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 = 0UL - 300UL, \
  72        .completed = 0UL - 300UL, \
  73        .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
  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};
 211
 212static long blimit = 10;        /* Maximum callbacks per rcu_do_batch. */
 213static long qhimark = 10000;    /* If this many pending, ignore blimit. */
 214static long qlowmark = 100;     /* Once only this many pending, use blimit. */
 215
 216module_param(blimit, long, 0444);
 217module_param(qhimark, long, 0444);
 218module_param(qlowmark, long, 0444);
 219
 220int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
 221int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
 222
 223module_param(rcu_cpu_stall_suppress, int, 0644);
 224module_param(rcu_cpu_stall_timeout, int, 0644);
 225
 226static ulong jiffies_till_first_fqs = RCU_JIFFIES_TILL_FORCE_QS;
 227static ulong jiffies_till_next_fqs = RCU_JIFFIES_TILL_FORCE_QS;
 228
 229module_param(jiffies_till_first_fqs, ulong, 0644);
 230module_param(jiffies_till_next_fqs, ulong, 0644);
 231
 232static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *));
 233static void force_quiescent_state(struct rcu_state *rsp);
 234static int rcu_pending(int cpu);
 235
 236/*
 237 * Return the number of RCU-sched batches processed thus far for debug & stats.
 238 */
 239long rcu_batches_completed_sched(void)
 240{
 241        return rcu_sched_state.completed;
 242}
 243EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
 244
 245/*
 246 * Return the number of RCU BH batches processed thus far for debug & stats.
 247 */
 248long rcu_batches_completed_bh(void)
 249{
 250        return rcu_bh_state.completed;
 251}
 252EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
 253
 254/*
 255 * Force a quiescent state for RCU BH.
 256 */
 257void rcu_bh_force_quiescent_state(void)
 258{
 259        force_quiescent_state(&rcu_bh_state);
 260}
 261EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
 262
 263/*
 264 * Record the number of times rcutorture tests have been initiated and
 265 * terminated.  This information allows the debugfs tracing stats to be
 266 * correlated to the rcutorture messages, even when the rcutorture module
 267 * is being repeatedly loaded and unloaded.  In other words, we cannot
 268 * store this state in rcutorture itself.
 269 */
 270void rcutorture_record_test_transition(void)
 271{
 272        rcutorture_testseq++;
 273        rcutorture_vernum = 0;
 274}
 275EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
 276
 277/*
 278 * Record the number of writer passes through the current rcutorture test.
 279 * This is also used to correlate debugfs tracing stats with the rcutorture
 280 * messages.
 281 */
 282void rcutorture_record_progress(unsigned long vernum)
 283{
 284        rcutorture_vernum++;
 285}
 286EXPORT_SYMBOL_GPL(rcutorture_record_progress);
 287
 288/*
 289 * Force a quiescent state for RCU-sched.
 290 */
 291void rcu_sched_force_quiescent_state(void)
 292{
 293        force_quiescent_state(&rcu_sched_state);
 294}
 295EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
 296
 297/*
 298 * Does the CPU have callbacks ready to be invoked?
 299 */
 300static int
 301cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
 302{
 303        return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
 304               rdp->nxttail[RCU_DONE_TAIL] != NULL;
 305}
 306
 307/*
 308 * Does the current CPU require a yet-as-unscheduled grace period?
 309 */
 310static int
 311cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
 312{
 313        struct rcu_head **ntp;
 314
 315        ntp = rdp->nxttail[RCU_DONE_TAIL +
 316                           (ACCESS_ONCE(rsp->completed) != rdp->completed)];
 317        return rdp->nxttail[RCU_DONE_TAIL] && ntp && *ntp &&
 318               !rcu_gp_in_progress(rsp);
 319}
 320
 321/*
 322 * Return the root node of the specified rcu_state structure.
 323 */
 324static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
 325{
 326        return &rsp->node[0];
 327}
 328
 329/*
 330 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
 331 *
 332 * If the new value of the ->dynticks_nesting counter now is zero,
 333 * we really have entered idle, and must do the appropriate accounting.
 334 * The caller must have disabled interrupts.
 335 */
 336static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
 337                                bool user)
 338{
 339        trace_rcu_dyntick("Start", oldval, 0);
 340        if (!user && !is_idle_task(current)) {
 341                struct task_struct *idle = idle_task(smp_processor_id());
 342
 343                trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
 344                ftrace_dump(DUMP_ORIG);
 345                WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
 346                          current->pid, current->comm,
 347                          idle->pid, idle->comm); /* must be idle task! */
 348        }
 349        rcu_prepare_for_idle(smp_processor_id());
 350        /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
 351        smp_mb__before_atomic_inc();  /* See above. */
 352        atomic_inc(&rdtp->dynticks);
 353        smp_mb__after_atomic_inc();  /* Force ordering with next sojourn. */
 354        WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
 355
 356        /*
 357         * It is illegal to enter an extended quiescent state while
 358         * in an RCU read-side critical section.
 359         */
 360        rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
 361                           "Illegal idle entry in RCU read-side critical section.");
 362        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
 363                           "Illegal idle entry in RCU-bh read-side critical section.");
 364        rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
 365                           "Illegal idle entry in RCU-sched read-side critical section.");
 366}
 367
 368/*
 369 * Enter an RCU extended quiescent state, which can be either the
 370 * idle loop or adaptive-tickless usermode execution.
 371 */
 372static void rcu_eqs_enter(bool user)
 373{
 374        long long oldval;
 375        struct rcu_dynticks *rdtp;
 376
 377        rdtp = &__get_cpu_var(rcu_dynticks);
 378        oldval = rdtp->dynticks_nesting;
 379        WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
 380        if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
 381                rdtp->dynticks_nesting = 0;
 382        else
 383                rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
 384        rcu_eqs_enter_common(rdtp, oldval, user);
 385}
 386
 387/**
 388 * rcu_idle_enter - inform RCU that current CPU is entering idle
 389 *
 390 * Enter idle mode, in other words, -leave- the mode in which RCU
 391 * read-side critical sections can occur.  (Though RCU read-side
 392 * critical sections can occur in irq handlers in idle, a possibility
 393 * handled by irq_enter() and irq_exit().)
 394 *
 395 * We crowbar the ->dynticks_nesting field to zero to allow for
 396 * the possibility of usermode upcalls having messed up our count
 397 * of interrupt nesting level during the prior busy period.
 398 */
 399void rcu_idle_enter(void)
 400{
 401        unsigned long flags;
 402
 403        local_irq_save(flags);
 404        rcu_eqs_enter(false);
 405        local_irq_restore(flags);
 406}
 407EXPORT_SYMBOL_GPL(rcu_idle_enter);
 408
 409#ifdef CONFIG_RCU_USER_QS
 410/**
 411 * rcu_user_enter - inform RCU that we are resuming userspace.
 412 *
 413 * Enter RCU idle mode right before resuming userspace.  No use of RCU
 414 * is permitted between this call and rcu_user_exit(). This way the
 415 * CPU doesn't need to maintain the tick for RCU maintenance purposes
 416 * when the CPU runs in userspace.
 417 */
 418void rcu_user_enter(void)
 419{
 420        rcu_eqs_enter(1);
 421}
 422
 423/**
 424 * rcu_user_enter_after_irq - inform RCU that we are going to resume userspace
 425 * after the current irq returns.
 426 *
 427 * This is similar to rcu_user_enter() but in the context of a non-nesting
 428 * irq. After this call, RCU enters into idle mode when the interrupt
 429 * returns.
 430 */
 431void rcu_user_enter_after_irq(void)
 432{
 433        unsigned long flags;
 434        struct rcu_dynticks *rdtp;
 435
 436        local_irq_save(flags);
 437        rdtp = &__get_cpu_var(rcu_dynticks);
 438        /* Ensure this irq is interrupting a non-idle RCU state.  */
 439        WARN_ON_ONCE(!(rdtp->dynticks_nesting & DYNTICK_TASK_MASK));
 440        rdtp->dynticks_nesting = 1;
 441        local_irq_restore(flags);
 442}
 443#endif /* CONFIG_RCU_USER_QS */
 444
 445/**
 446 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
 447 *
 448 * Exit from an interrupt handler, which might possibly result in entering
 449 * idle mode, in other words, leaving the mode in which read-side critical
 450 * sections can occur.
 451 *
 452 * This code assumes that the idle loop never does anything that might
 453 * result in unbalanced calls to irq_enter() and irq_exit().  If your
 454 * architecture violates this assumption, RCU will give you what you
 455 * deserve, good and hard.  But very infrequently and irreproducibly.
 456 *
 457 * Use things like work queues to work around this limitation.
 458 *
 459 * You have been warned.
 460 */
 461void rcu_irq_exit(void)
 462{
 463        unsigned long flags;
 464        long long oldval;
 465        struct rcu_dynticks *rdtp;
 466
 467        local_irq_save(flags);
 468        rdtp = &__get_cpu_var(rcu_dynticks);
 469        oldval = rdtp->dynticks_nesting;
 470        rdtp->dynticks_nesting--;
 471        WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
 472        if (rdtp->dynticks_nesting)
 473                trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
 474        else
 475                rcu_eqs_enter_common(rdtp, oldval, true);
 476        local_irq_restore(flags);
 477}
 478
 479/*
 480 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
 481 *
 482 * If the new value of the ->dynticks_nesting counter was previously zero,
 483 * we really have exited idle, and must do the appropriate accounting.
 484 * The caller must have disabled interrupts.
 485 */
 486static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
 487                               int user)
 488{
 489        smp_mb__before_atomic_inc();  /* Force ordering w/previous sojourn. */
 490        atomic_inc(&rdtp->dynticks);
 491        /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
 492        smp_mb__after_atomic_inc();  /* See above. */
 493        WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
 494        rcu_cleanup_after_idle(smp_processor_id());
 495        trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
 496        if (!user && !is_idle_task(current)) {
 497                struct task_struct *idle = idle_task(smp_processor_id());
 498
 499                trace_rcu_dyntick("Error on exit: not idle task",
 500                                  oldval, rdtp->dynticks_nesting);
 501                ftrace_dump(DUMP_ORIG);
 502                WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
 503                          current->pid, current->comm,
 504                          idle->pid, idle->comm); /* must be idle task! */
 505        }
 506}
 507
 508/*
 509 * Exit an RCU extended quiescent state, which can be either the
 510 * idle loop or adaptive-tickless usermode execution.
 511 */
 512static void rcu_eqs_exit(bool user)
 513{
 514        struct rcu_dynticks *rdtp;
 515        long long oldval;
 516
 517        rdtp = &__get_cpu_var(rcu_dynticks);
 518        oldval = rdtp->dynticks_nesting;
 519        WARN_ON_ONCE(oldval < 0);
 520        if (oldval & DYNTICK_TASK_NEST_MASK)
 521                rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
 522        else
 523                rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
 524        rcu_eqs_exit_common(rdtp, oldval, user);
 525}
 526
 527/**
 528 * rcu_idle_exit - inform RCU that current CPU is leaving idle
 529 *
 530 * Exit idle mode, in other words, -enter- the mode in which RCU
 531 * read-side critical sections can occur.
 532 *
 533 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
 534 * allow for the possibility of usermode upcalls messing up our count
 535 * of interrupt nesting level during the busy period that is just
 536 * now starting.
 537 */
 538void rcu_idle_exit(void)
 539{
 540        unsigned long flags;
 541
 542        local_irq_save(flags);
 543        rcu_eqs_exit(false);
 544        local_irq_restore(flags);
 545}
 546EXPORT_SYMBOL_GPL(rcu_idle_exit);
 547
 548#ifdef CONFIG_RCU_USER_QS
 549/**
 550 * rcu_user_exit - inform RCU that we are exiting userspace.
 551 *
 552 * Exit RCU idle mode while entering the kernel because it can
 553 * run a RCU read side critical section anytime.
 554 */
 555void rcu_user_exit(void)
 556{
 557        rcu_eqs_exit(1);
 558}
 559
 560/**
 561 * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace
 562 * idle mode after the current non-nesting irq returns.
 563 *
 564 * This is similar to rcu_user_exit() but in the context of an irq.
 565 * This is called when the irq has interrupted a userspace RCU idle mode
 566 * context. When the current non-nesting interrupt returns after this call,
 567 * the CPU won't restore the RCU idle mode.
