compile-time values are accurate, just leave. */
4998	if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4999	    nr_cpu_ids == NR_CPUS)
5000		return;
5001	pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
5002		rcu_fanout_leaf, nr_cpu_ids);
5003
5004	/*
5005	 * The boot-time rcu_fanout_leaf parameter must be at least two
5006	 * and cannot exceed the number of bits in the rcu_node masks.
5007	 * Complain and fall back to the compile-time values if this
5008	 * limit is exceeded.
5009	 */
5010	if (rcu_fanout_leaf < 2 ||
5011	    rcu_fanout_leaf > sizeof(unsigned long) * 8) {
5012		rcu_fanout_leaf = RCU_FANOUT_LEAF;
5013		WARN_ON(1);
5014		return;
5015	}
5016
5017	/*
5018	 * Compute number of nodes that can be handled an rcu_node tree
5019	 * with the given number of levels.
5020	 */
5021	rcu_capacity[0] = rcu_fanout_leaf;
5022	for (i = 1; i < RCU_NUM_LVLS; i++)
5023		rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
5024
5025	/*
5026	 * The tree must be able to accommodate the configured number of CPUs.
5027	 * If this limit is exceeded, fall back to the compile-time values.
5028	 */
5029	if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
5030		rcu_fanout_leaf = RCU_FANOUT_LEAF;
5031		WARN_ON(1);
5032		return;
5033	}
5034
5035	/* Calculate the number of levels in the tree. */
5036	for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
5037	}
5038	rcu_num_lvls = i + 1;
5039
5040	/* Calculate the number of rcu_nodes at each level of the tree. */
5041	for (i = 0; i < rcu_num_lvls; i++) {
5042		int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
5043		num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
5044	}
5045
5046	/* Calculate the total number of rcu_node structures. */
5047	rcu_num_nodes = 0;
5048	for (i = 0; i < rcu_num_lvls; i++)
5049		rcu_num_nodes += num_rcu_lvl[i];
5050}
5051
5052/*
5053 * Dump out the structure of the rcu_node combining tree associated
5054 * with the rcu_state structure.
5055 */
5056static void __init rcu_dump_rcu_node_tree(void)
5057{
5058	int level = 0;
5059	struct rcu_node *rnp;
5060
5061	pr_info("rcu_node tree layout dump\n");
5062	pr_info(" ");
5063	rcu_for_each_node_breadth_first(rnp) {
5064		if (rnp->level != level) {
5065			pr_cont("\n");
5066			pr_info(" ");
5067			level = rnp->level;
5068		}
5069		pr_cont("%d:%d ^%d  ", rnp->grplo, rnp->grphi, rnp->grpnum);
5070	}
5071	pr_cont("\n");
5072}
5073
5074struct workqueue_struct *rcu_gp_wq;
 
5075
5076static void __init kfree_rcu_batch_init(void)
5077{
5078	int cpu;
5079	int i, j;
5080	struct shrinker *kfree_rcu_shrinker;
5081
5082	/* Clamp it to [0:100] seconds interval. */
5083	if (rcu_delay_page_cache_fill_msec < 0 ||
5084		rcu_delay_page_cache_fill_msec > 100 * MSEC_PER_SEC) {
5085
5086		rcu_delay_page_cache_fill_msec =
5087			clamp(rcu_delay_page_cache_fill_msec, 0,
5088				(int) (100 * MSEC_PER_SEC));
5089
5090		pr_info("Adjusting rcutree.rcu_delay_page_cache_fill_msec to %d ms.\n",
5091			rcu_delay_page_cache_fill_msec);
5092	}
5093
5094	for_each_possible_cpu(cpu) {
5095		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
5096
5097		for (i = 0; i < KFREE_N_BATCHES; i++) {
5098			INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
5099			krcp->krw_arr[i].krcp = krcp;
5100
5101			for (j = 0; j < FREE_N_CHANNELS; j++)
5102				INIT_LIST_HEAD(&krcp->krw_arr[i].bulk_head_free[j]);
5103		}
5104
5105		for (i = 0; i < FREE_N_CHANNELS; i++)
5106			INIT_LIST_HEAD(&krcp->bulk_head[i]);
5107
5108		INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
5109		INIT_DELAYED_WORK(&krcp->page_cache_work, fill_page_cache_func);
5110		krcp->initialized = true;
5111	}
5112
5113	kfree_rcu_shrinker = shrinker_alloc(0, "rcu-kfree");
5114	if (!kfree_rcu_shrinker) {
5115		pr_err("Failed to allocate kfree_rcu() shrinker!\n");
5116		return;
5117	}
5118
5119	kfree_rcu_shrinker->count_objects = kfree_rcu_shrink_count;
5120	kfree_rcu_shrinker->scan_objects = kfree_rcu_shrink_scan;
5121
5122	shrinker_register(kfree_rcu_shrinker);
5123}
5124
5125void __init rcu_init(void)
5126{
5127	int cpu = smp_processor_id();
5128
5129	rcu_early_boot_tests();
5130
5131	kfree_rcu_batch_init();
5132	rcu_bootup_announce();
5133	sanitize_kthread_prio();
5134	rcu_init_geometry();
5135	rcu_init_one();
5136	if (dump_tree)
5137		rcu_dump_rcu_node_tree();
5138	if (use_softirq)
5139		open_softirq(RCU_SOFTIRQ, rcu_core_si);
5140
5141	/*
5142	 * We don't need protection against CPU-hotplug here because
5143	 * this is called early in boot, before either interrupts
5144	 * or the scheduler are operational.
5145	 */
5146	pm_notifier(rcu_pm_notify, 0);
5147	WARN_ON(num_online_cpus() > 1); // Only one CPU this early in boot.
5148	rcutree_prepare_cpu(cpu);
5149	rcutree_report_cpu_starting(cpu);
5150	rcutree_online_cpu(cpu);
 
5151
5152	/* Create workqueue for Tree SRCU and for expedited GPs. */
5153	rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
5154	WARN_ON(!rcu_gp_wq);
5155	rcu_alloc_par_gp_wq();
 
5156
5157	/* Fill in default value for rcutree.qovld boot parameter. */
5158	/* -After- the rcu_node ->lock fields are initialized! */
5159	if (qovld < 0)
5160		qovld_calc = DEFAULT_RCU_QOVLD_MULT * qhimark;
5161	else
5162		qovld_calc = qovld;
5163
5164	// Kick-start in case any polled grace periods started early.
5165	(void)start_poll_synchronize_rcu_expedited();
5166
5167	rcu_test_sync_prims();
5168}
5169
5170#include "tree_stall.h"
5171#include "tree_exp.h"
5172#include "tree_nocb.h"
5173#include "tree_plugin.h"
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0+
   2/*
   3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
   4 *
   5 * Copyright IBM Corporation, 2008
   6 *
   7 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
   8 *	    Manfred Spraul <manfred@colorfullife.com>
   9 *	    Paul E. McKenney <paulmck@linux.ibm.com>
  10 *
  11 * Based on the original work by Paul McKenney <paulmck@linux.ibm.com>
  12 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
  13 *
  14 * For detailed explanation of Read-Copy Update mechanism see -
  15 *	Documentation/RCU
  16 */
  17
  18#define pr_fmt(fmt) "rcu: " fmt
  19
  20#include <linux/types.h>
  21#include <linux/kernel.h>
  22#include <linux/init.h>
  23#include <linux/spinlock.h>
  24#include <linux/smp.h>
  25#include <linux/rcupdate_wait.h>
  26#include <linux/interrupt.h>
  27#include <linux/sched.h>
  28#include <linux/sched/debug.h>
  29#include <linux/nmi.h>
  30#include <linux/atomic.h>
  31#include <linux/bitops.h>
  32#include <linux/export.h>
  33#include <linux/completion.h>
 
  34#include <linux/moduleparam.h>
  35#include <linux/panic.h>
  36#include <linux/panic_notifier.h>
  37#include <linux/percpu.h>
  38#include <linux/notifier.h>
  39#include <linux/cpu.h>
  40#include <linux/mutex.h>
  41#include <linux/time.h>
  42#include <linux/kernel_stat.h>
  43#include <linux/wait.h>
  44#include <linux/kthread.h>
  45#include <uapi/linux/sched/types.h>
  46#include <linux/prefetch.h>
  47#include <linux/delay.h>
  48#include <linux/random.h>
  49#include <linux/trace_events.h>
  50#include <linux/suspend.h>
  51#include <linux/ftrace.h>
  52#include <linux/tick.h>
  53#include <linux/sysrq.h>
  54#include <linux/kprobes.h>
  55#include <linux/gfp.h>
  56#include <linux/oom.h>
  57#include <linux/smpboot.h>
  58#include <linux/jiffies.h>
  59#include <linux/slab.h>
  60#include <linux/sched/isolation.h>
  61#include <linux/sched/clock.h>
  62#include <linux/vmalloc.h>
  63#include <linux/mm.h>
  64#include <linux/kasan.h>
 
  65#include "../time/tick-internal.h"
  66
  67#include "tree.h"
  68#include "rcu.h"
  69
  70#ifdef MODULE_PARAM_PREFIX
  71#undef MODULE_PARAM_PREFIX
  72#endif
  73#define MODULE_PARAM_PREFIX "rcutree."
  74
  75/* Data structures. */
  76
  77/*
  78 * Steal a bit from the bottom of ->dynticks for idle entry/exit
  79 * control.  Initially this is for TLB flushing.
  80 */
  81#define RCU_DYNTICK_CTRL_MASK 0x1
  82#define RCU_DYNTICK_CTRL_CTR  (RCU_DYNTICK_CTRL_MASK + 1)
  83
  84static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = {
  85	.dynticks_nesting = 1,
  86	.dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE,
  87	.dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
  88#ifdef CONFIG_RCU_NOCB_CPU
  89	.cblist.flags = SEGCBLIST_SOFTIRQ_ONLY,
  90#endif
  91};
  92static struct rcu_state rcu_state = {
  93	.level = { &rcu_state.node[0] },
  94	.gp_state = RCU_GP_IDLE,
  95	.gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT,
  96	.barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex),
 
  97	.name = RCU_NAME,
  98	.abbr = RCU_ABBR,
  99	.exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
 100	.exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
 101	.ofl_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.ofl_lock),
 102};
 103
 104/* Dump rcu_node combining tree at boot to verify correct setup. */
 105static bool dump_tree;
 106module_param(dump_tree, bool, 0444);
 107/* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */
 108static bool use_softirq = !IS_ENABLED(CONFIG_PREEMPT_RT);
 109#ifndef CONFIG_PREEMPT_RT
 110module_param(use_softirq, bool, 0444);
 111#endif
 112/* Control rcu_node-tree auto-balancing at boot time. */
 113static bool rcu_fanout_exact;
 114module_param(rcu_fanout_exact, bool, 0444);
 115/* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
 116static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
 117module_param(rcu_fanout_leaf, int, 0444);
 118int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
 119/* Number of rcu_nodes at specified level. */
 120int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
 121int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
 122
 123/*
 124 * The rcu_scheduler_active variable is initialized to the value
 125 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
 126 * first task is spawned.  So when this variable is RCU_SCHEDULER_INACTIVE,
 127 * RCU can assume that there is but one task, allowing RCU to (for example)
 128 * optimize synchronize_rcu() to a simple barrier().  When this variable
 129 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
 130 * to detect real grace periods.  This variable is also used to suppress
 131 * boot-time false positives from lockdep-RCU error checking.  Finally, it
 132 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
 133 * is fully initialized, including all of its kthreads having been spawned.
 134 */
 135int rcu_scheduler_active __read_mostly;
 136EXPORT_SYMBOL_GPL(rcu_scheduler_active);
 137
 138/*
 139 * The rcu_scheduler_fully_active variable transitions from zero to one
 140 * during the early_initcall() processing, which is after the scheduler
 141 * is capable of creating new tasks.  So RCU processing (for example,
 142 * creating tasks for RCU priority boosting) must be delayed until after
 143 * rcu_scheduler_fully_active transitions from zero to one.  We also
 144 * currently delay invocation of any RCU callbacks until after this point.
 145 *
 146 * It might later prove better for people registering RCU callbacks during
 147 * early boot to take responsibility for these callbacks, but one step at
 148 * a time.
 149 */
 150static int rcu_scheduler_fully_active __read_mostly;
 151
 152static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
 153			      unsigned long gps, unsigned long flags);
 154static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
 155static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
 156static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
 157static void invoke_rcu_core(void);
 158static void rcu_report_exp_rdp(struct rcu_data *rdp);
 159static void sync_sched_exp_online_cleanup(int cpu);
 160static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp);
 161static bool rcu_rdp_is_offloaded(struct rcu_data *rdp);
 
 
 
 
 162
 163/* rcuc/rcub kthread realtime priority */
 
 
 
