Linux Audio

Check our new training course

Loading...
v4.6
 
   1/*
   2 * Read-Copy Update mechanism for mutual exclusion
   3 *
   4 * This program is free software; you can redistribute it and/or modify
   5 * it under the terms of the GNU General Public License as published by
   6 * the Free Software Foundation; either version 2 of the License, or
   7 * (at your option) any later version.
   8 *
   9 * This program is distributed in the hope that it will be useful,
  10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12 * GNU General Public License for more details.
  13 *
  14 * You should have received a copy of the GNU General Public License
  15 * along with this program; if not, you can access it online at
  16 * http://www.gnu.org/licenses/gpl-2.0.html.
  17 *
  18 * Copyright IBM Corporation, 2008
  19 *
  20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
  21 *	    Manfred Spraul <manfred@colorfullife.com>
  22 *	    Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
  23 *
  24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
  25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
  26 *
  27 * For detailed explanation of Read-Copy Update mechanism see -
  28 *	Documentation/RCU
  29 */
 
 
 
  30#include <linux/types.h>
  31#include <linux/kernel.h>
  32#include <linux/init.h>
  33#include <linux/spinlock.h>
  34#include <linux/smp.h>
  35#include <linux/rcupdate.h>
  36#include <linux/interrupt.h>
  37#include <linux/sched.h>
 
  38#include <linux/nmi.h>
  39#include <linux/atomic.h>
  40#include <linux/bitops.h>
  41#include <linux/export.h>
  42#include <linux/completion.h>
  43#include <linux/moduleparam.h>
  44#include <linux/module.h>
  45#include <linux/percpu.h>
  46#include <linux/notifier.h>
  47#include <linux/cpu.h>
  48#include <linux/mutex.h>
  49#include <linux/time.h>
  50#include <linux/kernel_stat.h>
  51#include <linux/wait.h>
  52#include <linux/kthread.h>
 
  53#include <linux/prefetch.h>
  54#include <linux/delay.h>
  55#include <linux/stop_machine.h>
  56#include <linux/random.h>
  57#include <linux/trace_events.h>
  58#include <linux/suspend.h>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  59
  60#include "tree.h"
  61#include "rcu.h"
  62
  63MODULE_ALIAS("rcutree");
  64#ifdef MODULE_PARAM_PREFIX
  65#undef MODULE_PARAM_PREFIX
  66#endif
  67#define MODULE_PARAM_PREFIX "rcutree."
  68
  69/* Data structures. */
  70
  71/*
  72 * In order to export the rcu_state name to the tracing tools, it
  73 * needs to be added in the __tracepoint_string section.
  74 * This requires defining a separate variable tp_<sname>_varname
  75 * that points to the string being used, and this will allow
  76 * the tracing userspace tools to be able to decipher the string
  77 * address to the matching string.
  78 */
  79#ifdef CONFIG_TRACING
  80# define DEFINE_RCU_TPS(sname) \
  81static char sname##_varname[] = #sname; \
  82static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
  83# define RCU_STATE_NAME(sname) sname##_varname
  84#else
  85# define DEFINE_RCU_TPS(sname)
  86# define RCU_STATE_NAME(sname) __stringify(sname)
  87#endif
  88
  89#define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
  90DEFINE_RCU_TPS(sname) \
  91static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
  92struct rcu_state sname##_state = { \
  93	.level = { &sname##_state.node[0] }, \
  94	.rda = &sname##_data, \
  95	.call = cr, \
  96	.gp_state = RCU_GP_IDLE, \
  97	.gpnum = 0UL - 300UL, \
  98	.completed = 0UL - 300UL, \
  99	.orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
 100	.orphan_nxttail = &sname##_state.orphan_nxtlist, \
 101	.orphan_donetail = &sname##_state.orphan_donelist, \
 102	.barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
 103	.name = RCU_STATE_NAME(sname), \
 104	.abbr = sabbr, \
 105}
 106
 107RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
 108RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
 
 
 
 
 109
 110static struct rcu_state *const rcu_state_p;
 111LIST_HEAD(rcu_struct_flavors);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 112
 113/* Dump rcu_node combining tree at boot to verify correct setup. */
 114static bool dump_tree;
 115module_param(dump_tree, bool, 0444);
 
 
 
 116/* Control rcu_node-tree auto-balancing at boot time. */
 117static bool rcu_fanout_exact;
 118module_param(rcu_fanout_exact, bool, 0444);
 119/* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
 120static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
 121module_param(rcu_fanout_leaf, int, 0444);
 122int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
 123/* Number of rcu_nodes at specified level. */
 124static int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
 125int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
 126
 127/*
 128 * The rcu_scheduler_active variable transitions from zero to one just
 129 * before the first task is spawned.  So when this variable is zero, RCU
 130 * can assume that there is but one task, allowing RCU to (for example)
 131 * optimize synchronize_sched() to a simple barrier().  When this variable
 132 * is one, RCU must actually do all the hard work required to detect real
 133 * grace periods.  This variable is also used to suppress boot-time false
 134 * positives from lockdep-RCU error checking.
 
 
 
 135 */
 136int rcu_scheduler_active __read_mostly;
 137EXPORT_SYMBOL_GPL(rcu_scheduler_active);
 138
 139/*
 140 * The rcu_scheduler_fully_active variable transitions from zero to one
 141 * during the early_initcall() processing, which is after the scheduler
 142 * is capable of creating new tasks.  So RCU processing (for example,
 143 * creating tasks for RCU priority boosting) must be delayed until after
 144 * rcu_scheduler_fully_active transitions from zero to one.  We also
 145 * currently delay invocation of any RCU callbacks until after this point.
 146 *
 147 * It might later prove better for people registering RCU callbacks during
 148 * early boot to take responsibility for these callbacks, but one step at
 149 * a time.
 150 */
 151static int rcu_scheduler_fully_active __read_mostly;
 152
 
 
 153static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
 154static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
 155static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
 156static void invoke_rcu_core(void);
 157static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
 158static void rcu_report_exp_rdp(struct rcu_state *rsp,
 159			       struct rcu_data *rdp, bool wake);
 160
 161/* rcuc/rcub kthread realtime priority */
 162#ifdef CONFIG_RCU_KTHREAD_PRIO
 163static int kthread_prio = CONFIG_RCU_KTHREAD_PRIO;
 164#else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */
 165static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
 166#endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */
 167module_param(kthread_prio, int, 0644);
 168
 169/* Delay in jiffies for grace-period initialization delays, debug only. */
 170
 171#ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT
 172static int gp_preinit_delay = CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY;
 173module_param(gp_preinit_delay, int, 0644);
 174#else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
 175static const int gp_preinit_delay;
 176#endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
 177
 178#ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT
 179static int gp_init_delay = CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY;
 180module_param(gp_init_delay, int, 0644);
 181#else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
 182static const int gp_init_delay;
 183#endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
 184
 185#ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP
 186static int gp_cleanup_delay = CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY;
 187module_param(gp_cleanup_delay, int, 0644);
 188#else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
 189static const int gp_cleanup_delay;
 190#endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
 
 
 191
 192/*
 193 * Number of grace periods between delays, normalized by the duration of
 194 * the delay.  The longer the the delay, the more the grace periods between
 195 * each delay.  The reason for this normalization is that it means that,
 196 * for non-zero delays, the overall slowdown of grace periods is constant
 197 * regardless of the duration of the delay.  This arrangement balances
 198 * the need for long delays to increase some race probabilities with the
 199 * need for fast grace periods to increase other race probabilities.
 200 */
 201#define PER_RCU_NODE_PERIOD 3	/* Number of grace periods between delays. */
 202
 203/*
 204 * Track the rcutorture test sequence number and the update version
 205 * number within a given test.  The rcutorture_testseq is incremented
 206 * on every rcutorture module load and unload, so has an odd value
 207 * when a test is running.  The rcutorture_vernum is set to zero
 208 * when rcutorture starts and is incremented on each rcutorture update.
 209 * These variables enable correlating rcutorture output with the
 210 * RCU tracing information.
 211 */
 212unsigned long rcutorture_testseq;
 213unsigned long rcutorture_vernum;
 214
 215/*
 216 * Compute the mask of online CPUs for the specified rcu_node structure.
 217 * This will not be stable unless the rcu_node structure's ->lock is
 218 * held, but the bit corresponding to the current CPU will be stable
 219 * in most contexts.
 220 */
 221unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
 222{
 223	return READ_ONCE(rnp->qsmaskinitnext);
 224}
 225
 226/*
 227 * Return true if an RCU grace period is in progress.  The READ_ONCE()s
 228 * permit this function to be invoked without holding the root rcu_node
 229 * structure's ->lock, but of course results can be subject to change.
 230 */
 231static int rcu_gp_in_progress(struct rcu_state *rsp)
 232{
 233	return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum);
 234}
 235
 236/*
 237 * Note a quiescent state.  Because we do not need to know
 238 * how many quiescent states passed, just if there was at least
 239 * one since the start of the grace period, this just sets a flag.
 240 * The caller must have disabled preemption.
 241 */
 242void rcu_sched_qs(void)
 243{
 244	if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.s))
 245		return;
 246	trace_rcu_grace_period(TPS("rcu_sched"),
 247			       __this_cpu_read(rcu_sched_data.gpnum),
 248			       TPS("cpuqs"));
 249	__this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
 250	if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
 251		return;
 252	__this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
 253	rcu_report_exp_rdp(&rcu_sched_state,
 254			   this_cpu_ptr(&rcu_sched_data), true);
 255}
 256
 257void rcu_bh_qs(void)
 258{
 259	if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) {
 260		trace_rcu_grace_period(TPS("rcu_bh"),
 261				       __this_cpu_read(rcu_bh_data.gpnum),
 262				       TPS("cpuqs"));
 263		__this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false);
 264	}
 265}
 266
 267static DEFINE_PER_CPU(int, rcu_sched_qs_mask);
 268
 269static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
 270	.dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
 271	.dynticks = ATOMIC_INIT(1),
 272#ifdef CONFIG_NO_HZ_FULL_SYSIDLE
 273	.dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
 274	.dynticks_idle = ATOMIC_INIT(1),
 275#endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
 276};
 277
 278DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr);
 279EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr);
 280
 281/*
 282 * Let the RCU core know that this CPU has gone through the scheduler,
 283 * which is a quiescent state.  This is called when the need for a
 284 * quiescent state is urgent, so we burn an atomic operation and full
 285 * memory barriers to let the RCU core know about it, regardless of what
 286 * this CPU might (or might not) do in the near future.
 287 *
 288 * We inform the RCU core by emulating a zero-duration dyntick-idle
 289 * period, which we in turn do by incrementing the ->dynticks counter
 290 * by two.
 291 *
 292 * The caller must have disabled interrupts.
 293 */
 294static void rcu_momentary_dyntick_idle(void)
 295{
 296	struct rcu_data *rdp;
 297	struct rcu_dynticks *rdtp;
 298	int resched_mask;
 299	struct rcu_state *rsp;
 300
 301	/*
 302	 * Yes, we can lose flag-setting operations.  This is OK, because
 303	 * the flag will be set again after some delay.
 
 304	 */
 305	resched_mask = raw_cpu_read(rcu_sched_qs_mask);
 306	raw_cpu_write(rcu_sched_qs_mask, 0);
 307
 308	/* Find the flavor that needs a quiescent state. */
 309	for_each_rcu_flavor(rsp) {
 310		rdp = raw_cpu_ptr(rsp->rda);
 311		if (!(resched_mask & rsp->flavor_mask))
 312			continue;
 313		smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
 314		if (READ_ONCE(rdp->mynode->completed) !=
 315		    READ_ONCE(rdp->cond_resched_completed))
 316			continue;
 317
 318		/*
 319		 * Pretend to be momentarily idle for the quiescent state.
 320		 * This allows the grace-period kthread to record the
 321		 * quiescent state, with no need for this CPU to do anything
 322		 * further.
 323		 */
 324		rdtp = this_cpu_ptr(&rcu_dynticks);
 325		smp_mb__before_atomic(); /* Earlier stuff before QS. */
 326		atomic_add(2, &rdtp->dynticks);  /* QS. */
 327		smp_mb__after_atomic(); /* Later stuff after QS. */
 328		break;
 329	}
 330}
 331
 332/*
 333 * Note a context switch.  This is a quiescent state for RCU-sched,
 334 * and requires special handling for preemptible RCU.
 335 * The caller must have disabled interrupts.
 336 */
 337void rcu_note_context_switch(void)
 338{
 339	barrier(); /* Avoid RCU read-side critical sections leaking down. */
 340	trace_rcu_utilization(TPS("Start context switch"));
 341	rcu_sched_qs();
 342	rcu_preempt_note_context_switch();
 343	if (unlikely(raw_cpu_read(rcu_sched_qs_mask)))
 344		rcu_momentary_dyntick_idle();
 345	trace_rcu_utilization(TPS("End context switch"));
 346	barrier(); /* Avoid RCU read-side critical sections leaking up. */
 
 
 
 
 
 
 
 
 
 347}
 348EXPORT_SYMBOL_GPL(rcu_note_context_switch);
 349
 350/*
 351 * Register a quiescent state for all RCU flavors.  If there is an
 352 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
 353 * dyntick-idle quiescent state visible to other CPUs (but only for those
 354 * RCU flavors in desperate need of a quiescent state, which will normally
 355 * be none of them).  Either way, do a lightweight quiescent state for
 356 * all RCU flavors.
 357 *
 358 * The barrier() calls are redundant in the common case when this is
 359 * called externally, but just in case this is called from within this
 360 * file.
 361 *
 
 
 
 362 */
 363void rcu_all_qs(void)
 364{
 365	unsigned long flags;
 366
 367	barrier(); /* Avoid RCU read-side critical sections leaking down. */
 368	if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) {
 369		local_irq_save(flags);
 370		rcu_momentary_dyntick_idle();
 371		local_irq_restore(flags);
 372	}
 373	this_cpu_inc(rcu_qs_ctr);
 374	barrier(); /* Avoid RCU read-side critical sections leaking up. */
 375}
 376EXPORT_SYMBOL_GPL(rcu_all_qs);
 377
 378static long blimit = 10;	/* Maximum callbacks per rcu_do_batch. */
 379static long qhimark = 10000;	/* If this many pending, ignore blimit. */
 380static long qlowmark = 100;	/* Once only this many pending, use blimit. */
 381
 382module_param(blimit, long, 0444);
 383module_param(qhimark, long, 0444);
 384module_param(qlowmark, long, 0444);
 385
 386static ulong jiffies_till_first_fqs = ULONG_MAX;
 387static ulong jiffies_till_next_fqs = ULONG_MAX;
 388
 389module_param(jiffies_till_first_fqs, ulong, 0644);
 390module_param(jiffies_till_next_fqs, ulong, 0644);
 391
 392/*
 393 * How long the grace period must be before we start recruiting
 394 * quiescent-state help from rcu_note_context_switch().
 395 */
 396static ulong jiffies_till_sched_qs = HZ / 20;
 397module_param(jiffies_till_sched_qs, ulong, 0644);
 398
 399static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
 400				  struct rcu_data *rdp);
 401static void force_qs_rnp(struct rcu_state *rsp,
 402			 int (*f)(struct rcu_data *rsp, bool *isidle,
 403				  unsigned long *maxj),
 404			 bool *isidle, unsigned long *maxj);
 405static void force_quiescent_state(struct rcu_state *rsp);
 406static int rcu_pending(void);
 407
 408/*
 409 * Return the number of RCU batches started thus far for debug & stats.
 410 */
 411unsigned long rcu_batches_started(void)
 412{
 413	return rcu_state_p->gpnum;
 414}
 415EXPORT_SYMBOL_GPL(rcu_batches_started);
 416
 417/*
 418 * Return the number of RCU-sched batches started thus far for debug & stats.
 419 */
 420unsigned long rcu_batches_started_sched(void)
 421{
 422	return rcu_sched_state.gpnum;
 423}
 424EXPORT_SYMBOL_GPL(rcu_batches_started_sched);
 425
 426/*
 427 * Return the number of RCU BH batches started thus far for debug & stats.
 
 428 */
 429unsigned long rcu_batches_started_bh(void)
 430{
 431	return rcu_bh_state.gpnum;
 432}
 433EXPORT_SYMBOL_GPL(rcu_batches_started_bh);
 434
 435/*
 436 * Return the number of RCU batches completed thus far for debug & stats.
 437 */
 438unsigned long rcu_batches_completed(void)
 439{
 440	return rcu_state_p->completed;
 441}
 442EXPORT_SYMBOL_GPL(rcu_batches_completed);
 443
 444/*
 445 * Return the number of RCU-sched batches completed thus far for debug & stats.
 
 446 */
 447unsigned long rcu_batches_completed_sched(void)
 448{
 449	return rcu_sched_state.completed;
 450}
 451EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
 452
 453/*
 454 * Return the number of RCU BH batches completed thus far for debug & stats.
 
 
 455 */
 456unsigned long rcu_batches_completed_bh(void)
 457{
 458	return rcu_bh_state.completed;
 459}
 460EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
 461
 462/*
 463 * Force a quiescent state.
 
 464 */
 465void rcu_force_quiescent_state(void)
 466{
 467	force_quiescent_state(rcu_state_p);
 
 
 
 
 
 
 
 
 
 
 
 
 
 468}
 469EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
 470
 471/*
 472 * Force a quiescent state for RCU BH.
 
 
 
 
 473 */
 474void rcu_bh_force_quiescent_state(void)
 475{
 476	force_quiescent_state(&rcu_bh_state);
 
 
 
 
 
 
 
 
 
 
 
 
 
 477}
 478EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
 479
 480/*
 481 * Force a quiescent state for RCU-sched.
 
 
 
 
 
 
 
 
 482 */
 483void rcu_sched_force_quiescent_state(void)
 484{
 485	force_quiescent_state(&rcu_sched_state);
 
 
 
 
 
 
 
 486}
 487EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
 488
 489/*
 490 * Show the state of the grace-period kthreads.
 
 
 
 
 
 491 */
 492void show_rcu_gp_kthreads(void)
 493{
 494	struct rcu_state *rsp;
 495
 496	for_each_rcu_flavor(rsp) {
 497		pr_info("%s: wait state: %d ->state: %#lx\n",
 498			rsp->name, rsp->gp_state, rsp->gp_kthread->state);
 499		/* sched_show_task(rsp->gp_kthread); */
 500	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 501}
 502EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 503
 504/*
 505 * Record the number of times rcutorture tests have been initiated and
 506 * terminated.  This information allows the debugfs tracing stats to be
 507 * correlated to the rcutorture messages, even when the rcutorture module
 508 * is being repeatedly loaded and unloaded.  In other words, we cannot
 509 * store this state in rcutorture itself.
 510 */
 511void rcutorture_record_test_transition(void)
 512{
 513	rcutorture_testseq++;
 514	rcutorture_vernum = 0;
 515}
 516EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
 517
 518/*
 519 * Send along grace-period-related data for rcutorture diagnostics.
 
 
 
 520 */
 521void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
 522			    unsigned long *gpnum, unsigned long *completed)
 523{
 524	struct rcu_state *rsp = NULL;
 525
 526	switch (test_type) {
 527	case RCU_FLAVOR:
 528		rsp = rcu_state_p;
 529		break;
 530	case RCU_BH_FLAVOR:
 531		rsp = &rcu_bh_state;
 532		break;
 533	case RCU_SCHED_FLAVOR:
 534		rsp = &rcu_sched_state;
 535		break;
 536	default:
 537		break;
 538	}
 539	if (rsp != NULL) {
 540		*flags = READ_ONCE(rsp->gp_flags);
 541		*gpnum = READ_ONCE(rsp->gpnum);
 542		*completed = READ_ONCE(rsp->completed);
 543		return;
 544	}
 545	*flags = 0;
 546	*gpnum = 0;
 547	*completed = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 548}
 549EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
 550
 551/*
 552 * Record the number of writer passes through the current rcutorture test.
 553 * This is also used to correlate debugfs tracing stats with the rcutorture
 554 * messages.
 555 */
 556void rcutorture_record_progress(unsigned long vernum)
 557{
 558	rcutorture_vernum++;
 559}
 560EXPORT_SYMBOL_GPL(rcutorture_record_progress);
 561
 562/*
 563 * Does the CPU have callbacks ready to be invoked?
 564 */
 565static int
 566cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
 567{
 568	return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
 569	       rdp->nxttail[RCU_DONE_TAIL] != NULL;
 570}
 571
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 572/*
 573 * Return the root node of the specified rcu_state structure.
 574 */
 575static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
 576{
 577	return &rsp->node[0];
 578}
 
 579
 580/*
 581 * Is there any need for future grace periods?
 582 * Interrupts must be disabled.  If the caller does not hold the root
 583 * rnp_node structure's ->lock, the results are advisory only.
 
 584 */
 585static int rcu_future_needs_gp(struct rcu_state *rsp)
 586{
 587	struct rcu_node *rnp = rcu_get_root(rsp);
 588	int idx = (READ_ONCE(rnp->completed) + 1) & 0x1;
 589	int *fp = &rnp->need_future_gp[idx];
 590
 591	return READ_ONCE(*fp);
 592}
 
 593
 594/*
 595 * Does the current CPU require a not-yet-started grace period?
 596 * The caller must have disabled interrupts to prevent races with
 597 * normal callback registry.
 598 */
 599static bool
 600cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
 601{
 602	int i;
 603
 604	if (rcu_gp_in_progress(rsp))
 605		return false;  /* No, a grace period is already in progress. */
 606	if (rcu_future_needs_gp(rsp))
 607		return true;  /* Yes, a no-CBs CPU needs one. */
 608	if (!rdp->nxttail[RCU_NEXT_TAIL])
 609		return false;  /* No, this is a no-CBs (or offline) CPU. */
 610	if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
 611		return true;  /* Yes, CPU has newly registered callbacks. */
 612	for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
 613		if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
 614		    ULONG_CMP_LT(READ_ONCE(rsp->completed),
 615				 rdp->nxtcompleted[i]))
 616			return true;  /* Yes, CBs for future grace period. */
 617	return false; /* No grace period needed. */
 618}
 619
 620/*
 621 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
 622 *
 623 * If the new value of the ->dynticks_nesting counter now is zero,
 624 * we really have entered idle, and must do the appropriate accounting.
 625 * The caller must have disabled interrupts.
 626 */
 627static void rcu_eqs_enter_common(long long oldval, bool user)
 
 628{
 629	struct rcu_state *rsp;
 630	struct rcu_data *rdp;
 631	struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
 632
 633	trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
 634	if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
 635	    !user && !is_idle_task(current)) {
 636		struct task_struct *idle __maybe_unused =
 637			idle_task(smp_processor_id());
 638
 639		trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
 640		ftrace_dump(DUMP_ORIG);
 641		WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
 642			  current->pid, current->comm,
 643			  idle->pid, idle->comm); /* must be idle task! */
 644	}
 645	for_each_rcu_flavor(rsp) {
 646		rdp = this_cpu_ptr(rsp->rda);
 647		do_nocb_deferred_wakeup(rdp);
 648	}
 649	rcu_prepare_for_idle();
 650	/* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
 651	smp_mb__before_atomic();  /* See above. */
 652	atomic_inc(&rdtp->dynticks);
 653	smp_mb__after_atomic();  /* Force ordering with next sojourn. */
 654	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
 655		     atomic_read(&rdtp->dynticks) & 0x1);
 656	rcu_dynticks_task_enter();
 657
 658	/*
 659	 * It is illegal to enter an extended quiescent state while
 660	 * in an RCU read-side critical section.
 661	 */
 662	RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map),
 663			 "Illegal idle entry in RCU read-side critical section.");
 664	RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map),
 665			 "Illegal idle entry in RCU-bh read-side critical section.");
 666	RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map),
 667			 "Illegal idle entry in RCU-sched read-side critical section.");
 668}
 
 669
 670/*
 671 * Enter an RCU extended quiescent state, which can be either the
 672 * idle loop or adaptive-tickless usermode execution.
 
 
 
 
 673 */
 674static void rcu_eqs_enter(bool user)
 675{
 676	long long oldval;
 677	struct rcu_dynticks *rdtp;
 678
 679	rdtp = this_cpu_ptr(&rcu_dynticks);
 680	oldval = rdtp->dynticks_nesting;
 681	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
 682		     (oldval & DYNTICK_TASK_NEST_MASK) == 0);
 683	if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
 684		rdtp->dynticks_nesting = 0;
 685		rcu_eqs_enter_common(oldval, user);
 686	} else {
 687		rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
 688	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 689}
 690
 691/**
 692 * rcu_idle_enter - inform RCU that current CPU is entering idle
 693 *
 694 * Enter idle mode, in other words, -leave- the mode in which RCU
 695 * read-side critical sections can occur.  (Though RCU read-side
 696 * critical sections can occur in irq handlers in idle, a possibility
 697 * handled by irq_enter() and irq_exit().)
 698 *
 699 * We crowbar the ->dynticks_nesting field to zero to allow for
 700 * the possibility of usermode upcalls having messed up our count
 701 * of interrupt nesting level during the prior busy period.
 702 */
 703void rcu_idle_enter(void)
 704{
 705	unsigned long flags;
 706
 707	local_irq_save(flags);
 708	rcu_eqs_enter(false);
 709	rcu_sysidle_enter(0);
 710	local_irq_restore(flags);
 711}
 712EXPORT_SYMBOL_GPL(rcu_idle_enter);
 713
 714#ifdef CONFIG_NO_HZ_FULL
 715/**
 716 * rcu_user_enter - inform RCU that we are resuming userspace.
 717 *
 718 * Enter RCU idle mode right before resuming userspace.  No use of RCU
 719 * is permitted between this call and rcu_user_exit(). This way the
 720 * CPU doesn't need to maintain the tick for RCU maintenance purposes
 721 * when the CPU runs in userspace.
 
 
 
 722 */
 723void rcu_user_enter(void)
 724{
 725	rcu_eqs_enter(1);
 
 726}
 727#endif /* CONFIG_NO_HZ_FULL */
 728
 729/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 730 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
 731 *
 732 * Exit from an interrupt handler, which might possibly result in entering
 733 * idle mode, in other words, leaving the mode in which read-side critical
 734 * sections can occur.  The caller must have disabled interrupts.
 735 *
 736 * This code assumes that the idle loop never does anything that might
 737 * result in unbalanced calls to irq_enter() and irq_exit().  If your
 738 * architecture violates this assumption, RCU will give you what you
 739 * deserve, good and hard.  But very infrequently and irreproducibly.
 740 *
 741 * Use things like work queues to work around this limitation.
 742 *
 743 * You have been warned.
 
 
 
 744 */
 745void rcu_irq_exit(void)
 746{
 747	long long oldval;
 748	struct rcu_dynticks *rdtp;
 
 749
 750	RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_exit() invoked with irqs enabled!!!");
 751	rdtp = this_cpu_ptr(&rcu_dynticks);
 752	oldval = rdtp->dynticks_nesting;
 753	rdtp->dynticks_nesting--;
 754	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
 755		     rdtp->dynticks_nesting < 0);
 756	if (rdtp->dynticks_nesting)
 757		trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
 758	else
 759		rcu_eqs_enter_common(oldval, true);
 760	rcu_sysidle_enter(1);
 
 
 
 
 
 
 
 
 
 761}
 762
 763/*
 764 * Wrapper for rcu_irq_exit() where interrupts are enabled.
 
 765 */
 766void rcu_irq_exit_irqson(void)
 767{
 768	unsigned long flags;
 769
 770	local_irq_save(flags);
 771	rcu_irq_exit();
 772	local_irq_restore(flags);
 
 
 
 
 773}
 
 774
 775/*
 776 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
 777 *
 778 * If the new value of the ->dynticks_nesting counter was previously zero,
 779 * we really have exited idle, and must do the appropriate accounting.
 780 * The caller must have disabled interrupts.
 781 */
 782static void rcu_eqs_exit_common(long long oldval, int user)
 783{
 784	struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
 785
 786	rcu_dynticks_task_exit();
 787	smp_mb__before_atomic();  /* Force ordering w/previous sojourn. */
 788	atomic_inc(&rdtp->dynticks);
 789	/* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
 790	smp_mb__after_atomic();  /* See above. */
 791	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
 792		     !(atomic_read(&rdtp->dynticks) & 0x1));
 793	rcu_cleanup_after_idle();
 794	trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
 795	if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
 796	    !user && !is_idle_task(current)) {
 797		struct task_struct *idle __maybe_unused =
 798			idle_task(smp_processor_id());
 799
 800		trace_rcu_dyntick(TPS("Error on exit: not idle task"),
 801				  oldval, rdtp->dynticks_nesting);
 802		ftrace_dump(DUMP_ORIG);
 803		WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
 804			  current->pid, current->comm,
 805			  idle->pid, idle->comm); /* must be idle task! */
 806	}
 807}
 808
 809/*
 810 * Exit an RCU extended quiescent state, which can be either the
 811 * idle loop or adaptive-tickless usermode execution.
 
 
 
 
 812 */
 813static void rcu_eqs_exit(bool user)
 814{
 815	struct rcu_dynticks *rdtp;
 816	long long oldval;
 817
 818	rdtp = this_cpu_ptr(&rcu_dynticks);
 819	oldval = rdtp->dynticks_nesting;
 
 820	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
 821	if (oldval & DYNTICK_TASK_NEST_MASK) {
 822		rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
 823	} else {
 824		rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
 825		rcu_eqs_exit_common(oldval, user);
 826	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 827}
 828
 829/**
 830 * rcu_idle_exit - inform RCU that current CPU is leaving idle
 831 *
 832 * Exit idle mode, in other words, -enter- the mode in which RCU
 833 * read-side critical sections can occur.
 834 *
 835 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
 836 * allow for the possibility of usermode upcalls messing up our count
 837 * of interrupt nesting level during the busy period that is just
 838 * now starting.
 839 */
 840void rcu_idle_exit(void)
 841{
 842	unsigned long flags;
 843
 844	local_irq_save(flags);
 845	rcu_eqs_exit(false);
 846	rcu_sysidle_exit(0);
 847	local_irq_restore(flags);
 848}
 849EXPORT_SYMBOL_GPL(rcu_idle_exit);
 850
 851#ifdef CONFIG_NO_HZ_FULL
 852/**
 853 * rcu_user_exit - inform RCU that we are exiting userspace.
 854 *
 855 * Exit RCU idle mode while entering the kernel because it can
 856 * run a RCU read side critical section anytime.
 
 
 
 857 */
 858void rcu_user_exit(void)
 859{
 860	rcu_eqs_exit(1);
 861}
 862#endif /* CONFIG_NO_HZ_FULL */
 863
 864/**
 865 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
 866 *
 867 * Enter an interrupt handler, which might possibly result in exiting
 868 * idle mode, in other words, entering the mode in which read-side critical
 869 * sections can occur.  The caller must have disabled interrupts.
 870 *
 871 * Note that the Linux kernel is fully capable of entering an interrupt
 872 * handler that it never exits, for example when doing upcalls to
 873 * user mode!  This code assumes that the idle loop never does upcalls to
 874 * user mode.  If your architecture does do upcalls from the idle loop (or
 875 * does anything else that results in unbalanced calls to the irq_enter()
 876 * and irq_exit() functions), RCU will give you what you deserve, good
 877 * and hard.  But very infrequently and irreproducibly.
 878 *
 879 * Use things like work queues to work around this limitation.
 880 *
 881 * You have been warned.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 882 */
 883void rcu_irq_enter(void)
 884{
 885	struct rcu_dynticks *rdtp;
 886	long long oldval;
 887
 888	RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_enter() invoked with irqs enabled!!!");
 889	rdtp = this_cpu_ptr(&rcu_dynticks);
 890	oldval = rdtp->dynticks_nesting;
 891	rdtp->dynticks_nesting++;
 892	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
 893		     rdtp->dynticks_nesting == 0);
 894	if (oldval)
 895		trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
 896	else
 897		rcu_eqs_exit_common(oldval, true);
 898	rcu_sysidle_exit(1);
 899}
 900
 901/*
 902 * Wrapper for rcu_irq_enter() where interrupts are enabled.
 903 */
 904void rcu_irq_enter_irqson(void)
 905{
 906	unsigned long flags;
 907
 908	local_irq_save(flags);
 909	rcu_irq_enter();
 910	local_irq_restore(flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 911}
 
 912
 913/**
 914 * rcu_nmi_enter - inform RCU of entry to NMI context
 915 *
 916 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
 917 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
 918 * that the CPU is active.  This implementation permits nested NMIs, as
 919 * long as the nesting level does not overflow an int.  (You will probably
 920 * run out of stack space first.)
 
 
 
 921 */
 922void rcu_nmi_enter(void)
 923{
 924	struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
 925	int incby = 2;
 926
 927	/* Complain about underflow. */
 928	WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
 929
 930	/*
 931	 * If idle from RCU viewpoint, atomically increment ->dynticks
 932	 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
 933	 * Otherwise, increment ->dynticks_nmi_nesting by two.  This means
 934	 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
 935	 * to be in the outermost NMI handler that interrupted an RCU-idle
 936	 * period (observation due to Andy Lutomirski).
 937	 */
 938	if (!(atomic_read(&rdtp->dynticks) & 0x1)) {
 939		smp_mb__before_atomic();  /* Force delay from prior write. */
 940		atomic_inc(&rdtp->dynticks);
 941		/* atomic_inc() before later RCU read-side crit sects */
 942		smp_mb__after_atomic();  /* See above. */
 943		WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 944		incby = 1;
 
 
 
 
 
 
 945	}
 946	rdtp->dynticks_nmi_nesting += incby;
 
 
 
 
 
 
 947	barrier();
 948}
 949
 950/**
 951 * rcu_nmi_exit - inform RCU of exit from NMI context
 952 *
 953 * If we are returning from the outermost NMI handler that interrupted an
 954 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
 955 * to let the RCU grace-period handling know that the CPU is back to
 956 * being RCU-idle.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 957 */
 958void rcu_nmi_exit(void)
 959{
 960	struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
 961
 962	/*
 963	 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
 964	 * (We are exiting an NMI handler, so RCU better be paying attention
 965	 * to us!)
 966	 */
 967	WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
 968	WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
 969
 970	/*
 971	 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
 972	 * leave it in non-RCU-idle state.
 973	 */
 974	if (rdtp->dynticks_nmi_nesting != 1) {
 975		rdtp->dynticks_nmi_nesting -= 2;
 976		return;
 977	}
 
 978
 979	/* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
 980	rdtp->dynticks_nmi_nesting = 0;
 981	/* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
 982	smp_mb__before_atomic();  /* See above. */
 983	atomic_inc(&rdtp->dynticks);
 984	smp_mb__after_atomic();  /* Force delay to next write. */
 985	WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
 986}
 987
 988/**
 989 * __rcu_is_watching - are RCU read-side critical sections safe?
 990 *
 991 * Return true if RCU is watching the running CPU, which means that
 992 * this CPU can safely enter RCU read-side critical sections.  Unlike
 993 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
 994 * least disabled preemption.
 995 */
 996bool notrace __rcu_is_watching(void)
 
 
 
 
 
 
 
 
 
 
 
 997{
 998	return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
 999}
1000
1001/**
1002 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1003 *
1004 * If the current CPU is in its idle loop and is neither in an interrupt
1005 * or NMI handler, return true.
 
 
1006 */
1007bool notrace rcu_is_watching(void)
1008{
1009	bool ret;
1010
1011	preempt_disable_notrace();
1012	ret = __rcu_is_watching();
1013	preempt_enable_notrace();
1014	return ret;
1015}
1016EXPORT_SYMBOL_GPL(rcu_is_watching);
1017
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1018#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1019
1020/*
1021 * Is the current CPU online?  Disable preemption to avoid false positives
1022 * that could otherwise happen due to the current CPU number being sampled,
1023 * this task being preempted, its old CPU being taken offline, resuming
1024 * on some other CPU, then determining that its old CPU is now offline.
1025 * It is OK to use RCU on an offline processor during initial boot, hence
1026 * the check for rcu_scheduler_fully_active.  Note also that it is OK
1027 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1028 * online in the cpu_online_mask.  Similarly, it is OK for a CPU going
1029 * offline to continue to use RCU for one jiffy after marking itself
1030 * offline in the cpu_online_mask.  This leniency is necessary given the
1031 * non-atomic nature of the online and offline processing, for example,
1032 * the fact that a CPU enters the scheduler after completing the CPU_DYING
1033 * notifiers.
1034 *
1035 * This is also why RCU internally marks CPUs online during the
1036 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
1037 *
1038 * Disable checking if in an NMI handler because we cannot safely report
1039 * errors from NMI handlers anyway.
 