 568 */
 569void rcu_user_exit_after_irq(void)
 570{
 571        unsigned long flags;
 572        struct rcu_dynticks *rdtp;
 573
 574        local_irq_save(flags);
 575        rdtp = &__get_cpu_var(rcu_dynticks);
 576        /* Ensure we are interrupting an RCU idle mode. */
 577        WARN_ON_ONCE(rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK);
 578        rdtp->dynticks_nesting += DYNTICK_TASK_EXIT_IDLE;
 579        local_irq_restore(flags);
 580}
 581#endif /* CONFIG_RCU_USER_QS */
 582
 583/**
 584 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
 585 *
 586 * Enter an interrupt handler, which might possibly result in exiting
 587 * idle mode, in other words, entering the mode in which read-side critical
 588 * sections can occur.
 589 *
 590 * Note that the Linux kernel is fully capable of entering an interrupt
 591 * handler that it never exits, for example when doing upcalls to
 592 * user mode!  This code assumes that the idle loop never does upcalls to
 593 * user mode.  If your architecture does do upcalls from the idle loop (or
 594 * does anything else that results in unbalanced calls to the irq_enter()
 595 * and irq_exit() functions), RCU will give you what you deserve, good
 596 * and hard.  But very infrequently and irreproducibly.
 597 *
 598 * Use things like work queues to work around this limitation.
 599 *
 600 * You have been warned.
 601 */
 602void rcu_irq_enter(void)
 603{
 604        unsigned long flags;
 605        struct rcu_dynticks *rdtp;
 606        long long oldval;
 607
 608        local_irq_save(flags);
 609        rdtp = &__get_cpu_var(rcu_dynticks);
 610        oldval = rdtp->dynticks_nesting;
 611        rdtp->dynticks_nesting++;
 612        WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
 613        if (oldval)
 614                trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
 615        else
 616                rcu_eqs_exit_common(rdtp, oldval, true);
 617        local_irq_restore(flags);
 618}
 619
 620/**
 621 * rcu_nmi_enter - inform RCU of entry to NMI context
 622 *
 623 * If the CPU was idle with dynamic ticks active, and there is no
 624 * irq handler running, this updates rdtp->dynticks_nmi to let the
 625 * RCU grace-period handling know that the CPU is active.
 626 */
 627void rcu_nmi_enter(void)
 628{
 629        struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
 630
 631        if (rdtp->dynticks_nmi_nesting == 0 &&
 632            (atomic_read(&rdtp->dynticks) & 0x1))
 633                return;
 634        rdtp->dynticks_nmi_nesting++;
 635        smp_mb__before_atomic_inc();  /* Force delay from prior write. */
 636        atomic_inc(&rdtp->dynticks);
 637        /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
 638        smp_mb__after_atomic_inc();  /* See above. */
 639        WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
 640}
 641
 642/**
 643 * rcu_nmi_exit - inform RCU of exit from NMI context
 644 *
 645 * If the CPU was idle with dynamic ticks active, and there is no
 646 * irq handler running, this updates rdtp->dynticks_nmi to let the
 647 * RCU grace-period handling know that the CPU is no longer active.
 648 */
 649void rcu_nmi_exit(void)
 650{
 651        struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
 652
 653        if (rdtp->dynticks_nmi_nesting == 0 ||
 654            --rdtp->dynticks_nmi_nesting != 0)
 655                return;
 656        /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
 657        smp_mb__before_atomic_inc();  /* See above. */
 658        atomic_inc(&rdtp->dynticks);
 659        smp_mb__after_atomic_inc();  /* Force delay to next write. */
 660        WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
 661}
 662
 663/**
 664 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
 665 *
 666 * If the current CPU is in its idle loop and is neither in an interrupt
 667 * or NMI handler, return true.
 668 */
 669int rcu_is_cpu_idle(void)
 670{
 671        int ret;
 672
 673        preempt_disable();
 674        ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
 675        preempt_enable();
 676        return ret;
 677}
 678EXPORT_SYMBOL(rcu_is_cpu_idle);
 679
 680#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
 681
 682/*
 683 * Is the current CPU online?  Disable preemption to avoid false positives
 684 * that could otherwise happen due to the current CPU number being sampled,
 685 * this task being preempted, its old CPU being taken offline, resuming
 686 * on some other CPU, then determining that its old CPU is now offline.
 687 * It is OK to use RCU on an offline processor during initial boot, hence
 688 * the check for rcu_scheduler_fully_active.  Note also that it is OK
 689 * for a CPU coming online to use RCU for one jiffy prior to marking itself
 690 * online in the cpu_online_mask.  Similarly, it is OK for a CPU going
 691 * offline to continue to use RCU for one jiffy after marking itself
 692 * offline in the cpu_online_mask.  This leniency is necessary given the
 693 * non-atomic nature of the online and offline processing, for example,
 694 * the fact that a CPU enters the scheduler after completing the CPU_DYING
 695 * notifiers.
 696 *
 697 * This is also why RCU internally marks CPUs online during the
 698 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
 699 *
 700 * Disable checking if in an NMI handler because we cannot safely report
 701 * errors from NMI handlers anyway.
 702 */
 703bool rcu_lockdep_current_cpu_online(void)
 704{
 705        struct rcu_data *rdp;
 706        struct rcu_node *rnp;
 707        bool ret;
 708
 709        if (in_nmi())
 710                return 1;
 711        preempt_disable();
 712        rdp = &__get_cpu_var(rcu_sched_data);
 713        rnp = rdp->mynode;
 714        ret = (rdp->grpmask & rnp->qsmaskinit) ||
 715              !rcu_scheduler_fully_active;
 716        preempt_enable();
 717        return ret;
 718}
 719EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
 720
 721#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
 722
 723/**
 724 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
 725 *
 726 * If the current CPU is idle or running at a first-level (not nested)
 727 * interrupt from idle, return true.  The caller must have at least
 728 * disabled preemption.
 729 */
 730int rcu_is_cpu_rrupt_from_idle(void)
 731{
 732        return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
 733}
 734
 735/*
 736 * Snapshot the specified CPU's dynticks counter so that we can later
 737 * credit them with an implicit quiescent state.  Return 1 if this CPU
 738 * is in dynticks idle mode, which is an extended quiescent state.
 739 */
 740static int dyntick_save_progress_counter(struct rcu_data *rdp)
 741{
 742        rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
 743        return (rdp->dynticks_snap & 0x1) == 0;
 744}
 745
 746/*
 747 * Return true if the specified CPU has passed through a quiescent
 748 * state by virtue of being in or having passed through an dynticks
 749 * idle state since the last call to dyntick_save_progress_counter()
 750 * for this same CPU, or by virtue of having been offline.
 751 */
 752static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
 753{
 754        unsigned int curr;
 755        unsigned int snap;
 756
 757        curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
 758        snap = (unsigned int)rdp->dynticks_snap;
 759
 760        /*
 761         * If the CPU passed through or entered a dynticks idle phase with
 762         * no active irq/NMI handlers, then we can safely pretend that the CPU
 763         * already acknowledged the request to pass through a quiescent
 764         * state.  Either way, that CPU cannot possibly be in an RCU
 765         * read-side critical section that started before the beginning
 766         * of the current RCU grace period.
 767         */
 768        if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
 769                trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
 770                rdp->dynticks_fqs++;
 771                return 1;
 772        }
 773
 774        /*
 775         * Check for the CPU being offline, but only if the grace period
 776         * is old enough.  We don't need to worry about the CPU changing
 777         * state: If we see it offline even once, it has been through a
 778         * quiescent state.
 779         *
 780         * The reason for insisting that the grace period be at least
 781         * one jiffy old is that CPUs that are not quite online and that
 782         * have just gone offline can still execute RCU read-side critical
 783         * sections.
 784         */
 785        if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
 786                return 0;  /* Grace period is not old enough. */
 787        barrier();
 788        if (cpu_is_offline(rdp->cpu)) {
 789                trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
 790                rdp->offline_fqs++;
 791                return 1;
 792        }
 793        return 0;
 794}
 795
 796static int jiffies_till_stall_check(void)
 797{
 798        int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
 799
 800        /*
 801         * Limit check must be consistent with the Kconfig limits
 802         * for CONFIG_RCU_CPU_STALL_TIMEOUT.
 803         */
 804        if (till_stall_check < 3) {
 805                ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
 806                till_stall_check = 3;
 807        } else if (till_stall_check > 300) {
 808                ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
 809                till_stall_check = 300;
 810        }
 811        return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
 812}
 813
 814static void record_gp_stall_check_time(struct rcu_state *rsp)
 815{
 816        rsp->gp_start = jiffies;
 817        rsp->jiffies_stall = jiffies + jiffies_till_stall_check();
 818}
 819
 820/*
 821 * Dump stacks of all tasks running on stalled CPUs.  This is a fallback
 822 * for architectures that do not implement trigger_all_cpu_backtrace().
 823 * The NMI-triggered stack traces are more accurate because they are
 824 * printed by the target CPU.
 825 */
 826static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
 827{
 828        int cpu;
 829        unsigned long flags;
 830        struct rcu_node *rnp;
 831
 832        rcu_for_each_leaf_node(rsp, rnp) {
 833                raw_spin_lock_irqsave(&rnp->lock, flags);
 834                if (rnp->qsmask != 0) {
 835                        for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
 836                                if (rnp->qsmask & (1UL << cpu))
 837                                        dump_cpu_task(rnp->grplo + cpu);
 838                }
 839                raw_spin_unlock_irqrestore(&rnp->lock, flags);
 840        }
 841}
 842
 843static void print_other_cpu_stall(struct rcu_state *rsp)
 844{
 845        int cpu;
 846        long delta;
 847        unsigned long flags;
 848        int ndetected = 0;
 849        struct rcu_node *rnp = rcu_get_root(rsp);
 850        long totqlen = 0;
 851
 852        /* Only let one CPU complain about others per time interval. */
 853
 854        raw_spin_lock_irqsave(&rnp->lock, flags);
 855        delta = jiffies - rsp->jiffies_stall;
 856        if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
 857                raw_spin_unlock_irqrestore(&rnp->lock, flags);
 858                return;
 859        }
 860        rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3;
 861        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 862
 863        /*
 864         * OK, time to rat on our buddy...
 865         * See Documentation/RCU/stallwarn.txt for info on how to debug
 866         * RCU CPU stall warnings.
 867         */
 868        printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
 869               rsp->name);
 870        print_cpu_stall_info_begin();
 871        rcu_for_each_leaf_node(rsp, rnp) {
 872                raw_spin_lock_irqsave(&rnp->lock, flags);
 873                ndetected += rcu_print_task_stall(rnp);
 874                if (rnp->qsmask != 0) {
 875                        for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
 876                                if (rnp->qsmask & (1UL << cpu)) {
 877                                        print_cpu_stall_info(rsp,
 878                                                             rnp->grplo + cpu);
 879                                        ndetected++;
 880                                }
 881                }
 882                raw_spin_unlock_irqrestore(&rnp->lock, flags);
 883        }
 884
 885        /*
 886         * Now rat on any tasks that got kicked up to the root rcu_node
 887         * due to CPU offlining.
 888         */
 889        rnp = rcu_get_root(rsp);
 890        raw_spin_lock_irqsave(&rnp->lock, flags);
 891        ndetected += rcu_print_task_stall(rnp);
 892        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 893
 894        print_cpu_stall_info_end();
 895        for_each_possible_cpu(cpu)
 896                totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
 897        pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
 898               smp_processor_id(), (long)(jiffies - rsp->gp_start),
 899               rsp->gpnum, rsp->completed, totqlen);
 900        if (ndetected == 0)
 901                printk(KERN_ERR "INFO: Stall ended before state dump start\n");
 902        else if (!trigger_all_cpu_backtrace())
 903                rcu_dump_cpu_stacks(rsp);
 904
 905        /* Complain about tasks blocking the grace period. */
 906
 907        rcu_print_detail_task_stall(rsp);
 908
 909        force_quiescent_state(rsp);  /* Kick them all. */
 910}
 911
 912static void print_cpu_stall(struct rcu_state *rsp)
 913{
 914        int cpu;
 915        unsigned long flags;
 916        struct rcu_node *rnp = rcu_get_root(rsp);
 917        long totqlen = 0;
 918
 919        /*
 920         * OK, time to rat on ourselves...
 921         * See Documentation/RCU/stallwarn.txt for info on how to debug
 922         * RCU CPU stall warnings.