 
 164static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
 165module_param(kthread_prio, int, 0444);
 166
 167/* Delay in jiffies for grace-period initialization delays, debug only. */
 168
 169static int gp_preinit_delay;
 170module_param(gp_preinit_delay, int, 0444);
 171static int gp_init_delay;
 172module_param(gp_init_delay, int, 0444);
 173static int gp_cleanup_delay;
 174module_param(gp_cleanup_delay, int, 0444);
 175
 176// Add delay to rcu_read_unlock() for strict grace periods.
 177static int rcu_unlock_delay;
 178#ifdef CONFIG_RCU_STRICT_GRACE_PERIOD
 179module_param(rcu_unlock_delay, int, 0444);
 180#endif
 181
 182/*
 183 * This rcu parameter is runtime-read-only. It reflects
 184 * a minimum allowed number of objects which can be cached
 185 * per-CPU. Object size is equal to one page. This value
 186 * can be changed at boot time.
 187 */
 188static int rcu_min_cached_objs = 5;
 189module_param(rcu_min_cached_objs, int, 0444);
 190
 191// A page shrinker can ask for pages to be freed to make them
 192// available for other parts of the system. This usually happens
 193// under low memory conditions, and in that case we should also
 194// defer page-cache filling for a short time period.
 195//
 196// The default value is 5 seconds, which is long enough to reduce
 197// interference with the shrinker while it asks other systems to
 198// drain their caches.
 199static int rcu_delay_page_cache_fill_msec = 5000;
 200module_param(rcu_delay_page_cache_fill_msec, int, 0444);
 201
 202/* Retrieve RCU kthreads priority for rcutorture */
 203int rcu_get_gp_kthreads_prio(void)
 204{
 205	return kthread_prio;
 206}
 207EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio);
 208
 209/*
 210 * Number of grace periods between delays, normalized by the duration of
 211 * the delay.  The longer the delay, the more the grace periods between
 212 * each delay.  The reason for this normalization is that it means that,
 213 * for non-zero delays, the overall slowdown of grace periods is constant
 214 * regardless of the duration of the delay.  This arrangement balances
 215 * the need for long delays to increase some race probabilities with the
 216 * need for fast grace periods to increase other race probabilities.
 217 */
 218#define PER_RCU_NODE_PERIOD 3	/* Number of grace periods between delays for debugging. */
 219
 220/*
 221 * Compute the mask of online CPUs for the specified rcu_node structure.
 222 * This will not be stable unless the rcu_node structure's ->lock is
 223 * held, but the bit corresponding to the current CPU will be stable
 224 * in most contexts.
 225 */
 226static unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
 227{
 228	return READ_ONCE(rnp->qsmaskinitnext);
 229}
 230
 231/*
 232 * Return true if an RCU grace period is in progress.  The READ_ONCE()s
 233 * permit this function to be invoked without holding the root rcu_node
 234 * structure's ->lock, but of course results can be subject to change.
 235 */
 236static int rcu_gp_in_progress(void)
 237{
 238	return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq));
 239}
 240
 241/*
 242 * Return the number of callbacks queued on the specified CPU.
 243 * Handles both the nocbs and normal cases.
 244 */
 245static long rcu_get_n_cbs_cpu(int cpu)
 246{
 247	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 248
 249	if (rcu_segcblist_is_enabled(&rdp->cblist))
 250		return rcu_segcblist_n_cbs(&rdp->cblist);
 251	return 0;
 252}
 253
 254void rcu_softirq_qs(void)
 255{
 256	rcu_qs();
 257	rcu_preempt_deferred_qs(current);
 258	rcu_tasks_qs(current, false);
 259}
 260
 261/*
 262 * Record entry into an extended quiescent state.  This is only to be
 263 * called when not already in an extended quiescent state, that is,
 264 * RCU is watching prior to the call to this function and is no longer
 265 * watching upon return.
 266 */
 267static noinstr void rcu_dynticks_eqs_enter(void)
 268{
 269	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 270	int seq;
 271
 272	/*
 273	 * CPUs seeing atomic_add_return() must see prior RCU read-side
 274	 * critical sections, and we also must force ordering with the
 275	 * next idle sojourn.
 276	 */
 277	rcu_dynticks_task_trace_enter();  // Before ->dynticks update!
 278	seq = arch_atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
 279	// RCU is no longer watching.  Better be in extended quiescent state!
 280	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
 281		     (seq & RCU_DYNTICK_CTRL_CTR));
 282	/* Better not have special action (TLB flush) pending! */
 283	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
 284		     (seq & RCU_DYNTICK_CTRL_MASK));
 285}
 286
 287/*
 288 * Record exit from an extended quiescent state.  This is only to be
 289 * called from an extended quiescent state, that is, RCU is not watching
 290 * prior to the call to this function and is watching upon return.
 291 */
 292static noinstr void rcu_dynticks_eqs_exit(void)
 293{
 294	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 295	int seq;
 296
 297	/*
 298	 * CPUs seeing atomic_add_return() must see prior idle sojourns,
 299	 * and we also must force ordering with the next RCU read-side
 300	 * critical section.
 301	 */
 302	seq = arch_atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
 303	// RCU is now watching.  Better not be in an extended quiescent state!
 304	rcu_dynticks_task_trace_exit();  // After ->dynticks update!
 305	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
 306		     !(seq & RCU_DYNTICK_CTRL_CTR));
 307	if (seq & RCU_DYNTICK_CTRL_MASK) {
 308		arch_atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdp->dynticks);
 309		smp_mb__after_atomic(); /* _exit after clearing mask. */
 310	}
 311}
 312
 313/*
 314 * Reset the current CPU's ->dynticks counter to indicate that the
 315 * newly onlined CPU is no longer in an extended quiescent state.
 316 * This will either leave the counter unchanged, or increment it
 317 * to the next non-quiescent value.
 318 *
 319 * The non-atomic test/increment sequence works because the upper bits
 320 * of the ->dynticks counter are manipulated only by the corresponding CPU,
 321 * or when the corresponding CPU is offline.
 322 */
 323static void rcu_dynticks_eqs_online(void)
 324{
 325	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 326
 327	if (atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR)
 328		return;
 329	atomic_add(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
 330}
 331
 332/*
 333 * Is the current CPU in an extended quiescent state?
 334 *
 335 * No ordering, as we are sampling CPU-local information.
 336 */
 337static __always_inline bool rcu_dynticks_curr_cpu_in_eqs(void)
 338{
 339	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 340
 341	return !(arch_atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR);
 342}
 343
 344/*
 345 * Snapshot the ->dynticks counter with full ordering so as to allow
 346 * stable comparison of this counter with past and future snapshots.
 347 */
 348static int rcu_dynticks_snap(struct rcu_data *rdp)
 349{
 350	int snap = atomic_add_return(0, &rdp->dynticks);
 351
 352	return snap & ~RCU_DYNTICK_CTRL_MASK;
 353}
 354
 355/*
 356 * Return true if the snapshot returned from rcu_dynticks_snap()
 357 * indicates that RCU is in an extended quiescent state.
 358 */
 359static bool rcu_dynticks_in_eqs(int snap)
 360{
 361	return !(snap & RCU_DYNTICK_CTRL_CTR);
 362}
 363
 364/* Return true if the specified CPU is currently idle from an RCU viewpoint.  */
 365bool rcu_is_idle_cpu(int cpu)
 366{
 367	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 368
 369	return rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp));
 370}
 371
 372/*
 373 * Return true if the CPU corresponding to the specified rcu_data
 374 * structure has spent some time in an extended quiescent state since
 375 * rcu_dynticks_snap() returned the specified snapshot.
 376 */
 377static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap)
 378{
 379	return snap != rcu_dynticks_snap(rdp);
 380}
 381
 382/*
 383 * Return true if the referenced integer is zero while the specified
 384 * CPU remains within a single extended quiescent state.
 385 */
 386bool rcu_dynticks_zero_in_eqs(int cpu, int *vp)
 387{
 388	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 389	int snap;
 390
 391	// If not quiescent, force back to earlier extended quiescent state.
 392	snap = atomic_read(&rdp->dynticks) & ~(RCU_DYNTICK_CTRL_MASK |
 393					       RCU_DYNTICK_CTRL_CTR);
 394
 395	smp_rmb(); // Order ->dynticks and *vp reads.
 396	if (READ_ONCE(*vp))
 397		return false;  // Non-zero, so report failure;
 398	smp_rmb(); // Order *vp read and ->dynticks re-read.
 399
 400	// If still in the same extended quiescent state, we are good!
 401	return snap == (atomic_read(&rdp->dynticks) & ~RCU_DYNTICK_CTRL_MASK);
 402}
 403
 404/*
 405 * Set the special (bottom) bit of the specified CPU so that it
 406 * will take special action (such as flushing its TLB) on the
 407 * next exit from an extended quiescent state.  Returns true if
 408 * the bit was successfully set, or false if the CPU was not in
 409 * an extended quiescent state.
 410 */
 411bool rcu_eqs_special_set(int cpu)
 412{
 413	int old;
 414	int new;
 415	int new_old;
 416	struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
 417
 418	new_old = atomic_read(&rdp->dynticks);
 419	do {
 420		old = new_old;
 421		if (old & RCU_DYNTICK_CTRL_CTR)
 422			return false;
 423		new = old | RCU_DYNTICK_CTRL_MASK;
 424		new_old = atomic_cmpxchg(&rdp->dynticks, old, new);
 425	} while (new_old != old);
 426	return true;
 427}
 428
 429/*
 430 * Let the RCU core know that this CPU has gone through the scheduler,
 431 * which is a quiescent state.  This is called when the need for a
 432 * quiescent state is urgent, so we burn an atomic operation and full
 433 * memory barriers to let the RCU core know about it, regardless of what
 434 * this CPU might (or might not) do in the near future.
 435 *
 436 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
 437 *
 438 * The caller must have disabled interrupts and must not be idle.
 439 */
 440notrace void rcu_momentary_dyntick_idle(void)
 441{
 442	int special;
 443
 444	raw_cpu_write(rcu_data.rcu_need_heavy_qs, false);
 445	special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
 446				    &this_cpu_ptr(&rcu_data)->dynticks);
 447	/* It is illegal to call this from idle state. */
 448	WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
 449	rcu_preempt_deferred_qs(current);
 450}
 451EXPORT_SYMBOL_GPL(rcu_momentary_dyntick_idle);
 452
 453/**
 454 * rcu_is_cpu_rrupt_from_idle - see if 'interrupted' from idle
 455 *
 456 * If the current CPU is idle and running at a first-level (not nested)
 457 * interrupt, or directly, from idle, return true.
 458 *
 459 * The caller must have at least disabled IRQs.
 460 */
 461static int rcu_is_cpu_rrupt_from_idle(void)
 462{
 463	long nesting;
 464
 465	/*
 466	 * Usually called from the tick; but also used from smp_function_call()
 467	 * for expedited grace periods. This latter can result in running from
 468	 * the idle task, instead of an actual IPI.
 469	 */
 470	lockdep_assert_irqs_disabled();
 471
 472	/* Check for counter underflows */
 473	RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) < 0,
 474			 "RCU dynticks_nesting counter underflow!");
 475	RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) <= 0,
 476			 "RCU dynticks_nmi_nesting counter underflow/zero!");
 477
 478	/* Are we at first interrupt nesting level? */
 479	nesting = __this_cpu_read(rcu_data.dynticks_nmi_nesting);
 480	if (nesting > 1)
 481		return false;
 482
 483	/*
 484	 * If we're not in an interrupt, we must be in the idle task!
 485	 */
 486	WARN_ON_ONCE(!nesting && !is_idle_task(current));
 487
 488	/* Does CPU appear to be idle from an RCU standpoint? */
 489	return __this_cpu_read(rcu_data.dynticks_nesting) == 0;
 490}
 491
 492#define DEFAULT_RCU_BLIMIT (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 1000 : 10)
 493				// Maximum callbacks per rcu_do_batch ...
 494#define DEFAULT_MAX_RCU_BLIMIT 10000 // ... even during callback flood.
 495static long blimit = DEFAULT_RCU_BLIMIT;
 496#define DEFAULT_RCU_QHIMARK 10000 // If this many pending, ignore blimit.
 497static long qhimark = DEFAULT_RCU_QHIMARK;
 498#define DEFAULT_RCU_QLOMARK 100   // Once only this many pending, use blimit.
 499static long qlowmark = DEFAULT_RCU_QLOMARK;
 500#define DEFAULT_RCU_QOVLD_MULT 2
 501#define DEFAULT_RCU_QOVLD (DEFAULT_RCU_QOVLD_MULT * DEFAULT_RCU_QHIMARK)
 502static long qovld = DEFAULT_RCU_QOVLD; // If this many pending, hammer QS.
 503static long qovld_calc = -1;	  // No pre-initialization lock acquisitions!
 504
 505module_param(blimit, long, 0444);
 506module_param(qhimark, long, 0444);
 507module_param(qlowmark, long, 0444);
 508module_param(qovld, long, 0444);
 509
 510static ulong jiffies_till_first_fqs = IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 0 : ULONG_MAX;
 511static ulong jiffies_till_next_fqs = ULONG_MAX;
 512static bool rcu_kick_kthreads;
 513static int rcu_divisor = 7;
 514module_param(rcu_divisor, int, 0644);
 515
 516/* Force an exit from rcu_do_batch() after 3 milliseconds. */
 517static long rcu_resched_ns = 3 * NSEC_PER_MSEC;
 518module_param(rcu_resched_ns, long, 0644);
 519
 520/*
 521 * How long the grace period must be before we start recruiting
 522 * quiescent-state help from rcu_note_context_switch().
 523 */
 524static ulong jiffies_till_sched_qs = ULONG_MAX;
 525module_param(jiffies_till_sched_qs, ulong, 0444);
 526static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */
 527module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
 528
 529/*
 530 * Make sure that we give the grace-period kthread time to detect any
 531 * idle CPUs before taking active measures to force quiescent states.
 532 * However, don't go below 100 milliseconds, adjusted upwards for really
 533 * large systems.
 534 */
 535static void adjust_jiffies_till_sched_qs(void)
 536{
 537	unsigned long j;
 538
 539	/* If jiffies_till_sched_qs was specified, respect the request. */
 540	if (jiffies_till_sched_qs != ULONG_MAX) {
 541		WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
 542		return;
 543	}
 544	/* Otherwise, set to third fqs scan, but bound below on large system. */
 545	j = READ_ONCE(jiffies_till_first_fqs) +
 546		      2 * READ_ONCE(jiffies_till_next_fqs);
 547	if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
 548		j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
 549	pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
 550	WRITE_ONCE(jiffies_to_sched_qs, j);
 551}
 552
 553static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
 554{
 555	ulong j;
 556	int ret = kstrtoul(val, 0, &j);
 557
 558	if (!ret) {
 559		WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
 560		adjust_jiffies_till_sched_qs();
 561	}
 562	return ret;
 563}
 564
 565static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
 566{
 567	ulong j;
 568	int ret = kstrtoul(val, 0, &j);
 569
 570	if (!ret) {
 571		WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
 572		adjust_jiffies_till_sched_qs();
 573	}
 574	return ret;
 575}
 576
 577static const struct kernel_param_ops first_fqs_jiffies_ops = {
 578	.set = param_set_first_fqs_jiffies,
 579	.get = param_get_ulong,
 580};
 581
 582static const struct kernel_param_ops next_fqs_jiffies_ops = {
 583	.set = param_set_next_fqs_jiffies,
 584	.get = param_get_ulong,
 585};
 586
 587module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
 588module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
 589module_param(rcu_kick_kthreads, bool, 0644);
 590
 591static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
 592static int rcu_pending(int user);
 593
 594/*
 595 * Return the number of RCU GPs completed thus far for debug & stats.
 596 */
 597unsigned long rcu_get_gp_seq(void)
 598{
 599	return READ_ONCE(rcu_state.gp_seq);
 600}
 601EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
 602
 603/*
 604 * Return the number of RCU expedited batches completed thus far for
 605 * debug & stats.  Odd numbers mean that a batch is in progress, even
 606 * numbers mean idle.  The value returned will thus be roughly double
 607 * the cumulative batches since boot.
 608 */
 609unsigned long rcu_exp_batches_completed(void)
 610{
 611	return rcu_state.expedited_sequence;
 612}
 613EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
 614
 615/*
 616 * Return the root node of the rcu_state structure.
 617 */
 618static struct rcu_node *rcu_get_root(void)
 619{
 620	return &rcu_state.node[0];
 621}
 622
 623/*
 624 * Send along grace-period-related data for rcutorture diagnostics.
 625 */
 626void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
 627			    unsigned long *gp_seq)
 628{
 629	switch (test_type) {
 630	case RCU_FLAVOR:
 631		*flags = READ_ONCE(rcu_state.gp_flags);
 632		*gp_seq = rcu_seq_current(&rcu_state.gp_seq);
 633		break;
 634	default:
 635		break;
 636	}
 637}
 638EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
 639
 640/*
 641 * Enter an RCU extended quiescent state, which can be either the
 642 * idle loop or adaptive-tickless usermode execution.
 643 *
 644 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
 645 * the possibility of usermode upcalls having messed up our count
 646 * of interrupt nesting level during the prior busy period.
 647 */
 648static noinstr void rcu_eqs_enter(bool user)
 649{
 650	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 651
 652	WARN_ON_ONCE(rdp->dynticks_nmi_nesting != DYNTICK_IRQ_NONIDLE);
 653	WRITE_ONCE(rdp->dynticks_nmi_nesting, 0);
 654	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
 655		     rdp->dynticks_nesting == 0);
 656	if (rdp->dynticks_nesting != 1) {
 657		// RCU will still be watching, so just do accounting and leave.
 658		rdp->dynticks_nesting--;
 659		return;
 660	}
 661
 662	lockdep_assert_irqs_disabled();
 663	instrumentation_begin();
 664	trace_rcu_dyntick(TPS("Start"), rdp->dynticks_nesting, 0, atomic_read(&rdp->dynticks));
 665	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
 666	rcu_prepare_for_idle();
 667	rcu_preempt_deferred_qs(current);
 668
 669	// instrumentation for the noinstr rcu_dynticks_eqs_enter()
 670	instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
 671
 672	instrumentation_end();
 673	WRITE_ONCE(rdp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
 674	// RCU is watching here ...
 675	rcu_dynticks_eqs_enter();
 676	// ... but is no longer watching here.
 677	rcu_dynticks_task_enter();
 678}
 679
 680/**
 681 * rcu_idle_enter - inform RCU that current CPU is entering idle
 682 *
 683 * Enter idle mode, in other words, -leave- the mode in which RCU
 684 * read-side critical sections can occur.  (Though RCU read-side
 685 * critical sections can occur in irq handlers in idle, a possibility
 686 * handled by irq_enter() and irq_exit().)
 687 *
 688 * If you add or remove a call to rcu_idle_enter(), be sure to test with
 689 * CONFIG_RCU_EQS_DEBUG=y.
 690 */
 691void rcu_idle_enter(void)
 692{
 693	lockdep_assert_irqs_disabled();
 694	rcu_eqs_enter(false);
 695}
 696EXPORT_SYMBOL_GPL(rcu_idle_enter);
 697
 698#ifdef CONFIG_NO_HZ_FULL
 699
 700#if !defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK)
 701/*
 702 * An empty function that will trigger a reschedule on
 703 * IRQ tail once IRQs get re-enabled on userspace/guest resume.
 704 */
 705static void late_wakeup_func(struct irq_work *work)
 706{
 707}
 708
 709static DEFINE_PER_CPU(struct irq_work, late_wakeup_work) =
 710	IRQ_WORK_INIT(late_wakeup_func);
 711
 712/*
 713 * If either:
 714 *
 715 * 1) the task is about to enter in guest mode and $ARCH doesn't support KVM generic work
 716 * 2) the task is about to enter in user mode and $ARCH doesn't support generic entry.
 717 *
 718 * In these cases the late RCU wake ups aren't supported in the resched loops and our
 719 * last resort is to fire a local irq_work that will trigger a reschedule once IRQs
 720 * get re-enabled again.
 721 */
 722noinstr static void rcu_irq_work_resched(void)
 723{
 724	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 725
 726	if (IS_ENABLED(CONFIG_GENERIC_ENTRY) && !(current->flags & PF_VCPU))
 727		return;
 728
 729	if (IS_ENABLED(CONFIG_KVM_XFER_TO_GUEST_WORK) && (current->flags & PF_VCPU))
 730		return;
 731
 732	instrumentation_begin();
 733	if (do_nocb_deferred_wakeup(rdp) && need_resched()) {
 734		irq_work_queue(this_cpu_ptr(&late_wakeup_work));
 735	}
 736	instrumentation_end();
 737}
 738
 739#else
 740static inline void rcu_irq_work_resched(void) { }
 741#endif
 742
 743/**
 744 * rcu_user_enter - inform RCU that we are resuming userspace.
 745 *
 746 * Enter RCU idle mode right before resuming userspace.  No use of RCU
 747 * is permitted between this call and rcu_user_exit(). This way the
 748 * CPU doesn't need to maintain the tick for RCU maintenance purposes
 749 * when the CPU runs in userspace.
 750 *
 751 * If you add or remove a call to rcu_user_enter(), be sure to test with
 752 * CONFIG_RCU_EQS_DEBUG=y.
 753 */
 754noinstr void rcu_user_enter(void)
 755{
 756	lockdep_assert_irqs_disabled();
 757
 758	/*
 759	 * Other than generic entry implementation, we may be past the last
 760	 * rescheduling opportunity in the entry code. Trigger a self IPI
 761	 * that will fire and reschedule once we resume in user/guest mode.
 762	 */
 763	rcu_irq_work_resched();
 764	rcu_eqs_enter(true);
 765}
 766
 767#endif /* CONFIG_NO_HZ_FULL */
 768
 769/**
 770 * rcu_nmi_exit - inform RCU of exit from NMI context
 771 *
 772 * If we are returning from the outermost NMI handler that interrupted an
 773 * RCU-idle period, update rdp->dynticks and rdp->dynticks_nmi_nesting
 774 * to let the RCU grace-period handling know that the CPU is back to
 775 * being RCU-idle.
 776 *
 777 * If you add or remove a call to rcu_nmi_exit(), be sure to test
 778 * with CONFIG_RCU_EQS_DEBUG=y.
 779 */
 780noinstr void rcu_nmi_exit(void)
 781{
 782	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 783
 784	instrumentation_begin();
 785	/*
 786	 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
 787	 * (We are exiting an NMI handler, so RCU better be paying attention
 788	 * to us!)
 789	 */
 790	WARN_ON_ONCE(rdp->dynticks_nmi_nesting <= 0);
 791	WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
 792
 793	/*
 794	 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
 795	 * leave it in non-RCU-idle state.
 796	 */
 797	if (rdp->dynticks_nmi_nesting != 1) {
 798		trace_rcu_dyntick(TPS("--="), rdp->dynticks_nmi_nesting, rdp->dynticks_nmi_nesting - 2,
 799				  atomic_read(&rdp->dynticks));
 800		WRITE_ONCE(rdp->dynticks_nmi_nesting, /* No store tearing. */
 801			   rdp->dynticks_nmi_nesting - 2);
 802		instrumentation_end();
 803		return;
 804	}
 805
 806	/* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
 807	trace_rcu_dyntick(TPS("Startirq"), rdp->dynticks_nmi_nesting, 0, atomic_read(&rdp->dynticks));
 808	WRITE_ONCE(rdp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
 809
 810	if (!in_nmi())
 811		rcu_prepare_for_idle();
 812
 813	// instrumentation for the noinstr rcu_dynticks_eqs_enter()
 814	instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
 815	instrumentation_end();
 816
 817	// RCU is watching here ...
 818	rcu_dynticks_eqs_enter();
 819	// ... but is no longer watching here.
 820
 821	if (!in_nmi())
 822		rcu_dynticks_task_enter();
 823}
 824
 825/**
 826 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
 827 *
 828 * Exit from an interrupt handler, which might possibly result in entering
 829 * idle mode, in other words, leaving the mode in which read-side critical
 830 * sections can occur.  The caller must have disabled interrupts.
 831 *
 832 * This code assumes that the idle loop never does anything that might
 833 * result in unbalanced calls to irq_enter() and irq_exit().  If your
 834 * architecture's idle loop violates this assumption, RCU will give you what
 835 * you deserve, good and hard.  But very infrequently and irreproducibly.
 836 *
 837 * Use things like work queues to work around this limitation.
 838 *
 839 * You have been warned.
 840 *
 841 * If you add or remove a call to rcu_irq_exit(), be sure to test with
 842 * CONFIG_RCU_EQS_DEBUG=y.
 843 */
 844void noinstr rcu_irq_exit(void)
 845{
 846	lockdep_assert_irqs_disabled();
 847	rcu_nmi_exit();
 848}
 849
 850#ifdef CONFIG_PROVE_RCU
 851/**
 852 * rcu_irq_exit_check_preempt - Validate that scheduling is possible
 853 */
 854void rcu_irq_exit_check_preempt(void)
 855{
 856	lockdep_assert_irqs_disabled();
 857
 858	RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) <= 0,
 859			 "RCU dynticks_nesting counter underflow/zero!");
 860	RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) !=
 861			 DYNTICK_IRQ_NONIDLE,
 862			 "Bad RCU  dynticks_nmi_nesting counter\n");
 863	RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
 864			 "RCU in extended quiescent state!");
 865}
 866#endif /* #ifdef CONFIG_PROVE_RCU */
 867
 868/*
 869 * Wrapper for rcu_irq_exit() where interrupts are enabled.
 870 *
 871 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
 872 * with CONFIG_RCU_EQS_DEBUG=y.
 873 */
 874void rcu_irq_exit_irqson(void)
 875{
 876	unsigned long flags;
 877
 878	local_irq_save(flags);
 879	rcu_irq_exit();
 880	local_irq_restore(flags);
 881}
 882
 883/*
 884 * Exit an RCU extended quiescent state, which can be either the
 885 * idle loop or adaptive-tickless usermode execution.
 886 *
 887 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
 888 * allow for the possibility of usermode upcalls messing up our count of
 889 * interrupt nesting level during the busy period that is just now starting.
 890 */
 891static void noinstr rcu_eqs_exit(bool user)
 892{
 893	struct rcu_data *rdp;
 894	long oldval;
 895
 896	lockdep_assert_irqs_disabled();
 897	rdp = this_cpu_ptr(&rcu_data);
 898	oldval = rdp->dynticks_nesting;
 899	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
 900	if (oldval) {
 901		// RCU was already watching, so just do accounting and leave.
 902		rdp->dynticks_nesting++;
 903		return;
 904	}
 905	rcu_dynticks_task_exit();
 906	// RCU is not watching here ...
 907	rcu_dynticks_eqs_exit();
 908	// ... but is watching here.
 909	instrumentation_begin();
 910
 911	// instrumentation for the noinstr rcu_dynticks_eqs_exit()
 912	instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
 913
 914	rcu_cleanup_after_idle();
 915	trace_rcu_dyntick(TPS("End"), rdp->dynticks_nesting, 1, atomic_read(&rdp->dynticks));
 916	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
 917	WRITE_ONCE(rdp->dynticks_nesting, 1);
 918	WARN_ON_ONCE(rdp->dynticks_nmi_nesting);
 919	WRITE_ONCE(rdp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
 920	instrumentation_end();
 921}
 922
 923/**
 924 * rcu_idle_exit - inform RCU that current CPU is leaving idle
 925 *
 926 * Exit idle mode, in other words, -enter- the mode in which RCU
 927 * read-side critical sections can occur.
 928 *
 929 * If you add or remove a call to rcu_idle_exit(), be sure to test with
 930 * CONFIG_RCU_EQS_DEBUG=y.
 931 */
 932void rcu_idle_exit(void)
 933{
 934	unsigned long flags;
 935
 936	local_irq_save(flags);
 937	rcu_eqs_exit(false);
 938	local_irq_restore(flags);
 939}
 940EXPORT_SYMBOL_GPL(rcu_idle_exit);
 941
 942#ifdef CONFIG_NO_HZ_FULL
 943/**
 944 * rcu_user_exit - inform RCU that we are exiting userspace.
 945 *
 946 * Exit RCU idle mode while entering the kernel because it can
 947 * run a RCU read side critical section anytime.
 948 *
 949 * If you add or remove a call to rcu_user_exit(), be sure to test with
 950 * CONFIG_RCU_EQS_DEBUG=y.
 951 */
 952void noinstr rcu_user_exit(void)
 953{
 954	rcu_eqs_exit(true);
 955}
 956
 957/**
 958 * __rcu_irq_enter_check_tick - Enable scheduler tick on CPU if RCU needs it.
 959 *
 960 * The scheduler tick is not normally enabled when CPUs enter the kernel
 961 * from nohz_full userspace execution.  After all, nohz_full userspace
 962 * execution is an RCU quiescent state and the time executing in the kernel
 963 * is quite short.  Except of course when it isn't.  And it is not hard to
 964 * cause a large system to spend tens of seconds or even minutes looping
 965 * in the kernel, which can cause a number of problems, include RCU CPU
 966 * stall warnings.
 967 *
 968 * Therefore, if a nohz_full CPU fails to report a quiescent state
 969 * in a timely manner, the RCU grace-period kthread sets that CPU's
 970 * ->rcu_urgent_qs flag with the expectation that the next interrupt or
 971 * exception will invoke this function, which will turn on the scheduler
 972 * tick, which will enable RCU to detect that CPU's quiescent states,
 973 * for example, due to cond_resched() calls in CONFIG_PREEMPT=n kernels.
 974 * The tick will be disabled once a quiescent state is reported for
 975 * this CPU.
 976 *
 977 * Of course, in carefully tuned systems, there might never be an
 978 * interrupt or exception.  In that case, the RCU grace-period kthread
 979 * will eventually cause one to happen.  However, in less carefully
 980 * controlled environments, this function allows RCU to get what it
 981 * needs without creating otherwise useless interruptions.
 982 */
 983void __rcu_irq_enter_check_tick(void)
 984{
 985	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 986
 987	// If we're here from NMI there's nothing to do.
 988	if (in_nmi())
 989		return;
 990
 991	RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
 992			 "Illegal rcu_irq_enter_check_tick() from extended quiescent state");
 993
 994	if (!tick_nohz_full_cpu(rdp->cpu) ||
 995	    !READ_ONCE(rdp->rcu_urgent_qs) ||
 996	    READ_ONCE(rdp->rcu_forced_tick)) {
 997		// RCU doesn't need nohz_full help from this CPU, or it is
 998		// already getting that help.
 999		return;
1000	}
1001
1002	// We get here only when not in an extended quiescent state and
1003	// from interrupts (as opposed to NMIs).  Therefore, (1) RCU is
1004	// already watching and (2) The fact that we are in an interrupt
1005	// handler and that the rcu_node lock is an irq-disabled lock
1006	// prevents self-deadlock.  So we can safely recheck under the lock.
1007	// Note that the nohz_full state currently cannot change.
1008	raw_spin_lock_rcu_node(rdp->mynode);
1009	if (rdp->rcu_urgent_qs && !rdp->rcu_forced_tick) {
1010		// A nohz_full CPU is in the kernel and RCU needs a
1011		// quiescent state.  Turn on the tick!
1012		WRITE_ONCE(rdp->rcu_forced_tick, true);
1013		tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
1014	}
1015	raw_spin_unlock_rcu_node(rdp->mynode);
1016}
 
1017#endif /* CONFIG_NO_HZ_FULL */
1018
1019/**
1020 * rcu_nmi_enter - inform RCU of entry to NMI context
1021 *
1022 * If the CPU was idle from RCU's viewpoint, update rdp->dynticks and
1023 * rdp->dynticks_nmi_nesting to let the RCU grace-period handling know
1024 * that the CPU is active.  This implementation permits nested NMIs, as
1025 * long as the nesting level does not overflow an int.  (You will probably
1026 * run out of stack space first.)
1027 *
1028 * If you add or remove a call to rcu_nmi_enter(), be sure to test
1029 * with CONFIG_RCU_EQS_DEBUG=y.
1030 */
1031noinstr void rcu_nmi_enter(void)
1032{
1033	long incby = 2;
1034	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1035
1036	/* Complain about underflow. */
1037	WARN_ON_ONCE(rdp->dynticks_nmi_nesting < 0);
1038
1039	/*
1040	 * If idle from RCU viewpoint, atomically increment ->dynticks
1041	 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
1042	 * Otherwise, increment ->dynticks_nmi_nesting by two.  This means
1043	 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
1044	 * to be in the outermost NMI handler that interrupted an RCU-idle
1045	 * period (observation due to Andy Lutomirski).
1046	 */
1047	if (rcu_dynticks_curr_cpu_in_eqs()) {
1048
1049		if (!in_nmi())
1050			rcu_dynticks_task_exit();
1051
1052		// RCU is not watching here ...
1053		rcu_dynticks_eqs_exit();
1054		// ... but is watching here.
1055
1056		if (!in_nmi()) {
1057			instrumentation_begin();
1058			rcu_cleanup_after_idle();
1059			instrumentation_end();
1060		}
1061
1062		instrumentation_begin();
1063		// instrumentation for the noinstr rcu_dynticks_curr_cpu_in_eqs()
1064		instrument_atomic_read(&rdp->dynticks, sizeof(rdp->dynticks));
1065		// instrumentation for the noinstr rcu_dynticks_eqs_exit()
1066		instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
1067
1068		incby = 1;
1069	} else if (!in_nmi()) {
1070		instrumentation_begin();
1071		rcu_irq_enter_check_tick();
1072	} else  {
1073		instrumentation_begin();
1074	}
1075
1076	trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
1077			  rdp->dynticks_nmi_nesting,
1078			  rdp->dynticks_nmi_nesting + incby, atomic_read(&rdp->dynticks));
1079	instrumentation_end();
1080	WRITE_ONCE(rdp->dynticks_nmi_nesting, /* Prevent store tearing. */
1081		   rdp->dynticks_nmi_nesting + incby);
1082	barrier();
1083}
1084
1085/**
1086 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
1087 *
1088 * Enter an interrupt handler, which might possibly result in exiting
1089 * idle mode, in other words, entering the mode in which read-side critical
1090 * sections can occur.  The caller must have disabled interrupts.
1091 *
1092 * Note that the Linux kernel is fully capable of entering an interrupt
1093 * handler that it never exits, for example when doing upcalls to user mode!
1094 * This code assumes that the idle loop never does upcalls to user mode.
1095 * If your architecture's idle loop does do upcalls to user mode (or does
1096 * anything else that results in unbalanced calls to the irq_enter() and
1097 * irq_exit() functions), RCU will give you what you deserve, good and hard.
1098 * But very infrequently and irreproducibly.
1099 *
1100 * Use things like work queues to work around this limitation.
1101 *
1102 * You have been warned.
1103 *
1104 * If you add or remove a call to rcu_irq_enter(), be sure to test with
1105 * CONFIG_RCU_EQS_DEBUG=y.
1106 */
1107noinstr void rcu_irq_enter(void)
1108{
1109	lockdep_assert_irqs_disabled();
1110	rcu_nmi_enter();
1111}
1112
1113/*
1114 * Wrapper for rcu_irq_enter() where interrupts are enabled.
 