 
 
 
1040 */
1041bool rcu_lockdep_current_cpu_online(void)
1042{
1043	struct rcu_data *rdp;
1044	struct rcu_node *rnp;
1045	bool ret;
1046
1047	if (in_nmi())
1048		return true;
1049	preempt_disable();
1050	rdp = this_cpu_ptr(&rcu_sched_data);
1051	rnp = rdp->mynode;
1052	ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
1053	      !rcu_scheduler_fully_active;
1054	preempt_enable();
1055	return ret;
1056}
1057EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1058
1059#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1060
1061/**
1062 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1063 *
1064 * If the current CPU is idle or running at a first-level (not nested)
1065 * interrupt from idle, return true.  The caller must have at least
1066 * disabled preemption.
1067 */
1068static int rcu_is_cpu_rrupt_from_idle(void)
1069{
1070	return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
 
 
 
 
 
1071}
1072
1073/*
1074 * Snapshot the specified CPU's dynticks counter so that we can later
1075 * credit them with an implicit quiescent state.  Return 1 if this CPU
1076 * is in dynticks idle mode, which is an extended quiescent state.
1077 */
1078static int dyntick_save_progress_counter(struct rcu_data *rdp,
1079					 bool *isidle, unsigned long *maxj)
1080{
1081	rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
1082	rcu_sysidle_check_cpu(rdp, isidle, maxj);
1083	if ((rdp->dynticks_snap & 0x1) == 0) {
1084		trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1085		if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4,
1086				 rdp->mynode->gpnum))
1087			WRITE_ONCE(rdp->gpwrap, true);
1088		return 1;
1089	}
1090	return 0;
1091}
1092
1093/*
1094 * Return true if the specified CPU has passed through a quiescent
1095 * state by virtue of being in or having passed through an dynticks
1096 * idle state since the last call to dyntick_save_progress_counter()
1097 * for this same CPU, or by virtue of having been offline.
1098 */
1099static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
1100				    bool *isidle, unsigned long *maxj)
1101{
1102	unsigned int curr;
1103	int *rcrmp;
1104	unsigned int snap;
1105
1106	curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
1107	snap = (unsigned int)rdp->dynticks_snap;
1108
1109	/*
1110	 * If the CPU passed through or entered a dynticks idle phase with
1111	 * no active irq/NMI handlers, then we can safely pretend that the CPU
1112	 * already acknowledged the request to pass through a quiescent
1113	 * state.  Either way, that CPU cannot possibly be in an RCU
1114	 * read-side critical section that started before the beginning
1115	 * of the current RCU grace period.
1116	 */
1117	if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
1118		trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1119		rdp->dynticks_fqs++;
1120		return 1;
1121	}
1122
1123	/*
1124	 * Check for the CPU being offline, but only if the grace period
1125	 * is old enough.  We don't need to worry about the CPU changing
1126	 * state: If we see it offline even once, it has been through a
1127	 * quiescent state.
1128	 *
1129	 * The reason for insisting that the grace period be at least
1130	 * one jiffy old is that CPUs that are not quite online and that
1131	 * have just gone offline can still execute RCU read-side critical
1132	 * sections.
1133	 */
1134	if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
1135		return 0;  /* Grace period is not old enough. */
1136	barrier();
1137	if (cpu_is_offline(rdp->cpu)) {
1138		trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
1139		rdp->offline_fqs++;
1140		return 1;
 
1141	}
1142
1143	/*
1144	 * A CPU running for an extended time within the kernel can
1145	 * delay RCU grace periods.  When the CPU is in NO_HZ_FULL mode,
1146	 * even context-switching back and forth between a pair of
1147	 * in-kernel CPU-bound tasks cannot advance grace periods.
1148	 * So if the grace period is old enough, make the CPU pay attention.
1149	 * Note that the unsynchronized assignments to the per-CPU
1150	 * rcu_sched_qs_mask variable are safe.  Yes, setting of
1151	 * bits can be lost, but they will be set again on the next
1152	 * force-quiescent-state pass.  So lost bit sets do not result
1153	 * in incorrect behavior, merely in a grace period lasting
1154	 * a few jiffies longer than it might otherwise.  Because
1155	 * there are at most four threads involved, and because the
1156	 * updates are only once every few jiffies, the probability of
1157	 * lossage (and thus of slight grace-period extension) is
1158	 * quite low.
1159	 *
1160	 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1161	 * is set too high, we override with half of the RCU CPU stall
1162	 * warning delay.
1163	 */
1164	rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu);
1165	if (ULONG_CMP_GE(jiffies,
1166			 rdp->rsp->gp_start + jiffies_till_sched_qs) ||
1167	    ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
1168		if (!(READ_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) {
1169			WRITE_ONCE(rdp->cond_resched_completed,
1170				   READ_ONCE(rdp->mynode->completed));
1171			smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1172			WRITE_ONCE(*rcrmp,
1173				   READ_ONCE(*rcrmp) + rdp->rsp->flavor_mask);
1174		}
1175		rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */
1176	}
1177
1178	/* And if it has been a really long time, kick the CPU as well. */
1179	if (ULONG_CMP_GE(jiffies,
1180			 rdp->rsp->gp_start + 2 * jiffies_till_sched_qs) ||
1181	    ULONG_CMP_GE(jiffies, rdp->rsp->gp_start + jiffies_till_sched_qs))
1182		resched_cpu(rdp->cpu);  /* Force CPU into scheduler. */
1183
1184	return 0;
1185}
1186
1187static void record_gp_stall_check_time(struct rcu_state *rsp)
1188{
1189	unsigned long j = jiffies;
1190	unsigned long j1;
1191
1192	rsp->gp_start = j;
1193	smp_wmb(); /* Record start time before stall time. */
1194	j1 = rcu_jiffies_till_stall_check();
1195	WRITE_ONCE(rsp->jiffies_stall, j + j1);
1196	rsp->jiffies_resched = j + j1 / 2;
1197	rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1198}
1199
1200/*
1201 * Convert a ->gp_state value to a character string.
1202 */
1203static const char *gp_state_getname(short gs)
1204{
1205	if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
1206		return "???";
1207	return gp_state_names[gs];
1208}
1209
1210/*
1211 * Complain about starvation of grace-period kthread.
1212 */
1213static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1214{
1215	unsigned long gpa;
1216	unsigned long j;
1217
1218	j = jiffies;
1219	gpa = READ_ONCE(rsp->gp_activity);
1220	if (j - gpa > 2 * HZ) {
1221		pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x %s(%d) ->state=%#lx\n",
1222		       rsp->name, j - gpa,
1223		       rsp->gpnum, rsp->completed,
1224		       rsp->gp_flags,
1225		       gp_state_getname(rsp->gp_state), rsp->gp_state,
1226		       rsp->gp_kthread ? rsp->gp_kthread->state : ~0);
1227		if (rsp->gp_kthread)
1228			sched_show_task(rsp->gp_kthread);
1229	}
1230}
1231
1232/*
1233 * Dump stacks of all tasks running on stalled CPUs.
1234 */
1235static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1236{
1237	int cpu;
1238	unsigned long flags;
1239	struct rcu_node *rnp;
1240
1241	rcu_for_each_leaf_node(rsp, rnp) {
1242		raw_spin_lock_irqsave_rcu_node(rnp, flags);
1243		if (rnp->qsmask != 0) {
1244			for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1245				if (rnp->qsmask & (1UL << cpu))
1246					dump_cpu_task(rnp->grplo + cpu);
1247		}
1248		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1249	}
1250}
1251
1252static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
1253{
1254	int cpu;
1255	long delta;
1256	unsigned long flags;
1257	unsigned long gpa;
1258	unsigned long j;
1259	int ndetected = 0;
1260	struct rcu_node *rnp = rcu_get_root(rsp);
1261	long totqlen = 0;
1262
1263	/* Only let one CPU complain about others per time interval. */
1264
1265	raw_spin_lock_irqsave_rcu_node(rnp, flags);
1266	delta = jiffies - READ_ONCE(rsp->jiffies_stall);
1267	if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1268		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1269		return;
1270	}
1271	WRITE_ONCE(rsp->jiffies_stall,
1272		   jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1273	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1274
1275	/*
1276	 * OK, time to rat on our buddy...
1277	 * See Documentation/RCU/stallwarn.txt for info on how to debug
1278	 * RCU CPU stall warnings.
1279	 */
1280	pr_err("INFO: %s detected stalls on CPUs/tasks:",
1281	       rsp->name);
1282	print_cpu_stall_info_begin();
1283	rcu_for_each_leaf_node(rsp, rnp) {
1284		raw_spin_lock_irqsave_rcu_node(rnp, flags);
1285		ndetected += rcu_print_task_stall(rnp);
1286		if (rnp->qsmask != 0) {
1287			for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
1288				if (rnp->qsmask & (1UL << cpu)) {
1289					print_cpu_stall_info(rsp,
1290							     rnp->grplo + cpu);
1291					ndetected++;
1292				}
1293		}
1294		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1295	}
1296
1297	print_cpu_stall_info_end();
1298	for_each_possible_cpu(cpu)
1299		totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1300	pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1301	       smp_processor_id(), (long)(jiffies - rsp->gp_start),
1302	       (long)rsp->gpnum, (long)rsp->completed, totqlen);
1303	if (ndetected) {
1304		rcu_dump_cpu_stacks(rsp);
1305	} else {
1306		if (READ_ONCE(rsp->gpnum) != gpnum ||
1307		    READ_ONCE(rsp->completed) == gpnum) {
1308			pr_err("INFO: Stall ended before state dump start\n");
1309		} else {
1310			j = jiffies;
1311			gpa = READ_ONCE(rsp->gp_activity);
1312			pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1313			       rsp->name, j - gpa, j, gpa,
1314			       jiffies_till_next_fqs,
1315			       rcu_get_root(rsp)->qsmask);
1316			/* In this case, the current CPU might be at fault. */
1317			sched_show_task(current);
1318		}
1319	}
1320
1321	/* Complain about tasks blocking the grace period. */
1322	rcu_print_detail_task_stall(rsp);
1323
1324	rcu_check_gp_kthread_starvation(rsp);
1325
1326	force_quiescent_state(rsp);  /* Kick them all. */
1327}
1328
1329static void print_cpu_stall(struct rcu_state *rsp)
1330{
1331	int cpu;
1332	unsigned long flags;
1333	struct rcu_node *rnp = rcu_get_root(rsp);
1334	long totqlen = 0;
1335
1336	/*
1337	 * OK, time to rat on ourselves...
1338	 * See Documentation/RCU/stallwarn.txt for info on how to debug
1339	 * RCU CPU stall warnings.
1340	 */
1341	pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1342	print_cpu_stall_info_begin();
1343	print_cpu_stall_info(rsp, smp_processor_id());
1344	print_cpu_stall_info_end();
1345	for_each_possible_cpu(cpu)
1346		totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1347	pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1348		jiffies - rsp->gp_start,
1349		(long)rsp->gpnum, (long)rsp->completed, totqlen);
1350
1351	rcu_check_gp_kthread_starvation(rsp);
1352
1353	rcu_dump_cpu_stacks(rsp);
1354
1355	raw_spin_lock_irqsave_rcu_node(rnp, flags);
1356	if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1357		WRITE_ONCE(rsp->jiffies_stall,
1358			   jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1359	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1360
1361	/*
1362	 * Attempt to revive the RCU machinery by forcing a context switch.
1363	 *
1364	 * A context switch would normally allow the RCU state machine to make
1365	 * progress and it could be we're stuck in kernel space without context
1366	 * switches for an entirely unreasonable amount of time.
1367	 */
1368	resched_cpu(smp_processor_id());
1369}
1370
1371static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1372{
1373	unsigned long completed;
1374	unsigned long gpnum;
1375	unsigned long gps;
1376	unsigned long j;
1377	unsigned long js;
1378	struct rcu_node *rnp;
1379
1380	if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
1381		return;
1382	j = jiffies;
1383
1384	/*
1385	 * Lots of memory barriers to reject false positives.
1386	 *
1387	 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1388	 * then rsp->gp_start, and finally rsp->completed.  These values
1389	 * are updated in the opposite order with memory barriers (or
1390	 * equivalent) during grace-period initialization and cleanup.
1391	 * Now, a false positive can occur if we get an new value of
1392	 * rsp->gp_start and a old value of rsp->jiffies_stall.  But given
1393	 * the memory barriers, the only way that this can happen is if one
1394	 * grace period ends and another starts between these two fetches.
1395	 * Detect this by comparing rsp->completed with the previous fetch
1396	 * from rsp->gpnum.
1397	 *
1398	 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1399	 * and rsp->gp_start suffice to forestall false positives.
1400	 */
1401	gpnum = READ_ONCE(rsp->gpnum);
1402	smp_rmb(); /* Pick up ->gpnum first... */
1403	js = READ_ONCE(rsp->jiffies_stall);
1404	smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1405	gps = READ_ONCE(rsp->gp_start);
1406	smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1407	completed = READ_ONCE(rsp->completed);
1408	if (ULONG_CMP_GE(completed, gpnum) ||
1409	    ULONG_CMP_LT(j, js) ||
1410	    ULONG_CMP_GE(gps, js))
1411		return; /* No stall or GP completed since entering function. */
1412	rnp = rdp->mynode;
1413	if (rcu_gp_in_progress(rsp) &&
1414	    (READ_ONCE(rnp->qsmask) & rdp->grpmask)) {
1415
1416		/* We haven't checked in, so go dump stack. */
1417		print_cpu_stall(rsp);
1418
1419	} else if (rcu_gp_in_progress(rsp) &&
1420		   ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1421
1422		/* They had a few time units to dump stack, so complain. */
1423		print_other_cpu_stall(rsp, gpnum);
1424	}
1425}
1426
1427/**
1428 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1429 *
1430 * Set the stall-warning timeout way off into the future, thus preventing
1431 * any RCU CPU stall-warning messages from appearing in the current set of
1432 * RCU grace periods.
1433 *
1434 * The caller must disable hard irqs.
1435 */
1436void rcu_cpu_stall_reset(void)
1437{
1438	struct rcu_state *rsp;
1439
1440	for_each_rcu_flavor(rsp)
1441		WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1442}
1443
1444/*
1445 * Initialize the specified rcu_data structure's default callback list
1446 * to empty.  The default callback list is the one that is not used by
1447 * no-callbacks CPUs.
1448 */
1449static void init_default_callback_list(struct rcu_data *rdp)
1450{
1451	int i;
1452
1453	rdp->nxtlist = NULL;
1454	for (i = 0; i < RCU_NEXT_SIZE; i++)
1455		rdp->nxttail[i] = &rdp->nxtlist;
1456}
1457
1458/*
1459 * Initialize the specified rcu_data structure's callback list to empty.
1460 */
1461static void init_callback_list(struct rcu_data *rdp)
1462{
1463	if (init_nocb_callback_list(rdp))
1464		return;
1465	init_default_callback_list(rdp);
1466}
1467
1468/*
1469 * Determine the value that ->completed will have at the end of the
1470 * next subsequent grace period.  This is used to tag callbacks so that
1471 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1472 * been dyntick-idle for an extended period with callbacks under the
1473 * influence of RCU_FAST_NO_HZ.
1474 *
1475 * The caller must hold rnp->lock with interrupts disabled.
1476 */
1477static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1478				       struct rcu_node *rnp)
1479{
1480	/*
1481	 * If RCU is idle, we just wait for the next grace period.
1482	 * But we can only be sure that RCU is idle if we are looking
1483	 * at the root rcu_node structure -- otherwise, a new grace
1484	 * period might have started, but just not yet gotten around
1485	 * to initializing the current non-root rcu_node structure.
1486	 */
1487	if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1488		return rnp->completed + 1;
1489
1490	/*
1491	 * Otherwise, wait for a possible partial grace period and
1492	 * then the subsequent full grace period.
1493	 */
1494	return rnp->completed + 2;
1495}
1496
1497/*
1498 * Trace-event helper function for rcu_start_future_gp() and
1499 * rcu_nocb_wait_gp().
1500 */
1501static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1502				unsigned long c, const char *s)
1503{
1504	trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1505				      rnp->completed, c, rnp->level,
1506				      rnp->grplo, rnp->grphi, s);
1507}
1508
1509/*
1510 * Start some future grace period, as needed to handle newly arrived
 
 
 
 
 
1511 * callbacks.  The required future grace periods are recorded in each
1512 * rcu_node structure's ->need_future_gp field.  Returns true if there
1513 * is reason to awaken the grace-period kthread.
1514 *
1515 * The caller must hold the specified rcu_node structure's ->lock.
 
 
 
1516 */
1517static bool __maybe_unused
1518rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1519		    unsigned long *c_out)
1520{
1521	unsigned long c;
1522	int i;
1523	bool ret = false;
1524	struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1525
1526	/*
1527	 * Pick up grace-period number for new callbacks.  If this
1528	 * grace period is already marked as needed, return to the caller.
1529	 */
1530	c = rcu_cbs_completed(rdp->rsp, rnp);
1531	trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1532	if (rnp->need_future_gp[c & 0x1]) {
1533		trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1534		goto out;
1535	}
1536
1537	/*
1538	 * If either this rcu_node structure or the root rcu_node structure
1539	 * believe that a grace period is in progress, then we must wait
1540	 * for the one following, which is in "c".  Because our request
1541	 * will be noticed at the end of the current grace period, we don't
1542	 * need to explicitly start one.  We only do the lockless check
1543	 * of rnp_root's fields if the current rcu_node structure thinks
1544	 * there is no grace period in flight, and because we hold rnp->lock,
1545	 * the only possible change is when rnp_root's two fields are
1546	 * equal, in which case rnp_root->gpnum might be concurrently
1547	 * incremented.  But that is OK, as it will just result in our
1548	 * doing some extra useless work.
1549	 */
1550	if (rnp->gpnum != rnp->completed ||
1551	    READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) {
1552		rnp->need_future_gp[c & 0x1]++;
1553		trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1554		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1555	}
1556
1557	/*
1558	 * There might be no grace period in progress.  If we don't already
1559	 * hold it, acquire the root rcu_node structure's lock in order to
1560	 * start one (if needed).
1561	 */
1562	if (rnp != rnp_root)
1563		raw_spin_lock_rcu_node(rnp_root);
1564
1565	/*
1566	 * Get a new grace-period number.  If there really is no grace
1567	 * period in progress, it will be smaller than the one we obtained
1568	 * earlier.  Adjust callbacks as needed.  Note that even no-CBs
1569	 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1570	 */
1571	c = rcu_cbs_completed(rdp->rsp, rnp_root);
1572	for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1573		if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1574			rdp->nxtcompleted[i] = c;
1575
1576	/*
1577	 * If the needed for the required grace period is already
1578	 * recorded, trace and leave.
1579	 */
1580	if (rnp_root->need_future_gp[c & 0x1]) {
1581		trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1582		goto unlock_out;
1583	}
1584
1585	/* Record the need for the future grace period. */
1586	rnp_root->need_future_gp[c & 0x1]++;
1587
1588	/* If a grace period is not already in progress, start one. */
1589	if (rnp_root->gpnum != rnp_root->completed) {
1590		trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1591	} else {
1592		trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1593		ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1594	}
 
 
1595unlock_out:
1596	if (rnp != rnp_root)
1597		raw_spin_unlock_rcu_node(rnp_root);
1598out:
1599	if (c_out != NULL)
1600		*c_out = c;
 
 
1601	return ret;
1602}
1603
1604/*
1605 * Clean up any old requests for the just-ended grace period.  Also return
1606 * whether any additional grace periods have been requested.  Also invoke
1607 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1608 * waiting for this grace period to complete.
1609 */
1610static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1611{
1612	int c = rnp->completed;
1613	int needmore;
1614	struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1615
1616	rnp->need_future_gp[c & 0x1] = 0;
1617	needmore = rnp->need_future_gp[(c + 1) & 0x1];
1618	trace_rcu_future_gp(rnp, rdp, c,
1619			    needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1620	return needmore;
1621}
1622
1623/*
1624 * Awaken the grace-period kthread for the specified flavor of RCU.
1625 * Don't do a self-awaken, and don't bother awakening when there is
1626 * nothing for the grace-period kthread to do (as in several CPUs
1627 * raced to awaken, and we lost), and finally don't try to awaken
1628 * a kthread that has not yet been created.
1629 */
1630static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1631{
1632	if (current == rsp->gp_kthread ||
1633	    !READ_ONCE(rsp->gp_flags) ||
1634	    !rsp->gp_kthread)
 
 
 
 
 
 
 
 
 
1635		return;
1636	swake_up(&rsp->gp_wq);
 
 
1637}
1638
1639/*
1640 * If there is room, assign a ->completed number to any callbacks on
1641 * this CPU that have not already been assigned.  Also accelerate any
1642 * callbacks that were previously assigned a ->completed number that has
1643 * since proven to be too conservative, which can happen if callbacks get
1644 * assigned a ->completed number while RCU is idle, but with reference to
1645 * a non-root rcu_node structure.  This function is idempotent, so it does
1646 * not hurt to call it repeatedly.  Returns an flag saying that we should
1647 * awaken the RCU grace-period kthread.
1648 *
1649 * The caller must hold rnp->lock with interrupts disabled.
1650 */
1651static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1652			       struct rcu_data *rdp)
1653{
1654	unsigned long c;
1655	int i;
1656	bool ret;
1657
1658	/* If the CPU has no callbacks, nothing to do. */
1659	if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1660		return false;
1661
1662	/*
1663	 * Starting from the sublist containing the callbacks most
1664	 * recently assigned a ->completed number and working down, find the
1665	 * first sublist that is not assignable to an upcoming grace period.
1666	 * Such a sublist has something in it (first two tests) and has
1667	 * a ->completed number assigned that will complete sooner than
1668	 * the ->completed number for newly arrived callbacks (last test).
1669	 *
1670	 * The key point is that any later sublist can be assigned the
1671	 * same ->completed number as the newly arrived callbacks, which
1672	 * means that the callbacks in any of these later sublist can be
1673	 * grouped into a single sublist, whether or not they have already
1674	 * been assigned a ->completed number.
1675	 */
1676	c = rcu_cbs_completed(rsp, rnp);
1677	for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1678		if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1679		    !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1680			break;
1681
1682	/*
1683	 * If there are no sublist for unassigned callbacks, leave.
1684	 * At the same time, advance "i" one sublist, so that "i" will
1685	 * index into the sublist where all the remaining callbacks should
1686	 * be grouped into.
1687	 */
1688	if (++i >= RCU_NEXT_TAIL)
1689		return false;
1690
1691	/*
1692	 * Assign all subsequent callbacks' ->completed number to the next
1693	 * full grace period and group them all in the sublist initially
1694	 * indexed by "i".
1695	 */
1696	for (; i <= RCU_NEXT_TAIL; i++) {
1697		rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1698		rdp->nxtcompleted[i] = c;
1699	}
1700	/* Record any needed additional grace periods. */
1701	ret = rcu_start_future_gp(rnp, rdp, NULL);
 
 
1702
1703	/* Trace depending on how much we were able to accelerate. */
1704	if (!*rdp->nxttail[RCU_WAIT_TAIL])
1705		trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1706	else
1707		trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1708	return ret;
1709}
1710
1711/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1712 * Move any callbacks whose grace period has completed to the
1713 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1714 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1715 * sublist.  This function is idempotent, so it does not hurt to
1716 * invoke it repeatedly.  As long as it is not invoked -too- often...
1717 * Returns true if the RCU grace-period kthread needs to be awakened.
1718 *
1719 * The caller must hold rnp->lock with interrupts disabled.
1720 */
1721static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1722			    struct rcu_data *rdp)
1723{
1724	int i, j;
 
1725
1726	/* If the CPU has no callbacks, nothing to do. */
1727	if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1728		return false;
1729
1730	/*
1731	 * Find all callbacks whose ->completed numbers indicate that they
1732	 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1733	 */
1734	for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1735		if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1736			break;
1737		rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1738	}
1739	/* Clean up any sublist tail pointers that were misordered above. */
1740	for (j = RCU_WAIT_TAIL; j < i; j++)
1741		rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1742
1743	/* Copy down callbacks to fill in empty sublists. */
1744	for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1745		if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1746			break;
1747		rdp->nxttail[j] = rdp->nxttail[i];
1748		rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1749	}
1750
1751	/* Classify any remaining callbacks. */
1752	return rcu_accelerate_cbs(rsp, rnp, rdp);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1753}
1754
1755/*
1756 * Update CPU-local rcu_data state to record the beginnings and ends of
1757 * grace periods.  The caller must hold the ->lock of the leaf rcu_node
1758 * structure corresponding to the current CPU, and must have irqs disabled.
1759 * Returns true if the grace-period kthread needs to be awakened.
1760 */
1761static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1762			      struct rcu_data *rdp)
1763{
1764	bool ret;
 
 
 
1765
1766	/* Handle the ends of any preceding grace periods first. */
1767	if (rdp->completed == rnp->completed &&
1768	    !unlikely(READ_ONCE(rdp->gpwrap))) {
1769
1770		/* No grace period end, so just accelerate recent callbacks. */
1771		ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1772
 
 
 
 
 
 
 
1773	} else {
1774
1775		/* Advance callbacks. */
1776		ret = rcu_advance_cbs(rsp, rnp, rdp);
1777
1778		/* Remember that we saw this grace-period completion. */
1779		rdp->completed = rnp->completed;
1780		trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1781	}
1782
1783	if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) {
 
 
1784		/*
1785		 * If the current grace period is waiting for this CPU,
1786		 * set up to detect a quiescent state, otherwise don't
1787		 * go looking for one.
1788		 */
1789		rdp->gpnum = rnp->gpnum;
1790		trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1791		rdp->cpu_no_qs.b.norm = true;
1792		rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr);
1793		rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask);
1794		zero_cpu_stall_ticks(rdp);
1795		WRITE_ONCE(rdp->gpwrap, false);
1796	}
 
 
 
 
 
1797	return ret;
1798}
1799
1800static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1801{
1802	unsigned long flags;
1803	bool needwake;
1804	struct rcu_node *rnp;
1805
1806	local_irq_save(flags);
1807	rnp = rdp->mynode;
1808	if ((rdp->gpnum == READ_ONCE(rnp->gpnum) &&
1809	     rdp->completed == READ_ONCE(rnp->completed) &&
1810	     !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1811	    !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1812		local_irq_restore(flags);
1813		return;
1814	}
1815	needwake = __note_gp_changes(rsp, rnp, rdp);
1816	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1817	if (needwake)
1818		rcu_gp_kthread_wake(rsp);
1819}
1820
1821static void rcu_gp_slow(struct rcu_state *rsp, int delay)
1822{
1823	if (delay > 0 &&
1824	    !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1825		schedule_timeout_uninterruptible(delay);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1826}
1827
1828/*
1829 * Initialize a new grace period.  Return false if no grace period required.
1830 */
1831static bool rcu_gp_init(struct rcu_state *rsp)
1832{
 
1833	unsigned long oldmask;
 
1834	struct rcu_data *rdp;
1835	struct rcu_node *rnp = rcu_get_root(rsp);
1836
1837	WRITE_ONCE(rsp->gp_activity, jiffies);
1838	raw_spin_lock_irq_rcu_node(rnp);
1839	if (!READ_ONCE(rsp->gp_flags)) {
1840		/* Spurious wakeup, tell caller to go back to sleep.  */
1841		raw_spin_unlock_irq_rcu_node(rnp);
1842		return false;
1843	}
1844	WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
1845
1846	if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1847		/*
1848		 * Grace period already in progress, don't start another.
1849		 * Not supposed to be able to happen.
1850		 */
1851		raw_spin_unlock_irq_rcu_node(rnp);
1852		return false;
1853	}
1854
1855	/* Advance to a new grace period and initialize state. */
1856	record_gp_stall_check_time(rsp);
1857	/* Record GP times before starting GP, hence smp_store_release(). */
1858	smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1859	trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
 
1860	raw_spin_unlock_irq_rcu_node(rnp);
1861
1862	/*
1863	 * Apply per-leaf buffered online and offline operations to the
1864	 * rcu_node tree.  Note that this new grace period need not wait
1865	 * for subsequent online CPUs, and that quiescent-state forcing
1866	 * will handle subsequent offline CPUs.
1867	 */
1868	rcu_for_each_leaf_node(rsp, rnp) {
1869		rcu_gp_slow(rsp, gp_preinit_delay);
 
1870		raw_spin_lock_irq_rcu_node(rnp);
1871		if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1872		    !rnp->wait_blkd_tasks) {
1873			/* Nothing to do on this leaf rcu_node structure. */
1874			raw_spin_unlock_irq_rcu_node(rnp);
 
1875			continue;
1876		}
1877
1878		/* Record old state, apply changes to ->qsmaskinit field. */
1879		oldmask = rnp->qsmaskinit;
1880		rnp->qsmaskinit = rnp->qsmaskinitnext;
1881
1882		/* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1883		if (!oldmask != !rnp->qsmaskinit) {
1884			if (!oldmask) /* First online CPU for this rcu_node. */
1885				rcu_init_new_rnp(rnp);
1886			else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */
1887				rnp->wait_blkd_tasks = true;
1888			else /* Last offline CPU and can propagate. */
 
1889				rcu_cleanup_dead_rnp(rnp);
 
1890		}
1891
1892		/*
1893		 * If all waited-on tasks from prior grace period are
1894		 * done, and if all this rcu_node structure's CPUs are
1895		 * still offline, propagate up the rcu_node tree and
1896		 * clear ->wait_blkd_tasks.  Otherwise, if one of this
1897		 * rcu_node structure's CPUs has since come back online,
1898		 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1899		 * checks for this, so just call it unconditionally).
1900		 */
1901		if (rnp->wait_blkd_tasks &&
1902		    (!rcu_preempt_has_tasks(rnp) ||
1903		     rnp->qsmaskinit)) {
1904			rnp->wait_blkd_tasks = false;
1905			rcu_cleanup_dead_rnp(rnp);
 
1906		}
1907
1908		raw_spin_unlock_irq_rcu_node(rnp);
 
1909	}
 
1910
1911	/*
1912	 * Set the quiescent-state-needed bits in all the rcu_node
1913	 * structures for all currently online CPUs in breadth-first order,
1914	 * starting from the root rcu_node structure, relying on the layout
1915	 * of the tree within the rsp->node[] array.  Note that other CPUs
1916	 * will access only the leaves of the hierarchy, thus seeing that no
1917	 * grace period is in progress, at least until the corresponding
1918	 * leaf node has been initialized.  In addition, we have excluded
1919	 * CPU-hotplug operations.
1920	 *
1921	 * The grace period cannot complete until the initialization
1922	 * process finishes, because this kthread handles both.
1923	 */
1924	rcu_for_each_node_breadth_first(rsp, rnp) {
1925		rcu_gp_slow(rsp, gp_init_delay);
1926		raw_spin_lock_irq_rcu_node(rnp);
1927		rdp = this_cpu_ptr(rsp->rda);
 
1928		rcu_preempt_check_blocked_tasks(rnp);
1929		rnp->qsmask = rnp->qsmaskinit;
1930		WRITE_ONCE(rnp->gpnum, rsp->gpnum);
1931		if (WARN_ON_ONCE(rnp->completed != rsp->completed))
1932			WRITE_ONCE(rnp->completed, rsp->completed);
1933		if (rnp == rdp->mynode)
1934			(void)__note_gp_changes(rsp, rnp, rdp);
1935		rcu_preempt_boost_start_gp(rnp);
1936		trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1937					    rnp->level, rnp->grplo,
1938					    rnp->grphi, rnp->qsmask);
1939		raw_spin_unlock_irq_rcu_node(rnp);
1940		cond_resched_rcu_qs();
1941		WRITE_ONCE(rsp->gp_activity, jiffies);
 
 
 
 
 
 
1942	}
1943
1944	return true;
1945}
1946
1947/*
1948 * Helper function for wait_event_interruptible_timeout() wakeup
1949 * at force-quiescent-state time.
1950 */
1951static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
1952{
1953	struct rcu_node *rnp = rcu_get_root(rsp);
 
 
 
 
1954
1955	/* Someone like call_rcu() requested a force-quiescent-state scan. */
1956	*gfp = READ_ONCE(rsp->gp_flags);
1957	if (*gfp & RCU_GP_FLAG_FQS)
1958		return true;
1959
1960	/* The current grace period has completed. */
1961	if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1962		return true;
1963
1964	return false;
1965}
1966
1967/*
1968 * Do one round of quiescent-state forcing.
1969 */
1970static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time)
1971{
1972	bool isidle = false;
1973	unsigned long maxj;
1974	struct rcu_node *rnp = rcu_get_root(rsp);
1975
1976	WRITE_ONCE(rsp->gp_activity, jiffies);
1977	rsp->n_force_qs++;
1978	if (first_time) {
1979		/* Collect dyntick-idle snapshots. */
1980		if (is_sysidle_rcu_state(rsp)) {
1981			isidle = true;
1982			maxj = jiffies - ULONG_MAX / 4;
1983		}
1984		force_qs_rnp(rsp, dyntick_save_progress_counter,
1985			     &isidle, &maxj);
1986		rcu_sysidle_report_gp(rsp, isidle, maxj);
1987	} else {
1988		/* Handle dyntick-idle and offline CPUs. */
1989		isidle = true;
1990		force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1991	}
1992	/* Clear flag to prevent immediate re-entry. */
1993	if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1994		raw_spin_lock_irq_rcu_node(rnp);
1995		WRITE_ONCE(rsp->gp_flags,
1996			   READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
1997		raw_spin_unlock_irq_rcu_node(rnp);
1998	}
1999}
2000
2001/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2002 * Clean up after the old grace period.
2003 */
2004static void rcu_gp_cleanup(struct rcu_state *rsp)
2005{
2006	unsigned long gp_duration;
2007	bool needgp = false;
2008	int nocb = 0;
 
 
2009	struct rcu_data *rdp;
2010	struct rcu_node *rnp = rcu_get_root(rsp);
2011	struct swait_queue_head *sq;
2012
2013	WRITE_ONCE(rsp->gp_activity, jiffies);
2014	raw_spin_lock_irq_rcu_node(rnp);
2015	gp_duration = jiffies - rsp->gp_start;
2016	if (gp_duration > rsp->gp_max)
2017		rsp->gp_max = gp_duration;
 
2018
2019	/*
2020	 * We know the grace period is complete, but to everyone else
2021	 * it appears to still be ongoing.  But it is also the case
2022	 * that to everyone else it looks like there is nothing that
2023	 * they can do to advance the grace period.  It is therefore
2024	 * safe for us to drop the lock in order to mark the grace
2025	 * period as completed in all of the rcu_node structures.
2026	 */
2027	raw_spin_unlock_irq_rcu_node(rnp);
2028
2029	/*
2030	 * Propagate new ->completed value to rcu_node structures so
2031	 * that other CPUs don't have to wait until the start of the next
2032	 * grace period to process their callbacks.  This also avoids
2033	 * some nasty RCU grace-period initialization races by forcing
2034	 * the end of the current grace period to be completely recorded in
2035	 * all of the rcu_node structures before the beginning of the next
2036	 * grace period is recorded in any of the rcu_node structures.
2037	 */
2038	rcu_for_each_node_breadth_first(rsp, rnp) {
 
 
2039		raw_spin_lock_irq_rcu_node(rnp);
2040		WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
 
2041		WARN_ON_ONCE(rnp->qsmask);
2042		WRITE_ONCE(rnp->completed, rsp->gpnum);
2043		rdp = this_cpu_ptr(rsp->rda);
2044		if (rnp == rdp->mynode)
2045			needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2046		/* smp_mb() provided by prior unlock-lock pair. */
2047		nocb += rcu_future_gp_cleanup(rsp, rnp);
 
 
 
 
 
 
2048		sq = rcu_nocb_gp_get(rnp);
2049		raw_spin_unlock_irq_rcu_node(rnp);
2050		rcu_nocb_gp_cleanup(sq);
2051		cond_resched_rcu_qs();
2052		WRITE_ONCE(rsp->gp_activity, jiffies);
2053		rcu_gp_slow(rsp, gp_cleanup_delay);
2054	}
2055	rnp = rcu_get_root(rsp);
2056	raw_spin_lock_irq_rcu_node(rnp); /* Order GP before ->completed update. */
2057	rcu_nocb_gp_set(rnp, nocb);
2058
2059	/* Declare grace period done. */
2060	WRITE_ONCE(rsp->completed, rsp->gpnum);
2061	trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
2062	rsp->gp_state = RCU_GP_IDLE;
2063	rdp = this_cpu_ptr(rsp->rda);
 
 
 
 
 
 
2064	/* Advance CBs to reduce false positives below. */
2065	needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
2066	if (needgp || cpu_needs_another_gp(rsp, rdp)) {
2067		WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2068		trace_rcu_grace_period(rsp->name,
2069				       READ_ONCE(rsp->gpnum),
 
 
2070				       TPS("newreq"));
 
 
 
2071	}
2072	raw_spin_unlock_irq_rcu_node(rnp);
2073}
2074
2075/*
2076 * Body of kthread that handles grace periods.
2077 */
2078static int __noreturn rcu_gp_kthread(void *arg)
2079{
2080	bool first_gp_fqs;
2081	int gf;
2082	unsigned long j;
2083	int ret;
2084	struct rcu_state *rsp = arg;
2085	struct rcu_node *rnp = rcu_get_root(rsp);
2086
2087	rcu_bind_gp_kthread();
2088	for (;;) {
2089
2090		/* Handle grace-period start. */
2091		for (;;) {
2092			trace_rcu_grace_period(rsp->name,
2093					       READ_ONCE(rsp->gpnum),
2094					       TPS("reqwait"));
2095			rsp->gp_state = RCU_GP_WAIT_GPS;
2096			swait_event_interruptible(rsp->gp_wq,
2097						 READ_ONCE(rsp->gp_flags) &
2098						 RCU_GP_FLAG_INIT);
2099			rsp->gp_state = RCU_GP_DONE_GPS;
 
2100			/* Locking provides needed memory barrier. */
2101			if (rcu_gp_init(rsp))
2102				break;
2103			cond_resched_rcu_qs();
2104			WRITE_ONCE(rsp->gp_activity, jiffies);
2105			WARN_ON(signal_pending(current));
2106			trace_rcu_grace_period(rsp->name,
2107					       READ_ONCE(rsp->gpnum),
2108					       TPS("reqwaitsig"));
2109		}
2110
2111		/* Handle quiescent-state forcing. */
2112		first_gp_fqs = true;
2113		j = jiffies_till_first_fqs;
2114		if (j > HZ) {
2115			j = HZ;
2116			jiffies_till_first_fqs = HZ;
2117		}
2118		ret = 0;
2119		for (;;) {
2120			if (!ret)
2121				rsp->jiffies_force_qs = jiffies + j;
2122			trace_rcu_grace_period(rsp->name,
2123					       READ_ONCE(rsp->gpnum),
2124					       TPS("fqswait"));
2125			rsp->gp_state = RCU_GP_WAIT_FQS;
2126			ret = swait_event_interruptible_timeout(rsp->gp_wq,
2127					rcu_gp_fqs_check_wake(rsp, &gf), j);
2128			rsp->gp_state = RCU_GP_DOING_FQS;
2129			/* Locking provides needed memory barriers. */
2130			/* If grace period done, leave loop. */
2131			if (!READ_ONCE(rnp->qsmask) &&
2132			    !rcu_preempt_blocked_readers_cgp(rnp))
2133				break;
2134			/* If time for quiescent-state forcing, do it. */
2135			if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2136			    (gf & RCU_GP_FLAG_FQS)) {
2137				trace_rcu_grace_period(rsp->name,
2138						       READ_ONCE(rsp->gpnum),
2139						       TPS("fqsstart"));
2140				rcu_gp_fqs(rsp, first_gp_fqs);
2141				first_gp_fqs = false;
2142				trace_rcu_grace_period(rsp->name,
2143						       READ_ONCE(rsp->gpnum),
2144						       TPS("fqsend"));
2145				cond_resched_rcu_qs();
2146				WRITE_ONCE(rsp->gp_activity, jiffies);
2147			} else {
2148				/* Deal with stray signal. */
2149				cond_resched_rcu_qs();
2150				WRITE_ONCE(rsp->gp_activity, jiffies);
2151				WARN_ON(signal_pending(current));
2152				trace_rcu_grace_period(rsp->name,
2153						       READ_ONCE(rsp->gpnum),
2154						       TPS("fqswaitsig"));
2155			}
2156			j = jiffies_till_next_fqs;
2157			if (j > HZ) {
2158				j = HZ;
2159				jiffies_till_next_fqs = HZ;
2160			} else if (j < 1) {
2161				j = 1;
2162				jiffies_till_next_fqs = 1;
2163			}
2164		}
2165
2166		/* Handle grace-period end. */
2167		rsp->gp_state = RCU_GP_CLEANUP;
2168		rcu_gp_cleanup(rsp);
2169		rsp->gp_state = RCU_GP_CLEANED;
2170	}
2171}
2172
2173/*
2174 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2175 * in preparation for detecting the next grace period.  The caller must hold
2176 * the root node's ->lock and hard irqs must be disabled.
2177 *
2178 * Note that it is legal for a dying CPU (which is marked as offline) to
2179 * invoke this function.  This can happen when the dying CPU reports its
2180 * quiescent state.
2181 *
2182 * Returns true if the grace-period kthread must be awakened.
2183 */
2184static bool
2185rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
2186		      struct rcu_data *rdp)
2187{
2188	if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
2189		/*
2190		 * Either we have not yet spawned the grace-period
2191		 * task, this CPU does not need another grace period,
2192		 * or a grace period is already in progress.
2193		 * Either way, don't start a new grace period.
2194		 */
2195		return false;
2196	}
2197	WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2198	trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum),
2199			       TPS("newreq"));
2200
2201	/*
2202	 * We can't do wakeups while holding the rnp->lock, as that
2203	 * could cause possible deadlocks with the rq->lock. Defer
2204	 * the wakeup to our caller.
2205	 */
2206	return true;
2207}
2208
2209/*
2210 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2211 * callbacks.  Note that rcu_start_gp_advanced() cannot do this because it
2212 * is invoked indirectly from rcu_advance_cbs(), which would result in
2213 * endless recursion -- or would do so if it wasn't for the self-deadlock
2214 * that is encountered beforehand.
2215 *
2216 * Returns true if the grace-period kthread needs to be awakened.
2217 */
2218static bool rcu_start_gp(struct rcu_state *rsp)
2219{
2220	struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
2221	struct rcu_node *rnp = rcu_get_root(rsp);
2222	bool ret = false;
2223
2224	/*
2225	 * If there is no grace period in progress right now, any
2226	 * callbacks we have up to this point will be satisfied by the
2227	 * next grace period.  Also, advancing the callbacks reduces the
2228	 * probability of false positives from cpu_needs_another_gp()
2229	 * resulting in pointless grace periods.  So, advance callbacks
2230	 * then start the grace period!
2231	 */
2232	ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
2233	ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
2234	return ret;
2235}
2236
2237/*
2238 * Report a full set of quiescent states to the specified rcu_state data
2239 * structure.  Invoke rcu_gp_kthread_wake() to awaken the grace-period
2240 * kthread if another grace period is required.  Whether we wake
2241 * the grace-period kthread or it awakens itself for the next round
2242 * of quiescent-state forcing, that kthread will clean up after the
2243 * just-completed grace period.  Note that the caller must hold rnp->lock,
2244 * which is released before return.
2245 */
2246static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2247	__releases(rcu_get_root(rsp)->lock)
2248{
2249	WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2250	WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2251	raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2252	swake_up(&rsp->gp_wq);  /* Memory barrier implied by swake_up() path. */
 
 
2253}
2254
2255/*
2256 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2257 * Allows quiescent states for a group of CPUs to be reported at one go
2258 * to the specified rcu_node structure, though all the CPUs in the group
2259 * must be represented by the same rcu_node structure (which need not be a
2260 * leaf rcu_node structure, though it often will be).  The gps parameter
2261 * is the grace-period snapshot, which means that the quiescent states
2262 * are valid only if rnp->gpnum is equal to gps.  That structure's lock
2263 * must be held upon entry, and it is released before return.
 