 923         */
 924        printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
 925        print_cpu_stall_info_begin();
 926        print_cpu_stall_info(rsp, smp_processor_id());
 927        print_cpu_stall_info_end();
 928        for_each_possible_cpu(cpu)
 929                totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
 930        pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
 931                jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
 932        if (!trigger_all_cpu_backtrace())
 933                dump_stack();
 934
 935        raw_spin_lock_irqsave(&rnp->lock, flags);
 936        if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
 937                rsp->jiffies_stall = jiffies +
 938                                     3 * jiffies_till_stall_check() + 3;
 939        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 940
 941        set_need_resched();  /* kick ourselves to get things going. */
 942}
 943
 944static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
 945{
 946        unsigned long j;
 947        unsigned long js;
 948        struct rcu_node *rnp;
 949
 950        if (rcu_cpu_stall_suppress)
 951                return;
 952        j = ACCESS_ONCE(jiffies);
 953        js = ACCESS_ONCE(rsp->jiffies_stall);
 954        rnp = rdp->mynode;
 955        if (rcu_gp_in_progress(rsp) &&
 956            (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
 957
 958                /* We haven't checked in, so go dump stack. */
 959                print_cpu_stall(rsp);
 960
 961        } else if (rcu_gp_in_progress(rsp) &&
 962                   ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
 963
 964                /* They had a few time units to dump stack, so complain. */
 965                print_other_cpu_stall(rsp);
 966        }
 967}
 968
 969static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
 970{
 971        rcu_cpu_stall_suppress = 1;
 972        return NOTIFY_DONE;
 973}
 974
 975/**
 976 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
 977 *
 978 * Set the stall-warning timeout way off into the future, thus preventing
 979 * any RCU CPU stall-warning messages from appearing in the current set of
 980 * RCU grace periods.
 981 *
 982 * The caller must disable hard irqs.
 983 */
 984void rcu_cpu_stall_reset(void)
 985{
 986        struct rcu_state *rsp;
 987
 988        for_each_rcu_flavor(rsp)
 989                rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
 990}
 991
 992static struct notifier_block rcu_panic_block = {
 993        .notifier_call = rcu_panic,
 994};
 995
 996static void __init check_cpu_stall_init(void)
 997{
 998        atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
 999}
1000
1001/*
1002 * Update CPU-local rcu_data state to record the newly noticed grace period.
1003 * This is used both when we started the grace period and when we notice
1004 * that someone else started the grace period.  The caller must hold the
1005 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
1006 *  and must have irqs disabled.
1007 */
1008static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1009{
1010        if (rdp->gpnum != rnp->gpnum) {
1011                /*
1012                 * If the current grace period is waiting for this CPU,
1013                 * set up to detect a quiescent state, otherwise don't
1014                 * go looking for one.
1015                 */
1016                rdp->gpnum = rnp->gpnum;
1017                trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
1018                rdp->passed_quiesce = 0;
1019                rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1020                zero_cpu_stall_ticks(rdp);
1021        }
1022}
1023
1024static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
1025{
1026        unsigned long flags;
1027        struct rcu_node *rnp;
1028
1029        local_irq_save(flags);
1030        rnp = rdp->mynode;
1031        if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
1032            !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1033                local_irq_restore(flags);
1034                return;
1035        }
1036        __note_new_gpnum(rsp, rnp, rdp);
1037        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1038}
1039
1040/*
1041 * Did someone else start a new RCU grace period start since we last
1042 * checked?  Update local state appropriately if so.  Must be called
1043 * on the CPU corresponding to rdp.
1044 */
1045static int
1046check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
1047{
1048        unsigned long flags;
1049        int ret = 0;
1050
1051        local_irq_save(flags);
1052        if (rdp->gpnum != rsp->gpnum) {
1053                note_new_gpnum(rsp, rdp);
1054                ret = 1;
1055        }
1056        local_irq_restore(flags);
1057        return ret;
1058}
1059
1060/*
1061 * Initialize the specified rcu_data structure's callback list to empty.
1062 */
1063static void init_callback_list(struct rcu_data *rdp)
1064{
1065        int i;
1066
1067        rdp->nxtlist = NULL;
1068        for (i = 0; i < RCU_NEXT_SIZE; i++)
1069                rdp->nxttail[i] = &rdp->nxtlist;
1070        init_nocb_callback_list(rdp);
1071}
1072
1073/*
1074 * Advance this CPU's callbacks, but only if the current grace period
1075 * has ended.  This may be called only from the CPU to whom the rdp
1076 * belongs.  In addition, the corresponding leaf rcu_node structure's
1077 * ->lock must be held by the caller, with irqs disabled.
1078 */
1079static void
1080__rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1081{
1082        /* Did another grace period end? */
1083        if (rdp->completed != rnp->completed) {
1084
1085                /* Advance callbacks.  No harm if list empty. */
1086                rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
1087                rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
1088                rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1089
1090                /* Remember that we saw this grace-period completion. */
1091                rdp->completed = rnp->completed;
1092                trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
1093
1094                /*
1095                 * If we were in an extended quiescent state, we may have
1096                 * missed some grace periods that others CPUs handled on
1097                 * our behalf. Catch up with this state to avoid noting
1098                 * spurious new grace periods.  If another grace period
1099                 * has started, then rnp->gpnum will have advanced, so
1100                 * we will detect this later on.  Of course, any quiescent
1101                 * states we found for the old GP are now invalid.
1102                 */
1103                if (ULONG_CMP_LT(rdp->gpnum, rdp->completed)) {
1104                        rdp->gpnum = rdp->completed;
1105                        rdp->passed_quiesce = 0;
1106                }
1107
1108                /*
1109                 * If RCU does not need a quiescent state from this CPU,
1110                 * then make sure that this CPU doesn't go looking for one.
1111                 */
1112                if ((rnp->qsmask & rdp->grpmask) == 0)
1113                        rdp->qs_pending = 0;
1114        }
1115}
1116
1117/*
1118 * Advance this CPU's callbacks, but only if the current grace period
1119 * has ended.  This may be called only from the CPU to whom the rdp
1120 * belongs.
1121 */
1122static void
1123rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
1124{
1125        unsigned long flags;
1126        struct rcu_node *rnp;
1127
1128        local_irq_save(flags);
1129        rnp = rdp->mynode;
1130        if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
1131            !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1132                local_irq_restore(flags);
1133                return;
1134        }
1135        __rcu_process_gp_end(rsp, rnp, rdp);
1136        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1137}
1138
1139/*
1140 * Do per-CPU grace-period initialization for running CPU.  The caller
1141 * must hold the lock of the leaf rcu_node structure corresponding to
1142 * this CPU.
1143 */
1144static void
1145rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1146{
1147        /* Prior grace period ended, so advance callbacks for current CPU. */
1148        __rcu_process_gp_end(rsp, rnp, rdp);
1149
1150        /* Set state so that this CPU will detect the next quiescent state. */
1151        __note_new_gpnum(rsp, rnp, rdp);
1152}
1153
1154/*
1155 * Initialize a new grace period.
1156 */
1157static int rcu_gp_init(struct rcu_state *rsp)
1158{
1159        struct rcu_data *rdp;
1160        struct rcu_node *rnp = rcu_get_root(rsp);
1161
1162        raw_spin_lock_irq(&rnp->lock);
1163        rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1164
1165        if (rcu_gp_in_progress(rsp)) {
1166                /* Grace period already in progress, don't start another.  */
1167                raw_spin_unlock_irq(&rnp->lock);
1168                return 0;
1169        }
1170
1171        /* Advance to a new grace period and initialize state. */
1172        rsp->gpnum++;
1173        trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1174        record_gp_stall_check_time(rsp);
1175        raw_spin_unlock_irq(&rnp->lock);
1176
1177        /* Exclude any concurrent CPU-hotplug operations. */
1178        mutex_lock(&rsp->onoff_mutex);
1179
1180        /*
1181         * Set the quiescent-state-needed bits in all the rcu_node
1182         * structures for all currently online CPUs in breadth-first order,
1183         * starting from the root rcu_node structure, relying on the layout
1184         * of the tree within the rsp->node[] array.  Note that other CPUs
1185         * will access only the leaves of the hierarchy, thus seeing that no
1186         * grace period is in progress, at least until the corresponding
1187         * leaf node has been initialized.  In addition, we have excluded
1188         * CPU-hotplug operations.
1189         *
1190         * The grace period cannot complete until the initialization
1191         * process finishes, because this kthread handles both.
1192         */
1193        rcu_for_each_node_breadth_first(rsp, rnp) {
1194                raw_spin_lock_irq(&rnp->lock);
1195                rdp = this_cpu_ptr(rsp->rda);
1196                rcu_preempt_check_blocked_tasks(rnp);
1197                rnp->qsmask = rnp->qsmaskinit;
1198                rnp->gpnum = rsp->gpnum;
1199                WARN_ON_ONCE(rnp->completed != rsp->completed);
1200                rnp->completed = rsp->completed;
1201                if (rnp == rdp->mynode)
1202                        rcu_start_gp_per_cpu(rsp, rnp, rdp);
1203                rcu_preempt_boost_start_gp(rnp);
1204                trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1205                                            rnp->level, rnp->grplo,
1206                                            rnp->grphi, rnp->qsmask);
1207                raw_spin_unlock_irq(&rnp->lock);
1208#ifdef CONFIG_PROVE_RCU_DELAY
1209                if ((random32() % (rcu_num_nodes * 8)) == 0)
1210                        schedule_timeout_uninterruptible(2);
1211#endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1212                cond_resched();
1213        }
1214
1215        mutex_unlock(&rsp->onoff_mutex);
1216        return 1;
1217}
1218
1219/*
1220 * Do one round of quiescent-state forcing.
1221 */
1222int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1223{
1224        int fqs_state = fqs_state_in;
1225        struct rcu_node *rnp = rcu_get_root(rsp);
1226
1227        rsp->n_force_qs++;
1228        if (fqs_state == RCU_SAVE_DYNTICK) {
1229                /* Collect dyntick-idle snapshots. */
1230                force_qs_rnp(rsp, dyntick_save_progress_counter);
1231                fqs_state = RCU_FORCE_QS;
1232        } else {
1233                /* Handle dyntick-idle and offline CPUs. */
1234                force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1235        }
1236        /* Clear flag to prevent immediate re-entry. */
1237        if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1238                raw_spin_lock_irq(&rnp->lock);
1239                rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1240                raw_spin_unlock_irq(&rnp->lock);
1241        }
1242        return fqs_state;
1243}
1244
1245/*
1246 * Clean up after the old grace period.
1247 */
1248static void rcu_gp_cleanup(struct rcu_state *rsp)
1249{
1250        unsigned long gp_duration;
1251        struct rcu_data *rdp;
1252        struct rcu_node *rnp = rcu_get_root(rsp);
1253
1254        raw_spin_lock_irq(&rnp->lock);
1255        gp_duration = jiffies - rsp->gp_start;
1256        if (gp_duration > rsp->gp_max)
1257                rsp->gp_max = gp_duration;
1258
1259        /*
1260         * We know the grace period is complete, but to everyone else
1261         * it appears to still be ongoing.  But it is also the case
1262         * that to everyone else it looks like there is nothing that
1263         * they can do to advance the grace period.  It is therefore
1264         * safe for us to drop the lock in order to mark the grace
1265         * period as completed in all of the rcu_node structures.
1266         */
1267        raw_spin_unlock_irq(&rnp->lock);
1268
1269        /*
1270         * Propagate new ->completed value to rcu_node structures so
1271         * that other CPUs don't have to wait until the start of the next
1272         * grace period to process their callbacks.  This also avoids
1273         * some nasty RCU grace-period initialization races by forcing
1274         * the end of the current grace period to be completely recorded in
1275         * all of the rcu_node structures before the beginning of the next
1276         * grace period is recorded in any of the rcu_node structures.
1277         */
1278        rcu_for_each_node_breadth_first(rsp, rnp) {
1279                raw_spin_lock_irq(&rnp->lock);
1280                rnp->completed = rsp->gpnum;
1281                raw_spin_unlock_irq(&rnp->lock);
1282                cond_resched();
1283        }
1284        rnp = rcu_get_root(rsp);
1285        raw_spin_lock_irq(&rnp->lock);
1286
1287        rsp->completed = rsp->gpnum; /* Declare grace period done. */
1288        trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1289        rsp->fqs_state = RCU_GP_IDLE;
1290        rdp = this_cpu_ptr(rsp->rda);
1291        if (cpu_needs_another_gp(rsp, rdp))
1292                rsp->gp_flags = 1;
1293        raw_spin_unlock_irq(&rnp->lock);
1294}
1295
1296/*
1297 * Body of kthread that handles grace periods.