 
 
 
 
1115 *
1116 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
1117 * with CONFIG_RCU_EQS_DEBUG=y.
1118 */
1119void rcu_irq_enter_irqson(void)
1120{
1121	unsigned long flags;
1122
1123	local_irq_save(flags);
1124	rcu_irq_enter();
1125	local_irq_restore(flags);
1126}
1127
1128/*
1129 * If any sort of urgency was applied to the current CPU (for example,
1130 * the scheduler-clock interrupt was enabled on a nohz_full CPU) in order
1131 * to get to a quiescent state, disable it.
1132 */
1133static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp)
1134{
1135	raw_lockdep_assert_held_rcu_node(rdp->mynode);
1136	WRITE_ONCE(rdp->rcu_urgent_qs, false);
1137	WRITE_ONCE(rdp->rcu_need_heavy_qs, false);
1138	if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) {
1139		tick_dep_clear_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
1140		WRITE_ONCE(rdp->rcu_forced_tick, false);
1141	}
1142}
1143
1144/**
1145 * rcu_is_watching - see if RCU thinks that the current CPU is not idle
1146 *
1147 * Return true if RCU is watching the running CPU, which means that this
1148 * CPU can safely enter RCU read-side critical sections.  In other words,
1149 * if the current CPU is not in its idle loop or is in an interrupt or
1150 * NMI handler, return true.
 
 
 
 
1151 *
1152 * Make notrace because it can be called by the internal functions of
1153 * ftrace, and making this notrace removes unnecessary recursion calls.
1154 */
1155notrace bool rcu_is_watching(void)
1156{
1157	bool ret;
1158
1159	preempt_disable_notrace();
1160	ret = !rcu_dynticks_curr_cpu_in_eqs();
1161	preempt_enable_notrace();
1162	return ret;
1163}
1164EXPORT_SYMBOL_GPL(rcu_is_watching);
1165
1166/*
1167 * If a holdout task is actually running, request an urgent quiescent
1168 * state from its CPU.  This is unsynchronized, so migrations can cause
1169 * the request to go to the wrong CPU.  Which is OK, all that will happen
1170 * is that the CPU's next context switch will be a bit slower and next
1171 * time around this task will generate another request.
1172 */
1173void rcu_request_urgent_qs_task(struct task_struct *t)
1174{
1175	int cpu;
1176
1177	barrier();
1178	cpu = task_cpu(t);
1179	if (!task_curr(t))
1180		return; /* This task is not running on that CPU. */
1181	smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
1182}
1183
1184#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1185
1186/*
1187 * Is the current CPU online as far as RCU is concerned?
1188 *
1189 * Disable preemption to avoid false positives that could otherwise
1190 * happen due to the current CPU number being sampled, this task being
1191 * preempted, its old CPU being taken offline, resuming on some other CPU,
1192 * then determining that its old CPU is now offline.
1193 *
1194 * Disable checking if in an NMI handler because we cannot safely
1195 * report errors from NMI handlers anyway.  In addition, it is OK to use
1196 * RCU on an offline processor during initial boot, hence the check for
1197 * rcu_scheduler_fully_active.
1198 */
1199bool rcu_lockdep_current_cpu_online(void)
1200{
1201	struct rcu_data *rdp;
1202	struct rcu_node *rnp;
1203	bool ret = false;
1204
1205	if (in_nmi() || !rcu_scheduler_fully_active)
1206		return true;
1207	preempt_disable_notrace();
1208	rdp = this_cpu_ptr(&rcu_data);
1209	rnp = rdp->mynode;
1210	if (rdp->grpmask & rcu_rnp_online_cpus(rnp) || READ_ONCE(rnp->ofl_seq) & 0x1)
1211		ret = true;
1212	preempt_enable_notrace();
1213	return ret;
1214}
1215EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1216
1217#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1218
1219/*
1220 * When trying to report a quiescent state on behalf of some other CPU,
1221 * it is our responsibility to check for and handle potential overflow
1222 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
1223 * After all, the CPU might be in deep idle state, and thus executing no
1224 * code whatsoever.
1225 */
1226static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
1227{
1228	raw_lockdep_assert_held_rcu_node(rnp);
1229	if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
1230			 rnp->gp_seq))
1231		WRITE_ONCE(rdp->gpwrap, true);
1232	if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
1233		rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
1234}
1235
1236/*
1237 * Snapshot the specified CPU's dynticks counter so that we can later
1238 * credit them with an implicit quiescent state.  Return 1 if this CPU
1239 * is in dynticks idle mode, which is an extended quiescent state.
1240 */
1241static int dyntick_save_progress_counter(struct rcu_data *rdp)
1242{
1243	rdp->dynticks_snap = rcu_dynticks_snap(rdp);
1244	if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1245		trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1246		rcu_gpnum_ovf(rdp->mynode, rdp);
1247		return 1;
1248	}
1249	return 0;
1250}
1251
1252/*
1253 * Return true if the specified CPU has passed through a quiescent
1254 * state by virtue of being in or having passed through an dynticks
1255 * idle state since the last call to dyntick_save_progress_counter()
1256 * for this same CPU, or by virtue of having been offline.
 
 
 
 
1257 */
1258static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1259{
1260	unsigned long jtsq;
1261	bool *rnhqp;
1262	bool *ruqp;
1263	struct rcu_node *rnp = rdp->mynode;
1264
1265	/*
1266	 * If the CPU passed through or entered a dynticks idle phase with
1267	 * no active irq/NMI handlers, then we can safely pretend that the CPU
1268	 * already acknowledged the request to pass through a quiescent
1269	 * state.  Either way, that CPU cannot possibly be in an RCU
1270	 * read-side critical section that started before the beginning
1271	 * of the current RCU grace period.
1272	 */
1273	if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
1274		trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1275		rcu_gpnum_ovf(rnp, rdp);
1276		return 1;
1277	}
1278
1279	/*
1280	 * Complain if a CPU that is considered to be offline from RCU's
1281	 * perspective has not yet reported a quiescent state.  After all,
1282	 * the offline CPU should have reported a quiescent state during
1283	 * the CPU-offline process, or, failing that, by rcu_gp_init()
1284	 * if it ran concurrently with either the CPU going offline or the
1285	 * last task on a leaf rcu_node structure exiting its RCU read-side
1286	 * critical section while all CPUs corresponding to that structure
1287	 * are offline.  This added warning detects bugs in any of these
1288	 * code paths.
1289	 *
1290	 * The rcu_node structure's ->lock is held here, which excludes
1291	 * the relevant portions the CPU-hotplug code, the grace-period
1292	 * initialization code, and the rcu_read_unlock() code paths.
1293	 *
1294	 * For more detail, please refer to the "Hotplug CPU" section
1295	 * of RCU's Requirements documentation.
1296	 */
1297	if (WARN_ON_ONCE(!(rdp->grpmask & rcu_rnp_online_cpus(rnp)))) {
1298		bool onl;
1299		struct rcu_node *rnp1;
1300
1301		pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
1302			__func__, rnp->grplo, rnp->grphi, rnp->level,
1303			(long)rnp->gp_seq, (long)rnp->completedqs);
1304		for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
1305			pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
1306				__func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
1307		onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
1308		pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
1309			__func__, rdp->cpu, ".o"[onl],
1310			(long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
1311			(long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
1312		return 1; /* Break things loose after complaining. */
1313	}
1314
1315	/*
1316	 * A CPU running for an extended time within the kernel can
1317	 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
1318	 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
1319	 * both .rcu_need_heavy_qs and .rcu_urgent_qs.  Note that the
1320	 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
1321	 * variable are safe because the assignments are repeated if this
1322	 * CPU failed to pass through a quiescent state.  This code
1323	 * also checks .jiffies_resched in case jiffies_to_sched_qs
1324	 * is set way high.
1325	 */
1326	jtsq = READ_ONCE(jiffies_to_sched_qs);
1327	ruqp = per_cpu_ptr(&rcu_data.rcu_urgent_qs, rdp->cpu);
1328	rnhqp = &per_cpu(rcu_data.rcu_need_heavy_qs, rdp->cpu);
1329	if (!READ_ONCE(*rnhqp) &&
1330	    (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
1331	     time_after(jiffies, rcu_state.jiffies_resched) ||
1332	     rcu_state.cbovld)) {
1333		WRITE_ONCE(*rnhqp, true);
1334		/* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1335		smp_store_release(ruqp, true);
1336	} else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
1337		WRITE_ONCE(*ruqp, true);
1338	}
1339
1340	/*
1341	 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
1342	 * The above code handles this, but only for straight cond_resched().
1343	 * And some in-kernel loops check need_resched() before calling
1344	 * cond_resched(), which defeats the above code for CPUs that are
1345	 * running in-kernel with scheduling-clock interrupts disabled.
1346	 * So hit them over the head with the resched_cpu() hammer!
1347	 */
1348	if (tick_nohz_full_cpu(rdp->cpu) &&
1349	    (time_after(jiffies, READ_ONCE(rdp->last_fqs_resched) + jtsq * 3) ||
1350	     rcu_state.cbovld)) {
1351		WRITE_ONCE(*ruqp, true);
1352		resched_cpu(rdp->cpu);
1353		WRITE_ONCE(rdp->last_fqs_resched, jiffies);
 
1354	}
1355
1356	/*
1357	 * If more than halfway to RCU CPU stall-warning time, invoke
1358	 * resched_cpu() more frequently to try to loosen things up a bit.
1359	 * Also check to see if the CPU is getting hammered with interrupts,
1360	 * but only once per grace period, just to keep the IPIs down to
1361	 * a dull roar.
1362	 */
1363	if (time_after(jiffies, rcu_state.jiffies_resched)) {
1364		if (time_after(jiffies,
1365			       READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
1366			resched_cpu(rdp->cpu);
1367			WRITE_ONCE(rdp->last_fqs_resched, jiffies);
 
1368		}
1369		if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1370		    !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
1371		    (rnp->ffmask & rdp->grpmask)) {
1372			rdp->rcu_iw_pending = true;
1373			rdp->rcu_iw_gp_seq = rnp->gp_seq;
1374			irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1375		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1376	}
1377
1378	return 0;
1379}
1380
1381/* Trace-event wrapper function for trace_rcu_future_grace_period.  */
1382static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1383			      unsigned long gp_seq_req, const char *s)
1384{
1385	trace_rcu_future_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
1386				      gp_seq_req, rnp->level,
1387				      rnp->grplo, rnp->grphi, s);
1388}
1389
1390/*
1391 * rcu_start_this_gp - Request the start of a particular grace period
1392 * @rnp_start: The leaf node of the CPU from which to start.
1393 * @rdp: The rcu_data corresponding to the CPU from which to start.
1394 * @gp_seq_req: The gp_seq of the grace period to start.
1395 *
1396 * Start the specified grace period, as needed to handle newly arrived
1397 * callbacks.  The required future grace periods are recorded in each
1398 * rcu_node structure's ->gp_seq_needed field.  Returns true if there
1399 * is reason to awaken the grace-period kthread.
1400 *
1401 * The caller must hold the specified rcu_node structure's ->lock, which
1402 * is why the caller is responsible for waking the grace-period kthread.
1403 *
1404 * Returns true if the GP thread needs to be awakened else false.
1405 */
1406static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
1407			      unsigned long gp_seq_req)
1408{
1409	bool ret = false;
1410	struct rcu_node *rnp;
1411
1412	/*
1413	 * Use funnel locking to either acquire the root rcu_node
1414	 * structure's lock or bail out if the need for this grace period
1415	 * has already been recorded -- or if that grace period has in
1416	 * fact already started.  If there is already a grace period in
1417	 * progress in a non-leaf node, no recording is needed because the
1418	 * end of the grace period will scan the leaf rcu_node structures.
1419	 * Note that rnp_start->lock must not be released.
1420	 */
1421	raw_lockdep_assert_held_rcu_node(rnp_start);
1422	trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
1423	for (rnp = rnp_start; 1; rnp = rnp->parent) {
1424		if (rnp != rnp_start)
1425			raw_spin_lock_rcu_node(rnp);
1426		if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
1427		    rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
1428		    (rnp != rnp_start &&
1429		     rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
1430			trace_rcu_this_gp(rnp, rdp, gp_seq_req,
1431					  TPS("Prestarted"));
1432			goto unlock_out;
1433		}
1434		WRITE_ONCE(rnp->gp_seq_needed, gp_seq_req);
1435		if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
1436			/*
1437			 * We just marked the leaf or internal node, and a
1438			 * grace period is in progress, which means that
1439			 * rcu_gp_cleanup() will see the marking.  Bail to
1440			 * reduce contention.
1441			 */
1442			trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
1443					  TPS("Startedleaf"));
1444			goto unlock_out;
1445		}
1446		if (rnp != rnp_start && rnp->parent != NULL)
1447			raw_spin_unlock_rcu_node(rnp);
1448		if (!rnp->parent)
1449			break;  /* At root, and perhaps also leaf. */
1450	}
1451
1452	/* If GP already in progress, just leave, otherwise start one. */
1453	if (rcu_gp_in_progress()) {
1454		trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
1455		goto unlock_out;
1456	}
1457	trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
1458	WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
1459	WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
1460	if (!READ_ONCE(rcu_state.gp_kthread)) {
1461		trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
1462		goto unlock_out;
1463	}
1464	trace_rcu_grace_period(rcu_state.name, data_race(rcu_state.gp_seq), TPS("newreq"));
1465	ret = true;  /* Caller must wake GP kthread. */
1466unlock_out:
1467	/* Push furthest requested GP to leaf node and rcu_data structure. */
1468	if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
1469		WRITE_ONCE(rnp_start->gp_seq_needed, rnp->gp_seq_needed);
1470		WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1471	}
1472	if (rnp != rnp_start)
1473		raw_spin_unlock_rcu_node(rnp);
1474	return ret;
1475}
1476
1477/*
1478 * Clean up any old requests for the just-ended grace period.  Also return
1479 * whether any additional grace periods have been requested.
1480 */
1481static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
1482{
1483	bool needmore;
1484	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1485
1486	needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1487	if (!needmore)
1488		rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1489	trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1490			  needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1491	return needmore;
1492}
1493
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1494/*
1495 * Awaken the grace-period kthread.  Don't do a self-awaken (unless in an
1496 * interrupt or softirq handler, in which case we just might immediately
1497 * sleep upon return, resulting in a grace-period hang), and don't bother
1498 * awakening when there is nothing for the grace-period kthread to do
1499 * (as in several CPUs raced to awaken, we lost), and finally don't try
1500 * to awaken a kthread that has not yet been created.  If all those checks
1501 * are passed, track some debug information and awaken.
1502 *
1503 * So why do the self-wakeup when in an interrupt or softirq handler
1504 * in the grace-period kthread's context?  Because the kthread might have
1505 * been interrupted just as it was going to sleep, and just after the final
1506 * pre-sleep check of the awaken condition.  In this case, a wakeup really
1507 * is required, and is therefore supplied.
1508 */
1509static void rcu_gp_kthread_wake(void)
1510{
1511	struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
1512
1513	if ((current == t && !in_irq() && !in_serving_softirq()) ||
1514	    !READ_ONCE(rcu_state.gp_flags) || !t)
1515		return;
1516	WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
1517	WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
1518	swake_up_one(&rcu_state.gp_wq);
1519}
1520
1521/*
1522 * If there is room, assign a ->gp_seq number to any callbacks on this
1523 * CPU that have not already been assigned.  Also accelerate any callbacks
1524 * that were previously assigned a ->gp_seq number that has since proven
1525 * to be too conservative, which can happen if callbacks get assigned a
1526 * ->gp_seq number while RCU is idle, but with reference to a non-root
1527 * rcu_node structure.  This function is idempotent, so it does not hurt
1528 * to call it repeatedly.  Returns an flag saying that we should awaken
1529 * the RCU grace-period kthread.
1530 *
1531 * The caller must hold rnp->lock with interrupts disabled.
1532 */
1533static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1534{
1535	unsigned long gp_seq_req;
1536	bool ret = false;
1537
1538	rcu_lockdep_assert_cblist_protected(rdp);
1539	raw_lockdep_assert_held_rcu_node(rnp);
1540
1541	/* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1542	if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1543		return false;
1544
1545	trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPreAcc"));
1546
1547	/*
1548	 * Callbacks are often registered with incomplete grace-period
1549	 * information.  Something about the fact that getting exact
1550	 * information requires acquiring a global lock...  RCU therefore
1551	 * makes a conservative estimate of the grace period number at which
1552	 * a given callback will become ready to invoke.	The following
1553	 * code checks this estimate and improves it when possible, thus
1554	 * accelerating callback invocation to an earlier grace-period
1555	 * number.
1556	 */
1557	gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1558	if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1559		ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1560
1561	/* Trace depending on how much we were able to accelerate. */
1562	if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1563		trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccWaitCB"));
1564	else
1565		trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccReadyCB"));
1566
1567	trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPostAcc"));
1568
1569	return ret;
1570}
1571
1572/*
1573 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1574 * rcu_node structure's ->lock be held.  It consults the cached value
1575 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1576 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1577 * while holding the leaf rcu_node structure's ->lock.
1578 */
1579static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1580					struct rcu_data *rdp)
1581{
1582	unsigned long c;
1583	bool needwake;
1584
1585	rcu_lockdep_assert_cblist_protected(rdp);
1586	c = rcu_seq_snap(&rcu_state.gp_seq);
1587	if (!READ_ONCE(rdp->gpwrap) && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1588		/* Old request still live, so mark recent callbacks. */
1589		(void)rcu_segcblist_accelerate(&rdp->cblist, c);
1590		return;
1591	}
1592	raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1593	needwake = rcu_accelerate_cbs(rnp, rdp);
1594	raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1595	if (needwake)
1596		rcu_gp_kthread_wake();
1597}
1598
1599/*
1600 * Move any callbacks whose grace period has completed to the
1601 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1602 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1603 * sublist.  This function is idempotent, so it does not hurt to
1604 * invoke it repeatedly.  As long as it is not invoked -too- often...
1605 * Returns true if the RCU grace-period kthread needs to be awakened.
1606 *
1607 * The caller must hold rnp->lock with interrupts disabled.
1608 */
1609static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1610{
1611	rcu_lockdep_assert_cblist_protected(rdp);
1612	raw_lockdep_assert_held_rcu_node(rnp);
1613
1614	/* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1615	if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1616		return false;
1617
1618	/*
1619	 * Find all callbacks whose ->gp_seq numbers indicate that they
1620	 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1621	 */
1622	rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1623
1624	/* Classify any remaining callbacks. */
1625	return rcu_accelerate_cbs(rnp, rdp);
1626}
1627
1628/*
1629 * Move and classify callbacks, but only if doing so won't require
1630 * that the RCU grace-period kthread be awakened.
1631 */
1632static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
1633						  struct rcu_data *rdp)
1634{
1635	rcu_lockdep_assert_cblist_protected(rdp);
1636	if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) ||
1637	    !raw_spin_trylock_rcu_node(rnp))
1638		return;
1639	WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
 
 
1640	raw_spin_unlock_rcu_node(rnp);
1641}
1642
1643/*
1644 * In CONFIG_RCU_STRICT_GRACE_PERIOD=y kernels, attempt to generate a
1645 * quiescent state.  This is intended to be invoked when the CPU notices
1646 * a new grace period.
1647 */
1648static void rcu_strict_gp_check_qs(void)
1649{
1650	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
1651		rcu_read_lock();
1652		rcu_read_unlock();
1653	}
1654}
1655
1656/*
1657 * Update CPU-local rcu_data state to record the beginnings and ends of
1658 * grace periods.  The caller must hold the ->lock of the leaf rcu_node
1659 * structure corresponding to the current CPU, and must have irqs disabled.
1660 * Returns true if the grace-period kthread needs to be awakened.
1661 */
1662static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
1663{
1664	bool ret = false;
1665	bool need_qs;
1666	const bool offloaded = rcu_rdp_is_offloaded(rdp);
1667
1668	raw_lockdep_assert_held_rcu_node(rnp);
1669
1670	if (rdp->gp_seq == rnp->gp_seq)
1671		return false; /* Nothing to do. */
1672
1673	/* Handle the ends of any preceding grace periods first. */
1674	if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1675	    unlikely(READ_ONCE(rdp->gpwrap))) {
1676		if (!offloaded)
1677			ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
1678		rdp->core_needs_qs = false;
1679		trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1680	} else {
1681		if (!offloaded)
1682			ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
1683		if (rdp->core_needs_qs)
1684			rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask);
1685	}
1686
1687	/* Now handle the beginnings of any new-to-this-CPU grace periods. */
1688	if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1689	    unlikely(READ_ONCE(rdp->gpwrap))) {
1690		/*
1691		 * If the current grace period is waiting for this CPU,
1692		 * set up to detect a quiescent state, otherwise don't
1693		 * go looking for one.
1694		 */
1695		trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1696		need_qs = !!(rnp->qsmask & rdp->grpmask);
1697		rdp->cpu_no_qs.b.norm = need_qs;
1698		rdp->core_needs_qs = need_qs;
1699		zero_cpu_stall_ticks(rdp);
1700	}
1701	rdp->gp_seq = rnp->gp_seq;  /* Remember new grace-period state. */
1702	if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
1703		WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
 