 
 
 
2264 */
2265static void
2266rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2267		  struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2268	__releases(rnp->lock)
2269{
2270	unsigned long oldmask = 0;
2271	struct rcu_node *rnp_c;
2272
 
 
2273	/* Walk up the rcu_node hierarchy. */
2274	for (;;) {
2275		if (!(rnp->qsmask & mask) || rnp->gpnum != gps) {
2276
2277			/*
2278			 * Our bit has already been cleared, or the
2279			 * relevant grace period is already over, so done.
2280			 */
2281			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2282			return;
2283		}
2284		WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2285		rnp->qsmask &= ~mask;
2286		trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
 
 
2287						 mask, rnp->qsmask, rnp->level,
2288						 rnp->grplo, rnp->grphi,
2289						 !!rnp->gp_tasks);
2290		if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2291
2292			/* Other bits still set at this level, so done. */
2293			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2294			return;
2295		}
 
2296		mask = rnp->grpmask;
2297		if (rnp->parent == NULL) {
2298
2299			/* No more levels.  Exit loop holding root lock. */
2300
2301			break;
2302		}
2303		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2304		rnp_c = rnp;
2305		rnp = rnp->parent;
2306		raw_spin_lock_irqsave_rcu_node(rnp, flags);
2307		oldmask = rnp_c->qsmask;
2308	}
2309
2310	/*
2311	 * Get here if we are the last CPU to pass through a quiescent
2312	 * state for this grace period.  Invoke rcu_report_qs_rsp()
2313	 * to clean up and start the next grace period if one is needed.
2314	 */
2315	rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2316}
2317
2318/*
2319 * Record a quiescent state for all tasks that were previously queued
2320 * on the specified rcu_node structure and that were blocking the current
2321 * RCU grace period.  The caller must hold the specified rnp->lock with
2322 * irqs disabled, and this lock is released upon return, but irqs remain
2323 * disabled.
2324 */
2325static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2326				      struct rcu_node *rnp, unsigned long flags)
2327	__releases(rnp->lock)
2328{
2329	unsigned long gps;
2330	unsigned long mask;
2331	struct rcu_node *rnp_p;
2332
2333	if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
2334	    rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
 
 
2335		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2336		return;  /* Still need more quiescent states! */
2337	}
2338
 
2339	rnp_p = rnp->parent;
2340	if (rnp_p == NULL) {
2341		/*
2342		 * Only one rcu_node structure in the tree, so don't
2343		 * try to report up to its nonexistent parent!
2344		 */
2345		rcu_report_qs_rsp(rsp, flags);
2346		return;
2347	}
2348
2349	/* Report up the rest of the hierarchy, tracking current ->gpnum. */
2350	gps = rnp->gpnum;
2351	mask = rnp->grpmask;
2352	raw_spin_unlock_rcu_node(rnp);	/* irqs remain disabled. */
2353	raw_spin_lock_rcu_node(rnp_p);	/* irqs already disabled. */
2354	rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2355}
2356
2357/*
2358 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2359 * structure.  This must be called from the specified CPU.
2360 */
2361static void
2362rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2363{
2364	unsigned long flags;
2365	unsigned long mask;
2366	bool needwake;
 
 
2367	struct rcu_node *rnp;
2368
2369	rnp = rdp->mynode;
2370	raw_spin_lock_irqsave_rcu_node(rnp, flags);
2371	if ((rdp->cpu_no_qs.b.norm &&
2372	     rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) ||
2373	    rdp->gpnum != rnp->gpnum || rnp->completed == rnp->gpnum ||
2374	    rdp->gpwrap) {
2375
2376		/*
2377		 * The grace period in which this quiescent state was
2378		 * recorded has ended, so don't report it upwards.
2379		 * We will instead need a new quiescent state that lies
2380		 * within the current grace period.
2381		 */
2382		rdp->cpu_no_qs.b.norm = true;	/* need qs for new gp. */
2383		rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr);
2384		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2385		return;
2386	}
2387	mask = rdp->grpmask;
 
 
2388	if ((rnp->qsmask & mask) == 0) {
2389		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2390	} else {
2391		rdp->core_needs_qs = false;
2392
2393		/*
2394		 * This GP can't end until cpu checks in, so all of our
2395		 * callbacks can be processed during the next GP.
2396		 */
2397		needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
 
2398
2399		rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
 
2400		/* ^^^ Released rnp->lock */
2401		if (needwake)
2402			rcu_gp_kthread_wake(rsp);
2403	}
2404}
2405
2406/*
2407 * Check to see if there is a new grace period of which this CPU
2408 * is not yet aware, and if so, set up local rcu_data state for it.
2409 * Otherwise, see if this CPU has just passed through its first
2410 * quiescent state for this grace period, and record that fact if so.
2411 */
2412static void
2413rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2414{
2415	/* Check for grace-period ends and beginnings. */
2416	note_gp_changes(rsp, rdp);
2417
2418	/*
2419	 * Does this CPU still need to do its part for current grace period?
2420	 * If no, return and let the other CPUs do their part as well.
2421	 */
2422	if (!rdp->core_needs_qs)
2423		return;
2424
2425	/*
2426	 * Was there a quiescent state since the beginning of the grace
2427	 * period? If no, then exit and wait for the next call.
2428	 */
2429	if (rdp->cpu_no_qs.b.norm &&
2430	    rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr))
2431		return;
2432
2433	/*
2434	 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2435	 * judge of that).
2436	 */
2437	rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2438}
2439
2440/*
2441 * Send the specified CPU's RCU callbacks to the orphanage.  The
2442 * specified CPU must be offline, and the caller must hold the
2443 * ->orphan_lock.
2444 */
2445static void
2446rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
2447			  struct rcu_node *rnp, struct rcu_data *rdp)
2448{
2449	/* No-CBs CPUs do not have orphanable callbacks. */
2450	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu))
2451		return;
2452
2453	/*
2454	 * Orphan the callbacks.  First adjust the counts.  This is safe
2455	 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2456	 * cannot be running now.  Thus no memory barrier is required.
2457	 */
2458	if (rdp->nxtlist != NULL) {
2459		rsp->qlen_lazy += rdp->qlen_lazy;
2460		rsp->qlen += rdp->qlen;
2461		rdp->n_cbs_orphaned += rdp->qlen;
2462		rdp->qlen_lazy = 0;
2463		WRITE_ONCE(rdp->qlen, 0);
2464	}
2465
2466	/*
2467	 * Next, move those callbacks still needing a grace period to
2468	 * the orphanage, where some other CPU will pick them up.
2469	 * Some of the callbacks might have gone partway through a grace
2470	 * period, but that is too bad.  They get to start over because we
2471	 * cannot assume that grace periods are synchronized across CPUs.
2472	 * We don't bother updating the ->nxttail[] array yet, instead
2473	 * we just reset the whole thing later on.
2474	 */
2475	if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
2476		*rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
2477		rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
2478		*rdp->nxttail[RCU_DONE_TAIL] = NULL;
2479	}
2480
2481	/*
2482	 * Then move the ready-to-invoke callbacks to the orphanage,
2483	 * where some other CPU will pick them up.  These will not be
2484	 * required to pass though another grace period: They are done.
2485	 */
2486	if (rdp->nxtlist != NULL) {
2487		*rsp->orphan_donetail = rdp->nxtlist;
2488		rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
2489	}
2490
2491	/*
2492	 * Finally, initialize the rcu_data structure's list to empty and
2493	 * disallow further callbacks on this CPU.
2494	 */
2495	init_callback_list(rdp);
2496	rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2497}
2498
2499/*
2500 * Adopt the RCU callbacks from the specified rcu_state structure's
2501 * orphanage.  The caller must hold the ->orphan_lock.
2502 */
2503static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
2504{
2505	int i;
2506	struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2507
2508	/* No-CBs CPUs are handled specially. */
2509	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2510	    rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
2511		return;
2512
2513	/* Do the accounting first. */
2514	rdp->qlen_lazy += rsp->qlen_lazy;
2515	rdp->qlen += rsp->qlen;
2516	rdp->n_cbs_adopted += rsp->qlen;
2517	if (rsp->qlen_lazy != rsp->qlen)
2518		rcu_idle_count_callbacks_posted();
2519	rsp->qlen_lazy = 0;
2520	rsp->qlen = 0;
2521
2522	/*
2523	 * We do not need a memory barrier here because the only way we
2524	 * can get here if there is an rcu_barrier() in flight is if
2525	 * we are the task doing the rcu_barrier().
2526	 */
2527
2528	/* First adopt the ready-to-invoke callbacks. */
2529	if (rsp->orphan_donelist != NULL) {
2530		*rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
2531		*rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
2532		for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
2533			if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2534				rdp->nxttail[i] = rsp->orphan_donetail;
2535		rsp->orphan_donelist = NULL;
2536		rsp->orphan_donetail = &rsp->orphan_donelist;
2537	}
2538
2539	/* And then adopt the callbacks that still need a grace period. */
2540	if (rsp->orphan_nxtlist != NULL) {
2541		*rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
2542		rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
2543		rsp->orphan_nxtlist = NULL;
2544		rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2545	}
2546}
2547
2548/*
2549 * Trace the fact that this CPU is going offline.
 
2550 */
2551static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2552{
2553	RCU_TRACE(unsigned long mask);
2554	RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
2555	RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
2556
2557	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2558		return;
2559
2560	RCU_TRACE(mask = rdp->grpmask);
2561	trace_rcu_grace_period(rsp->name,
2562			       rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2563			       TPS("cpuofl"));
2564}
2565
2566/*
2567 * All CPUs for the specified rcu_node structure have gone offline,
2568 * and all tasks that were preempted within an RCU read-side critical
2569 * section while running on one of those CPUs have since exited their RCU
2570 * read-side critical section.  Some other CPU is reporting this fact with
2571 * the specified rcu_node structure's ->lock held and interrupts disabled.
2572 * This function therefore goes up the tree of rcu_node structures,
2573 * clearing the corresponding bits in the ->qsmaskinit fields.  Note that
2574 * the leaf rcu_node structure's ->qsmaskinit field has already been
2575 * updated
2576 *
2577 * This function does check that the specified rcu_node structure has
2578 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2579 * prematurely.  That said, invoking it after the fact will cost you
2580 * a needless lock acquisition.  So once it has done its work, don't
2581 * invoke it again.
2582 */
2583static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2584{
2585	long mask;
2586	struct rcu_node *rnp = rnp_leaf;
2587
 
2588	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2589	    rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
 
2590		return;
2591	for (;;) {
2592		mask = rnp->grpmask;
2593		rnp = rnp->parent;
2594		if (!rnp)
2595			break;
2596		raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2597		rnp->qsmaskinit &= ~mask;
2598		rnp->qsmask &= ~mask;
 
2599		if (rnp->qsmaskinit) {
2600			raw_spin_unlock_rcu_node(rnp);
2601			/* irqs remain disabled. */
2602			return;
2603		}
2604		raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2605	}
2606}
2607
2608/*
2609 * The CPU has been completely removed, and some other CPU is reporting
2610 * this fact from process context.  Do the remainder of the cleanup,
2611 * including orphaning the outgoing CPU's RCU callbacks, and also
2612 * adopting them.  There can only be one CPU hotplug operation at a time,
2613 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2614 */
2615static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2616{
2617	unsigned long flags;
2618	struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2619	struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
2620
2621	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2622		return;
2623
2624	/* Adjust any no-longer-needed kthreads. */
2625	rcu_boost_kthread_setaffinity(rnp, -1);
 
 
2626
2627	/* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2628	raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2629	rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2630	rcu_adopt_orphan_cbs(rsp, flags);
2631	raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags);
2632
2633	WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2634		  "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2635		  cpu, rdp->qlen, rdp->nxtlist);
2636}
2637
2638/*
2639 * Invoke any RCU callbacks that have made it to the end of their grace
2640 * period.  Thottle as specified by rdp->blimit.
2641 */
2642static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2643{
2644	unsigned long flags;
2645	struct rcu_head *next, *list, **tail;
2646	long bl, count, count_lazy;
2647	int i;
 
 
 
2648
2649	/* If no callbacks are ready, just return. */
2650	if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2651		trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2652		trace_rcu_batch_end(rsp->name, 0, !!READ_ONCE(rdp->nxtlist),
 
 
2653				    need_resched(), is_idle_task(current),
2654				    rcu_is_callbacks_kthread());
2655		return;
2656	}
2657
2658	/*
2659	 * Extract the list of ready callbacks, disabling to prevent
2660	 * races with call_rcu() from interrupt handlers.
 
2661	 */
2662	local_irq_save(flags);
 
2663	WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2664	bl = rdp->blimit;
2665	trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2666	list = rdp->nxtlist;
2667	rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2668	*rdp->nxttail[RCU_DONE_TAIL] = NULL;
2669	tail = rdp->nxttail[RCU_DONE_TAIL];
2670	for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2671		if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2672			rdp->nxttail[i] = &rdp->nxtlist;
2673	local_irq_restore(flags);
2674
2675	/* Invoke callbacks. */
2676	count = count_lazy = 0;
2677	while (list) {
2678		next = list->next;
2679		prefetch(next);
2680		debug_rcu_head_unqueue(list);
2681		if (__rcu_reclaim(rsp->name, list))
2682			count_lazy++;
2683		list = next;
2684		/* Stop only if limit reached and CPU has something to do. */
2685		if (++count >= bl &&
 
 
 
 
 
 
 
 
 
 
 
2686		    (need_resched() ||
2687		     (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2688			break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2689	}
2690
2691	local_irq_save(flags);
2692	trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2693			    is_idle_task(current),
2694			    rcu_is_callbacks_kthread());
2695
2696	/* Update count, and requeue any remaining callbacks. */
2697	if (list != NULL) {
2698		*tail = rdp->nxtlist;
2699		rdp->nxtlist = list;
2700		for (i = 0; i < RCU_NEXT_SIZE; i++)
2701			if (&rdp->nxtlist == rdp->nxttail[i])
2702				rdp->nxttail[i] = tail;
2703			else
2704				break;
2705	}
2706	smp_mb(); /* List handling before counting for rcu_barrier(). */
2707	rdp->qlen_lazy -= count_lazy;
2708	WRITE_ONCE(rdp->qlen, rdp->qlen - count);
2709	rdp->n_cbs_invoked += count;
 
 
 
 
 
 
 
2710
2711	/* Reinstate batch limit if we have worked down the excess. */
2712	if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
 
2713		rdp->blimit = blimit;
2714
2715	/* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2716	if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2717		rdp->qlen_last_fqs_check = 0;
2718		rdp->n_force_qs_snap = rsp->n_force_qs;
2719	} else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2720		rdp->qlen_last_fqs_check = rdp->qlen;
2721	WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2722
2723	local_irq_restore(flags);
 
 
 
 
 
 
 
 
2724
2725	/* Re-invoke RCU core processing if there are callbacks remaining. */
2726	if (cpu_has_callbacks_ready_to_invoke(rdp))
2727		invoke_rcu_core();
 
2728}
2729
2730/*
2731 * Check to see if this CPU is in a non-context-switch quiescent state
2732 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2733 * Also schedule RCU core processing.
2734 *
2735 * This function must be called from hardirq context.  It is normally
2736 * invoked from the scheduling-clock interrupt.  If rcu_pending returns
2737 * false, there is no point in invoking rcu_check_callbacks().
2738 */
2739void rcu_check_callbacks(int user)
2740{
2741	trace_rcu_utilization(TPS("Start scheduler-tick"));
2742	increment_cpu_stall_ticks();
2743	if (user || rcu_is_cpu_rrupt_from_idle()) {
2744
2745		/*
2746		 * Get here if this CPU took its interrupt from user
2747		 * mode or from the idle loop, and if this is not a
2748		 * nested interrupt.  In this case, the CPU is in
2749		 * a quiescent state, so note it.
2750		 *
2751		 * No memory barrier is required here because both
2752		 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2753		 * variables that other CPUs neither access nor modify,
2754		 * at least not while the corresponding CPU is online.
2755		 */
2756
2757		rcu_sched_qs();
2758		rcu_bh_qs();
2759
2760	} else if (!in_softirq()) {
2761
2762		/*
2763		 * Get here if this CPU did not take its interrupt from
2764		 * softirq, in other words, if it is not interrupting
2765		 * a rcu_bh read-side critical section.  This is an _bh
2766		 * critical section, so note it.
2767		 */
2768
2769		rcu_bh_qs();
2770	}
2771	rcu_preempt_check_callbacks();
2772	if (rcu_pending())
2773		invoke_rcu_core();
2774	if (user)
2775		rcu_note_voluntary_context_switch(current);
2776	trace_rcu_utilization(TPS("End scheduler-tick"));
2777}
2778
2779/*
2780 * Scan the leaf rcu_node structures, processing dyntick state for any that
2781 * have not yet encountered a quiescent state, using the function specified.
2782 * Also initiate boosting for any threads blocked on the root rcu_node.
2783 *
2784 * The caller must have suppressed start of new grace periods.
2785 */
2786static void force_qs_rnp(struct rcu_state *rsp,
2787			 int (*f)(struct rcu_data *rsp, bool *isidle,
2788				  unsigned long *maxj),
2789			 bool *isidle, unsigned long *maxj)
2790{
2791	unsigned long bit;
2792	int cpu;
2793	unsigned long flags;
2794	unsigned long mask;
 
2795	struct rcu_node *rnp;
2796
2797	rcu_for_each_leaf_node(rsp, rnp) {
2798		cond_resched_rcu_qs();
 
 
2799		mask = 0;
2800		raw_spin_lock_irqsave_rcu_node(rnp, flags);
 
2801		if (rnp->qsmask == 0) {
2802			if (rcu_state_p == &rcu_sched_state ||
2803			    rsp != rcu_state_p ||
2804			    rcu_preempt_blocked_readers_cgp(rnp)) {
2805				/*
2806				 * No point in scanning bits because they
2807				 * are all zero.  But we might need to
2808				 * priority-boost blocked readers.
2809				 */
2810				rcu_initiate_boost(rnp, flags);
2811				/* rcu_initiate_boost() releases rnp->lock */
2812				continue;
2813			}
2814			if (rnp->parent &&
2815			    (rnp->parent->qsmask & rnp->grpmask)) {
2816				/*
2817				 * Race between grace-period
2818				 * initialization and task exiting RCU
2819				 * read-side critical section: Report.
2820				 */
2821				rcu_report_unblock_qs_rnp(rsp, rnp, flags);
2822				/* rcu_report_unblock_qs_rnp() rlses ->lock */
2823				continue;
2824			}
2825		}
2826		cpu = rnp->grplo;
2827		bit = 1;
2828		for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2829			if ((rnp->qsmask & bit) != 0) {
2830				if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2831					mask |= bit;
2832			}
2833		}
2834		if (mask != 0) {
2835			/* Idle/offline CPUs, report (releases rnp->lock. */
2836			rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2837		} else {
2838			/* Nothing to do here, so just drop the lock. */
2839			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2840		}
2841	}
2842}
2843
2844/*
2845 * Force quiescent states on reluctant CPUs, and also detect which
2846 * CPUs are in dyntick-idle mode.
2847 */
2848static void force_quiescent_state(struct rcu_state *rsp)
2849{
2850	unsigned long flags;
2851	bool ret;
2852	struct rcu_node *rnp;
2853	struct rcu_node *rnp_old = NULL;
2854
2855	/* Funnel through hierarchy to reduce memory contention. */
2856	rnp = __this_cpu_read(rsp->rda->mynode);
2857	for (; rnp != NULL; rnp = rnp->parent) {
2858		ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2859		      !raw_spin_trylock(&rnp->fqslock);
2860		if (rnp_old != NULL)
2861			raw_spin_unlock(&rnp_old->fqslock);
2862		if (ret) {
2863			rsp->n_force_qs_lh++;
2864			return;
2865		}
2866		rnp_old = rnp;
2867	}
2868	/* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2869
2870	/* Reached the root of the rcu_node tree, acquire lock. */
2871	raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2872	raw_spin_unlock(&rnp_old->fqslock);
2873	if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2874		rsp->n_force_qs_lh++;
2875		raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2876		return;  /* Someone beat us to it. */
2877	}
2878	WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
 
2879	raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2880	swake_up(&rsp->gp_wq); /* Memory barrier implied by swake_up() path. */
2881}
 
2882
2883/*
2884 * This does the RCU core processing work for the specified rcu_state
2885 * and rcu_data structures.  This may be called only from the CPU to
2886 * whom the rdp belongs.
2887 */
2888static void
2889__rcu_process_callbacks(struct rcu_state *rsp)
2890{
2891	unsigned long flags;
2892	bool needwake;
2893	struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
 
 
 
 
 
 
 
2894
2895	WARN_ON_ONCE(rdp->beenonline == 0);
 
 
 
 
 
 
2896
2897	/* Update RCU state based on any recent quiescent states. */
2898	rcu_check_quiescent_state(rsp, rdp);
2899
2900	/* Does this CPU require a not-yet-started grace period? */
2901	local_irq_save(flags);
2902	if (cpu_needs_another_gp(rsp, rdp)) {
2903		raw_spin_lock_rcu_node(rcu_get_root(rsp)); /* irqs disabled. */
2904		needwake = rcu_start_gp(rsp);
2905		raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2906		if (needwake)
2907			rcu_gp_kthread_wake(rsp);
2908	} else {
2909		local_irq_restore(flags);
2910	}
2911
 
 
2912	/* If there are callbacks ready, invoke them. */
2913	if (cpu_has_callbacks_ready_to_invoke(rdp))
2914		invoke_rcu_callbacks(rsp, rdp);
 
2915
2916	/* Do any needed deferred wakeups of rcuo kthreads. */
2917	do_nocb_deferred_wakeup(rdp);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2918}
2919
2920/*
2921 * Do RCU core processing for the current CPU.
2922 */
2923static void rcu_process_callbacks(struct softirq_action *unused)
2924{
2925	struct rcu_state *rsp;
2926
2927	if (cpu_is_offline(smp_processor_id()))
2928		return;
2929	trace_rcu_utilization(TPS("Start RCU core"));
2930	for_each_rcu_flavor(rsp)
2931		__rcu_process_callbacks(rsp);
2932	trace_rcu_utilization(TPS("End RCU core"));
 
 
 
 
 
 
 
 
 
 
2933}
2934
2935/*
2936 * Schedule RCU callback invocation.  If the specified type of RCU
2937 * does not support RCU priority boosting, just do a direct call,
2938 * otherwise wake up the per-CPU kernel kthread.  Note that because we
2939 * are running on the current CPU with softirqs disabled, the
2940 * rcu_cpu_kthread_task cannot disappear out from under us.
2941 */
2942static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2943{
2944	if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
2945		return;
2946	if (likely(!rsp->boost)) {
2947		rcu_do_batch(rsp, rdp);
2948		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2949	}
2950	invoke_rcu_callbacks_kthread();
 
 
 
 
2951}
2952
2953static void invoke_rcu_core(void)
 
 
 
 
 
 
 
 
 
 
 
 
2954{
2955	if (cpu_online(smp_processor_id()))
2956		raise_softirq(RCU_SOFTIRQ);
 
 
 
 
 
 
 
2957}
 
2958
2959/*
2960 * Handle any core-RCU processing required by a call_rcu() invocation.
2961 */
2962static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2963			    struct rcu_head *head, unsigned long flags)
2964{
2965	bool needwake;
2966
2967	/*
2968	 * If called from an extended quiescent state, invoke the RCU
2969	 * core in order to force a re-evaluation of RCU's idleness.
2970	 */
2971	if (!rcu_is_watching())
2972		invoke_rcu_core();
2973
2974	/* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2975	if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2976		return;
2977
2978	/*
2979	 * Force the grace period if too many callbacks or too long waiting.
2980	 * Enforce hysteresis, and don't invoke force_quiescent_state()
2981	 * if some other CPU has recently done so.  Also, don't bother
2982	 * invoking force_quiescent_state() if the newly enqueued callback
2983	 * is the only one waiting for a grace period to complete.
2984	 */
2985	if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
 
2986
2987		/* Are we ignoring a completed grace period? */
2988		note_gp_changes(rsp, rdp);
2989
2990		/* Start a new grace period if one not already started. */
2991		if (!rcu_gp_in_progress(rsp)) {
2992			struct rcu_node *rnp_root = rcu_get_root(rsp);
2993
2994			raw_spin_lock_rcu_node(rnp_root);
2995			needwake = rcu_start_gp(rsp);
2996			raw_spin_unlock_rcu_node(rnp_root);
2997			if (needwake)
2998				rcu_gp_kthread_wake(rsp);
2999		} else {
3000			/* Give the grace period a kick. */
3001			rdp->blimit = LONG_MAX;
3002			if (rsp->n_force_qs == rdp->n_force_qs_snap &&
3003			    *rdp->nxttail[RCU_DONE_TAIL] != head)
3004				force_quiescent_state(rsp);
3005			rdp->n_force_qs_snap = rsp->n_force_qs;
3006			rdp->qlen_last_fqs_check = rdp->qlen;
3007		}
3008	}
3009}
3010
3011/*
3012 * RCU callback function to leak a callback.
3013 */
3014static void rcu_leak_callback(struct rcu_head *rhp)
3015{
3016}
3017
3018/*
3019 * Helper function for call_rcu() and friends.  The cpu argument will
3020 * normally be -1, indicating "currently running CPU".  It may specify
3021 * a CPU only if that CPU is a no-CBs CPU.  Currently, only _rcu_barrier()
3022 * is expected to specify a CPU.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3023 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3024static void
3025__call_rcu(struct rcu_head *head, rcu_callback_t func,
3026	   struct rcu_state *rsp, int cpu, bool lazy)
3027{
3028	unsigned long flags;
3029	struct rcu_data *rdp;
 
 
 
 
3030
3031	WARN_ON_ONCE((unsigned long)head & 0x1); /* Misaligned rcu_head! */
3032	if (debug_rcu_head_queue(head)) {
3033		/* Probable double call_rcu(), so leak the callback. */
 
 
 
 
 
 
3034		WRITE_ONCE(head->func, rcu_leak_callback);
3035		WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
3036		return;
3037	}
3038	head->func = func;
3039	head->next = NULL;
3040
3041	/*
3042	 * Opportunistically note grace-period endings and beginnings.
3043	 * Note that we might see a beginning right after we see an
3044	 * end, but never vice versa, since this CPU has to pass through
3045	 * a quiescent state betweentimes.
3046	 */
3047	local_irq_save(flags);
3048	rdp = this_cpu_ptr(rsp->rda);
 
3049
3050	/* Add the callback to our list. */
3051	if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
3052		int offline;
3053
3054		if (cpu != -1)
3055			rdp = per_cpu_ptr(rsp->rda, cpu);
3056		if (likely(rdp->mynode)) {
3057			/* Post-boot, so this should be for a no-CBs CPU. */
3058			offline = !__call_rcu_nocb(rdp, head, lazy, flags);
3059			WARN_ON_ONCE(offline);
3060			/* Offline CPU, _call_rcu() illegal, leak callback.  */
3061			local_irq_restore(flags);
3062			return;
3063		}
3064		/*
3065		 * Very early boot, before rcu_init().  Initialize if needed
3066		 * and then drop through to queue the callback.
3067		 */
3068		BUG_ON(cpu != -1);
3069		WARN_ON_ONCE(!rcu_is_watching());
3070		if (!likely(rdp->nxtlist))
3071			init_default_callback_list(rdp);
 
 
3072	}
3073	WRITE_ONCE(rdp->qlen, rdp->qlen + 1);
3074	if (lazy)
3075		rdp->qlen_lazy++;
3076	else
3077		rcu_idle_count_callbacks_posted();
3078	smp_mb();  /* Count before adding callback for rcu_barrier(). */
3079	*rdp->nxttail[RCU_NEXT_TAIL] = head;
3080	rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
3081
3082	if (__is_kfree_rcu_offset((unsigned long)func))
3083		trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
3084					 rdp->qlen_lazy, rdp->qlen);
3085	else
3086		trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
 
3087
3088	/* Go handle any RCU core processing required. */
3089	__call_rcu_core(rsp, rdp, head, flags);
3090	local_irq_restore(flags);
3091}
3092
3093/*
3094 * Queue an RCU-sched callback for invocation after a grace period.
3095 */
3096void call_rcu_sched(struct rcu_head *head, rcu_callback_t func)
3097{
3098	__call_rcu(head, func, &rcu_sched_state, -1, 0);
3099}
3100EXPORT_SYMBOL_GPL(call_rcu_sched);
3101
3102/*
3103 * Queue an RCU callback for invocation after a quicker grace period.
3104 */
3105void call_rcu_bh(struct rcu_head *head, rcu_callback_t func)
3106{
3107	__call_rcu(head, func, &rcu_bh_state, -1, 0);
3108}
3109EXPORT_SYMBOL_GPL(call_rcu_bh);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3110
3111/*
3112 * Queue an RCU callback for lazy invocation after a grace period.
3113 * This will likely be later named something like "call_rcu_lazy()",
3114 * but this change will require some way of tagging the lazy RCU
3115 * callbacks in the list of pending callbacks. Until then, this
3116 * function may only be called from __kfree_rcu().
3117 */
3118void kfree_call_rcu(struct rcu_head *head,
3119		    rcu_callback_t func)
3120{
3121	__call_rcu(head, func, rcu_state_p, -1, 1);
3122}
3123EXPORT_SYMBOL_GPL(kfree_call_rcu);
3124
3125/*
3126 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3127 * any blocking grace-period wait automatically implies a grace period
3128 * if there is only one CPU online at any point time during execution
3129 * of either synchronize_sched() or synchronize_rcu_bh().  It is OK to
3130 * occasionally incorrectly indicate that there are multiple CPUs online
3131 * when there was in fact only one the whole time, as this just adds
3132 * some overhead: RCU still operates correctly.
3133 */
3134static inline int rcu_blocking_is_gp(void)
3135{
3136	int ret;
3137
3138	might_sleep();  /* Check for RCU read-side critical section. */
3139	preempt_disable();
3140	ret = num_online_cpus() <= 1;
3141	preempt_enable();
3142	return ret;
3143}
3144
3145/**
3146 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3147 *
3148 * Control will return to the caller some time after a full rcu-sched
3149 * grace period has elapsed, in other words after all currently executing
3150 * rcu-sched read-side critical sections have completed.   These read-side
3151 * critical sections are delimited by rcu_read_lock_sched() and
3152 * rcu_read_unlock_sched(), and may be nested.  Note that preempt_disable(),
3153 * local_irq_disable(), and so on may be used in place of
3154 * rcu_read_lock_sched().
3155 *
3156 * This means that all preempt_disable code sequences, including NMI and
3157 * non-threaded hardware-interrupt handlers, in progress on entry will
3158 * have completed before this primitive returns.  However, this does not
3159 * guarantee that softirq handlers will have completed, since in some
3160 * kernels, these handlers can run in process context, and can block.
3161 *
3162 * Note that this guarantee implies further memory-ordering guarantees.
3163 * On systems with more than one CPU, when synchronize_sched() returns,
3164 * each CPU is guaranteed to have executed a full memory barrier since the
3165 * end of its last RCU-sched read-side critical section whose beginning
3166 * preceded the call to synchronize_sched().  In addition, each CPU having
3167 * an RCU read-side critical section that extends beyond the return from
3168 * synchronize_sched() is guaranteed to have executed a full memory barrier
3169 * after the beginning of synchronize_sched() and before the beginning of
3170 * that RCU read-side critical section.  Note that these guarantees include
3171 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3172 * that are executing in the kernel.
3173 *
3174 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3175 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3176 * to have executed a full memory barrier during the execution of
3177 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3178 * again only if the system has more than one CPU).
3179 *
3180 * This primitive provides the guarantees made by the (now removed)
3181 * synchronize_kernel() API.  In contrast, synchronize_rcu() only
3182 * guarantees that rcu_read_lock() sections will have completed.
3183 * In "classic RCU", these two guarantees happen to be one and
3184 * the same, but can differ in realtime RCU implementations.
3185 */
3186void synchronize_sched(void)
3187{
3188	RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3189			 lock_is_held(&rcu_lock_map) ||
3190			 lock_is_held(&rcu_sched_lock_map),
3191			 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3192	if (rcu_blocking_is_gp())
3193		return;
3194	if (rcu_gp_is_expedited())
3195		synchronize_sched_expedited();
3196	else
3197		wait_rcu_gp(call_rcu_sched);
3198}
3199EXPORT_SYMBOL_GPL(synchronize_sched);
3200
3201/**
3202 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3203 *
3204 * Control will return to the caller some time after a full rcu_bh grace
3205 * period has elapsed, in other words after all currently executing rcu_bh
3206 * read-side critical sections have completed.  RCU read-side critical
3207 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3208 * and may be nested.
3209 *
3210 * See the description of synchronize_sched() for more detailed information
3211 * on memory ordering guarantees.
3212 */
3213void synchronize_rcu_bh(void)
3214{
3215	RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3216			 lock_is_held(&rcu_lock_map) ||
3217			 lock_is_held(&rcu_sched_lock_map),
3218			 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3219	if (rcu_blocking_is_gp())
3220		return;
3221	if (rcu_gp_is_expedited())
3222		synchronize_rcu_bh_expedited();
3223	else
3224		wait_rcu_gp(call_rcu_bh);
3225}
3226EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3227
3228/**
3229 * get_state_synchronize_rcu - Snapshot current RCU state
3230 *
3231 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3232 * to determine whether or not a full grace period has elapsed in the
3233 * meantime.
3234 */
3235unsigned long get_state_synchronize_rcu(void)
3236{
3237	/*
3238	 * Any prior manipulation of RCU-protected data must happen
3239	 * before the load from ->gpnum.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3240	 */
3241	smp_mb();  /* ^^^ */
 
 
3242
3243	/*
3244	 * Make sure this load happens before the purportedly
3245	 * time-consuming work between get_state_synchronize_rcu()
3246	 * and cond_synchronize_rcu().
3247	 */
3248	return smp_load_acquire(&rcu_state_p->gpnum);
3249}
3250EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3251
3252/**
3253 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3254 *
3255 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3256 *
3257 * If a full RCU grace period has elapsed since the earlier call to
3258 * get_state_synchronize_rcu(), just return.  Otherwise, invoke
3259 * synchronize_rcu() to wait for a full grace period.
3260 *
3261 * Yes, this function does not take counter wrap into account.  But
3262 * counter wrap is harmless.  If the counter wraps, we have waited for
3263 * more than 2 billion grace periods (and way more on a 64-bit system!),
3264 * so waiting for one additional grace period should be just fine.
3265 */
3266void cond_synchronize_rcu(unsigned long oldstate)
3267{
3268	unsigned long newstate;
 
3269
3270	/*
3271	 * Ensure that this load happens before any RCU-destructive
3272	 * actions the caller might carry out after we return.
3273	 */
3274	newstate = smp_load_acquire(&rcu_state_p->completed);
3275	if (ULONG_CMP_GE(oldstate, newstate))
3276		synchronize_rcu();
3277}
3278EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3279
3280/**
3281 * get_state_synchronize_sched - Snapshot current RCU-sched state
3282 *
3283 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3284 * to determine whether or not a full grace period has elapsed in the
3285 * meantime.
3286 */
3287unsigned long get_state_synchronize_sched(void)
3288{
3289	/*
3290	 * Any prior manipulation of RCU-protected data must happen
3291	 * before the load from ->gpnum.
3292	 */
3293	smp_mb();  /* ^^^ */
3294
3295	/*
3296	 * Make sure this load happens before the purportedly
3297	 * time-consuming work between get_state_synchronize_sched()
3298	 * and cond_synchronize_sched().
3299	 */
3300	return smp_load_acquire(&rcu_sched_state.gpnum);
3301}
3302EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3303
3304/**
3305 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3306 *
3307 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3308 *
3309 * If a full RCU-sched grace period has elapsed since the earlier call to
3310 * get_state_synchronize_sched(), just return.  Otherwise, invoke
3311 * synchronize_sched() to wait for a full grace period.
3312 *
3313 * Yes, this function does not take counter wrap into account.  But
3314 * counter wrap is harmless.  If the counter wraps, we have waited for
3315 * more than 2 billion grace periods (and way more on a 64-bit system!),
3316 * so waiting for one additional grace period should be just fine.
3317 */
3318void cond_synchronize_sched(unsigned long oldstate)
3319{
3320	unsigned long newstate;
3321
3322	/*
3323	 * Ensure that this load happens before any RCU-destructive
3324	 * actions the caller might carry out after we return.
3325	 */
3326	newstate = smp_load_acquire(&rcu_sched_state.completed);
3327	if (ULONG_CMP_GE(oldstate, newstate))
3328		synchronize_sched();
3329}
3330EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3331
3332/* Adjust sequence number for start of update-side operation. */
3333static void rcu_seq_start(unsigned long *sp)
3334{
3335	WRITE_ONCE(*sp, *sp + 1);
3336	smp_mb(); /* Ensure update-side operation after counter increment. */
3337	WARN_ON_ONCE(!(*sp & 0x1));
3338}
3339
3340/* Adjust sequence number for end of update-side operation. */
3341static void rcu_seq_end(unsigned long *sp)
3342{
3343	smp_mb(); /* Ensure update-side operation before counter increment. */
3344	WRITE_ONCE(*sp, *sp + 1);
3345	WARN_ON_ONCE(*sp & 0x1);
3346}
3347
3348/* Take a snapshot of the update side's sequence number. */
3349static unsigned long rcu_seq_snap(unsigned long *sp)
3350{
3351	unsigned long s;
 
3352
3353	s = (READ_ONCE(*sp) + 3) & ~0x1;
3354	smp_mb(); /* Above access must not bleed into critical section. */
3355	return s;
3356}
3357
3358/*
3359 * Given a snapshot from rcu_seq_snap(), determine whether or not a
3360 * full update-side operation has occurred.
3361 */
3362static bool rcu_seq_done(unsigned long *sp, unsigned long s)
3363{
3364	return ULONG_CMP_GE(READ_ONCE(*sp), s);
3365}
 
 
 
 
 
3366
3367/* Wrapper functions for expedited grace periods.  */
3368static void rcu_exp_gp_seq_start(struct rcu_state *rsp)
3369{
3370	rcu_seq_start(&rsp->expedited_sequence);
3371}
3372static void rcu_exp_gp_seq_end(struct rcu_state *rsp)
3373{
3374	rcu_seq_end(&rsp->expedited_sequence);
3375	smp_mb(); /* Ensure that consecutive grace periods serialize. */
3376}
3377static unsigned long rcu_exp_gp_seq_snap(struct rcu_state *rsp)
3378{
3379	smp_mb(); /* Caller's modifications seen first by other CPUs. */
3380	return rcu_seq_snap(&rsp->expedited_sequence);
3381}
3382static bool rcu_exp_gp_seq_done(struct rcu_state *rsp, unsigned long s)
3383{
3384	return rcu_seq_done(&rsp->expedited_sequence, s);
3385}
3386
3387/*
3388 * Reset the ->expmaskinit values in the rcu_node tree to reflect any
3389 * recent CPU-online activity.  Note that these masks are not cleared
3390 * when CPUs go offline, so they reflect the union of all CPUs that have
3391 * ever been online.  This means that this function normally takes its
3392 * no-work-to-do fastpath.
3393 */
3394static void sync_exp_reset_tree_hotplug(struct rcu_state *rsp)
3395{
3396	bool done;
3397	unsigned long flags;
3398	unsigned long mask;
3399	unsigned long oldmask;
3400	int ncpus = READ_ONCE(rsp->ncpus);
3401	struct rcu_node *rnp;
3402	struct rcu_node *rnp_up;
3403
3404	/* If no new CPUs onlined since last time, nothing to do. */
3405	if (likely(ncpus == rsp->ncpus_snap))
3406		return;
3407	rsp->ncpus_snap = ncpus;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3408
3409	/*
3410	 * Each pass through the following loop propagates newly onlined
3411	 * CPUs for the current rcu_node structure up the rcu_node tree.
3412	 */
3413	rcu_for_each_leaf_node(rsp, rnp) {
3414		raw_spin_lock_irqsave_rcu_node(rnp, flags);
3415		if (rnp->expmaskinit == rnp->expmaskinitnext) {
3416			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3417			continue;  /* No new CPUs, nothing to do. */
3418		}
3419
3420		/* Update this node's mask, track old value for propagation. */
3421		oldmask = rnp->expmaskinit;
3422		rnp->expmaskinit = rnp->expmaskinitnext;
3423		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3424
3425		/* If was already nonzero, nothing to propagate. */
3426		if (oldmask)
3427			continue;
3428
3429		/* Propagate the new CPU up the tree. */
3430		mask = rnp->grpmask;
3431		rnp_up = rnp->parent;
3432		done = false;
3433		while (rnp_up) {
3434			raw_spin_lock_irqsave_rcu_node(rnp_up, flags);
3435			if (rnp_up->expmaskinit)
3436				done = true;
3437			rnp_up->expmaskinit |= mask;
3438			raw_spin_unlock_irqrestore_rcu_node(rnp_up, flags);
3439			if (done)
3440				break;
3441			mask = rnp_up->grpmask;
3442			rnp_up = rnp_up->parent;
3443		}
3444	}
3445}
3446
3447/*
3448 * Reset the ->expmask values in the rcu_node tree in preparation for
3449 * a new expedited grace period.
3450 */
3451static void __maybe_unused sync_exp_reset_tree(struct rcu_state *rsp)
3452{
3453	unsigned long flags;
3454	struct rcu_node *rnp;
3455
3456	sync_exp_reset_tree_hotplug(rsp);
3457	rcu_for_each_node_breadth_first(rsp, rnp) {
3458		raw_spin_lock_irqsave_rcu_node(rnp, flags);
3459		WARN_ON_ONCE(rnp->expmask);
3460		rnp->expmask = rnp->expmaskinit;
3461		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 
 
 
 
3462	}
3463}
3464
3465/*
3466 * Return non-zero if there is no RCU expedited grace period in progress
3467 * for the specified rcu_node structure, in other words, if all CPUs and
3468 * tasks covered by the specified rcu_node structure have done their bit
3469 * for the current expedited grace period.  Works only for preemptible
3470 * RCU -- other RCU implementation use other means.
3471 *
3472 * Caller must hold the root rcu_node's exp_funnel_mutex.
 