1298 */
1299static int __noreturn rcu_gp_kthread(void *arg)
1300{
1301        int fqs_state;
1302        unsigned long j;
1303        int ret;
1304        struct rcu_state *rsp = arg;
1305        struct rcu_node *rnp = rcu_get_root(rsp);
1306
1307        for (;;) {
1308
1309                /* Handle grace-period start. */
1310                for (;;) {
1311                        wait_event_interruptible(rsp->gp_wq,
1312                                                 rsp->gp_flags &
1313                                                 RCU_GP_FLAG_INIT);
1314                        if ((rsp->gp_flags & RCU_GP_FLAG_INIT) &&
1315                            rcu_gp_init(rsp))
1316                                break;
1317                        cond_resched();
1318                        flush_signals(current);
1319                }
1320
1321                /* Handle quiescent-state forcing. */
1322                fqs_state = RCU_SAVE_DYNTICK;
1323                j = jiffies_till_first_fqs;
1324                if (j > HZ) {
1325                        j = HZ;
1326                        jiffies_till_first_fqs = HZ;
1327                }
1328                for (;;) {
1329                        rsp->jiffies_force_qs = jiffies + j;
1330                        ret = wait_event_interruptible_timeout(rsp->gp_wq,
1331                                        (rsp->gp_flags & RCU_GP_FLAG_FQS) ||
1332                                        (!ACCESS_ONCE(rnp->qsmask) &&
1333                                         !rcu_preempt_blocked_readers_cgp(rnp)),
1334                                        j);
1335                        /* If grace period done, leave loop. */
1336                        if (!ACCESS_ONCE(rnp->qsmask) &&
1337                            !rcu_preempt_blocked_readers_cgp(rnp))
1338                                break;
1339                        /* If time for quiescent-state forcing, do it. */
1340                        if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) {
1341                                fqs_state = rcu_gp_fqs(rsp, fqs_state);
1342                                cond_resched();
1343                        } else {
1344                                /* Deal with stray signal. */
1345                                cond_resched();
1346                                flush_signals(current);
1347                        }
1348                        j = jiffies_till_next_fqs;
1349                        if (j > HZ) {
1350                                j = HZ;
1351                                jiffies_till_next_fqs = HZ;
1352                        } else if (j < 1) {
1353                                j = 1;
1354                                jiffies_till_next_fqs = 1;
1355                        }
1356                }
1357
1358                /* Handle grace-period end. */
1359                rcu_gp_cleanup(rsp);
1360        }
1361}
1362
1363/*
1364 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1365 * in preparation for detecting the next grace period.  The caller must hold
1366 * the root node's ->lock, which is released before return.  Hard irqs must
1367 * be disabled.
1368 *
1369 * Note that it is legal for a dying CPU (which is marked as offline) to
1370 * invoke this function.  This can happen when the dying CPU reports its
1371 * quiescent state.
1372 */
1373static void
1374rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1375        __releases(rcu_get_root(rsp)->lock)
1376{
1377        struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1378        struct rcu_node *rnp = rcu_get_root(rsp);
1379
1380        if (!rsp->gp_kthread ||
1381            !cpu_needs_another_gp(rsp, rdp)) {
1382                /*
1383                 * Either we have not yet spawned the grace-period
1384                 * task, this CPU does not need another grace period,
1385                 * or a grace period is already in progress.
1386                 * Either way, don't start a new grace period.
1387                 */
1388                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1389                return;
1390        }
1391
1392        /*
1393         * Because there is no grace period in progress right now,
1394         * any callbacks we have up to this point will be satisfied
1395         * by the next grace period.  So promote all callbacks to be
1396         * handled after the end of the next grace period.  If the
1397         * CPU is not yet aware of the end of the previous grace period,
1398         * we need to allow for the callback advancement that will
1399         * occur when it does become aware.  Deadlock prevents us from
1400         * making it aware at this point: We cannot acquire a leaf
1401         * rcu_node ->lock while holding the root rcu_node ->lock.
1402         */
1403        rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1404        if (rdp->completed == rsp->completed)
1405                rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1406
1407        rsp->gp_flags = RCU_GP_FLAG_INIT;
1408        raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */
1409
1410        /* Ensure that CPU is aware of completion of last grace period. */
1411        rcu_process_gp_end(rsp, rdp);
1412        local_irq_restore(flags);
1413
1414        /* Wake up rcu_gp_kthread() to start the grace period. */
1415        wake_up(&rsp->gp_wq);
1416}
1417
1418/*
1419 * Report a full set of quiescent states to the specified rcu_state
1420 * data structure.  This involves cleaning up after the prior grace
1421 * period and letting rcu_start_gp() start up the next grace period
1422 * if one is needed.  Note that the caller must hold rnp->lock, as
1423 * required by rcu_start_gp(), which will release it.
1424 */
1425static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1426        __releases(rcu_get_root(rsp)->lock)
1427{
1428        WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1429        raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1430        wake_up(&rsp->gp_wq);  /* Memory barrier implied by wake_up() path. */
1431}
1432
1433/*
1434 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1435 * Allows quiescent states for a group of CPUs to be reported at one go
1436 * to the specified rcu_node structure, though all the CPUs in the group
1437 * must be represented by the same rcu_node structure (which need not be
1438 * a leaf rcu_node structure, though it often will be).  That structure's
1439 * lock must be held upon entry, and it is released before return.
1440 */
1441static void
1442rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1443                  struct rcu_node *rnp, unsigned long flags)
1444        __releases(rnp->lock)
1445{
1446        struct rcu_node *rnp_c;
1447
1448        /* Walk up the rcu_node hierarchy. */
1449        for (;;) {
1450                if (!(rnp->qsmask & mask)) {
1451
1452                        /* Our bit has already been cleared, so done. */
1453                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1454                        return;
1455                }
1456                rnp->qsmask &= ~mask;
1457                trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1458                                                 mask, rnp->qsmask, rnp->level,
1459                                                 rnp->grplo, rnp->grphi,
1460                                                 !!rnp->gp_tasks);
1461                if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1462
1463                        /* Other bits still set at this level, so done. */
1464                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1465                        return;
1466                }
1467                mask = rnp->grpmask;
1468                if (rnp->parent == NULL) {
1469
1470                        /* No more levels.  Exit loop holding root lock. */
1471
1472                        break;
1473                }
1474                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1475                rnp_c = rnp;
1476                rnp = rnp->parent;
1477                raw_spin_lock_irqsave(&rnp->lock, flags);
1478                WARN_ON_ONCE(rnp_c->qsmask);
1479        }
1480
1481        /*
1482         * Get here if we are the last CPU to pass through a quiescent
1483         * state for this grace period.  Invoke rcu_report_qs_rsp()
1484         * to clean up and start the next grace period if one is needed.
1485         */
1486        rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1487}
1488
1489/*
1490 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1491 * structure.  This must be either called from the specified CPU, or
1492 * called when the specified CPU is known to be offline (and when it is
1493 * also known that no other CPU is concurrently trying to help the offline
1494 * CPU).  The lastcomp argument is used to make sure we are still in the
1495 * grace period of interest.  We don't want to end the current grace period
1496 * based on quiescent states detected in an earlier grace period!
1497 */
1498static void
1499rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1500{
1501        unsigned long flags;
1502        unsigned long mask;
1503        struct rcu_node *rnp;
1504
1505        rnp = rdp->mynode;
1506        raw_spin_lock_irqsave(&rnp->lock, flags);
1507        if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1508            rnp->completed == rnp->gpnum) {
1509
1510                /*
1511                 * The grace period in which this quiescent state was
1512                 * recorded has ended, so don't report it upwards.
1513                 * We will instead need a new quiescent state that lies
1514                 * within the current grace period.
1515                 */
1516                rdp->passed_quiesce = 0;        /* need qs for new gp. */
1517                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1518                return;
1519        }
1520        mask = rdp->grpmask;
1521        if ((rnp->qsmask & mask) == 0) {
1522                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1523        } else {
1524                rdp->qs_pending = 0;
1525
1526                /*
1527                 * This GP can't end until cpu checks in, so all of our
1528                 * callbacks can be processed during the next GP.
1529                 */
1530                rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1531
1532                rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1533        }
1534}
1535
1536/*
1537 * Check to see if there is a new grace period of which this CPU
1538 * is not yet aware, and if so, set up local rcu_data state for it.
1539 * Otherwise, see if this CPU has just passed through its first
1540 * quiescent state for this grace period, and record that fact if so.
1541 */
1542static void
1543rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1544{
1545        /* If there is now a new grace period, record and return. */
1546        if (check_for_new_grace_period(rsp, rdp))
1547                return;
1548
1549        /*
1550         * Does this CPU still need to do its part for current grace period?
1551         * If no, return and let the other CPUs do their part as well.
1552         */
1553        if (!rdp->qs_pending)
1554                return;
1555
1556        /*
1557         * Was there a quiescent state since the beginning of the grace
1558         * period? If no, then exit and wait for the next call.
1559         */
1560        if (!rdp->passed_quiesce)
1561                return;
1562
1563        /*
1564         * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1565         * judge of that).
1566         */
1567        rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1568}
1569
1570#ifdef CONFIG_HOTPLUG_CPU
1571
1572/*
1573 * Send the specified CPU's RCU callbacks to the orphanage.  The
1574 * specified CPU must be offline, and the caller must hold the
1575 * ->orphan_lock.
1576 */
1577static void
1578rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1579                          struct rcu_node *rnp, struct rcu_data *rdp)
1580{
1581        /* No-CBs CPUs do not have orphanable callbacks. */
1582        if (is_nocb_cpu(rdp->cpu))
1583                return;
1584
1585        /*
1586         * Orphan the callbacks.  First adjust the counts.  This is safe
1587         * because _rcu_barrier() excludes CPU-hotplug operations, so it
1588         * cannot be running now.  Thus no memory barrier is required.
1589         */
1590        if (rdp->nxtlist != NULL) {
1591                rsp->qlen_lazy += rdp->qlen_lazy;
1592                rsp->qlen += rdp->qlen;
1593                rdp->n_cbs_orphaned += rdp->qlen;
1594                rdp->qlen_lazy = 0;
1595                ACCESS_ONCE(rdp->qlen) = 0;
1596        }
1597
1598        /*
1599         * Next, move those callbacks still needing a grace period to
1600         * the orphanage, where some other CPU will pick them up.
1601         * Some of the callbacks might have gone partway through a grace
1602         * period, but that is too bad.  They get to start over because we
1603         * cannot assume that grace periods are synchronized across CPUs.
1604         * We don't bother updating the ->nxttail[] array yet, instead
1605         * we just reset the whole thing later on.
1606         */
1607        if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1608                *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1609                rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1610                *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1611        }
1612
1613        /*
1614         * Then move the ready-to-invoke callbacks to the orphanage,
1615         * where some other CPU will pick them up.  These will not be
1616         * required to pass though another grace period: They are done.
1617         */
1618        if (rdp->nxtlist != NULL) {
1619                *rsp->orphan_donetail = rdp->nxtlist;
1620                rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1621        }
1622
1623        /* Finally, initialize the rcu_data structure's list to empty.  */
1624        init_callback_list(rdp);
1625}
1626
1627/*
1628 * Adopt the RCU callbacks from the specified rcu_state structure's
1629 * orphanage.  The caller must hold the ->orphan_lock.
1630 */
1631static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1632{
1633        int i;
1634        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1635
1636        /* No-CBs CPUs are handled specially. */
1637        if (rcu_nocb_adopt_orphan_cbs(rsp, rdp))
1638                return;
1639
1640        /* Do the accounting first. */
1641        rdp->qlen_lazy += rsp->qlen_lazy;
1642        rdp->qlen += rsp->qlen;
1643        rdp->n_cbs_adopted += rsp->qlen;
1644        if (rsp->qlen_lazy != rsp->qlen)
1645                rcu_idle_count_callbacks_posted();
1646        rsp->qlen_lazy = 0;
1647        rsp->qlen = 0;
1648
1649        /*
1650         * We do not need a memory barrier here because the only way we
1651         * can get here if there is an rcu_barrier() in flight is if
1652         * we are the task doing the rcu_barrier().