 
1704	WRITE_ONCE(rdp->gpwrap, false);
1705	rcu_gpnum_ovf(rnp, rdp);
1706	return ret;
1707}
1708
1709static void note_gp_changes(struct rcu_data *rdp)
1710{
1711	unsigned long flags;
1712	bool needwake;
1713	struct rcu_node *rnp;
1714
1715	local_irq_save(flags);
1716	rnp = rdp->mynode;
1717	if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1718	     !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1719	    !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1720		local_irq_restore(flags);
1721		return;
1722	}
1723	needwake = __note_gp_changes(rnp, rdp);
1724	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1725	rcu_strict_gp_check_qs();
1726	if (needwake)
1727		rcu_gp_kthread_wake();
1728}
1729
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1730static void rcu_gp_slow(int delay)
1731{
1732	if (delay > 0 &&
1733	    !(rcu_seq_ctr(rcu_state.gp_seq) %
1734	      (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1735		schedule_timeout_idle(delay);
1736}
1737
1738static unsigned long sleep_duration;
1739
1740/* Allow rcutorture to stall the grace-period kthread. */
1741void rcu_gp_set_torture_wait(int duration)
1742{
1743	if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST) && duration > 0)
1744		WRITE_ONCE(sleep_duration, duration);
1745}
1746EXPORT_SYMBOL_GPL(rcu_gp_set_torture_wait);
1747
1748/* Actually implement the aforementioned wait. */
1749static void rcu_gp_torture_wait(void)
1750{
1751	unsigned long duration;
1752
1753	if (!IS_ENABLED(CONFIG_RCU_TORTURE_TEST))
1754		return;
1755	duration = xchg(&sleep_duration, 0UL);
1756	if (duration > 0) {
1757		pr_alert("%s: Waiting %lu jiffies\n", __func__, duration);
1758		schedule_timeout_idle(duration);
1759		pr_alert("%s: Wait complete\n", __func__);
1760	}
1761}
1762
1763/*
1764 * Handler for on_each_cpu() to invoke the target CPU's RCU core
1765 * processing.
1766 */
1767static void rcu_strict_gp_boundary(void *unused)
1768{
1769	invoke_rcu_core();
1770}
1771
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1772/*
1773 * Initialize a new grace period.  Return false if no grace period required.
1774 */
1775static bool rcu_gp_init(void)
1776{
1777	unsigned long firstseq;
1778	unsigned long flags;
1779	unsigned long oldmask;
1780	unsigned long mask;
1781	struct rcu_data *rdp;
1782	struct rcu_node *rnp = rcu_get_root();
1783
1784	WRITE_ONCE(rcu_state.gp_activity, jiffies);
1785	raw_spin_lock_irq_rcu_node(rnp);
1786	if (!READ_ONCE(rcu_state.gp_flags)) {
1787		/* Spurious wakeup, tell caller to go back to sleep.  */
1788		raw_spin_unlock_irq_rcu_node(rnp);
1789		return false;
1790	}
1791	WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1792
1793	if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1794		/*
1795		 * Grace period already in progress, don't start another.
1796		 * Not supposed to be able to happen.
1797		 */
1798		raw_spin_unlock_irq_rcu_node(rnp);
1799		return false;
1800	}
1801
1802	/* Advance to a new grace period and initialize state. */
1803	record_gp_stall_check_time();
1804	/* Record GP times before starting GP, hence rcu_seq_start(). */
1805	rcu_seq_start(&rcu_state.gp_seq);
1806	ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
1807	trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
 
1808	raw_spin_unlock_irq_rcu_node(rnp);
1809
1810	/*
1811	 * Apply per-leaf buffered online and offline operations to
1812	 * the rcu_node tree. Note that this new grace period need not
1813	 * wait for subsequent online CPUs, and that RCU hooks in the CPU
1814	 * offlining path, when combined with checks in this function,
1815	 * will handle CPUs that are currently going offline or that will
1816	 * go offline later.  Please also refer to "Hotplug CPU" section
1817	 * of RCU's Requirements documentation.
1818	 */
1819	WRITE_ONCE(rcu_state.gp_state, RCU_GP_ONOFF);
 
1820	rcu_for_each_leaf_node(rnp) {
1821		smp_mb(); // Pair with barriers used when updating ->ofl_seq to odd values.
1822		firstseq = READ_ONCE(rnp->ofl_seq);
1823		if (firstseq & 0x1)
1824			while (firstseq == READ_ONCE(rnp->ofl_seq))
1825				schedule_timeout_idle(1);  // Can't wake unless RCU is watching.
1826		smp_mb(); // Pair with barriers used when updating ->ofl_seq to even values.
1827		raw_spin_lock(&rcu_state.ofl_lock);
1828		raw_spin_lock_irq_rcu_node(rnp);
1829		if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1830		    !rnp->wait_blkd_tasks) {
1831			/* Nothing to do on this leaf rcu_node structure. */
1832			raw_spin_unlock_irq_rcu_node(rnp);
1833			raw_spin_unlock(&rcu_state.ofl_lock);
 
1834			continue;
1835		}
1836
1837		/* Record old state, apply changes to ->qsmaskinit field. */
1838		oldmask = rnp->qsmaskinit;
1839		rnp->qsmaskinit = rnp->qsmaskinitnext;
1840
1841		/* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1842		if (!oldmask != !rnp->qsmaskinit) {
1843			if (!oldmask) { /* First online CPU for rcu_node. */
1844				if (!rnp->wait_blkd_tasks) /* Ever offline? */
1845					rcu_init_new_rnp(rnp);
1846			} else if (rcu_preempt_has_tasks(rnp)) {
1847				rnp->wait_blkd_tasks = true; /* blocked tasks */
1848			} else { /* Last offline CPU and can propagate. */
1849				rcu_cleanup_dead_rnp(rnp);
1850			}
1851		}
1852
1853		/*
1854		 * If all waited-on tasks from prior grace period are
1855		 * done, and if all this rcu_node structure's CPUs are
1856		 * still offline, propagate up the rcu_node tree and
1857		 * clear ->wait_blkd_tasks.  Otherwise, if one of this
1858		 * rcu_node structure's CPUs has since come back online,
1859		 * simply clear ->wait_blkd_tasks.
1860		 */
1861		if (rnp->wait_blkd_tasks &&
1862		    (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1863			rnp->wait_blkd_tasks = false;
1864			if (!rnp->qsmaskinit)
1865				rcu_cleanup_dead_rnp(rnp);
1866		}
1867
1868		raw_spin_unlock_irq_rcu_node(rnp);
1869		raw_spin_unlock(&rcu_state.ofl_lock);
 
1870	}
1871	rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1872
1873	/*
1874	 * Set the quiescent-state-needed bits in all the rcu_node
1875	 * structures for all currently online CPUs in breadth-first
1876	 * order, starting from the root rcu_node structure, relying on the
1877	 * layout of the tree within the rcu_state.node[] array.  Note that
1878	 * other CPUs will access only the leaves of the hierarchy, thus
1879	 * seeing that no grace period is in progress, at least until the
1880	 * corresponding leaf node has been initialized.
1881	 *
1882	 * The grace period cannot complete until the initialization
1883	 * process finishes, because this kthread handles both.
1884	 */
1885	WRITE_ONCE(rcu_state.gp_state, RCU_GP_INIT);
1886	rcu_for_each_node_breadth_first(rnp) {
1887		rcu_gp_slow(gp_init_delay);
1888		raw_spin_lock_irqsave_rcu_node(rnp, flags);
1889		rdp = this_cpu_ptr(&rcu_data);
1890		rcu_preempt_check_blocked_tasks(rnp);
1891		rnp->qsmask = rnp->qsmaskinit;
1892		WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
1893		if (rnp == rdp->mynode)
1894			(void)__note_gp_changes(rnp, rdp);
1895		rcu_preempt_boost_start_gp(rnp);
1896		trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1897					    rnp->level, rnp->grplo,
1898					    rnp->grphi, rnp->qsmask);
1899		/* Quiescent states for tasks on any now-offline CPUs. */
1900		mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1901		rnp->rcu_gp_init_mask = mask;
1902		if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1903			rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1904		else
1905			raw_spin_unlock_irq_rcu_node(rnp);
1906		cond_resched_tasks_rcu_qs();
1907		WRITE_ONCE(rcu_state.gp_activity, jiffies);
1908	}
1909
1910	// If strict, make all CPUs aware of new grace period.
1911	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
1912		on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
1913
1914	return true;
1915}
1916
1917/*
1918 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1919 * time.
1920 */
1921static bool rcu_gp_fqs_check_wake(int *gfp)
1922{
1923	struct rcu_node *rnp = rcu_get_root();
1924
1925	// If under overload conditions, force an immediate FQS scan.
1926	if (*gfp & RCU_GP_FLAG_OVLD)
1927		return true;
1928
1929	// Someone like call_rcu() requested a force-quiescent-state scan.
1930	*gfp = READ_ONCE(rcu_state.gp_flags);
1931	if (*gfp & RCU_GP_FLAG_FQS)
1932		return true;
1933
1934	// The current grace period has completed.
1935	if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1936		return true;
1937
1938	return false;
1939}
1940
1941/*
1942 * Do one round of quiescent-state forcing.
1943 */
1944static void rcu_gp_fqs(bool first_time)
1945{
 
1946	struct rcu_node *rnp = rcu_get_root();
1947
1948	WRITE_ONCE(rcu_state.gp_activity, jiffies);
1949	rcu_state.n_force_qs++;
 
 
 
 
 
 
 
 
 
 
 
1950	if (first_time) {
1951		/* Collect dyntick-idle snapshots. */
1952		force_qs_rnp(dyntick_save_progress_counter);
1953	} else {
1954		/* Handle dyntick-idle and offline CPUs. */
1955		force_qs_rnp(rcu_implicit_dynticks_qs);
1956	}
1957	/* Clear flag to prevent immediate re-entry. */
1958	if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
1959		raw_spin_lock_irq_rcu_node(rnp);
1960		WRITE_ONCE(rcu_state.gp_flags,
1961			   READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
1962		raw_spin_unlock_irq_rcu_node(rnp);
1963	}
1964}
1965
1966/*
1967 * Loop doing repeated quiescent-state forcing until the grace period ends.
1968 */
1969static void rcu_gp_fqs_loop(void)
1970{
1971	bool first_gp_fqs;
1972	int gf = 0;
1973	unsigned long j;
1974	int ret;
1975	struct rcu_node *rnp = rcu_get_root();
1976
1977	first_gp_fqs = true;
1978	j = READ_ONCE(jiffies_till_first_fqs);
1979	if (rcu_state.cbovld)
1980		gf = RCU_GP_FLAG_OVLD;
1981	ret = 0;
1982	for (;;) {
1983		if (!ret) {
 
 
 
 
 
1984			WRITE_ONCE(rcu_state.jiffies_force_qs, jiffies + j);
1985			/*
1986			 * jiffies_force_qs before RCU_GP_WAIT_FQS state
1987			 * update; required for stall checks.
1988			 */
1989			smp_wmb();
1990			WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
1991				   jiffies + (j ? 3 * j : 2));
1992		}
1993		trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1994				       TPS("fqswait"));
1995		WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_FQS);
1996		ret = swait_event_idle_timeout_exclusive(
1997				rcu_state.gp_wq, rcu_gp_fqs_check_wake(&gf), j);
1998		rcu_gp_torture_wait();
1999		WRITE_ONCE(rcu_state.gp_state, RCU_GP_DOING_FQS);
2000		/* Locking provides needed memory barriers. */
2001		/* If grace period done, leave loop. */
 
 
 
 
 
 
 
 
2002		if (!READ_ONCE(rnp->qsmask) &&
2003		    !rcu_preempt_blocked_readers_cgp(rnp))
2004			break;
2005		/* If time for quiescent-state forcing, do it. */
2006		if (!time_after(rcu_state.jiffies_force_qs, jiffies) ||
2007		    (gf & (RCU_GP_FLAG_FQS | RCU_GP_FLAG_OVLD))) {
2008			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2009					       TPS("fqsstart"));
2010			rcu_gp_fqs(first_gp_fqs);
2011			gf = 0;
2012			if (first_gp_fqs) {
2013				first_gp_fqs = false;
2014				gf = rcu_state.cbovld ? RCU_GP_FLAG_OVLD : 0;
2015			}
2016			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2017					       TPS("fqsend"));
2018			cond_resched_tasks_rcu_qs();
2019			WRITE_ONCE(rcu_state.gp_activity, jiffies);
2020			ret = 0; /* Force full wait till next FQS. */
2021			j = READ_ONCE(jiffies_till_next_fqs);
2022		} else {
2023			/* Deal with stray signal. */
2024			cond_resched_tasks_rcu_qs();
2025			WRITE_ONCE(rcu_state.gp_activity, jiffies);
2026			WARN_ON(signal_pending(current));
2027			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2028					       TPS("fqswaitsig"));
2029			ret = 1; /* Keep old FQS timing. */
2030			j = jiffies;
2031			if (time_after(jiffies, rcu_state.jiffies_force_qs))
2032				j = 1;
2033			else
2034				j = rcu_state.jiffies_force_qs - j;
2035			gf = 0;
2036		}
2037	}
2038}
2039
2040/*
2041 * Clean up after the old grace period.
2042 */
2043static noinline void rcu_gp_cleanup(void)
2044{
2045	int cpu;
2046	bool needgp = false;
2047	unsigned long gp_duration;
2048	unsigned long new_gp_seq;
2049	bool offloaded;
2050	struct rcu_data *rdp;
2051	struct rcu_node *rnp = rcu_get_root();
2052	struct swait_queue_head *sq;
2053
2054	WRITE_ONCE(rcu_state.gp_activity, jiffies);
2055	raw_spin_lock_irq_rcu_node(rnp);
2056	rcu_state.gp_end = jiffies;
2057	gp_duration = rcu_state.gp_end - rcu_state.gp_start;
2058	if (gp_duration > rcu_state.gp_max)
2059		rcu_state.gp_max = gp_duration;
2060
2061	/*
2062	 * We know the grace period is complete, but to everyone else
2063	 * it appears to still be ongoing.  But it is also the case
2064	 * that to everyone else it looks like there is nothing that
2065	 * they can do to advance the grace period.  It is therefore
2066	 * safe for us to drop the lock in order to mark the grace
2067	 * period as completed in all of the rcu_node structures.
2068	 */
 
2069	raw_spin_unlock_irq_rcu_node(rnp);
2070
2071	/*
2072	 * Propagate new ->gp_seq value to rcu_node structures so that
2073	 * other CPUs don't have to wait until the start of the next grace
2074	 * period to process their callbacks.  This also avoids some nasty
2075	 * RCU grace-period initialization races by forcing the end of
2076	 * the current grace period to be completely recorded in all of
2077	 * the rcu_node structures before the beginning of the next grace
2078	 * period is recorded in any of the rcu_node structures.
2079	 */
2080	new_gp_seq = rcu_state.gp_seq;
2081	rcu_seq_end(&new_gp_seq);
2082	rcu_for_each_node_breadth_first(rnp) {
2083		raw_spin_lock_irq_rcu_node(rnp);
2084		if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
2085			dump_blkd_tasks(rnp, 10);
2086		WARN_ON_ONCE(rnp->qsmask);
2087		WRITE_ONCE(rnp->gp_seq, new_gp_seq);
 
 
2088		rdp = this_cpu_ptr(&rcu_data);
2089		if (rnp == rdp->mynode)
2090			needgp = __note_gp_changes(rnp, rdp) || needgp;
2091		/* smp_mb() provided by prior unlock-lock pair. */
2092		needgp = rcu_future_gp_cleanup(rnp) || needgp;
2093		// Reset overload indication for CPUs no longer overloaded
2094		if (rcu_is_leaf_node(rnp))
2095			for_each_leaf_node_cpu_mask(rnp, cpu, rnp->cbovldmask) {
2096				rdp = per_cpu_ptr(&rcu_data, cpu);
2097				check_cb_ovld_locked(rdp, rnp);
2098			}
2099		sq = rcu_nocb_gp_get(rnp);
2100		raw_spin_unlock_irq_rcu_node(rnp);
2101		rcu_nocb_gp_cleanup(sq);
2102		cond_resched_tasks_rcu_qs();
2103		WRITE_ONCE(rcu_state.gp_activity, jiffies);
2104		rcu_gp_slow(gp_cleanup_delay);
2105	}
2106	rnp = rcu_get_root();
2107	raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
2108
2109	/* Declare grace period done, trace first to use old GP number. */
2110	trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
2111	rcu_seq_end(&rcu_state.gp_seq);
2112	ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
2113	WRITE_ONCE(rcu_state.gp_state, RCU_GP_IDLE);
2114	/* Check for GP requests since above loop. */
2115	rdp = this_cpu_ptr(&rcu_data);
2116	if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
2117		trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
2118				  TPS("CleanupMore"));
2119		needgp = true;
2120	}
2121	/* Advance CBs to reduce false positives below. */
2122	offloaded = rcu_rdp_is_offloaded(rdp);
2123	if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
 
 
 
 
 
 
 
 
 
 
 
2124		WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
2125		WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
2126		trace_rcu_grace_period(rcu_state.name,
2127				       rcu_state.gp_seq,
2128				       TPS("newreq"));
2129	} else {
2130		WRITE_ONCE(rcu_state.gp_flags,
2131			   rcu_state.gp_flags & RCU_GP_FLAG_INIT);
 
 
 
 
 
 
2132	}
2133	raw_spin_unlock_irq_rcu_node(rnp);
2134
2135	// If strict, make all CPUs aware of the end of the old grace period.
2136	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
2137		on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
2138}
2139
2140/*
2141 * Body of kthread that handles grace periods.
2142 */
2143static int __noreturn rcu_gp_kthread(void *unused)
2144{
2145	rcu_bind_gp_kthread();
2146	for (;;) {
2147
2148		/* Handle grace-period start. */
2149		for (;;) {
2150			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2151					       TPS("reqwait"));
2152			WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_GPS);
2153			swait_event_idle_exclusive(rcu_state.gp_wq,
2154					 READ_ONCE(rcu_state.gp_flags) &
2155					 RCU_GP_FLAG_INIT);
2156			rcu_gp_torture_wait();
2157			WRITE_ONCE(rcu_state.gp_state, RCU_GP_DONE_GPS);
2158			/* Locking provides needed memory barrier. */
2159			if (rcu_gp_init())
2160				break;
2161			cond_resched_tasks_rcu_qs();
2162			WRITE_ONCE(rcu_state.gp_activity, jiffies);
2163			WARN_ON(signal_pending(current));
2164			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2165					       TPS("reqwaitsig"));
2166		}
2167
2168		/* Handle quiescent-state forcing. */
2169		rcu_gp_fqs_loop();
2170
2171		/* Handle grace-period end. */
2172		WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANUP);
2173		rcu_gp_cleanup();
2174		WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANED);
2175	}
2176}
2177
2178/*
2179 * Report a full set of quiescent states to the rcu_state data structure.
2180 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
2181 * another grace period is required.  Whether we wake the grace-period
2182 * kthread or it awakens itself for the next round of quiescent-state
2183 * forcing, that kthread will clean up after the just-completed grace
2184 * period.  Note that the caller must hold rnp->lock, which is released
2185 * before return.
2186 */
2187static void rcu_report_qs_rsp(unsigned long flags)
2188	__releases(rcu_get_root()->lock)
2189{
2190	raw_lockdep_assert_held_rcu_node(rcu_get_root());
2191	WARN_ON_ONCE(!rcu_gp_in_progress());
2192	WRITE_ONCE(rcu_state.gp_flags,
2193		   READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2194	raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
2195	rcu_gp_kthread_wake();
2196}
2197
2198/*
2199 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2200 * Allows quiescent states for a group of CPUs to be reported at one go
2201 * to the specified rcu_node structure, though all the CPUs in the group
2202 * must be represented by the same rcu_node structure (which need not be a
2203 * leaf rcu_node structure, though it often will be).  The gps parameter
2204 * is the grace-period snapshot, which means that the quiescent states
2205 * are valid only if rnp->gp_seq is equal to gps.  That structure's lock
2206 * must be held upon entry, and it is released before return.
2207 *
2208 * As a special case, if mask is zero, the bit-already-cleared check is
2209 * disabled.  This allows propagating quiescent state due to resumed tasks
2210 * during grace-period initialization.
2211 */
2212static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
2213			      unsigned long gps, unsigned long flags)
2214	__releases(rnp->lock)
2215{
2216	unsigned long oldmask = 0;
2217	struct rcu_node *rnp_c;
2218
2219	raw_lockdep_assert_held_rcu_node(rnp);
2220
2221	/* Walk up the rcu_node hierarchy. */
2222	for (;;) {
2223		if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
2224
2225			/*
2226			 * Our bit has already been cleared, or the
2227			 * relevant grace period is already over, so done.
2228			 */
2229			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2230			return;
2231		}
2232		WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2233		WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
2234			     rcu_preempt_blocked_readers_cgp(rnp));
2235		WRITE_ONCE(rnp->qsmask, rnp->qsmask & ~mask);
2236		trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
2237						 mask, rnp->qsmask, rnp->level,
2238						 rnp->grplo, rnp->grphi,
2239						 !!rnp->gp_tasks);
2240		if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2241
2242			/* Other bits still set at this level, so done. */
2243			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2244			return;
2245		}
2246		rnp->completedqs = rnp->gp_seq;
2247		mask = rnp->grpmask;
2248		if (rnp->parent == NULL) {
2249
2250			/* No more levels.  Exit loop holding root lock. */
2251
2252			break;
2253		}
2254		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2255		rnp_c = rnp;
2256		rnp = rnp->parent;
2257		raw_spin_lock_irqsave_rcu_node(rnp, flags);
2258		oldmask = READ_ONCE(rnp_c->qsmask);
2259	}
2260
2261	/*
2262	 * Get here if we are the last CPU to pass through a quiescent
2263	 * state for this grace period.  Invoke rcu_report_qs_rsp()
2264	 * to clean up and start the next grace period if one is needed.
2265	 */
2266	rcu_report_qs_rsp(flags); /* releases rnp->lock. */
2267}
2268
2269/*
2270 * Record a quiescent state for all tasks that were previously queued
2271 * on the specified rcu_node structure and that were blocking the current
2272 * RCU grace period.  The caller must hold the corresponding rnp->lock with
2273 * irqs disabled, and this lock is released upon return, but irqs remain
2274 * disabled.
2275 */
2276static void __maybe_unused
2277rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
2278	__releases(rnp->lock)
2279{
2280	unsigned long gps;
2281	unsigned long mask;
2282	struct rcu_node *rnp_p;
2283
2284	raw_lockdep_assert_held_rcu_node(rnp);
2285	if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RCU)) ||
2286	    WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
2287	    rnp->qsmask != 0) {
2288		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2289		return;  /* Still need more quiescent states! */
2290	}
2291
2292	rnp->completedqs = rnp->gp_seq;
2293	rnp_p = rnp->parent;
2294	if (rnp_p == NULL) {
2295		/*
2296		 * Only one rcu_node structure in the tree, so don't
2297		 * try to report up to its nonexistent parent!
2298		 */
2299		rcu_report_qs_rsp(flags);
2300		return;
2301	}
2302
2303	/* Report up the rest of the hierarchy, tracking current ->gp_seq. */
2304	gps = rnp->gp_seq;
2305	mask = rnp->grpmask;
2306	raw_spin_unlock_rcu_node(rnp);	/* irqs remain disabled. */
2307	raw_spin_lock_rcu_node(rnp_p);	/* irqs already disabled. */
2308	rcu_report_qs_rnp(mask, rnp_p, gps, flags);
2309}
2310
2311/*
2312 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2313 * structure.  This must be called from the specified CPU.
2314 */
2315static void
2316rcu_report_qs_rdp(struct rcu_data *rdp)
2317{
2318	unsigned long flags;
2319	unsigned long mask;
2320	bool needwake = false;
2321	const bool offloaded = rcu_rdp_is_offloaded(rdp);
2322	struct rcu_node *rnp;
2323
2324	WARN_ON_ONCE(rdp->cpu != smp_processor_id());
2325	rnp = rdp->mynode;
2326	raw_spin_lock_irqsave_rcu_node(rnp, flags);
2327	if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
2328	    rdp->gpwrap) {
2329
2330		/*
2331		 * The grace period in which this quiescent state was
2332		 * recorded has ended, so don't report it upwards.
2333		 * We will instead need a new quiescent state that lies
2334		 * within the current grace period.
2335		 */
2336		rdp->cpu_no_qs.b.norm = true;	/* need qs for new gp. */
2337		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2338		return;
2339	}
2340	mask = rdp->grpmask;
2341	rdp->core_needs_qs = false;
2342	if ((rnp->qsmask & mask) == 0) {
2343		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2344	} else {
2345		/*
2346		 * This GP can't end until cpu checks in, so all of our
2347		 * callbacks can be processed during the next GP.
 
