linux/kernel/rcutree.c
<<
>>
Prefs
   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, write to the Free Software
  16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  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/percpu.h>
  45#include <linux/notifier.h>
  46#include <linux/cpu.h>
  47#include <linux/mutex.h>
  48#include <linux/time.h>
  49#include <linux/kernel_stat.h>
  50#include <linux/wait.h>
  51#include <linux/kthread.h>
  52#include <linux/prefetch.h>
  53#include <linux/delay.h>
  54#include <linux/stop_machine.h>
  55#include <linux/random.h>
  56
  57#include "rcutree.h"
  58#include <trace/events/rcu.h>
  59
  60#include "rcu.h"
  61
  62/* Data structures. */
  63
  64static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
  65static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
  66
  67#define RCU_STATE_INITIALIZER(sname, cr) { \
  68        .level = { &sname##_state.node[0] }, \
  69        .call = cr, \
  70        .fqs_state = RCU_GP_IDLE, \
  71        .gpnum = -300, \
  72        .completed = -300, \
  73        .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.onofflock), \
  74        .orphan_nxttail = &sname##_state.orphan_nxtlist, \
  75        .orphan_donetail = &sname##_state.orphan_donelist, \
  76        .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
  77        .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
  78        .name = #sname, \
  79}
  80
  81struct rcu_state rcu_sched_state =
  82        RCU_STATE_INITIALIZER(rcu_sched, call_rcu_sched);
  83DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
  84
  85struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh, call_rcu_bh);
  86DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
  87
  88static struct rcu_state *rcu_state;
  89LIST_HEAD(rcu_struct_flavors);
  90
  91/* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
  92static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
  93module_param(rcu_fanout_leaf, int, 0444);
  94int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
  95static int num_rcu_lvl[] = {  /* Number of rcu_nodes at specified level. */
  96        NUM_RCU_LVL_0,
  97        NUM_RCU_LVL_1,
  98        NUM_RCU_LVL_2,
  99        NUM_RCU_LVL_3,
 100        NUM_RCU_LVL_4,
 101};
 102int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
 103
 104/*
 105 * The rcu_scheduler_active variable transitions from zero to one just
 106 * before the first task is spawned.  So when this variable is zero, RCU
 107 * can assume that there is but one task, allowing RCU to (for example)
 108 * optimized synchronize_sched() to a simple barrier().  When this variable
 109 * is one, RCU must actually do all the hard work required to detect real
 110 * grace periods.  This variable is also used to suppress boot-time false
 111 * positives from lockdep-RCU error checking.
 112 */
 113int rcu_scheduler_active __read_mostly;
 114EXPORT_SYMBOL_GPL(rcu_scheduler_active);
 115
 116/*
 117 * The rcu_scheduler_fully_active variable transitions from zero to one
 118 * during the early_initcall() processing, which is after the scheduler
 119 * is capable of creating new tasks.  So RCU processing (for example,
 120 * creating tasks for RCU priority boosting) must be delayed until after
 121 * rcu_scheduler_fully_active transitions from zero to one.  We also
 122 * currently delay invocation of any RCU callbacks until after this point.
 123 *
 124 * It might later prove better for people registering RCU callbacks during
 125 * early boot to take responsibility for these callbacks, but one step at
 126 * a time.
 127 */
 128static int rcu_scheduler_fully_active __read_mostly;
 129
 130#ifdef CONFIG_RCU_BOOST
 131
 132/*
 133 * Control variables for per-CPU and per-rcu_node kthreads.  These
 134 * handle all flavors of RCU.
 135 */
 136static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
 137DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
 138DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
 139DEFINE_PER_CPU(char, rcu_cpu_has_work);
 140
 141#endif /* #ifdef CONFIG_RCU_BOOST */
 142
 143static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
 144static void invoke_rcu_core(void);
 145static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
 146
 147/*
 148 * Track the rcutorture test sequence number and the update version
 149 * number within a given test.  The rcutorture_testseq is incremented
 150 * on every rcutorture module load and unload, so has an odd value
 151 * when a test is running.  The rcutorture_vernum is set to zero
 152 * when rcutorture starts and is incremented on each rcutorture update.
 153 * These variables enable correlating rcutorture output with the
 154 * RCU tracing information.
 155 */
 156unsigned long rcutorture_testseq;
 157unsigned long rcutorture_vernum;
 158
 159/*
 160 * Return true if an RCU grace period is in progress.  The ACCESS_ONCE()s
 161 * permit this function to be invoked without holding the root rcu_node
 162 * structure's ->lock, but of course results can be subject to change.
 163 */
 164static int rcu_gp_in_progress(struct rcu_state *rsp)
 165{
 166        return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
 167}
 168
 169/*
 170 * Note a quiescent state.  Because we do not need to know
 171 * how many quiescent states passed, just if there was at least
 172 * one since the start of the grace period, this just sets a flag.
 173 * The caller must have disabled preemption.
 174 */
 175void rcu_sched_qs(int cpu)
 176{
 177        struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
 178
 179        if (rdp->passed_quiesce == 0)
 180                trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
 181        rdp->passed_quiesce = 1;
 182}
 183
 184void rcu_bh_qs(int cpu)
 185{
 186        struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
 187
 188        if (rdp->passed_quiesce == 0)
 189                trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
 190        rdp->passed_quiesce = 1;
 191}
 192
 193/*
 194 * Note a context switch.  This is a quiescent state for RCU-sched,
 195 * and requires special handling for preemptible RCU.
 196 * The caller must have disabled preemption.
 197 */
 198void rcu_note_context_switch(int cpu)
 199{
 200        trace_rcu_utilization("Start context switch");
 201        rcu_sched_qs(cpu);
 202        rcu_preempt_note_context_switch(cpu);
 203        trace_rcu_utilization("End context switch");
 204}
 205EXPORT_SYMBOL_GPL(rcu_note_context_switch);
 206
 207DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
 208        .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
 209        .dynticks = ATOMIC_INIT(1),
 210#if defined(CONFIG_RCU_USER_QS) && !defined(CONFIG_RCU_USER_QS_FORCE)
 211        .ignore_user_qs = true,
 212#endif
 213};
 214
 215static long blimit = 10;        /* Maximum callbacks per rcu_do_batch. */
 216static long qhimark = 10000;    /* If this many pending, ignore blimit. */
 217static long qlowmark = 100;     /* Once only this many pending, use blimit. */
 218
 219module_param(blimit, long, 0444);
 220module_param(qhimark, long, 0444);
 221module_param(qlowmark, long, 0444);
 222
 223int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
 224int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
 225
 226module_param(rcu_cpu_stall_suppress, int, 0644);
 227module_param(rcu_cpu_stall_timeout, int, 0644);
 228
 229static ulong jiffies_till_first_fqs = RCU_JIFFIES_TILL_FORCE_QS;
 230static ulong jiffies_till_next_fqs = RCU_JIFFIES_TILL_FORCE_QS;
 231
 232module_param(jiffies_till_first_fqs, ulong, 0644);
 233module_param(jiffies_till_next_fqs, ulong, 0644);
 234
 235static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *));
 236static void force_quiescent_state(struct rcu_state *rsp);
 237static int rcu_pending(int cpu);
 238
 239/*
 240 * Return the number of RCU-sched batches processed thus far for debug & stats.
 241 */
 242long rcu_batches_completed_sched(void)
 243{
 244        return rcu_sched_state.completed;
 245}
 246EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
 247
 248/*
 249 * Return the number of RCU BH batches processed thus far for debug & stats.
 250 */
 251long rcu_batches_completed_bh(void)
 252{
 253        return rcu_bh_state.completed;
 254}
 255EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
 256
 257/*
 258 * Force a quiescent state for RCU BH.
 259 */
 260void rcu_bh_force_quiescent_state(void)
 261{
 262        force_quiescent_state(&rcu_bh_state);
 263}
 264EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
 265
 266/*
 267 * Record the number of times rcutorture tests have been initiated and
 268 * terminated.  This information allows the debugfs tracing stats to be
 269 * correlated to the rcutorture messages, even when the rcutorture module
 270 * is being repeatedly loaded and unloaded.  In other words, we cannot
 271 * store this state in rcutorture itself.
 272 */
 273void rcutorture_record_test_transition(void)
 274{
 275        rcutorture_testseq++;
 276        rcutorture_vernum = 0;
 277}
 278EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
 279
 280/*
 281 * Record the number of writer passes through the current rcutorture test.
 282 * This is also used to correlate debugfs tracing stats with the rcutorture
 283 * messages.
 284 */
 285void rcutorture_record_progress(unsigned long vernum)
 286{
 287        rcutorture_vernum++;
 288}
 289EXPORT_SYMBOL_GPL(rcutorture_record_progress);
 290
 291/*
 292 * Force a quiescent state for RCU-sched.
 293 */
 294void rcu_sched_force_quiescent_state(void)
 295{
 296        force_quiescent_state(&rcu_sched_state);
 297}
 298EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
 299
 300/*
 301 * Does the CPU have callbacks ready to be invoked?
 302 */
 303static int
 304cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
 305{
 306        return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
 307}
 308
 309/*
 310 * Does the current CPU require a yet-as-unscheduled grace period?
 311 */
 312static int
 313cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
 314{
 315        return *rdp->nxttail[RCU_DONE_TAIL +
 316                             ACCESS_ONCE(rsp->completed) != rdp->completed] &&
 317               !rcu_gp_in_progress(rsp);
 318}
 319
 320/*
 321 * Return the root node of the specified rcu_state structure.
 322 */
 323static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
 324{
 325        return &rsp->node[0];
 326}
 327
 328/*
 329 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
 330 *
 331 * If the new value of the ->dynticks_nesting counter now is zero,
 332 * we really have entered idle, and must do the appropriate accounting.
 333 * The caller must have disabled interrupts.
 334 */
 335static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
 336                                bool user)
 337{
 338        trace_rcu_dyntick("Start", oldval, 0);
 339        if (!user && !is_idle_task(current)) {
 340                struct task_struct *idle = idle_task(smp_processor_id());
 341
 342                trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
 343                ftrace_dump(DUMP_ORIG);
 344                WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
 345                          current->pid, current->comm,
 346                          idle->pid, idle->comm); /* must be idle task! */
 347        }
 348        rcu_prepare_for_idle(smp_processor_id());
 349        /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
 350        smp_mb__before_atomic_inc();  /* See above. */
 351        atomic_inc(&rdtp->dynticks);
 352        smp_mb__after_atomic_inc();  /* Force ordering with next sojourn. */
 353        WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
 354
 355        /*
 356         * It is illegal to enter an extended quiescent state while
 357         * in an RCU read-side critical section.
 358         */
 359        rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
 360                           "Illegal idle entry in RCU read-side critical section.");
 361        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
 362                           "Illegal idle entry in RCU-bh read-side critical section.");
 363        rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
 364                           "Illegal idle entry in RCU-sched read-side critical section.");
 365}
 366
 367/*
 368 * Enter an RCU extended quiescent state, which can be either the
 369 * idle loop or adaptive-tickless usermode execution.
 370 */
 371static void rcu_eqs_enter(bool user)
 372{
 373        long long oldval;
 374        struct rcu_dynticks *rdtp;
 375
 376        rdtp = &__get_cpu_var(rcu_dynticks);
 377        oldval = rdtp->dynticks_nesting;
 378        WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
 379        if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
 380                rdtp->dynticks_nesting = 0;
 381        else
 382                rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
 383        rcu_eqs_enter_common(rdtp, oldval, user);
 384}
 385
 386/**
 387 * rcu_idle_enter - inform RCU that current CPU is entering idle
 388 *
 389 * Enter idle mode, in other words, -leave- the mode in which RCU
 390 * read-side critical sections can occur.  (Though RCU read-side
 391 * critical sections can occur in irq handlers in idle, a possibility
 392 * handled by irq_enter() and irq_exit().)
 393 *
 394 * We crowbar the ->dynticks_nesting field to zero to allow for
 395 * the possibility of usermode upcalls having messed up our count
 396 * of interrupt nesting level during the prior busy period.
 397 */
 398void rcu_idle_enter(void)
 399{
 400        unsigned long flags;
 401
 402        local_irq_save(flags);
 403        rcu_eqs_enter(false);
 404        local_irq_restore(flags);
 405}
 406EXPORT_SYMBOL_GPL(rcu_idle_enter);
 407
 408#ifdef CONFIG_RCU_USER_QS
 409/**
 410 * rcu_user_enter - inform RCU that we are resuming userspace.
 411 *
 412 * Enter RCU idle mode right before resuming userspace.  No use of RCU
 413 * is permitted between this call and rcu_user_exit(). This way the
 414 * CPU doesn't need to maintain the tick for RCU maintenance purposes
 415 * when the CPU runs in userspace.