3473 */
3474static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
3475{
3476	return rnp->exp_tasks == NULL &&
3477	       READ_ONCE(rnp->expmask) == 0;
3478}
3479
3480/*
3481 * Report the exit from RCU read-side critical section for the last task
3482 * that queued itself during or before the current expedited preemptible-RCU
3483 * grace period.  This event is reported either to the rcu_node structure on
3484 * which the task was queued or to one of that rcu_node structure's ancestors,
3485 * recursively up the tree.  (Calm down, calm down, we do the recursion
3486 * iteratively!)
3487 *
3488 * Caller must hold the root rcu_node's exp_funnel_mutex and the
3489 * specified rcu_node structure's ->lock.
3490 */
3491static void __rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
3492				 bool wake, unsigned long flags)
3493	__releases(rnp->lock)
3494{
3495	unsigned long mask;
3496
3497	for (;;) {
3498		if (!sync_rcu_preempt_exp_done(rnp)) {
3499			if (!rnp->expmask)
3500				rcu_initiate_boost(rnp, flags);
3501			else
3502				raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3503			break;
3504		}
3505		if (rnp->parent == NULL) {
3506			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3507			if (wake) {
3508				smp_mb(); /* EGP done before wake_up(). */
3509				swake_up(&rsp->expedited_wq);
 
 
 
 
 
 
 
3510			}
3511			break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3512		}
3513		mask = rnp->grpmask;
3514		raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled */
3515		rnp = rnp->parent;
3516		raw_spin_lock_rcu_node(rnp); /* irqs already disabled */
3517		WARN_ON_ONCE(!(rnp->expmask & mask));
3518		rnp->expmask &= ~mask;
3519	}
 
 
3520}
3521
3522/*
3523 * Report expedited quiescent state for specified node.  This is a
3524 * lock-acquisition wrapper function for __rcu_report_exp_rnp().
3525 *
3526 * Caller must hold the root rcu_node's exp_funnel_mutex.
3527 */
3528static void __maybe_unused rcu_report_exp_rnp(struct rcu_state *rsp,
3529					      struct rcu_node *rnp, bool wake)
3530{
3531	unsigned long flags;
 
 
 
 
 
 
3532
3533	raw_spin_lock_irqsave_rcu_node(rnp, flags);
3534	__rcu_report_exp_rnp(rsp, rnp, wake, flags);
 
 
3535}
3536
3537/*
3538 * Report expedited quiescent state for multiple CPUs, all covered by the
3539 * specified leaf rcu_node structure.  Caller must hold the root
3540 * rcu_node's exp_funnel_mutex.
3541 */
3542static void rcu_report_exp_cpu_mult(struct rcu_state *rsp, struct rcu_node *rnp,
3543				    unsigned long mask, bool wake)
3544{
3545	unsigned long flags;
 
 
3546
3547	raw_spin_lock_irqsave_rcu_node(rnp, flags);
3548	if (!(rnp->expmask & mask)) {
3549		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3550		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3551	}
3552	rnp->expmask &= ~mask;
3553	__rcu_report_exp_rnp(rsp, rnp, wake, flags); /* Releases rnp->lock. */
 
 
 
 
3554}
3555
3556/*
3557 * Report expedited quiescent state for specified rcu_data (CPU).
3558 * Caller must hold the root rcu_node's exp_funnel_mutex.
 
 
 
 
 
 
 
 
3559 */
3560static void rcu_report_exp_rdp(struct rcu_state *rsp, struct rcu_data *rdp,
3561			       bool wake)
3562{
3563	rcu_report_exp_cpu_mult(rsp, rdp->mynode, rdp->grpmask, wake);
3564}
 
 
3565
3566/* Common code for synchronize_{rcu,sched}_expedited() work-done checking. */
3567static bool sync_exp_work_done(struct rcu_state *rsp, struct rcu_node *rnp,
3568			       struct rcu_data *rdp,
3569			       atomic_long_t *stat, unsigned long s)
3570{
3571	if (rcu_exp_gp_seq_done(rsp, s)) {
3572		if (rnp)
3573			mutex_unlock(&rnp->exp_funnel_mutex);
3574		else if (rdp)
3575			mutex_unlock(&rdp->exp_funnel_mutex);
3576		/* Ensure test happens before caller kfree(). */
3577		smp_mb__before_atomic(); /* ^^^ */
3578		atomic_long_inc(stat);
3579		return true;
3580	}
3581	return false;
3582}
3583
3584/*
3585 * Funnel-lock acquisition for expedited grace periods.  Returns a
3586 * pointer to the root rcu_node structure, or NULL if some other
3587 * task did the expedited grace period for us.
3588 */
3589static struct rcu_node *exp_funnel_lock(struct rcu_state *rsp, unsigned long s)
3590{
3591	struct rcu_data *rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id());
3592	struct rcu_node *rnp0;
3593	struct rcu_node *rnp1 = NULL;
3594
3595	/*
3596	 * First try directly acquiring the root lock in order to reduce
3597	 * latency in the common case where expedited grace periods are
3598	 * rare.  We check mutex_is_locked() to avoid pathological levels of
3599	 * memory contention on ->exp_funnel_mutex in the heavy-load case.
3600	 */
3601	rnp0 = rcu_get_root(rsp);
3602	if (!mutex_is_locked(&rnp0->exp_funnel_mutex)) {
3603		if (mutex_trylock(&rnp0->exp_funnel_mutex)) {
3604			if (sync_exp_work_done(rsp, rnp0, NULL,
3605					       &rdp->expedited_workdone0, s))
3606				return NULL;
3607			return rnp0;
3608		}
3609	}
3610
3611	/*
3612	 * Each pass through the following loop works its way
3613	 * up the rcu_node tree, returning if others have done the
3614	 * work or otherwise falls through holding the root rnp's
3615	 * ->exp_funnel_mutex.  The mapping from CPU to rcu_node structure
3616	 * can be inexact, as it is just promoting locality and is not
3617	 * strictly needed for correctness.
3618	 */
3619	if (sync_exp_work_done(rsp, NULL, NULL, &rdp->expedited_workdone1, s))
3620		return NULL;
3621	mutex_lock(&rdp->exp_funnel_mutex);
3622	rnp0 = rdp->mynode;
3623	for (; rnp0 != NULL; rnp0 = rnp0->parent) {
3624		if (sync_exp_work_done(rsp, rnp1, rdp,
3625				       &rdp->expedited_workdone2, s))
3626			return NULL;
3627		mutex_lock(&rnp0->exp_funnel_mutex);
3628		if (rnp1)
3629			mutex_unlock(&rnp1->exp_funnel_mutex);
3630		else
3631			mutex_unlock(&rdp->exp_funnel_mutex);
3632		rnp1 = rnp0;
3633	}
3634	if (sync_exp_work_done(rsp, rnp1, rdp,
3635			       &rdp->expedited_workdone3, s))
3636		return NULL;
3637	return rnp1;
3638}
3639
3640/* Invoked on each online non-idle CPU for expedited quiescent state. */
3641static void sync_sched_exp_handler(void *data)
3642{
3643	struct rcu_data *rdp;
3644	struct rcu_node *rnp;
3645	struct rcu_state *rsp = data;
3646
3647	rdp = this_cpu_ptr(rsp->rda);
3648	rnp = rdp->mynode;
3649	if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
3650	    __this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
3651		return;
3652	__this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, true);
3653	resched_cpu(smp_processor_id());
 
 
 
 
 
 
 
 
 
 
 
 
 
3654}
 
3655
3656/* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
3657static void sync_sched_exp_online_cleanup(int cpu)
3658{
3659	struct rcu_data *rdp;
3660	int ret;
3661	struct rcu_node *rnp;
3662	struct rcu_state *rsp = &rcu_sched_state;
3663
3664	rdp = per_cpu_ptr(rsp->rda, cpu);
3665	rnp = rdp->mynode;
3666	if (!(READ_ONCE(rnp->expmask) & rdp->grpmask))
3667		return;
3668	ret = smp_call_function_single(cpu, sync_sched_exp_handler, rsp, 0);
3669	WARN_ON_ONCE(ret);
 
 
3670}
3671
3672/*
3673 * Select the nodes that the upcoming expedited grace period needs
3674 * to wait for.
3675 */
3676static void sync_rcu_exp_select_cpus(struct rcu_state *rsp,
3677				     smp_call_func_t func)
3678{
3679	int cpu;
3680	unsigned long flags;
3681	unsigned long mask;
3682	unsigned long mask_ofl_test;
3683	unsigned long mask_ofl_ipi;
3684	int ret;
3685	struct rcu_node *rnp;
3686
3687	sync_exp_reset_tree(rsp);
3688	rcu_for_each_leaf_node(rsp, rnp) {
3689		raw_spin_lock_irqsave_rcu_node(rnp, flags);
3690
3691		/* Each pass checks a CPU for identity, offline, and idle. */
3692		mask_ofl_test = 0;
3693		for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
3694			struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3695			struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
3696
3697			if (raw_smp_processor_id() == cpu ||
3698			    !(atomic_add_return(0, &rdtp->dynticks) & 0x1))
3699				mask_ofl_test |= rdp->grpmask;
3700		}
3701		mask_ofl_ipi = rnp->expmask & ~mask_ofl_test;
3702
3703		/*
3704		 * Need to wait for any blocked tasks as well.  Note that
3705		 * additional blocking tasks will also block the expedited
3706		 * GP until such time as the ->expmask bits are cleared.
3707		 */
3708		if (rcu_preempt_has_tasks(rnp))
3709			rnp->exp_tasks = rnp->blkd_tasks.next;
3710		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3711
3712		/* IPI the remaining CPUs for expedited quiescent state. */
3713		mask = 1;
3714		for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) {
3715			if (!(mask_ofl_ipi & mask))
3716				continue;
3717retry_ipi:
3718			ret = smp_call_function_single(cpu, func, rsp, 0);
3719			if (!ret) {
3720				mask_ofl_ipi &= ~mask;
3721				continue;
3722			}
3723			/* Failed, raced with offline. */
3724			raw_spin_lock_irqsave_rcu_node(rnp, flags);
3725			if (cpu_online(cpu) &&
3726			    (rnp->expmask & mask)) {
3727				raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3728				schedule_timeout_uninterruptible(1);
3729				if (cpu_online(cpu) &&
3730				    (rnp->expmask & mask))
3731					goto retry_ipi;
3732				raw_spin_lock_irqsave_rcu_node(rnp, flags);
3733			}
3734			if (!(rnp->expmask & mask))
3735				mask_ofl_ipi &= ~mask;
3736			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3737		}
3738		/* Report quiescent states for those that went offline. */
3739		mask_ofl_test |= mask_ofl_ipi;
3740		if (mask_ofl_test)
3741			rcu_report_exp_cpu_mult(rsp, rnp, mask_ofl_test, false);
3742	}
 
 
3743}
3744
3745static void synchronize_sched_expedited_wait(struct rcu_state *rsp)
 
 
 
 
 
 
 
3746{
3747	int cpu;
3748	unsigned long jiffies_stall;
3749	unsigned long jiffies_start;
3750	unsigned long mask;
3751	int ndetected;
3752	struct rcu_node *rnp;
3753	struct rcu_node *rnp_root = rcu_get_root(rsp);
3754	int ret;
3755
3756	jiffies_stall = rcu_jiffies_till_stall_check();
3757	jiffies_start = jiffies;
3758
3759	for (;;) {
3760		ret = swait_event_timeout(
3761				rsp->expedited_wq,
3762				sync_rcu_preempt_exp_done(rnp_root),
3763				jiffies_stall);
3764		if (ret > 0 || sync_rcu_preempt_exp_done(rnp_root))
3765			return;
3766		if (ret < 0) {
3767			/* Hit a signal, disable CPU stall warnings. */
3768			swait_event(rsp->expedited_wq,
3769				   sync_rcu_preempt_exp_done(rnp_root));
3770			return;
3771		}
3772		pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
3773		       rsp->name);
3774		ndetected = 0;
3775		rcu_for_each_leaf_node(rsp, rnp) {
3776			ndetected = rcu_print_task_exp_stall(rnp);
3777			mask = 1;
3778			for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) {
3779				struct rcu_data *rdp;
3780
3781				if (!(rnp->expmask & mask))
3782					continue;
3783				ndetected++;
3784				rdp = per_cpu_ptr(rsp->rda, cpu);
3785				pr_cont(" %d-%c%c%c", cpu,
3786					"O."[cpu_online(cpu)],
3787					"o."[!!(rdp->grpmask & rnp->expmaskinit)],
3788					"N."[!!(rdp->grpmask & rnp->expmaskinitnext)]);
3789			}
3790			mask <<= 1;
3791		}
3792		pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
3793			jiffies - jiffies_start, rsp->expedited_sequence,
3794			rnp_root->expmask, ".T"[!!rnp_root->exp_tasks]);
3795		if (!ndetected) {
3796			pr_err("blocking rcu_node structures:");
3797			rcu_for_each_node_breadth_first(rsp, rnp) {
3798				if (rnp == rnp_root)
3799					continue; /* printed unconditionally */
3800				if (sync_rcu_preempt_exp_done(rnp))
3801					continue;
3802				pr_cont(" l=%u:%d-%d:%#lx/%c",
3803					rnp->level, rnp->grplo, rnp->grphi,
3804					rnp->expmask,
3805					".T"[!!rnp->exp_tasks]);
3806			}
3807			pr_cont("\n");
3808		}
3809		rcu_for_each_leaf_node(rsp, rnp) {
3810			mask = 1;
3811			for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) {
3812				if (!(rnp->expmask & mask))
3813					continue;
3814				dump_cpu_task(cpu);
3815			}
3816		}
3817		jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3;
 
 
 
3818	}
3819}
3820
3821/**
3822 * synchronize_sched_expedited - Brute-force RCU-sched grace period
 
3823 *
3824 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
3825 * approach to force the grace period to end quickly.  This consumes
3826 * significant time on all CPUs and is unfriendly to real-time workloads,
3827 * so is thus not recommended for any sort of common-case code.  In fact,
3828 * if you are using synchronize_sched_expedited() in a loop, please
3829 * restructure your code to batch your updates, and then use a single
3830 * synchronize_sched() instead.
3831 *
3832 * This implementation can be thought of as an application of sequence
3833 * locking to expedited grace periods, but using the sequence counter to
3834 * determine when someone else has already done the work instead of for
3835 * retrying readers.
3836 */
3837void synchronize_sched_expedited(void)
3838{
3839	unsigned long s;
3840	struct rcu_node *rnp;
3841	struct rcu_state *rsp = &rcu_sched_state;
3842
3843	/* If only one CPU, this is automatically a grace period. */
3844	if (rcu_blocking_is_gp())
3845		return;
 
 
 
 
 
3846
3847	/* If expedited grace periods are prohibited, fall back to normal. */
3848	if (rcu_gp_is_normal()) {
3849		wait_rcu_gp(call_rcu_sched);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3850		return;
3851	}
3852
3853	/* Take a snapshot of the sequence number.  */
3854	s = rcu_exp_gp_seq_snap(rsp);
3855
3856	rnp = exp_funnel_lock(rsp, s);
3857	if (rnp == NULL)
3858		return;  /* Someone else did our work for us. */
3859
3860	rcu_exp_gp_seq_start(rsp);
3861	sync_rcu_exp_select_cpus(rsp, sync_sched_exp_handler);
3862	synchronize_sched_expedited_wait(rsp);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3863
3864	rcu_exp_gp_seq_end(rsp);
3865	mutex_unlock(&rnp->exp_funnel_mutex);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3866}
3867EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
3868
3869/*
3870 * Check to see if there is any immediate RCU-related work to be done
3871 * by the current CPU, for the specified type of RCU, returning 1 if so.
3872 * The checks are in order of increasing expense: checks that can be
3873 * carried out against CPU-local state are performed first.  However,
3874 * we must check for CPU stalls first, else we might not get a chance.
3875 */
3876static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3877{
 
 
3878	struct rcu_node *rnp = rdp->mynode;
3879
3880	rdp->n_rcu_pending++;
3881
3882	/* Check for CPU stalls, if enabled. */
3883	check_cpu_stall(rsp, rdp);
 
 
 
 
3884
3885	/* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3886	if (rcu_nohz_full_cpu(rsp))
3887		return 0;
3888
3889	/* Is the RCU core waiting for a quiescent state from this CPU? */
3890	if (rcu_scheduler_fully_active &&
3891	    rdp->core_needs_qs && rdp->cpu_no_qs.b.norm &&
3892	    rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) {
3893		rdp->n_rp_core_needs_qs++;
3894	} else if (rdp->core_needs_qs &&
3895		   (!rdp->cpu_no_qs.b.norm ||
3896		    rdp->rcu_qs_ctr_snap != __this_cpu_read(rcu_qs_ctr))) {
3897		rdp->n_rp_report_qs++;
3898		return 1;
3899	}
3900
3901	/* Does this CPU have callbacks ready to invoke? */
3902	if (cpu_has_callbacks_ready_to_invoke(rdp)) {
3903		rdp->n_rp_cb_ready++;
3904		return 1;
3905	}
3906
3907	/* Has RCU gone idle with this CPU needing another grace period? */
3908	if (cpu_needs_another_gp(rsp, rdp)) {
3909		rdp->n_rp_cpu_needs_gp++;
3910		return 1;
3911	}
3912
3913	/* Has another RCU grace period completed?  */
3914	if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3915		rdp->n_rp_gp_completed++;
3916		return 1;
3917	}
3918
3919	/* Has a new RCU grace period started? */
3920	if (READ_ONCE(rnp->gpnum) != rdp->gpnum ||
3921	    unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */
3922		rdp->n_rp_gp_started++;
3923		return 1;
3924	}
3925
3926	/* Does this CPU need a deferred NOCB wakeup? */
3927	if (rcu_nocb_need_deferred_wakeup(rdp)) {
3928		rdp->n_rp_nocb_defer_wakeup++;
3929		return 1;
3930	}
3931
3932	/* nothing to do */
3933	rdp->n_rp_need_nothing++;
3934	return 0;
3935}
3936
3937/*
3938 * Check to see if there is any immediate RCU-related work to be done
3939 * by the current CPU, returning 1 if so.  This function is part of the
3940 * RCU implementation; it is -not- an exported member of the RCU API.
3941 */
3942static int rcu_pending(void)
3943{
3944	struct rcu_state *rsp;
3945
3946	for_each_rcu_flavor(rsp)
3947		if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3948			return 1;
3949	return 0;
3950}
3951
3952/*
3953 * Return true if the specified CPU has any callback.  If all_lazy is
3954 * non-NULL, store an indication of whether all callbacks are lazy.
3955 * (If there are no callbacks, all of them are deemed to be lazy.)
3956 */
3957static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3958{
3959	bool al = true;
3960	bool hc = false;
3961	struct rcu_data *rdp;
3962	struct rcu_state *rsp;
3963
3964	for_each_rcu_flavor(rsp) {
3965		rdp = this_cpu_ptr(rsp->rda);
3966		if (!rdp->nxtlist)
3967			continue;
3968		hc = true;
3969		if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
3970			al = false;
3971			break;
3972		}
3973	}
3974	if (all_lazy)
3975		*all_lazy = al;
3976	return hc;
3977}
3978
3979/*
3980 * Helper function for _rcu_barrier() tracing.  If tracing is disabled,
3981 * the compiler is expected to optimize this away.
3982 */
3983static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3984			       int cpu, unsigned long done)
3985{
3986	trace_rcu_barrier(rsp->name, s, cpu,
3987			  atomic_read(&rsp->barrier_cpu_count), done);
3988}
3989
3990/*
3991 * RCU callback function for _rcu_barrier().  If we are last, wake
3992 * up the task executing _rcu_barrier().
 
 
 
 
 
 
3993 */
3994static void rcu_barrier_callback(struct rcu_head *rhp)
3995{
3996	struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3997	struct rcu_state *rsp = rdp->rsp;
3998
3999	if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
4000		_rcu_barrier_trace(rsp, "LastCB", -1, rsp->barrier_sequence);
4001		complete(&rsp->barrier_completion);
4002	} else {
4003		_rcu_barrier_trace(rsp, "CB", -1, rsp->barrier_sequence);
4004	}
4005}
4006
4007/*
4008 * Called with preemption disabled, and from cross-cpu IRQ context.
4009 */
4010static void rcu_barrier_func(void *type)
4011{
4012	struct rcu_state *rsp = type;
4013	struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
4014
4015	_rcu_barrier_trace(rsp, "IRQ", -1, rsp->barrier_sequence);
4016	atomic_inc(&rsp->barrier_cpu_count);
4017	rsp->call(&rdp->barrier_head, rcu_barrier_callback);
 
 
 
 
 
 
 
 
 
 
4018}
4019
4020/*
4021 * Orchestrate the specified type of RCU barrier, waiting for all
4022 * RCU callbacks of the specified type to complete.
 
 
 
 
4023 */
4024static void _rcu_barrier(struct rcu_state *rsp)
4025{
4026	int cpu;
4027	struct rcu_data *rdp;
4028	unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
4029
4030	_rcu_barrier_trace(rsp, "Begin", -1, s);
4031
4032	/* Take mutex to serialize concurrent rcu_barrier() requests. */
4033	mutex_lock(&rsp->barrier_mutex);
4034
4035	/* Did someone else do our work for us? */
4036	if (rcu_seq_done(&rsp->barrier_sequence, s)) {
4037		_rcu_barrier_trace(rsp, "EarlyExit", -1, rsp->barrier_sequence);
 
4038		smp_mb(); /* caller's subsequent code after above check. */
4039		mutex_unlock(&rsp->barrier_mutex);
4040		return;
4041	}
4042
4043	/* Mark the start of the barrier operation. */
4044	rcu_seq_start(&rsp->barrier_sequence);
4045	_rcu_barrier_trace(rsp, "Inc1", -1, rsp->barrier_sequence);
4046
4047	/*
4048	 * Initialize the count to one rather than to zero in order to
4049	 * avoid a too-soon return to zero in case of a short grace period
4050	 * (or preemption of this task).  Exclude CPU-hotplug operations
4051	 * to ensure that no offline CPU has callbacks queued.
 
4052	 */
4053	init_completion(&rsp->barrier_completion);
4054	atomic_set(&rsp->barrier_cpu_count, 1);
4055	get_online_cpus();
4056
4057	/*
4058	 * Force each CPU with callbacks to register a new callback.
4059	 * When that callback is invoked, we will know that all of the
4060	 * corresponding CPU's preceding callbacks have been invoked.
4061	 */
4062	for_each_possible_cpu(cpu) {
4063		if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
 
 
4064			continue;
4065		rdp = per_cpu_ptr(rsp->rda, cpu);
4066		if (rcu_is_nocb_cpu(cpu)) {
4067			if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
4068				_rcu_barrier_trace(rsp, "OfflineNoCB", cpu,
4069						   rsp->barrier_sequence);
4070			} else {
4071				_rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
4072						   rsp->barrier_sequence);
4073				smp_mb__before_atomic();
4074				atomic_inc(&rsp->barrier_cpu_count);
4075				__call_rcu(&rdp->barrier_head,
4076					   rcu_barrier_callback, rsp, cpu, 0);
4077			}
4078		} else if (READ_ONCE(rdp->qlen)) {
4079			_rcu_barrier_trace(rsp, "OnlineQ", cpu,
4080					   rsp->barrier_sequence);
4081			smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
4082		} else {
4083			_rcu_barrier_trace(rsp, "OnlineNQ", cpu,
4084					   rsp->barrier_sequence);
4085		}
4086	}
4087	put_online_cpus();
4088
4089	/*
4090	 * Now that we have an rcu_barrier_callback() callback on each
4091	 * CPU, and thus each counted, remove the initial count.
4092	 */
4093	if (atomic_dec_and_test(&rsp->barrier_cpu_count))
4094		complete(&rsp->barrier_completion);
4095
4096	/* Wait for all rcu_barrier_callback() callbacks to be invoked. */
4097	wait_for_completion(&rsp->barrier_completion);
4098
4099	/* Mark the end of the barrier operation. */
4100	_rcu_barrier_trace(rsp, "Inc2", -1, rsp->barrier_sequence);
4101	rcu_seq_end(&rsp->barrier_sequence);
4102
4103	/* Other rcu_barrier() invocations can now safely proceed. */
4104	mutex_unlock(&rsp->barrier_mutex);
4105}
4106
4107/**
4108 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
4109 */
4110void rcu_barrier_bh(void)
4111{
4112	_rcu_barrier(&rcu_bh_state);
4113}
4114EXPORT_SYMBOL_GPL(rcu_barrier_bh);
4115
4116/**
4117 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
4118 */
4119void rcu_barrier_sched(void)
4120{
4121	_rcu_barrier(&rcu_sched_state);
4122}
4123EXPORT_SYMBOL_GPL(rcu_barrier_sched);
4124
4125/*
4126 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
4127 * first CPU in a given leaf rcu_node structure coming online.  The caller
4128 * must hold the corresponding leaf rcu_node ->lock with interrrupts
4129 * disabled.
4130 */
4131static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
4132{
4133	long mask;
 
4134	struct rcu_node *rnp = rnp_leaf;
4135
 
 
4136	for (;;) {
4137		mask = rnp->grpmask;
4138		rnp = rnp->parent;
4139		if (rnp == NULL)
4140			return;
4141		raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
 
4142		rnp->qsmaskinit |= mask;
4143		raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
 
 
4144	}
4145}
4146
4147/*
4148 * Do boot-time initialization of a CPU's per-CPU RCU data.
4149 */
4150static void __init
4151rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
4152{
4153	unsigned long flags;
4154	struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
4155	struct rcu_node *rnp = rcu_get_root(rsp);
4156
4157	/* Set up local state, ensuring consistent view of global state. */
4158	raw_spin_lock_irqsave_rcu_node(rnp, flags);
4159	rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
4160	rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
4161	WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
4162	WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
 
 
4163	rdp->cpu = cpu;
4164	rdp->rsp = rsp;
4165	mutex_init(&rdp->exp_funnel_mutex);
4166	rcu_boot_init_nocb_percpu_data(rdp);
4167	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4168}
4169
4170/*
4171 * Initialize a CPU's per-CPU RCU data.  Note that only one online or
4172 * offline event can be happening at a given time.  Note also that we
4173 * can accept some slop in the rsp->completed access due to the fact
4174 * that this CPU cannot possibly have any RCU callbacks in flight yet.
 
 
 
 
4175 */
4176static void
4177rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
4178{
4179	unsigned long flags;
4180	unsigned long mask;
4181	struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
4182	struct rcu_node *rnp = rcu_get_root(rsp);
4183
4184	/* Set up local state, ensuring consistent view of global state. */
4185	raw_spin_lock_irqsave_rcu_node(rnp, flags);
4186	rdp->qlen_last_fqs_check = 0;
4187	rdp->n_force_qs_snap = rsp->n_force_qs;
4188	rdp->blimit = blimit;
4189	if (!rdp->nxtlist)
4190		init_callback_list(rdp);  /* Re-enable callbacks on this CPU. */
4191	rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
4192	rcu_sysidle_init_percpu_data(rdp->dynticks);
4193	atomic_set(&rdp->dynticks->dynticks,
4194		   (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
4195	raw_spin_unlock_rcu_node(rnp);		/* irqs remain disabled. */
4196
4197	/*
4198	 * Add CPU to leaf rcu_node pending-online bitmask.  Any needed
4199	 * propagation up the rcu_node tree will happen at the beginning
4200	 * of the next grace period.
4201	 */
4202	rnp = rdp->mynode;
4203	mask = rdp->grpmask;
4204	raw_spin_lock_rcu_node(rnp);		/* irqs already disabled. */
4205	rnp->qsmaskinitnext |= mask;
4206	rnp->expmaskinitnext |= mask;
4207	if (!rdp->beenonline)
4208		WRITE_ONCE(rsp->ncpus, READ_ONCE(rsp->ncpus) + 1);
4209	rdp->beenonline = true;	 /* We have now been online. */
4210	rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */
4211	rdp->completed = rnp->completed;
4212	rdp->cpu_no_qs.b.norm = true;
4213	rdp->rcu_qs_ctr_snap = per_cpu(rcu_qs_ctr, cpu);
4214	rdp->core_needs_qs = false;
4215	trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
 
 
4216	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 
 
 
 
4217}
4218
4219static void rcu_prepare_cpu(int cpu)
 
 
 
4220{
4221	struct rcu_state *rsp;
4222
4223	for_each_rcu_flavor(rsp)
4224		rcu_init_percpu_data(cpu, rsp);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4225}
4226
4227#ifdef CONFIG_HOTPLUG_CPU
4228/*
4229 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
4230 * function.  We now remove it from the rcu_node tree's ->qsmaskinit
4231 * bit masks.
4232 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
4233 * function.  We now remove it from the rcu_node tree's ->qsmaskinit
4234 * bit masks.
4235 */
4236static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4237{
4238	unsigned long flags;
4239	unsigned long mask;
4240	struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
4241	struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
 
4242
4243	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
4244		return;
4245
4246	/* Remove outgoing CPU from mask in the leaf rcu_node structure. */
 
 
 
4247	mask = rdp->grpmask;
4248	raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
4249	rnp->qsmaskinitnext &= ~mask;
4250	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4251}
4252
 
 
 
 
 
 
 
 
 
4253void rcu_report_dead(unsigned int cpu)
4254{
4255	struct rcu_state *rsp;
 
 
 
4256
4257	/* QS for any half-done expedited RCU-sched GP. */
4258	preempt_disable();
4259	rcu_report_exp_rdp(&rcu_sched_state,
4260			   this_cpu_ptr(rcu_sched_state.rda), true);
4261	preempt_enable();
4262	for_each_rcu_flavor(rsp)
4263		rcu_cleanup_dying_idle_cpu(cpu, rsp);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4264}
4265#endif
4266
4267/*
4268 * Handle CPU online/offline notification events.
 
 
4269 */
4270int rcu_cpu_notify(struct notifier_block *self,
4271		   unsigned long action, void *hcpu)
4272{
4273	long cpu = (long)hcpu;
4274	struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
4275	struct rcu_node *rnp = rdp->mynode;
4276	struct rcu_state *rsp;
 
4277
4278	switch (action) {
4279	case CPU_UP_PREPARE:
4280	case CPU_UP_PREPARE_FROZEN:
4281		rcu_prepare_cpu(cpu);
4282		rcu_prepare_kthreads(cpu);
4283		rcu_spawn_all_nocb_kthreads(cpu);
4284		break;
4285	case CPU_ONLINE:
4286	case CPU_DOWN_FAILED:
4287		sync_sched_exp_online_cleanup(cpu);
4288		rcu_boost_kthread_setaffinity(rnp, -1);
4289		break;
4290	case CPU_DOWN_PREPARE:
4291		rcu_boost_kthread_setaffinity(rnp, cpu);
4292		break;
4293	case CPU_DYING:
4294	case CPU_DYING_FROZEN:
4295		for_each_rcu_flavor(rsp)
4296			rcu_cleanup_dying_cpu(rsp);
4297		break;
4298	case CPU_DEAD:
4299	case CPU_DEAD_FROZEN:
4300	case CPU_UP_CANCELED:
4301	case CPU_UP_CANCELED_FROZEN:
4302		for_each_rcu_flavor(rsp) {
4303			rcu_cleanup_dead_cpu(cpu, rsp);
4304			do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
4305		}
4306		break;
4307	default:
4308		break;
4309	}
4310	return NOTIFY_OK;
 
 
 
 
 
 
 
4311}
 
4312
 
 
 
 
4313static int rcu_pm_notify(struct notifier_block *self,
4314			 unsigned long action, void *hcpu)
4315{
4316	switch (action) {
4317	case PM_HIBERNATION_PREPARE:
4318	case PM_SUSPEND_PREPARE:
4319		if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
4320			rcu_expedite_gp();
4321		break;
4322	case PM_POST_HIBERNATION:
4323	case PM_POST_SUSPEND:
4324		if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
4325			rcu_unexpedite_gp();
4326		break;
4327	default:
4328		break;
4329	}
4330	return NOTIFY_OK;
4331}
4332
4333/*
4334 * Spawn the kthreads that handle each RCU flavor's grace periods.
4335 */
4336static int __init rcu_spawn_gp_kthread(void)
4337{
4338	unsigned long flags;
4339	int kthread_prio_in = kthread_prio;
4340	struct rcu_node *rnp;
4341	struct rcu_state *rsp;
4342	struct sched_param sp;
4343	struct task_struct *t;
4344
4345	/* Force priority into range. */
4346	if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
 
 
 
4347		kthread_prio = 1;
4348	else if (kthread_prio < 0)
4349		kthread_prio = 0;
4350	else if (kthread_prio > 99)
4351		kthread_prio = 99;
 
4352	if (kthread_prio != kthread_prio_in)
4353		pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4354			 kthread_prio, kthread_prio_in);
4355
4356	rcu_scheduler_fully_active = 1;
4357	for_each_rcu_flavor(rsp) {
4358		t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
4359		BUG_ON(IS_ERR(t));
4360		rnp = rcu_get_root(rsp);
4361		raw_spin_lock_irqsave_rcu_node(rnp, flags);
4362		rsp->gp_kthread = t;
4363		if (kthread_prio) {
4364			sp.sched_priority = kthread_prio;
4365			sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4366		}
4367		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4368		wake_up_process(t);
4369	}
 
 
 
 
 
 
 
 
4370	rcu_spawn_nocb_kthreads();
4371	rcu_spawn_boost_kthreads();
4372	return 0;
4373}
4374early_initcall(rcu_spawn_gp_kthread);
4375
4376/*
4377 * This function is invoked towards the end of the scheduler's initialization
4378 * process.  Before this is called, the idle task might contain
4379 * RCU read-side critical sections (during which time, this idle
4380 * task is booting the system).  After this function is called, the
4381 * idle tasks are prohibited from containing RCU read-side critical
4382 * sections.  This function also enables RCU lockdep checking.
 
 
4383 */
4384void rcu_scheduler_starting(void)
4385{
4386	WARN_ON(num_online_cpus() != 1);
4387	WARN_ON(nr_context_switches() > 0);
4388	rcu_scheduler_active = 1;
4389}
4390
4391/*
4392 * Compute the per-level fanout, either using the exact fanout specified
4393 * or balancing the tree, depending on the rcu_fanout_exact boot parameter.
4394 */
4395static void __init rcu_init_levelspread(int *levelspread, const int *levelcnt)
4396{
4397	int i;
4398
4399	if (rcu_fanout_exact) {
4400		levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
4401		for (i = rcu_num_lvls - 2; i >= 0; i--)
4402			levelspread[i] = RCU_FANOUT;
4403	} else {
4404		int ccur;
4405		int cprv;
4406
4407		cprv = nr_cpu_ids;
4408		for (i = rcu_num_lvls - 1; i >= 0; i--) {
4409			ccur = levelcnt[i];
4410			levelspread[i] = (cprv + ccur - 1) / ccur;
4411			cprv = ccur;
4412		}
4413	}
4414}
4415
4416/*
4417 * Helper function for rcu_init() that initializes one rcu_state structure.
4418 */
4419static void __init rcu_init_one(struct rcu_state *rsp)
4420{
4421	static const char * const buf[] = RCU_NODE_NAME_INIT;
4422	static const char * const fqs[] = RCU_FQS_NAME_INIT;
4423	static const char * const exp[] = RCU_EXP_NAME_INIT;
4424	static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4425	static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4426	static struct lock_class_key rcu_exp_class[RCU_NUM_LVLS];
4427	static u8 fl_mask = 0x1;
4428
4429	int levelcnt[RCU_NUM_LVLS];		/* # nodes in each level. */
4430	int levelspread[RCU_NUM_LVLS];		/* kids/node in each level. */
4431	int cpustride = 1;
4432	int i;
4433	int j;
4434	struct rcu_node *rnp;
4435
4436	BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf));  /* Fix buf[] init! */
4437
4438	/* Silence gcc 4.8 false positive about array index out of range. */
4439	if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4440		panic("rcu_init_one: rcu_num_lvls out of range");
4441
4442	/* Initialize the level-tracking arrays. */
4443
4444	for (i = 0; i < rcu_num_lvls; i++)
4445		levelcnt[i] = num_rcu_lvl[i];
4446	for (i = 1; i < rcu_num_lvls; i++)
4447		rsp->level[i] = rsp->level[i - 1] + levelcnt[i - 1];
4448	rcu_init_levelspread(levelspread, levelcnt);
4449	rsp->flavor_mask = fl_mask;
4450	fl_mask <<= 1;
4451
4452	/* Initialize the elements themselves, starting from the leaves. */
4453
4454	for (i = rcu_num_lvls - 1; i >= 0; i--) {
4455		cpustride *= levelspread[i];
4456		rnp = rsp->level[i];
4457		for (j = 0; j < levelcnt[i]; j++, rnp++) {
4458			raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4459			lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4460						   &rcu_node_class[i], buf[i]);
4461			raw_spin_lock_init(&rnp->fqslock);
4462			lockdep_set_class_and_name(&rnp->fqslock,
4463						   &rcu_fqs_class[i], fqs[i]);
4464			rnp->gpnum = rsp->gpnum;
4465			rnp->completed = rsp->completed;
 
4466			rnp->qsmask = 0;
4467			rnp->qsmaskinit = 0;
4468			rnp->grplo = j * cpustride;
4469			rnp->grphi = (j + 1) * cpustride - 1;
4470			if (rnp->grphi >= nr_cpu_ids)
4471				rnp->grphi = nr_cpu_ids - 1;
4472			if (i == 0) {
4473				rnp->grpnum = 0;
4474				rnp->grpmask = 0;
4475				rnp->parent = NULL;
4476			} else {
4477				rnp->grpnum = j % levelspread[i - 1];
4478				rnp->grpmask = 1UL << rnp->grpnum;
4479				rnp->parent = rsp->level[i - 1] +
4480					      j / levelspread[i - 1];
4481			}
4482			rnp->level = i;
4483			INIT_LIST_HEAD(&rnp->blkd_tasks);
4484			rcu_init_one_nocb(rnp);
4485			mutex_init(&rnp->exp_funnel_mutex);
4486			lockdep_set_class_and_name(&rnp->exp_funnel_mutex,
4487						   &rcu_exp_class[i], exp[i]);
 
 
4488		}
4489	}
4490
4491	init_swait_queue_head(&rsp->gp_wq);
4492	init_swait_queue_head(&rsp->expedited_wq);
4493	rnp = rsp->level[rcu_num_lvls - 1];
4494	for_each_possible_cpu(i) {
4495		while (i > rnp->grphi)
4496			rnp++;
4497		per_cpu_ptr(rsp->rda, i)->mynode = rnp;
4498		rcu_boot_init_percpu_data(i, rsp);
4499	}
4500	list_add(&rsp->flavors, &rcu_struct_flavors);
4501}
4502
4503/*
4504 * Compute the rcu_node tree geometry from kernel parameters.  This cannot
4505 * replace the definitions in tree.h because those are needed to size
4506 * the ->node array in the rcu_state structure.
4507 */
4508static void __init rcu_init_geometry(void)
4509{
4510	ulong d;
4511	int i;
4512	int rcu_capacity[RCU_NUM_LVLS];
4513
4514	/*
4515	 * Initialize any unspecified boot parameters.
4516	 * The default values of jiffies_till_first_fqs and
4517	 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4518	 * value, which is a function of HZ, then adding one for each
4519	 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4520	 */
4521	d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4522	if (jiffies_till_first_fqs == ULONG_MAX)
4523		jiffies_till_first_fqs = d;
4524	if (jiffies_till_next_fqs == ULONG_MAX)
4525		jiffies_till_next_fqs = d;
 
4526
4527	/* If the compile-time values are accurate, just leave. */
4528	if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4529	    nr_cpu_ids == NR_CPUS)
4530		return;
4531	pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4532		rcu_fanout_leaf, nr_cpu_ids);
4533
4534	/*
4535	 * The boot-time rcu_fanout_leaf parameter must be at least two
4536	 * and cannot exceed the number of bits in the rcu_node masks.
4537	 * Complain and fall back to the compile-time values if this
4538	 * limit is exceeded.
4539	 */
4540	if (rcu_fanout_leaf < 2 ||
4541	    rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4542		rcu_fanout_leaf = RCU_FANOUT_LEAF;
4543		WARN_ON(1);
4544		return;
4545	}
4546
4547	/*
4548	 * Compute number of nodes that can be handled an rcu_node tree
4549	 * with the given number of levels.
4550	 */
4551	rcu_capacity[0] = rcu_fanout_leaf;
4552	for (i = 1; i < RCU_NUM_LVLS; i++)
4553		rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4554
4555	/*
4556	 * The tree must be able to accommodate the configured number of CPUs.
4557	 * If this limit is exceeded, fall back to the compile-time values.
4558	 */
4559	if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4560		rcu_fanout_leaf = RCU_FANOUT_LEAF;
4561		WARN_ON(1);
4562		return;
4563	}
4564
4565	/* Calculate the number of levels in the tree. */
4566	for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4567	}
4568	rcu_num_lvls = i + 1;
4569
4570	/* Calculate the number of rcu_nodes at each level of the tree. */
4571	for (i = 0; i < rcu_num_lvls; i++) {
4572		int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4573		num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4574	}
4575
4576	/* Calculate the total number of rcu_node structures. */
4577	rcu_num_nodes = 0;
4578	for (i = 0; i < rcu_num_lvls; i++)
4579		rcu_num_nodes += num_rcu_lvl[i];
4580}
4581
4582/*
4583 * Dump out the structure of the rcu_node combining tree associated
4584 * with the rcu_state structure referenced by rsp.
4585 */
4586static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4587{
4588	int level = 0;
4589	struct rcu_node *rnp;
4590
4591	pr_info("rcu_node tree layout dump\n");
4592	pr_info(" ");
4593	rcu_for_each_node_breadth_first(rsp, rnp) {
4594		if (rnp->level != level) {
4595			pr_cont("\n");
4596			pr_info(" ");
4597			level = rnp->level;
4598		}
4599		pr_cont("%d:%d ^%d  ", rnp->grplo, rnp->grphi, rnp->grpnum);
4600	}
4601	pr_cont("\n");
4602}
4603
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4604void __init rcu_init(void)
4605{
4606	int cpu;
4607
4608	rcu_early_boot_tests();
4609
 