 
2348		 */
2349		if (!offloaded)
2350			needwake = rcu_accelerate_cbs(rnp, rdp);
 
 
 
 
 
 
 
 
 
 
 
 
2351
2352		rcu_disable_urgency_upon_qs(rdp);
2353		rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2354		/* ^^^ Released rnp->lock */
2355		if (needwake)
2356			rcu_gp_kthread_wake();
 
 
 
 
2357	}
2358}
2359
2360/*
2361 * Check to see if there is a new grace period of which this CPU
2362 * is not yet aware, and if so, set up local rcu_data state for it.
2363 * Otherwise, see if this CPU has just passed through its first
2364 * quiescent state for this grace period, and record that fact if so.
2365 */
2366static void
2367rcu_check_quiescent_state(struct rcu_data *rdp)
2368{
2369	/* Check for grace-period ends and beginnings. */
2370	note_gp_changes(rdp);
2371
2372	/*
2373	 * Does this CPU still need to do its part for current grace period?
2374	 * If no, return and let the other CPUs do their part as well.
2375	 */
2376	if (!rdp->core_needs_qs)
2377		return;
2378
2379	/*
2380	 * Was there a quiescent state since the beginning of the grace
2381	 * period? If no, then exit and wait for the next call.
2382	 */
2383	if (rdp->cpu_no_qs.b.norm)
2384		return;
2385
2386	/*
2387	 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2388	 * judge of that).
2389	 */
2390	rcu_report_qs_rdp(rdp);
2391}
2392
2393/*
2394 * Near the end of the offline process.  Trace the fact that this CPU
2395 * is going offline.
2396 */
2397int rcutree_dying_cpu(unsigned int cpu)
2398{
2399	bool blkd;
2400	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
2401	struct rcu_node *rnp = rdp->mynode;
2402
2403	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2404		return 0;
2405
2406	blkd = !!(rnp->qsmask & rdp->grpmask);
2407	trace_rcu_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
2408			       blkd ? TPS("cpuofl-bgp") : TPS("cpuofl"));
2409	return 0;
2410}
2411
2412/*
2413 * All CPUs for the specified rcu_node structure have gone offline,
2414 * and all tasks that were preempted within an RCU read-side critical
2415 * section while running on one of those CPUs have since exited their RCU
2416 * read-side critical section.  Some other CPU is reporting this fact with
2417 * the specified rcu_node structure's ->lock held and interrupts disabled.
2418 * This function therefore goes up the tree of rcu_node structures,
2419 * clearing the corresponding bits in the ->qsmaskinit fields.  Note that
2420 * the leaf rcu_node structure's ->qsmaskinit field has already been
2421 * updated.
2422 *
2423 * This function does check that the specified rcu_node structure has
2424 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2425 * prematurely.  That said, invoking it after the fact will cost you
2426 * a needless lock acquisition.  So once it has done its work, don't
2427 * invoke it again.
2428 */
2429static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2430{
2431	long mask;
2432	struct rcu_node *rnp = rnp_leaf;
2433
2434	raw_lockdep_assert_held_rcu_node(rnp_leaf);
2435	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2436	    WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
2437	    WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
2438		return;
2439	for (;;) {
2440		mask = rnp->grpmask;
2441		rnp = rnp->parent;
2442		if (!rnp)
2443			break;
2444		raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2445		rnp->qsmaskinit &= ~mask;
2446		/* Between grace periods, so better already be zero! */
2447		WARN_ON_ONCE(rnp->qsmask);
2448		if (rnp->qsmaskinit) {
2449			raw_spin_unlock_rcu_node(rnp);
2450			/* irqs remain disabled. */
2451			return;
2452		}
2453		raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2454	}
2455}
2456
2457/*
2458 * The CPU has been completely removed, and some other CPU is reporting
2459 * this fact from process context.  Do the remainder of the cleanup.
2460 * There can only be one CPU hotplug operation at a time, so no need for
2461 * explicit locking.
2462 */
2463int rcutree_dead_cpu(unsigned int cpu)
2464{
2465	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2466	struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
2467
2468	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2469		return 0;
2470
2471	WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus - 1);
2472	/* Adjust any no-longer-needed kthreads. */
2473	rcu_boost_kthread_setaffinity(rnp, -1);
2474	/* Do any needed no-CB deferred wakeups from this CPU. */
2475	do_nocb_deferred_wakeup(per_cpu_ptr(&rcu_data, cpu));
2476
2477	// Stop-machine done, so allow nohz_full to disable tick.
2478	tick_dep_clear(TICK_DEP_BIT_RCU);
2479	return 0;
2480}
2481
2482/*
2483 * Invoke any RCU callbacks that have made it to the end of their grace
2484 * period.  Throttle as specified by rdp->blimit.
2485 */
2486static void rcu_do_batch(struct rcu_data *rdp)
2487{
 
 
2488	int div;
2489	bool __maybe_unused empty;
2490	unsigned long flags;
2491	const bool offloaded = rcu_rdp_is_offloaded(rdp);
 
 
 
2492	struct rcu_head *rhp;
2493	struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2494	long bl, count = 0;
2495	long pending, tlimit = 0;
2496
2497	/* If no callbacks are ready, just return. */
2498	if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2499		trace_rcu_batch_start(rcu_state.name,
2500				      rcu_segcblist_n_cbs(&rdp->cblist), 0);
2501		trace_rcu_batch_end(rcu_state.name, 0,
2502				    !rcu_segcblist_empty(&rdp->cblist),
2503				    need_resched(), is_idle_task(current),
2504				    rcu_is_callbacks_kthread());
2505		return;
2506	}
2507
2508	/*
2509	 * Extract the list of ready callbacks, disabling to prevent
2510	 * races with call_rcu() from interrupt handlers.  Leave the
2511	 * callback counts, as rcu_barrier() needs to be conservative.
2512	 */
2513	local_irq_save(flags);
2514	rcu_nocb_lock(rdp);
2515	WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2516	pending = rcu_segcblist_n_cbs(&rdp->cblist);
2517	div = READ_ONCE(rcu_divisor);
2518	div = div < 0 ? 7 : div > sizeof(long) * 8 - 2 ? sizeof(long) * 8 - 2 : div;
2519	bl = max(rdp->blimit, pending >> div);
2520	if (unlikely(bl > 100)) {
 
 
2521		long rrn = READ_ONCE(rcu_resched_ns);
2522
2523		rrn = rrn < NSEC_PER_MSEC ? NSEC_PER_MSEC : rrn > NSEC_PER_SEC ? NSEC_PER_SEC : rrn;
2524		tlimit = local_clock() + rrn;
 
 
2525	}
2526	trace_rcu_batch_start(rcu_state.name,
2527			      rcu_segcblist_n_cbs(&rdp->cblist), bl);
2528	rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2529	if (offloaded)
2530		rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2531
2532	trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbDequeued"));
2533	rcu_nocb_unlock_irqrestore(rdp, flags);
2534
2535	/* Invoke callbacks. */
2536	tick_dep_set_task(current, TICK_DEP_BIT_RCU);
2537	rhp = rcu_cblist_dequeue(&rcl);
2538
2539	for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2540		rcu_callback_t f;
2541
2542		count++;
2543		debug_rcu_head_unqueue(rhp);
2544
2545		rcu_lock_acquire(&rcu_callback_map);
2546		trace_rcu_invoke_callback(rcu_state.name, rhp);
2547
2548		f = rhp->func;
 
2549		WRITE_ONCE(rhp->func, (rcu_callback_t)0L);
2550		f(rhp);
2551
2552		rcu_lock_release(&rcu_callback_map);
2553
2554		/*
2555		 * Stop only if limit reached and CPU has something to do.
2556		 */
2557		if (count >= bl && !offloaded &&
2558		    (need_resched() ||
2559		     (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2560			break;
2561		if (unlikely(tlimit)) {
2562			/* only call local_clock() every 32 callbacks */
2563			if (likely((count & 31) || local_clock() < tlimit))
2564				continue;
2565			/* Exceeded the time limit, so leave. */
2566			break;
2567		}
2568		if (!in_serving_softirq()) {
2569			local_bh_enable();
2570			lockdep_assert_irqs_enabled();
2571			cond_resched_tasks_rcu_qs();
2572			lockdep_assert_irqs_enabled();
2573			local_bh_disable();
 
 
 
 
 
 
 
2574		}
2575	}
2576
2577	local_irq_save(flags);
2578	rcu_nocb_lock(rdp);
2579	rdp->n_cbs_invoked += count;
2580	trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2581			    is_idle_task(current), rcu_is_callbacks_kthread());
2582
2583	/* Update counts and requeue any remaining callbacks. */
2584	rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2585	rcu_segcblist_add_len(&rdp->cblist, -count);
2586
2587	/* Reinstate batch limit if we have worked down the excess. */
2588	count = rcu_segcblist_n_cbs(&rdp->cblist);
2589	if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
2590		rdp->blimit = blimit;
2591
2592	/* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2593	if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2594		rdp->qlen_last_fqs_check = 0;
2595		rdp->n_force_qs_snap = rcu_state.n_force_qs;
2596	} else if (count < rdp->qlen_last_fqs_check - qhimark)
2597		rdp->qlen_last_fqs_check = count;
2598
2599	/*
2600	 * The following usually indicates a double call_rcu().  To track
2601	 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2602	 */
2603	empty = rcu_segcblist_empty(&rdp->cblist);
2604	WARN_ON_ONCE(count == 0 && !empty);
2605	WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2606		     count != 0 && empty);
2607	WARN_ON_ONCE(count == 0 && rcu_segcblist_n_segment_cbs(&rdp->cblist) != 0);
2608	WARN_ON_ONCE(!empty && rcu_segcblist_n_segment_cbs(&rdp->cblist) == 0);
2609
2610	rcu_nocb_unlock_irqrestore(rdp, flags);
2611
2612	/* Re-invoke RCU core processing if there are callbacks remaining. */
2613	if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist))
2614		invoke_rcu_core();
2615	tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
2616}
2617
2618/*
2619 * This function is invoked from each scheduling-clock interrupt,
2620 * and checks to see if this CPU is in a non-context-switch quiescent
2621 * state, for example, user mode or idle loop.  It also schedules RCU
2622 * core processing.  If the current grace period has gone on too long,
2623 * it will ask the scheduler to manufacture a context switch for the sole
2624 * purpose of providing the needed quiescent state.
2625 */
2626void rcu_sched_clock_irq(int user)
2627{
 
 
 
 
 
 
 
2628	trace_rcu_utilization(TPS("Start scheduler-tick"));
2629	lockdep_assert_irqs_disabled();
2630	raw_cpu_inc(rcu_data.ticks_this_gp);
2631	/* The load-acquire pairs with the store-release setting to true. */
2632	if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
2633		/* Idle and userspace execution already are quiescent states. */
2634		if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2635			set_tsk_need_resched(current);
2636			set_preempt_need_resched();
2637		}
2638		__this_cpu_write(rcu_data.rcu_urgent_qs, false);
2639	}
2640	rcu_flavor_sched_clock_irq(user);
2641	if (rcu_pending(user))
2642		invoke_rcu_core();
 
 
2643	lockdep_assert_irqs_disabled();
2644
2645	trace_rcu_utilization(TPS("End scheduler-tick"));
2646}
2647
2648/*
2649 * Scan the leaf rcu_node structures.  For each structure on which all
2650 * CPUs have reported a quiescent state and on which there are tasks
2651 * blocking the current grace period, initiate RCU priority boosting.
2652 * Otherwise, invoke the specified function to check dyntick state for
2653 * each CPU that has not yet reported a quiescent state.
2654 */
2655static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
2656{
2657	int cpu;
2658	unsigned long flags;
2659	unsigned long mask;
2660	struct rcu_data *rdp;
2661	struct rcu_node *rnp;
2662
2663	rcu_state.cbovld = rcu_state.cbovldnext;
2664	rcu_state.cbovldnext = false;
2665	rcu_for_each_leaf_node(rnp) {
 
 
 
2666		cond_resched_tasks_rcu_qs();
2667		mask = 0;
2668		raw_spin_lock_irqsave_rcu_node(rnp, flags);
2669		rcu_state.cbovldnext |= !!rnp->cbovldmask;
2670		if (rnp->qsmask == 0) {
2671			if (rcu_preempt_blocked_readers_cgp(rnp)) {
2672				/*
2673				 * No point in scanning bits because they
2674				 * are all zero.  But we might need to
2675				 * priority-boost blocked readers.
2676				 */
2677				rcu_initiate_boost(rnp, flags);
2678				/* rcu_initiate_boost() releases rnp->lock */
2679				continue;
2680			}
2681			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2682			continue;
2683		}
2684		for_each_leaf_node_cpu_mask(rnp, cpu, rnp->qsmask) {
 
 
 
2685			rdp = per_cpu_ptr(&rcu_data, cpu);
2686			if (f(rdp)) {
 
2687				mask |= rdp->grpmask;
2688				rcu_disable_urgency_upon_qs(rdp);
2689			}
 
 
2690		}
2691		if (mask != 0) {
2692			/* Idle/offline CPUs, report (releases rnp->lock). */
2693			rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2694		} else {
2695			/* Nothing to do here, so just drop the lock. */
2696			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2697		}
 
 
 
2698	}
2699}
2700
2701/*
2702 * Force quiescent states on reluctant CPUs, and also detect which
2703 * CPUs are in dyntick-idle mode.
2704 */
2705void rcu_force_quiescent_state(void)
2706{
2707	unsigned long flags;
2708	bool ret;
2709	struct rcu_node *rnp;
2710	struct rcu_node *rnp_old = NULL;
2711
 
 
2712	/* Funnel through hierarchy to reduce memory contention. */
2713	rnp = __this_cpu_read(rcu_data.mynode);
2714	for (; rnp != NULL; rnp = rnp->parent) {
2715		ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2716		       !raw_spin_trylock(&rnp->fqslock);
2717		if (rnp_old != NULL)
2718			raw_spin_unlock(&rnp_old->fqslock);
2719		if (ret)
2720			return;
2721		rnp_old = rnp;
2722	}
2723	/* rnp_old == rcu_get_root(), rnp == NULL. */
2724
2725	/* Reached the root of the rcu_node tree, acquire lock. */
2726	raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2727	raw_spin_unlock(&rnp_old->fqslock);
2728	if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2729		raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2730		return;  /* Someone beat us to it. */
2731	}
2732	WRITE_ONCE(rcu_state.gp_flags,
2733		   READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2734	raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2735	rcu_gp_kthread_wake();
2736}
2737EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
2738
2739// Workqueue handler for an RCU reader for kernels enforcing struct RCU
2740// grace periods.
2741static void strict_work_handler(struct work_struct *work)
2742{
2743	rcu_read_lock();
2744	rcu_read_unlock();
2745}
2746
2747/* Perform RCU core processing work for the current CPU.  */
2748static __latent_entropy void rcu_core(void)
2749{
2750	unsigned long flags;
2751	struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2752	struct rcu_node *rnp = rdp->mynode;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2753	const bool do_batch = !rcu_segcblist_completely_offloaded(&rdp->cblist);
2754
2755	if (cpu_is_offline(smp_processor_id()))
2756		return;
2757	trace_rcu_utilization(TPS("Start RCU core"));
2758	WARN_ON_ONCE(!rdp->beenonline);
2759
2760	/* Report any deferred quiescent states if preemption enabled. */
2761	if (!(preempt_count() & PREEMPT_MASK)) {
2762		rcu_preempt_deferred_qs(current);
2763	} else if (rcu_preempt_need_deferred_qs(current)) {
2764		set_tsk_need_resched(current);
2765		set_preempt_need_resched();
2766	}
2767
2768	/* Update RCU state based on any recent quiescent states. */
2769	rcu_check_quiescent_state(rdp);
2770
2771	/* No grace period and unregistered callbacks? */
2772	if (!rcu_gp_in_progress() &&
2773	    rcu_segcblist_is_enabled(&rdp->cblist) && do_batch) {
2774		rcu_nocb_lock_irqsave(rdp, flags);
2775		if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2776			rcu_accelerate_cbs_unlocked(rnp, rdp);
2777		rcu_nocb_unlock_irqrestore(rdp, flags);
2778	}
2779
2780	rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
2781
2782	/* If there are callbacks ready, invoke them. */
2783	if (do_batch && rcu_segcblist_ready_cbs(&rdp->cblist) &&
2784	    likely(READ_ONCE(rcu_scheduler_fully_active)))
2785		rcu_do_batch(rdp);
 
 
 
 
2786
2787	/* Do any needed deferred wakeups of rcuo kthreads. */
2788	do_nocb_deferred_wakeup(rdp);
2789	trace_rcu_utilization(TPS("End RCU core"));
2790
2791	// If strict GPs, schedule an RCU reader in a clean environment.
2792	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
2793		queue_work_on(rdp->cpu, rcu_gp_wq, &rdp->strict_work);
2794}
2795
2796static void rcu_core_si(struct softirq_action *h)
2797{
2798	rcu_core();
2799}
2800
2801static void rcu_wake_cond(struct task_struct *t, int status)
2802{
2803	/*
2804	 * If the thread is yielding, only wake it when this
2805	 * is invoked from idle
2806	 */
2807	if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
2808		wake_up_process(t);
2809}
2810
2811static void invoke_rcu_core_kthread(void)
2812{
2813	struct task_struct *t;
2814	unsigned long flags;
2815
2816	local_irq_save(flags);
2817	__this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
2818	t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
2819	if (t != NULL && t != current)
2820		rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
2821	local_irq_restore(flags);
2822}
2823
2824/*
2825 * Wake up this CPU's rcuc kthread to do RCU core processing.
2826 */
2827static void invoke_rcu_core(void)
2828{
2829	if (!cpu_online(smp_processor_id()))
2830		return;
2831	if (use_softirq)
2832		raise_softirq(RCU_SOFTIRQ);
2833	else
2834		invoke_rcu_core_kthread();
2835}
2836
2837static void rcu_cpu_kthread_park(unsigned int cpu)
2838{
2839	per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
2840}
2841
2842static int rcu_cpu_kthread_should_run(unsigned int cpu)
2843{
2844	return __this_cpu_read(rcu_data.rcu_cpu_has_work);
2845}
2846
2847/*
2848 * Per-CPU kernel thread that invokes RCU callbacks.  This replaces
2849 * the RCU softirq used in configurations of RCU that do not support RCU
2850 * priority boosting.
2851 */
2852static void rcu_cpu_kthread(unsigned int cpu)
2853{
2854	unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
2855	char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
 
2856	int spincnt;
2857
2858	trace_rcu_utilization(TPS("Start CPU kthread@rcu_run"));
2859	for (spincnt = 0; spincnt < 10; spincnt++) {
 
2860		local_bh_disable();
2861		*statusp = RCU_KTHREAD_RUNNING;
2862		local_irq_disable();
2863		work = *workp;
2864		*workp = 0;
2865		local_irq_enable();
2866		if (work)
2867			rcu_core();
2868		local_bh_enable();
2869		if (*workp == 0) {
2870			trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2871			*statusp = RCU_KTHREAD_WAITING;
2872			return;
2873		}
2874	}
2875	*statusp = RCU_KTHREAD_YIELDING;
2876	trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2877	schedule_timeout_idle(2);
2878	trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2879	*statusp = RCU_KTHREAD_WAITING;
 
2880}
2881
2882static struct smp_hotplug_thread rcu_cpu_thread_spec = {
2883	.store			= &rcu_data.rcu_cpu_kthread_task,
2884	.thread_should_run	= rcu_cpu_kthread_should_run,
2885	.thread_fn		= rcu_cpu_kthread,
2886	.thread_comm		= "rcuc/%u",
2887	.setup			= rcu_cpu_kthread_setup,
2888	.park			= rcu_cpu_kthread_park,
2889};
2890
2891/*
2892 * Spawn per-CPU RCU core processing kthreads.
2893 */
2894static int __init rcu_spawn_core_kthreads(void)
2895{
2896	int cpu;
2897
2898	for_each_possible_cpu(cpu)
2899		per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
2900	if (!IS_ENABLED(CONFIG_RCU_BOOST) && use_softirq)
2901		return 0;
2902	WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
2903		  "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
2904	return 0;
2905}
2906
2907/*
2908 * Handle any core-RCU processing required by a call_rcu() invocation.
2909 */
2910static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2911			    unsigned long flags)
2912{
2913	/*
2914	 * If called from an extended quiescent state, invoke the RCU
2915	 * core in order to force a re-evaluation of RCU's idleness.
2916	 */
2917	if (!rcu_is_watching())
2918		invoke_rcu_core();
2919
2920	/* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2921	if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2922		return;
2923
2924	/*
2925	 * Force the grace period if too many callbacks or too long waiting.
2926	 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2927	 * if some other CPU has recently done so.  Also, don't bother
2928	 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2929	 * is the only one waiting for a grace period to complete.
2930	 */
2931	if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2932		     rdp->qlen_last_fqs_check + qhimark)) {
2933
2934		/* Are we ignoring a completed grace period? */
2935		note_gp_changes(rdp);
2936
2937		/* Start a new grace period if one not already started. */
2938		if (!rcu_gp_in_progress()) {
2939			rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
2940		} else {
2941			/* Give the grace period a kick. */
2942			rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
2943			if (rcu_state.n_force_qs == rdp->n_force_qs_snap &&
2944			    rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2945				rcu_force_quiescent_state();
2946			rdp->n_force_qs_snap = rcu_state.n_force_qs;
2947			rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2948		}
2949	}
2950}
2951
2952/*
2953 * RCU callback function to leak a callback.
2954 */
2955static void rcu_leak_callback(struct rcu_head *rhp)
2956{
2957}
2958
2959/*
2960 * Check and if necessary update the leaf rcu_node structure's
2961 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2962 * number of queued RCU callbacks.  The caller must hold the leaf rcu_node
2963 * structure's ->lock.
2964 */
2965static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp)
2966{
2967	raw_lockdep_assert_held_rcu_node(rnp);
2968	if (qovld_calc <= 0)
2969		return; // Early boot and wildcard value set.
2970	if (rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc)
2971		WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask | rdp->grpmask);
2972	else
2973		WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask & ~rdp->grpmask);
2974}
2975
2976/*
2977 * Check and if necessary update the leaf rcu_node structure's
2978 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2979 * number of queued RCU callbacks.  No locks need be held, but the
2980 * caller must have disabled interrupts.
2981 *
2982 * Note that this function ignores the possibility that there are a lot
2983 * of callbacks all of which have already seen the end of their respective
2984 * grace periods.  This omission is due to the need for no-CBs CPUs to
2985 * be holding ->nocb_lock to do this check, which is too heavy for a
2986 * common-case operation.
2987 */
2988static void check_cb_ovld(struct rcu_data *rdp)
2989{
2990	struct rcu_node *const rnp = rdp->mynode;
2991
2992	if (qovld_calc <= 0 ||
2993	    ((rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) ==
2994	     !!(READ_ONCE(rnp->cbovldmask) & rdp->grpmask)))
2995		return; // Early boot wildcard value or already set correctly.
2996	raw_spin_lock_rcu_node(rnp);
2997	check_cb_ovld_locked(rdp, rnp);
2998	raw_spin_unlock_rcu_node(rnp);
2999}
3000
3001/* Helper function for call_rcu() and friends.  */
3002static void
3003__call_rcu(struct rcu_head *head, rcu_callback_t func)
3004{
3005	static atomic_t doublefrees;
3006	unsigned long flags;
 