 416 */
 417void rcu_user_enter(void)
 418{
 419        unsigned long flags;
 420        struct rcu_dynticks *rdtp;
 421
 422        /*
 423         * Some contexts may involve an exception occuring in an irq,
 424         * leading to that nesting:
 425         * rcu_irq_enter() rcu_user_exit() rcu_user_exit() rcu_irq_exit()
 426         * This would mess up the dyntick_nesting count though. And rcu_irq_*()
 427         * helpers are enough to protect RCU uses inside the exception. So
 428         * just return immediately if we detect we are in an IRQ.
 429         */
 430        if (in_interrupt())
 431                return;
 432
 433        WARN_ON_ONCE(!current->mm);
 434
 435        local_irq_save(flags);
 436        rdtp = &__get_cpu_var(rcu_dynticks);
 437        if (!rdtp->ignore_user_qs && !rdtp->in_user) {
 438                rdtp->in_user = true;
 439                rcu_eqs_enter(true);
 440        }
 441        local_irq_restore(flags);
 442}
 443
 444/**
 445 * rcu_user_enter_after_irq - inform RCU that we are going to resume userspace
 446 * after the current irq returns.
 447 *
 448 * This is similar to rcu_user_enter() but in the context of a non-nesting
 449 * irq. After this call, RCU enters into idle mode when the interrupt
 450 * returns.
 451 */
 452void rcu_user_enter_after_irq(void)
 453{
 454        unsigned long flags;
 455        struct rcu_dynticks *rdtp;
 456
 457        local_irq_save(flags);
 458        rdtp = &__get_cpu_var(rcu_dynticks);
 459        /* Ensure this irq is interrupting a non-idle RCU state.  */
 460        WARN_ON_ONCE(!(rdtp->dynticks_nesting & DYNTICK_TASK_MASK));
 461        rdtp->dynticks_nesting = 1;
 462        local_irq_restore(flags);
 463}
 464#endif /* CONFIG_RCU_USER_QS */
 465
 466/**
 467 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
 468 *
 469 * Exit from an interrupt handler, which might possibly result in entering
 470 * idle mode, in other words, leaving the mode in which read-side critical
 471 * sections can occur.
 472 *
 473 * This code assumes that the idle loop never does anything that might
 474 * result in unbalanced calls to irq_enter() and irq_exit().  If your
 475 * architecture violates this assumption, RCU will give you what you
 476 * deserve, good and hard.  But very infrequently and irreproducibly.
 477 *
 478 * Use things like work queues to work around this limitation.
 479 *
 480 * You have been warned.
 481 */
 482void rcu_irq_exit(void)
 483{
 484        unsigned long flags;
 485        long long oldval;
 486        struct rcu_dynticks *rdtp;
 487
 488        local_irq_save(flags);
 489        rdtp = &__get_cpu_var(rcu_dynticks);
 490        oldval = rdtp->dynticks_nesting;
 491        rdtp->dynticks_nesting--;
 492        WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
 493        if (rdtp->dynticks_nesting)
 494                trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
 495        else
 496                rcu_eqs_enter_common(rdtp, oldval, true);
 497        local_irq_restore(flags);
 498}
 499
 500/*
 501 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
 502 *
 503 * If the new value of the ->dynticks_nesting counter was previously zero,
 504 * we really have exited idle, and must do the appropriate accounting.
 505 * The caller must have disabled interrupts.
 506 */
 507static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
 508                               int user)
 509{
 510        smp_mb__before_atomic_inc();  /* Force ordering w/previous sojourn. */
 511        atomic_inc(&rdtp->dynticks);
 512        /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
 513        smp_mb__after_atomic_inc();  /* See above. */
 514        WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
 515        rcu_cleanup_after_idle(smp_processor_id());
 516        trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
 517        if (!user && !is_idle_task(current)) {
 518                struct task_struct *idle = idle_task(smp_processor_id());
 519
 520                trace_rcu_dyntick("Error on exit: not idle task",
 521                                  oldval, rdtp->dynticks_nesting);
 522                ftrace_dump(DUMP_ORIG);
 523                WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
 524                          current->pid, current->comm,
 525                          idle->pid, idle->comm); /* must be idle task! */
 526        }
 527}
 528
 529/*
 530 * Exit an RCU extended quiescent state, which can be either the
 531 * idle loop or adaptive-tickless usermode execution.
 532 */
 533static void rcu_eqs_exit(bool user)
 534{
 535        struct rcu_dynticks *rdtp;
 536        long long oldval;
 537
 538        rdtp = &__get_cpu_var(rcu_dynticks);
 539        oldval = rdtp->dynticks_nesting;
 540        WARN_ON_ONCE(oldval < 0);
 541        if (oldval & DYNTICK_TASK_NEST_MASK)
 542                rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
 543        else
 544                rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
 545        rcu_eqs_exit_common(rdtp, oldval, user);
 546}
 547
 548/**
 549 * rcu_idle_exit - inform RCU that current CPU is leaving idle
 550 *
 551 * Exit idle mode, in other words, -enter- the mode in which RCU
 552 * read-side critical sections can occur.
 553 *
 554 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
 555 * allow for the possibility of usermode upcalls messing up our count
 556 * of interrupt nesting level during the busy period that is just
 557 * now starting.
 558 */
 559void rcu_idle_exit(void)
 560{
 561        unsigned long flags;
 562
 563        local_irq_save(flags);
 564        rcu_eqs_exit(false);
 565        local_irq_restore(flags);
 566}
 567EXPORT_SYMBOL_GPL(rcu_idle_exit);
 568
 569#ifdef CONFIG_RCU_USER_QS
 570/**
 571 * rcu_user_exit - inform RCU that we are exiting userspace.
 572 *
 573 * Exit RCU idle mode while entering the kernel because it can
 574 * run a RCU read side critical section anytime.
 575 */
 576void rcu_user_exit(void)
 577{
 578        unsigned long flags;
 579        struct rcu_dynticks *rdtp;
 580
 581        /*
 582         * Some contexts may involve an exception occuring in an irq,
 583         * leading to that nesting:
 584         * rcu_irq_enter() rcu_user_exit() rcu_user_exit() rcu_irq_exit()
 585         * This would mess up the dyntick_nesting count though. And rcu_irq_*()
 586         * helpers are enough to protect RCU uses inside the exception. So
 587         * just return immediately if we detect we are in an IRQ.
 588         */
 589        if (in_interrupt())
 590                return;
 591
 592        local_irq_save(flags);
 593        rdtp = &__get_cpu_var(rcu_dynticks);
 594        if (rdtp->in_user) {
 595                rdtp->in_user = false;
 596                rcu_eqs_exit(true);
 597        }
 598        local_irq_restore(flags);
 599}
 600
 601/**
 602 * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace
 603 * idle mode after the current non-nesting irq returns.
 604 *
 605 * This is similar to rcu_user_exit() but in the context of an irq.
 606 * This is called when the irq has interrupted a userspace RCU idle mode
 607 * context. When the current non-nesting interrupt returns after this call,
 608 * the CPU won't restore the RCU idle mode.
 609 */
 610void rcu_user_exit_after_irq(void)
 611{
 612        unsigned long flags;
 613        struct rcu_dynticks *rdtp;
 614
 615        local_irq_save(flags);
 616        rdtp = &__get_cpu_var(rcu_dynticks);
 617        /* Ensure we are interrupting an RCU idle mode. */
 618        WARN_ON_ONCE(rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK);
 619        rdtp->dynticks_nesting += DYNTICK_TASK_EXIT_IDLE;
 620        local_irq_restore(flags);
 621}
 622#endif /* CONFIG_RCU_USER_QS */
 623
 624/**
 625 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
 626 *
 627 * Enter an interrupt handler, which might possibly result in exiting
 628 * idle mode, in other words, entering the mode in which read-side critical
 629 * sections can occur.
 630 *
 631 * Note that the Linux kernel is fully capable of entering an interrupt
 632 * handler that it never exits, for example when doing upcalls to
 633 * user mode!  This code assumes that the idle loop never does upcalls to
 634 * user mode.  If your architecture does do upcalls from the idle loop (or
 635 * does anything else that results in unbalanced calls to the irq_enter()
 636 * and irq_exit() functions), RCU will give you what you deserve, good
 637 * and hard.  But very infrequently and irreproducibly.
 638 *
 639 * Use things like work queues to work around this limitation.
 640 *
 641 * You have been warned.
 642 */
 643void rcu_irq_enter(void)
 644{
 645        unsigned long flags;
 646        struct rcu_dynticks *rdtp;
 647        long long oldval;
 648
 649        local_irq_save(flags);
 650        rdtp = &__get_cpu_var(rcu_dynticks);
 651        oldval = rdtp->dynticks_nesting;
 652        rdtp->dynticks_nesting++;
 653        WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
 654        if (oldval)
 655                trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
 656        else
 657                rcu_eqs_exit_common(rdtp, oldval, true);
 658        local_irq_restore(flags);
 659}
 660
 661/**
 662 * rcu_nmi_enter - inform RCU of entry to NMI context
 663 *
 664 * If the CPU was idle with dynamic ticks active, and there is no
 665 * irq handler running, this updates rdtp->dynticks_nmi to let the
 666 * RCU grace-period handling know that the CPU is active.
 667 */
 668void rcu_nmi_enter(void)
 669{
 670        struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
 671
 672        if (rdtp->dynticks_nmi_nesting == 0 &&
 673            (atomic_read(&rdtp->dynticks) & 0x1))
 674                return;
 675        rdtp->dynticks_nmi_nesting++;
 676        smp_mb__before_atomic_inc();  /* Force delay from prior write. */
 677        atomic_inc(&rdtp->dynticks);
 678        /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
 679        smp_mb__after_atomic_inc();  /* See above. */
 680        WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
 681}
 682
 683/**
 684 * rcu_nmi_exit - inform RCU of exit from NMI context
 685 *
 686 * If the CPU was idle with dynamic ticks active, and there is no
 687 * irq handler running, this updates rdtp->dynticks_nmi to let the
 688 * RCU grace-period handling know that the CPU is no longer active.
 689 */
 690void rcu_nmi_exit(void)
 691{
 692        struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
 693
 694        if (rdtp->dynticks_nmi_nesting == 0 ||
 695            --rdtp->dynticks_nmi_nesting != 0)
 696                return;
 697        /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
 698        smp_mb__before_atomic_inc();  /* See above. */
 699        atomic_inc(&rdtp->dynticks);
 700        smp_mb__after_atomic_inc();  /* Force delay to next write. */
 701        WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
 702}
 703
 704/**
 705 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
 706 *
 707 * If the current CPU is in its idle loop and is neither in an interrupt
 708 * or NMI handler, return true.
 709 */
 710int rcu_is_cpu_idle(void)
 711{
 712        int ret;
 713
 714        preempt_disable();
 715        ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
 716        preempt_enable();
 717        return ret;
 718}
 719EXPORT_SYMBOL(rcu_is_cpu_idle);
 720
 721#ifdef CONFIG_RCU_USER_QS
 722void rcu_user_hooks_switch(struct task_struct *prev,
 723                           struct task_struct *next)
 724{
 725        struct rcu_dynticks *rdtp;
 726
 727        /* Interrupts are disabled in context switch */
 728        rdtp = &__get_cpu_var(rcu_dynticks);
 729        if (!rdtp->ignore_user_qs) {
 730                clear_tsk_thread_flag(prev, TIF_NOHZ);
 731                set_tsk_thread_flag(next, TIF_NOHZ);
 732        }
 733}
 734#endif /* #ifdef CONFIG_RCU_USER_QS */
 735
 736#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
 737
 738/*
 739 * Is the current CPU online?  Disable preemption to avoid false positives
 740 * that could otherwise happen due to the current CPU number being sampled,
 741 * this task being preempted, its old CPU being taken offline, resuming
 742 * on some other CPU, then determining that its old CPU is now offline.
 743 * It is OK to use RCU on an offline processor during initial boot, hence
 744 * the check for rcu_scheduler_fully_active.  Note also that it is OK
 745 * for a CPU coming online to use RCU for one jiffy prior to marking itself
 746 * online in the cpu_online_mask.  Similarly, it is OK for a CPU going
 747 * offline to continue to use RCU for one jiffy after marking itself
 748 * offline in the cpu_online_mask.  This leniency is necessary given the
 749 * non-atomic nature of the online and offline processing, for example,
 750 * the fact that a CPU enters the scheduler after completing the CPU_DYING
 751 * notifiers.
 752 *
 753 * This is also why RCU internally marks CPUs online during the
 754 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
 755 *
 756 * Disable checking if in an NMI handler because we cannot safely report
 757 * errors from NMI handlers anyway.