1653         */
1654
1655        /* First adopt the ready-to-invoke callbacks. */
1656        if (rsp->orphan_donelist != NULL) {
1657                *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1658                *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1659                for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1660                        if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1661                                rdp->nxttail[i] = rsp->orphan_donetail;
1662                rsp->orphan_donelist = NULL;
1663                rsp->orphan_donetail = &rsp->orphan_donelist;
1664        }
1665
1666        /* And then adopt the callbacks that still need a grace period. */
1667        if (rsp->orphan_nxtlist != NULL) {
1668                *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1669                rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1670                rsp->orphan_nxtlist = NULL;
1671                rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1672        }
1673}
1674
1675/*
1676 * Trace the fact that this CPU is going offline.
1677 */
1678static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1679{
1680        RCU_TRACE(unsigned long mask);
1681        RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1682        RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1683
1684        RCU_TRACE(mask = rdp->grpmask);
1685        trace_rcu_grace_period(rsp->name,
1686                               rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1687                               "cpuofl");
1688}
1689
1690/*
1691 * The CPU has been completely removed, and some other CPU is reporting
1692 * this fact from process context.  Do the remainder of the cleanup,
1693 * including orphaning the outgoing CPU's RCU callbacks, and also
1694 * adopting them.  There can only be one CPU hotplug operation at a time,
1695 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1696 */
1697static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1698{
1699        unsigned long flags;
1700        unsigned long mask;
1701        int need_report = 0;
1702        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1703        struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
1704
1705        /* Adjust any no-longer-needed kthreads. */
1706        rcu_boost_kthread_setaffinity(rnp, -1);
1707
1708        /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1709
1710        /* Exclude any attempts to start a new grace period. */
1711        mutex_lock(&rsp->onoff_mutex);
1712        raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
1713
1714        /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1715        rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1716        rcu_adopt_orphan_cbs(rsp);
1717
1718        /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1719        mask = rdp->grpmask;    /* rnp->grplo is constant. */
1720        do {
1721                raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
1722                rnp->qsmaskinit &= ~mask;
1723                if (rnp->qsmaskinit != 0) {
1724                        if (rnp != rdp->mynode)
1725                                raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1726                        break;
1727                }
1728                if (rnp == rdp->mynode)
1729                        need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1730                else
1731                        raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1732                mask = rnp->grpmask;
1733                rnp = rnp->parent;
1734        } while (rnp != NULL);
1735
1736        /*
1737         * We still hold the leaf rcu_node structure lock here, and
1738         * irqs are still disabled.  The reason for this subterfuge is
1739         * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
1740         * held leads to deadlock.
1741         */
1742        raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
1743        rnp = rdp->mynode;
1744        if (need_report & RCU_OFL_TASKS_NORM_GP)
1745                rcu_report_unblock_qs_rnp(rnp, flags);
1746        else
1747                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1748        if (need_report & RCU_OFL_TASKS_EXP_GP)
1749                rcu_report_exp_rnp(rsp, rnp, true);
1750        WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1751                  "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1752                  cpu, rdp->qlen, rdp->nxtlist);
1753        init_callback_list(rdp);
1754        /* Disallow further callbacks on this CPU. */
1755        rdp->nxttail[RCU_NEXT_TAIL] = NULL;
1756        mutex_unlock(&rsp->onoff_mutex);
1757}
1758
1759#else /* #ifdef CONFIG_HOTPLUG_CPU */
1760
1761static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1762{
1763}
1764
1765static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1766{
1767}
1768
1769#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1770
1771/*
1772 * Invoke any RCU callbacks that have made it to the end of their grace
1773 * period.  Thottle as specified by rdp->blimit.
1774 */
1775static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1776{
1777        unsigned long flags;
1778        struct rcu_head *next, *list, **tail;
1779        long bl, count, count_lazy;
1780        int i;
1781
1782        /* If no callbacks are ready, just return.*/
1783        if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1784                trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1785                trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1786                                    need_resched(), is_idle_task(current),
1787                                    rcu_is_callbacks_kthread());
1788                return;
1789        }
1790
1791        /*
1792         * Extract the list of ready callbacks, disabling to prevent
1793         * races with call_rcu() from interrupt handlers.
1794         */
1795        local_irq_save(flags);
1796        WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1797        bl = rdp->blimit;
1798        trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1799        list = rdp->nxtlist;
1800        rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1801        *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1802        tail = rdp->nxttail[RCU_DONE_TAIL];
1803        for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1804                if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1805                        rdp->nxttail[i] = &rdp->nxtlist;
1806        local_irq_restore(flags);
1807
1808        /* Invoke callbacks. */
1809        count = count_lazy = 0;
1810        while (list) {
1811                next = list->next;
1812                prefetch(next);
1813                debug_rcu_head_unqueue(list);
1814                if (__rcu_reclaim(rsp->name, list))
1815                        count_lazy++;
1816                list = next;
1817                /* Stop only if limit reached and CPU has something to do. */
1818                if (++count >= bl &&
1819                    (need_resched() ||
1820                     (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1821                        break;
1822        }
1823
1824        local_irq_save(flags);
1825        trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1826                            is_idle_task(current),
1827                            rcu_is_callbacks_kthread());
1828
1829        /* Update count, and requeue any remaining callbacks. */
1830        if (list != NULL) {
1831                *tail = rdp->nxtlist;
1832                rdp->nxtlist = list;
1833                for (i = 0; i < RCU_NEXT_SIZE; i++)
1834                        if (&rdp->nxtlist == rdp->nxttail[i])
1835                                rdp->nxttail[i] = tail;
1836                        else
1837                                break;
1838        }
1839        smp_mb(); /* List handling before counting for rcu_barrier(). */
1840        rdp->qlen_lazy -= count_lazy;
1841        ACCESS_ONCE(rdp->qlen) -= count;
1842        rdp->n_cbs_invoked += count;
1843
1844        /* Reinstate batch limit if we have worked down the excess. */
1845        if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1846                rdp->blimit = blimit;
1847
1848        /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1849        if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1850                rdp->qlen_last_fqs_check = 0;
1851                rdp->n_force_qs_snap = rsp->n_force_qs;
1852        } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1853                rdp->qlen_last_fqs_check = rdp->qlen;
1854        WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
1855
1856        local_irq_restore(flags);
1857
1858        /* Re-invoke RCU core processing if there are callbacks remaining. */
1859        if (cpu_has_callbacks_ready_to_invoke(rdp))
1860                invoke_rcu_core();
1861}
1862
1863/*
1864 * Check to see if this CPU is in a non-context-switch quiescent state
1865 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1866 * Also schedule RCU core processing.
1867 *
1868 * This function must be called from hardirq context.  It is normally
1869 * invoked from the scheduling-clock interrupt.  If rcu_pending returns
1870 * false, there is no point in invoking rcu_check_callbacks().
1871 */
1872void rcu_check_callbacks(int cpu, int user)
1873{
1874        trace_rcu_utilization("Start scheduler-tick");
1875        increment_cpu_stall_ticks();
1876        if (user || rcu_is_cpu_rrupt_from_idle()) {
1877
1878                /*
1879                 * Get here if this CPU took its interrupt from user
1880                 * mode or from the idle loop, and if this is not a
1881                 * nested interrupt.  In this case, the CPU is in
1882                 * a quiescent state, so note it.
1883                 *
1884                 * No memory barrier is required here because both
1885                 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1886                 * variables that other CPUs neither access nor modify,
1887                 * at least not while the corresponding CPU is online.
1888                 */
1889
1890                rcu_sched_qs(cpu);
1891                rcu_bh_qs(cpu);
1892
1893        } else if (!in_softirq()) {
1894
1895                /*
1896                 * Get here if this CPU did not take its interrupt from
1897                 * softirq, in other words, if it is not interrupting
1898                 * a rcu_bh read-side critical section.  This is an _bh
1899                 * critical section, so note it.
1900                 */
1901
1902                rcu_bh_qs(cpu);
1903        }
1904        rcu_preempt_check_callbacks(cpu);
1905        if (rcu_pending(cpu))
1906                invoke_rcu_core();
1907        trace_rcu_utilization("End scheduler-tick");
1908}
1909
1910/*
1911 * Scan the leaf rcu_node structures, processing dyntick state for any that
1912 * have not yet encountered a quiescent state, using the function specified.
1913 * Also initiate boosting for any threads blocked on the root rcu_node.
1914 *
1915 * The caller must have suppressed start of new grace periods.
1916 */
1917static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1918{
1919        unsigned long bit;
1920        int cpu;
1921        unsigned long flags;
1922        unsigned long mask;
1923        struct rcu_node *rnp;
1924
1925        rcu_for_each_leaf_node(rsp, rnp) {
1926                cond_resched();
1927                mask = 0;
1928                raw_spin_lock_irqsave(&rnp->lock, flags);
1929                if (!rcu_gp_in_progress(rsp)) {
1930                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1931                        return;
1932                }
1933                if (rnp->qsmask == 0) {
1934                        rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1935                        continue;
1936                }
1937                cpu = rnp->grplo;
1938                bit = 1;
1939                for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1940                        if ((rnp->qsmask & bit) != 0 &&
1941                            f(per_cpu_ptr(rsp->rda, cpu)))
1942                                mask |= bit;
1943                }
1944                if (mask != 0) {
1945
1946                        /* rcu_report_qs_rnp() releases rnp->lock. */
1947                        rcu_report_qs_rnp(mask, rsp, rnp, flags);
1948                        continue;
1949                }
1950                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1951        }
1952        rnp = rcu_get_root(rsp);
1953        if (rnp->qsmask == 0) {
1954                raw_spin_lock_irqsave(&rnp->lock, flags);
1955                rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1956        }
1957}
1958
1959/*
1960 * Force quiescent states on reluctant CPUs, and also detect which
1961 * CPUs are in dyntick-idle mode.
1962 */
1963static void force_quiescent_state(struct rcu_state *rsp)
1964{
1965        unsigned long flags;
1966        bool ret;
1967        struct rcu_node *rnp;
1968        struct rcu_node *rnp_old = NULL;
1969
1970        /* Funnel through hierarchy to reduce memory contention. */
1971        rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
1972        for (; rnp != NULL; rnp = rnp->parent) {
1973                ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
1974                      !raw_spin_trylock(&rnp->fqslock);
1975                if (rnp_old != NULL)
1976                        raw_spin_unlock(&rnp_old->fqslock);
1977                if (ret) {
1978                        rsp->n_force_qs_lh++;
1979                        return;
1980                }
1981                rnp_old = rnp;
1982        }
1983        /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
1984
1985        /* Reached the root of the rcu_node tree, acquire lock. */
1986        raw_spin_lock_irqsave(&rnp_old->lock, flags);
1987        raw_spin_unlock(&rnp_old->fqslock);
1988        if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1989                rsp->n_force_qs_lh++;
1990                raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
1991                return;  /* Someone beat us to it. */
1992        }
1993        rsp->gp_flags |= RCU_GP_FLAG_FQS;
1994        raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
1995        wake_up(&rsp->gp_wq);  /* Memory barrier implied by wake_up() path. */
1996}
1997
1998/*
1999 * This does the RCU core processing work for the specified rcu_state
2000 * and rcu_data structures.  This may be called only from the CPU to
2001 * whom the rdp belongs.
2002 */
2003static void
2004__rcu_process_callbacks(struct rcu_state *rsp)
2005{
2006        unsigned long flags;
2007        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2008
2009        WARN_ON_ONCE(rdp->beenonline == 0);
2010
2011        /*
2012         * Advance callbacks in response to end of earlier grace
2013         * period that some other CPU ended.
2014         */
2015        rcu_process_gp_end(rsp, rdp);
2016
2017        /* Update RCU state based on any recent quiescent states. */
2018        rcu_check_quiescent_state(rsp, rdp);
2019
2020        /* Does this CPU require a not-yet-started grace period? */
2021        if (cpu_needs_another_gp(rsp, rdp)) {
2022                raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
2023                rcu_start_gp(rsp, flags);  /* releases above lock */
2024        }
2025
2026        /* If there are callbacks ready, invoke them. */
2027        if (cpu_has_callbacks_ready_to_invoke(rdp))
2028                invoke_rcu_callbacks(rsp, rdp);
2029}
2030
2031/*
2032 * Do RCU core processing for the current CPU.
2033 */
2034static void rcu_process_callbacks(struct softirq_action *unused)
2035{
2036        struct rcu_state *rsp;
2037
2038        if (cpu_is_offline(smp_processor_id()))
2039                return;
2040        trace_rcu_utilization("Start RCU core");
2041        for_each_rcu_flavor(rsp)
2042                __rcu_process_callbacks(rsp);
2043        trace_rcu_utilization("End RCU core");
2044}
2045
2046/*
2047 * Schedule RCU callback invocation.  If the specified type of RCU
2048 * does not support RCU priority boosting, just do a direct call,
2049 * otherwise wake up the per-CPU kernel kthread.  Note that because we
2050 * are running on the current CPU with interrupts disabled, the
2051 * rcu_cpu_kthread_task cannot disappear out from under us.
