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
  56#include "rcutree.h"
  57#include <trace/events/rcu.h>
  58
  59#include "rcu.h"
  60
  61/* Data structures. */
  62
  63static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
  64
  65#define RCU_STATE_INITIALIZER(sname, cr) { \
  66        .level = { &sname##_state.node[0] }, \
  67        .call = cr, \
  68        .fqs_state = RCU_GP_IDLE, \
  69        .gpnum = -300, \
  70        .completed = -300, \
  71        .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.onofflock), \
  72        .orphan_nxttail = &sname##_state.orphan_nxtlist, \
  73        .orphan_donetail = &sname##_state.orphan_donelist, \
  74        .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
  75        .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.fqslock), \
  76        .name = #sname, \
  77}
  78
  79struct rcu_state rcu_sched_state =
  80        RCU_STATE_INITIALIZER(rcu_sched, call_rcu_sched);
  81DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
  82
  83struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh, call_rcu_bh);
  84DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
  85
  86static struct rcu_state *rcu_state;
  87LIST_HEAD(rcu_struct_flavors);
  88
  89/* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
  90static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
  91module_param(rcu_fanout_leaf, int, 0);
  92int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
  93static int num_rcu_lvl[] = {  /* Number of rcu_nodes at specified level. */
  94        NUM_RCU_LVL_0,
  95        NUM_RCU_LVL_1,
  96        NUM_RCU_LVL_2,
  97        NUM_RCU_LVL_3,
  98        NUM_RCU_LVL_4,
  99};
 100int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
 101
 102/*
 103 * The rcu_scheduler_active variable transitions from zero to one just
 104 * before the first task is spawned.  So when this variable is zero, RCU
 105 * can assume that there is but one task, allowing RCU to (for example)
 106 * optimized synchronize_sched() to a simple barrier().  When this variable
 107 * is one, RCU must actually do all the hard work required to detect real
 108 * grace periods.  This variable is also used to suppress boot-time false
 109 * positives from lockdep-RCU error checking.
 110 */
 111int rcu_scheduler_active __read_mostly;
 112EXPORT_SYMBOL_GPL(rcu_scheduler_active);
 113
 114/*
 115 * The rcu_scheduler_fully_active variable transitions from zero to one
 116 * during the early_initcall() processing, which is after the scheduler
 117 * is capable of creating new tasks.  So RCU processing (for example,
 118 * creating tasks for RCU priority boosting) must be delayed until after
 119 * rcu_scheduler_fully_active transitions from zero to one.  We also
 120 * currently delay invocation of any RCU callbacks until after this point.
 121 *
 122 * It might later prove better for people registering RCU callbacks during
 123 * early boot to take responsibility for these callbacks, but one step at
 124 * a time.
 125 */
 126static int rcu_scheduler_fully_active __read_mostly;
 127
 128#ifdef CONFIG_RCU_BOOST
 129
 130/*
 131 * Control variables for per-CPU and per-rcu_node kthreads.  These
 132 * handle all flavors of RCU.
 133 */
 134static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
 135DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
 136DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu);
 137DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
 138DEFINE_PER_CPU(char, rcu_cpu_has_work);
 139
 140#endif /* #ifdef CONFIG_RCU_BOOST */
 141
 142static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
 143static void invoke_rcu_core(void);
 144static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
 145
 146/*
 147 * Track the rcutorture test sequence number and the update version
 148 * number within a given test.  The rcutorture_testseq is incremented
 149 * on every rcutorture module load and unload, so has an odd value
 150 * when a test is running.  The rcutorture_vernum is set to zero
 151 * when rcutorture starts and is incremented on each rcutorture update.
 152 * These variables enable correlating rcutorture output with the
 153 * RCU tracing information.
 154 */
 155unsigned long rcutorture_testseq;
 156unsigned long rcutorture_vernum;
 157
 158/*
 159 * Return true if an RCU grace period is in progress.  The ACCESS_ONCE()s
 160 * permit this function to be invoked without holding the root rcu_node
 161 * structure's ->lock, but of course results can be subject to change.
 162 */
 163static int rcu_gp_in_progress(struct rcu_state *rsp)
 164{
 165        return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
 166}
 167
 168/*
 169 * Note a quiescent state.  Because we do not need to know
 170 * how many quiescent states passed, just if there was at least
 171 * one since the start of the grace period, this just sets a flag.
 172 * The caller must have disabled preemption.
 173 */
 174void rcu_sched_qs(int cpu)
 175{
 176        struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
 177
 178        rdp->passed_quiesce_gpnum = rdp->gpnum;
 179        barrier();
 180        if (rdp->passed_quiesce == 0)
 181                trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
 182        rdp->passed_quiesce = 1;
 183}
 184
 185void rcu_bh_qs(int cpu)
 186{
 187        struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
 188
 189        rdp->passed_quiesce_gpnum = rdp->gpnum;
 190        barrier();
 191        if (rdp->passed_quiesce == 0)
 192                trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
 193        rdp->passed_quiesce = 1;
 194}
 195
 196/*
 197 * Note a context switch.  This is a quiescent state for RCU-sched,
 198 * and requires special handling for preemptible RCU.