4610	rcu_bootup_announce();
4611	rcu_init_geometry();
4612	rcu_init_one(&rcu_bh_state);
4613	rcu_init_one(&rcu_sched_state);
4614	if (dump_tree)
4615		rcu_dump_rcu_node_tree(&rcu_sched_state);
4616	__rcu_init_preempt();
4617	open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4618
4619	/*
4620	 * We don't need protection against CPU-hotplug here because
4621	 * this is called early in boot, before either interrupts
4622	 * or the scheduler are operational.
4623	 */
4624	cpu_notifier(rcu_cpu_notify, 0);
4625	pm_notifier(rcu_pm_notify, 0);
4626	for_each_online_cpu(cpu)
4627		rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4628}
4629
 
 
4630#include "tree_plugin.h"
v5.9
   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/percpu.h>
  36#include <linux/notifier.h>
  37#include <linux/cpu.h>
  38#include <linux/mutex.h>
  39#include <linux/time.h>
  40#include <linux/kernel_stat.h>
  41#include <linux/wait.h>
  42#include <linux/kthread.h>
  43#include <uapi/linux/sched/types.h>
  44#include <linux/prefetch.h>
  45#include <linux/delay.h>
 
  46#include <linux/random.h>
  47#include <linux/trace_events.h>
  48#include <linux/suspend.h>
  49#include <linux/ftrace.h>
  50#include <linux/tick.h>
  51#include <linux/sysrq.h>
  52#include <linux/kprobes.h>
  53#include <linux/gfp.h>
  54#include <linux/oom.h>
  55#include <linux/smpboot.h>
  56#include <linux/jiffies.h>
  57#include <linux/slab.h>
  58#include <linux/sched/isolation.h>
  59#include <linux/sched/clock.h>
  60#include <linux/vmalloc.h>
  61#include <linux/mm.h>
  62#include <linux/kasan.h>
  63#include "../time/tick-internal.h"
  64
  65#include "tree.h"
  66#include "rcu.h"
  67
 
  68#ifdef MODULE_PARAM_PREFIX
  69#undef MODULE_PARAM_PREFIX
  70#endif
  71#define MODULE_PARAM_PREFIX "rcutree."
  72
  73#ifndef data_race
  74#define data_race(expr)							\
  75	({								\
  76		expr;							\
  77	})
  78#endif
  79#ifndef ASSERT_EXCLUSIVE_WRITER
  80#define ASSERT_EXCLUSIVE_WRITER(var) do { } while (0)
  81#endif
  82#ifndef ASSERT_EXCLUSIVE_ACCESS
  83#define ASSERT_EXCLUSIVE_ACCESS(var) do { } while (0)
 
 
 
 
 
 
 
  84#endif
  85
  86/* Data structures. */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  87
  88/*
  89 * Steal a bit from the bottom of ->dynticks for idle entry/exit
  90 * control.  Initially this is for TLB flushing.
  91 */
  92#define RCU_DYNTICK_CTRL_MASK 0x1
  93#define RCU_DYNTICK_CTRL_CTR  (RCU_DYNTICK_CTRL_MASK + 1)
  94
  95static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = {
  96	.dynticks_nesting = 1,
  97	.dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE,
  98	.dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
  99};
 100static struct rcu_state rcu_state = {
 101	.level = { &rcu_state.node[0] },
 102	.gp_state = RCU_GP_IDLE,
 103	.gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT,
 104	.barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex),
 105	.name = RCU_NAME,
 106	.abbr = RCU_ABBR,
 107	.exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
 108	.exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
 109	.ofl_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.ofl_lock),
 110};
 111
 112/* Dump rcu_node combining tree at boot to verify correct setup. */
 113static bool dump_tree;
 114module_param(dump_tree, bool, 0444);
 115/* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */
 116static bool use_softirq = true;
 117module_param(use_softirq, bool, 0444);
 118/* Control rcu_node-tree auto-balancing at boot time. */
 119static bool rcu_fanout_exact;
 120module_param(rcu_fanout_exact, bool, 0444);
 121/* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
 122static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
 123module_param(rcu_fanout_leaf, int, 0444);
 124int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
 125/* Number of rcu_nodes at specified level. */
 126int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
 127int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
 128
 129/*
 130 * The rcu_scheduler_active variable is initialized to the value
 131 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
 132 * first task is spawned.  So when this variable is RCU_SCHEDULER_INACTIVE,
 133 * RCU can assume that there is but one task, allowing RCU to (for example)
 134 * optimize synchronize_rcu() to a simple barrier().  When this variable
 135 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
 136 * to detect real grace periods.  This variable is also used to suppress
 137 * boot-time false positives from lockdep-RCU error checking.  Finally, it
 138 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
 139 * is fully initialized, including all of its kthreads having been spawned.
 140 */
 141int rcu_scheduler_active __read_mostly;
 142EXPORT_SYMBOL_GPL(rcu_scheduler_active);
 143
 144/*
 145 * The rcu_scheduler_fully_active variable transitions from zero to one
 146 * during the early_initcall() processing, which is after the scheduler
 147 * is capable of creating new tasks.  So RCU processing (for example,
 148 * creating tasks for RCU priority boosting) must be delayed until after
 149 * rcu_scheduler_fully_active transitions from zero to one.  We also
 150 * currently delay invocation of any RCU callbacks until after this point.
 151 *
 152 * It might later prove better for people registering RCU callbacks during
 153 * early boot to take responsibility for these callbacks, but one step at
 154 * a time.
 155 */
 156static int rcu_scheduler_fully_active __read_mostly;
 157
 158static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
 159			      unsigned long gps, unsigned long flags);
 160static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
 161static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
 162static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
 163static void invoke_rcu_core(void);
 164static void rcu_report_exp_rdp(struct rcu_data *rdp);
 165static void sync_sched_exp_online_cleanup(int cpu);
 166static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp);
 167
 168/* rcuc/rcub kthread realtime priority */
 
 
 
 169static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
 170module_param(kthread_prio, int, 0444);
 
 171
 172/* Delay in jiffies for grace-period initialization delays, debug only. */
 173
 174static int gp_preinit_delay;
 175module_param(gp_preinit_delay, int, 0444);
 176static int gp_init_delay;
 177module_param(gp_init_delay, int, 0444);
 178static int gp_cleanup_delay;
 179module_param(gp_cleanup_delay, int, 0444);
 180
 181/*
 182 * This rcu parameter is runtime-read-only. It reflects
 183 * a minimum allowed number of objects which can be cached
 184 * per-CPU. Object size is equal to one page. This value
 185 * can be changed at boot time.
 186 */
 187static int rcu_min_cached_objs = 2;
 188module_param(rcu_min_cached_objs, int, 0444);
 189
 190/* Retrieve RCU kthreads priority for rcutorture */
 191int rcu_get_gp_kthreads_prio(void)
 192{
 193	return kthread_prio;
 194}
 195EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio);
 196
 197/*
 198 * Number of grace periods between delays, normalized by the duration of
 199 * the delay.  The longer the delay, the more the grace periods between
 200 * each delay.  The reason for this normalization is that it means that,
 201 * for non-zero delays, the overall slowdown of grace periods is constant
 202 * regardless of the duration of the delay.  This arrangement balances
 203 * the need for long delays to increase some race probabilities with the
 204 * need for fast grace periods to increase other race probabilities.
 205 */
 206#define PER_RCU_NODE_PERIOD 3	/* Number of grace periods between delays. */
 207
 208/*
 
 
 
 
 
 
 
 
 
 
 
 
 209 * Compute the mask of online CPUs for the specified rcu_node structure.
 210 * This will not be stable unless the rcu_node structure's ->lock is
 211 * held, but the bit corresponding to the current CPU will be stable
 212 * in most contexts.
 213 */
 214static unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
 215{
 216	return READ_ONCE(rnp->qsmaskinitnext);
 217}
 218
 219/*
 220 * Return true if an RCU grace period is in progress.  The READ_ONCE()s
 221 * permit this function to be invoked without holding the root rcu_node
 222 * structure's ->lock, but of course results can be subject to change.
 223 */
 224static int rcu_gp_in_progress(void)
 225{
 226	return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq));
 227}
 228
 229/*
 230 * Return the number of callbacks queued on the specified CPU.
 231 * Handles both the nocbs and normal cases.
 
 
 232 */
 233static long rcu_get_n_cbs_cpu(int cpu)
 234{
 235	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 236
 237	if (rcu_segcblist_is_enabled(&rdp->cblist))
 238		return rcu_segcblist_n_cbs(&rdp->cblist);
 239	return 0;
 
 
 
 
 
 
 240}
 241
 242void rcu_softirq_qs(void)
 243{
 244	rcu_qs();
 245	rcu_preempt_deferred_qs(current);
 
 
 
 
 246}
 247
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 248/*
 249 * Record entry into an extended quiescent state.  This is only to be
 250 * called when not already in an extended quiescent state, that is,
 251 * RCU is watching prior to the call to this function and is no longer
 252 * watching upon return.
 
 
 
 
 
 
 
 253 */
 254static noinstr void rcu_dynticks_eqs_enter(void)
 255{
 256	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 257	int seq;
 
 
 258
 259	/*
 260	 * CPUs seeing atomic_add_return() must see prior RCU read-side
 261	 * critical sections, and we also must force ordering with the
 262	 * next idle sojourn.
 263	 */
 264	rcu_dynticks_task_trace_enter();  // Before ->dynticks update!
 265	seq = arch_atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
 266	// RCU is no longer watching.  Better be in extended quiescent state!
 267	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
 268		     (seq & RCU_DYNTICK_CTRL_CTR));
 269	/* Better not have special action (TLB flush) pending! */
 270	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
 271		     (seq & RCU_DYNTICK_CTRL_MASK));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 272}
 273
 274/*
 275 * Record exit from an extended quiescent state.  This is only to be
 276 * called from an extended quiescent state, that is, RCU is not watching
 277 * prior to the call to this function and is watching upon return.
 278 */
 279static noinstr void rcu_dynticks_eqs_exit(void)
 280{
 281	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 282	int seq;
 283
 284	/*
 285	 * CPUs seeing atomic_add_return() must see prior idle sojourns,
 286	 * and we also must force ordering with the next RCU read-side
 287	 * critical section.
 288	 */
 289	seq = arch_atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
 290	// RCU is now watching.  Better not be in an extended quiescent state!
 291	rcu_dynticks_task_trace_exit();  // After ->dynticks update!
 292	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
 293		     !(seq & RCU_DYNTICK_CTRL_CTR));
 294	if (seq & RCU_DYNTICK_CTRL_MASK) {
 295		arch_atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdp->dynticks);
 296		smp_mb__after_atomic(); /* _exit after clearing mask. */
 297	}
 298}
 
 299
 300/*
 301 * Reset the current CPU's ->dynticks counter to indicate that the
 302 * newly onlined CPU is no longer in an extended quiescent state.
 303 * This will either leave the counter unchanged, or increment it
 304 * to the next non-quiescent value.
 
 
 
 
 
 
 305 *
 306 * The non-atomic test/increment sequence works because the upper bits
 307 * of the ->dynticks counter are manipulated only by the corresponding CPU,
 308 * or when the corresponding CPU is offline.
 309 */
 310static void rcu_dynticks_eqs_online(void)
 311{
 312	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 313
 314	if (atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR)
 315		return;
 316	atomic_add(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
 
 
 
 
 
 317}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 318
 319/*
 320 * Is the current CPU in an extended quiescent state?
 321 *
 322 * No ordering, as we are sampling CPU-local information.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 323 */
 324static __always_inline bool rcu_dynticks_curr_cpu_in_eqs(void)
 325{
 326	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 
 
 327
 328	return !(arch_atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR);
 
 
 
 
 
 329}
 
 330
 331/*
 332 * Snapshot the ->dynticks counter with full ordering so as to allow
 333 * stable comparison of this counter with past and future snapshots.
 334 */
 335static int rcu_dynticks_snap(struct rcu_data *rdp)
 336{
 337	int snap = atomic_add_return(0, &rdp->dynticks);
 
 
 338
 339	return snap & ~RCU_DYNTICK_CTRL_MASK;
 
 
 
 
 
 340}
 
 341
 342/*
 343 * Return true if the snapshot returned from rcu_dynticks_snap()
 344 * indicates that RCU is in an extended quiescent state.
 345 */
 346static bool rcu_dynticks_in_eqs(int snap)
 347{
 348	return !(snap & RCU_DYNTICK_CTRL_CTR);
 349}
 
 350
 351/*
 352 * Return true if the CPU corresponding to the specified rcu_data
 353 * structure has spent some time in an extended quiescent state since
 354 * rcu_dynticks_snap() returned the specified snapshot.
 355 */
 356static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap)
 357{
 358	return snap != rcu_dynticks_snap(rdp);
 359}
 
 360
 361/*
 362 * Return true if the referenced integer is zero while the specified
 363 * CPU remains within a single extended quiescent state.
 364 */
 365bool rcu_dynticks_zero_in_eqs(int cpu, int *vp)
 366{
 367	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 368	int snap;
 369
 370	// If not quiescent, force back to earlier extended quiescent state.
 371	snap = atomic_read(&rdp->dynticks) & ~(RCU_DYNTICK_CTRL_MASK |
 372					       RCU_DYNTICK_CTRL_CTR);
 373
 374	smp_rmb(); // Order ->dynticks and *vp reads.
 375	if (READ_ONCE(*vp))
 376		return false;  // Non-zero, so report failure;
 377	smp_rmb(); // Order *vp read and ->dynticks re-read.
 378
 379	// If still in the same extended quiescent state, we are good!
 380	return snap == (atomic_read(&rdp->dynticks) & ~RCU_DYNTICK_CTRL_MASK);
 381}
 
 382
 383/*
 384 * Set the special (bottom) bit of the specified CPU so that it
 385 * will take special action (such as flushing its TLB) on the
 386 * next exit from an extended quiescent state.  Returns true if
 387 * the bit was successfully set, or false if the CPU was not in
 388 * an extended quiescent state.
 389 */
 390bool rcu_eqs_special_set(int cpu)
 391{
 392	int old;
 393	int new;
 394	int new_old;
 395	struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
 396
 397	new_old = atomic_read(&rdp->dynticks);
 398	do {
 399		old = new_old;
 400		if (old & RCU_DYNTICK_CTRL_CTR)
 401			return false;
 402		new = old | RCU_DYNTICK_CTRL_MASK;
 403		new_old = atomic_cmpxchg(&rdp->dynticks, old, new);
 404	} while (new_old != old);
 405	return true;
 406}
 
 407
 408/*
 409 * Let the RCU core know that this CPU has gone through the scheduler,
 410 * which is a quiescent state.  This is called when the need for a
 411 * quiescent state is urgent, so we burn an atomic operation and full
 412 * memory barriers to let the RCU core know about it, regardless of what
 413 * this CPU might (or might not) do in the near future.
 414 *
 415 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
 416 *
 417 * The caller must have disabled interrupts and must not be idle.
 418 */
 419void rcu_momentary_dyntick_idle(void)
 420{
 421	int special;
 422
 423	raw_cpu_write(rcu_data.rcu_need_heavy_qs, false);
 424	special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
 425				    &this_cpu_ptr(&rcu_data)->dynticks);
 426	/* It is illegal to call this from idle state. */
 427	WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
 428	rcu_preempt_deferred_qs(current);
 429}
 430EXPORT_SYMBOL_GPL(rcu_momentary_dyntick_idle);
 431
 432/**
 433 * rcu_is_cpu_rrupt_from_idle - see if 'interrupted' from idle
 434 *
 435 * If the current CPU is idle and running at a first-level (not nested)
 436 * interrupt, or directly, from idle, return true.
 437 *
 438 * The caller must have at least disabled IRQs.
 439 */
 440static int rcu_is_cpu_rrupt_from_idle(void)
 441{
 442	long nesting;
 443
 444	/*
 445	 * Usually called from the tick; but also used from smp_function_call()
 446	 * for expedited grace periods. This latter can result in running from
 447	 * the idle task, instead of an actual IPI.
 448	 */
 449	lockdep_assert_irqs_disabled();
 450
 451	/* Check for counter underflows */
 452	RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) < 0,
 453			 "RCU dynticks_nesting counter underflow!");
 454	RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) <= 0,
 455			 "RCU dynticks_nmi_nesting counter underflow/zero!");
 456
 457	/* Are we at first interrupt nesting level? */
 458	nesting = __this_cpu_read(rcu_data.dynticks_nmi_nesting);
 459	if (nesting > 1)
 460		return false;
 461
 462	/*
 463	 * If we're not in an interrupt, we must be in the idle task!
 464	 */
 465	WARN_ON_ONCE(!nesting && !is_idle_task(current));
 466
 467	/* Does CPU appear to be idle from an RCU standpoint? */
 468	return __this_cpu_read(rcu_data.dynticks_nesting) == 0;
 469}
 470
 471#define DEFAULT_RCU_BLIMIT 10     /* Maximum callbacks per rcu_do_batch ... */
 472#define DEFAULT_MAX_RCU_BLIMIT 10000 /* ... even during callback flood. */
 473static long blimit = DEFAULT_RCU_BLIMIT;
 474#define DEFAULT_RCU_QHIMARK 10000 /* If this many pending, ignore blimit. */
 475static long qhimark = DEFAULT_RCU_QHIMARK;
 476#define DEFAULT_RCU_QLOMARK 100   /* Once only this many pending, use blimit. */
 477static long qlowmark = DEFAULT_RCU_QLOMARK;
 478#define DEFAULT_RCU_QOVLD_MULT 2
 479#define DEFAULT_RCU_QOVLD (DEFAULT_RCU_QOVLD_MULT * DEFAULT_RCU_QHIMARK)
 480static long qovld = DEFAULT_RCU_QOVLD; /* If this many pending, hammer QS. */
 481static long qovld_calc = -1;	  /* No pre-initialization lock acquisitions! */
 482
 483module_param(blimit, long, 0444);
 484module_param(qhimark, long, 0444);
 485module_param(qlowmark, long, 0444);
 486module_param(qovld, long, 0444);
 487
 488static ulong jiffies_till_first_fqs = ULONG_MAX;
 489static ulong jiffies_till_next_fqs = ULONG_MAX;
 490static bool rcu_kick_kthreads;
 491static int rcu_divisor = 7;
 492module_param(rcu_divisor, int, 0644);
 493
 494/* Force an exit from rcu_do_batch() after 3 milliseconds. */
 495static long rcu_resched_ns = 3 * NSEC_PER_MSEC;
 496module_param(rcu_resched_ns, long, 0644);
 497
 498/*
 499 * How long the grace period must be before we start recruiting
 500 * quiescent-state help from rcu_note_context_switch().
 
 
 
 501 */
 502static ulong jiffies_till_sched_qs = ULONG_MAX;
 503module_param(jiffies_till_sched_qs, ulong, 0444);
 504static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */
 505module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
 
 
 506
 507/*
 508 * Make sure that we give the grace-period kthread time to detect any
 509 * idle CPUs before taking active measures to force quiescent states.
 510 * However, don't go below 100 milliseconds, adjusted upwards for really
 511 * large systems.
 512 */
 513static void adjust_jiffies_till_sched_qs(void)
 
 514{
 515	unsigned long j;
 516
 517	/* If jiffies_till_sched_qs was specified, respect the request. */
 518	if (jiffies_till_sched_qs != ULONG_MAX) {
 519		WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 520		return;
 521	}
 522	/* Otherwise, set to third fqs scan, but bound below on large system. */
 523	j = READ_ONCE(jiffies_till_first_fqs) +
 524		      2 * READ_ONCE(jiffies_till_next_fqs);
 525	if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
 526		j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
 527	pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
 528	WRITE_ONCE(jiffies_to_sched_qs, j);
 529}
 530
 531static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
 532{
 533	ulong j;
 534	int ret = kstrtoul(val, 0, &j);
 535
 536	if (!ret) {
 537		WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
 538		adjust_jiffies_till_sched_qs();
 539	}
 540	return ret;
 541}
 
 542
 543static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
 
 
 
 
 
 544{
 545	ulong j;
 546	int ret = kstrtoul(val, 0, &j);
 
 547
 548	if (!ret) {
 549		WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
 550		adjust_jiffies_till_sched_qs();
 551	}
 552	return ret;
 
 
 
 553}
 554
 555static struct kernel_param_ops first_fqs_jiffies_ops = {
 556	.set = param_set_first_fqs_jiffies,
 557	.get = param_get_ulong,
 558};
 559
 560static struct kernel_param_ops next_fqs_jiffies_ops = {
 561	.set = param_set_next_fqs_jiffies,
 562	.get = param_get_ulong,
 563};
 564
 565module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
 566module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
 567module_param(rcu_kick_kthreads, bool, 0644);
 568
 569static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
 570static int rcu_pending(int user);
 571
 572/*
 573 * Return the number of RCU GPs completed thus far for debug & stats.
 574 */
 575unsigned long rcu_get_gp_seq(void)
 576{
 577	return READ_ONCE(rcu_state.gp_seq);
 578}
 579EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
 580
 581/*
 582 * Return the number of RCU expedited batches completed thus far for
 583 * debug & stats.  Odd numbers mean that a batch is in progress, even
 584 * numbers mean idle.  The value returned will thus be roughly double
 585 * the cumulative batches since boot.
 586 */
 587unsigned long rcu_exp_batches_completed(void)
 588{
 589	return rcu_state.expedited_sequence;
 
 
 
 
 590}
 591EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
 592
 593/*
 594 * Return the root node of the rcu_state structure.
 
 
 595 */
 596static struct rcu_node *rcu_get_root(void)
 
 597{
 598	return &rcu_state.node[0];
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 599}
 600
 601/*
 602 * Send along grace-period-related data for rcutorture diagnostics.
 
 
 
 
 603 */
 604void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
 605			    unsigned long *gp_seq)
 606{
 607	switch (test_type) {
 608	case RCU_FLAVOR:
 609		*flags = READ_ONCE(rcu_state.gp_flags);
 610		*gp_seq = rcu_seq_current(&rcu_state.gp_seq);
 611		break;
 612	default:
 613		break;
 
 
 
 
 
 
 
 
 
 
 
 
 614	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 615}
 616EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
 617
 618/*
 619 * Enter an RCU extended quiescent state, which can be either the
 620 * idle loop or adaptive-tickless usermode execution.
 621 *
 622 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
 623 * the possibility of usermode upcalls having messed up our count
 624 * of interrupt nesting level during the prior busy period.
 625 */
 626static noinstr void rcu_eqs_enter(bool user)
 627{
 628	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 
 629
 630	WARN_ON_ONCE(rdp->dynticks_nmi_nesting != DYNTICK_IRQ_NONIDLE);
 631	WRITE_ONCE(rdp->dynticks_nmi_nesting, 0);
 632	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
 633		     rdp->dynticks_nesting == 0);
 634	if (rdp->dynticks_nesting != 1) {
 635		// RCU will still be watching, so just do accounting and leave.
 636		rdp->dynticks_nesting--;
 637		return;
 
 638	}
 639
 640	lockdep_assert_irqs_disabled();
 641	instrumentation_begin();
 642	trace_rcu_dyntick(TPS("Start"), rdp->dynticks_nesting, 0, atomic_read(&rdp->dynticks));
 643	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
 644	rdp = this_cpu_ptr(&rcu_data);
 645	do_nocb_deferred_wakeup(rdp);
 646	rcu_prepare_for_idle();
 647	rcu_preempt_deferred_qs(current);
 648
 649	// instrumentation for the noinstr rcu_dynticks_eqs_enter()
 650	instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
 651
 652	instrumentation_end();
 653	WRITE_ONCE(rdp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
 654	// RCU is watching here ...
 655	rcu_dynticks_eqs_enter();
 656	// ... but is no longer watching here.
 657	rcu_dynticks_task_enter();
 658}
 659
 660/**
 661 * rcu_idle_enter - inform RCU that current CPU is entering idle
 662 *
 663 * Enter idle mode, in other words, -leave- the mode in which RCU
 664 * read-side critical sections can occur.  (Though RCU read-side
 665 * critical sections can occur in irq handlers in idle, a possibility
 666 * handled by irq_enter() and irq_exit().)
 667 *
 668 * If you add or remove a call to rcu_idle_enter(), be sure to test with
 669 * CONFIG_RCU_EQS_DEBUG=y.
 
 670 */
 671void rcu_idle_enter(void)
 672{
 673	lockdep_assert_irqs_disabled();
 
 
 674	rcu_eqs_enter(false);
 
 
 675}
 676EXPORT_SYMBOL_GPL(rcu_idle_enter);
 677
 678#ifdef CONFIG_NO_HZ_FULL
 679/**
 680 * rcu_user_enter - inform RCU that we are resuming userspace.
 681 *
 682 * Enter RCU idle mode right before resuming userspace.  No use of RCU
 683 * is permitted between this call and rcu_user_exit(). This way the
 684 * CPU doesn't need to maintain the tick for RCU maintenance purposes
 685 * when the CPU runs in userspace.
 686 *
 687 * If you add or remove a call to rcu_user_enter(), be sure to test with
 688 * CONFIG_RCU_EQS_DEBUG=y.
 689 */
 690noinstr void rcu_user_enter(void)
 691{
 692	lockdep_assert_irqs_disabled();
 693	rcu_eqs_enter(true);
 694}
 695#endif /* CONFIG_NO_HZ_FULL */
 696
 697/**
 698 * rcu_nmi_exit - inform RCU of exit from NMI context
 699 *
 700 * If we are returning from the outermost NMI handler that interrupted an
 701 * RCU-idle period, update rdp->dynticks and rdp->dynticks_nmi_nesting
 702 * to let the RCU grace-period handling know that the CPU is back to
 703 * being RCU-idle.
 704 *
 705 * If you add or remove a call to rcu_nmi_exit(), be sure to test
 706 * with CONFIG_RCU_EQS_DEBUG=y.
 707 */
 708noinstr void rcu_nmi_exit(void)
 709{
 710	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 711
 712	instrumentation_begin();
 713	/*
 714	 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
 715	 * (We are exiting an NMI handler, so RCU better be paying attention
 716	 * to us!)
 717	 */
 718	WARN_ON_ONCE(rdp->dynticks_nmi_nesting <= 0);
 719	WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
 720
 721	/*
 722	 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
 723	 * leave it in non-RCU-idle state.
 724	 */
 725	if (rdp->dynticks_nmi_nesting != 1) {
 726		trace_rcu_dyntick(TPS("--="), rdp->dynticks_nmi_nesting, rdp->dynticks_nmi_nesting - 2,
 727				  atomic_read(&rdp->dynticks));
 728		WRITE_ONCE(rdp->dynticks_nmi_nesting, /* No store tearing. */
 729			   rdp->dynticks_nmi_nesting - 2);
 730		instrumentation_end();
 731		return;
 732	}
 733
 734	/* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
 735	trace_rcu_dyntick(TPS("Startirq"), rdp->dynticks_nmi_nesting, 0, atomic_read(&rdp->dynticks));
 736	WRITE_ONCE(rdp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
 737
 738	if (!in_nmi())
 739		rcu_prepare_for_idle();
 740
 741	// instrumentation for the noinstr rcu_dynticks_eqs_enter()
 742	instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
 743	instrumentation_end();
 744
 745	// RCU is watching here ...
 746	rcu_dynticks_eqs_enter();
 747	// ... but is no longer watching here.
 748
 749	if (!in_nmi())
 750		rcu_dynticks_task_enter();
 751}
 752
 753/**
 754 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
 755 *
 756 * Exit from an interrupt handler, which might possibly result in entering
 757 * idle mode, in other words, leaving the mode in which read-side critical
 758 * sections can occur.  The caller must have disabled interrupts.
 759 *
 760 * This code assumes that the idle loop never does anything that might
 761 * result in unbalanced calls to irq_enter() and irq_exit().  If your
 762 * architecture's idle loop violates this assumption, RCU will give you what
 763 * you deserve, good and hard.  But very infrequently and irreproducibly.
 764 *
 765 * Use things like work queues to work around this limitation.
 766 *
 767 * You have been warned.
 768 *
 769 * If you add or remove a call to rcu_irq_exit(), be sure to test with
 770 * CONFIG_RCU_EQS_DEBUG=y.
 771 */
 772void noinstr rcu_irq_exit(void)
 773{
 774	lockdep_assert_irqs_disabled();
 775	rcu_nmi_exit();
 776}
 777
 778/**
 779 * rcu_irq_exit_preempt - Inform RCU that current CPU is exiting irq
 780 *			  towards in kernel preemption
 781 *
 782 * Same as rcu_irq_exit() but has a sanity check that scheduling is safe
 783 * from RCU point of view. Invoked from return from interrupt before kernel
 784 * preemption.
 785 */
 786void rcu_irq_exit_preempt(void)
 787{
 788	lockdep_assert_irqs_disabled();
 789	rcu_nmi_exit();
 790
 791	RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) <= 0,
 792			 "RCU dynticks_nesting counter underflow/zero!");
 793	RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) !=
 794			 DYNTICK_IRQ_NONIDLE,
 795			 "Bad RCU  dynticks_nmi_nesting counter\n");
 796	RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
 797			 "RCU in extended quiescent state!");
 798}
 799
 800#ifdef CONFIG_PROVE_RCU
 801/**
 802 * rcu_irq_exit_check_preempt - Validate that scheduling is possible
 803 */
 804void rcu_irq_exit_check_preempt(void)
 805{
 806	lockdep_assert_irqs_disabled();
 807
 808	RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) <= 0,
 809			 "RCU dynticks_nesting counter underflow/zero!");
 810	RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) !=
 811			 DYNTICK_IRQ_NONIDLE,
 812			 "Bad RCU  dynticks_nmi_nesting counter\n");
 813	RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
 814			 "RCU in extended quiescent state!");
 815}
 816#endif /* #ifdef CONFIG_PROVE_RCU */
 817
 818/*
 819 * Wrapper for rcu_irq_exit() where interrupts are enabled.
 820 *
 821 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
 822 * with CONFIG_RCU_EQS_DEBUG=y.
 
 823 */
 824void rcu_irq_exit_irqson(void)
 825{
 826	unsigned long flags;
 827
 828	local_irq_save(flags);
 829	rcu_irq_exit();
 830	local_irq_restore(flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 831}
 832
 833/*
 834 * Exit an RCU extended quiescent state, which can be either the
 835 * idle loop or adaptive-tickless usermode execution.
 836 *
 837 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
 838 * allow for the possibility of usermode upcalls messing up our count of
 839 * interrupt nesting level during the busy period that is just now starting.
 840 */
 841static void noinstr rcu_eqs_exit(bool user)
 842{
 843	struct rcu_data *rdp;
 844	long oldval;
 845
 846	lockdep_assert_irqs_disabled();
 847	rdp = this_cpu_ptr(&rcu_data);
 848	oldval = rdp->dynticks_nesting;
 849	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
 850	if (oldval) {
 851		// RCU was already watching, so just do accounting and leave.
 852		rdp->dynticks_nesting++;
 853		return;
 
 854	}
 855	rcu_dynticks_task_exit();
 856	// RCU is not watching here ...
 857	rcu_dynticks_eqs_exit();
 858	// ... but is watching here.
 859	instrumentation_begin();
 860
 861	// instrumentation for the noinstr rcu_dynticks_eqs_exit()
 862	instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
 863
 864	rcu_cleanup_after_idle();
 865	trace_rcu_dyntick(TPS("End"), rdp->dynticks_nesting, 1, atomic_read(&rdp->dynticks));
 866	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
 867	WRITE_ONCE(rdp->dynticks_nesting, 1);
 868	WARN_ON_ONCE(rdp->dynticks_nmi_nesting);
 869	WRITE_ONCE(rdp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
 870	instrumentation_end();
 871}
 872
 873/**
 874 * rcu_idle_exit - inform RCU that current CPU is leaving idle
 875 *
 876 * Exit idle mode, in other words, -enter- the mode in which RCU
 877 * read-side critical sections can occur.
 878 *
 879 * If you add or remove a call to rcu_idle_exit(), be sure to test with
 880 * CONFIG_RCU_EQS_DEBUG=y.
 
 
 881 */
 882void rcu_idle_exit(void)
 883{
 884	unsigned long flags;
 885
 886	local_irq_save(flags);
 887	rcu_eqs_exit(false);
 
 888	local_irq_restore(flags);
 889}
 890EXPORT_SYMBOL_GPL(rcu_idle_exit);
 891
 892#ifdef CONFIG_NO_HZ_FULL
 893/**
 894 * rcu_user_exit - inform RCU that we are exiting userspace.
 895 *
 896 * Exit RCU idle mode while entering the kernel because it can
 897 * run a RCU read side critical section anytime.
 898 *
 899 * If you add or remove a call to rcu_user_exit(), be sure to test with
 900 * CONFIG_RCU_EQS_DEBUG=y.
 901 */
 902void noinstr rcu_user_exit(void)
 903{
 904	rcu_eqs_exit(1);
 905}
 
 906
 907/**
 908 * __rcu_irq_enter_check_tick - Enable scheduler tick on CPU if RCU needs it.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 909 *
 910 * The scheduler tick is not normally enabled when CPUs enter the kernel
 911 * from nohz_full userspace execution.  After all, nohz_full userspace
 912 * execution is an RCU quiescent state and the time executing in the kernel
 913 * is quite short.  Except of course when it isn't.  And it is not hard to
 914 * cause a large system to spend tens of seconds or even minutes looping
 915 * in the kernel, which can cause a number of problems, include RCU CPU
 916 * stall warnings.
 917 *
 918 * Therefore, if a nohz_full CPU fails to report a quiescent state
 919 * in a timely manner, the RCU grace-period kthread sets that CPU's
 920 * ->rcu_urgent_qs flag with the expectation that the next interrupt or
 921 * exception will invoke this function, which will turn on the scheduler
 922 * tick, which will enable RCU to detect that CPU's quiescent states,
 923 * for example, due to cond_resched() calls in CONFIG_PREEMPT=n kernels.
 924 * The tick will be disabled once a quiescent state is reported for
 925 * this CPU.
 926 *
 927 * Of course, in carefully tuned systems, there might never be an
 928 * interrupt or exception.  In that case, the RCU grace-period kthread
 929 * will eventually cause one to happen.  However, in less carefully
 930 * controlled environments, this function allows RCU to get what it
 931 * needs without creating otherwise useless interruptions.
 932 */
 933void __rcu_irq_enter_check_tick(void)
 934{
 935	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 
 936
 937	 // Enabling the tick is unsafe in NMI handlers.
 938	if (WARN_ON_ONCE(in_nmi()))
 939		return;
 
 
 
 
 
 
 
 
 
 940
 941	RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
 942			 "Illegal rcu_irq_enter_check_tick() from extended quiescent state");
 
 
 
 
 943
 944	if (!tick_nohz_full_cpu(rdp->cpu) ||
 945	    !READ_ONCE(rdp->rcu_urgent_qs) ||
 946	    READ_ONCE(rdp->rcu_forced_tick)) {
 947		// RCU doesn't need nohz_full help from this CPU, or it is
 948		// already getting that help.
 949		return;
 950	}
 951
 952	// We get here only when not in an extended quiescent state and
 953	// from interrupts (as opposed to NMIs).  Therefore, (1) RCU is
 954	// already watching and (2) The fact that we are in an interrupt
 955	// handler and that the rcu_node lock is an irq-disabled lock
 956	// prevents self-deadlock.  So we can safely recheck under the lock.
 957	// Note that the nohz_full state currently cannot change.
 958	raw_spin_lock_rcu_node(rdp->mynode);
 959	if (rdp->rcu_urgent_qs && !rdp->rcu_forced_tick) {
 960		// A nohz_full CPU is in the kernel and RCU needs a
 961		// quiescent state.  Turn on the tick!
 962		WRITE_ONCE(rdp->rcu_forced_tick, true);
 963		tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
 964	}
 965	raw_spin_unlock_rcu_node(rdp->mynode);
 966}
 967#endif /* CONFIG_NO_HZ_FULL */
 968
 969/**
 970 * rcu_nmi_enter - inform RCU of entry to NMI context
 971 *
 972 * If the CPU was idle from RCU's viewpoint, update rdp->dynticks and
 973 * rdp->dynticks_nmi_nesting to let the RCU grace-period handling know
 974 * that the CPU is active.  This implementation permits nested NMIs, as
 975 * long as the nesting level does not overflow an int.  (You will probably
 976 * run out of stack space first.)
 977 *
 978 * If you add or remove a call to rcu_nmi_enter(), be sure to test
 979 * with CONFIG_RCU_EQS_DEBUG=y.
 980 */
 981noinstr void rcu_nmi_enter(void)
 982{
 983	long incby = 2;
 984	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 985
 986	/* Complain about underflow. */
 987	WARN_ON_ONCE(rdp->dynticks_nmi_nesting < 0);
 988
 989	/*
 990	 * If idle from RCU viewpoint, atomically increment ->dynticks
 991	 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
 992	 * Otherwise, increment ->dynticks_nmi_nesting by two.  This means
 993	 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
 994	 * to be in the outermost NMI handler that interrupted an RCU-idle
 995	 * period (observation due to Andy Lutomirski).
 996	 */
 997	if (rcu_dynticks_curr_cpu_in_eqs()) {
 998
 999		if (!in_nmi())
1000			rcu_dynticks_task_exit();
1001
1002		// RCU is not watching here ...
1003		rcu_dynticks_eqs_exit();
1004		// ... but is watching here.
1005
1006		if (!in_nmi()) {
1007			instrumentation_begin();
1008			rcu_cleanup_after_idle();
1009			instrumentation_end();
1010		}
1011
1012		instrumentation_begin();
1013		// instrumentation for the noinstr rcu_dynticks_curr_cpu_in_eqs()
1014		instrument_atomic_read(&rdp->dynticks, sizeof(rdp->dynticks));
1015		// instrumentation for the noinstr rcu_dynticks_eqs_exit()
1016		instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
1017
1018		incby = 1;
1019	} else if (!in_nmi()) {
1020		instrumentation_begin();
1021		rcu_irq_enter_check_tick();
1022		instrumentation_end();
1023	} else  {
1024		instrumentation_begin();
1025	}
1026
1027	trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
1028			  rdp->dynticks_nmi_nesting,
1029			  rdp->dynticks_nmi_nesting + incby, atomic_read(&rdp->dynticks));
1030	instrumentation_end();
1031	WRITE_ONCE(rdp->dynticks_nmi_nesting, /* Prevent store tearing. */
1032		   rdp->dynticks_nmi_nesting + incby);
1033	barrier();
1034}
1035
1036/**
1037 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
1038 *
1039 * Enter an interrupt handler, which might possibly result in exiting
1040 * idle mode, in other words, entering the mode in which read-side critical
1041 * sections can occur.  The caller must have disabled interrupts.
1042 *
1043 * Note that the Linux kernel is fully capable of entering an interrupt
1044 * handler that it never exits, for example when doing upcalls to user mode!
1045 * This code assumes that the idle loop never does upcalls to user mode.
1046 * If your architecture's idle loop does do upcalls to user mode (or does
1047 * anything else that results in unbalanced calls to the irq_enter() and
1048 * irq_exit() functions), RCU will give you what you deserve, good and hard.
1049 * But very infrequently and irreproducibly.
1050 *
1051 * Use things like work queues to work around this limitation.
1052 *
1053 * You have been warned.
1054 *
1055 * If you add or remove a call to rcu_irq_enter(), be sure to test with
1056 * CONFIG_RCU_EQS_DEBUG=y.
1057 */
1058noinstr void rcu_irq_enter(void)
1059{
1060	lockdep_assert_irqs_disabled();
1061	rcu_nmi_enter();
1062}
 
 
 
 
 
 
1063
1064/*
1065 * Wrapper for rcu_irq_enter() where interrupts are enabled.
1066 *
1067 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
1068 * with CONFIG_RCU_EQS_DEBUG=y.
1069 */
1070void rcu_irq_enter_irqson(void)
1071{
1072	unsigned long flags;
1073
1074	local_irq_save(flags);
1075	rcu_irq_enter();
1076	local_irq_restore(flags);
 
 
 
 
1077}
1078
1079/*
1080 * If any sort of urgency was applied to the current CPU (for example,
1081 * the scheduler-clock interrupt was enabled on a nohz_full CPU) in order
1082 * to get to a quiescent state, disable it.
 
 
 
1083 */
1084static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp)
1085{
1086	raw_lockdep_assert_held_rcu_node(rdp->mynode);
1087	WRITE_ONCE(rdp->rcu_urgent_qs, false);
1088	WRITE_ONCE(rdp->rcu_need_heavy_qs, false);
1089	if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) {
1090		tick_dep_clear_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
1091		WRITE_ONCE(rdp->rcu_forced_tick, false);
1092	}
1093}
1094
1095noinstr bool __rcu_is_watching(void)
1096{
1097	return !rcu_dynticks_curr_cpu_in_eqs();
1098}
1099
1100/**
1101 * rcu_is_watching - see if RCU thinks that the current CPU is not idle
1102 *
1103 * Return true if RCU is watching the running CPU, which means that this
1104 * CPU can safely enter RCU read-side critical sections.  In other words,
1105 * if the current CPU is not in its idle loop or is in an interrupt or
1106 * NMI handler, return true.
1107 */
1108bool rcu_is_watching(void)
1109{
1110	bool ret;
1111
1112	preempt_disable_notrace();
1113	ret = !rcu_dynticks_curr_cpu_in_eqs();
1114	preempt_enable_notrace();
1115	return ret;
1116}
1117EXPORT_SYMBOL_GPL(rcu_is_watching);
1118
1119/*
1120 * If a holdout task is actually running, request an urgent quiescent
1121 * state from its CPU.  This is unsynchronized, so migrations can cause
1122 * the request to go to the wrong CPU.  Which is OK, all that will happen
1123 * is that the CPU's next context switch will be a bit slower and next
1124 * time around this task will generate another request.
1125 */
1126void rcu_request_urgent_qs_task(struct task_struct *t)
1127{
1128	int cpu;
1129
1130	barrier();
1131	cpu = task_cpu(t);
1132	if (!task_curr(t))
1133		return; /* This task is not running on that CPU. */
1134	smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
1135}
1136
1137#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1138
1139/*
1140 * Is the current CPU online as far as RCU is concerned?
 