 758 */
 759bool rcu_lockdep_current_cpu_online(void)
 760{
 761        struct rcu_data *rdp;
 762        struct rcu_node *rnp;
 763        bool ret;
 764
 765        if (in_nmi())
 766                return 1;
 767        preempt_disable();
 768        rdp = &__get_cpu_var(rcu_sched_data);
 769        rnp = rdp->mynode;
 770        ret = (rdp->grpmask & rnp->qsmaskinit) ||
 771              !rcu_scheduler_fully_active;
 772        preempt_enable();
 773        return ret;
 774}
 775EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
 776
 777#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
 778
 779/**
 780 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
 781 *
 782 * If the current CPU is idle or running at a first-level (not nested)
 783 * interrupt from idle, return true.  The caller must have at least
 784 * disabled preemption.
 785 */
 786int rcu_is_cpu_rrupt_from_idle(void)
 787{
 788        return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
 789}
 790
 791/*
 792 * Snapshot the specified CPU's dynticks counter so that we can later
 793 * credit them with an implicit quiescent state.  Return 1 if this CPU
 794 * is in dynticks idle mode, which is an extended quiescent state.
 795 */
 796static int dyntick_save_progress_counter(struct rcu_data *rdp)
 797{
 798        rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
 799        return (rdp->dynticks_snap & 0x1) == 0;
 800}
 801
 802/*
 803 * Return true if the specified CPU has passed through a quiescent
 804 * state by virtue of being in or having passed through an dynticks
 805 * idle state since the last call to dyntick_save_progress_counter()
 806 * for this same CPU, or by virtue of having been offline.
 807 */
 808static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
 809{
 810        unsigned int curr;
 811        unsigned int snap;
 812
 813        curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
 814        snap = (unsigned int)rdp->dynticks_snap;
 815
 816        /*
 817         * If the CPU passed through or entered a dynticks idle phase with
 818         * no active irq/NMI handlers, then we can safely pretend that the CPU
 819         * already acknowledged the request to pass through a quiescent
 820         * state.  Either way, that CPU cannot possibly be in an RCU
 821         * read-side critical section that started before the beginning
 822         * of the current RCU grace period.
 823         */
 824        if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
 825                trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
 826                rdp->dynticks_fqs++;
 827                return 1;
 828        }
 829
 830        /*
 831         * Check for the CPU being offline, but only if the grace period
 832         * is old enough.  We don't need to worry about the CPU changing
 833         * state: If we see it offline even once, it has been through a
 834         * quiescent state.
 835         *
 836         * The reason for insisting that the grace period be at least
 837         * one jiffy old is that CPUs that are not quite online and that
 838         * have just gone offline can still execute RCU read-side critical
 839         * sections.
 840         */
 841        if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
 842                return 0;  /* Grace period is not old enough. */
 843        barrier();
 844        if (cpu_is_offline(rdp->cpu)) {
 845                trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
 846                rdp->offline_fqs++;
 847                return 1;
 848        }
 849        return 0;
 850}
 851
 852static int jiffies_till_stall_check(void)
 853{
 854        int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
 855
 856        /*
 857         * Limit check must be consistent with the Kconfig limits
 858         * for CONFIG_RCU_CPU_STALL_TIMEOUT.
 859         */
 860        if (till_stall_check < 3) {
 861                ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
 862                till_stall_check = 3;
 863        } else if (till_stall_check > 300) {
 864                ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
 865                till_stall_check = 300;
 866        }
 867        return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
 868}
 869
 870static void record_gp_stall_check_time(struct rcu_state *rsp)
 871{
 872        rsp->gp_start = jiffies;
 873        rsp->jiffies_stall = jiffies + jiffies_till_stall_check();
 874}
 875
 876static void print_other_cpu_stall(struct rcu_state *rsp)
 877{
 878        int cpu;
 879        long delta;
 880        unsigned long flags;
 881        int ndetected = 0;
 882        struct rcu_node *rnp = rcu_get_root(rsp);
 883
 884        /* Only let one CPU complain about others per time interval. */
 885
 886        raw_spin_lock_irqsave(&rnp->lock, flags);
 887        delta = jiffies - rsp->jiffies_stall;
 888        if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
 889                raw_spin_unlock_irqrestore(&rnp->lock, flags);
 890                return;
 891        }
 892        rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3;
 893        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 894
 895        /*
 896         * OK, time to rat on our buddy...
 897         * See Documentation/RCU/stallwarn.txt for info on how to debug
 898         * RCU CPU stall warnings.
 899         */
 900        printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
 901               rsp->name);
 902        print_cpu_stall_info_begin();
 903        rcu_for_each_leaf_node(rsp, rnp) {
 904                raw_spin_lock_irqsave(&rnp->lock, flags);
 905                ndetected += rcu_print_task_stall(rnp);
 906                if (rnp->qsmask != 0) {
 907                        for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
 908                                if (rnp->qsmask & (1UL << cpu)) {
 909                                        print_cpu_stall_info(rsp,
 910                                                             rnp->grplo + cpu);
 911                                        ndetected++;
 912                                }
 913                }
 914                raw_spin_unlock_irqrestore(&rnp->lock, flags);
 915        }
 916
 917        /*
 918         * Now rat on any tasks that got kicked up to the root rcu_node
 919         * due to CPU offlining.
 920         */
 921        rnp = rcu_get_root(rsp);
 922        raw_spin_lock_irqsave(&rnp->lock, flags);
 923        ndetected += rcu_print_task_stall(rnp);
 924        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 925
 926        print_cpu_stall_info_end();
 927        printk(KERN_CONT "(detected by %d, t=%ld jiffies)\n",
 928               smp_processor_id(), (long)(jiffies - rsp->gp_start));
 929        if (ndetected == 0)
 930                printk(KERN_ERR "INFO: Stall ended before state dump start\n");
 931        else if (!trigger_all_cpu_backtrace())
 932                dump_stack();
 933
 934        /* Complain about tasks blocking the grace period. */
 935
 936        rcu_print_detail_task_stall(rsp);
 937
 938        force_quiescent_state(rsp);  /* Kick them all. */
 939}
 940
 941static void print_cpu_stall(struct rcu_state *rsp)
 942{
 943        unsigned long flags;
 944        struct rcu_node *rnp = rcu_get_root(rsp);
 945
 946        /*
 947         * OK, time to rat on ourselves...
 948         * See Documentation/RCU/stallwarn.txt for info on how to debug
 949         * RCU CPU stall warnings.
 950         */
 951        printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
 952        print_cpu_stall_info_begin();
 953        print_cpu_stall_info(rsp, smp_processor_id());
 954        print_cpu_stall_info_end();
 955        printk(KERN_CONT " (t=%lu jiffies)\n", jiffies - rsp->gp_start);
 956        if (!trigger_all_cpu_backtrace())
 957                dump_stack();
 958
 959        raw_spin_lock_irqsave(&rnp->lock, flags);
 960        if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
 961                rsp->jiffies_stall = jiffies +
 962                                     3 * jiffies_till_stall_check() + 3;
 963        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 964
 965        set_need_resched();  /* kick ourselves to get things going. */
 966}
 967
 968static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
 969{
 970        unsigned long j;
 971        unsigned long js;
 972        struct rcu_node *rnp;
 973
 974        if (rcu_cpu_stall_suppress)
 975                return;
 976        j = ACCESS_ONCE(jiffies);
 977        js = ACCESS_ONCE(rsp->jiffies_stall);
 978        rnp = rdp->mynode;
 979        if (rcu_gp_in_progress(rsp) &&
 980            (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
 981
 982                /* We haven't checked in, so go dump stack. */
 983                print_cpu_stall(rsp);
 984
 985        } else if (rcu_gp_in_progress(rsp) &&
 986                   ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
 987
 988                /* They had a few time units to dump stack, so complain. */
 989                print_other_cpu_stall(rsp);
 990        }
 991}
 992
 993static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
 994{
 995        rcu_cpu_stall_suppress = 1;
 996        return NOTIFY_DONE;
 997}
 998
 999/**
1000 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1001 *
1002 * Set the stall-warning timeout way off into the future, thus preventing
1003 * any RCU CPU stall-warning messages from appearing in the current set of
1004 * RCU grace periods.
1005 *
1006 * The caller must disable hard irqs.
1007 */
1008void rcu_cpu_stall_reset(void)
1009{
1010        struct rcu_state *rsp;
1011
1012        for_each_rcu_flavor(rsp)
1013                rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
1014}
1015
1016static struct notifier_block rcu_panic_block = {
1017        .notifier_call = rcu_panic,
1018};
1019
1020static void __init check_cpu_stall_init(void)
1021{
1022        atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
1023}
1024
1025/*
1026 * Update CPU-local rcu_data state to record the newly noticed grace period.
1027 * This is used both when we started the grace period and when we notice
1028 * that someone else started the grace period.  The caller must hold the
1029 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
1030 *  and must have irqs disabled.
1031 */
1032static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1033{
1034        if (rdp->gpnum != rnp->gpnum) {
1035                /*
1036                 * If the current grace period is waiting for this CPU,
1037                 * set up to detect a quiescent state, otherwise don't
1038                 * go looking for one.
1039                 */
1040                rdp->gpnum = rnp->gpnum;
1041                trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
1042                rdp->passed_quiesce = 0;
1043                rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1044                zero_cpu_stall_ticks(rdp);
1045        }
1046}
1047
1048static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
1049{
1050        unsigned long flags;
1051        struct rcu_node *rnp;
1052
1053        local_irq_save(flags);
1054        rnp = rdp->mynode;
1055        if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
1056            !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1057                local_irq_restore(flags);
1058                return;
1059        }
1060        __note_new_gpnum(rsp, rnp, rdp);
1061        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1062}
1063
1064/*
1065 * Did someone else start a new RCU grace period start since we last
1066 * checked?  Update local state appropriately if so.  Must be called
1067 * on the CPU corresponding to rdp.
1068 */
1069static int
1070check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
1071{
1072        unsigned long flags;
1073        int ret = 0;
1074
1075        local_irq_save(flags);
1076        if (rdp->gpnum != rsp->gpnum) {
1077                note_new_gpnum(rsp, rdp);
1078                ret = 1;
1079        }
1080        local_irq_restore(flags);
1081        return ret;
1082}
1083
1084/*
1085 * Initialize the specified rcu_data structure's callback list to empty.
1086 */
1087static void init_callback_list(struct rcu_data *rdp)
1088{
1089        int i;
1090
1091        rdp->nxtlist = NULL;
1092        for (i = 0; i < RCU_NEXT_SIZE; i++)
1093                rdp->nxttail[i] = &rdp->nxtlist;
1094}
1095
1096/*
1097 * Advance this CPU's callbacks, but only if the current grace period
1098 * has ended.  This may be called only from the CPU to whom the rdp
1099 * belongs.  In addition, the corresponding leaf rcu_node structure's
1100 * ->lock must be held by the caller, with irqs disabled.
1101 */
1102static void
1103__rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1104{
1105        /* Did another grace period end? */
1106        if (rdp->completed != rnp->completed) {
1107
1108                /* Advance callbacks.  No harm if list empty. */
1109                rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
1110                rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
1111                rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1112
1113                /* Remember that we saw this grace-period completion. */
1114                rdp->completed = rnp->completed;
1115                trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
1116
1117                /*
1118                 * If we were in an extended quiescent state, we may have
1119                 * missed some grace periods that others CPUs handled on
1120                 * our behalf. Catch up with this state to avoid noting
1121                 * spurious new grace periods.  If another grace period
1122                 * has started, then rnp->gpnum will have advanced, so
1123                 * we will detect this later on.  Of course, any quiescent
1124                 * states we found for the old GP are now invalid.
1125                 */
1126                if (ULONG_CMP_LT(rdp->gpnum, rdp->completed)) {
1127                        rdp->gpnum = rdp->completed;
1128                        rdp->passed_quiesce = 0;
1129                }
1130
1131                /*
1132                 * If RCU does not need a quiescent state from this CPU,
1133                 * then make sure that this CPU doesn't go looking for one.
1134                 */
1135                if ((rnp->qsmask & rdp->grpmask) == 0)
1136                        rdp->qs_pending = 0;
1137        }
1138}
1139
1140/*
1141 * Advance this CPU's callbacks, but only if the current grace period
1142 * has ended.  This may be called only from the CPU to whom the rdp
1143 * belongs.
1144 */
1145static void
1146rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
1147{
1148        unsigned long flags;
1149        struct rcu_node *rnp;
1150
1151        local_irq_save(flags);
1152        rnp = rdp->mynode;
1153        if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
1154            !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1155                local_irq_restore(flags);
1156                return;
1157        }
1158        __rcu_process_gp_end(rsp, rnp, rdp);
1159        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1160}
1161
1162/*
1163 * Do per-CPU grace-period initialization for running CPU.  The caller
1164 * must hold the lock of the leaf rcu_node structure corresponding to
1165 * this CPU.