3007	struct rcu_data *rdp;
3008	bool was_alldone;
3009
3010	/* Misaligned rcu_head! */
3011	WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
3012
3013	if (debug_rcu_head_queue(head)) {
3014		/*
3015		 * Probable double call_rcu(), so leak the callback.
3016		 * Use rcu:rcu_callback trace event to find the previous
3017		 * time callback was passed to __call_rcu().
3018		 */
3019		if (atomic_inc_return(&doublefrees) < 4) {
3020			pr_err("%s(): Double-freed CB %p->%pS()!!!  ", __func__, head, head->func);
3021			mem_dump_obj(head);
3022		}
3023		WRITE_ONCE(head->func, rcu_leak_callback);
3024		return;
3025	}
3026	head->func = func;
3027	head->next = NULL;
 
3028	local_irq_save(flags);
3029	kasan_record_aux_stack(head);
3030	rdp = this_cpu_ptr(&rcu_data);
 
3031
3032	/* Add the callback to our list. */
3033	if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
3034		// This can trigger due to call_rcu() from offline CPU:
3035		WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
3036		WARN_ON_ONCE(!rcu_is_watching());
3037		// Very early boot, before rcu_init().  Initialize if needed
3038		// and then drop through to queue the callback.
3039		if (rcu_segcblist_empty(&rdp->cblist))
3040			rcu_segcblist_init(&rdp->cblist);
3041	}
3042
3043	check_cb_ovld(rdp);
3044	if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags))
3045		return; // Enqueued onto ->nocb_bypass, so just leave.
3046	// If no-CBs CPU gets here, rcu_nocb_try_bypass() acquired ->nocb_lock.
3047	rcu_segcblist_enqueue(&rdp->cblist, head);
3048	if (__is_kvfree_rcu_offset((unsigned long)func))
3049		trace_rcu_kvfree_callback(rcu_state.name, head,
3050					 (unsigned long)func,
3051					 rcu_segcblist_n_cbs(&rdp->cblist));
3052	else
3053		trace_rcu_callback(rcu_state.name, head,
3054				   rcu_segcblist_n_cbs(&rdp->cblist));
3055
3056	trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCBQueued"));
3057
3058	/* Go handle any RCU core processing required. */
3059	if (unlikely(rcu_rdp_is_offloaded(rdp))) {
3060		__call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */
3061	} else {
3062		__call_rcu_core(rdp, head, flags);
3063		local_irq_restore(flags);
3064	}
3065}
3066
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3067/**
3068 * call_rcu() - Queue an RCU callback for invocation after a grace period.
 
 
 
 
3069 * @head: structure to be used for queueing the RCU updates.
3070 * @func: actual callback function to be invoked after the grace period
3071 *
3072 * The callback function will be invoked some time after a full grace
3073 * period elapses, in other words after all pre-existing RCU read-side
3074 * critical sections have completed.  However, the callback function
3075 * might well execute concurrently with RCU read-side critical sections
3076 * that started after call_rcu() was invoked.
3077 *
3078 * RCU read-side critical sections are delimited by rcu_read_lock()
3079 * and rcu_read_unlock(), and may be nested.  In addition, but only in
3080 * v5.0 and later, regions of code across which interrupts, preemption,
3081 * or softirqs have been disabled also serve as RCU read-side critical
3082 * sections.  This includes hardware interrupt handlers, softirq handlers,
3083 * and NMI handlers.
3084 *
3085 * Note that all CPUs must agree that the grace period extended beyond
3086 * all pre-existing RCU read-side critical section.  On systems with more
3087 * than one CPU, this means that when "func()" is invoked, each CPU is
3088 * guaranteed to have executed a full memory barrier since the end of its
3089 * last RCU read-side critical section whose beginning preceded the call
3090 * to call_rcu().  It also means that each CPU executing an RCU read-side
3091 * critical section that continues beyond the start of "func()" must have
3092 * executed a memory barrier after the call_rcu() but before the beginning
3093 * of that RCU read-side critical section.  Note that these guarantees
3094 * include CPUs that are offline, idle, or executing in user mode, as
3095 * well as CPUs that are executing in the kernel.
3096 *
3097 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
3098 * resulting RCU callback function "func()", then both CPU A and CPU B are
3099 * guaranteed to execute a full memory barrier during the time interval
3100 * between the call to call_rcu() and the invocation of "func()" -- even
3101 * if CPU A and CPU B are the same CPU (but again only if the system has
3102 * more than one CPU).
3103 *
3104 * Implementation of these memory-ordering guarantees is described here:
3105 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
3106 */
3107void call_rcu(struct rcu_head *head, rcu_callback_t func)
3108{
3109	__call_rcu(head, func);
3110}
3111EXPORT_SYMBOL_GPL(call_rcu);
3112
3113
3114/* Maximum number of jiffies to wait before draining a batch. */
3115#define KFREE_DRAIN_JIFFIES (HZ / 50)
3116#define KFREE_N_BATCHES 2
3117#define FREE_N_CHANNELS 2
3118
3119/**
3120 * struct kvfree_rcu_bulk_data - single block to store kvfree_rcu() pointers
 
 
3121 * @nr_records: Number of active pointers in the array
3122 * @next: Next bulk object in the block chain
3123 * @records: Array of the kvfree_rcu() pointers
3124 */
3125struct kvfree_rcu_bulk_data {
 
 
3126	unsigned long nr_records;
3127	struct kvfree_rcu_bulk_data *next;
3128	void *records[];
3129};
3130
3131/*
3132 * This macro defines how many entries the "records" array
3133 * will contain. It is based on the fact that the size of
3134 * kvfree_rcu_bulk_data structure becomes exactly one page.
3135 */
3136#define KVFREE_BULK_MAX_ENTR \
3137	((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *))
3138
3139/**
3140 * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
3141 * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
3142 * @head_free: List of kfree_rcu() objects waiting for a grace period
3143 * @bkvhead_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
 
3144 * @krcp: Pointer to @kfree_rcu_cpu structure
3145 */
3146
3147struct kfree_rcu_cpu_work {
3148	struct rcu_work rcu_work;
3149	struct rcu_head *head_free;
3150	struct kvfree_rcu_bulk_data *bkvhead_free[FREE_N_CHANNELS];
 
3151	struct kfree_rcu_cpu *krcp;
3152};
3153
3154/**
3155 * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
3156 * @head: List of kfree_rcu() objects not yet waiting for a grace period
3157 * @bkvhead: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period
 
3158 * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
3159 * @lock: Synchronize access to this structure
3160 * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES
3161 * @monitor_todo: Tracks whether a @monitor_work delayed work is pending
3162 * @initialized: The @rcu_work fields have been initialized
3163 * @count: Number of objects for which GP not started
 
3164 * @bkvcache:
3165 *	A simple cache list that contains objects for reuse purpose.
3166 *	In order to save some per-cpu space the list is singular.
3167 *	Even though it is lockless an access has to be protected by the
3168 *	per-cpu lock.
3169 * @page_cache_work: A work to refill the cache when it is empty
3170 * @backoff_page_cache_fill: Delay cache refills
3171 * @work_in_progress: Indicates that page_cache_work is running
3172 * @hrtimer: A hrtimer for scheduling a page_cache_work
3173 * @nr_bkv_objs: number of allocated objects at @bkvcache.
3174 *
3175 * This is a per-CPU structure.  The reason that it is not included in
3176 * the rcu_data structure is to permit this code to be extracted from
3177 * the RCU files.  Such extraction could allow further optimization of
3178 * the interactions with the slab allocators.
3179 */
3180struct kfree_rcu_cpu {
 
 
3181	struct rcu_head *head;
3182	struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS];
 
 
 
 
 
 
3183	struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
3184	raw_spinlock_t lock;
3185	struct delayed_work monitor_work;
3186	bool monitor_todo;
3187	bool initialized;
3188	int count;
3189
3190	struct delayed_work page_cache_work;
3191	atomic_t backoff_page_cache_fill;
3192	atomic_t work_in_progress;
3193	struct hrtimer hrtimer;
3194
3195	struct llist_head bkvcache;
3196	int nr_bkv_objs;
3197};
3198
3199static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc) = {
3200	.lock = __RAW_SPIN_LOCK_UNLOCKED(krc.lock),
3201};
3202
3203static __always_inline void
3204debug_rcu_bhead_unqueue(struct kvfree_rcu_bulk_data *bhead)
3205{
3206#ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
3207	int i;
3208
3209	for (i = 0; i < bhead->nr_records; i++)
3210		debug_rcu_head_unqueue((struct rcu_head *)(bhead->records[i]));
3211#endif
3212}
3213
3214static inline struct kfree_rcu_cpu *
3215krc_this_cpu_lock(unsigned long *flags)
3216{
3217	struct kfree_rcu_cpu *krcp;
3218
3219	local_irq_save(*flags);	// For safely calling this_cpu_ptr().
3220	krcp = this_cpu_ptr(&krc);
3221	raw_spin_lock(&krcp->lock);
3222
3223	return krcp;
3224}
3225
3226static inline void
3227krc_this_cpu_unlock(struct kfree_rcu_cpu *krcp, unsigned long flags)
3228{
3229	raw_spin_unlock_irqrestore(&krcp->lock, flags);
3230}
3231
3232static inline struct kvfree_rcu_bulk_data *
3233get_cached_bnode(struct kfree_rcu_cpu *krcp)
3234{
3235	if (!krcp->nr_bkv_objs)
3236		return NULL;
3237
3238	WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs - 1);
3239	return (struct kvfree_rcu_bulk_data *)
3240		llist_del_first(&krcp->bkvcache);
3241}
3242
3243static inline bool
3244put_cached_bnode(struct kfree_rcu_cpu *krcp,
3245	struct kvfree_rcu_bulk_data *bnode)
3246{
3247	// Check the limit.
3248	if (krcp->nr_bkv_objs >= rcu_min_cached_objs)
3249		return false;
3250
3251	llist_add((struct llist_node *) bnode, &krcp->bkvcache);
3252	WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs + 1);
3253	return true;
3254}
3255
3256static int
3257drain_page_cache(struct kfree_rcu_cpu *krcp)
3258{
3259	unsigned long flags;
3260	struct llist_node *page_list, *pos, *n;
3261	int freed = 0;
3262
 
 
 
3263	raw_spin_lock_irqsave(&krcp->lock, flags);
3264	page_list = llist_del_all(&krcp->bkvcache);
3265	WRITE_ONCE(krcp->nr_bkv_objs, 0);
3266	raw_spin_unlock_irqrestore(&krcp->lock, flags);
3267
3268	llist_for_each_safe(pos, n, page_list) {
3269		free_page((unsigned long)pos);
3270		freed++;
3271	}
3272
3273	return freed;
3274}
3275
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3276/*
3277 * This function is invoked in workqueue context after a grace period.
3278 * It frees all the objects queued on ->bkvhead_free or ->head_free.
3279 */
3280static void kfree_rcu_work(struct work_struct *work)
3281{
3282	unsigned long flags;
3283	struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS], *bnext;
3284	struct rcu_head *head, *next;
 
3285	struct kfree_rcu_cpu *krcp;
3286	struct kfree_rcu_cpu_work *krwp;
3287	int i, j;
 
3288
3289	krwp = container_of(to_rcu_work(work),
3290			    struct kfree_rcu_cpu_work, rcu_work);
3291	krcp = krwp->krcp;
3292
3293	raw_spin_lock_irqsave(&krcp->lock, flags);
3294	// Channels 1 and 2.
3295	for (i = 0; i < FREE_N_CHANNELS; i++) {
3296		bkvhead[i] = krwp->bkvhead_free[i];
3297		krwp->bkvhead_free[i] = NULL;
3298	}
3299
3300	// Channel 3.
3301	head = krwp->head_free;
3302	krwp->head_free = NULL;
 
3303	raw_spin_unlock_irqrestore(&krcp->lock, flags);
3304
3305	// Handle the first two channels.
3306	for (i = 0; i < FREE_N_CHANNELS; i++) {
3307		for (; bkvhead[i]; bkvhead[i] = bnext) {
3308			bnext = bkvhead[i]->next;
3309			debug_rcu_bhead_unqueue(bkvhead[i]);
3310
3311			rcu_lock_acquire(&rcu_callback_map);
3312			if (i == 0) { // kmalloc() / kfree().
3313				trace_rcu_invoke_kfree_bulk_callback(
3314					rcu_state.name, bkvhead[i]->nr_records,
3315					bkvhead[i]->records);
3316
3317				kfree_bulk(bkvhead[i]->nr_records,
3318					bkvhead[i]->records);
3319			} else { // vmalloc() / vfree().
3320				for (j = 0; j < bkvhead[i]->nr_records; j++) {
3321					trace_rcu_invoke_kvfree_callback(
3322						rcu_state.name,
3323						bkvhead[i]->records[j], 0);
3324
3325					vfree(bkvhead[i]->records[j]);
3326				}
3327			}
3328			rcu_lock_release(&rcu_callback_map);
3329
3330			raw_spin_lock_irqsave(&krcp->lock, flags);
3331			if (put_cached_bnode(krcp, bkvhead[i]))
3332				bkvhead[i] = NULL;
3333			raw_spin_unlock_irqrestore(&krcp->lock, flags);
3334
3335			if (bkvhead[i])
3336				free_page((unsigned long) bkvhead[i]);
3337
3338			cond_resched_tasks_rcu_qs();
3339		}
3340	}
3341
3342	/*
3343	 * This is used when the "bulk" path can not be used for the
3344	 * double-argument of kvfree_rcu().  This happens when the
3345	 * page-cache is empty, which means that objects are instead
3346	 * queued on a linked list through their rcu_head structures.
3347	 * This list is named "Channel 3".
3348	 */
3349	for (; head; head = next) {
3350		unsigned long offset = (unsigned long)head->func;
3351		void *ptr = (void *)head - offset;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3352
3353		next = head->next;
3354		debug_rcu_head_unqueue((struct rcu_head *)ptr);
3355		rcu_lock_acquire(&rcu_callback_map);
3356		trace_rcu_invoke_kvfree_callback(rcu_state.name, head, offset);
3357
3358		if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset)))
3359			kvfree(ptr);
 
 
 
 
3360
3361		rcu_lock_release(&rcu_callback_map);
3362		cond_resched_tasks_rcu_qs();
 
3363	}
 
 
 
3364}
3365
3366/*
3367 * This function is invoked after the KFREE_DRAIN_JIFFIES timeout.
3368 */
3369static void kfree_rcu_monitor(struct work_struct *work)
3370{
3371	struct kfree_rcu_cpu *krcp = container_of(work,
3372		struct kfree_rcu_cpu, monitor_work.work);
3373	unsigned long flags;
3374	int i, j;
3375
 
 
 
3376	raw_spin_lock_irqsave(&krcp->lock, flags);
3377
3378	// Attempt to start a new batch.
3379	for (i = 0; i < KFREE_N_BATCHES; i++) {
3380		struct kfree_rcu_cpu_work *krwp = &(krcp->krw_arr[i]);
3381
3382		// Try to detach bkvhead or head and attach it over any
3383		// available corresponding free channel. It can be that
3384		// a previous RCU batch is in progress, it means that
3385		// immediately to queue another one is not possible so
3386		// in that case the monitor work is rearmed.
3387		if ((krcp->bkvhead[0] && !krwp->bkvhead_free[0]) ||
3388			(krcp->bkvhead[1] && !krwp->bkvhead_free[1]) ||
3389				(krcp->head && !krwp->head_free)) {
3390			// Channel 1 corresponds to the SLAB-pointer bulk path.
3391			// Channel 2 corresponds to vmalloc-pointer bulk path.
3392			for (j = 0; j < FREE_N_CHANNELS; j++) {
3393				if (!krwp->bkvhead_free[j]) {
3394					krwp->bkvhead_free[j] = krcp->bkvhead[j];
3395					krcp->bkvhead[j] = NULL;
 
3396				}
3397			}
3398
3399			// Channel 3 corresponds to both SLAB and vmalloc
3400			// objects queued on the linked list.
3401			if (!krwp->head_free) {
3402				krwp->head_free = krcp->head;
3403				krcp->head = NULL;
 
 
3404			}
3405
3406			WRITE_ONCE(krcp->count, 0);
3407
3408			// One work is per one batch, so there are three
3409			// "free channels", the batch can handle. It can
3410			// be that the work is in the pending state when
3411			// channels have been detached following by each
3412			// other.
3413			queue_rcu_work(system_wq, &krwp->rcu_work);
3414		}
3415	}
3416
 
 
3417	// If there is nothing to detach, it means that our job is
3418	// successfully done here. In case of having at least one
3419	// of the channels that is still busy we should rearm the
3420	// work to repeat an attempt. Because previous batches are
3421	// still in progress.
3422	if (!krcp->bkvhead[0] && !krcp->bkvhead[1] && !krcp->head)
3423		krcp->monitor_todo = false;
3424	else
3425		schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES);
3426
3427	raw_spin_unlock_irqrestore(&krcp->lock, flags);
3428}
3429
3430static enum hrtimer_restart
3431schedule_page_work_fn(struct hrtimer *t)
3432{
3433	struct kfree_rcu_cpu *krcp =
3434		container_of(t, struct kfree_rcu_cpu, hrtimer);
3435
3436	queue_delayed_work(system_highpri_wq, &krcp->page_cache_work, 0);
3437	return HRTIMER_NORESTART;
3438}
3439
3440static void fill_page_cache_func(struct work_struct *work)
3441{
3442	struct kvfree_rcu_bulk_data *bnode;
3443	struct kfree_rcu_cpu *krcp =
3444		container_of(work, struct kfree_rcu_cpu,
3445			page_cache_work.work);
3446	unsigned long flags;
3447	int nr_pages;
3448	bool pushed;
3449	int i;
3450
3451	nr_pages = atomic_read(&krcp->backoff_page_cache_fill) ?
3452		1 : rcu_min_cached_objs;
3453
3454	for (i = 0; i < nr_pages; i++) {
3455		bnode = (struct kvfree_rcu_bulk_data *)
3456			__get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3457
3458		if (bnode) {
3459			raw_spin_lock_irqsave(&krcp->lock, flags);
3460			pushed = put_cached_bnode(krcp, bnode);
3461			raw_spin_unlock_irqrestore(&krcp->lock, flags);
 
 
3462
3463			if (!pushed) {
3464				free_page((unsigned long) bnode);
3465				break;
3466			}
3467		}
3468	}
3469
3470	atomic_set(&krcp->work_in_progress, 0);
3471	atomic_set(&krcp->backoff_page_cache_fill, 0);
3472}
3473
3474static void
3475run_page_cache_worker(struct kfree_rcu_cpu *krcp)
3476{
 
 
 
 
3477	if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
3478			!atomic_xchg(&krcp->work_in_progress, 1)) {
3479		if (atomic_read(&krcp->backoff_page_cache_fill)) {
3480			queue_delayed_work(system_wq,
3481				&krcp->page_cache_work,
3482					msecs_to_jiffies(rcu_delay_page_cache_fill_msec));
3483		} else {
3484			hrtimer_init(&krcp->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3485			krcp->hrtimer.function = schedule_page_work_fn;
3486			hrtimer_start(&krcp->hrtimer, 0, HRTIMER_MODE_REL);
3487		}
3488	}
3489}
3490
3491// Record ptr in a page managed by krcp, with the pre-krc_this_cpu_lock()
3492// state specified by flags.  If can_alloc is true, the caller must
3493// be schedulable and not be holding any locks or mutexes that might be
3494// acquired by the memory allocator or anything that it might invoke.
3495// Returns true if ptr was successfully recorded, else the caller must
3496// use a fallback.
3497static inline bool
3498add_ptr_to_bulk_krc_lock(struct kfree_rcu_cpu **krcp,
3499	unsigned long *flags, void *ptr, bool can_alloc)
3500{
3501	struct kvfree_rcu_bulk_data *bnode;
3502	int idx;
3503
3504	*krcp = krc_this_cpu_lock(flags);
3505	if (unlikely(!(*krcp)->initialized))
3506		return false;
3507
3508	idx = !!is_vmalloc_addr(ptr);
 
 
3509
3510	/* Check if a new block is required. */
3511	if (!(*krcp)->bkvhead[idx] ||
3512			(*krcp)->bkvhead[idx]->nr_records == KVFREE_BULK_MAX_ENTR) {
3513		bnode = get_cached_bnode(*krcp);
3514		if (!bnode && can_alloc) {
3515			krc_this_cpu_unlock(*krcp, *flags);
3516
3517			// __GFP_NORETRY - allows a light-weight direct reclaim
3518			// what is OK from minimizing of fallback hitting point of
3519			// view. Apart of that it forbids any OOM invoking what is
3520			// also beneficial since we are about to release memory soon.
3521			//
3522			// __GFP_NOMEMALLOC - prevents from consuming of all the
3523			// memory reserves. Please note we have a fallback path.
3524			//
3525			// __GFP_NOWARN - it is supposed that an allocation can
3526			// be failed under low memory or high memory pressure
3527			// scenarios.
3528			bnode = (struct kvfree_rcu_bulk_data *)
3529				__get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3530			*krcp = krc_this_cpu_lock(flags);
3531		}
3532
3533		if (!bnode)
3534			return false;
3535
3536		/* Initialize the new block. */
3537		bnode->nr_records = 0;
3538		bnode->next = (*krcp)->bkvhead[idx];
3539
3540		/* Attach it to the head. */
3541		(*krcp)->bkvhead[idx] = bnode;
3542	}
3543
3544	/* Finally insert. */
3545	(*krcp)->bkvhead[idx]->records
3546		[(*krcp)->bkvhead[idx]->nr_records++] = ptr;
 