 
 
 
 
 
 
 
 
 
 
 
1141 *
1142 * Disable preemption to avoid false positives that could otherwise
1143 * happen due to the current CPU number being sampled, this task being
1144 * preempted, its old CPU being taken offline, resuming on some other CPU,
1145 * then determining that its old CPU is now offline.
1146 *
1147 * Disable checking if in an NMI handler because we cannot safely
1148 * report errors from NMI handlers anyway.  In addition, it is OK to use
1149 * RCU on an offline processor during initial boot, hence the check for
1150 * rcu_scheduler_fully_active.
1151 */
1152bool rcu_lockdep_current_cpu_online(void)
1153{
1154	struct rcu_data *rdp;
1155	struct rcu_node *rnp;
1156	bool ret = false;
1157
1158	if (in_nmi() || !rcu_scheduler_fully_active)
1159		return true;
1160	preempt_disable_notrace();
1161	rdp = this_cpu_ptr(&rcu_data);
1162	rnp = rdp->mynode;
1163	if (rdp->grpmask & rcu_rnp_online_cpus(rnp))
1164		ret = true;
1165	preempt_enable_notrace();
1166	return ret;
1167}
1168EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1169
1170#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1171
1172/*
1173 * We are reporting a quiescent state on behalf of some other CPU, so
1174 * it is our responsibility to check for and handle potential overflow
1175 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
1176 * After all, the CPU might be in deep idle state, and thus executing no
1177 * code whatsoever.
1178 */
1179static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
1180{
1181	raw_lockdep_assert_held_rcu_node(rnp);
1182	if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
1183			 rnp->gp_seq))
1184		WRITE_ONCE(rdp->gpwrap, true);
1185	if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
1186		rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
1187}
1188
1189/*
1190 * Snapshot the specified CPU's dynticks counter so that we can later
1191 * credit them with an implicit quiescent state.  Return 1 if this CPU
1192 * is in dynticks idle mode, which is an extended quiescent state.
1193 */
1194static int dyntick_save_progress_counter(struct rcu_data *rdp)
 
1195{
1196	rdp->dynticks_snap = rcu_dynticks_snap(rdp);
1197	if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1198		trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1199		rcu_gpnum_ovf(rdp->mynode, rdp);
 
 
 
1200		return 1;
1201	}
1202	return 0;
1203}
1204
1205/*
1206 * Return true if the specified CPU has passed through a quiescent
1207 * state by virtue of being in or having passed through an dynticks
1208 * idle state since the last call to dyntick_save_progress_counter()
1209 * for this same CPU, or by virtue of having been offline.
1210 */
1211static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
 
1212{
1213	unsigned long jtsq;
1214	bool *rnhqp;
1215	bool *ruqp;
1216	struct rcu_node *rnp = rdp->mynode;
 
 
1217
1218	/*
1219	 * If the CPU passed through or entered a dynticks idle phase with
1220	 * no active irq/NMI handlers, then we can safely pretend that the CPU
1221	 * already acknowledged the request to pass through a quiescent
1222	 * state.  Either way, that CPU cannot possibly be in an RCU
1223	 * read-side critical section that started before the beginning
1224	 * of the current RCU grace period.
1225	 */
1226	if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
1227		trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1228		rcu_gpnum_ovf(rnp, rdp);
1229		return 1;
1230	}
1231
1232	/* If waiting too long on an offline CPU, complain. */
1233	if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp)) &&
1234	    time_after(jiffies, rcu_state.gp_start + HZ)) {
1235		bool onl;
1236		struct rcu_node *rnp1;
1237
1238		WARN_ON(1);  /* Offline CPUs are supposed to report QS! */
1239		pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
1240			__func__, rnp->grplo, rnp->grphi, rnp->level,
1241			(long)rnp->gp_seq, (long)rnp->completedqs);
1242		for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
1243			pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
1244				__func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
1245		onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
1246		pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
1247			__func__, rdp->cpu, ".o"[onl],
1248			(long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
1249			(long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
1250		return 1; /* Break things loose after complaining. */
1251	}
1252
1253	/*
1254	 * A CPU running for an extended time within the kernel can
1255	 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
1256	 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
1257	 * both .rcu_need_heavy_qs and .rcu_urgent_qs.  Note that the
1258	 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
1259	 * variable are safe because the assignments are repeated if this
1260	 * CPU failed to pass through a quiescent state.  This code
1261	 * also checks .jiffies_resched in case jiffies_to_sched_qs
1262	 * is set way high.
1263	 */
1264	jtsq = READ_ONCE(jiffies_to_sched_qs);
1265	ruqp = per_cpu_ptr(&rcu_data.rcu_urgent_qs, rdp->cpu);
1266	rnhqp = &per_cpu(rcu_data.rcu_need_heavy_qs, rdp->cpu);
1267	if (!READ_ONCE(*rnhqp) &&
1268	    (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
1269	     time_after(jiffies, rcu_state.jiffies_resched) ||
1270	     rcu_state.cbovld)) {
1271		WRITE_ONCE(*rnhqp, true);
1272		/* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1273		smp_store_release(ruqp, true);
1274	} else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
1275		WRITE_ONCE(*ruqp, true);
 
 
 
 
 
 
 
 
 
 
1276	}
1277
1278	/*
1279	 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
1280	 * The above code handles this, but only for straight cond_resched().
1281	 * And some in-kernel loops check need_resched() before calling
1282	 * cond_resched(), which defeats the above code for CPUs that are
1283	 * running in-kernel with scheduling-clock interrupts disabled.
1284	 * So hit them over the head with the resched_cpu() hammer!
1285	 */
1286	if (tick_nohz_full_cpu(rdp->cpu) &&
1287	    (time_after(jiffies, READ_ONCE(rdp->last_fqs_resched) + jtsq * 3) ||
1288	     rcu_state.cbovld)) {
1289		WRITE_ONCE(*ruqp, true);
1290		resched_cpu(rdp->cpu);
1291		WRITE_ONCE(rdp->last_fqs_resched, jiffies);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1292	}
 
 
 
1293
1294	/*
1295	 * If more than halfway to RCU CPU stall-warning time, invoke
1296	 * resched_cpu() more frequently to try to loosen things up a bit.
1297	 * Also check to see if the CPU is getting hammered with interrupts,
1298	 * but only once per grace period, just to keep the IPIs down to
1299	 * a dull roar.
1300	 */
1301	if (time_after(jiffies, rcu_state.jiffies_resched)) {
1302		if (time_after(jiffies,
1303			       READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
1304			resched_cpu(rdp->cpu);
1305			WRITE_ONCE(rdp->last_fqs_resched, jiffies);
 
 
 
 
 
 
1306		}
1307		if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1308		    !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
1309		    (rnp->ffmask & rdp->grpmask)) {
1310			init_irq_work(&rdp->rcu_iw, rcu_iw_handler);
1311			atomic_set(&rdp->rcu_iw.flags, IRQ_WORK_HARD_IRQ);
1312			rdp->rcu_iw_pending = true;
1313			rdp->rcu_iw_gp_seq = rnp->gp_seq;
1314			irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1315		}
1316	}
1317
1318	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1319}
1320
1321/* Trace-event wrapper function for trace_rcu_future_grace_period.  */
1322static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1323			      unsigned long gp_seq_req, const char *s)
 
 
 
1324{
1325	trace_rcu_future_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
1326				      gp_seq_req, rnp->level,
1327				      rnp->grplo, rnp->grphi, s);
1328}
1329
1330/*
1331 * rcu_start_this_gp - Request the start of a particular grace period
1332 * @rnp_start: The leaf node of the CPU from which to start.
1333 * @rdp: The rcu_data corresponding to the CPU from which to start.
1334 * @gp_seq_req: The gp_seq of the grace period to start.
1335 *
1336 * Start the specified grace period, as needed to handle newly arrived
1337 * callbacks.  The required future grace periods are recorded in each
1338 * rcu_node structure's ->gp_seq_needed field.  Returns true if there
1339 * is reason to awaken the grace-period kthread.
1340 *
1341 * The caller must hold the specified rcu_node structure's ->lock, which
1342 * is why the caller is responsible for waking the grace-period kthread.
1343 *
1344 * Returns true if the GP thread needs to be awakened else false.
1345 */
1346static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
1347			      unsigned long gp_seq_req)
 
1348{
 
 
1349	bool ret = false;
1350	struct rcu_node *rnp;
 
 
 
 
 
 
 
 
 
 
 
1351
1352	/*
1353	 * Use funnel locking to either acquire the root rcu_node
1354	 * structure's lock or bail out if the need for this grace period
1355	 * has already been recorded -- or if that grace period has in
1356	 * fact already started.  If there is already a grace period in
1357	 * progress in a non-leaf node, no recording is needed because the
1358	 * end of the grace period will scan the leaf rcu_node structures.
1359	 * Note that rnp_start->lock must not be released.
1360	 */
1361	raw_lockdep_assert_held_rcu_node(rnp_start);
1362	trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
1363	for (rnp = rnp_start; 1; rnp = rnp->parent) {
1364		if (rnp != rnp_start)
1365			raw_spin_lock_rcu_node(rnp);
1366		if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
1367		    rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
1368		    (rnp != rnp_start &&
1369		     rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
1370			trace_rcu_this_gp(rnp, rdp, gp_seq_req,
1371					  TPS("Prestarted"));
1372			goto unlock_out;
1373		}
1374		WRITE_ONCE(rnp->gp_seq_needed, gp_seq_req);
1375		if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
1376			/*
1377			 * We just marked the leaf or internal node, and a
1378			 * grace period is in progress, which means that
1379			 * rcu_gp_cleanup() will see the marking.  Bail to
1380			 * reduce contention.
1381			 */
1382			trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
1383					  TPS("Startedleaf"));
1384			goto unlock_out;
1385		}
1386		if (rnp != rnp_start && rnp->parent != NULL)
1387			raw_spin_unlock_rcu_node(rnp);
1388		if (!rnp->parent)
1389			break;  /* At root, and perhaps also leaf. */
1390	}
1391
1392	/* If GP already in progress, just leave, otherwise start one. */
1393	if (rcu_gp_in_progress()) {
1394		trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1395		goto unlock_out;
1396	}
1397	trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
1398	WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
1399	WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
1400	if (!READ_ONCE(rcu_state.gp_kthread)) {
1401		trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
1402		goto unlock_out;
 
 
 
 
1403	}
1404	trace_rcu_grace_period(rcu_state.name, data_race(rcu_state.gp_seq), TPS("newreq"));
1405	ret = true;  /* Caller must wake GP kthread. */
1406unlock_out:
1407	/* Push furthest requested GP to leaf node and rcu_data structure. */
1408	if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
1409		WRITE_ONCE(rnp_start->gp_seq_needed, rnp->gp_seq_needed);
1410		WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1411	}
1412	if (rnp != rnp_start)
1413		raw_spin_unlock_rcu_node(rnp);
1414	return ret;
1415}
1416
1417/*
1418 * Clean up any old requests for the just-ended grace period.  Also return
1419 * whether any additional grace periods have been requested.
1420 */
1421static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
1422{
1423	bool needmore;
1424	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1425
1426	needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1427	if (!needmore)
1428		rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1429	trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1430			  needmore ? TPS("CleanupMore") : TPS("Cleanup"));
 
 
1431	return needmore;
1432}
1433
1434/*
1435 * Awaken the grace-period kthread.  Don't do a self-awaken (unless in an
1436 * interrupt or softirq handler, in which case we just might immediately
1437 * sleep upon return, resulting in a grace-period hang), and don't bother
1438 * awakening when there is nothing for the grace-period kthread to do
1439 * (as in several CPUs raced to awaken, we lost), and finally don't try
1440 * to awaken a kthread that has not yet been created.  If all those checks
1441 * are passed, track some debug information and awaken.
1442 *
1443 * So why do the self-wakeup when in an interrupt or softirq handler
1444 * in the grace-period kthread's context?  Because the kthread might have
1445 * been interrupted just as it was going to sleep, and just after the final
1446 * pre-sleep check of the awaken condition.  In this case, a wakeup really
1447 * is required, and is therefore supplied.
1448 */
1449static void rcu_gp_kthread_wake(void)
1450{
1451	struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
1452
1453	if ((current == t && !in_irq() && !in_serving_softirq()) ||
1454	    !READ_ONCE(rcu_state.gp_flags) || !t)
1455		return;
1456	WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
1457	WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
1458	swake_up_one(&rcu_state.gp_wq);
1459}
1460
1461/*
1462 * If there is room, assign a ->gp_seq number to any callbacks on this
1463 * CPU that have not already been assigned.  Also accelerate any callbacks
1464 * that were previously assigned a ->gp_seq number that has since proven
1465 * to be too conservative, which can happen if callbacks get assigned a
1466 * ->gp_seq number while RCU is idle, but with reference to a non-root
1467 * rcu_node structure.  This function is idempotent, so it does not hurt
1468 * to call it repeatedly.  Returns an flag saying that we should awaken
1469 * the RCU grace-period kthread.
1470 *
1471 * The caller must hold rnp->lock with interrupts disabled.
1472 */
1473static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
 
1474{
1475	unsigned long gp_seq_req;
1476	bool ret = false;
 
 
 
 
 
1477
1478	rcu_lockdep_assert_cblist_protected(rdp);
1479	raw_lockdep_assert_held_rcu_node(rnp);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1480
1481	/* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1482	if (!rcu_segcblist_pend_cbs(&rdp->cblist))
 
 
 
 
 
1483		return false;
1484
1485	/*
1486	 * Callbacks are often registered with incomplete grace-period
1487	 * information.  Something about the fact that getting exact
1488	 * information requires acquiring a global lock...  RCU therefore
1489	 * makes a conservative estimate of the grace period number at which
1490	 * a given callback will become ready to invoke.	The following
1491	 * code checks this estimate and improves it when possible, thus
1492	 * accelerating callback invocation to an earlier grace-period
1493	 * number.
1494	 */
1495	gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1496	if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1497		ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1498
1499	/* Trace depending on how much we were able to accelerate. */
1500	if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1501		trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccWaitCB"));
1502	else
1503		trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccReadyCB"));
1504	return ret;
1505}
1506
1507/*
1508 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1509 * rcu_node structure's ->lock be held.  It consults the cached value
1510 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1511 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1512 * while holding the leaf rcu_node structure's ->lock.
1513 */
1514static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1515					struct rcu_data *rdp)
1516{
1517	unsigned long c;
1518	bool needwake;
1519
1520	rcu_lockdep_assert_cblist_protected(rdp);
1521	c = rcu_seq_snap(&rcu_state.gp_seq);
1522	if (!READ_ONCE(rdp->gpwrap) && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1523		/* Old request still live, so mark recent callbacks. */
1524		(void)rcu_segcblist_accelerate(&rdp->cblist, c);
1525		return;
1526	}
1527	raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1528	needwake = rcu_accelerate_cbs(rnp, rdp);
1529	raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1530	if (needwake)
1531		rcu_gp_kthread_wake();
1532}
1533
1534/*
1535 * Move any callbacks whose grace period has completed to the
1536 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1537 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1538 * sublist.  This function is idempotent, so it does not hurt to
1539 * invoke it repeatedly.  As long as it is not invoked -too- often...
1540 * Returns true if the RCU grace-period kthread needs to be awakened.
1541 *
1542 * The caller must hold rnp->lock with interrupts disabled.
1543 */
1544static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
 
1545{
1546	rcu_lockdep_assert_cblist_protected(rdp);
1547	raw_lockdep_assert_held_rcu_node(rnp);
1548
1549	/* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1550	if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1551		return false;
1552
1553	/*
1554	 * Find all callbacks whose ->gp_seq numbers indicate that they
1555	 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1556	 */
1557	rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1558
1559	/* Classify any remaining callbacks. */
1560	return rcu_accelerate_cbs(rnp, rdp);
1561}
1562
1563/*
1564 * Move and classify callbacks, but only if doing so won't require
1565 * that the RCU grace-period kthread be awakened.
1566 */
1567static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
1568						  struct rcu_data *rdp)
1569{
1570	rcu_lockdep_assert_cblist_protected(rdp);
1571	if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) ||
1572	    !raw_spin_trylock_rcu_node(rnp))
1573		return;
1574	WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
1575	raw_spin_unlock_rcu_node(rnp);
1576}
1577
1578/*
1579 * Update CPU-local rcu_data state to record the beginnings and ends of
1580 * grace periods.  The caller must hold the ->lock of the leaf rcu_node
1581 * structure corresponding to the current CPU, and must have irqs disabled.
1582 * Returns true if the grace-period kthread needs to be awakened.
1583 */
1584static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
 
1585{
1586	bool ret = false;
1587	bool need_qs;
1588	const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
1589			       rcu_segcblist_is_offloaded(&rdp->cblist);
1590
1591	raw_lockdep_assert_held_rcu_node(rnp);
 
 
1592
1593	if (rdp->gp_seq == rnp->gp_seq)
1594		return false; /* Nothing to do. */
1595
1596	/* Handle the ends of any preceding grace periods first. */
1597	if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1598	    unlikely(READ_ONCE(rdp->gpwrap))) {
1599		if (!offloaded)
1600			ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
1601		rdp->core_needs_qs = false;
1602		trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1603	} else {
1604		if (!offloaded)
1605			ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
1606		if (rdp->core_needs_qs)
1607			rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask);
 
 
 
1608	}
1609
1610	/* Now handle the beginnings of any new-to-this-CPU grace periods. */
1611	if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1612	    unlikely(READ_ONCE(rdp->gpwrap))) {
1613		/*
1614		 * If the current grace period is waiting for this CPU,
1615		 * set up to detect a quiescent state, otherwise don't
1616		 * go looking for one.
1617		 */
1618		trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1619		need_qs = !!(rnp->qsmask & rdp->grpmask);
1620		rdp->cpu_no_qs.b.norm = need_qs;
1621		rdp->core_needs_qs = need_qs;
 
1622		zero_cpu_stall_ticks(rdp);
 
1623	}
1624	rdp->gp_seq = rnp->gp_seq;  /* Remember new grace-period state. */
1625	if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
1626		WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1627	WRITE_ONCE(rdp->gpwrap, false);
1628	rcu_gpnum_ovf(rnp, rdp);
1629	return ret;
1630}
1631
1632static void note_gp_changes(struct rcu_data *rdp)
1633{
1634	unsigned long flags;
1635	bool needwake;
1636	struct rcu_node *rnp;
1637
1638	local_irq_save(flags);
1639	rnp = rdp->mynode;
1640	if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
 
1641	     !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1642	    !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1643		local_irq_restore(flags);
1644		return;
1645	}
1646	needwake = __note_gp_changes(rnp, rdp);
1647	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1648	if (needwake)
1649		rcu_gp_kthread_wake();
1650}
1651
1652static void rcu_gp_slow(int delay)
1653{
1654	if (delay > 0 &&
1655	    !(rcu_seq_ctr(rcu_state.gp_seq) %
1656	      (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1657		schedule_timeout_idle(delay);
1658}
1659
1660static unsigned long sleep_duration;
1661
1662/* Allow rcutorture to stall the grace-period kthread. */
1663void rcu_gp_set_torture_wait(int duration)
1664{
1665	if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST) && duration > 0)
1666		WRITE_ONCE(sleep_duration, duration);
1667}
1668EXPORT_SYMBOL_GPL(rcu_gp_set_torture_wait);
1669
1670/* Actually implement the aforementioned wait. */
1671static void rcu_gp_torture_wait(void)
1672{
1673	unsigned long duration;
1674
1675	if (!IS_ENABLED(CONFIG_RCU_TORTURE_TEST))
1676		return;
1677	duration = xchg(&sleep_duration, 0UL);
1678	if (duration > 0) {
1679		pr_alert("%s: Waiting %lu jiffies\n", __func__, duration);
1680		schedule_timeout_idle(duration);
1681		pr_alert("%s: Wait complete\n", __func__);
1682	}
1683}
1684
1685/*
1686 * Initialize a new grace period.  Return false if no grace period required.
1687 */
1688static bool rcu_gp_init(void)
1689{
1690	unsigned long flags;
1691	unsigned long oldmask;
1692	unsigned long mask;
1693	struct rcu_data *rdp;
1694	struct rcu_node *rnp = rcu_get_root();
1695
1696	WRITE_ONCE(rcu_state.gp_activity, jiffies);
1697	raw_spin_lock_irq_rcu_node(rnp);
1698	if (!READ_ONCE(rcu_state.gp_flags)) {
1699		/* Spurious wakeup, tell caller to go back to sleep.  */
1700		raw_spin_unlock_irq_rcu_node(rnp);
1701		return false;
1702	}
1703	WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1704
1705	if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1706		/*
1707		 * Grace period already in progress, don't start another.
1708		 * Not supposed to be able to happen.
1709		 */
1710		raw_spin_unlock_irq_rcu_node(rnp);
1711		return false;
1712	}
1713
1714	/* Advance to a new grace period and initialize state. */
1715	record_gp_stall_check_time();
1716	/* Record GP times before starting GP, hence rcu_seq_start(). */
1717	rcu_seq_start(&rcu_state.gp_seq);
1718	ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
1719	trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
1720	raw_spin_unlock_irq_rcu_node(rnp);
1721
1722	/*
1723	 * Apply per-leaf buffered online and offline operations to the
1724	 * rcu_node tree.  Note that this new grace period need not wait
1725	 * for subsequent online CPUs, and that quiescent-state forcing
1726	 * will handle subsequent offline CPUs.
1727	 */
1728	rcu_state.gp_state = RCU_GP_ONOFF;
1729	rcu_for_each_leaf_node(rnp) {
1730		raw_spin_lock(&rcu_state.ofl_lock);
1731		raw_spin_lock_irq_rcu_node(rnp);
1732		if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1733		    !rnp->wait_blkd_tasks) {
1734			/* Nothing to do on this leaf rcu_node structure. */
1735			raw_spin_unlock_irq_rcu_node(rnp);
1736			raw_spin_unlock(&rcu_state.ofl_lock);
1737			continue;
1738		}
1739
1740		/* Record old state, apply changes to ->qsmaskinit field. */
1741		oldmask = rnp->qsmaskinit;
1742		rnp->qsmaskinit = rnp->qsmaskinitnext;
1743
1744		/* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1745		if (!oldmask != !rnp->qsmaskinit) {
1746			if (!oldmask) { /* First online CPU for rcu_node. */
1747				if (!rnp->wait_blkd_tasks) /* Ever offline? */
1748					rcu_init_new_rnp(rnp);
1749			} else if (rcu_preempt_has_tasks(rnp)) {
1750				rnp->wait_blkd_tasks = true; /* blocked tasks */
1751			} else { /* Last offline CPU and can propagate. */
1752				rcu_cleanup_dead_rnp(rnp);
1753			}
1754		}
1755
1756		/*
1757		 * If all waited-on tasks from prior grace period are
1758		 * done, and if all this rcu_node structure's CPUs are
1759		 * still offline, propagate up the rcu_node tree and
1760		 * clear ->wait_blkd_tasks.  Otherwise, if one of this
1761		 * rcu_node structure's CPUs has since come back online,
1762		 * simply clear ->wait_blkd_tasks.
 
1763		 */
1764		if (rnp->wait_blkd_tasks &&
1765		    (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
 
1766			rnp->wait_blkd_tasks = false;
1767			if (!rnp->qsmaskinit)
1768				rcu_cleanup_dead_rnp(rnp);
1769		}
1770
1771		raw_spin_unlock_irq_rcu_node(rnp);
1772		raw_spin_unlock(&rcu_state.ofl_lock);
1773	}
1774	rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1775
1776	/*
1777	 * Set the quiescent-state-needed bits in all the rcu_node
1778	 * structures for all currently online CPUs in breadth-first
1779	 * order, starting from the root rcu_node structure, relying on the
1780	 * layout of the tree within the rcu_state.node[] array.  Note that
1781	 * other CPUs will access only the leaves of the hierarchy, thus
1782	 * seeing that no grace period is in progress, at least until the
1783	 * corresponding leaf node has been initialized.
 
1784	 *
1785	 * The grace period cannot complete until the initialization
1786	 * process finishes, because this kthread handles both.
1787	 */
1788	rcu_state.gp_state = RCU_GP_INIT;
1789	rcu_for_each_node_breadth_first(rnp) {
1790		rcu_gp_slow(gp_init_delay);
1791		raw_spin_lock_irqsave_rcu_node(rnp, flags);
1792		rdp = this_cpu_ptr(&rcu_data);
1793		rcu_preempt_check_blocked_tasks(rnp);
1794		rnp->qsmask = rnp->qsmaskinit;
1795		WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
 
 
1796		if (rnp == rdp->mynode)
1797			(void)__note_gp_changes(rnp, rdp);
1798		rcu_preempt_boost_start_gp(rnp);
1799		trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1800					    rnp->level, rnp->grplo,
1801					    rnp->grphi, rnp->qsmask);
1802		/* Quiescent states for tasks on any now-offline CPUs. */
1803		mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1804		rnp->rcu_gp_init_mask = mask;
1805		if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1806			rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1807		else
1808			raw_spin_unlock_irq_rcu_node(rnp);
1809		cond_resched_tasks_rcu_qs();
1810		WRITE_ONCE(rcu_state.gp_activity, jiffies);
1811	}
1812
1813	return true;
1814}
1815
1816/*
1817 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1818 * time.
1819 */
1820static bool rcu_gp_fqs_check_wake(int *gfp)
1821{
1822	struct rcu_node *rnp = rcu_get_root();
1823
1824	// If under overload conditions, force an immediate FQS scan.
1825	if (*gfp & RCU_GP_FLAG_OVLD)
1826		return true;
1827
1828	// Someone like call_rcu() requested a force-quiescent-state scan.
1829	*gfp = READ_ONCE(rcu_state.gp_flags);
1830	if (*gfp & RCU_GP_FLAG_FQS)
1831		return true;
1832
1833	// The current grace period has completed.
1834	if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1835		return true;
1836
1837	return false;
1838}
1839
1840/*
1841 * Do one round of quiescent-state forcing.
1842 */
1843static void rcu_gp_fqs(bool first_time)
1844{
1845	struct rcu_node *rnp = rcu_get_root();
 
 
1846
1847	WRITE_ONCE(rcu_state.gp_activity, jiffies);
1848	rcu_state.n_force_qs++;
1849	if (first_time) {
1850		/* Collect dyntick-idle snapshots. */
1851		force_qs_rnp(dyntick_save_progress_counter);
 
 
 
 
 
 
1852	} else {
1853		/* Handle dyntick-idle and offline CPUs. */
1854		force_qs_rnp(rcu_implicit_dynticks_qs);
 
1855	}
1856	/* Clear flag to prevent immediate re-entry. */
1857	if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
1858		raw_spin_lock_irq_rcu_node(rnp);
1859		WRITE_ONCE(rcu_state.gp_flags,
1860			   READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
1861		raw_spin_unlock_irq_rcu_node(rnp);
1862	}
1863}
1864
1865/*
1866 * Loop doing repeated quiescent-state forcing until the grace period ends.
1867 */
1868static void rcu_gp_fqs_loop(void)
1869{
1870	bool first_gp_fqs;
1871	int gf = 0;
1872	unsigned long j;
1873	int ret;
1874	struct rcu_node *rnp = rcu_get_root();
1875
1876	first_gp_fqs = true;
1877	j = READ_ONCE(jiffies_till_first_fqs);
1878	if (rcu_state.cbovld)
1879		gf = RCU_GP_FLAG_OVLD;
1880	ret = 0;
1881	for (;;) {
1882		if (!ret) {
1883			rcu_state.jiffies_force_qs = jiffies + j;
1884			WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
1885				   jiffies + (j ? 3 * j : 2));
1886		}
1887		trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1888				       TPS("fqswait"));
1889		rcu_state.gp_state = RCU_GP_WAIT_FQS;
1890		ret = swait_event_idle_timeout_exclusive(
1891				rcu_state.gp_wq, rcu_gp_fqs_check_wake(&gf), j);
1892		rcu_gp_torture_wait();
1893		rcu_state.gp_state = RCU_GP_DOING_FQS;
1894		/* Locking provides needed memory barriers. */
1895		/* If grace period done, leave loop. */
1896		if (!READ_ONCE(rnp->qsmask) &&
1897		    !rcu_preempt_blocked_readers_cgp(rnp))
1898			break;
1899		/* If time for quiescent-state forcing, do it. */
1900		if (!time_after(rcu_state.jiffies_force_qs, jiffies) ||
1901		    (gf & RCU_GP_FLAG_FQS)) {
1902			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1903					       TPS("fqsstart"));
1904			rcu_gp_fqs(first_gp_fqs);
1905			gf = 0;
1906			if (first_gp_fqs) {
1907				first_gp_fqs = false;
1908				gf = rcu_state.cbovld ? RCU_GP_FLAG_OVLD : 0;
1909			}
1910			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1911					       TPS("fqsend"));
1912			cond_resched_tasks_rcu_qs();
1913			WRITE_ONCE(rcu_state.gp_activity, jiffies);
1914			ret = 0; /* Force full wait till next FQS. */
1915			j = READ_ONCE(jiffies_till_next_fqs);
1916		} else {
1917			/* Deal with stray signal. */
1918			cond_resched_tasks_rcu_qs();
1919			WRITE_ONCE(rcu_state.gp_activity, jiffies);
1920			WARN_ON(signal_pending(current));
1921			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1922					       TPS("fqswaitsig"));
1923			ret = 1; /* Keep old FQS timing. */
1924			j = jiffies;
1925			if (time_after(jiffies, rcu_state.jiffies_force_qs))
1926				j = 1;
1927			else
1928				j = rcu_state.jiffies_force_qs - j;
1929			gf = 0;
1930		}
1931	}
1932}
1933
1934/*
1935 * Clean up after the old grace period.
1936 */
1937static void rcu_gp_cleanup(void)
1938{
1939	int cpu;
1940	bool needgp = false;
1941	unsigned long gp_duration;
1942	unsigned long new_gp_seq;
1943	bool offloaded;
1944	struct rcu_data *rdp;
1945	struct rcu_node *rnp = rcu_get_root();
1946	struct swait_queue_head *sq;
1947
1948	WRITE_ONCE(rcu_state.gp_activity, jiffies);
1949	raw_spin_lock_irq_rcu_node(rnp);
1950	rcu_state.gp_end = jiffies;
1951	gp_duration = rcu_state.gp_end - rcu_state.gp_start;
1952	if (gp_duration > rcu_state.gp_max)
1953		rcu_state.gp_max = gp_duration;
1954
1955	/*
1956	 * We know the grace period is complete, but to everyone else
1957	 * it appears to still be ongoing.  But it is also the case
1958	 * that to everyone else it looks like there is nothing that
1959	 * they can do to advance the grace period.  It is therefore
1960	 * safe for us to drop the lock in order to mark the grace
1961	 * period as completed in all of the rcu_node structures.
1962	 */
1963	raw_spin_unlock_irq_rcu_node(rnp);
1964
1965	/*
1966	 * Propagate new ->gp_seq value to rcu_node structures so that
1967	 * other CPUs don't have to wait until the start of the next grace
1968	 * period to process their callbacks.  This also avoids some nasty
1969	 * RCU grace-period initialization races by forcing the end of
1970	 * the current grace period to be completely recorded in all of
1971	 * the rcu_node structures before the beginning of the next grace
1972	 * period is recorded in any of the rcu_node structures.
1973	 */
1974	new_gp_seq = rcu_state.gp_seq;
1975	rcu_seq_end(&new_gp_seq);
1976	rcu_for_each_node_breadth_first(rnp) {
1977		raw_spin_lock_irq_rcu_node(rnp);
1978		if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
1979			dump_blkd_tasks(rnp, 10);
1980		WARN_ON_ONCE(rnp->qsmask);
1981		WRITE_ONCE(rnp->gp_seq, new_gp_seq);
1982		rdp = this_cpu_ptr(&rcu_data);
1983		if (rnp == rdp->mynode)
1984			needgp = __note_gp_changes(rnp, rdp) || needgp;
1985		/* smp_mb() provided by prior unlock-lock pair. */
1986		needgp = rcu_future_gp_cleanup(rnp) || needgp;
1987		// Reset overload indication for CPUs no longer overloaded
1988		if (rcu_is_leaf_node(rnp))
1989			for_each_leaf_node_cpu_mask(rnp, cpu, rnp->cbovldmask) {
1990				rdp = per_cpu_ptr(&rcu_data, cpu);
1991				check_cb_ovld_locked(rdp, rnp);
1992			}
1993		sq = rcu_nocb_gp_get(rnp);
1994		raw_spin_unlock_irq_rcu_node(rnp);
1995		rcu_nocb_gp_cleanup(sq);
1996		cond_resched_tasks_rcu_qs();
1997		WRITE_ONCE(rcu_state.gp_activity, jiffies);
1998		rcu_gp_slow(gp_cleanup_delay);
1999	}
2000	rnp = rcu_get_root();
2001	raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
2002
2003	/* Declare grace period done, trace first to use old GP number. */
2004	trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
2005	rcu_seq_end(&rcu_state.gp_seq);
2006	ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
2007	rcu_state.gp_state = RCU_GP_IDLE;
2008	/* Check for GP requests since above loop. */
2009	rdp = this_cpu_ptr(&rcu_data);
2010	if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
2011		trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
2012				  TPS("CleanupMore"));
2013		needgp = true;
2014	}
2015	/* Advance CBs to reduce false positives below. */
2016	offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2017		    rcu_segcblist_is_offloaded(&rdp->cblist);
2018	if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
2019		WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
2020		WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
2021		trace_rcu_grace_period(rcu_state.name,
2022				       rcu_state.gp_seq,
2023				       TPS("newreq"));
2024	} else {
2025		WRITE_ONCE(rcu_state.gp_flags,
2026			   rcu_state.gp_flags & RCU_GP_FLAG_INIT);
2027	}
2028	raw_spin_unlock_irq_rcu_node(rnp);
2029}
2030
2031/*
2032 * Body of kthread that handles grace periods.
2033 */
2034static int __noreturn rcu_gp_kthread(void *unused)
2035{
 
 
 
 
 
 
 
2036	rcu_bind_gp_kthread();
2037	for (;;) {
2038
2039		/* Handle grace-period start. */
2040		for (;;) {
2041			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
 
2042					       TPS("reqwait"));
2043			rcu_state.gp_state = RCU_GP_WAIT_GPS;
2044			swait_event_idle_exclusive(rcu_state.gp_wq,
2045					 READ_ONCE(rcu_state.gp_flags) &
2046					 RCU_GP_FLAG_INIT);
2047			rcu_gp_torture_wait();
2048			rcu_state.gp_state = RCU_GP_DONE_GPS;
2049			/* Locking provides needed memory barrier. */
2050			if (rcu_gp_init())
2051				break;
2052			cond_resched_tasks_rcu_qs();
2053			WRITE_ONCE(rcu_state.gp_activity, jiffies);
2054			WARN_ON(signal_pending(current));
2055			trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
 
2056					       TPS("reqwaitsig"));
2057		}
2058
2059		/* Handle quiescent-state forcing. */
2060		rcu_gp_fqs_loop();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2061
2062		/* Handle grace-period end. */
2063		rcu_state.gp_state = RCU_GP_CLEANUP;
2064		rcu_gp_cleanup();
2065		rcu_state.gp_state = RCU_GP_CLEANED;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2066	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2067}
2068
2069/*
2070 * Report a full set of quiescent states to the rcu_state data structure.
2071 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
2072 * another grace period is required.  Whether we wake the grace-period
2073 * kthread or it awakens itself for the next round of quiescent-state
2074 * forcing, that kthread will clean up after the just-completed grace
2075 * period.  Note that the caller must hold rnp->lock, which is released
2076 * before return.
2077 */
2078static void rcu_report_qs_rsp(unsigned long flags)
2079	__releases(rcu_get_root()->lock)
2080{
2081	raw_lockdep_assert_held_rcu_node(rcu_get_root());
2082	WARN_ON_ONCE(!rcu_gp_in_progress());
2083	WRITE_ONCE(rcu_state.gp_flags,
2084		   READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2085	raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
2086	rcu_gp_kthread_wake();
2087}
2088
2089/*
2090 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2091 * Allows quiescent states for a group of CPUs to be reported at one go
2092 * to the specified rcu_node structure, though all the CPUs in the group
2093 * must be represented by the same rcu_node structure (which need not be a
2094 * leaf rcu_node structure, though it often will be).  The gps parameter
2095 * is the grace-period snapshot, which means that the quiescent states
2096 * are valid only if rnp->gp_seq is equal to gps.  That structure's lock
2097 * must be held upon entry, and it is released before return.
2098 *
2099 * As a special case, if mask is zero, the bit-already-cleared check is
2100 * disabled.  This allows propagating quiescent state due to resumed tasks
2101 * during grace-period initialization.
2102 */
2103static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
2104			      unsigned long gps, unsigned long flags)
 
2105	__releases(rnp->lock)
2106{
2107	unsigned long oldmask = 0;
2108	struct rcu_node *rnp_c;
2109
2110	raw_lockdep_assert_held_rcu_node(rnp);
2111
2112	/* Walk up the rcu_node hierarchy. */
2113	for (;;) {
2114		if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
2115
2116			/*
2117			 * Our bit has already been cleared, or the
2118			 * relevant grace period is already over, so done.
2119			 */
2120			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2121			return;
2122		}
2123		WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2124		WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
2125			     rcu_preempt_blocked_readers_cgp(rnp));
2126		WRITE_ONCE(rnp->qsmask, rnp->qsmask & ~mask);
2127		trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
2128						 mask, rnp->qsmask, rnp->level,
2129						 rnp->grplo, rnp->grphi,
2130						 !!rnp->gp_tasks);
2131		if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2132
2133			/* Other bits still set at this level, so done. */
2134			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2135			return;
2136		}
2137		rnp->completedqs = rnp->gp_seq;
2138		mask = rnp->grpmask;
2139		if (rnp->parent == NULL) {
2140
2141			/* No more levels.  Exit loop holding root lock. */
2142
2143			break;
2144		}
2145		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2146		rnp_c = rnp;
2147		rnp = rnp->parent;
2148		raw_spin_lock_irqsave_rcu_node(rnp, flags);
2149		oldmask = READ_ONCE(rnp_c->qsmask);
2150	}
2151
2152	/*
2153	 * Get here if we are the last CPU to pass through a quiescent
2154	 * state for this grace period.  Invoke rcu_report_qs_rsp()
2155	 * to clean up and start the next grace period if one is needed.
2156	 */
2157	rcu_report_qs_rsp(flags); /* releases rnp->lock. */
2158}
2159
2160/*
2161 * Record a quiescent state for all tasks that were previously queued
2162 * on the specified rcu_node structure and that were blocking the current
2163 * RCU grace period.  The caller must hold the corresponding rnp->lock with
2164 * irqs disabled, and this lock is released upon return, but irqs remain
2165 * disabled.
2166 */
2167static void __maybe_unused
2168rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
2169	__releases(rnp->lock)
2170{
2171	unsigned long gps;
2172	unsigned long mask;
2173	struct rcu_node *rnp_p;
2174
2175	raw_lockdep_assert_held_rcu_node(rnp);
2176	if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RCU)) ||
2177	    WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
2178	    rnp->qsmask != 0) {
2179		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2180		return;  /* Still need more quiescent states! */
2181	}
2182
2183	rnp->completedqs = rnp->gp_seq;
2184	rnp_p = rnp->parent;
2185	if (rnp_p == NULL) {
2186		/*
2187		 * Only one rcu_node structure in the tree, so don't
2188		 * try to report up to its nonexistent parent!
2189		 */
2190		rcu_report_qs_rsp(flags);
2191		return;
2192	}
2193
2194	/* Report up the rest of the hierarchy, tracking current ->gp_seq. */
2195	gps = rnp->gp_seq;
2196	mask = rnp->grpmask;
2197	raw_spin_unlock_rcu_node(rnp);	/* irqs remain disabled. */
2198	raw_spin_lock_rcu_node(rnp_p);	/* irqs already disabled. */
2199	rcu_report_qs_rnp(mask, rnp_p, gps, flags);
2200}
2201
2202/*
2203 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2204 * structure.  This must be called from the specified CPU.
2205 */
2206static void
2207rcu_report_qs_rdp(int cpu, struct rcu_data *rdp)
2208{
2209	unsigned long flags;
2210	unsigned long mask;
2211	bool needwake = false;
2212	const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2213			       rcu_segcblist_is_offloaded(&rdp->cblist);
2214	struct rcu_node *rnp;
2215
2216	rnp = rdp->mynode;
2217	raw_spin_lock_irqsave_rcu_node(rnp, flags);
2218	if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
 
 
2219	    rdp->gpwrap) {
2220
2221		/*
2222		 * The grace period in which this quiescent state was
2223		 * recorded has ended, so don't report it upwards.
2224		 * We will instead need a new quiescent state that lies
2225		 * within the current grace period.
2226		 */
2227		rdp->cpu_no_qs.b.norm = true;	/* need qs for new gp. */
 
2228		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2229		return;
2230	}
2231	mask = rdp->grpmask;
2232	if (rdp->cpu == smp_processor_id())
2233		rdp->core_needs_qs = false;
2234	if ((rnp->qsmask & mask) == 0) {
2235		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2236	} else {
 
 
2237		/*
2238		 * This GP can't end until cpu checks in, so all of our
2239		 * callbacks can be processed during the next GP.
2240		 */
2241		if (!offloaded)
2242			needwake = rcu_accelerate_cbs(rnp, rdp);
2243
2244		rcu_disable_urgency_upon_qs(rdp);
2245		rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2246		/* ^^^ Released rnp->lock */
2247		if (needwake)
2248			rcu_gp_kthread_wake();
2249	}
2250}
2251
2252/*
2253 * Check to see if there is a new grace period of which this CPU
2254 * is not yet aware, and if so, set up local rcu_data state for it.
2255 * Otherwise, see if this CPU has just passed through its first
2256 * quiescent state for this grace period, and record that fact if so.
2257 */
2258static void
2259rcu_check_quiescent_state(struct rcu_data *rdp)
2260{
2261	/* Check for grace-period ends and beginnings. */
2262	note_gp_changes(rdp);
2263
2264	/*
2265	 * Does this CPU still need to do its part for current grace period?
2266	 * If no, return and let the other CPUs do their part as well.
2267	 */
2268	if (!rdp->core_needs_qs)
2269		return;
2270
2271	/*
2272	 * Was there a quiescent state since the beginning of the grace
2273	 * period? If no, then exit and wait for the next call.
2274	 */
2275	if (rdp->cpu_no_qs.b.norm)
 
2276		return;
2277
2278	/*
2279	 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2280	 * judge of that).
2281	 */
2282	rcu_report_qs_rdp(rdp->cpu, rdp);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2283}
2284
2285/*
2286 * Near the end of the offline process.  Trace the fact that this CPU
2287 * is going offline.
2288 */
2289int rcutree_dying_cpu(unsigned int cpu)
2290{
2291	bool blkd;
2292	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
2293	struct rcu_node *rnp = rdp->mynode;
2294
2295	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2296		return 0;
2297
2298	blkd = !!(rnp->qsmask & rdp->grpmask);
2299	trace_rcu_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
2300			       blkd ? TPS("cpuofl") : TPS("cpuofl-bgp"));
2301	return 0;
2302}
2303
2304/*
2305 * All CPUs for the specified rcu_node structure have gone offline,
2306 * and all tasks that were preempted within an RCU read-side critical
2307 * section while running on one of those CPUs have since exited their RCU
2308 * read-side critical section.  Some other CPU is reporting this fact with
2309 * the specified rcu_node structure's ->lock held and interrupts disabled.
2310 * This function therefore goes up the tree of rcu_node structures,
2311 * clearing the corresponding bits in the ->qsmaskinit fields.  Note that
2312 * the leaf rcu_node structure's ->qsmaskinit field has already been
2313 * updated.
2314 *
2315 * This function does check that the specified rcu_node structure has
2316 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2317 * prematurely.  That said, invoking it after the fact will cost you
2318 * a needless lock acquisition.  So once it has done its work, don't
2319 * invoke it again.
2320 */
2321static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2322{
2323	long mask;
2324	struct rcu_node *rnp = rnp_leaf;
2325
2326	raw_lockdep_assert_held_rcu_node(rnp_leaf);
2327	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2328	    WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
2329	    WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
2330		return;
2331	for (;;) {
2332		mask = rnp->grpmask;
2333		rnp = rnp->parent;
2334		if (!rnp)
2335			break;
2336		raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2337		rnp->qsmaskinit &= ~mask;
2338		/* Between grace periods, so better already be zero! */
2339		WARN_ON_ONCE(rnp->qsmask);
2340		if (rnp->qsmaskinit) {
2341			raw_spin_unlock_rcu_node(rnp);
2342			/* irqs remain disabled. */
2343			return;
2344		}
2345		raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2346	}
2347}
2348
2349/*
2350 * The CPU has been completely removed, and some other CPU is reporting
2351 * this fact from process context.  Do the remainder of the cleanup.
2352 * There can only be one CPU hotplug operation at a time, so no need for
2353 * explicit locking.
 