1166 */
1167static void
1168rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1169{
1170        /* Prior grace period ended, so advance callbacks for current CPU. */
1171        __rcu_process_gp_end(rsp, rnp, rdp);
1172
1173        /* Set state so that this CPU will detect the next quiescent state. */
1174        __note_new_gpnum(rsp, rnp, rdp);
1175}
1176
1177/*
1178 * Initialize a new grace period.
1179 */
1180static int rcu_gp_init(struct rcu_state *rsp)
1181{
1182        struct rcu_data *rdp;
1183        struct rcu_node *rnp = rcu_get_root(rsp);
1184
1185        raw_spin_lock_irq(&rnp->lock);
1186        rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1187
1188        if (rcu_gp_in_progress(rsp)) {
1189                /* Grace period already in progress, don't start another.  */
1190                raw_spin_unlock_irq(&rnp->lock);
1191                return 0;
1192        }
1193
1194        /* Advance to a new grace period and initialize state. */
1195        rsp->gpnum++;
1196        trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1197        record_gp_stall_check_time(rsp);
1198        raw_spin_unlock_irq(&rnp->lock);
1199
1200        /* Exclude any concurrent CPU-hotplug operations. */
1201        mutex_lock(&rsp->onoff_mutex);
1202
1203        /*
1204         * Set the quiescent-state-needed bits in all the rcu_node
1205         * structures for all currently online CPUs in breadth-first order,
1206         * starting from the root rcu_node structure, relying on the layout
1207         * of the tree within the rsp->node[] array.  Note that other CPUs
1208         * will access only the leaves of the hierarchy, thus seeing that no
1209         * grace period is in progress, at least until the corresponding
1210         * leaf node has been initialized.  In addition, we have excluded
1211         * CPU-hotplug operations.
1212         *
1213         * The grace period cannot complete until the initialization
1214         * process finishes, because this kthread handles both.
1215         */
1216        rcu_for_each_node_breadth_first(rsp, rnp) {
1217                raw_spin_lock_irq(&rnp->lock);
1218                rdp = this_cpu_ptr(rsp->rda);
1219                rcu_preempt_check_blocked_tasks(rnp);
1220                rnp->qsmask = rnp->qsmaskinit;
1221                rnp->gpnum = rsp->gpnum;
1222                WARN_ON_ONCE(rnp->completed != rsp->completed);
1223                rnp->completed = rsp->completed;
1224                if (rnp == rdp->mynode)
1225                        rcu_start_gp_per_cpu(rsp, rnp, rdp);
1226                rcu_preempt_boost_start_gp(rnp);
1227                trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1228                                            rnp->level, rnp->grplo,
1229                                            rnp->grphi, rnp->qsmask);
1230                raw_spin_unlock_irq(&rnp->lock);
1231#ifdef CONFIG_PROVE_RCU_DELAY
1232                if ((random32() % (rcu_num_nodes * 8)) == 0)
1233                        schedule_timeout_uninterruptible(2);
1234#endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1235                cond_resched();
1236        }
1237
1238        mutex_unlock(&rsp->onoff_mutex);
1239        return 1;
1240}
1241
1242/*
1243 * Do one round of quiescent-state forcing.
1244 */
1245int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1246{
1247        int fqs_state = fqs_state_in;
1248        struct rcu_node *rnp = rcu_get_root(rsp);
1249
1250        rsp->n_force_qs++;
1251        if (fqs_state == RCU_SAVE_DYNTICK) {
1252                /* Collect dyntick-idle snapshots. */
1253                force_qs_rnp(rsp, dyntick_save_progress_counter);
1254                fqs_state = RCU_FORCE_QS;
1255        } else {
1256                /* Handle dyntick-idle and offline CPUs. */
1257                force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1258        }
1259        /* Clear flag to prevent immediate re-entry. */
1260        if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1261                raw_spin_lock_irq(&rnp->lock);
1262                rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1263                raw_spin_unlock_irq(&rnp->lock);
1264        }
1265        return fqs_state;
1266}
1267
1268/*
1269 * Clean up after the old grace period.
1270 */
1271static void rcu_gp_cleanup(struct rcu_state *rsp)
1272{
1273        unsigned long gp_duration;
1274        struct rcu_data *rdp;
1275        struct rcu_node *rnp = rcu_get_root(rsp);
1276
1277        raw_spin_lock_irq(&rnp->lock);
1278        gp_duration = jiffies - rsp->gp_start;
1279        if (gp_duration > rsp->gp_max)
1280                rsp->gp_max = gp_duration;
1281
1282        /*
1283         * We know the grace period is complete, but to everyone else
1284         * it appears to still be ongoing.  But it is also the case
1285         * that to everyone else it looks like there is nothing that
1286         * they can do to advance the grace period.  It is therefore
1287         * safe for us to drop the lock in order to mark the grace
1288         * period as completed in all of the rcu_node structures.
1289         */
1290        raw_spin_unlock_irq(&rnp->lock);
1291
1292        /*
1293         * Propagate new ->completed value to rcu_node structures so
1294         * that other CPUs don't have to wait until the start of the next
1295         * grace period to process their callbacks.  This also avoids
1296         * some nasty RCU grace-period initialization races by forcing
1297         * the end of the current grace period to be completely recorded in
1298         * all of the rcu_node structures before the beginning of the next
1299         * grace period is recorded in any of the rcu_node structures.
1300         */
1301        rcu_for_each_node_breadth_first(rsp, rnp) {
1302                raw_spin_lock_irq(&rnp->lock);
1303                rnp->completed = rsp->gpnum;
1304                raw_spin_unlock_irq(&rnp->lock);
1305                cond_resched();
1306        }
1307        rnp = rcu_get_root(rsp);
1308        raw_spin_lock_irq(&rnp->lock);
1309
1310        rsp->completed = rsp->gpnum; /* Declare grace period done. */
1311        trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1312        rsp->fqs_state = RCU_GP_IDLE;
1313        rdp = this_cpu_ptr(rsp->rda);
1314        if (cpu_needs_another_gp(rsp, rdp))
1315                rsp->gp_flags = 1;
1316        raw_spin_unlock_irq(&rnp->lock);
1317}
1318
1319/*
1320 * Body of kthread that handles grace periods.
1321 */
1322static int __noreturn rcu_gp_kthread(void *arg)
1323{
1324        int fqs_state;
1325        unsigned long j;
1326        int ret;
1327        struct rcu_state *rsp = arg;
1328        struct rcu_node *rnp = rcu_get_root(rsp);
1329
1330        for (;;) {
1331
1332                /* Handle grace-period start. */
1333                for (;;) {
1334                        wait_event_interruptible(rsp->gp_wq,
1335                                                 rsp->gp_flags &
1336                                                 RCU_GP_FLAG_INIT);
1337                        if ((rsp->gp_flags & RCU_GP_FLAG_INIT) &&
1338                            rcu_gp_init(rsp))
1339                                break;
1340                        cond_resched();
1341                        flush_signals(current);
1342                }
1343
1344                /* Handle quiescent-state forcing. */
1345                fqs_state = RCU_SAVE_DYNTICK;
1346                j = jiffies_till_first_fqs;
1347                if (j > HZ) {
1348                        j = HZ;
1349                        jiffies_till_first_fqs = HZ;
1350                }
1351                for (;;) {
1352                        rsp->jiffies_force_qs = jiffies + j;
1353                        ret = wait_event_interruptible_timeout(rsp->gp_wq,
1354                                        (rsp->gp_flags & RCU_GP_FLAG_FQS) ||
1355                                        (!ACCESS_ONCE(rnp->qsmask) &&
1356                                         !rcu_preempt_blocked_readers_cgp(rnp)),
1357                                        j);
1358                        /* If grace period done, leave loop. */
1359                        if (!ACCESS_ONCE(rnp->qsmask) &&
1360                            !rcu_preempt_blocked_readers_cgp(rnp))
1361                                break;
1362                        /* If time for quiescent-state forcing, do it. */
1363                        if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) {
1364                                fqs_state = rcu_gp_fqs(rsp, fqs_state);
1365                                cond_resched();
1366                        } else {
1367                                /* Deal with stray signal. */
1368                                cond_resched();
1369                                flush_signals(current);
1370                        }
1371                        j = jiffies_till_next_fqs;
1372                        if (j > HZ) {
1373                                j = HZ;
1374                                jiffies_till_next_fqs = HZ;
1375                        } else if (j < 1) {
1376                                j = 1;
1377                                jiffies_till_next_fqs = 1;
1378                        }
1379                }
1380
1381                /* Handle grace-period end. */
1382                rcu_gp_cleanup(rsp);
1383        }
1384}
1385
1386/*
1387 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1388 * in preparation for detecting the next grace period.  The caller must hold
1389 * the root node's ->lock, which is released before return.  Hard irqs must
1390 * be disabled.
1391 *
1392 * Note that it is legal for a dying CPU (which is marked as offline) to
1393 * invoke this function.  This can happen when the dying CPU reports its
1394 * quiescent state.
1395 */
1396static void
1397rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1398        __releases(rcu_get_root(rsp)->lock)
1399{
1400        struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1401        struct rcu_node *rnp = rcu_get_root(rsp);
1402
1403        if (!rsp->gp_kthread ||
1404            !cpu_needs_another_gp(rsp, rdp)) {
1405                /*
1406                 * Either we have not yet spawned the grace-period
1407                 * task or this CPU does not need another grace period.
1408                 * Either way, don't start a new grace period.
1409                 */
1410                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1411                return;
1412        }
1413
1414        rsp->gp_flags = RCU_GP_FLAG_INIT;
1415        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1416        wake_up(&rsp->gp_wq);
1417}
1418
1419/*
1420 * Report a full set of quiescent states to the specified rcu_state
1421 * data structure.  This involves cleaning up after the prior grace
1422 * period and letting rcu_start_gp() start up the next grace period
1423 * if one is needed.  Note that the caller must hold rnp->lock, as
1424 * required by rcu_start_gp(), which will release it.
1425 */
1426static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1427        __releases(rcu_get_root(rsp)->lock)
1428{
1429        WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1430        raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1431        wake_up(&rsp->gp_wq);  /* Memory barrier implied by wake_up() path. */
1432}
1433
1434/*
1435 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1436 * Allows quiescent states for a group of CPUs to be reported at one go
1437 * to the specified rcu_node structure, though all the CPUs in the group
1438 * must be represented by the same rcu_node structure (which need not be
1439 * a leaf rcu_node structure, though it often will be).  That structure's
1440 * lock must be held upon entry, and it is released before return.
1441 */
1442static void
1443rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1444                  struct rcu_node *rnp, unsigned long flags)
1445        __releases(rnp->lock)
1446{
1447        struct rcu_node *rnp_c;
1448
1449        /* Walk up the rcu_node hierarchy. */
1450        for (;;) {
1451                if (!(rnp->qsmask & mask)) {
1452
1453                        /* Our bit has already been cleared, so done. */
1454                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1455                        return;
1456                }
1457                rnp->qsmask &= ~mask;
1458                trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1459                                                 mask, rnp->qsmask, rnp->level,
1460                                                 rnp->grplo, rnp->grphi,
1461                                                 !!rnp->gp_tasks);
1462                if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1463
1464                        /* Other bits still set at this level, so done. */
1465                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1466                        return;
1467                }
1468                mask = rnp->grpmask;
1469                if (rnp->parent == NULL) {
1470
1471                        /* No more levels.  Exit loop holding root lock. */
1472
1473                        break;
1474                }
1475                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1476                rnp_c = rnp;
1477                rnp = rnp->parent;
1478                raw_spin_lock_irqsave(&rnp->lock, flags);
1479                WARN_ON_ONCE(rnp_c->qsmask);
1480        }
1481
1482        /*
1483         * Get here if we are the last CPU to pass through a quiescent
1484         * state for this grace period.  Invoke rcu_report_qs_rsp()
1485         * to clean up and start the next grace period if one is needed.
1486         */
1487        rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1488}
1489
1490/*
1491 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1492 * structure.  This must be either called from the specified CPU, or
1493 * called when the specified CPU is known to be offline (and when it is
1494 * also known that no other CPU is concurrently trying to help the offline
1495 * CPU).  The lastcomp argument is used to make sure we are still in the
1496 * grace period of interest.  We don't want to end the current grace period
1497 * based on quiescent states detected in an earlier grace period!