3547
3548	return true;
3549}
3550
3551/*
3552 * Queue a request for lazy invocation of the appropriate free routine
3553 * after a grace period.  Please note that three paths are maintained,
3554 * two for the common case using arrays of pointers and a third one that
3555 * is used only when the main paths cannot be used, for example, due to
3556 * memory pressure.
3557 *
3558 * Each kvfree_call_rcu() request is added to a batch. The batch will be drained
3559 * every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch will
3560 * be free'd in workqueue context. This allows us to: batch requests together to
3561 * reduce the number of grace periods during heavy kfree_rcu()/kvfree_rcu() load.
3562 */
3563void kvfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
3564{
3565	unsigned long flags;
3566	struct kfree_rcu_cpu *krcp;
3567	bool success;
3568	void *ptr;
3569
3570	if (head) {
3571		ptr = (void *) head - (unsigned long) func;
3572	} else {
3573		/*
3574		 * Please note there is a limitation for the head-less
3575		 * variant, that is why there is a clear rule for such
3576		 * objects: it can be used from might_sleep() context
3577		 * only. For other places please embed an rcu_head to
3578		 * your data.
3579		 */
3580		might_sleep();
3581		ptr = (unsigned long *) func;
3582	}
3583
3584	// Queue the object but don't yet schedule the batch.
3585	if (debug_rcu_head_queue(ptr)) {
3586		// Probable double kfree_rcu(), just leak.
3587		WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n",
3588			  __func__, head);
3589
3590		// Mark as success and leave.
3591		return;
3592	}
3593
3594	kasan_record_aux_stack(ptr);
3595	success = add_ptr_to_bulk_krc_lock(&krcp, &flags, ptr, !head);
3596	if (!success) {
3597		run_page_cache_worker(krcp);
3598
3599		if (head == NULL)
3600			// Inline if kvfree_rcu(one_arg) call.
3601			goto unlock_return;
3602
3603		head->func = func;
3604		head->next = krcp->head;
3605		krcp->head = head;
 
 
 
 
3606		success = true;
3607	}
3608
3609	WRITE_ONCE(krcp->count, krcp->count + 1);
 
 
 
 
 
 
3610
3611	// Set timer to drain after KFREE_DRAIN_JIFFIES.
3612	if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
3613	    !krcp->monitor_todo) {
3614		krcp->monitor_todo = true;
3615		schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES);
3616	}
3617
3618unlock_return:
3619	krc_this_cpu_unlock(krcp, flags);
3620
3621	/*
3622	 * Inline kvfree() after synchronize_rcu(). We can do
3623	 * it from might_sleep() context only, so the current
3624	 * CPU can pass the QS state.
3625	 */
3626	if (!success) {
3627		debug_rcu_head_unqueue((struct rcu_head *) ptr);
3628		synchronize_rcu();
3629		kvfree(ptr);
3630	}
3631}
3632EXPORT_SYMBOL_GPL(kvfree_call_rcu);
3633
3634static unsigned long
3635kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
3636{
3637	int cpu;
3638	unsigned long count = 0;
3639
3640	/* Snapshot count of all CPUs */
3641	for_each_possible_cpu(cpu) {
3642		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3643
3644		count += READ_ONCE(krcp->count);
3645		count += READ_ONCE(krcp->nr_bkv_objs);
3646		atomic_set(&krcp->backoff_page_cache_fill, 1);
3647	}
3648
3649	return count;
3650}
3651
3652static unsigned long
3653kfree_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
3654{
3655	int cpu, freed = 0;
3656
3657	for_each_possible_cpu(cpu) {
3658		int count;
3659		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3660
3661		count = krcp->count;
3662		count += drain_page_cache(krcp);
3663		kfree_rcu_monitor(&krcp->monitor_work.work);
3664
3665		sc->nr_to_scan -= count;
3666		freed += count;
3667
3668		if (sc->nr_to_scan <= 0)
3669			break;
3670	}
3671
3672	return freed == 0 ? SHRINK_STOP : freed;
3673}
3674
3675static struct shrinker kfree_rcu_shrinker = {
3676	.count_objects = kfree_rcu_shrink_count,
3677	.scan_objects = kfree_rcu_shrink_scan,
3678	.batch = 0,
3679	.seeks = DEFAULT_SEEKS,
3680};
3681
3682void __init kfree_rcu_scheduler_running(void)
3683{
3684	int cpu;
3685	unsigned long flags;
3686
3687	for_each_possible_cpu(cpu) {
3688		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3689
3690		raw_spin_lock_irqsave(&krcp->lock, flags);
3691		if ((!krcp->bkvhead[0] && !krcp->bkvhead[1] && !krcp->head) ||
3692				krcp->monitor_todo) {
3693			raw_spin_unlock_irqrestore(&krcp->lock, flags);
3694			continue;
3695		}
3696		krcp->monitor_todo = true;
3697		schedule_delayed_work_on(cpu, &krcp->monitor_work,
3698					 KFREE_DRAIN_JIFFIES);
3699		raw_spin_unlock_irqrestore(&krcp->lock, flags);
3700	}
3701}
3702
3703/*
3704 * During early boot, any blocking grace-period wait automatically
3705 * implies a grace period.  Later on, this is never the case for PREEMPTION.
3706 *
3707 * However, because a context switch is a grace period for !PREEMPTION, any
3708 * blocking grace-period wait automatically implies a grace period if
3709 * there is only one CPU online at any point time during execution of
3710 * either synchronize_rcu() or synchronize_rcu_expedited().  It is OK to
3711 * occasionally incorrectly indicate that there are multiple CPUs online
3712 * when there was in fact only one the whole time, as this just adds some
3713 * overhead: RCU still operates correctly.
3714 */
3715static int rcu_blocking_is_gp(void)
3716{
3717	int ret;
3718
3719	if (IS_ENABLED(CONFIG_PREEMPTION))
3720		return rcu_scheduler_active == RCU_SCHEDULER_INACTIVE;
3721	might_sleep();  /* Check for RCU read-side critical section. */
3722	preempt_disable();
3723	/*
3724	 * If the rcu_state.n_online_cpus counter is equal to one,
3725	 * there is only one CPU, and that CPU sees all prior accesses
3726	 * made by any CPU that was online at the time of its access.
3727	 * Furthermore, if this counter is equal to one, its value cannot
3728	 * change until after the preempt_enable() below.
3729	 *
3730	 * Furthermore, if rcu_state.n_online_cpus is equal to one here,
3731	 * all later CPUs (both this one and any that come online later
3732	 * on) are guaranteed to see all accesses prior to this point
3733	 * in the code, without the need for additional memory barriers.
3734	 * Those memory barriers are provided by CPU-hotplug code.
3735	 */
3736	ret = READ_ONCE(rcu_state.n_online_cpus) <= 1;
3737	preempt_enable();
3738	return ret;
3739}
3740
3741/**
3742 * synchronize_rcu - wait until a grace period has elapsed.
3743 *
3744 * Control will return to the caller some time after a full grace
3745 * period has elapsed, in other words after all currently executing RCU
3746 * read-side critical sections have completed.  Note, however, that
3747 * upon return from synchronize_rcu(), the caller might well be executing
3748 * concurrently with new RCU read-side critical sections that began while
3749 * synchronize_rcu() was waiting.
3750 *
3751 * RCU read-side critical sections are delimited by rcu_read_lock()
3752 * and rcu_read_unlock(), and may be nested.  In addition, but only in
3753 * v5.0 and later, regions of code across which interrupts, preemption,
3754 * or softirqs have been disabled also serve as RCU read-side critical
3755 * sections.  This includes hardware interrupt handlers, softirq handlers,
3756 * and NMI handlers.
3757 *
3758 * Note that this guarantee implies further memory-ordering guarantees.
3759 * On systems with more than one CPU, when synchronize_rcu() returns,
3760 * each CPU is guaranteed to have executed a full memory barrier since
3761 * the end of its last RCU read-side critical section whose beginning
3762 * preceded the call to synchronize_rcu().  In addition, each CPU having
3763 * an RCU read-side critical section that extends beyond the return from
3764 * synchronize_rcu() is guaranteed to have executed a full memory barrier
3765 * after the beginning of synchronize_rcu() and before the beginning of
3766 * that RCU read-side critical section.  Note that these guarantees include
3767 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3768 * that are executing in the kernel.
3769 *
3770 * Furthermore, if CPU A invoked synchronize_rcu(), which returned
3771 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3772 * to have executed a full memory barrier during the execution of
3773 * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
3774 * again only if the system has more than one CPU).
3775 *
3776 * Implementation of these memory-ordering guarantees is described here:
3777 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
3778 */
3779void synchronize_rcu(void)
3780{
 
 
 
3781	RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3782			 lock_is_held(&rcu_lock_map) ||
3783			 lock_is_held(&rcu_sched_lock_map),
3784			 "Illegal synchronize_rcu() in RCU read-side critical section");
3785	if (rcu_blocking_is_gp())
3786		return;  // Context allows vacuous grace periods.
3787	if (rcu_gp_is_expedited())
3788		synchronize_rcu_expedited();
3789	else
3790		wait_rcu_gp(call_rcu);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3791}
3792EXPORT_SYMBOL_GPL(synchronize_rcu);
3793
3794/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3795 * get_state_synchronize_rcu - Snapshot current RCU state
3796 *
3797 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3798 * or poll_state_synchronize_rcu() to determine whether or not a full
3799 * grace period has elapsed in the meantime.
3800 */
3801unsigned long get_state_synchronize_rcu(void)
3802{
3803	/*
3804	 * Any prior manipulation of RCU-protected data must happen
3805	 * before the load from ->gp_seq.
3806	 */
3807	smp_mb();  /* ^^^ */
3808	return rcu_seq_snap(&rcu_state.gp_seq);
3809}
3810EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3811
3812/**
3813 * start_poll_synchronize_rcu - Snapshot and start RCU grace period
 
3814 *
3815 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3816 * or poll_state_synchronize_rcu() to determine whether or not a full
3817 * grace period has elapsed in the meantime.  If the needed grace period
3818 * is not already slated to start, notifies RCU core of the need for that
3819 * grace period.
 
 
 
3820 *
3821 * Interrupts must be enabled for the case where it is necessary to awaken
3822 * the grace-period kthread.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3823 */
3824unsigned long start_poll_synchronize_rcu(void)
3825{
3826	unsigned long flags;
3827	unsigned long gp_seq = get_state_synchronize_rcu();
3828	bool needwake;
3829	struct rcu_data *rdp;
3830	struct rcu_node *rnp;
3831
3832	lockdep_assert_irqs_enabled();
3833	local_irq_save(flags);
3834	rdp = this_cpu_ptr(&rcu_data);
3835	rnp = rdp->mynode;
3836	raw_spin_lock_rcu_node(rnp); // irqs already disabled.
3837	needwake = rcu_start_this_gp(rnp, rdp, gp_seq);
 
 
 
 
 
 
3838	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3839	if (needwake)
3840		rcu_gp_kthread_wake();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3841	return gp_seq;
3842}
3843EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu);
3844
3845/**
3846 * poll_state_synchronize_rcu - Conditionally wait for an RCU grace period
 
3847 *
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3848 * @oldstate: value from get_state_synchronize_rcu() or start_poll_synchronize_rcu()
3849 *
3850 * If a full RCU grace period has elapsed since the earlier call from
3851 * which oldstate was obtained, return @true, otherwise return @false.
3852 * If @false is returned, it is the caller's responsibility to invoke this
3853 * function later on until it does return @true.  Alternatively, the caller
3854 * can explicitly wait for a grace period, for example, by passing @oldstate
3855 * to cond_synchronize_rcu() or by directly invoking synchronize_rcu().
 
 
3856 *
3857 * Yes, this function does not take counter wrap into account.
3858 * But counter wrap is harmless.  If the counter wraps, we have waited for
3859 * more than 2 billion grace periods (and way more on a 64-bit system!).
3860 * Those needing to keep oldstate values for very long time periods
3861 * (many hours even on 32-bit systems) should check them occasionally
3862 * and either refresh them or set a flag indicating that the grace period
3863 * has completed.
 
 
 
 
 
 
 
3864 *
3865 * This function provides the same memory-ordering guarantees that
3866 * would be provided by a synchronize_rcu() that was invoked at the call
3867 * to the function that provided @oldstate, and that returned at the end
3868 * of this function.
3869 */
3870bool poll_state_synchronize_rcu(unsigned long oldstate)
3871{
3872	if (rcu_seq_done(&rcu_state.gp_seq, oldstate)) {
 
3873		smp_mb(); /* Ensure GP ends before subsequent accesses. */
3874		return true;
3875	}
3876	return false;
3877}
3878EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu);
3879
3880/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3881 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3882 *
3883 * @oldstate: value from get_state_synchronize_rcu() or start_poll_synchronize_rcu()
3884 *
3885 * If a full RCU grace period has elapsed since the earlier call to
3886 * get_state_synchronize_rcu() or start_poll_synchronize_rcu(), just return.
3887 * Otherwise, invoke synchronize_rcu() to wait for a full grace period.
3888 *
3889 * Yes, this function does not take counter wrap into account.  But
3890 * counter wrap is harmless.  If the counter wraps, we have waited for
3891 * more than 2 billion grace periods (and way more on a 64-bit system!),
3892 * so waiting for one additional grace period should be just fine.
3893 *
3894 * This function provides the same memory-ordering guarantees that
3895 * would be provided by a synchronize_rcu() that was invoked at the call
3896 * to the function that provided @oldstate, and that returned at the end
3897 * of this function.
3898 */
3899void cond_synchronize_rcu(unsigned long oldstate)
3900{
3901	if (!poll_state_synchronize_rcu(oldstate))
3902		synchronize_rcu();
3903}
3904EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3905
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3906/*
3907 * Check to see if there is any immediate RCU-related work to be done by
3908 * the current CPU, returning 1 if so and zero otherwise.  The checks are
3909 * in order of increasing expense: checks that can be carried out against
3910 * CPU-local state are performed first.  However, we must check for CPU
3911 * stalls first, else we might not get a chance.
3912 */
3913static int rcu_pending(int user)
3914{
3915	bool gp_in_progress;
3916	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
3917	struct rcu_node *rnp = rdp->mynode;
3918
3919	lockdep_assert_irqs_disabled();
3920
3921	/* Check for CPU stalls, if enabled. */
3922	check_cpu_stall(rdp);
3923
3924	/* Does this CPU need a deferred NOCB wakeup? */
3925	if (rcu_nocb_need_deferred_wakeup(rdp, RCU_NOCB_WAKE))
3926		return 1;
3927
3928	/* Is this a nohz_full CPU in userspace or idle?  (Ignore RCU if so.) */
3929	if ((user || rcu_is_cpu_rrupt_from_idle()) && rcu_nohz_full_cpu())
3930		return 0;
3931
3932	/* Is the RCU core waiting for a quiescent state from this CPU? */
3933	gp_in_progress = rcu_gp_in_progress();
3934	if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm && gp_in_progress)
3935		return 1;
3936
3937	/* Does this CPU have callbacks ready to invoke? */
3938	if (!rcu_rdp_is_offloaded(rdp) &&
3939	    rcu_segcblist_ready_cbs(&rdp->cblist))
3940		return 1;
3941
3942	/* Has RCU gone idle with this CPU needing another grace period? */
3943	if (!gp_in_progress && rcu_segcblist_is_enabled(&rdp->cblist) &&
3944	    !rcu_rdp_is_offloaded(rdp) &&
3945	    !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
3946		return 1;
3947
3948	/* Have RCU grace period completed or started?  */
3949	if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
3950	    unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
3951		return 1;
3952
3953	/* nothing to do */
3954	return 0;
3955}
3956
3957/*
3958 * Helper function for rcu_barrier() tracing.  If tracing is disabled,
3959 * the compiler is expected to optimize this away.
3960 */
3961static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
3962{
3963	trace_rcu_barrier(rcu_state.name, s, cpu,
3964			  atomic_read(&rcu_state.barrier_cpu_count), done);
3965}
3966
3967/*
3968 * RCU callback function for rcu_barrier().  If we are last, wake
3969 * up the task executing rcu_barrier().
3970 *
3971 * Note that the value of rcu_state.barrier_sequence must be captured
3972 * before the atomic_dec_and_test().  Otherwise, if this CPU is not last,
3973 * other CPUs might count the value down to zero before this CPU gets
3974 * around to invoking rcu_barrier_trace(), which might result in bogus
3975 * data from the next instance of rcu_barrier().
3976 */
3977static void rcu_barrier_callback(struct rcu_head *rhp)
3978{
3979	unsigned long __maybe_unused s = rcu_state.barrier_sequence;
3980
3981	if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
3982		rcu_barrier_trace(TPS("LastCB"), -1, s);
3983		complete(&rcu_state.barrier_completion);
3984	} else {
3985		rcu_barrier_trace(TPS("CB"), -1, s);
3986	}
3987}
3988
3989/*
3990 * Called with preemption disabled, and from cross-cpu IRQ context.
3991 */
3992static void rcu_barrier_func(void *cpu_in)
3993{
3994	uintptr_t cpu = (uintptr_t)cpu_in;
3995	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 
 
3996
 
 
 
3997	rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
3998	rdp->barrier_head.func = rcu_barrier_callback;
3999	debug_rcu_head_queue(&rdp->barrier_head);
4000	rcu_nocb_lock(rdp);
4001	WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
 
 
 
 
 
 
 
4002	if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head)) {
4003		atomic_inc(&rcu_state.barrier_cpu_count);
4004	} else {
4005		debug_rcu_head_unqueue(&rdp->barrier_head);
4006		rcu_barrier_trace(TPS("IRQNQ"), -1,
4007				  rcu_state.barrier_sequence);
4008	}
4009	rcu_nocb_unlock(rdp);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4010}
4011
4012/**
4013 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
4014 *
4015 * Note that this primitive does not necessarily wait for an RCU grace period
4016 * to complete.  For example, if there are no RCU callbacks queued anywhere
4017 * in the system, then rcu_barrier() is within its rights to return
4018 * immediately, without waiting for anything, much less an RCU grace period.
4019 */
4020void rcu_barrier(void)
4021{
4022	uintptr_t cpu;
 
 
4023	struct rcu_data *rdp;
4024	unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
4025
4026	rcu_barrier_trace(TPS("Begin"), -1, s);
4027
4028	/* Take mutex to serialize concurrent rcu_barrier() requests. */
4029	mutex_lock(&rcu_state.barrier_mutex);
4030
4031	/* Did someone else do our work for us? */
4032	if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
4033		rcu_barrier_trace(TPS("EarlyExit"), -1,
4034				  rcu_state.barrier_sequence);
4035		smp_mb(); /* caller's subsequent code after above check. */
4036		mutex_unlock(&rcu_state.barrier_mutex);
4037		return;
4038	}
4039
4040	/* Mark the start of the barrier operation. */
 
4041	rcu_seq_start(&rcu_state.barrier_sequence);
 
4042	rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
4043
4044	/*
4045	 * Initialize the count to two rather than to zero in order
4046	 * to avoid a too-soon return to zero in case of an immediate
4047	 * invocation of the just-enqueued callback (or preemption of
4048	 * this task).  Exclude CPU-hotplug operations to ensure that no
4049	 * offline non-offloaded CPU has callbacks queued.
4050	 */
4051	init_completion(&rcu_state.barrier_completion);
4052	atomic_set(&rcu_state.barrier_cpu_count, 2);
4053	get_online_cpus();
4054
4055	/*
4056	 * Force each CPU with callbacks to register a new callback.
4057	 * When that callback is invoked, we will know that all of the
4058	 * corresponding CPU's preceding callbacks have been invoked.
4059	 */
4060	for_each_possible_cpu(cpu) {
4061		rdp = per_cpu_ptr(&rcu_data, cpu);
4062		if (cpu_is_offline(cpu) &&
4063		    !rcu_rdp_is_offloaded(rdp))
 
 
 
 
 
 
 
 
 
 
 
 
 
4064			continue;
4065		if (rcu_segcblist_n_cbs(&rdp->cblist) && cpu_online(cpu)) {
4066			rcu_barrier_trace(TPS("OnlineQ"), cpu,
4067					  rcu_state.barrier_sequence);
4068			smp_call_function_single(cpu, rcu_barrier_func, (void *)cpu, 1);
4069		} else if (rcu_segcblist_n_cbs(&rdp->cblist) &&
4070			   cpu_is_offline(cpu)) {
4071			rcu_barrier_trace(TPS("OfflineNoCBQ"), cpu,
4072					  rcu_state.barrier_sequence);
4073			local_irq_disable();
4074			rcu_barrier_func((void *)cpu);
4075			local_irq_enable();
4076		} else if (cpu_is_offline(cpu)) {
4077			rcu_barrier_trace(TPS("OfflineNoCBNoQ"), cpu,
4078					  rcu_state.barrier_sequence);
4079		} else {
4080			rcu_barrier_trace(TPS("OnlineNQ"), cpu,
4081					  rcu_state.barrier_sequence);
4082		}
 
 
 
 
 
 
 
4083	}
4084	put_online_cpus();
4085
4086	/*
4087	 * Now that we have an rcu_barrier_callback() callback on each
4088	 * CPU, and thus each counted, remove the initial count.
4089	 */
4090	if (atomic_sub_and_test(2, &rcu_state.barrier_cpu_count))
4091		complete(&rcu_state.barrier_completion);
4092
4093	/* Wait for all rcu_barrier_callback() callbacks to be invoked. */
4094	wait_for_completion(&rcu_state.barrier_completion);
4095
4096	/* Mark the end of the barrier operation. */
4097	rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
4098	rcu_seq_end(&rcu_state.barrier_sequence);
 
 
 
 
 
 
4099
4100	/* Other rcu_barrier() invocations can now safely proceed. */
4101	mutex_unlock(&rcu_state.barrier_mutex);
4102}
4103EXPORT_SYMBOL_GPL(rcu_barrier);
4104
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4105/*
4106 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
4107 * first CPU in a given leaf rcu_node structure coming online.  The caller
4108 * must hold the corresponding leaf rcu_node ->lock with interrupts
4109 * disabled.
4110 */
4111static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
4112{
4113	long mask;
4114	long oldmask;
4115	struct rcu_node *rnp = rnp_leaf;
4116
4117	raw_lockdep_assert_held_rcu_node(rnp_leaf);
4118	WARN_ON_ONCE(rnp->wait_blkd_tasks);
4119	for (;;) {
4120		mask = rnp->grpmask;
4121		rnp = rnp->parent;
4122		if (rnp == NULL)
4123			return;
4124		raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
4125		oldmask = rnp->qsmaskinit;
4126		rnp->qsmaskinit |= mask;
4127		raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
4128		if (oldmask)
4129			return;
4130	}
4131}
4132
4133/*
4134 * Do boot-time initialization of a CPU's per-CPU RCU data.
4135 */
4136static void __init
4137rcu_boot_init_percpu_data(int cpu)
4138{
 
4139	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4140
4141	/* Set up local state, ensuring consistent view of global state. */
4142	rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
4143	INIT_WORK(&rdp->strict_work, strict_work_handler);
4144	WARN_ON_ONCE(rdp->dynticks_nesting != 1);
4145	WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp)));
 
4146	rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
4147	rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
4148	rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
4149	rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
 
4150	rdp->cpu = cpu;
4151	rcu_boot_init_nocb_percpu_data(rdp);
4152}
4153
4154/*
4155 * Invoked early in the CPU-online process, when pretty much all services
4156 * are available.  The incoming CPU is not present.
4157 *
4158 * Initializes a CPU's per-CPU RCU data.  Note that only one online or
4159 * offline event can be happening at a given time.  Note also that we can
4160 * accept some slop in the rsp->gp_seq access due to the fact that this
4161 * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
4162 * And any offloaded callbacks are being numbered elsewhere.
4163 */
4164int rcutree_prepare_cpu(unsigned int cpu)
4165{
4166	unsigned long flags;
 