2354 */
2355int rcutree_dead_cpu(unsigned int cpu)
2356{
2357	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 
2358	struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
2359
2360	if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2361		return 0;
2362
2363	/* Adjust any no-longer-needed kthreads. */
2364	rcu_boost_kthread_setaffinity(rnp, -1);
2365	/* Do any needed no-CB deferred wakeups from this CPU. */
2366	do_nocb_deferred_wakeup(per_cpu_ptr(&rcu_data, cpu));
2367
2368	// Stop-machine done, so allow nohz_full to disable tick.
2369	tick_dep_clear(TICK_DEP_BIT_RCU);
2370	return 0;
 
 
 
 
 
 
2371}
2372
2373/*
2374 * Invoke any RCU callbacks that have made it to the end of their grace
2375 * period.  Thottle as specified by rdp->blimit.
2376 */
2377static void rcu_do_batch(struct rcu_data *rdp)
2378{
2379	unsigned long flags;
2380	const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2381			       rcu_segcblist_is_offloaded(&rdp->cblist);
2382	struct rcu_head *rhp;
2383	struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2384	long bl, count;
2385	long pending, tlimit = 0;
2386
2387	/* If no callbacks are ready, just return. */
2388	if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2389		trace_rcu_batch_start(rcu_state.name,
2390				      rcu_segcblist_n_cbs(&rdp->cblist), 0);
2391		trace_rcu_batch_end(rcu_state.name, 0,
2392				    !rcu_segcblist_empty(&rdp->cblist),
2393				    need_resched(), is_idle_task(current),
2394				    rcu_is_callbacks_kthread());
2395		return;
2396	}
2397
2398	/*
2399	 * Extract the list of ready callbacks, disabling to prevent
2400	 * races with call_rcu() from interrupt handlers.  Leave the
2401	 * callback counts, as rcu_barrier() needs to be conservative.
2402	 */
2403	local_irq_save(flags);
2404	rcu_nocb_lock(rdp);
2405	WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2406	pending = rcu_segcblist_n_cbs(&rdp->cblist);
2407	bl = max(rdp->blimit, pending >> rcu_divisor);
2408	if (unlikely(bl > 100))
2409		tlimit = local_clock() + rcu_resched_ns;
2410	trace_rcu_batch_start(rcu_state.name,
2411			      rcu_segcblist_n_cbs(&rdp->cblist), bl);
2412	rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2413	if (offloaded)
2414		rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2415	rcu_nocb_unlock_irqrestore(rdp, flags);
2416
2417	/* Invoke callbacks. */
2418	tick_dep_set_task(current, TICK_DEP_BIT_RCU);
2419	rhp = rcu_cblist_dequeue(&rcl);
2420	for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2421		rcu_callback_t f;
2422
2423		debug_rcu_head_unqueue(rhp);
2424
2425		rcu_lock_acquire(&rcu_callback_map);
2426		trace_rcu_invoke_callback(rcu_state.name, rhp);
2427
2428		f = rhp->func;
2429		WRITE_ONCE(rhp->func, (rcu_callback_t)0L);
2430		f(rhp);
2431
2432		rcu_lock_release(&rcu_callback_map);
2433
2434		/*
2435		 * Stop only if limit reached and CPU has something to do.
2436		 * Note: The rcl structure counts down from zero.
2437		 */
2438		if (-rcl.len >= bl && !offloaded &&
2439		    (need_resched() ||
2440		     (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2441			break;
2442		if (unlikely(tlimit)) {
2443			/* only call local_clock() every 32 callbacks */
2444			if (likely((-rcl.len & 31) || local_clock() < tlimit))
2445				continue;
2446			/* Exceeded the time limit, so leave. */
2447			break;
2448		}
2449		if (offloaded) {
2450			WARN_ON_ONCE(in_serving_softirq());
2451			local_bh_enable();
2452			lockdep_assert_irqs_enabled();
2453			cond_resched_tasks_rcu_qs();
2454			lockdep_assert_irqs_enabled();
2455			local_bh_disable();
2456		}
2457	}
2458
2459	local_irq_save(flags);
2460	rcu_nocb_lock(rdp);
2461	count = -rcl.len;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2462	rdp->n_cbs_invoked += count;
2463	trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2464			    is_idle_task(current), rcu_is_callbacks_kthread());
2465
2466	/* Update counts and requeue any remaining callbacks. */
2467	rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2468	smp_mb(); /* List handling before counting for rcu_barrier(). */
2469	rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2470
2471	/* Reinstate batch limit if we have worked down the excess. */
2472	count = rcu_segcblist_n_cbs(&rdp->cblist);
2473	if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
2474		rdp->blimit = blimit;
2475
2476	/* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2477	if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2478		rdp->qlen_last_fqs_check = 0;
2479		rdp->n_force_qs_snap = rcu_state.n_force_qs;
2480	} else if (count < rdp->qlen_last_fqs_check - qhimark)
2481		rdp->qlen_last_fqs_check = count;
 
2482
2483	/*
2484	 * The following usually indicates a double call_rcu().  To track
2485	 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2486	 */
2487	WARN_ON_ONCE(count == 0 && !rcu_segcblist_empty(&rdp->cblist));
2488	WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2489		     count != 0 && rcu_segcblist_empty(&rdp->cblist));
2490
2491	rcu_nocb_unlock_irqrestore(rdp, flags);
2492
2493	/* Re-invoke RCU core processing if there are callbacks remaining. */
2494	if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist))
2495		invoke_rcu_core();
2496	tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
2497}
2498
2499/*
2500 * This function is invoked from each scheduling-clock interrupt,
2501 * and checks to see if this CPU is in a non-context-switch quiescent
2502 * state, for example, user mode or idle loop.  It also schedules RCU
2503 * core processing.  If the current grace period has gone on too long,
2504 * it will ask the scheduler to manufacture a context switch for the sole
2505 * purpose of providing a providing the needed quiescent state.
 
2506 */
2507void rcu_sched_clock_irq(int user)
2508{
2509	trace_rcu_utilization(TPS("Start scheduler-tick"));
2510	raw_cpu_inc(rcu_data.ticks_this_gp);
2511	/* The load-acquire pairs with the store-release setting to true. */
2512	if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
2513		/* Idle and userspace execution already are quiescent states. */
2514		if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2515			set_tsk_need_resched(current);
2516			set_preempt_need_resched();
2517		}
2518		__this_cpu_write(rcu_data.rcu_urgent_qs, false);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2519	}
2520	rcu_flavor_sched_clock_irq(user);
2521	if (rcu_pending(user))
2522		invoke_rcu_core();
2523
 
2524	trace_rcu_utilization(TPS("End scheduler-tick"));
2525}
2526
2527/*
2528 * Scan the leaf rcu_node structures.  For each structure on which all
2529 * CPUs have reported a quiescent state and on which there are tasks
2530 * blocking the current grace period, initiate RCU priority boosting.
2531 * Otherwise, invoke the specified function to check dyntick state for
2532 * each CPU that has not yet reported a quiescent state.
2533 */
2534static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
 
 
 
2535{
 
2536	int cpu;
2537	unsigned long flags;
2538	unsigned long mask;
2539	struct rcu_data *rdp;
2540	struct rcu_node *rnp;
2541
2542	rcu_state.cbovld = rcu_state.cbovldnext;
2543	rcu_state.cbovldnext = false;
2544	rcu_for_each_leaf_node(rnp) {
2545		cond_resched_tasks_rcu_qs();
2546		mask = 0;
2547		raw_spin_lock_irqsave_rcu_node(rnp, flags);
2548		rcu_state.cbovldnext |= !!rnp->cbovldmask;
2549		if (rnp->qsmask == 0) {
2550			if (!IS_ENABLED(CONFIG_PREEMPT_RCU) ||
 
2551			    rcu_preempt_blocked_readers_cgp(rnp)) {
2552				/*
2553				 * No point in scanning bits because they
2554				 * are all zero.  But we might need to
2555				 * priority-boost blocked readers.
2556				 */
2557				rcu_initiate_boost(rnp, flags);
2558				/* rcu_initiate_boost() releases rnp->lock */
2559				continue;
2560			}
2561			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2562			continue;
 
 
 
 
 
 
 
 
 
2563		}
2564		for_each_leaf_node_cpu_mask(rnp, cpu, rnp->qsmask) {
2565			rdp = per_cpu_ptr(&rcu_data, cpu);
2566			if (f(rdp)) {
2567				mask |= rdp->grpmask;
2568				rcu_disable_urgency_upon_qs(rdp);
 
2569			}
2570		}
2571		if (mask != 0) {
2572			/* Idle/offline CPUs, report (releases rnp->lock). */
2573			rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2574		} else {
2575			/* Nothing to do here, so just drop the lock. */
2576			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2577		}
2578	}
2579}
2580
2581/*
2582 * Force quiescent states on reluctant CPUs, and also detect which
2583 * CPUs are in dyntick-idle mode.
2584 */
2585void rcu_force_quiescent_state(void)
2586{
2587	unsigned long flags;
2588	bool ret;
2589	struct rcu_node *rnp;
2590	struct rcu_node *rnp_old = NULL;
2591
2592	/* Funnel through hierarchy to reduce memory contention. */
2593	rnp = __this_cpu_read(rcu_data.mynode);
2594	for (; rnp != NULL; rnp = rnp->parent) {
2595		ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2596		       !raw_spin_trylock(&rnp->fqslock);
2597		if (rnp_old != NULL)
2598			raw_spin_unlock(&rnp_old->fqslock);
2599		if (ret)
 
2600			return;
 
2601		rnp_old = rnp;
2602	}
2603	/* rnp_old == rcu_get_root(), rnp == NULL. */
2604
2605	/* Reached the root of the rcu_node tree, acquire lock. */
2606	raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2607	raw_spin_unlock(&rnp_old->fqslock);
2608	if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
 
2609		raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2610		return;  /* Someone beat us to it. */
2611	}
2612	WRITE_ONCE(rcu_state.gp_flags,
2613		   READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2614	raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2615	rcu_gp_kthread_wake();
2616}
2617EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
2618
2619/* Perform RCU core processing work for the current CPU.  */
2620static __latent_entropy void rcu_core(void)
 
 
 
 
 
2621{
2622	unsigned long flags;
2623	struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2624	struct rcu_node *rnp = rdp->mynode;
2625	const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2626			       rcu_segcblist_is_offloaded(&rdp->cblist);
2627
2628	if (cpu_is_offline(smp_processor_id()))
2629		return;
2630	trace_rcu_utilization(TPS("Start RCU core"));
2631	WARN_ON_ONCE(!rdp->beenonline);
2632
2633	/* Report any deferred quiescent states if preemption enabled. */
2634	if (!(preempt_count() & PREEMPT_MASK)) {
2635		rcu_preempt_deferred_qs(current);
2636	} else if (rcu_preempt_need_deferred_qs(current)) {
2637		set_tsk_need_resched(current);
2638		set_preempt_need_resched();
2639	}
2640
2641	/* Update RCU state based on any recent quiescent states. */
2642	rcu_check_quiescent_state(rdp);
2643
2644	/* No grace period and unregistered callbacks? */
2645	if (!rcu_gp_in_progress() &&
2646	    rcu_segcblist_is_enabled(&rdp->cblist) && !offloaded) {
2647		local_irq_save(flags);
2648		if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2649			rcu_accelerate_cbs_unlocked(rnp, rdp);
 
 
 
2650		local_irq_restore(flags);
2651	}
2652
2653	rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
2654
2655	/* If there are callbacks ready, invoke them. */
2656	if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist) &&
2657	    likely(READ_ONCE(rcu_scheduler_fully_active)))
2658		rcu_do_batch(rdp);
2659
2660	/* Do any needed deferred wakeups of rcuo kthreads. */
2661	do_nocb_deferred_wakeup(rdp);
2662	trace_rcu_utilization(TPS("End RCU core"));
2663}
2664
2665static void rcu_core_si(struct softirq_action *h)
2666{
2667	rcu_core();
2668}
2669
2670static void rcu_wake_cond(struct task_struct *t, int status)
2671{
2672	/*
2673	 * If the thread is yielding, only wake it when this
2674	 * is invoked from idle
2675	 */
2676	if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
2677		wake_up_process(t);
2678}
2679
2680static void invoke_rcu_core_kthread(void)
2681{
2682	struct task_struct *t;
2683	unsigned long flags;
2684
2685	local_irq_save(flags);
2686	__this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
2687	t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
2688	if (t != NULL && t != current)
2689		rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
2690	local_irq_restore(flags);
2691}
2692
2693/*
2694 * Wake up this CPU's rcuc kthread to do RCU core processing.
2695 */
2696static void invoke_rcu_core(void)
2697{
2698	if (!cpu_online(smp_processor_id()))
 
 
2699		return;
2700	if (use_softirq)
2701		raise_softirq(RCU_SOFTIRQ);
2702	else
2703		invoke_rcu_core_kthread();
2704}
2705
2706static void rcu_cpu_kthread_park(unsigned int cpu)
2707{
2708	per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
2709}
2710
2711static int rcu_cpu_kthread_should_run(unsigned int cpu)
2712{
2713	return __this_cpu_read(rcu_data.rcu_cpu_has_work);
2714}
2715
2716/*
2717 * Per-CPU kernel thread that invokes RCU callbacks.  This replaces
2718 * the RCU softirq used in configurations of RCU that do not support RCU
2719 * priority boosting.
 
 
2720 */
2721static void rcu_cpu_kthread(unsigned int cpu)
2722{
2723	unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
2724	char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
2725	int spincnt;
2726
2727	trace_rcu_utilization(TPS("Start CPU kthread@rcu_run"));
2728	for (spincnt = 0; spincnt < 10; spincnt++) {
2729		local_bh_disable();
2730		*statusp = RCU_KTHREAD_RUNNING;
2731		local_irq_disable();
2732		work = *workp;
2733		*workp = 0;
2734		local_irq_enable();
2735		if (work)
2736			rcu_core();
2737		local_bh_enable();
2738		if (*workp == 0) {
2739			trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2740			*statusp = RCU_KTHREAD_WAITING;
2741			return;
2742		}
2743	}
2744	*statusp = RCU_KTHREAD_YIELDING;
2745	trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2746	schedule_timeout_idle(2);
2747	trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2748	*statusp = RCU_KTHREAD_WAITING;
2749}
2750
2751static struct smp_hotplug_thread rcu_cpu_thread_spec = {
2752	.store			= &rcu_data.rcu_cpu_kthread_task,
2753	.thread_should_run	= rcu_cpu_kthread_should_run,
2754	.thread_fn		= rcu_cpu_kthread,
2755	.thread_comm		= "rcuc/%u",
2756	.setup			= rcu_cpu_kthread_setup,
2757	.park			= rcu_cpu_kthread_park,
2758};
2759
2760/*
2761 * Spawn per-CPU RCU core processing kthreads.
2762 */
2763static int __init rcu_spawn_core_kthreads(void)
2764{
2765	int cpu;
2766
2767	for_each_possible_cpu(cpu)
2768		per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
2769	if (!IS_ENABLED(CONFIG_RCU_BOOST) && use_softirq)
2770		return 0;
2771	WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
2772		  "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
2773	return 0;
2774}
2775early_initcall(rcu_spawn_core_kthreads);
2776
2777/*
2778 * Handle any core-RCU processing required by a call_rcu() invocation.
2779 */
2780static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2781			    unsigned long flags)
2782{
 
 
2783	/*
2784	 * If called from an extended quiescent state, invoke the RCU
2785	 * core in order to force a re-evaluation of RCU's idleness.
2786	 */
2787	if (!rcu_is_watching())
2788		invoke_rcu_core();
2789
2790	/* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2791	if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2792		return;
2793
2794	/*
2795	 * Force the grace period if too many callbacks or too long waiting.
2796	 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2797	 * if some other CPU has recently done so.  Also, don't bother
2798	 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2799	 * is the only one waiting for a grace period to complete.
2800	 */
2801	if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2802		     rdp->qlen_last_fqs_check + qhimark)) {
2803
2804		/* Are we ignoring a completed grace period? */
2805		note_gp_changes(rdp);
2806
2807		/* Start a new grace period if one not already started. */
2808		if (!rcu_gp_in_progress()) {
2809			rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
 
 
 
 
 
 
2810		} else {
2811			/* Give the grace period a kick. */
2812			rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
2813			if (rcu_state.n_force_qs == rdp->n_force_qs_snap &&
2814			    rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2815				rcu_force_quiescent_state();
2816			rdp->n_force_qs_snap = rcu_state.n_force_qs;
2817			rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2818		}
2819	}
2820}
2821
2822/*
2823 * RCU callback function to leak a callback.
2824 */
2825static void rcu_leak_callback(struct rcu_head *rhp)
2826{
2827}
2828
2829/*
2830 * Check and if necessary update the leaf rcu_node structure's
2831 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2832 * number of queued RCU callbacks.  The caller must hold the leaf rcu_node
2833 * structure's ->lock.
2834 */
2835static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp)
2836{
2837	raw_lockdep_assert_held_rcu_node(rnp);
2838	if (qovld_calc <= 0)
2839		return; // Early boot and wildcard value set.
2840	if (rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc)
2841		WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask | rdp->grpmask);
2842	else
2843		WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask & ~rdp->grpmask);
2844}
2845
2846/*
2847 * Check and if necessary update the leaf rcu_node structure's
2848 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2849 * number of queued RCU callbacks.  No locks need be held, but the
2850 * caller must have disabled interrupts.
2851 *
2852 * Note that this function ignores the possibility that there are a lot
2853 * of callbacks all of which have already seen the end of their respective
2854 * grace periods.  This omission is due to the need for no-CBs CPUs to
2855 * be holding ->nocb_lock to do this check, which is too heavy for a
2856 * common-case operation.
2857 */
2858static void check_cb_ovld(struct rcu_data *rdp)
2859{
2860	struct rcu_node *const rnp = rdp->mynode;
2861
2862	if (qovld_calc <= 0 ||
2863	    ((rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) ==
2864	     !!(READ_ONCE(rnp->cbovldmask) & rdp->grpmask)))
2865		return; // Early boot wildcard value or already set correctly.
2866	raw_spin_lock_rcu_node(rnp);
2867	check_cb_ovld_locked(rdp, rnp);
2868	raw_spin_unlock_rcu_node(rnp);
2869}
2870
2871/* Helper function for call_rcu() and friends.  */
2872static void
2873__call_rcu(struct rcu_head *head, rcu_callback_t func)
 
2874{
2875	unsigned long flags;
2876	struct rcu_data *rdp;
2877	bool was_alldone;
2878
2879	/* Misaligned rcu_head! */
2880	WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2881
 
2882	if (debug_rcu_head_queue(head)) {
2883		/*
2884		 * Probable double call_rcu(), so leak the callback.
2885		 * Use rcu:rcu_callback trace event to find the previous
2886		 * time callback was passed to __call_rcu().
2887		 */
2888		WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pS()!!!\n",
2889			  head, head->func);
2890		WRITE_ONCE(head->func, rcu_leak_callback);
 
2891		return;
2892	}
2893	head->func = func;
2894	head->next = NULL;
 
 
 
 
 
 
 
2895	local_irq_save(flags);
2896	kasan_record_aux_stack(head);
2897	rdp = this_cpu_ptr(&rcu_data);
2898
2899	/* Add the callback to our list. */
2900	if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
2901		// This can trigger due to call_rcu() from offline CPU:
2902		WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2903		WARN_ON_ONCE(!rcu_is_watching());
2904		// Very early boot, before rcu_init().  Initialize if needed
2905		// and then drop through to queue the callback.
2906		if (rcu_segcblist_empty(&rdp->cblist))
2907			rcu_segcblist_init(&rdp->cblist);
2908	}
2909
2910	check_cb_ovld(rdp);
2911	if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags))
2912		return; // Enqueued onto ->nocb_bypass, so just leave.
2913	// If no-CBs CPU gets here, rcu_nocb_try_bypass() acquired ->nocb_lock.
2914	rcu_segcblist_enqueue(&rdp->cblist, head);
2915	if (__is_kvfree_rcu_offset((unsigned long)func))
2916		trace_rcu_kvfree_callback(rcu_state.name, head,
2917					 (unsigned long)func,
2918					 rcu_segcblist_n_cbs(&rdp->cblist));
 
 
2919	else
2920		trace_rcu_callback(rcu_state.name, head,
2921				   rcu_segcblist_n_cbs(&rdp->cblist));
2922
2923	/* Go handle any RCU core processing required. */
2924	if (IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2925	    unlikely(rcu_segcblist_is_offloaded(&rdp->cblist))) {
2926		__call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */
2927	} else {
2928		__call_rcu_core(rdp, head, flags);
2929		local_irq_restore(flags);
2930	}
 
 
 
2931}
 
2932
2933/**
2934 * call_rcu() - Queue an RCU callback for invocation after a grace period.
2935 * @head: structure to be used for queueing the RCU updates.
2936 * @func: actual callback function to be invoked after the grace period
2937 *
2938 * The callback function will be invoked some time after a full grace
2939 * period elapses, in other words after all pre-existing RCU read-side
2940 * critical sections have completed.  However, the callback function
2941 * might well execute concurrently with RCU read-side critical sections
2942 * that started after call_rcu() was invoked.  RCU read-side critical
2943 * sections are delimited by rcu_read_lock() and rcu_read_unlock(), and
2944 * may be nested.  In addition, regions of code across which interrupts,
2945 * preemption, or softirqs have been disabled also serve as RCU read-side
2946 * critical sections.  This includes hardware interrupt handlers, softirq
2947 * handlers, and NMI handlers.
2948 *
2949 * Note that all CPUs must agree that the grace period extended beyond
2950 * all pre-existing RCU read-side critical section.  On systems with more
2951 * than one CPU, this means that when "func()" is invoked, each CPU is
2952 * guaranteed to have executed a full memory barrier since the end of its
2953 * last RCU read-side critical section whose beginning preceded the call
2954 * to call_rcu().  It also means that each CPU executing an RCU read-side
2955 * critical section that continues beyond the start of "func()" must have
2956 * executed a memory barrier after the call_rcu() but before the beginning
2957 * of that RCU read-side critical section.  Note that these guarantees
2958 * include CPUs that are offline, idle, or executing in user mode, as
2959 * well as CPUs that are executing in the kernel.
2960 *
2961 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
2962 * resulting RCU callback function "func()", then both CPU A and CPU B are
2963 * guaranteed to execute a full memory barrier during the time interval
2964 * between the call to call_rcu() and the invocation of "func()" -- even
2965 * if CPU A and CPU B are the same CPU (but again only if the system has
2966 * more than one CPU).
2967 */
2968void call_rcu(struct rcu_head *head, rcu_callback_t func)
2969{
2970	__call_rcu(head, func);
2971}
2972EXPORT_SYMBOL_GPL(call_rcu);
2973
2974
2975/* Maximum number of jiffies to wait before draining a batch. */
2976#define KFREE_DRAIN_JIFFIES (HZ / 50)
2977#define KFREE_N_BATCHES 2
2978#define FREE_N_CHANNELS 2
2979
2980/**
2981 * struct kvfree_rcu_bulk_data - single block to store kvfree_rcu() pointers
2982 * @nr_records: Number of active pointers in the array
2983 * @next: Next bulk object in the block chain
2984 * @records: Array of the kvfree_rcu() pointers
2985 */
2986struct kvfree_rcu_bulk_data {
2987	unsigned long nr_records;
2988	struct kvfree_rcu_bulk_data *next;
2989	void *records[];
2990};
 
 
2991
2992/*
2993 * This macro defines how many entries the "records" array
2994 * will contain. It is based on the fact that the size of
2995 * kvfree_rcu_bulk_data structure becomes exactly one page.
 
 
 
 
2996 */
2997#define KVFREE_BULK_MAX_ENTR \
2998	((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *))
 
 
 
 
 
 
 
 
2999
3000/**
3001 * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
3002 * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
3003 * @head_free: List of kfree_rcu() objects waiting for a grace period
3004 * @bkvhead_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
3005 * @krcp: Pointer to @kfree_rcu_cpu structure
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3006 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3007
3008struct kfree_rcu_cpu_work {
3009	struct rcu_work rcu_work;
3010	struct rcu_head *head_free;
3011	struct kvfree_rcu_bulk_data *bkvhead_free[FREE_N_CHANNELS];
3012	struct kfree_rcu_cpu *krcp;
3013};
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3014
3015/**
3016 * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
3017 * @head: List of kfree_rcu() objects not yet waiting for a grace period
3018 * @bkvhead: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period
3019 * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
3020 * @lock: Synchronize access to this structure
3021 * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES
3022 * @monitor_todo: Tracks whether a @monitor_work delayed work is pending
3023 * @initialized: The @rcu_work fields have been initialized
3024 * @count: Number of objects for which GP not started
3025 *
3026 * This is a per-CPU structure.  The reason that it is not included in
3027 * the rcu_data structure is to permit this code to be extracted from
3028 * the RCU files.  Such extraction could allow further optimization of
3029 * the interactions with the slab allocators.
3030 */
3031struct kfree_rcu_cpu {
3032	struct rcu_head *head;
3033	struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS];
3034	struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
3035	raw_spinlock_t lock;
3036	struct delayed_work monitor_work;
3037	bool monitor_todo;
3038	bool initialized;
3039	int count;
3040
3041	/*
3042	 * A simple cache list that contains objects for
3043	 * reuse purpose. In order to save some per-cpu
3044	 * space the list is singular. Even though it is
3045	 * lockless an access has to be protected by the
3046	 * per-cpu lock.
3047	 */
3048	struct llist_head bkvcache;
3049	int nr_bkv_objs;
3050};
3051
3052static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc) = {
3053	.lock = __RAW_SPIN_LOCK_UNLOCKED(krc.lock),
3054};
 
 
 
 
 
3055
3056static __always_inline void
3057debug_rcu_bhead_unqueue(struct kvfree_rcu_bulk_data *bhead)
 
 
 
 
 
 
 
 
 
 
 
 
 
3058{
3059#ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
3060	int i;
3061
3062	for (i = 0; i < bhead->nr_records; i++)
3063		debug_rcu_head_unqueue((struct rcu_head *)(bhead->records[i]));
3064#endif
 
 
 
 
3065}
 
3066
3067static inline struct kfree_rcu_cpu *
3068krc_this_cpu_lock(unsigned long *flags)
 
 
 
 
 
 
3069{
3070	struct kfree_rcu_cpu *krcp;
 
 
 
 
 
 
 
 
 
 
 
 
 
3071
3072	local_irq_save(*flags);	// For safely calling this_cpu_ptr().
3073	krcp = this_cpu_ptr(&krc);
3074	raw_spin_lock(&krcp->lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3075
3076	return krcp;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3077}
3078
3079static inline void
3080krc_this_cpu_unlock(struct kfree_rcu_cpu *krcp, unsigned long flags)
3081{
3082	raw_spin_unlock(&krcp->lock);
3083	local_irq_restore(flags);
 
3084}
3085
3086static inline struct kvfree_rcu_bulk_data *
3087get_cached_bnode(struct kfree_rcu_cpu *krcp)
3088{
3089	if (!krcp->nr_bkv_objs)
3090		return NULL;
3091
3092	krcp->nr_bkv_objs--;
3093	return (struct kvfree_rcu_bulk_data *)
3094		llist_del_first(&krcp->bkvcache);
3095}
3096
3097static inline bool
3098put_cached_bnode(struct kfree_rcu_cpu *krcp,
3099	struct kvfree_rcu_bulk_data *bnode)
 
 
3100{
3101	// Check the limit.
3102	if (krcp->nr_bkv_objs >= rcu_min_cached_objs)
3103		return false;
3104
3105	llist_add((struct llist_node *) bnode, &krcp->bkvcache);
3106	krcp->nr_bkv_objs++;
3107	return true;
3108
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3109}
3110
3111/*
3112 * This function is invoked in workqueue context after a grace period.
3113 * It frees all the objects queued on ->bhead_free or ->head_free.
 
 
 
3114 */
3115static void kfree_rcu_work(struct work_struct *work)
3116{
 
3117	unsigned long flags;
3118	struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS], *bnext;
3119	struct rcu_head *head, *next;
3120	struct kfree_rcu_cpu *krcp;
3121	struct kfree_rcu_cpu_work *krwp;
3122	int i, j;
3123
3124	krwp = container_of(to_rcu_work(work),
3125			    struct kfree_rcu_cpu_work, rcu_work);
3126	krcp = krwp->krcp;
3127
3128	raw_spin_lock_irqsave(&krcp->lock, flags);
3129	// Channels 1 and 2.
3130	for (i = 0; i < FREE_N_CHANNELS; i++) {
3131		bkvhead[i] = krwp->bkvhead_free[i];
3132		krwp->bkvhead_free[i] = NULL;
3133	}
3134
3135	// Channel 3.
3136	head = krwp->head_free;
3137	krwp->head_free = NULL;
3138	raw_spin_unlock_irqrestore(&krcp->lock, flags);
3139
3140	// Handle two first channels.
3141	for (i = 0; i < FREE_N_CHANNELS; i++) {
3142		for (; bkvhead[i]; bkvhead[i] = bnext) {
3143			bnext = bkvhead[i]->next;
3144			debug_rcu_bhead_unqueue(bkvhead[i]);
3145
3146			rcu_lock_acquire(&rcu_callback_map);
3147			if (i == 0) { // kmalloc() / kfree().
3148				trace_rcu_invoke_kfree_bulk_callback(
3149					rcu_state.name, bkvhead[i]->nr_records,
3150					bkvhead[i]->records);
3151
3152				kfree_bulk(bkvhead[i]->nr_records,
3153					bkvhead[i]->records);
3154			} else { // vmalloc() / vfree().
3155				for (j = 0; j < bkvhead[i]->nr_records; j++) {
3156					trace_rcu_invoke_kvfree_callback(
3157						rcu_state.name,
3158						bkvhead[i]->records[j], 0);
3159
3160					vfree(bkvhead[i]->records[j]);
3161				}
3162			}
3163			rcu_lock_release(&rcu_callback_map);
 
 
 
 
 
 
3164
3165			krcp = krc_this_cpu_lock(&flags);
3166			if (put_cached_bnode(krcp, bkvhead[i]))
3167				bkvhead[i] = NULL;
3168			krc_this_cpu_unlock(krcp, flags);
3169
3170			if (bkvhead[i])
3171				free_page((unsigned long) bkvhead[i]);
 
3172
3173			cond_resched_tasks_rcu_qs();
 
 
 
 
 
 
 
 
 
 
 
 
 
3174		}
3175	}
 
3176
3177	/*
3178	 * Emergency case only. It can happen under low memory
3179	 * condition when an allocation gets failed, so the "bulk"
3180	 * path can not be temporary maintained.
3181	 */
3182	for (; head; head = next) {
3183		unsigned long offset = (unsigned long)head->func;
3184		void *ptr = (void *)head - offset;
3185
3186		next = head->next;
3187		debug_rcu_head_unqueue((struct rcu_head *)ptr);
3188		rcu_lock_acquire(&rcu_callback_map);
3189		trace_rcu_invoke_kvfree_callback(rcu_state.name, head, offset);
3190
3191		if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset)))
3192			kvfree(ptr);
3193
3194		rcu_lock_release(&rcu_callback_map);
3195		cond_resched_tasks_rcu_qs();
3196	}
3197}
3198
3199/*
3200 * Schedule the kfree batch RCU work to run in workqueue context after a GP.
 
 
 
 
3201 *
3202 * This function is invoked by kfree_rcu_monitor() when the KFREE_DRAIN_JIFFIES
3203 * timeout has been reached.
3204 */
3205static inline bool queue_kfree_rcu_work(struct kfree_rcu_cpu *krcp)
3206{
3207	struct kfree_rcu_cpu_work *krwp;
3208	bool repeat = false;
3209	int i, j;
3210
3211	lockdep_assert_held(&krcp->lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3212
3213	for (i = 0; i < KFREE_N_BATCHES; i++) {
3214		krwp = &(krcp->krw_arr[i]);
3215
3216		/*
3217		 * Try to detach bkvhead or head and attach it over any
3218		 * available corresponding free channel. It can be that
3219		 * a previous RCU batch is in progress, it means that
3220		 * immediately to queue another one is not possible so
3221		 * return false to tell caller to retry.
3222		 */
3223		if ((krcp->bkvhead[0] && !krwp->bkvhead_free[0]) ||
3224			(krcp->bkvhead[1] && !krwp->bkvhead_free[1]) ||
3225				(krcp->head && !krwp->head_free)) {
3226			// Channel 1 corresponds to SLAB ptrs.
3227			// Channel 2 corresponds to vmalloc ptrs.
3228			for (j = 0; j < FREE_N_CHANNELS; j++) {
3229				if (!krwp->bkvhead_free[j]) {
3230					krwp->bkvhead_free[j] = krcp->bkvhead[j];
3231					krcp->bkvhead[j] = NULL;
3232				}
3233			}
3234
3235			// Channel 3 corresponds to emergency path.
3236			if (!krwp->head_free) {
3237				krwp->head_free = krcp->head;
3238				krcp->head = NULL;
3239			}
3240
3241			WRITE_ONCE(krcp->count, 0);
3242
3243			/*
3244			 * One work is per one batch, so there are three
3245			 * "free channels", the batch can handle. It can
3246			 * be that the work is in the pending state when
3247			 * channels have been detached following by each
3248			 * other.
3249			 */
3250			queue_rcu_work(system_wq, &krwp->rcu_work);
3251		}
3252
3253		// Repeat if any "free" corresponding channel is still busy.
3254		if (krcp->bkvhead[0] || krcp->bkvhead[1] || krcp->head)
3255			repeat = true;
 
 
3256	}
3257
3258	return !repeat;
3259}
3260
3261static inline void kfree_rcu_drain_unlock(struct kfree_rcu_cpu *krcp,
3262					  unsigned long flags)
 
 
 
 
 
 
3263{
3264	// Attempt to start a new batch.
3265	krcp->monitor_todo = false;
3266	if (queue_kfree_rcu_work(krcp)) {
3267		// Success! Our job is done here.
3268		raw_spin_unlock_irqrestore(&krcp->lock, flags);
3269		return;
3270	}
3271
3272	// Previous RCU batch still in progress, try again later.
3273	krcp->monitor_todo = true;
3274	schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES);
3275	raw_spin_unlock_irqrestore(&krcp->lock, flags);
3276}
3277
3278/*
3279 * This function is invoked after the KFREE_DRAIN_JIFFIES timeout.
3280 * It invokes kfree_rcu_drain_unlock() to attempt to start another batch.
 
3281 */
3282static void kfree_rcu_monitor(struct work_struct *work)
 
3283{
3284	unsigned long flags;
3285	struct kfree_rcu_cpu *krcp = container_of(work, struct kfree_rcu_cpu,
3286						 monitor_work.work);
3287
3288	raw_spin_lock_irqsave(&krcp->lock, flags);
3289	if (krcp->monitor_todo)
3290		kfree_rcu_drain_unlock(krcp, flags);
3291	else
3292		raw_spin_unlock_irqrestore(&krcp->lock, flags);
3293}
3294
3295static inline bool
3296kvfree_call_rcu_add_ptr_to_bulk(struct kfree_rcu_cpu *krcp, void *ptr)
3297{
3298	struct kvfree_rcu_bulk_data *bnode;
3299	int idx;
3300
3301	if (unlikely(!krcp->initialized))
3302		return false;
3303
3304	lockdep_assert_held(&krcp->lock);
3305	idx = !!is_vmalloc_addr(ptr);
3306
3307	/* Check if a new block is required. */
3308	if (!krcp->bkvhead[idx] ||
3309			krcp->bkvhead[idx]->nr_records == KVFREE_BULK_MAX_ENTR) {
3310		bnode = get_cached_bnode(krcp);
3311		if (!bnode) {
3312			/*
3313			 * To keep this path working on raw non-preemptible
3314			 * sections, prevent the optional entry into the
3315			 * allocator as it uses sleeping locks. In fact, even
3316			 * if the caller of kfree_rcu() is preemptible, this
3317			 * path still is not, as krcp->lock is a raw spinlock.
3318			 * With additional page pre-allocation in the works,
3319			 * hitting this return is going to be much less likely.
3320			 */
3321			if (IS_ENABLED(CONFIG_PREEMPT_RT))
3322				return false;
3323
3324			/*
3325			 * NOTE: For one argument of kvfree_rcu() we can
3326			 * drop the lock and get the page in sleepable
3327			 * context. That would allow to maintain an array
3328			 * for the CONFIG_PREEMPT_RT as well if no cached
3329			 * pages are available.
3330			 */
3331			bnode = (struct kvfree_rcu_bulk_data *)
3332				__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
3333		}
3334
3335		/* Switch to emergency path. */
3336		if (unlikely(!bnode))
3337			return false;
3338
3339		/* Initialize the new block. */
3340		bnode->nr_records = 0;
3341		bnode->next = krcp->bkvhead[idx];
3342
3343		/* Attach it to the head. */
3344		krcp->bkvhead[idx] = bnode;
3345	}
3346
3347	/* Finally insert. */
3348	krcp->bkvhead[idx]->records
3349		[krcp->bkvhead[idx]->nr_records++] = ptr;
3350
3351	return true;
3352}
3353
3354/*
3355 * Queue a request for lazy invocation of appropriate free routine after a
3356 * grace period. Please note there are three paths are maintained, two are the
3357 * main ones that use array of pointers interface and third one is emergency
3358 * one, that is used only when the main path can not be maintained temporary,
3359 * due to memory pressure.
3360 *
3361 * Each kvfree_call_rcu() request is added to a batch. The batch will be drained
3362 * every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch will
3363 * be free'd in workqueue context. This allows us to: batch requests together to
3364 * reduce the number of grace periods during heavy kfree_rcu()/kvfree_rcu() load.
3365 */
3366void kvfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
 
3367{
3368	unsigned long flags;
3369	struct kfree_rcu_cpu *krcp;
3370	bool success;
3371	void *ptr;
3372
3373	if (head) {
3374		ptr = (void *) head - (unsigned long) func;
3375	} else {
3376		/*
3377		 * Please note there is a limitation for the head-less
3378		 * variant, that is why there is a clear rule for such
3379		 * objects: it can be used from might_sleep() context
3380		 * only. For other places please embed an rcu_head to
3381		 * your data.
3382		 */
3383		might_sleep();
3384		ptr = (unsigned long *) func;
 
 
3385	}
 
 
3386
3387	krcp = krc_this_cpu_lock(&flags);
3388
3389	// Queue the object but don't yet schedule the batch.
3390	if (debug_rcu_head_queue(ptr)) {
3391		// Probable double kfree_rcu(), just leak.
3392		WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n",
3393			  __func__, head);
3394
3395		// Mark as success and leave.
3396		success = true;
3397		goto unlock_return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3398	}
3399
3400	/*
3401	 * Under high memory pressure GFP_NOWAIT can fail,
3402	 * in that case the emergency path is maintained.
 