2052 */
2053static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2054{
2055        if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2056                return;
2057        if (likely(!rsp->boost)) {
2058                rcu_do_batch(rsp, rdp);
2059                return;
2060        }
2061        invoke_rcu_callbacks_kthread();
2062}
2063
2064static void invoke_rcu_core(void)
2065{
2066        raise_softirq(RCU_SOFTIRQ);
2067}
2068
2069/*
2070 * Handle any core-RCU processing required by a call_rcu() invocation.
2071 */
2072static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2073                            struct rcu_head *head, unsigned long flags)
2074{
2075        /*
2076         * If called from an extended quiescent state, invoke the RCU
2077         * core in order to force a re-evaluation of RCU's idleness.
2078         */
2079        if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2080                invoke_rcu_core();
2081
2082        /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2083        if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2084                return;
2085
2086        /*
2087         * Force the grace period if too many callbacks or too long waiting.
2088         * Enforce hysteresis, and don't invoke force_quiescent_state()
2089         * if some other CPU has recently done so.  Also, don't bother
2090         * invoking force_quiescent_state() if the newly enqueued callback
2091         * is the only one waiting for a grace period to complete.
2092         */
2093        if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2094
2095                /* Are we ignoring a completed grace period? */
2096                rcu_process_gp_end(rsp, rdp);
2097                check_for_new_grace_period(rsp, rdp);
2098
2099                /* Start a new grace period if one not already started. */
2100                if (!rcu_gp_in_progress(rsp)) {
2101                        unsigned long nestflag;
2102                        struct rcu_node *rnp_root = rcu_get_root(rsp);
2103
2104                        raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
2105                        rcu_start_gp(rsp, nestflag);  /* rlses rnp_root->lock */
2106                } else {
2107                        /* Give the grace period a kick. */
2108                        rdp->blimit = LONG_MAX;
2109                        if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2110                            *rdp->nxttail[RCU_DONE_TAIL] != head)
2111                                force_quiescent_state(rsp);
2112                        rdp->n_force_qs_snap = rsp->n_force_qs;
2113                        rdp->qlen_last_fqs_check = rdp->qlen;
2114                }
2115        }
2116}
2117
2118/*
2119 * Helper function for call_rcu() and friends.  The cpu argument will
2120 * normally be -1, indicating "currently running CPU".  It may specify
2121 * a CPU only if that CPU is a no-CBs CPU.  Currently, only _rcu_barrier()
2122 * is expected to specify a CPU.
2123 */
2124static void
2125__call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2126           struct rcu_state *rsp, int cpu, bool lazy)
2127{
2128        unsigned long flags;
2129        struct rcu_data *rdp;
2130
2131        WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2132        debug_rcu_head_queue(head);
2133        head->func = func;
2134        head->next = NULL;
2135
2136        /*
2137         * Opportunistically note grace-period endings and beginnings.
2138         * Note that we might see a beginning right after we see an
2139         * end, but never vice versa, since this CPU has to pass through
2140         * a quiescent state betweentimes.
2141         */
2142        local_irq_save(flags);
2143        rdp = this_cpu_ptr(rsp->rda);
2144
2145        /* Add the callback to our list. */
2146        if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2147                int offline;
2148
2149                if (cpu != -1)
2150                        rdp = per_cpu_ptr(rsp->rda, cpu);
2151                offline = !__call_rcu_nocb(rdp, head, lazy);
2152                WARN_ON_ONCE(offline);
2153                /* _call_rcu() is illegal on offline CPU; leak the callback. */
2154                local_irq_restore(flags);
2155                return;
2156        }
2157        ACCESS_ONCE(rdp->qlen)++;
2158        if (lazy)
2159                rdp->qlen_lazy++;
2160        else
2161                rcu_idle_count_callbacks_posted();
2162        smp_mb();  /* Count before adding callback for rcu_barrier(). */
2163        *rdp->nxttail[RCU_NEXT_TAIL] = head;
2164        rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2165
2166        if (__is_kfree_rcu_offset((unsigned long)func))
2167                trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2168                                         rdp->qlen_lazy, rdp->qlen);
2169        else
2170                trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2171
2172        /* Go handle any RCU core processing required. */
2173        __call_rcu_core(rsp, rdp, head, flags);
2174        local_irq_restore(flags);
2175}
2176
2177/*
2178 * Queue an RCU-sched callback for invocation after a grace period.
2179 */
2180void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2181{
2182        __call_rcu(head, func, &rcu_sched_state, -1, 0);
2183}
2184EXPORT_SYMBOL_GPL(call_rcu_sched);
2185
2186/*
2187 * Queue an RCU callback for invocation after a quicker grace period.
2188 */
2189void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2190{
2191        __call_rcu(head, func, &rcu_bh_state, -1, 0);
2192}
2193EXPORT_SYMBOL_GPL(call_rcu_bh);
2194
2195/*
2196 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2197 * any blocking grace-period wait automatically implies a grace period
2198 * if there is only one CPU online at any point time during execution
2199 * of either synchronize_sched() or synchronize_rcu_bh().  It is OK to
2200 * occasionally incorrectly indicate that there are multiple CPUs online
2201 * when there was in fact only one the whole time, as this just adds
2202 * some overhead: RCU still operates correctly.
2203 */
2204static inline int rcu_blocking_is_gp(void)
2205{
2206        int ret;
2207
2208        might_sleep();  /* Check for RCU read-side critical section. */
2209        preempt_disable();
2210        ret = num_online_cpus() <= 1;
2211        preempt_enable();
2212        return ret;
2213}
2214
2215/**
2216 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2217 *
2218 * Control will return to the caller some time after a full rcu-sched
2219 * grace period has elapsed, in other words after all currently executing
2220 * rcu-sched read-side critical sections have completed.   These read-side
2221 * critical sections are delimited by rcu_read_lock_sched() and
2222 * rcu_read_unlock_sched(), and may be nested.  Note that preempt_disable(),
2223 * local_irq_disable(), and so on may be used in place of
2224 * rcu_read_lock_sched().
2225 *
2226 * This means that all preempt_disable code sequences, including NMI and
2227 * non-threaded hardware-interrupt handlers, in progress on entry will
2228 * have completed before this primitive returns.  However, this does not
2229 * guarantee that softirq handlers will have completed, since in some
2230 * kernels, these handlers can run in process context, and can block.
2231 *
2232 * Note that this guarantee implies further memory-ordering guarantees.
2233 * On systems with more than one CPU, when synchronize_sched() returns,
2234 * each CPU is guaranteed to have executed a full memory barrier since the
2235 * end of its last RCU-sched read-side critical section whose beginning
2236 * preceded the call to synchronize_sched().  In addition, each CPU having
2237 * an RCU read-side critical section that extends beyond the return from
2238 * synchronize_sched() is guaranteed to have executed a full memory barrier
2239 * after the beginning of synchronize_sched() and before the beginning of
2240 * that RCU read-side critical section.  Note that these guarantees include
2241 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2242 * that are executing in the kernel.
2243 *
2244 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2245 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2246 * to have executed a full memory barrier during the execution of
2247 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2248 * again only if the system has more than one CPU).
2249 *
2250 * This primitive provides the guarantees made by the (now removed)
2251 * synchronize_kernel() API.  In contrast, synchronize_rcu() only
2252 * guarantees that rcu_read_lock() sections will have completed.
2253 * In "classic RCU", these two guarantees happen to be one and
2254 * the same, but can differ in realtime RCU implementations.
2255 */
2256void synchronize_sched(void)
2257{
2258        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2259                           !lock_is_held(&rcu_lock_map) &&
2260                           !lock_is_held(&rcu_sched_lock_map),
2261                           "Illegal synchronize_sched() in RCU-sched read-side critical section");
2262        if (rcu_blocking_is_gp())
2263                return;
2264        if (rcu_expedited)
2265                synchronize_sched_expedited();
2266        else
2267                wait_rcu_gp(call_rcu_sched);
2268}
2269EXPORT_SYMBOL_GPL(synchronize_sched);
2270
2271/**
2272 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2273 *
2274 * Control will return to the caller some time after a full rcu_bh grace
2275 * period has elapsed, in other words after all currently executing rcu_bh
2276 * read-side critical sections have completed.  RCU read-side critical
2277 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2278 * and may be nested.
2279 *
2280 * See the description of synchronize_sched() for more detailed information
2281 * on memory ordering guarantees.
2282 */
2283void synchronize_rcu_bh(void)
2284{
2285        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2286                           !lock_is_held(&rcu_lock_map) &&
2287                           !lock_is_held(&rcu_sched_lock_map),
2288                           "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2289        if (rcu_blocking_is_gp())
2290                return;
2291        if (rcu_expedited)
2292                synchronize_rcu_bh_expedited();
2293        else
2294                wait_rcu_gp(call_rcu_bh);
2295}
2296EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2297
2298static int synchronize_sched_expedited_cpu_stop(void *data)
2299{
2300        /*
2301         * There must be a full memory barrier on each affected CPU
2302         * between the time that try_stop_cpus() is called and the
2303         * time that it returns.
2304         *
2305         * In the current initial implementation of cpu_stop, the
2306         * above condition is already met when the control reaches
2307         * this point and the following smp_mb() is not strictly
2308         * necessary.  Do smp_mb() anyway for documentation and
2309         * robustness against future implementation changes.
2310         */
2311        smp_mb(); /* See above comment block. */
2312        return 0;
2313}
2314
2315/**
2316 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2317 *
2318 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2319 * approach to force the grace period to end quickly.  This consumes
2320 * significant time on all CPUs and is unfriendly to real-time workloads,
2321 * so is thus not recommended for any sort of common-case code.  In fact,
2322 * if you are using synchronize_sched_expedited() in a loop, please
2323 * restructure your code to batch your updates, and then use a single
2324 * synchronize_sched() instead.
2325 *
2326 * Note that it is illegal to call this function while holding any lock
2327 * that is acquired by a CPU-hotplug notifier.  And yes, it is also illegal
2328 * to call this function from a CPU-hotplug notifier.  Failing to observe
2329 * these restriction will result in deadlock.
2330 *
2331 * This implementation can be thought of as an application of ticket
2332 * locking to RCU, with sync_sched_expedited_started and
2333 * sync_sched_expedited_done taking on the roles of the halves
2334 * of the ticket-lock word.  Each task atomically increments
2335 * sync_sched_expedited_started upon entry, snapshotting the old value,
2336 * then attempts to stop all the CPUs.  If this succeeds, then each
2337 * CPU will have executed a context switch, resulting in an RCU-sched
2338 * grace period.  We are then done, so we use atomic_cmpxchg() to
2339 * update sync_sched_expedited_done to match our snapshot -- but
2340 * only if someone else has not already advanced past our snapshot.
2341 *
2342 * On the other hand, if try_stop_cpus() fails, we check the value
2343 * of sync_sched_expedited_done.  If it has advanced past our
2344 * initial snapshot, then someone else must have forced a grace period
2345 * some time after we took our snapshot.  In this case, our work is
2346 * done for us, and we can simply return.  Otherwise, we try again,
2347 * but keep our initial snapshot for purposes of checking for someone
2348 * doing our work for us.
2349 *
2350 * If we fail too many times in a row, we fall back to synchronize_sched().
2351 */
2352void synchronize_sched_expedited(void)
2353{
2354        long firstsnap, s, snap;
2355        int trycount = 0;
2356        struct rcu_state *rsp = &rcu_sched_state;
2357
2358        /*
2359         * If we are in danger of counter wrap, just do synchronize_sched().
2360         * By allowing sync_sched_expedited_started to advance no more than
2361         * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2362         * that more than 3.5 billion CPUs would be required to force a
2363         * counter wrap on a 32-bit system.  Quite a few more CPUs would of
2364         * course be required on a 64-bit system.
2365         */
2366        if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2367                         (ulong)atomic_long_read(&rsp->expedited_done) +
2368                         ULONG_MAX / 8)) {
2369                synchronize_sched();
2370                atomic_long_inc(&rsp->expedited_wrap);
2371                return;
2372        }
2373
2374        /*
2375         * Take a ticket.  Note that atomic_inc_return() implies a
2376         * full memory barrier.