1498 */
1499static void
1500rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1501{
1502        unsigned long flags;
1503        unsigned long mask;
1504        struct rcu_node *rnp;
1505
1506        rnp = rdp->mynode;
1507        raw_spin_lock_irqsave(&rnp->lock, flags);
1508        if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1509            rnp->completed == rnp->gpnum) {
1510
1511                /*
1512                 * The grace period in which this quiescent state was
1513                 * recorded has ended, so don't report it upwards.
1514                 * We will instead need a new quiescent state that lies
1515                 * within the current grace period.
1516                 */
1517                rdp->passed_quiesce = 0;        /* need qs for new gp. */
1518                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1519                return;
1520        }
1521        mask = rdp->grpmask;
1522        if ((rnp->qsmask & mask) == 0) {
1523                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1524        } else {
1525                rdp->qs_pending = 0;
1526
1527                /*
1528                 * This GP can't end until cpu checks in, so all of our
1529                 * callbacks can be processed during the next GP.
1530                 */
1531                rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1532
1533                rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1534        }
1535}
1536
1537/*
1538 * Check to see if there is a new grace period of which this CPU
1539 * is not yet aware, and if so, set up local rcu_data state for it.
1540 * Otherwise, see if this CPU has just passed through its first
1541 * quiescent state for this grace period, and record that fact if so.
1542 */
1543static void
1544rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1545{
1546        /* If there is now a new grace period, record and return. */
1547        if (check_for_new_grace_period(rsp, rdp))
1548                return;
1549
1550        /*
1551         * Does this CPU still need to do its part for current grace period?
1552         * If no, return and let the other CPUs do their part as well.
1553         */
1554        if (!rdp->qs_pending)
1555                return;
1556
1557        /*
1558         * Was there a quiescent state since the beginning of the grace
1559         * period? If no, then exit and wait for the next call.
1560         */
1561        if (!rdp->passed_quiesce)
1562                return;
1563
1564        /*
1565         * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1566         * judge of that).
1567         */
1568        rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1569}
1570
1571#ifdef CONFIG_HOTPLUG_CPU
1572
1573/*
1574 * Send the specified CPU's RCU callbacks to the orphanage.  The
1575 * specified CPU must be offline, and the caller must hold the
1576 * ->onofflock.
1577 */
1578static void
1579rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1580                          struct rcu_node *rnp, struct rcu_data *rdp)
1581{
1582        /*
1583         * Orphan the callbacks.  First adjust the counts.  This is safe
1584         * because ->onofflock excludes _rcu_barrier()'s adoption of
1585         * the callbacks, thus no memory barrier is required.
1586         */
1587        if (rdp->nxtlist != NULL) {
1588                rsp->qlen_lazy += rdp->qlen_lazy;
1589                rsp->qlen += rdp->qlen;
1590                rdp->n_cbs_orphaned += rdp->qlen;
1591                rdp->qlen_lazy = 0;
1592                ACCESS_ONCE(rdp->qlen) = 0;
1593        }
1594
1595        /*
1596         * Next, move those callbacks still needing a grace period to
1597         * the orphanage, where some other CPU will pick them up.
1598         * Some of the callbacks might have gone partway through a grace
1599         * period, but that is too bad.  They get to start over because we
1600         * cannot assume that grace periods are synchronized across CPUs.
1601         * We don't bother updating the ->nxttail[] array yet, instead
1602         * we just reset the whole thing later on.
1603         */
1604        if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1605                *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1606                rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1607                *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1608        }
1609
1610        /*
1611         * Then move the ready-to-invoke callbacks to the orphanage,
1612         * where some other CPU will pick them up.  These will not be
1613         * required to pass though another grace period: They are done.
1614         */
1615        if (rdp->nxtlist != NULL) {
1616                *rsp->orphan_donetail = rdp->nxtlist;
1617                rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1618        }
1619
1620        /* Finally, initialize the rcu_data structure's list to empty.  */
1621        init_callback_list(rdp);
1622}
1623
1624/*
1625 * Adopt the RCU callbacks from the specified rcu_state structure's
1626 * orphanage.  The caller must hold the ->onofflock.
1627 */
1628static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1629{
1630        int i;
1631        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1632
1633        /* Do the accounting first. */
1634        rdp->qlen_lazy += rsp->qlen_lazy;
1635        rdp->qlen += rsp->qlen;
1636        rdp->n_cbs_adopted += rsp->qlen;
1637        if (rsp->qlen_lazy != rsp->qlen)
1638                rcu_idle_count_callbacks_posted();
1639        rsp->qlen_lazy = 0;
1640        rsp->qlen = 0;
1641
1642        /*
1643         * We do not need a memory barrier here because the only way we
1644         * can get here if there is an rcu_barrier() in flight is if
1645         * we are the task doing the rcu_barrier().
1646         */
1647
1648        /* First adopt the ready-to-invoke callbacks. */
1649        if (rsp->orphan_donelist != NULL) {
1650                *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1651                *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1652                for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1653                        if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1654                                rdp->nxttail[i] = rsp->orphan_donetail;
1655                rsp->orphan_donelist = NULL;
1656                rsp->orphan_donetail = &rsp->orphan_donelist;
1657        }
1658
1659        /* And then adopt the callbacks that still need a grace period. */
1660        if (rsp->orphan_nxtlist != NULL) {
1661                *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1662                rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1663                rsp->orphan_nxtlist = NULL;
1664                rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1665        }
1666}
1667
1668/*
1669 * Trace the fact that this CPU is going offline.
1670 */
1671static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1672{
1673        RCU_TRACE(unsigned long mask);
1674        RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1675        RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1676
1677        RCU_TRACE(mask = rdp->grpmask);
1678        trace_rcu_grace_period(rsp->name,
1679                               rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1680                               "cpuofl");
1681}
1682
1683/*
1684 * The CPU has been completely removed, and some other CPU is reporting
1685 * this fact from process context.  Do the remainder of the cleanup,
1686 * including orphaning the outgoing CPU's RCU callbacks, and also
1687 * adopting them.  There can only be one CPU hotplug operation at a time,
1688 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1689 */
1690static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1691{
1692        unsigned long flags;
1693        unsigned long mask;
1694        int need_report = 0;
1695        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1696        struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
1697
1698        /* Adjust any no-longer-needed kthreads. */
1699        rcu_boost_kthread_setaffinity(rnp, -1);
1700
1701        /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1702
1703        /* Exclude any attempts to start a new grace period. */
1704        mutex_lock(&rsp->onoff_mutex);
1705        raw_spin_lock_irqsave(&rsp->onofflock, flags);
1706
1707        /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1708        rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1709        rcu_adopt_orphan_cbs(rsp);
1710
1711        /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1712        mask = rdp->grpmask;    /* rnp->grplo is constant. */
1713        do {
1714                raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
1715                rnp->qsmaskinit &= ~mask;
1716                if (rnp->qsmaskinit != 0) {
1717                        if (rnp != rdp->mynode)
1718                                raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1719                        break;
1720                }
1721                if (rnp == rdp->mynode)
1722                        need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1723                else
1724                        raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1725                mask = rnp->grpmask;
1726                rnp = rnp->parent;
1727        } while (rnp != NULL);
1728
1729        /*
1730         * We still hold the leaf rcu_node structure lock here, and
1731         * irqs are still disabled.  The reason for this subterfuge is
1732         * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1733         * held leads to deadlock.
1734         */
1735        raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1736        rnp = rdp->mynode;
1737        if (need_report & RCU_OFL_TASKS_NORM_GP)
1738                rcu_report_unblock_qs_rnp(rnp, flags);
1739        else
1740                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1741        if (need_report & RCU_OFL_TASKS_EXP_GP)
1742                rcu_report_exp_rnp(rsp, rnp, true);
1743        WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1744                  "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1745                  cpu, rdp->qlen, rdp->nxtlist);
1746        init_callback_list(rdp);
1747        /* Disallow further callbacks on this CPU. */
1748        rdp->nxttail[RCU_NEXT_TAIL] = NULL;
1749        mutex_unlock(&rsp->onoff_mutex);
1750}
1751
1752#else /* #ifdef CONFIG_HOTPLUG_CPU */
1753
1754static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1755{
1756}
1757
1758static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1759{
1760}
1761
1762#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1763
1764/*
1765 * Invoke any RCU callbacks that have made it to the end of their grace
1766 * period.  Thottle as specified by rdp->blimit.
1767 */
1768static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1769{
1770        unsigned long flags;
1771        struct rcu_head *next, *list, **tail;
1772        long bl, count, count_lazy;
1773        int i;
1774
1775        /* If no callbacks are ready, just return.*/
1776        if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1777                trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1778                trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1779                                    need_resched(), is_idle_task(current),
1780                                    rcu_is_callbacks_kthread());
1781                return;
1782        }
1783
1784        /*
1785         * Extract the list of ready callbacks, disabling to prevent
1786         * races with call_rcu() from interrupt handlers.
1787         */
1788        local_irq_save(flags);
1789        WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1790        bl = rdp->blimit;
1791        trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1792        list = rdp->nxtlist;
1793        rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1794        *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1795        tail = rdp->nxttail[RCU_DONE_TAIL];
1796        for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1797                if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1798                        rdp->nxttail[i] = &rdp->nxtlist;
1799        local_irq_restore(flags);
1800
1801        /* Invoke callbacks. */
1802        count = count_lazy = 0;
1803        while (list) {
1804                next = list->next;
1805                prefetch(next);
1806                debug_rcu_head_unqueue(list);
1807                if (__rcu_reclaim(rsp->name, list))
1808                        count_lazy++;
1809                list = next;
1810                /* Stop only if limit reached and CPU has something to do. */
1811                if (++count >= bl &&
1812                    (need_resched() ||
1813                     (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1814                        break;
1815        }
1816
1817        local_irq_save(flags);
1818        trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1819                            is_idle_task(current),
1820                            rcu_is_callbacks_kthread());
1821
1822        /* Update count, and requeue any remaining callbacks. */
1823        if (list != NULL) {
1824                *tail = rdp->nxtlist;
1825                rdp->nxtlist = list;
1826                for (i = 0; i < RCU_NEXT_SIZE; i++)
1827                        if (&rdp->nxtlist == rdp->nxttail[i])
1828                                rdp->nxttail[i] = tail;
1829                        else
1830                                break;
1831        }
1832        smp_mb(); /* List handling before counting for rcu_barrier(). */
1833        rdp->qlen_lazy -= count_lazy;
1834        ACCESS_ONCE(rdp->qlen) -= count;
1835        rdp->n_cbs_invoked += count;
1836
1837        /* Reinstate batch limit if we have worked down the excess. */
1838        if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1839                rdp->blimit = blimit;
1840
1841        /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1842        if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1843                rdp->qlen_last_fqs_check = 0;
1844                rdp->n_force_qs_snap = rsp->n_force_qs;
1845        } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1846                rdp->qlen_last_fqs_check = rdp->qlen;
1847        WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
1848
1849        local_irq_restore(flags);
1850
1851        /* Re-invoke RCU core processing if there are callbacks remaining. */
1852        if (cpu_has_callbacks_ready_to_invoke(rdp))
1853                invoke_rcu_core();
1854}
1855
1856/*
1857 * Check to see if this CPU is in a non-context-switch quiescent state
1858 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1859 * Also schedule RCU core processing.
1860 *
1861 * This function must be called from hardirq context.  It is normally
1862 * invoked from the scheduling-clock interrupt.  If rcu_pending returns
1863 * false, there is no point in invoking rcu_check_callbacks().
1864 */
1865void rcu_check_callbacks(int cpu, int user)
1866{
1867        trace_rcu_utilization("Start scheduler-tick");
1868        increment_cpu_stall_ticks();
1869        if (user || rcu_is_cpu_rrupt_from_idle()) {
1870
1871                /*
1872                 * Get here if this CPU took its interrupt from user
1873                 * mode or from the idle loop, and if this is not a
1874                 * nested interrupt.  In this case, the CPU is in
1875                 * a quiescent state, so note it.
1876                 *
1877                 * No memory barrier is required here because both
1878                 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1879                 * variables that other CPUs neither access nor modify,
1880                 * at least not while the corresponding CPU is online.
1881                 */
1882
1883                rcu_sched_qs(cpu);
1884                rcu_bh_qs(cpu);
1885
1886        } else if (!in_softirq()) {
1887
1888                /*
1889                 * Get here if this CPU did not take its interrupt from
1890                 * softirq, in other words, if it is not interrupting
1891                 * a rcu_bh read-side critical section.  This is an _bh
1892                 * critical section, so note it.
1893                 */
1894
1895                rcu_bh_qs(cpu);
1896        }
1897        rcu_preempt_check_callbacks(cpu);
1898        if (rcu_pending(cpu))
1899                invoke_rcu_core();
1900        trace_rcu_utilization("End scheduler-tick");
1901}
1902
1903/*
1904 * Scan the leaf rcu_node structures, processing dyntick state for any that
1905 * have not yet encountered a quiescent state, using the function specified.