 
 
 
3403	 */
3404	success = kvfree_call_rcu_add_ptr_to_bulk(krcp, ptr);
3405	if (!success) {
3406		if (head == NULL)
3407			// Inline if kvfree_rcu(one_arg) call.
3408			goto unlock_return;
3409
3410		head->func = func;
3411		head->next = krcp->head;
3412		krcp->head = head;
3413		success = true;
 
 
 
 
3414	}
 
 
 
 
 
3415
3416	WRITE_ONCE(krcp->count, krcp->count + 1);
 
 
 
 
 
3417
3418	// Set timer to drain after KFREE_DRAIN_JIFFIES.
3419	if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
3420	    !krcp->monitor_todo) {
3421		krcp->monitor_todo = true;
3422		schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES);
3423	}
3424
3425unlock_return:
3426	krc_this_cpu_unlock(krcp, flags);
3427
3428	/*
3429	 * Inline kvfree() after synchronize_rcu(). We can do
3430	 * it from might_sleep() context only, so the current
3431	 * CPU can pass the QS state.
3432	 */
3433	if (!success) {
3434		debug_rcu_head_unqueue((struct rcu_head *) ptr);
3435		synchronize_rcu();
3436		kvfree(ptr);
3437	}
3438}
3439EXPORT_SYMBOL_GPL(kvfree_call_rcu);
3440
3441static unsigned long
3442kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
3443{
3444	int cpu;
3445	unsigned long count = 0;
 
 
3446
3447	/* Snapshot count of all CPUs */
3448	for_each_online_cpu(cpu) {
3449		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3450
3451		count += READ_ONCE(krcp->count);
3452	}
3453
3454	return count;
3455}
3456
3457static unsigned long
3458kfree_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
 
 
 
 
3459{
3460	int cpu, freed = 0;
3461	unsigned long flags;
 
 
 
 
 
 
 
 
 
3462
3463	for_each_online_cpu(cpu) {
3464		int count;
3465		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3466
3467		count = krcp->count;
3468		raw_spin_lock_irqsave(&krcp->lock, flags);
3469		if (krcp->monitor_todo)
3470			kfree_rcu_drain_unlock(krcp, flags);
3471		else
3472			raw_spin_unlock_irqrestore(&krcp->lock, flags);
 
3473
3474		sc->nr_to_scan -= count;
3475		freed += count;
 
 
 
 
 
 
3476
3477		if (sc->nr_to_scan <= 0)
3478			break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3479	}
3480
3481	return freed == 0 ? SHRINK_STOP : freed;
3482}
3483
3484static struct shrinker kfree_rcu_shrinker = {
3485	.count_objects = kfree_rcu_shrink_count,
3486	.scan_objects = kfree_rcu_shrink_scan,
3487	.batch = 0,
3488	.seeks = DEFAULT_SEEKS,
3489};
3490
3491void __init kfree_rcu_scheduler_running(void)
3492{
3493	int cpu;
3494	unsigned long flags;
 
 
 
 
 
 
3495
3496	for_each_online_cpu(cpu) {
3497		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3498
3499		raw_spin_lock_irqsave(&krcp->lock, flags);
3500		if (!krcp->head || krcp->monitor_todo) {
3501			raw_spin_unlock_irqrestore(&krcp->lock, flags);
3502			continue;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3503		}
3504		krcp->monitor_todo = true;
3505		schedule_delayed_work_on(cpu, &krcp->monitor_work,
3506					 KFREE_DRAIN_JIFFIES);
3507		raw_spin_unlock_irqrestore(&krcp->lock, flags);
3508	}
3509}
3510
3511/*
3512 * During early boot, any blocking grace-period wait automatically
3513 * implies a grace period.  Later on, this is never the case for PREEMPTION.
3514 *
3515 * Howevr, because a context switch is a grace period for !PREEMPTION, any
3516 * blocking grace-period wait automatically implies a grace period if
3517 * there is only one CPU online at any point time during execution of
3518 * either synchronize_rcu() or synchronize_rcu_expedited().  It is OK to
3519 * occasionally incorrectly indicate that there are multiple CPUs online
3520 * when there was in fact only one the whole time, as this just adds some
3521 * overhead: RCU still operates correctly.
 
 
 
 
 
3522 */
3523static int rcu_blocking_is_gp(void)
3524{
3525	int ret;
 
 
3526
3527	if (IS_ENABLED(CONFIG_PREEMPTION))
3528		return rcu_scheduler_active == RCU_SCHEDULER_INACTIVE;
3529	might_sleep();  /* Check for RCU read-side critical section. */
3530	preempt_disable();
3531	ret = num_online_cpus() <= 1;
3532	preempt_enable();
3533	return ret;
3534}
3535
3536/**
3537 * synchronize_rcu - wait until a grace period has elapsed.
3538 *
3539 * Control will return to the caller some time after a full grace
3540 * period has elapsed, in other words after all currently executing RCU
3541 * read-side critical sections have completed.  Note, however, that
3542 * upon return from synchronize_rcu(), the caller might well be executing
3543 * concurrently with new RCU read-side critical sections that began while
3544 * synchronize_rcu() was waiting.  RCU read-side critical sections are
3545 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
3546 * In addition, regions of code across which interrupts, preemption, or
3547 * softirqs have been disabled also serve as RCU read-side critical
3548 * sections.  This includes hardware interrupt handlers, softirq handlers,
3549 * and NMI handlers.
3550 *
3551 * Note that this guarantee implies further memory-ordering guarantees.
3552 * On systems with more than one CPU, when synchronize_rcu() returns,
3553 * each CPU is guaranteed to have executed a full memory barrier since
3554 * the end of its last RCU read-side critical section whose beginning
3555 * preceded the call to synchronize_rcu().  In addition, each CPU having
3556 * an RCU read-side critical section that extends beyond the return from
3557 * synchronize_rcu() is guaranteed to have executed a full memory barrier
3558 * after the beginning of synchronize_rcu() and before the beginning of
3559 * that RCU read-side critical section.  Note that these guarantees include
3560 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3561 * that are executing in the kernel.
3562 *
3563 * Furthermore, if CPU A invoked synchronize_rcu(), which returned
3564 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3565 * to have executed a full memory barrier during the execution of
3566 * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
3567 * again only if the system has more than one CPU).
3568 */
3569void synchronize_rcu(void)
3570{
3571	RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3572			 lock_is_held(&rcu_lock_map) ||
3573			 lock_is_held(&rcu_sched_lock_map),
3574			 "Illegal synchronize_rcu() in RCU read-side critical section");
3575	if (rcu_blocking_is_gp())
3576		return;
3577	if (rcu_gp_is_expedited())
3578		synchronize_rcu_expedited();
3579	else
3580		wait_rcu_gp(call_rcu);
3581}
3582EXPORT_SYMBOL_GPL(synchronize_rcu);
 
 
3583
3584/**
3585 * get_state_synchronize_rcu - Snapshot current RCU state
3586 *
3587 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3588 * to determine whether or not a full grace period has elapsed in the
3589 * meantime.
3590 */
3591unsigned long get_state_synchronize_rcu(void)
3592{
3593	/*
3594	 * Any prior manipulation of RCU-protected data must happen
3595	 * before the load from ->gp_seq.
3596	 */
3597	smp_mb();  /* ^^^ */
3598	return rcu_seq_snap(&rcu_state.gp_seq);
3599}
3600EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3601
3602/**
3603 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3604 *
3605 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3606 *
3607 * If a full RCU grace period has elapsed since the earlier call to
3608 * get_state_synchronize_rcu(), just return.  Otherwise, invoke
3609 * synchronize_rcu() to wait for a full grace period.
3610 *
3611 * Yes, this function does not take counter wrap into account.  But
3612 * counter wrap is harmless.  If the counter wraps, we have waited for
3613 * more than 2 billion grace periods (and way more on a 64-bit system!),
3614 * so waiting for one additional grace period should be just fine.
3615 */
3616void cond_synchronize_rcu(unsigned long oldstate)
3617{
3618	if (!rcu_seq_done(&rcu_state.gp_seq, oldstate))
3619		synchronize_rcu();
3620	else
3621		smp_mb(); /* Ensure GP ends before subsequent accesses. */
3622}
3623EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3624
3625/*
3626 * Check to see if there is any immediate RCU-related work to be done by
3627 * the current CPU, returning 1 if so and zero otherwise.  The checks are
3628 * in order of increasing expense: checks that can be carried out against
3629 * CPU-local state are performed first.  However, we must check for CPU
3630 * stalls first, else we might not get a chance.
3631 */
3632static int rcu_pending(int user)
3633{
3634	bool gp_in_progress;
3635	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
3636	struct rcu_node *rnp = rdp->mynode;
3637
 
 
3638	/* Check for CPU stalls, if enabled. */
3639	check_cpu_stall(rdp);
3640
3641	/* Does this CPU need a deferred NOCB wakeup? */
3642	if (rcu_nocb_need_deferred_wakeup(rdp))
3643		return 1;
3644
3645	/* Is this a nohz_full CPU in userspace or idle?  (Ignore RCU if so.) */
3646	if ((user || rcu_is_cpu_rrupt_from_idle()) && rcu_nohz_full_cpu())
3647		return 0;
3648
3649	/* Is the RCU core waiting for a quiescent state from this CPU? */
3650	gp_in_progress = rcu_gp_in_progress();
3651	if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm && gp_in_progress)
 
 
 
 
 
 
3652		return 1;
 
3653
3654	/* Does this CPU have callbacks ready to invoke? */
3655	if (rcu_segcblist_ready_cbs(&rdp->cblist))
 
3656		return 1;
 
3657
3658	/* Has RCU gone idle with this CPU needing another grace period? */
3659	if (!gp_in_progress && rcu_segcblist_is_enabled(&rdp->cblist) &&
3660	    (!IS_ENABLED(CONFIG_RCU_NOCB_CPU) ||
3661	     !rcu_segcblist_is_offloaded(&rdp->cblist)) &&
3662	    !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
 
 
 
 
3663		return 1;
 
 
 
 
 
 
 
 
3664
3665	/* Have RCU grace period completed or started?  */
3666	if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
3667	    unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
3668		return 1;
 
3669
3670	/* nothing to do */
 
3671	return 0;
3672}
3673
3674/*
3675 * Helper function for rcu_barrier() tracing.  If tracing is disabled,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3676 * the compiler is expected to optimize this away.
3677 */
3678static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
 
3679{
3680	trace_rcu_barrier(rcu_state.name, s, cpu,
3681			  atomic_read(&rcu_state.barrier_cpu_count), done);
3682}
3683
3684/*
3685 * RCU callback function for rcu_barrier().  If we are last, wake
3686 * up the task executing rcu_barrier().
3687 *
3688 * Note that the value of rcu_state.barrier_sequence must be captured
3689 * before the atomic_dec_and_test().  Otherwise, if this CPU is not last,
3690 * other CPUs might count the value down to zero before this CPU gets
3691 * around to invoking rcu_barrier_trace(), which might result in bogus
3692 * data from the next instance of rcu_barrier().
3693 */
3694static void rcu_barrier_callback(struct rcu_head *rhp)
3695{
3696	unsigned long __maybe_unused s = rcu_state.barrier_sequence;
 
3697
3698	if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
3699		rcu_barrier_trace(TPS("LastCB"), -1, s);
3700		complete(&rcu_state.barrier_completion);
3701	} else {
3702		rcu_barrier_trace(TPS("CB"), -1, s);
3703	}
3704}
3705
3706/*
3707 * Called with preemption disabled, and from cross-cpu IRQ context.
3708 */
3709static void rcu_barrier_func(void *cpu_in)
3710{
3711	uintptr_t cpu = (uintptr_t)cpu_in;
3712	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3713
3714	rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
3715	rdp->barrier_head.func = rcu_barrier_callback;
3716	debug_rcu_head_queue(&rdp->barrier_head);
3717	rcu_nocb_lock(rdp);
3718	WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
3719	if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head)) {
3720		atomic_inc(&rcu_state.barrier_cpu_count);
3721	} else {
3722		debug_rcu_head_unqueue(&rdp->barrier_head);
3723		rcu_barrier_trace(TPS("IRQNQ"), -1,
3724				  rcu_state.barrier_sequence);
3725	}
3726	rcu_nocb_unlock(rdp);
3727}
3728
3729/**
3730 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
3731 *
3732 * Note that this primitive does not necessarily wait for an RCU grace period
3733 * to complete.  For example, if there are no RCU callbacks queued anywhere
3734 * in the system, then rcu_barrier() is within its rights to return
3735 * immediately, without waiting for anything, much less an RCU grace period.
3736 */
3737void rcu_barrier(void)
3738{
3739	uintptr_t cpu;
3740	struct rcu_data *rdp;
3741	unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
3742
3743	rcu_barrier_trace(TPS("Begin"), -1, s);
3744
3745	/* Take mutex to serialize concurrent rcu_barrier() requests. */
3746	mutex_lock(&rcu_state.barrier_mutex);
3747
3748	/* Did someone else do our work for us? */
3749	if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
3750		rcu_barrier_trace(TPS("EarlyExit"), -1,
3751				  rcu_state.barrier_sequence);
3752		smp_mb(); /* caller's subsequent code after above check. */
3753		mutex_unlock(&rcu_state.barrier_mutex);
3754		return;
3755	}
3756
3757	/* Mark the start of the barrier operation. */
3758	rcu_seq_start(&rcu_state.barrier_sequence);
3759	rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
3760
3761	/*
3762	 * Initialize the count to two rather than to zero in order
3763	 * to avoid a too-soon return to zero in case of an immediate
3764	 * invocation of the just-enqueued callback (or preemption of
3765	 * this task).  Exclude CPU-hotplug operations to ensure that no
3766	 * offline non-offloaded CPU has callbacks queued.
3767	 */
3768	init_completion(&rcu_state.barrier_completion);
3769	atomic_set(&rcu_state.barrier_cpu_count, 2);
3770	get_online_cpus();
3771
3772	/*
3773	 * Force each CPU with callbacks to register a new callback.
3774	 * When that callback is invoked, we will know that all of the
3775	 * corresponding CPU's preceding callbacks have been invoked.
3776	 */
3777	for_each_possible_cpu(cpu) {
3778		rdp = per_cpu_ptr(&rcu_data, cpu);
3779		if (cpu_is_offline(cpu) &&
3780		    !rcu_segcblist_is_offloaded(&rdp->cblist))
3781			continue;
3782		if (rcu_segcblist_n_cbs(&rdp->cblist) && cpu_online(cpu)) {
3783			rcu_barrier_trace(TPS("OnlineQ"), cpu,
3784					  rcu_state.barrier_sequence);
3785			smp_call_function_single(cpu, rcu_barrier_func, (void *)cpu, 1);
3786		} else if (rcu_segcblist_n_cbs(&rdp->cblist) &&
3787			   cpu_is_offline(cpu)) {
3788			rcu_barrier_trace(TPS("OfflineNoCBQ"), cpu,
3789					  rcu_state.barrier_sequence);
3790			local_irq_disable();
3791			rcu_barrier_func((void *)cpu);
3792			local_irq_enable();
3793		} else if (cpu_is_offline(cpu)) {
3794			rcu_barrier_trace(TPS("OfflineNoCBNoQ"), cpu,
3795					  rcu_state.barrier_sequence);
 
 
 
3796		} else {
3797			rcu_barrier_trace(TPS("OnlineNQ"), cpu,
3798					  rcu_state.barrier_sequence);
3799		}
3800	}
3801	put_online_cpus();
3802
3803	/*
3804	 * Now that we have an rcu_barrier_callback() callback on each
3805	 * CPU, and thus each counted, remove the initial count.
3806	 */
3807	if (atomic_sub_and_test(2, &rcu_state.barrier_cpu_count))
3808		complete(&rcu_state.barrier_completion);
3809
3810	/* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3811	wait_for_completion(&rcu_state.barrier_completion);
3812
3813	/* Mark the end of the barrier operation. */
3814	rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
3815	rcu_seq_end(&rcu_state.barrier_sequence);
3816
3817	/* Other rcu_barrier() invocations can now safely proceed. */
3818	mutex_unlock(&rcu_state.barrier_mutex);
 
 
 
 
 
 
 
 
3819}
3820EXPORT_SYMBOL_GPL(rcu_barrier);
 
 
 
 
 
 
 
 
 
3821
3822/*
3823 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3824 * first CPU in a given leaf rcu_node structure coming online.  The caller
3825 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3826 * disabled.
3827 */
3828static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3829{
3830	long mask;
3831	long oldmask;
3832	struct rcu_node *rnp = rnp_leaf;
3833
3834	raw_lockdep_assert_held_rcu_node(rnp_leaf);
3835	WARN_ON_ONCE(rnp->wait_blkd_tasks);
3836	for (;;) {
3837		mask = rnp->grpmask;
3838		rnp = rnp->parent;
3839		if (rnp == NULL)
3840			return;
3841		raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3842		oldmask = rnp->qsmaskinit;
3843		rnp->qsmaskinit |= mask;
3844		raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3845		if (oldmask)
3846			return;
3847	}
3848}
3849
3850/*
3851 * Do boot-time initialization of a CPU's per-CPU RCU data.
3852 */
3853static void __init
3854rcu_boot_init_percpu_data(int cpu)
3855{
3856	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 
 
3857
3858	/* Set up local state, ensuring consistent view of global state. */
3859	rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3860	WARN_ON_ONCE(rdp->dynticks_nesting != 1);
3861	WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp)));
3862	rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
3863	rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
3864	rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
3865	rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
3866	rdp->cpu = cpu;
 
 
3867	rcu_boot_init_nocb_percpu_data(rdp);
 
3868}
3869
3870/*
3871 * Invoked early in the CPU-online process, when pretty much all services
3872 * are available.  The incoming CPU is not present.
3873 *
3874 * Initializes a CPU's per-CPU RCU data.  Note that only one online or
3875 * offline event can be happening at a given time.  Note also that we can
3876 * accept some slop in the rsp->gp_seq access due to the fact that this
3877 * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
3878 * And any offloaded callbacks are being numbered elsewhere.
3879 */
3880int rcutree_prepare_cpu(unsigned int cpu)
 
3881{
3882	unsigned long flags;
3883	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3884	struct rcu_node *rnp = rcu_get_root();
 
3885
3886	/* Set up local state, ensuring consistent view of global state. */
3887	raw_spin_lock_irqsave_rcu_node(rnp, flags);
3888	rdp->qlen_last_fqs_check = 0;
3889	rdp->n_force_qs_snap = rcu_state.n_force_qs;
3890	rdp->blimit = blimit;
3891	if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3892	    !rcu_segcblist_is_offloaded(&rdp->cblist))
3893		rcu_segcblist_init(&rdp->cblist);  /* Re-enable callbacks. */
3894	rdp->dynticks_nesting = 1;	/* CPU not up, no tearing. */
3895	rcu_dynticks_eqs_online();
 
3896	raw_spin_unlock_rcu_node(rnp);		/* irqs remain disabled. */
3897
3898	/*
3899	 * Add CPU to leaf rcu_node pending-online bitmask.  Any needed
3900	 * propagation up the rcu_node tree will happen at the beginning
3901	 * of the next grace period.
3902	 */
3903	rnp = rdp->mynode;
 
3904	raw_spin_lock_rcu_node(rnp);		/* irqs already disabled. */
 
 
 
 
3905	rdp->beenonline = true;	 /* We have now been online. */
3906	rdp->gp_seq = READ_ONCE(rnp->gp_seq);
3907	rdp->gp_seq_needed = rdp->gp_seq;
3908	rdp->cpu_no_qs.b.norm = true;
 
3909	rdp->core_needs_qs = false;
3910	rdp->rcu_iw_pending = false;
3911	rdp->rcu_iw_gp_seq = rdp->gp_seq - 1;
3912	trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
3913	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3914	rcu_prepare_kthreads(cpu);
3915	rcu_spawn_cpu_nocb_kthread(cpu);
3916
3917	return 0;
3918}
3919
3920/*
3921 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3922 */
3923static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3924{
3925	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3926
3927	rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3928}
3929
3930/*
3931 * Near the end of the CPU-online process.  Pretty much all services
3932 * enabled, and the CPU is now very much alive.
3933 */
3934int rcutree_online_cpu(unsigned int cpu)
3935{
3936	unsigned long flags;
3937	struct rcu_data *rdp;
3938	struct rcu_node *rnp;
3939
3940	rdp = per_cpu_ptr(&rcu_data, cpu);
3941	rnp = rdp->mynode;
3942	raw_spin_lock_irqsave_rcu_node(rnp, flags);
3943	rnp->ffmask |= rdp->grpmask;
3944	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3945	if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
3946		return 0; /* Too early in boot for scheduler work. */
3947	sync_sched_exp_online_cleanup(cpu);
3948	rcutree_affinity_setting(cpu, -1);
3949
3950	// Stop-machine done, so allow nohz_full to disable tick.
3951	tick_dep_clear(TICK_DEP_BIT_RCU);
3952	return 0;
3953}
3954
 
3955/*
3956 * Near the beginning of the process.  The CPU is still very much alive
3957 * with pretty much all services enabled.
 
 
 
 
3958 */
3959int rcutree_offline_cpu(unsigned int cpu)
3960{
3961	unsigned long flags;
3962	struct rcu_data *rdp;
3963	struct rcu_node *rnp;
3964
3965	rdp = per_cpu_ptr(&rcu_data, cpu);
3966	rnp = rdp->mynode;
3967	raw_spin_lock_irqsave_rcu_node(rnp, flags);
3968	rnp->ffmask &= ~rdp->grpmask;
3969	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3970
3971	rcutree_affinity_setting(cpu, cpu);
3972
3973	// nohz_full CPUs need the tick for stop-machine to work quickly
3974	tick_dep_set(TICK_DEP_BIT_RCU);
3975	return 0;
3976}
3977
3978static DEFINE_PER_CPU(int, rcu_cpu_started);
3979
3980/*
3981 * Mark the specified CPU as being online so that subsequent grace periods
3982 * (both expedited and normal) will wait on it.  Note that this means that
3983 * incoming CPUs are not allowed to use RCU read-side critical sections
3984 * until this function is called.  Failing to observe this restriction
3985 * will result in lockdep splats.
3986 *
3987 * Note that this function is special in that it is invoked directly
3988 * from the incoming CPU rather than from the cpuhp_step mechanism.
3989 * This is because this function must be invoked at a precise location.
3990 */
3991void rcu_cpu_starting(unsigned int cpu)
3992{
3993	unsigned long flags;
3994	unsigned long mask;
3995	struct rcu_data *rdp;
3996	struct rcu_node *rnp;
3997	bool newcpu;
3998
3999	if (per_cpu(rcu_cpu_started, cpu))
4000		return;
4001
4002	per_cpu(rcu_cpu_started, cpu) = 1;
4003
4004	rdp = per_cpu_ptr(&rcu_data, cpu);
4005	rnp = rdp->mynode;
4006	mask = rdp->grpmask;
4007	raw_spin_lock_irqsave_rcu_node(rnp, flags);
4008	WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask);
4009	newcpu = !(rnp->expmaskinitnext & mask);
4010	rnp->expmaskinitnext |= mask;
4011	/* Allow lockless access for expedited grace periods. */
4012	smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + newcpu); /* ^^^ */
4013	ASSERT_EXCLUSIVE_WRITER(rcu_state.ncpus);
4014	rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
4015	rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4016	rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4017	if (rnp->qsmask & mask) { /* RCU waiting on incoming CPU? */
4018		rcu_disable_urgency_upon_qs(rdp);
4019		/* Report QS -after- changing ->qsmaskinitnext! */
4020		rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4021	} else {
4022		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4023	}
4024	smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
4025}
4026
4027#ifdef CONFIG_HOTPLUG_CPU
4028/*
4029 * The outgoing function has no further need of RCU, so remove it from
4030 * the rcu_node tree's ->qsmaskinitnext bit masks.
4031 *
4032 * Note that this function is special in that it is invoked directly
4033 * from the outgoing CPU rather than from the cpuhp_step mechanism.
4034 * This is because this function must be invoked at a precise location.
4035 */
4036void rcu_report_dead(unsigned int cpu)
4037{
4038	unsigned long flags;
4039	unsigned long mask;
4040	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4041	struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
4042
4043	/* QS for any half-done expedited grace period. */
4044	preempt_disable();
4045	rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
 
4046	preempt_enable();
4047	rcu_preempt_deferred_qs(current);
4048
4049	/* Remove outgoing CPU from mask in the leaf rcu_node structure. */
4050	mask = rdp->grpmask;
4051	raw_spin_lock(&rcu_state.ofl_lock);
4052	raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
4053	rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4054	rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4055	if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
4056		/* Report quiescent state -before- changing ->qsmaskinitnext! */
4057		rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4058		raw_spin_lock_irqsave_rcu_node(rnp, flags);
4059	}
4060	WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask);
4061	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4062	raw_spin_unlock(&rcu_state.ofl_lock);
4063
4064	per_cpu(rcu_cpu_started, cpu) = 0;
4065}
 
4066
4067/*
4068 * The outgoing CPU has just passed through the dying-idle state, and we
4069 * are being invoked from the CPU that was IPIed to continue the offline
4070 * operation.  Migrate the outgoing CPU's callbacks to the current CPU.
4071 */
4072void rcutree_migrate_callbacks(int cpu)
 
4073{
4074	unsigned long flags;
4075	struct rcu_data *my_rdp;
4076	struct rcu_node *my_rnp;
4077	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4078	bool needwake;
4079
4080	if (rcu_segcblist_is_offloaded(&rdp->cblist) ||
4081	    rcu_segcblist_empty(&rdp->cblist))
4082		return;  /* No callbacks to migrate. */
4083
4084	local_irq_save(flags);
4085	my_rdp = this_cpu_ptr(&rcu_data);
4086	my_rnp = my_rdp->mynode;
4087	rcu_nocb_lock(my_rdp); /* irqs already disabled. */
4088	WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies));
4089	raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
4090	/* Leverage recent GPs and set GP for new callbacks. */
4091	needwake = rcu_advance_cbs(my_rnp, rdp) ||
4092		   rcu_advance_cbs(my_rnp, my_rdp);
4093	rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
4094	needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
4095	rcu_segcblist_disable(&rdp->cblist);
4096	WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
4097		     !rcu_segcblist_n_cbs(&my_rdp->cblist));
4098	if (rcu_segcblist_is_offloaded(&my_rdp->cblist)) {
4099		raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
4100		__call_rcu_nocb_wake(my_rdp, true, flags);
4101	} else {
4102		rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
4103		raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags);
 
 
 
 
 
 
 
4104	}
4105	if (needwake)
4106		rcu_gp_kthread_wake();
4107	lockdep_assert_irqs_enabled();
4108	WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
4109		  !rcu_segcblist_empty(&rdp->cblist),
4110		  "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
4111		  cpu, rcu_segcblist_n_cbs(&rdp->cblist),
4112		  rcu_segcblist_first_cb(&rdp->cblist));
4113}
4114#endif
4115
4116/*
4117 * On non-huge systems, use expedited RCU grace periods to make suspend
4118 * and hibernation run faster.
4119 */
4120static int rcu_pm_notify(struct notifier_block *self,
4121			 unsigned long action, void *hcpu)
4122{
4123	switch (action) {
4124	case PM_HIBERNATION_PREPARE:
4125	case PM_SUSPEND_PREPARE:
4126		rcu_expedite_gp();
 
4127		break;
4128	case PM_POST_HIBERNATION:
4129	case PM_POST_SUSPEND:
4130		rcu_unexpedite_gp();
 
4131		break;
4132	default:
4133		break;
4134	}
4135	return NOTIFY_OK;
4136}
4137
4138/*
4139 * Spawn the kthreads that handle RCU's grace periods.
4140 */
4141static int __init rcu_spawn_gp_kthread(void)
4142{
4143	unsigned long flags;
4144	int kthread_prio_in = kthread_prio;
4145	struct rcu_node *rnp;
 
4146	struct sched_param sp;
4147	struct task_struct *t;
4148
4149	/* Force priority into range. */
4150	if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
4151	    && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
4152		kthread_prio = 2;
4153	else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4154		kthread_prio = 1;
4155	else if (kthread_prio < 0)
4156		kthread_prio = 0;
4157	else if (kthread_prio > 99)
4158		kthread_prio = 99;
4159
4160	if (kthread_prio != kthread_prio_in)
4161		pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4162			 kthread_prio, kthread_prio_in);
4163
4164	rcu_scheduler_fully_active = 1;
4165	t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
4166	if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
4167		return 0;
4168	if (kthread_prio) {
4169		sp.sched_priority = kthread_prio;
4170		sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
 
 
 
 
 
 
4171	}
4172	rnp = rcu_get_root();
4173	raw_spin_lock_irqsave_rcu_node(rnp, flags);
4174	WRITE_ONCE(rcu_state.gp_activity, jiffies);
4175	WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
4176	// Reset .gp_activity and .gp_req_activity before setting .gp_kthread.
4177	smp_store_release(&rcu_state.gp_kthread, t);  /* ^^^ */
4178	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4179	wake_up_process(t);
4180	rcu_spawn_nocb_kthreads();
4181	rcu_spawn_boost_kthreads();
4182	return 0;
4183}
4184early_initcall(rcu_spawn_gp_kthread);
4185
4186/*
4187 * This function is invoked towards the end of the scheduler's
4188 * initialization process.  Before this is called, the idle task might
4189 * contain synchronous grace-period primitives (during which time, this idle
4190 * task is booting the system, and such primitives are no-ops).  After this
4191 * function is called, any synchronous grace-period primitives are run as
4192 * expedited, with the requesting task driving the grace period forward.
4193 * A later core_initcall() rcu_set_runtime_mode() will switch to full
4194 * runtime RCU functionality.
4195 */
4196void rcu_scheduler_starting(void)
4197{
4198	WARN_ON(num_online_cpus() != 1);
4199	WARN_ON(nr_context_switches() > 0);
4200	rcu_test_sync_prims();
4201	rcu_scheduler_active = RCU_SCHEDULER_INIT;
4202	rcu_test_sync_prims();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4203}
4204
4205/*
4206 * Helper function for rcu_init() that initializes the rcu_state structure.
4207 */
4208static void __init rcu_init_one(void)
4209{
4210	static const char * const buf[] = RCU_NODE_NAME_INIT;
4211	static const char * const fqs[] = RCU_FQS_NAME_INIT;
 
4212	static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4213	static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
 
 
4214
 
4215	int levelspread[RCU_NUM_LVLS];		/* kids/node in each level. */
4216	int cpustride = 1;
4217	int i;
4218	int j;
4219	struct rcu_node *rnp;
4220
4221	BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf));  /* Fix buf[] init! */
4222
4223	/* Silence gcc 4.8 false positive about array index out of range. */
4224	if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4225		panic("rcu_init_one: rcu_num_lvls out of range");
4226
4227	/* Initialize the level-tracking arrays. */
4228
 
 
4229	for (i = 1; i < rcu_num_lvls; i++)
4230		rcu_state.level[i] =
4231			rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
4232	rcu_init_levelspread(levelspread, num_rcu_lvl);
 
4233
4234	/* Initialize the elements themselves, starting from the leaves. */
4235
4236	for (i = rcu_num_lvls - 1; i >= 0; i--) {
4237		cpustride *= levelspread[i];
4238		rnp = rcu_state.level[i];
4239		for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
4240			raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4241			lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4242						   &rcu_node_class[i], buf[i]);
4243			raw_spin_lock_init(&rnp->fqslock);
4244			lockdep_set_class_and_name(&rnp->fqslock,
4245						   &rcu_fqs_class[i], fqs[i]);
4246			rnp->gp_seq = rcu_state.gp_seq;
4247			rnp->gp_seq_needed = rcu_state.gp_seq;
4248			rnp->completedqs = rcu_state.gp_seq;
4249			rnp->qsmask = 0;
4250			rnp->qsmaskinit = 0;
4251			rnp->grplo = j * cpustride;
4252			rnp->grphi = (j + 1) * cpustride - 1;
4253			if (rnp->grphi >= nr_cpu_ids)
4254				rnp->grphi = nr_cpu_ids - 1;
4255			if (i == 0) {
4256				rnp->grpnum = 0;
4257				rnp->grpmask = 0;
4258				rnp->parent = NULL;
4259			} else {
4260				rnp->grpnum = j % levelspread[i - 1];
4261				rnp->grpmask = BIT(rnp->grpnum);
4262				rnp->parent = rcu_state.level[i - 1] +
4263					      j / levelspread[i - 1];
4264			}
4265			rnp->level = i;
4266			INIT_LIST_HEAD(&rnp->blkd_tasks);
4267			rcu_init_one_nocb(rnp);
4268			init_waitqueue_head(&rnp->exp_wq[0]);
4269			init_waitqueue_head(&rnp->exp_wq[1]);
4270			init_waitqueue_head(&rnp->exp_wq[2]);
4271			init_waitqueue_head(&rnp->exp_wq[3]);
4272			spin_lock_init(&rnp->exp_lock);
4273		}
4274	}
4275
4276	init_swait_queue_head(&rcu_state.gp_wq);
4277	init_swait_queue_head(&rcu_state.expedited_wq);
4278	rnp = rcu_first_leaf_node();
4279	for_each_possible_cpu(i) {
4280		while (i > rnp->grphi)
4281			rnp++;
4282		per_cpu_ptr(&rcu_data, i)->mynode = rnp;
4283		rcu_boot_init_percpu_data(i);
4284	}
 
4285}
4286
4287/*
4288 * Compute the rcu_node tree geometry from kernel parameters.  This cannot
4289 * replace the definitions in tree.h because those are needed to size
4290 * the ->node array in the rcu_state structure.
4291 */
4292static void __init rcu_init_geometry(void)
4293{
4294	ulong d;
4295	int i;
4296	int rcu_capacity[RCU_NUM_LVLS];
4297
4298	/*
4299	 * Initialize any unspecified boot parameters.
4300	 * The default values of jiffies_till_first_fqs and
4301	 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4302	 * value, which is a function of HZ, then adding one for each
4303	 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4304	 */
4305	d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4306	if (jiffies_till_first_fqs == ULONG_MAX)
4307		jiffies_till_first_fqs = d;
4308	if (jiffies_till_next_fqs == ULONG_MAX)
4309		jiffies_till_next_fqs = d;
4310	adjust_jiffies_till_sched_qs();
4311
4312	/* If the compile-time values are accurate, just leave. */
4313	if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4314	    nr_cpu_ids == NR_CPUS)
4315		return;
4316	pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4317		rcu_fanout_leaf, nr_cpu_ids);
4318
4319	/*
4320	 * The boot-time rcu_fanout_leaf parameter must be at least two
4321	 * and cannot exceed the number of bits in the rcu_node masks.
4322	 * Complain and fall back to the compile-time values if this
4323	 * limit is exceeded.
4324	 */
4325	if (rcu_fanout_leaf < 2 ||
4326	    rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4327		rcu_fanout_leaf = RCU_FANOUT_LEAF;
4328		WARN_ON(1);
4329		return;
4330	}
4331
4332	/*
4333	 * Compute number of nodes that can be handled an rcu_node tree
4334	 * with the given number of levels.
4335	 */
4336	rcu_capacity[0] = rcu_fanout_leaf;
4337	for (i = 1; i < RCU_NUM_LVLS; i++)
4338		rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4339
4340	/*
4341	 * The tree must be able to accommodate the configured number of CPUs.
4342	 * If this limit is exceeded, fall back to the compile-time values.
4343	 */
4344	if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4345		rcu_fanout_leaf = RCU_FANOUT_LEAF;
4346		WARN_ON(1);
4347		return;
4348	}
4349
4350	/* Calculate the number of levels in the tree. */
4351	for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4352	}
4353	rcu_num_lvls = i + 1;
4354
4355	/* Calculate the number of rcu_nodes at each level of the tree. */
4356	for (i = 0; i < rcu_num_lvls; i++) {
4357		int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4358		num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4359	}
4360
4361	/* Calculate the total number of rcu_node structures. */
4362	rcu_num_nodes = 0;
4363	for (i = 0; i < rcu_num_lvls; i++)
4364		rcu_num_nodes += num_rcu_lvl[i];
4365}
4366
4367/*
4368 * Dump out the structure of the rcu_node combining tree associated
4369 * with the rcu_state structure.
4370 */
4371static void __init rcu_dump_rcu_node_tree(void)
4372{
4373	int level = 0;
4374	struct rcu_node *rnp;
4375
4376	pr_info("rcu_node tree layout dump\n");
4377	pr_info(" ");
4378	rcu_for_each_node_breadth_first(rnp) {
4379		if (rnp->level != level) {
4380			pr_cont("\n");
4381			pr_info(" ");
4382			level = rnp->level;
4383		}
4384		pr_cont("%d:%d ^%d  ", rnp->grplo, rnp->grphi, rnp->grpnum);
4385	}
4386	pr_cont("\n");
4387}
4388
4389struct workqueue_struct *rcu_gp_wq;
4390struct workqueue_struct *rcu_par_gp_wq;
4391
4392static void __init kfree_rcu_batch_init(void)
4393{
4394	int cpu;
4395	int i;
4396
4397	for_each_possible_cpu(cpu) {
4398		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
4399		struct kvfree_rcu_bulk_data *bnode;
4400
4401		for (i = 0; i < KFREE_N_BATCHES; i++) {
4402			INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
4403			krcp->krw_arr[i].krcp = krcp;
4404		}
4405
4406		for (i = 0; i < rcu_min_cached_objs; i++) {
4407			bnode = (struct kvfree_rcu_bulk_data *)
4408				__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
4409
4410			if (bnode)
4411				put_cached_bnode(krcp, bnode);
4412			else
4413				pr_err("Failed to preallocate for %d CPU!\n", cpu);
4414		}
4415
4416		INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
4417		krcp->initialized = true;
4418	}
4419	if (register_shrinker(&kfree_rcu_shrinker))
4420		pr_err("Failed to register kfree_rcu() shrinker!\n");
4421}
4422
4423void __init rcu_init(void)
4424{
4425	int cpu;
4426
4427	rcu_early_boot_tests();
4428
4429	kfree_rcu_batch_init();
4430	rcu_bootup_announce();
4431	rcu_init_geometry();
4432	rcu_init_one();
 
4433	if (dump_tree)
4434		rcu_dump_rcu_node_tree();
4435	if (use_softirq)
4436		open_softirq(RCU_SOFTIRQ, rcu_core_si);
4437
4438	/*
4439	 * We don't need protection against CPU-hotplug here because
4440	 * this is called early in boot, before either interrupts
4441	 * or the scheduler are operational.
4442	 */
 
4443	pm_notifier(rcu_pm_notify, 0);
4444	for_each_online_cpu(cpu) {
4445		rcutree_prepare_cpu(cpu);
4446		rcu_cpu_starting(cpu);
4447		rcutree_online_cpu(cpu);
4448	}
4449
4450	/* Create workqueue for expedited GPs and for Tree SRCU. */
4451	rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
4452	WARN_ON(!rcu_gp_wq);
4453	rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
4454	WARN_ON(!rcu_par_gp_wq);
4455	srcu_init();
4456
4457	/* Fill in default value for rcutree.qovld boot parameter. */
4458	/* -After- the rcu_node ->lock fields are initialized! */
4459	if (qovld < 0)
4460		qovld_calc = DEFAULT_RCU_QOVLD_MULT * qhimark;
4461	else
4462		qovld_calc = qovld;
4463}
4464
4465#include "tree_stall.h"
4466#include "tree_exp.h"
4467#include "tree_plugin.h"