4167	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4168	struct rcu_node *rnp = rcu_get_root();
4169
4170	/* Set up local state, ensuring consistent view of global state. */
4171	raw_spin_lock_irqsave_rcu_node(rnp, flags);
4172	rdp->qlen_last_fqs_check = 0;
4173	rdp->n_force_qs_snap = rcu_state.n_force_qs;
4174	rdp->blimit = blimit;
4175	rdp->dynticks_nesting = 1;	/* CPU not up, no tearing. */
4176	rcu_dynticks_eqs_online();
4177	raw_spin_unlock_rcu_node(rnp);		/* irqs remain disabled. */
4178
4179	/*
4180	 * Only non-NOCB CPUs that didn't have early-boot callbacks need to be
4181	 * (re-)initialized.
4182	 */
4183	if (!rcu_segcblist_is_enabled(&rdp->cblist))
4184		rcu_segcblist_init(&rdp->cblist);  /* Re-enable callbacks. */
4185
4186	/*
4187	 * Add CPU to leaf rcu_node pending-online bitmask.  Any needed
4188	 * propagation up the rcu_node tree will happen at the beginning
4189	 * of the next grace period.
4190	 */
4191	rnp = rdp->mynode;
4192	raw_spin_lock_rcu_node(rnp);		/* irqs already disabled. */
4193	rdp->beenonline = true;	 /* We have now been online. */
4194	rdp->gp_seq = READ_ONCE(rnp->gp_seq);
4195	rdp->gp_seq_needed = rdp->gp_seq;
4196	rdp->cpu_no_qs.b.norm = true;
4197	rdp->core_needs_qs = false;
4198	rdp->rcu_iw_pending = false;
4199	rdp->rcu_iw = IRQ_WORK_INIT_HARD(rcu_iw_handler);
4200	rdp->rcu_iw_gp_seq = rdp->gp_seq - 1;
4201	trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
4202	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4203	rcu_spawn_one_boost_kthread(rnp);
4204	rcu_spawn_cpu_nocb_kthread(cpu);
4205	WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus + 1);
4206
4207	return 0;
4208}
4209
4210/*
4211 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
4212 */
4213static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
4214{
4215	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4216
4217	rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
4218}
4219
4220/*
 
 
 
 
 
 
 
 
 
 
4221 * Near the end of the CPU-online process.  Pretty much all services
4222 * enabled, and the CPU is now very much alive.
4223 */
4224int rcutree_online_cpu(unsigned int cpu)
4225{
4226	unsigned long flags;
4227	struct rcu_data *rdp;
4228	struct rcu_node *rnp;
4229
4230	rdp = per_cpu_ptr(&rcu_data, cpu);
4231	rnp = rdp->mynode;
4232	raw_spin_lock_irqsave_rcu_node(rnp, flags);
4233	rnp->ffmask |= rdp->grpmask;
4234	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4235	if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
4236		return 0; /* Too early in boot for scheduler work. */
4237	sync_sched_exp_online_cleanup(cpu);
4238	rcutree_affinity_setting(cpu, -1);
4239
4240	// Stop-machine done, so allow nohz_full to disable tick.
4241	tick_dep_clear(TICK_DEP_BIT_RCU);
4242	return 0;
4243}
4244
4245/*
4246 * Near the beginning of the process.  The CPU is still very much alive
4247 * with pretty much all services enabled.
4248 */
4249int rcutree_offline_cpu(unsigned int cpu)
4250{
4251	unsigned long flags;
4252	struct rcu_data *rdp;
4253	struct rcu_node *rnp;
4254
4255	rdp = per_cpu_ptr(&rcu_data, cpu);
4256	rnp = rdp->mynode;
4257	raw_spin_lock_irqsave_rcu_node(rnp, flags);
4258	rnp->ffmask &= ~rdp->grpmask;
4259	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4260
4261	rcutree_affinity_setting(cpu, cpu);
4262
4263	// nohz_full CPUs need the tick for stop-machine to work quickly
4264	tick_dep_set(TICK_DEP_BIT_RCU);
4265	return 0;
4266}
4267
4268/*
4269 * Mark the specified CPU as being online so that subsequent grace periods
4270 * (both expedited and normal) will wait on it.  Note that this means that
4271 * incoming CPUs are not allowed to use RCU read-side critical sections
4272 * until this function is called.  Failing to observe this restriction
4273 * will result in lockdep splats.
4274 *
4275 * Note that this function is special in that it is invoked directly
4276 * from the incoming CPU rather than from the cpuhp_step mechanism.
4277 * This is because this function must be invoked at a precise location.
 
 
 
4278 */
4279void rcu_cpu_starting(unsigned int cpu)
4280{
4281	unsigned long flags;
4282	unsigned long mask;
4283	struct rcu_data *rdp;
4284	struct rcu_node *rnp;
4285	bool newcpu;
4286
 
4287	rdp = per_cpu_ptr(&rcu_data, cpu);
4288	if (rdp->cpu_started)
4289		return;
4290	rdp->cpu_started = true;
4291
4292	rnp = rdp->mynode;
4293	mask = rdp->grpmask;
4294	WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
4295	WARN_ON_ONCE(!(rnp->ofl_seq & 0x1));
4296	smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
4297	raw_spin_lock_irqsave_rcu_node(rnp, flags);
4298	WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask);
 
4299	newcpu = !(rnp->expmaskinitnext & mask);
4300	rnp->expmaskinitnext |= mask;
4301	/* Allow lockless access for expedited grace periods. */
4302	smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + newcpu); /* ^^^ */
4303	ASSERT_EXCLUSIVE_WRITER(rcu_state.ncpus);
4304	rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
4305	rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4306	rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4307
4308	/* An incoming CPU should never be blocking a grace period. */
4309	if (WARN_ON_ONCE(rnp->qsmask & mask)) { /* RCU waiting on incoming CPU? */
 
 
 
 
4310		rcu_disable_urgency_upon_qs(rdp);
4311		/* Report QS -after- changing ->qsmaskinitnext! */
4312		rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4313	} else {
4314		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4315	}
4316	smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
4317	WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
4318	WARN_ON_ONCE(rnp->ofl_seq & 0x1);
4319	smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
4320}
4321
4322/*
4323 * The outgoing function has no further need of RCU, so remove it from
4324 * the rcu_node tree's ->qsmaskinitnext bit masks.
4325 *
4326 * Note that this function is special in that it is invoked directly
4327 * from the outgoing CPU rather than from the cpuhp_step mechanism.
4328 * This is because this function must be invoked at a precise location.
 
 
4329 */
4330void rcu_report_dead(unsigned int cpu)
4331{
4332	unsigned long flags;
4333	unsigned long mask;
4334	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4335	struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
4336
 
 
 
 
 
4337	// Do any dangling deferred wakeups.
4338	do_nocb_deferred_wakeup(rdp);
4339
4340	/* QS for any half-done expedited grace period. */
4341	preempt_disable();
4342	rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
4343	preempt_enable();
4344	rcu_preempt_deferred_qs(current);
4345
4346	/* Remove outgoing CPU from mask in the leaf rcu_node structure. */
4347	mask = rdp->grpmask;
4348	WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
4349	WARN_ON_ONCE(!(rnp->ofl_seq & 0x1));
4350	smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
4351	raw_spin_lock(&rcu_state.ofl_lock);
4352	raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
4353	rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4354	rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4355	if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
4356		/* Report quiescent state -before- changing ->qsmaskinitnext! */
 
4357		rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4358		raw_spin_lock_irqsave_rcu_node(rnp, flags);
4359	}
4360	WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask);
4361	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4362	raw_spin_unlock(&rcu_state.ofl_lock);
4363	smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
4364	WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
4365	WARN_ON_ONCE(rnp->ofl_seq & 0x1);
4366
4367	rdp->cpu_started = false;
4368}
4369
4370#ifdef CONFIG_HOTPLUG_CPU
4371/*
4372 * The outgoing CPU has just passed through the dying-idle state, and we
4373 * are being invoked from the CPU that was IPIed to continue the offline
4374 * operation.  Migrate the outgoing CPU's callbacks to the current CPU.
4375 */
4376void rcutree_migrate_callbacks(int cpu)
4377{
4378	unsigned long flags;
4379	struct rcu_data *my_rdp;
4380	struct rcu_node *my_rnp;
4381	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4382	bool needwake;
4383
4384	if (rcu_rdp_is_offloaded(rdp) ||
4385	    rcu_segcblist_empty(&rdp->cblist))
4386		return;  /* No callbacks to migrate. */
4387
4388	local_irq_save(flags);
 
 
4389	my_rdp = this_cpu_ptr(&rcu_data);
4390	my_rnp = my_rdp->mynode;
4391	rcu_nocb_lock(my_rdp); /* irqs already disabled. */
4392	WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies));
4393	raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
4394	/* Leverage recent GPs and set GP for new callbacks. */
4395	needwake = rcu_advance_cbs(my_rnp, rdp) ||
4396		   rcu_advance_cbs(my_rnp, my_rdp);
4397	rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
 
4398	needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
4399	rcu_segcblist_disable(&rdp->cblist);
4400	WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
4401		     !rcu_segcblist_n_cbs(&my_rdp->cblist));
4402	if (rcu_rdp_is_offloaded(my_rdp)) {
4403		raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
4404		__call_rcu_nocb_wake(my_rdp, true, flags);
4405	} else {
4406		rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
4407		raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags);
4408	}
4409	if (needwake)
4410		rcu_gp_kthread_wake();
4411	lockdep_assert_irqs_enabled();
4412	WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
4413		  !rcu_segcblist_empty(&rdp->cblist),
4414		  "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
4415		  cpu, rcu_segcblist_n_cbs(&rdp->cblist),
4416		  rcu_segcblist_first_cb(&rdp->cblist));
4417}
4418#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4419
4420/*
4421 * On non-huge systems, use expedited RCU grace periods to make suspend
4422 * and hibernation run faster.
4423 */
4424static int rcu_pm_notify(struct notifier_block *self,
4425			 unsigned long action, void *hcpu)
4426{
4427	switch (action) {
4428	case PM_HIBERNATION_PREPARE:
4429	case PM_SUSPEND_PREPARE:
 
4430		rcu_expedite_gp();
4431		break;
4432	case PM_POST_HIBERNATION:
4433	case PM_POST_SUSPEND:
4434		rcu_unexpedite_gp();
 
4435		break;
4436	default:
4437		break;
4438	}
4439	return NOTIFY_OK;
4440}
4441
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4442/*
4443 * Spawn the kthreads that handle RCU's grace periods.
4444 */
4445static int __init rcu_spawn_gp_kthread(void)
4446{
4447	unsigned long flags;
4448	int kthread_prio_in = kthread_prio;
4449	struct rcu_node *rnp;
4450	struct sched_param sp;
4451	struct task_struct *t;
4452
4453	/* Force priority into range. */
4454	if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
4455	    && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
4456		kthread_prio = 2;
4457	else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4458		kthread_prio = 1;
4459	else if (kthread_prio < 0)
4460		kthread_prio = 0;
4461	else if (kthread_prio > 99)
4462		kthread_prio = 99;
4463
4464	if (kthread_prio != kthread_prio_in)
4465		pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4466			 kthread_prio, kthread_prio_in);
4467
4468	rcu_scheduler_fully_active = 1;
4469	t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
4470	if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
4471		return 0;
4472	if (kthread_prio) {
4473		sp.sched_priority = kthread_prio;
4474		sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4475	}
4476	rnp = rcu_get_root();
4477	raw_spin_lock_irqsave_rcu_node(rnp, flags);
4478	WRITE_ONCE(rcu_state.gp_activity, jiffies);
4479	WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
4480	// Reset .gp_activity and .gp_req_activity before setting .gp_kthread.
4481	smp_store_release(&rcu_state.gp_kthread, t);  /* ^^^ */
4482	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4483	wake_up_process(t);
4484	rcu_spawn_nocb_kthreads();
4485	rcu_spawn_boost_kthreads();
 
 
 
 
 
 
4486	rcu_spawn_core_kthreads();
 
 
4487	return 0;
4488}
4489early_initcall(rcu_spawn_gp_kthread);
4490
4491/*
4492 * This function is invoked towards the end of the scheduler's
4493 * initialization process.  Before this is called, the idle task might
4494 * contain synchronous grace-period primitives (during which time, this idle
4495 * task is booting the system, and such primitives are no-ops).  After this
4496 * function is called, any synchronous grace-period primitives are run as
4497 * expedited, with the requesting task driving the grace period forward.
4498 * A later core_initcall() rcu_set_runtime_mode() will switch to full
4499 * runtime RCU functionality.
4500 */
4501void rcu_scheduler_starting(void)
4502{
 
 
 
4503	WARN_ON(num_online_cpus() != 1);
4504	WARN_ON(nr_context_switches() > 0);
4505	rcu_test_sync_prims();
 
 
 
 
 
 
 
 
4506	rcu_scheduler_active = RCU_SCHEDULER_INIT;
4507	rcu_test_sync_prims();
4508}
4509
4510/*
4511 * Helper function for rcu_init() that initializes the rcu_state structure.
4512 */
4513static void __init rcu_init_one(void)
4514{
4515	static const char * const buf[] = RCU_NODE_NAME_INIT;
4516	static const char * const fqs[] = RCU_FQS_NAME_INIT;
4517	static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4518	static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4519
4520	int levelspread[RCU_NUM_LVLS];		/* kids/node in each level. */
4521	int cpustride = 1;
4522	int i;
4523	int j;
4524	struct rcu_node *rnp;
4525
4526	BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf));  /* Fix buf[] init! */
4527
4528	/* Silence gcc 4.8 false positive about array index out of range. */
4529	if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4530		panic("rcu_init_one: rcu_num_lvls out of range");
4531
4532	/* Initialize the level-tracking arrays. */
4533
4534	for (i = 1; i < rcu_num_lvls; i++)
4535		rcu_state.level[i] =
4536			rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
4537	rcu_init_levelspread(levelspread, num_rcu_lvl);
4538
4539	/* Initialize the elements themselves, starting from the leaves. */
4540
4541	for (i = rcu_num_lvls - 1; i >= 0; i--) {
4542		cpustride *= levelspread[i];
4543		rnp = rcu_state.level[i];
4544		for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
4545			raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4546			lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4547						   &rcu_node_class[i], buf[i]);
4548			raw_spin_lock_init(&rnp->fqslock);
4549			lockdep_set_class_and_name(&rnp->fqslock,
4550						   &rcu_fqs_class[i], fqs[i]);
4551			rnp->gp_seq = rcu_state.gp_seq;
4552			rnp->gp_seq_needed = rcu_state.gp_seq;
4553			rnp->completedqs = rcu_state.gp_seq;
4554			rnp->qsmask = 0;
4555			rnp->qsmaskinit = 0;
4556			rnp->grplo = j * cpustride;
4557			rnp->grphi = (j + 1) * cpustride - 1;
4558			if (rnp->grphi >= nr_cpu_ids)
4559				rnp->grphi = nr_cpu_ids - 1;
4560			if (i == 0) {
4561				rnp->grpnum = 0;
4562				rnp->grpmask = 0;
4563				rnp->parent = NULL;
4564			} else {
4565				rnp->grpnum = j % levelspread[i - 1];
4566				rnp->grpmask = BIT(rnp->grpnum);
4567				rnp->parent = rcu_state.level[i - 1] +
4568					      j / levelspread[i - 1];
4569			}
4570			rnp->level = i;
4571			INIT_LIST_HEAD(&rnp->blkd_tasks);
4572			rcu_init_one_nocb(rnp);
4573			init_waitqueue_head(&rnp->exp_wq[0]);
4574			init_waitqueue_head(&rnp->exp_wq[1]);
4575			init_waitqueue_head(&rnp->exp_wq[2]);
4576			init_waitqueue_head(&rnp->exp_wq[3]);
4577			spin_lock_init(&rnp->exp_lock);
 
 
 
 
4578		}
4579	}
4580
4581	init_swait_queue_head(&rcu_state.gp_wq);
4582	init_swait_queue_head(&rcu_state.expedited_wq);
4583	rnp = rcu_first_leaf_node();
4584	for_each_possible_cpu(i) {
4585		while (i > rnp->grphi)
4586			rnp++;
4587		per_cpu_ptr(&rcu_data, i)->mynode = rnp;
4588		rcu_boot_init_percpu_data(i);
4589	}
4590}
4591
4592/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4593 * Compute the rcu_node tree geometry from kernel parameters.  This cannot
4594 * replace the definitions in tree.h because those are needed to size
4595 * the ->node array in the rcu_state structure.
4596 */
4597void rcu_init_geometry(void)
4598{
4599	ulong d;
4600	int i;
4601	static unsigned long old_nr_cpu_ids;
4602	int rcu_capacity[RCU_NUM_LVLS];
4603	static bool initialized;
4604
4605	if (initialized) {
4606		/*
4607		 * Warn if setup_nr_cpu_ids() had not yet been invoked,
4608		 * unless nr_cpus_ids == NR_CPUS, in which case who cares?
4609		 */
4610		WARN_ON_ONCE(old_nr_cpu_ids != nr_cpu_ids);
4611		return;
4612	}
4613
4614	old_nr_cpu_ids = nr_cpu_ids;
4615	initialized = true;
4616
4617	/*
4618	 * Initialize any unspecified boot parameters.
4619	 * The default values of jiffies_till_first_fqs and
4620	 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4621	 * value, which is a function of HZ, then adding one for each
4622	 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4623	 */
4624	d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4625	if (jiffies_till_first_fqs == ULONG_MAX)
4626		jiffies_till_first_fqs = d;
4627	if (jiffies_till_next_fqs == ULONG_MAX)
4628		jiffies_till_next_fqs = d;
4629	adjust_jiffies_till_sched_qs();
4630
4631	/* If the compile-time values are accurate, just leave. */
4632	if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4633	    nr_cpu_ids == NR_CPUS)
4634		return;
4635	pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4636		rcu_fanout_leaf, nr_cpu_ids);
4637
4638	/*
4639	 * The boot-time rcu_fanout_leaf parameter must be at least two
4640	 * and cannot exceed the number of bits in the rcu_node masks.
4641	 * Complain and fall back to the compile-time values if this
4642	 * limit is exceeded.
4643	 */
4644	if (rcu_fanout_leaf < 2 ||
4645	    rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4646		rcu_fanout_leaf = RCU_FANOUT_LEAF;
4647		WARN_ON(1);
4648		return;
4649	}
4650
4651	/*
4652	 * Compute number of nodes that can be handled an rcu_node tree
4653	 * with the given number of levels.
4654	 */
4655	rcu_capacity[0] = rcu_fanout_leaf;
4656	for (i = 1; i < RCU_NUM_LVLS; i++)
4657		rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4658
4659	/*
4660	 * The tree must be able to accommodate the configured number of CPUs.
4661	 * If this limit is exceeded, fall back to the compile-time values.
4662	 */
4663	if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4664		rcu_fanout_leaf = RCU_FANOUT_LEAF;
4665		WARN_ON(1);
4666		return;
4667	}
4668
4669	/* Calculate the number of levels in the tree. */
4670	for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4671	}
4672	rcu_num_lvls = i + 1;
4673
4674	/* Calculate the number of rcu_nodes at each level of the tree. */
4675	for (i = 0; i < rcu_num_lvls; i++) {
4676		int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4677		num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4678	}
4679
4680	/* Calculate the total number of rcu_node structures. */
4681	rcu_num_nodes = 0;
4682	for (i = 0; i < rcu_num_lvls; i++)
4683		rcu_num_nodes += num_rcu_lvl[i];
4684}
4685
4686/*
4687 * Dump out the structure of the rcu_node combining tree associated
4688 * with the rcu_state structure.
4689 */
4690static void __init rcu_dump_rcu_node_tree(void)
4691{
4692	int level = 0;
4693	struct rcu_node *rnp;
4694
4695	pr_info("rcu_node tree layout dump\n");
4696	pr_info(" ");
4697	rcu_for_each_node_breadth_first(rnp) {
4698		if (rnp->level != level) {
4699			pr_cont("\n");
4700			pr_info(" ");
4701			level = rnp->level;
4702		}
4703		pr_cont("%d:%d ^%d  ", rnp->grplo, rnp->grphi, rnp->grpnum);
4704	}
4705	pr_cont("\n");
4706}
4707
4708struct workqueue_struct *rcu_gp_wq;
4709struct workqueue_struct *rcu_par_gp_wq;
4710
4711static void __init kfree_rcu_batch_init(void)
4712{
4713	int cpu;
4714	int i;
 
4715
4716	/* Clamp it to [0:100] seconds interval. */
4717	if (rcu_delay_page_cache_fill_msec < 0 ||
4718		rcu_delay_page_cache_fill_msec > 100 * MSEC_PER_SEC) {
4719
4720		rcu_delay_page_cache_fill_msec =
4721			clamp(rcu_delay_page_cache_fill_msec, 0,
4722				(int) (100 * MSEC_PER_SEC));
4723
4724		pr_info("Adjusting rcutree.rcu_delay_page_cache_fill_msec to %d ms.\n",
4725			rcu_delay_page_cache_fill_msec);
4726	}
4727
4728	for_each_possible_cpu(cpu) {
4729		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
4730
4731		for (i = 0; i < KFREE_N_BATCHES; i++) {
4732			INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
4733			krcp->krw_arr[i].krcp = krcp;
 
 
 
4734		}
4735
 
 
 
4736		INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
4737		INIT_DELAYED_WORK(&krcp->page_cache_work, fill_page_cache_func);
4738		krcp->initialized = true;
4739	}
4740	if (register_shrinker(&kfree_rcu_shrinker))
4741		pr_err("Failed to register kfree_rcu() shrinker!\n");
 
 
 
 
 
 
 
 
 
4742}
4743
4744void __init rcu_init(void)
4745{
4746	int cpu;
4747
4748	rcu_early_boot_tests();
4749
4750	kfree_rcu_batch_init();
4751	rcu_bootup_announce();
 
4752	rcu_init_geometry();
4753	rcu_init_one();
4754	if (dump_tree)
4755		rcu_dump_rcu_node_tree();
4756	if (use_softirq)
4757		open_softirq(RCU_SOFTIRQ, rcu_core_si);
4758
4759	/*
4760	 * We don't need protection against CPU-hotplug here because
4761	 * this is called early in boot, before either interrupts
4762	 * or the scheduler are operational.
4763	 */
4764	pm_notifier(rcu_pm_notify, 0);
4765	for_each_online_cpu(cpu) {
4766		rcutree_prepare_cpu(cpu);
4767		rcu_cpu_starting(cpu);
4768		rcutree_online_cpu(cpu);
4769	}
4770
4771	/* Create workqueue for Tree SRCU and for expedited GPs. */
4772	rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
4773	WARN_ON(!rcu_gp_wq);
4774	rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
4775	WARN_ON(!rcu_par_gp_wq);
4776
4777	/* Fill in default value for rcutree.qovld boot parameter. */
4778	/* -After- the rcu_node ->lock fields are initialized! */
4779	if (qovld < 0)
4780		qovld_calc = DEFAULT_RCU_QOVLD_MULT * qhimark;
4781	else
4782		qovld_calc = qovld;
 
 
 
 
 
4783}
4784
4785#include "tree_stall.h"
4786#include "tree_exp.h"
 
4787#include "tree_plugin.h"