1906 * Also initiate boosting for any threads blocked on the root rcu_node.
1907 *
1908 * The caller must have suppressed start of new grace periods.
1909 */
1910static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1911{
1912        unsigned long bit;
1913        int cpu;
1914        unsigned long flags;
1915        unsigned long mask;
1916        struct rcu_node *rnp;
1917
1918        rcu_for_each_leaf_node(rsp, rnp) {
1919                cond_resched();
1920                mask = 0;
1921                raw_spin_lock_irqsave(&rnp->lock, flags);
1922                if (!rcu_gp_in_progress(rsp)) {
1923                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1924                        return;
1925                }
1926                if (rnp->qsmask == 0) {
1927                        rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1928                        continue;
1929                }
1930                cpu = rnp->grplo;
1931                bit = 1;
1932                for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1933                        if ((rnp->qsmask & bit) != 0 &&
1934                            f(per_cpu_ptr(rsp->rda, cpu)))
1935                                mask |= bit;
1936                }
1937                if (mask != 0) {
1938
1939                        /* rcu_report_qs_rnp() releases rnp->lock. */
1940                        rcu_report_qs_rnp(mask, rsp, rnp, flags);
1941                        continue;
1942                }
1943                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1944        }
1945        rnp = rcu_get_root(rsp);
1946        if (rnp->qsmask == 0) {
1947                raw_spin_lock_irqsave(&rnp->lock, flags);
1948                rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1949        }
1950}
1951
1952/*
1953 * Force quiescent states on reluctant CPUs, and also detect which
1954 * CPUs are in dyntick-idle mode.
1955 */
1956static void force_quiescent_state(struct rcu_state *rsp)
1957{
1958        unsigned long flags;
1959        bool ret;
1960        struct rcu_node *rnp;
1961        struct rcu_node *rnp_old = NULL;
1962
1963        /* Funnel through hierarchy to reduce memory contention. */
1964        rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
1965        for (; rnp != NULL; rnp = rnp->parent) {
1966                ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
1967                      !raw_spin_trylock(&rnp->fqslock);
1968                if (rnp_old != NULL)
1969                        raw_spin_unlock(&rnp_old->fqslock);
1970                if (ret) {
1971                        rsp->n_force_qs_lh++;
1972                        return;
1973                }
1974                rnp_old = rnp;
1975        }
1976        /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
1977
1978        /* Reached the root of the rcu_node tree, acquire lock. */
1979        raw_spin_lock_irqsave(&rnp_old->lock, flags);
1980        raw_spin_unlock(&rnp_old->fqslock);
1981        if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1982                rsp->n_force_qs_lh++;
1983                raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
1984                return;  /* Someone beat us to it. */
1985        }
1986        rsp->gp_flags |= RCU_GP_FLAG_FQS;
1987        raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
1988        wake_up(&rsp->gp_wq);  /* Memory barrier implied by wake_up() path. */
1989}
1990
1991/*
1992 * This does the RCU core processing work for the specified rcu_state
1993 * and rcu_data structures.  This may be called only from the CPU to
1994 * whom the rdp belongs.
1995 */
1996static void
1997__rcu_process_callbacks(struct rcu_state *rsp)
1998{
1999        unsigned long flags;
2000        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2001
2002        WARN_ON_ONCE(rdp->beenonline == 0);
2003
2004        /*
2005         * Advance callbacks in response to end of earlier grace
2006         * period that some other CPU ended.
2007         */
2008        rcu_process_gp_end(rsp, rdp);
2009
2010        /* Update RCU state based on any recent quiescent states. */
2011        rcu_check_quiescent_state(rsp, rdp);
2012
2013        /* Does this CPU require a not-yet-started grace period? */
2014        if (cpu_needs_another_gp(rsp, rdp)) {
2015                raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
2016                rcu_start_gp(rsp, flags);  /* releases above lock */
2017        }
2018
2019        /* If there are callbacks ready, invoke them. */
2020        if (cpu_has_callbacks_ready_to_invoke(rdp))
2021                invoke_rcu_callbacks(rsp, rdp);
2022}
2023
2024/*
2025 * Do RCU core processing for the current CPU.
2026 */
2027static void rcu_process_callbacks(struct softirq_action *unused)
2028{
2029        struct rcu_state *rsp;
2030
2031        if (cpu_is_offline(smp_processor_id()))
2032                return;
2033        trace_rcu_utilization("Start RCU core");
2034        for_each_rcu_flavor(rsp)
2035                __rcu_process_callbacks(rsp);
2036        trace_rcu_utilization("End RCU core");
2037}
2038
2039/*
2040 * Schedule RCU callback invocation.  If the specified type of RCU
2041 * does not support RCU priority boosting, just do a direct call,
2042 * otherwise wake up the per-CPU kernel kthread.  Note that because we
2043 * are running on the current CPU with interrupts disabled, the
2044 * rcu_cpu_kthread_task cannot disappear out from under us.
2045 */
2046static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2047{
2048        if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2049                return;
2050        if (likely(!rsp->boost)) {
2051                rcu_do_batch(rsp, rdp);
2052                return;
2053        }
2054        invoke_rcu_callbacks_kthread();
2055}
2056
2057static void invoke_rcu_core(void)
2058{
2059        raise_softirq(RCU_SOFTIRQ);
2060}
2061
2062/*
2063 * Handle any core-RCU processing required by a call_rcu() invocation.
2064 */
2065static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2066                            struct rcu_head *head, unsigned long flags)
2067{
2068        /*
2069         * If called from an extended quiescent state, invoke the RCU
2070         * core in order to force a re-evaluation of RCU's idleness.
2071         */
2072        if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2073                invoke_rcu_core();
2074
2075        /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2076        if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2077                return;
2078
2079        /*
2080         * Force the grace period if too many callbacks or too long waiting.
2081         * Enforce hysteresis, and don't invoke force_quiescent_state()
2082         * if some other CPU has recently done so.  Also, don't bother
2083         * invoking force_quiescent_state() if the newly enqueued callback
2084         * is the only one waiting for a grace period to complete.
2085         */
2086        if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2087
2088                /* Are we ignoring a completed grace period? */
2089                rcu_process_gp_end(rsp, rdp);
2090                check_for_new_grace_period(rsp, rdp);
2091
2092                /* Start a new grace period if one not already started. */
2093                if (!rcu_gp_in_progress(rsp)) {
2094                        unsigned long nestflag;
2095                        struct rcu_node *rnp_root = rcu_get_root(rsp);
2096
2097                        raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
2098                        rcu_start_gp(rsp, nestflag);  /* rlses rnp_root->lock */
2099                } else {
2100                        /* Give the grace period a kick. */
2101                        rdp->blimit = LONG_MAX;
2102                        if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2103                            *rdp->nxttail[RCU_DONE_TAIL] != head)
2104                                force_quiescent_state(rsp);
2105                        rdp->n_force_qs_snap = rsp->n_force_qs;
2106                        rdp->qlen_last_fqs_check = rdp->qlen;
2107                }
2108        }
2109}
2110
2111static void
2112__call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2113           struct rcu_state *rsp, bool lazy)
2114{
2115        unsigned long flags;
2116        struct rcu_data *rdp;
2117
2118        WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2119        debug_rcu_head_queue(head);
2120        head->func = func;
2121        head->next = NULL;
2122
2123        /*
2124         * Opportunistically note grace-period endings and beginnings.
2125         * Note that we might see a beginning right after we see an
2126         * end, but never vice versa, since this CPU has to pass through
2127         * a quiescent state betweentimes.
2128         */
2129        local_irq_save(flags);
2130        rdp = this_cpu_ptr(rsp->rda);
2131
2132        /* Add the callback to our list. */
2133        if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL)) {
2134                /* _call_rcu() is illegal on offline CPU; leak the callback. */
2135                WARN_ON_ONCE(1);
2136                local_irq_restore(flags);
2137                return;
2138        }
2139        ACCESS_ONCE(rdp->qlen)++;
2140        if (lazy)
2141                rdp->qlen_lazy++;
2142        else
2143                rcu_idle_count_callbacks_posted();
2144        smp_mb();  /* Count before adding callback for rcu_barrier(). */
2145        *rdp->nxttail[RCU_NEXT_TAIL] = head;
2146        rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2147
2148        if (__is_kfree_rcu_offset((unsigned long)func))
2149                trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2150                                         rdp->qlen_lazy, rdp->qlen);
2151        else
2152                trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2153
2154        /* Go handle any RCU core processing required. */
2155        __call_rcu_core(rsp, rdp, head, flags);
2156        local_irq_restore(flags);
2157}
2158
2159/*
2160 * Queue an RCU-sched callback for invocation after a grace period.
2161 */
2162void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2163{
2164        __call_rcu(head, func, &rcu_sched_state, 0);
2165}
2166EXPORT_SYMBOL_GPL(call_rcu_sched);
2167
2168/*
2169 * Queue an RCU callback for invocation after a quicker grace period.
2170 */
2171void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2172{
2173        __call_rcu(head, func, &rcu_bh_state, 0);
2174}
2175EXPORT_SYMBOL_GPL(call_rcu_bh);
2176
2177/*
2178 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2179 * any blocking grace-period wait automatically implies a grace period
2180 * if there is only one CPU online at any point time during execution
2181 * of either synchronize_sched() or synchronize_rcu_bh().  It is OK to
2182 * occasionally incorrectly indicate that there are multiple CPUs online
2183 * when there was in fact only one the whole time, as this just adds
2184 * some overhead: RCU still operates correctly.
2185 */
2186static inline int rcu_blocking_is_gp(void)
2187{
2188        int ret;
2189
2190        might_sleep();  /* Check for RCU read-side critical section. */
2191        preempt_disable();
2192        ret = num_online_cpus() <= 1;
2193        preempt_enable();
2194        return ret;
2195}
2196
2197/**
2198 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2199 *
2200 * Control will return to the caller some time after a full rcu-sched
2201 * grace period has elapsed, in other words after all currently executing
2202 * rcu-sched read-side critical sections have completed.   These read-side
2203 * critical sections are delimited by rcu_read_lock_sched() and
2204 * rcu_read_unlock_sched(), and may be nested.  Note that preempt_disable(),
2205 * local_irq_disable(), and so on may be used in place of
2206 * rcu_read_lock_sched().
2207 *
2208 * This means that all preempt_disable code sequences, including NMI and
2209 * hardware-interrupt handlers, in progress on entry will have completed
2210 * before this primitive returns.  However, this does not guarantee that
2211 * softirq handlers will have completed, since in some kernels, these
2212 * handlers can run in process context, and can block.
2213 *
2214 * This primitive provides the guarantees made by the (now removed)
2215 * synchronize_kernel() API.  In contrast, synchronize_rcu() only
2216 * guarantees that rcu_read_lock() sections will have completed.
2217 * In "classic RCU", these two guarantees happen to be one and
2218 * the same, but can differ in realtime RCU implementations.
2219 */
2220void synchronize_sched(void)
2221{
2222        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2223                           !lock_is_held(&rcu_lock_map) &&
2224                           !lock_is_held(&rcu_sched_lock_map),
2225                           "Illegal synchronize_sched() in RCU-sched read-side critical section");
2226        if (rcu_blocking_is_gp())
2227                return;
2228        wait_rcu_gp(call_rcu_sched);
2229}
2230EXPORT_SYMBOL_GPL(synchronize_sched);
2231
2232/**
2233 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2234 *
2235 * Control will return to the caller some time after a full rcu_bh grace
2236 * period has elapsed, in other words after all currently executing rcu_bh
2237 * read-side critical sections have completed.  RCU read-side critical
2238 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2239 * and may be nested.
2240 */
2241void synchronize_rcu_bh(void)
2242{
2243        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2244                           !lock_is_held(&rcu_lock_map) &&
2245                           !lock_is_held(&rcu_sched_lock_map),
2246                           "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2247        if (rcu_blocking_is_gp())
2248                return;
2249        wait_rcu_gp(call_rcu_bh);
2250}
2251EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2252
2253static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
2254static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
2255
2256static int synchronize_sched_expedited_cpu_stop(void *data)
2257{
2258        /*
2259         * There must be a full memory barrier on each affected CPU
2260         * between the time that try_stop_cpus() is called and the
2261         * time that it returns.
2262         *
2263         * In the current initial implementation of cpu_stop, the
2264         * above condition is already met when the control reaches
2265         * this point and the following smp_mb() is not strictly
2266         * necessary.  Do smp_mb() anyway for documentation and
2267         * robustness against future implementation changes.