2377         */
2378        snap = atomic_long_inc_return(&rsp->expedited_start);
2379        firstsnap = snap;
2380        get_online_cpus();
2381        WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2382
2383        /*
2384         * Each pass through the following loop attempts to force a
2385         * context switch on each CPU.
2386         */
2387        while (try_stop_cpus(cpu_online_mask,
2388                             synchronize_sched_expedited_cpu_stop,
2389                             NULL) == -EAGAIN) {
2390                put_online_cpus();
2391                atomic_long_inc(&rsp->expedited_tryfail);
2392
2393                /* Check to see if someone else did our work for us. */
2394                s = atomic_long_read(&rsp->expedited_done);
2395                if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2396                        /* ensure test happens before caller kfree */
2397                        smp_mb__before_atomic_inc(); /* ^^^ */
2398                        atomic_long_inc(&rsp->expedited_workdone1);
2399                        return;
2400                }
2401
2402                /* No joy, try again later.  Or just synchronize_sched(). */
2403                if (trycount++ < 10) {
2404                        udelay(trycount * num_online_cpus());
2405                } else {
2406                        wait_rcu_gp(call_rcu_sched);
2407                        atomic_long_inc(&rsp->expedited_normal);
2408                        return;
2409                }
2410
2411                /* Recheck to see if someone else did our work for us. */
2412                s = atomic_long_read(&rsp->expedited_done);
2413                if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2414                        /* ensure test happens before caller kfree */
2415                        smp_mb__before_atomic_inc(); /* ^^^ */
2416                        atomic_long_inc(&rsp->expedited_workdone2);
2417                        return;
2418                }
2419
2420                /*
2421                 * Refetching sync_sched_expedited_started allows later
2422                 * callers to piggyback on our grace period.  We retry
2423                 * after they started, so our grace period works for them,
2424                 * and they started after our first try, so their grace
2425                 * period works for us.
2426                 */
2427                get_online_cpus();
2428                snap = atomic_long_read(&rsp->expedited_start);
2429                smp_mb(); /* ensure read is before try_stop_cpus(). */
2430        }
2431        atomic_long_inc(&rsp->expedited_stoppedcpus);
2432
2433        /*
2434         * Everyone up to our most recent fetch is covered by our grace
2435         * period.  Update the counter, but only if our work is still
2436         * relevant -- which it won't be if someone who started later
2437         * than we did already did their update.
2438         */
2439        do {
2440                atomic_long_inc(&rsp->expedited_done_tries);
2441                s = atomic_long_read(&rsp->expedited_done);
2442                if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2443                        /* ensure test happens before caller kfree */
2444                        smp_mb__before_atomic_inc(); /* ^^^ */
2445                        atomic_long_inc(&rsp->expedited_done_lost);
2446                        break;
2447                }
2448        } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2449        atomic_long_inc(&rsp->expedited_done_exit);
2450
2451        put_online_cpus();
2452}
2453EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2454
2455/*
2456 * Check to see if there is any immediate RCU-related work to be done
2457 * by the current CPU, for the specified type of RCU, returning 1 if so.
2458 * The checks are in order of increasing expense: checks that can be
2459 * carried out against CPU-local state are performed first.  However,
2460 * we must check for CPU stalls first, else we might not get a chance.
2461 */
2462static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2463{
2464        struct rcu_node *rnp = rdp->mynode;
2465
2466        rdp->n_rcu_pending++;
2467
2468        /* Check for CPU stalls, if enabled. */
2469        check_cpu_stall(rsp, rdp);
2470
2471        /* Is the RCU core waiting for a quiescent state from this CPU? */
2472        if (rcu_scheduler_fully_active &&
2473            rdp->qs_pending && !rdp->passed_quiesce) {
2474                rdp->n_rp_qs_pending++;
2475        } else if (rdp->qs_pending && rdp->passed_quiesce) {
2476                rdp->n_rp_report_qs++;
2477                return 1;
2478        }
2479
2480        /* Does this CPU have callbacks ready to invoke? */
2481        if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2482                rdp->n_rp_cb_ready++;
2483                return 1;
2484        }
2485
2486        /* Has RCU gone idle with this CPU needing another grace period? */
2487        if (cpu_needs_another_gp(rsp, rdp)) {
2488                rdp->n_rp_cpu_needs_gp++;
2489                return 1;
2490        }
2491
2492        /* Has another RCU grace period completed?  */
2493        if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2494                rdp->n_rp_gp_completed++;
2495                return 1;
2496        }
2497
2498        /* Has a new RCU grace period started? */
2499        if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2500                rdp->n_rp_gp_started++;
2501                return 1;
2502        }
2503
2504        /* nothing to do */
2505        rdp->n_rp_need_nothing++;
2506        return 0;
2507}
2508
2509/*
2510 * Check to see if there is any immediate RCU-related work to be done
2511 * by the current CPU, returning 1 if so.  This function is part of the
2512 * RCU implementation; it is -not- an exported member of the RCU API.
2513 */
2514static int rcu_pending(int cpu)
2515{
2516        struct rcu_state *rsp;
2517
2518        for_each_rcu_flavor(rsp)
2519                if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2520                        return 1;
2521        return 0;
2522}
2523
2524/*
2525 * Check to see if any future RCU-related work will need to be done
2526 * by the current CPU, even if none need be done immediately, returning
2527 * 1 if so.
2528 */
2529static int rcu_cpu_has_callbacks(int cpu)
2530{
2531        struct rcu_state *rsp;
2532
2533        /* RCU callbacks either ready or pending? */
2534        for_each_rcu_flavor(rsp)
2535                if (per_cpu_ptr(rsp->rda, cpu)->nxtlist)
2536                        return 1;
2537        return 0;
2538}
2539
2540/*
2541 * Helper function for _rcu_barrier() tracing.  If tracing is disabled,
2542 * the compiler is expected to optimize this away.
2543 */
2544static void _rcu_barrier_trace(struct rcu_state *rsp, char *s,
2545                               int cpu, unsigned long done)
2546{
2547        trace_rcu_barrier(rsp->name, s, cpu,
2548                          atomic_read(&rsp->barrier_cpu_count), done);
2549}
2550
2551/*
2552 * RCU callback function for _rcu_barrier().  If we are last, wake
2553 * up the task executing _rcu_barrier().
2554 */
2555static void rcu_barrier_callback(struct rcu_head *rhp)
2556{
2557        struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2558        struct rcu_state *rsp = rdp->rsp;
2559
2560        if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2561                _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2562                complete(&rsp->barrier_completion);
2563        } else {
2564                _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2565        }
2566}
2567
2568/*
2569 * Called with preemption disabled, and from cross-cpu IRQ context.
2570 */
2571static void rcu_barrier_func(void *type)
2572{
2573        struct rcu_state *rsp = type;
2574        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2575
2576        _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2577        atomic_inc(&rsp->barrier_cpu_count);
2578        rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2579}
2580
2581/*
2582 * Orchestrate the specified type of RCU barrier, waiting for all
2583 * RCU callbacks of the specified type to complete.
2584 */
2585static void _rcu_barrier(struct rcu_state *rsp)
2586{
2587        int cpu;
2588        struct rcu_data *rdp;
2589        unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2590        unsigned long snap_done;
2591
2592        _rcu_barrier_trace(rsp, "Begin", -1, snap);
2593
2594        /* Take mutex to serialize concurrent rcu_barrier() requests. */
2595        mutex_lock(&rsp->barrier_mutex);
2596
2597        /*
2598         * Ensure that all prior references, including to ->n_barrier_done,
2599         * are ordered before the _rcu_barrier() machinery.
2600         */
2601        smp_mb();  /* See above block comment. */
2602
2603        /*
2604         * Recheck ->n_barrier_done to see if others did our work for us.
2605         * This means checking ->n_barrier_done for an even-to-odd-to-even
2606         * transition.  The "if" expression below therefore rounds the old
2607         * value up to the next even number and adds two before comparing.
2608         */
2609        snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2610        _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2611        if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2612                _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2613                smp_mb(); /* caller's subsequent code after above check. */
2614                mutex_unlock(&rsp->barrier_mutex);
2615                return;
2616        }
2617
2618        /*
2619         * Increment ->n_barrier_done to avoid duplicate work.  Use
2620         * ACCESS_ONCE() to prevent the compiler from speculating
2621         * the increment to precede the early-exit check.
2622         */
2623        ACCESS_ONCE(rsp->n_barrier_done)++;
2624        WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2625        _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2626        smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2627
2628        /*
2629         * Initialize the count to one rather than to zero in order to
2630         * avoid a too-soon return to zero in case of a short grace period
2631         * (or preemption of this task).  Exclude CPU-hotplug operations
2632         * to ensure that no offline CPU has callbacks queued.
2633         */
2634        init_completion(&rsp->barrier_completion);
2635        atomic_set(&rsp->barrier_cpu_count, 1);
2636        get_online_cpus();
2637
2638        /*
2639         * Force each CPU with callbacks to register a new callback.
2640         * When that callback is invoked, we will know that all of the
2641         * corresponding CPU's preceding callbacks have been invoked.
2642         */
2643        for_each_possible_cpu(cpu) {
2644                if (!cpu_online(cpu) && !is_nocb_cpu(cpu))
2645                        continue;
2646                rdp = per_cpu_ptr(rsp->rda, cpu);
2647                if (is_nocb_cpu(cpu)) {
2648                        _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
2649                                           rsp->n_barrier_done);
2650                        atomic_inc(&rsp->barrier_cpu_count);
2651                        __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
2652                                   rsp, cpu, 0);
2653                } else if (ACCESS_ONCE(rdp->qlen)) {
2654                        _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2655                                           rsp->n_barrier_done);
2656                        smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2657                } else {
2658                        _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2659                                           rsp->n_barrier_done);
2660                }
2661        }
2662        put_online_cpus();
2663
2664        /*
2665         * Now that we have an rcu_barrier_callback() callback on each
2666         * CPU, and thus each counted, remove the initial count.
2667         */
2668        if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2669                complete(&rsp->barrier_completion);
2670
2671        /* Increment ->n_barrier_done to prevent duplicate work. */
2672        smp_mb(); /* Keep increment after above mechanism. */
2673        ACCESS_ONCE(rsp->n_barrier_done)++;
2674        WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2675        _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2676        smp_mb(); /* Keep increment before caller's subsequent code. */
2677
2678        /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2679        wait_for_completion(&rsp->barrier_completion);
2680
2681        /* Other rcu_barrier() invocations can now safely proceed. */
2682        mutex_unlock(&rsp->barrier_mutex);
2683}
2684
2685/**
2686 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2687 */
2688void rcu_barrier_bh(void)
2689{
2690        _rcu_barrier(&rcu_bh_state);
2691}
2692EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2693
2694/**
2695 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2696 */
2697void rcu_barrier_sched(void)
2698{
2699        _rcu_barrier(&rcu_sched_state);
2700}
2701EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2702
2703/*
2704 * Do boot-time initialization of a CPU's per-CPU RCU data.
2705 */
2706static void __init
2707rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2708{
2709        unsigned long flags;
2710        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2711        struct rcu_node *rnp = rcu_get_root(rsp);
2712
2713        /* Set up local state, ensuring consistent view of global state. */
2714        raw_spin_lock_irqsave(&rnp->lock, flags);
2715        rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2716        init_callback_list(rdp);
2717        rdp->qlen_lazy = 0;
2718        ACCESS_ONCE(rdp->qlen) = 0;
2719        rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2720        WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2721        WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2722#ifdef CONFIG_RCU_USER_QS
2723        WARN_ON_ONCE(rdp->dynticks->in_user);
2724#endif
2725        rdp->cpu = cpu;
2726        rdp->rsp = rsp;
2727        rcu_boot_init_nocb_percpu_data(rdp);
2728        raw_spin_unlock_irqrestore(&rnp->lock, flags);
2729}
2730
2731/*
2732 * Initialize a CPU's per-CPU RCU data.  Note that only one online or
2733 * offline event can be happening at a given time.  Note also that we
2734 * can accept some slop in the rsp->completed access due to the fact
2735 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2736 */
2737static void __cpuinit
2738rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2739{
2740        unsigned long flags;
2741        unsigned long mask;
2742        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2743        struct rcu_node *rnp = rcu_get_root(rsp);
2744
2745        /* Exclude new grace periods. */
2746        mutex_lock(&rsp->onoff_mutex);
2747
2748        /* Set up local state, ensuring consistent view of global state. */
2749        raw_spin_lock_irqsave(&rnp->lock, flags);
2750        rdp->beenonline = 1;     /* We have now been online. */
2751        rdp->preemptible = preemptible;
2752        rdp->qlen_last_fqs_check = 0;
2753        rdp->n_force_qs_snap = rsp->n_force_qs;
2754        rdp->blimit = blimit;
2755        init_callback_list(rdp);  /* Re-enable callbacks on this CPU. */
2756        rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2757        atomic_set(&rdp->dynticks->dynticks,
2758                   (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2759        rcu_prepare_for_idle_init(cpu);
2760        raw_spin_unlock(&rnp->lock);            /* irqs remain disabled. */
2761
2762        /* Add CPU to rcu_node bitmasks. */
2763        rnp = rdp->mynode;
2764        mask = rdp->grpmask;
2765        do {
2766                /* Exclude any attempts to start a new GP on small systems. */
2767                raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
2768                rnp->qsmaskinit |= mask;
2769                mask = rnp->grpmask;
2770                if (rnp == rdp->mynode) {
2771                        /*
2772                         * If there is a grace period in progress, we will
2773                         * set up to wait for it next time we run the
2774                         * RCU core code.
