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