2268         */
2269        smp_mb(); /* See above comment block. */
2270        return 0;
2271}
2272
2273/**
2274 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2275 *
2276 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2277 * approach to force the grace period to end quickly.  This consumes
2278 * significant time on all CPUs and is unfriendly to real-time workloads,
2279 * so is thus not recommended for any sort of common-case code.  In fact,
2280 * if you are using synchronize_sched_expedited() in a loop, please
2281 * restructure your code to batch your updates, and then use a single
2282 * synchronize_sched() instead.
2283 *
2284 * Note that it is illegal to call this function while holding any lock
2285 * that is acquired by a CPU-hotplug notifier.  And yes, it is also illegal
2286 * to call this function from a CPU-hotplug notifier.  Failing to observe
2287 * these restriction will result in deadlock.
2288 *
2289 * This implementation can be thought of as an application of ticket
2290 * locking to RCU, with sync_sched_expedited_started and
2291 * sync_sched_expedited_done taking on the roles of the halves
2292 * of the ticket-lock word.  Each task atomically increments
2293 * sync_sched_expedited_started upon entry, snapshotting the old value,
2294 * then attempts to stop all the CPUs.  If this succeeds, then each
2295 * CPU will have executed a context switch, resulting in an RCU-sched
2296 * grace period.  We are then done, so we use atomic_cmpxchg() to
2297 * update sync_sched_expedited_done to match our snapshot -- but
2298 * only if someone else has not already advanced past our snapshot.
2299 *
2300 * On the other hand, if try_stop_cpus() fails, we check the value
2301 * of sync_sched_expedited_done.  If it has advanced past our
2302 * initial snapshot, then someone else must have forced a grace period
2303 * some time after we took our snapshot.  In this case, our work is
2304 * done for us, and we can simply return.  Otherwise, we try again,
2305 * but keep our initial snapshot for purposes of checking for someone
2306 * doing our work for us.
2307 *
2308 * If we fail too many times in a row, we fall back to synchronize_sched().
2309 */
2310void synchronize_sched_expedited(void)
2311{
2312        int firstsnap, s, snap, trycount = 0;
2313
2314        /* Note that atomic_inc_return() implies full memory barrier. */
2315        firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
2316        get_online_cpus();
2317        WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2318
2319        /*
2320         * Each pass through the following loop attempts to force a
2321         * context switch on each CPU.
2322         */
2323        while (try_stop_cpus(cpu_online_mask,
2324                             synchronize_sched_expedited_cpu_stop,
2325                             NULL) == -EAGAIN) {
2326                put_online_cpus();
2327
2328                /* No joy, try again later.  Or just synchronize_sched(). */
2329                if (trycount++ < 10) {
2330                        udelay(trycount * num_online_cpus());
2331                } else {
2332                        synchronize_sched();
2333                        return;
2334                }
2335
2336                /* Check to see if someone else did our work for us. */
2337                s = atomic_read(&sync_sched_expedited_done);
2338                if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
2339                        smp_mb(); /* ensure test happens before caller kfree */
2340                        return;
2341                }
2342
2343                /*
2344                 * Refetching sync_sched_expedited_started allows later
2345                 * callers to piggyback on our grace period.  We subtract
2346                 * 1 to get the same token that the last incrementer got.
2347                 * We retry after they started, so our grace period works
2348                 * for them, and they started after our first try, so their
2349                 * grace period works for us.
2350                 */
2351                get_online_cpus();
2352                snap = atomic_read(&sync_sched_expedited_started);
2353                smp_mb(); /* ensure read is before try_stop_cpus(). */
2354        }
2355
2356        /*
2357         * Everyone up to our most recent fetch is covered by our grace
2358         * period.  Update the counter, but only if our work is still
2359         * relevant -- which it won't be if someone who started later
2360         * than we did beat us to the punch.
2361         */
2362        do {
2363                s = atomic_read(&sync_sched_expedited_done);
2364                if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
2365                        smp_mb(); /* ensure test happens before caller kfree */
2366                        break;
2367                }
2368        } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
2369
2370        put_online_cpus();
2371}
2372EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2373
2374/*
2375 * Check to see if there is any immediate RCU-related work to be done
2376 * by the current CPU, for the specified type of RCU, returning 1 if so.
2377 * The checks are in order of increasing expense: checks that can be
2378 * carried out against CPU-local state are performed first.  However,
2379 * we must check for CPU stalls first, else we might not get a chance.
2380 */
2381static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2382{
2383        struct rcu_node *rnp = rdp->mynode;
2384
2385        rdp->n_rcu_pending++;
2386
2387        /* Check for CPU stalls, if enabled. */
2388        check_cpu_stall(rsp, rdp);
2389
2390        /* Is the RCU core waiting for a quiescent state from this CPU? */
2391        if (rcu_scheduler_fully_active &&
2392            rdp->qs_pending && !rdp->passed_quiesce) {
2393                rdp->n_rp_qs_pending++;
2394        } else if (rdp->qs_pending && rdp->passed_quiesce) {
2395                rdp->n_rp_report_qs++;
2396                return 1;
2397        }
2398
2399        /* Does this CPU have callbacks ready to invoke? */
2400        if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2401                rdp->n_rp_cb_ready++;
2402                return 1;
2403        }
2404
2405        /* Has RCU gone idle with this CPU needing another grace period? */
2406        if (cpu_needs_another_gp(rsp, rdp)) {
2407                rdp->n_rp_cpu_needs_gp++;
2408                return 1;
2409        }
2410
2411        /* Has another RCU grace period completed?  */
2412        if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2413                rdp->n_rp_gp_completed++;
2414                return 1;
2415        }
2416
2417        /* Has a new RCU grace period started? */
2418        if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2419                rdp->n_rp_gp_started++;
2420                return 1;
2421        }
2422
2423        /* nothing to do */
2424        rdp->n_rp_need_nothing++;
2425        return 0;
2426}
2427
2428/*
2429 * Check to see if there is any immediate RCU-related work to be done
2430 * by the current CPU, returning 1 if so.  This function is part of the
2431 * RCU implementation; it is -not- an exported member of the RCU API.
2432 */
2433static int rcu_pending(int cpu)
2434{
2435        struct rcu_state *rsp;
2436
2437        for_each_rcu_flavor(rsp)
2438                if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2439                        return 1;
2440        return 0;
2441}
2442
2443/*
2444 * Check to see if any future RCU-related work will need to be done
2445 * by the current CPU, even if none need be done immediately, returning
2446 * 1 if so.
2447 */
2448static int rcu_cpu_has_callbacks(int cpu)
2449{
2450        struct rcu_state *rsp;
2451
2452        /* RCU callbacks either ready or pending? */
2453        for_each_rcu_flavor(rsp)
2454                if (per_cpu_ptr(rsp->rda, cpu)->nxtlist)
2455                        return 1;
2456        return 0;
2457}
2458
2459/*
2460 * Helper function for _rcu_barrier() tracing.  If tracing is disabled,
2461 * the compiler is expected to optimize this away.
2462 */
2463static void _rcu_barrier_trace(struct rcu_state *rsp, char *s,
2464                               int cpu, unsigned long done)
2465{
2466        trace_rcu_barrier(rsp->name, s, cpu,
2467                          atomic_read(&rsp->barrier_cpu_count), done);
2468}
2469
2470/*
2471 * RCU callback function for _rcu_barrier().  If we are last, wake
2472 * up the task executing _rcu_barrier().
2473 */
2474static void rcu_barrier_callback(struct rcu_head *rhp)
2475{
2476        struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2477        struct rcu_state *rsp = rdp->rsp;
2478
2479        if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2480                _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2481                complete(&rsp->barrier_completion);
2482        } else {
2483                _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2484        }
2485}
2486
2487/*
2488 * Called with preemption disabled, and from cross-cpu IRQ context.
2489 */
2490static void rcu_barrier_func(void *type)
2491{
2492        struct rcu_state *rsp = type;
2493        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2494
2495        _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2496        atomic_inc(&rsp->barrier_cpu_count);
2497        rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2498}
2499
2500/*
2501 * Orchestrate the specified type of RCU barrier, waiting for all
2502 * RCU callbacks of the specified type to complete.
2503 */
2504static void _rcu_barrier(struct rcu_state *rsp)
2505{
2506        int cpu;
2507        struct rcu_data *rdp;
2508        unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2509        unsigned long snap_done;
2510
2511        _rcu_barrier_trace(rsp, "Begin", -1, snap);
2512
2513        /* Take mutex to serialize concurrent rcu_barrier() requests. */
2514        mutex_lock(&rsp->barrier_mutex);
2515
2516        /*
2517         * Ensure that all prior references, including to ->n_barrier_done,
2518         * are ordered before the _rcu_barrier() machinery.
2519         */
2520        smp_mb();  /* See above block comment. */
2521
2522        /*
2523         * Recheck ->n_barrier_done to see if others did our work for us.
2524         * This means checking ->n_barrier_done for an even-to-odd-to-even
2525         * transition.  The "if" expression below therefore rounds the old
2526         * value up to the next even number and adds two before comparing.
2527         */
2528        snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2529        _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2530        if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2531                _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2532                smp_mb(); /* caller's subsequent code after above check. */
2533                mutex_unlock(&rsp->barrier_mutex);
2534                return;
2535        }
2536
2537        /*
2538         * Increment ->n_barrier_done to avoid duplicate work.  Use
2539         * ACCESS_ONCE() to prevent the compiler from speculating
2540         * the increment to precede the early-exit check.
2541         */
2542        ACCESS_ONCE(rsp->n_barrier_done)++;
2543        WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2544        _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2545        smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2546
2547        /*
2548         * Initialize the count to one rather than to zero in order to
2549         * avoid a too-soon return to zero in case of a short grace period
2550         * (or preemption of this task).  Exclude CPU-hotplug operations
2551         * to ensure that no offline CPU has callbacks queued.
2552         */
2553        init_completion(&rsp->barrier_completion);
2554        atomic_set(&rsp->barrier_cpu_count, 1);
2555        get_online_cpus();
2556
2557        /*
2558         * Force each CPU with callbacks to register a new callback.
2559         * When that callback is invoked, we will know that all of the
2560         * corresponding CPU's preceding callbacks have been invoked.
2561         */
2562        for_each_online_cpu(cpu) {
2563                rdp = per_cpu_ptr(rsp->rda, cpu);
2564                if (ACCESS_ONCE(rdp->qlen)) {
2565                        _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2566                                           rsp->n_barrier_done);
2567                        smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2568                } else {
2569                        _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2570                                           rsp->n_barrier_done);
2571                }
2572        }
2573        put_online_cpus();
2574
2575        /*
2576         * Now that we have an rcu_barrier_callback() callback on each
2577         * CPU, and thus each counted, remove the initial count.
2578         */
2579        if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2580                complete(&rsp->barrier_completion);
2581
2582        /* Increment ->n_barrier_done to prevent duplicate work. */
2583        smp_mb(); /* Keep increment after above mechanism. */
2584        ACCESS_ONCE(rsp->n_barrier_done)++;
2585        WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2586        _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2587        smp_mb(); /* Keep increment before caller's subsequent code. */
2588
2589        /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2590        wait_for_completion(&rsp->barrier_completion);
2591
2592        /* Other rcu_barrier() invocations can now safely proceed. */
2593        mutex_unlock(&rsp->barrier_mutex);
2594}
2595
2596/**
2597 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2598 */
2599void rcu_barrier_bh(void)
2600{
2601        _rcu_barrier(&rcu_bh_state);
2602}
2603EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2604
2605/**
2606 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2607 */
2608void rcu_barrier_sched(void)
2609{
2610        _rcu_barrier(&rcu_sched_state);
2611}
2612EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2613
2614/*
2615 * Do boot-time initialization of a CPU's per-CPU RCU data.