2775                         */
2776                        rdp->gpnum = rnp->completed;
2777                        rdp->completed = rnp->completed;
2778                        rdp->passed_quiesce = 0;
2779                        rdp->qs_pending = 0;
2780                        trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2781                }
2782                raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2783                rnp = rnp->parent;
2784        } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2785        local_irq_restore(flags);
2786
2787        mutex_unlock(&rsp->onoff_mutex);
2788}
2789
2790static void __cpuinit rcu_prepare_cpu(int cpu)
2791{
2792        struct rcu_state *rsp;
2793
2794        for_each_rcu_flavor(rsp)
2795                rcu_init_percpu_data(cpu, rsp,
2796                                     strcmp(rsp->name, "rcu_preempt") == 0);
2797}
2798
2799/*
2800 * Handle CPU online/offline notification events.
2801 */
2802static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2803                                    unsigned long action, void *hcpu)
2804{
2805        long cpu = (long)hcpu;
2806        struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2807        struct rcu_node *rnp = rdp->mynode;
2808        struct rcu_state *rsp;
2809        int ret = NOTIFY_OK;
2810
2811        trace_rcu_utilization("Start CPU hotplug");
2812        switch (action) {
2813        case CPU_UP_PREPARE:
2814        case CPU_UP_PREPARE_FROZEN:
2815                rcu_prepare_cpu(cpu);
2816                rcu_prepare_kthreads(cpu);
2817                break;
2818        case CPU_ONLINE:
2819        case CPU_DOWN_FAILED:
2820                rcu_boost_kthread_setaffinity(rnp, -1);
2821                break;
2822        case CPU_DOWN_PREPARE:
2823                if (nocb_cpu_expendable(cpu))
2824                        rcu_boost_kthread_setaffinity(rnp, cpu);
2825                else
2826                        ret = NOTIFY_BAD;
2827                break;
2828        case CPU_DYING:
2829        case CPU_DYING_FROZEN:
2830                /*
2831                 * The whole machine is "stopped" except this CPU, so we can
2832                 * touch any data without introducing corruption. We send the
2833                 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2834                 */
2835                for_each_rcu_flavor(rsp)
2836                        rcu_cleanup_dying_cpu(rsp);
2837                rcu_cleanup_after_idle(cpu);
2838                break;
2839        case CPU_DEAD:
2840        case CPU_DEAD_FROZEN:
2841        case CPU_UP_CANCELED:
2842        case CPU_UP_CANCELED_FROZEN:
2843                for_each_rcu_flavor(rsp)
2844                        rcu_cleanup_dead_cpu(cpu, rsp);
2845                break;
2846        default:
2847                break;
2848        }
2849        trace_rcu_utilization("End CPU hotplug");
2850        return ret;
2851}
2852
2853/*
2854 * Spawn the kthread that handles this RCU flavor's grace periods.
2855 */
2856static int __init rcu_spawn_gp_kthread(void)
2857{
2858        unsigned long flags;
2859        struct rcu_node *rnp;
2860        struct rcu_state *rsp;
2861        struct task_struct *t;
2862
2863        for_each_rcu_flavor(rsp) {
2864                t = kthread_run(rcu_gp_kthread, rsp, rsp->name);
2865                BUG_ON(IS_ERR(t));
2866                rnp = rcu_get_root(rsp);
2867                raw_spin_lock_irqsave(&rnp->lock, flags);
2868                rsp->gp_kthread = t;
2869                raw_spin_unlock_irqrestore(&rnp->lock, flags);
2870                rcu_spawn_nocb_kthreads(rsp);
2871        }
2872        return 0;
2873}
2874early_initcall(rcu_spawn_gp_kthread);
2875
2876/*
2877 * This function is invoked towards the end of the scheduler's initialization
2878 * process.  Before this is called, the idle task might contain
2879 * RCU read-side critical sections (during which time, this idle
2880 * task is booting the system).  After this function is called, the
2881 * idle tasks are prohibited from containing RCU read-side critical
2882 * sections.  This function also enables RCU lockdep checking.
2883 */
2884void rcu_scheduler_starting(void)
2885{
2886        WARN_ON(num_online_cpus() != 1);
2887        WARN_ON(nr_context_switches() > 0);
2888        rcu_scheduler_active = 1;
2889}
2890
2891/*
2892 * Compute the per-level fanout, either using the exact fanout specified
2893 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2894 */
2895#ifdef CONFIG_RCU_FANOUT_EXACT
2896static void __init rcu_init_levelspread(struct rcu_state *rsp)
2897{
2898        int i;
2899
2900        for (i = rcu_num_lvls - 1; i > 0; i--)
2901                rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2902        rsp->levelspread[0] = rcu_fanout_leaf;
2903}
2904#else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2905static void __init rcu_init_levelspread(struct rcu_state *rsp)
2906{
2907        int ccur;
2908        int cprv;
2909        int i;
2910
2911        cprv = nr_cpu_ids;
2912        for (i = rcu_num_lvls - 1; i >= 0; i--) {
2913                ccur = rsp->levelcnt[i];
2914                rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2915                cprv = ccur;
2916        }
2917}
2918#endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2919
2920/*
2921 * Helper function for rcu_init() that initializes one rcu_state structure.
2922 */
2923static void __init rcu_init_one(struct rcu_state *rsp,
2924                struct rcu_data __percpu *rda)
2925{
2926        static char *buf[] = { "rcu_node_0",
2927                               "rcu_node_1",
2928                               "rcu_node_2",
2929                               "rcu_node_3" };  /* Match MAX_RCU_LVLS */
2930        static char *fqs[] = { "rcu_node_fqs_0",
2931                               "rcu_node_fqs_1",
2932                               "rcu_node_fqs_2",
2933                               "rcu_node_fqs_3" };  /* Match MAX_RCU_LVLS */
2934        int cpustride = 1;
2935        int i;
2936        int j;
2937        struct rcu_node *rnp;
2938
2939        BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf));  /* Fix buf[] init! */
2940
2941        /* Initialize the level-tracking arrays. */
2942
2943        for (i = 0; i < rcu_num_lvls; i++)
2944                rsp->levelcnt[i] = num_rcu_lvl[i];
2945        for (i = 1; i < rcu_num_lvls; i++)
2946                rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2947        rcu_init_levelspread(rsp);
2948
2949        /* Initialize the elements themselves, starting from the leaves. */
2950
2951        for (i = rcu_num_lvls - 1; i >= 0; i--) {
2952                cpustride *= rsp->levelspread[i];
2953                rnp = rsp->level[i];
2954                for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2955                        raw_spin_lock_init(&rnp->lock);
2956                        lockdep_set_class_and_name(&rnp->lock,
2957                                                   &rcu_node_class[i], buf[i]);
2958                        raw_spin_lock_init(&rnp->fqslock);
2959                        lockdep_set_class_and_name(&rnp->fqslock,
2960                                                   &rcu_fqs_class[i], fqs[i]);
2961                        rnp->gpnum = rsp->gpnum;
2962                        rnp->completed = rsp->completed;
2963                        rnp->qsmask = 0;
2964                        rnp->qsmaskinit = 0;
2965                        rnp->grplo = j * cpustride;
2966                        rnp->grphi = (j + 1) * cpustride - 1;
2967                        if (rnp->grphi >= NR_CPUS)
2968                                rnp->grphi = NR_CPUS - 1;
2969                        if (i == 0) {
2970                                rnp->grpnum = 0;
2971                                rnp->grpmask = 0;
2972                                rnp->parent = NULL;
2973                        } else {
2974                                rnp->grpnum = j % rsp->levelspread[i - 1];
2975                                rnp->grpmask = 1UL << rnp->grpnum;
2976                                rnp->parent = rsp->level[i - 1] +
2977                                              j / rsp->levelspread[i - 1];
2978                        }
2979                        rnp->level = i;
2980                        INIT_LIST_HEAD(&rnp->blkd_tasks);
2981                }
2982        }
2983
2984        rsp->rda = rda;
2985        init_waitqueue_head(&rsp->gp_wq);
2986        rnp = rsp->level[rcu_num_lvls - 1];
2987        for_each_possible_cpu(i) {
2988                while (i > rnp->grphi)
2989                        rnp++;
2990                per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2991                rcu_boot_init_percpu_data(i, rsp);
2992        }
2993        list_add(&rsp->flavors, &rcu_struct_flavors);
2994}
2995
2996/*
2997 * Compute the rcu_node tree geometry from kernel parameters.  This cannot
2998 * replace the definitions in rcutree.h because those are needed to size
2999 * the ->node array in the rcu_state structure.
3000 */
3001static void __init rcu_init_geometry(void)
3002{
3003        int i;
3004        int j;
3005        int n = nr_cpu_ids;
3006        int rcu_capacity[MAX_RCU_LVLS + 1];
3007
3008        /* If the compile-time values are accurate, just leave. */
3009        if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3010            nr_cpu_ids == NR_CPUS)
3011                return;
3012
3013        /*
3014         * Compute number of nodes that can be handled an rcu_node tree
3015         * with the given number of levels.  Setting rcu_capacity[0] makes
3016         * some of the arithmetic easier.
3017         */
3018        rcu_capacity[0] = 1;
3019        rcu_capacity[1] = rcu_fanout_leaf;
3020        for (i = 2; i <= MAX_RCU_LVLS; i++)
3021                rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3022
3023        /*
3024         * The boot-time rcu_fanout_leaf parameter is only permitted
3025         * to increase the leaf-level fanout, not decrease it.  Of course,
3026         * the leaf-level fanout cannot exceed the number of bits in
3027         * the rcu_node masks.  Finally, the tree must be able to accommodate
3028         * the configured number of CPUs.  Complain and fall back to the
3029         * compile-time values if these limits are exceeded.
3030         */
3031        if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3032            rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3033            n > rcu_capacity[MAX_RCU_LVLS]) {
3034                WARN_ON(1);
3035                return;
3036        }
3037
3038        /* Calculate the number of rcu_nodes at each level of the tree. */
3039        for (i = 1; i <= MAX_RCU_LVLS; i++)
3040                if (n <= rcu_capacity[i]) {
3041                        for (j = 0; j <= i; j++)
3042                                num_rcu_lvl[j] =
3043                                        DIV_ROUND_UP(n, rcu_capacity[i - j]);
3044                        rcu_num_lvls = i;
3045                        for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3046                                num_rcu_lvl[j] = 0;
3047                        break;
3048                }
3049
3050        /* Calculate the total number of rcu_node structures. */
3051        rcu_num_nodes = 0;
3052        for (i = 0; i <= MAX_RCU_LVLS; i++)
3053                rcu_num_nodes += num_rcu_lvl[i];
3054        rcu_num_nodes -= n;
3055}
3056
3057void __init rcu_init(void)
3058{
3059        int cpu;
3060
3061        rcu_bootup_announce();
3062        rcu_init_geometry();
3063        rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3064        rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3065        __rcu_init_preempt();
3066        rcu_init_nocb();
3067         open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3068
3069        /*
3070         * We don't need protection against CPU-hotplug here because
3071         * this is called early in boot, before either interrupts
3072         * or the scheduler are operational.
3073         */
3074        cpu_notifier(rcu_cpu_notify, 0);
3075        for_each_online_cpu(cpu)
3076                rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3077        check_cpu_stall_init();
3078}
3079
3080#include "rcutree_plugin.h"
3081
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.