2616 */
2617static void __init
2618rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2619{
2620        unsigned long flags;
2621        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2622        struct rcu_node *rnp = rcu_get_root(rsp);
2623
2624        /* Set up local state, ensuring consistent view of global state. */
2625        raw_spin_lock_irqsave(&rnp->lock, flags);
2626        rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2627        init_callback_list(rdp);
2628        rdp->qlen_lazy = 0;
2629        ACCESS_ONCE(rdp->qlen) = 0;
2630        rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2631        WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2632        WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2633#ifdef CONFIG_RCU_USER_QS
2634        WARN_ON_ONCE(rdp->dynticks->in_user);
2635#endif
2636        rdp->cpu = cpu;
2637        rdp->rsp = rsp;
2638        raw_spin_unlock_irqrestore(&rnp->lock, flags);
2639}
2640
2641/*
2642 * Initialize a CPU's per-CPU RCU data.  Note that only one online or
2643 * offline event can be happening at a given time.  Note also that we
2644 * can accept some slop in the rsp->completed access due to the fact
2645 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2646 */
2647static void __cpuinit
2648rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2649{
2650        unsigned long flags;
2651        unsigned long mask;
2652        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2653        struct rcu_node *rnp = rcu_get_root(rsp);
2654
2655        /* Exclude new grace periods. */
2656        mutex_lock(&rsp->onoff_mutex);
2657
2658        /* Set up local state, ensuring consistent view of global state. */
2659        raw_spin_lock_irqsave(&rnp->lock, flags);
2660        rdp->beenonline = 1;     /* We have now been online. */
2661        rdp->preemptible = preemptible;
2662        rdp->qlen_last_fqs_check = 0;
2663        rdp->n_force_qs_snap = rsp->n_force_qs;
2664        rdp->blimit = blimit;
2665        init_callback_list(rdp);  /* Re-enable callbacks on this CPU. */
2666        rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2667        atomic_set(&rdp->dynticks->dynticks,
2668                   (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2669        rcu_prepare_for_idle_init(cpu);
2670        raw_spin_unlock(&rnp->lock);            /* irqs remain disabled. */
2671
2672        /* Add CPU to rcu_node bitmasks. */
2673        rnp = rdp->mynode;
2674        mask = rdp->grpmask;
2675        do {
2676                /* Exclude any attempts to start a new GP on small systems. */
2677                raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
2678                rnp->qsmaskinit |= mask;
2679                mask = rnp->grpmask;
2680                if (rnp == rdp->mynode) {
2681                        /*
2682                         * If there is a grace period in progress, we will
2683                         * set up to wait for it next time we run the
2684                         * RCU core code.
2685                         */
2686                        rdp->gpnum = rnp->completed;
2687                        rdp->completed = rnp->completed;
2688                        rdp->passed_quiesce = 0;
2689                        rdp->qs_pending = 0;
2690                        trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2691                }
2692                raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2693                rnp = rnp->parent;
2694        } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2695        local_irq_restore(flags);
2696
2697        mutex_unlock(&rsp->onoff_mutex);
2698}
2699
2700static void __cpuinit rcu_prepare_cpu(int cpu)
2701{
2702        struct rcu_state *rsp;
2703
2704        for_each_rcu_flavor(rsp)
2705                rcu_init_percpu_data(cpu, rsp,
2706                                     strcmp(rsp->name, "rcu_preempt") == 0);
2707}
2708
2709/*
2710 * Handle CPU online/offline notification events.
2711 */
2712static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2713                                    unsigned long action, void *hcpu)
2714{
2715        long cpu = (long)hcpu;
2716        struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2717        struct rcu_node *rnp = rdp->mynode;
2718        struct rcu_state *rsp;
2719
2720        trace_rcu_utilization("Start CPU hotplug");
2721        switch (action) {
2722        case CPU_UP_PREPARE:
2723        case CPU_UP_PREPARE_FROZEN:
2724                rcu_prepare_cpu(cpu);
2725                rcu_prepare_kthreads(cpu);
2726                break;
2727        case CPU_ONLINE:
2728        case CPU_DOWN_FAILED:
2729                rcu_boost_kthread_setaffinity(rnp, -1);
2730                break;
2731        case CPU_DOWN_PREPARE:
2732                rcu_boost_kthread_setaffinity(rnp, cpu);
2733                break;
2734        case CPU_DYING:
2735        case CPU_DYING_FROZEN:
2736                /*
2737                 * The whole machine is "stopped" except this CPU, so we can
2738                 * touch any data without introducing corruption. We send the
2739                 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2740                 */
2741                for_each_rcu_flavor(rsp)
2742                        rcu_cleanup_dying_cpu(rsp);
2743                rcu_cleanup_after_idle(cpu);
2744                break;
2745        case CPU_DEAD:
2746        case CPU_DEAD_FROZEN:
2747        case CPU_UP_CANCELED:
2748        case CPU_UP_CANCELED_FROZEN:
2749                for_each_rcu_flavor(rsp)
2750                        rcu_cleanup_dead_cpu(cpu, rsp);
2751                break;
2752        default:
2753                break;
2754        }
2755        trace_rcu_utilization("End CPU hotplug");
2756        return NOTIFY_OK;
2757}
2758
2759/*
2760 * Spawn the kthread that handles this RCU flavor's grace periods.
2761 */
2762static int __init rcu_spawn_gp_kthread(void)
2763{
2764        unsigned long flags;
2765        struct rcu_node *rnp;
2766        struct rcu_state *rsp;
2767        struct task_struct *t;
2768
2769        for_each_rcu_flavor(rsp) {
2770                t = kthread_run(rcu_gp_kthread, rsp, rsp->name);
2771                BUG_ON(IS_ERR(t));
2772                rnp = rcu_get_root(rsp);
2773                raw_spin_lock_irqsave(&rnp->lock, flags);
2774                rsp->gp_kthread = t;
2775                raw_spin_unlock_irqrestore(&rnp->lock, flags);
2776        }
2777        return 0;
2778}
2779early_initcall(rcu_spawn_gp_kthread);
2780
2781/*
2782 * This function is invoked towards the end of the scheduler's initialization
2783 * process.  Before this is called, the idle task might contain
2784 * RCU read-side critical sections (during which time, this idle
2785 * task is booting the system).  After this function is called, the
2786 * idle tasks are prohibited from containing RCU read-side critical
2787 * sections.  This function also enables RCU lockdep checking.
2788 */
2789void rcu_scheduler_starting(void)
2790{
2791        WARN_ON(num_online_cpus() != 1);
2792        WARN_ON(nr_context_switches() > 0);
2793        rcu_scheduler_active = 1;
2794}
2795
2796/*
2797 * Compute the per-level fanout, either using the exact fanout specified
2798 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2799 */
2800#ifdef CONFIG_RCU_FANOUT_EXACT
2801static void __init rcu_init_levelspread(struct rcu_state *rsp)
2802{
2803        int i;
2804
2805        for (i = rcu_num_lvls - 1; i > 0; i--)
2806                rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2807        rsp->levelspread[0] = rcu_fanout_leaf;
2808}
2809#else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2810static void __init rcu_init_levelspread(struct rcu_state *rsp)
2811{
2812        int ccur;
2813        int cprv;
2814        int i;
2815
2816        cprv = nr_cpu_ids;
2817        for (i = rcu_num_lvls - 1; i >= 0; i--) {
2818                ccur = rsp->levelcnt[i];
2819                rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2820                cprv = ccur;
2821        }
2822}
2823#endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2824
2825/*
2826 * Helper function for rcu_init() that initializes one rcu_state structure.
2827 */
2828static void __init rcu_init_one(struct rcu_state *rsp,
2829                struct rcu_data __percpu *rda)
2830{
2831        static char *buf[] = { "rcu_node_0",
2832                               "rcu_node_1",
2833                               "rcu_node_2",
2834                               "rcu_node_3" };  /* Match MAX_RCU_LVLS */
2835        static char *fqs[] = { "rcu_node_fqs_0",
2836                               "rcu_node_fqs_1",
2837                               "rcu_node_fqs_2",
2838                               "rcu_node_fqs_3" };  /* Match MAX_RCU_LVLS */
2839        int cpustride = 1;
2840        int i;
2841        int j;
2842        struct rcu_node *rnp;
2843
2844        BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf));  /* Fix buf[] init! */
2845
2846        /* Initialize the level-tracking arrays. */
2847
2848        for (i = 0; i < rcu_num_lvls; i++)
2849                rsp->levelcnt[i] = num_rcu_lvl[i];
2850        for (i = 1; i < rcu_num_lvls; i++)
2851                rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2852        rcu_init_levelspread(rsp);
2853
2854        /* Initialize the elements themselves, starting from the leaves. */
2855
2856        for (i = rcu_num_lvls - 1; i >= 0; i--) {
2857                cpustride *= rsp->levelspread[i];
2858                rnp = rsp->level[i];
2859                for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2860                        raw_spin_lock_init(&rnp->lock);
2861                        lockdep_set_class_and_name(&rnp->lock,
2862                                                   &rcu_node_class[i], buf[i]);
2863                        raw_spin_lock_init(&rnp->fqslock);
2864                        lockdep_set_class_and_name(&rnp->fqslock,
2865                                                   &rcu_fqs_class[i], fqs[i]);
2866                        rnp->gpnum = rsp->gpnum;
2867                        rnp->completed = rsp->completed;
2868                        rnp->qsmask = 0;
2869                        rnp->qsmaskinit = 0;
2870                        rnp->grplo = j * cpustride;
2871                        rnp->grphi = (j + 1) * cpustride - 1;
2872                        if (rnp->grphi >= NR_CPUS)
2873                                rnp->grphi = NR_CPUS - 1;
2874                        if (i == 0) {
2875                                rnp->grpnum = 0;
2876                                rnp->grpmask = 0;
2877                                rnp->parent = NULL;
2878                        } else {
2879                                rnp->grpnum = j % rsp->levelspread[i - 1];
2880                                rnp->grpmask = 1UL << rnp->grpnum;
2881                                rnp->parent = rsp->level[i - 1] +
2882                                              j / rsp->levelspread[i - 1];
2883                        }
2884                        rnp->level = i;
2885                        INIT_LIST_HEAD(&rnp->blkd_tasks);
2886                }
2887        }
2888
2889        rsp->rda = rda;
2890        init_waitqueue_head(&rsp->gp_wq);
2891        rnp = rsp->level[rcu_num_lvls - 1];
2892        for_each_possible_cpu(i) {
2893                while (i > rnp->grphi)
2894                        rnp++;
2895                per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2896                rcu_boot_init_percpu_data(i, rsp);
2897        }
2898        list_add(&rsp->flavors, &rcu_struct_flavors);
2899}
2900
2901/*
2902 * Compute the rcu_node tree geometry from kernel parameters.  This cannot
2903 * replace the definitions in rcutree.h because those are needed to size
2904 * the ->node array in the rcu_state structure.
2905 */
2906static void __init rcu_init_geometry(void)
2907{
2908        int i;
2909        int j;
2910        int n = nr_cpu_ids;
2911        int rcu_capacity[MAX_RCU_LVLS + 1];
2912
2913        /* If the compile-time values are accurate, just leave. */
2914        if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
2915            nr_cpu_ids == NR_CPUS)
2916                return;
2917
2918        /*
2919         * Compute number of nodes that can be handled an rcu_node tree
2920         * with the given number of levels.  Setting rcu_capacity[0] makes
2921         * some of the arithmetic easier.
2922         */
2923        rcu_capacity[0] = 1;
2924        rcu_capacity[1] = rcu_fanout_leaf;
2925        for (i = 2; i <= MAX_RCU_LVLS; i++)
2926                rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
2927
2928        /*
2929         * The boot-time rcu_fanout_leaf parameter is only permitted
2930         * to increase the leaf-level fanout, not decrease it.  Of course,
2931         * the leaf-level fanout cannot exceed the number of bits in
2932         * the rcu_node masks.  Finally, the tree must be able to accommodate
2933         * the configured number of CPUs.  Complain and fall back to the
2934         * compile-time values if these limits are exceeded.
2935         */
2936        if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
2937            rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
2938            n > rcu_capacity[MAX_RCU_LVLS]) {
2939                WARN_ON(1);
2940                return;
2941        }
2942
2943        /* Calculate the number of rcu_nodes at each level of the tree. */
2944        for (i = 1; i <= MAX_RCU_LVLS; i++)
2945                if (n <= rcu_capacity[i]) {
2946                        for (j = 0; j <= i; j++)
2947                                num_rcu_lvl[j] =
2948                                        DIV_ROUND_UP(n, rcu_capacity[i - j]);
2949                        rcu_num_lvls = i;
2950                        for (j = i + 1; j <= MAX_RCU_LVLS; j++)
2951                                num_rcu_lvl[j] = 0;
2952                        break;
2953                }
2954
2955        /* Calculate the total number of rcu_node structures. */
2956        rcu_num_nodes = 0;
2957        for (i = 0; i <= MAX_RCU_LVLS; i++)
2958                rcu_num_nodes += num_rcu_lvl[i];
2959        rcu_num_nodes -= n;
2960}
2961
2962void __init rcu_init(void)
2963{
2964        int cpu;
2965
2966        rcu_bootup_announce();
2967        rcu_init_geometry();
2968        rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2969        rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2970        __rcu_init_preempt();
2971         open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2972
2973        /*
2974         * We don't need protection against CPU-hotplug here because
2975         * this is called early in boot, before either interrupts
2976         * or the scheduler are operational.
2977         */
2978        cpu_notifier(rcu_cpu_notify, 0);
2979        for_each_online_cpu(cpu)
2980                rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2981        check_cpu_stall_init();
2982}
2983
2984#include "rcutree_plugin.h"
2985
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.