 199 * The caller must have disabled preemption.
 200 */
 201void rcu_note_context_switch(int cpu)
 202{
 203        trace_rcu_utilization("Start context switch");
 204        rcu_sched_qs(cpu);
 205        rcu_preempt_note_context_switch(cpu);
 206        trace_rcu_utilization("End context switch");
 207}
 208EXPORT_SYMBOL_GPL(rcu_note_context_switch);
 209
 210DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
 211        .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
 212        .dynticks = ATOMIC_INIT(1),
 213};
 214
 215static int blimit = 10;         /* Maximum callbacks per rcu_do_batch. */
 216static int qhimark = 10000;     /* If this many pending, ignore blimit. */
 217static int qlowmark = 100;      /* Once only this many pending, use blimit. */
 218
 219module_param(blimit, int, 0);
 220module_param(qhimark, int, 0);
 221module_param(qlowmark, int, 0);
 222
 223int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
 224int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
 225
 226module_param(rcu_cpu_stall_suppress, int, 0644);
 227module_param(rcu_cpu_stall_timeout, int, 0644);
 228
 229static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
 230static int rcu_pending(int cpu);
 231
 232/*
 233 * Return the number of RCU-sched batches processed thus far for debug & stats.
 234 */
 235long rcu_batches_completed_sched(void)
 236{
 237        return rcu_sched_state.completed;
 238}
 239EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
 240
 241/*
 242 * Return the number of RCU BH batches processed thus far for debug & stats.
 243 */
 244long rcu_batches_completed_bh(void)
 245{
 246        return rcu_bh_state.completed;
 247}
 248EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
 249
 250/*
 251 * Force a quiescent state for RCU BH.
 252 */
 253void rcu_bh_force_quiescent_state(void)
 254{
 255        force_quiescent_state(&rcu_bh_state, 0);
 256}
 257EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
 258
 259/*
 260 * Record the number of times rcutorture tests have been initiated and
 261 * terminated.  This information allows the debugfs tracing stats to be
 262 * correlated to the rcutorture messages, even when the rcutorture module
 263 * is being repeatedly loaded and unloaded.  In other words, we cannot
 264 * store this state in rcutorture itself.
 265 */
 266void rcutorture_record_test_transition(void)
 267{
 268        rcutorture_testseq++;
 269        rcutorture_vernum = 0;
 270}
 271EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
 272
 273/*
 274 * Record the number of writer passes through the current rcutorture test.
 275 * This is also used to correlate debugfs tracing stats with the rcutorture
 276 * messages.
 277 */
 278void rcutorture_record_progress(unsigned long vernum)
 279{
 280        rcutorture_vernum++;
 281}
 282EXPORT_SYMBOL_GPL(rcutorture_record_progress);
 283
 284/*
 285 * Force a quiescent state for RCU-sched.
 286 */
 287void rcu_sched_force_quiescent_state(void)
 288{
 289        force_quiescent_state(&rcu_sched_state, 0);
 290}
 291EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
 292
 293/*
 294 * Does the CPU have callbacks ready to be invoked?
 295 */
 296static int
 297cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
 298{
 299        return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
 300}
 301
 302/*
 303 * Does the current CPU require a yet-as-unscheduled grace period?
 304 */
 305static int
 306cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
 307{
 308        return *rdp->nxttail[RCU_DONE_TAIL +
 309                             ACCESS_ONCE(rsp->completed) != rdp->completed] &&
 310               !rcu_gp_in_progress(rsp);
 311}
 312
 313/*
 314 * Return the root node of the specified rcu_state structure.
 315 */
 316static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
 317{
 318        return &rsp->node[0];
 319}
 320
 321/*
 322 * If the specified CPU is offline, tell the caller that it is in
 323 * a quiescent state.  Otherwise, whack it with a reschedule IPI.
 324 * Grace periods can end up waiting on an offline CPU when that
 325 * CPU is in the process of coming online -- it will be added to the
 326 * rcu_node bitmasks before it actually makes it online.  The same thing
 327 * can happen while a CPU is in the process of coming online.  Because this
 328 * race is quite rare, we check for it after detecting that the grace
 329 * period has been delayed rather than checking each and every CPU
 330 * each and every time we start a new grace period.
 331 */
 332static int rcu_implicit_offline_qs(struct rcu_data *rdp)
 333{
 334        /*
 335         * If the CPU is offline for more than a jiffy, it is in a quiescent
 336         * state.  We can trust its state not to change because interrupts
 337         * are disabled.  The reason for the jiffy's worth of slack is to
 338         * handle CPUs initializing on the way up and finding their way
 339         * to the idle loop on the way down.
 340         */
 341        if (cpu_is_offline(rdp->cpu) &&
 342            ULONG_CMP_LT(rdp->rsp->gp_start + 2, jiffies)) {
 343                trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
 344                rdp->offline_fqs++;
 345                return 1;
 346        }
 347        return 0;
 348}
 349
 350/*
 351 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
 352 *
 353 * If the new value of the ->dynticks_nesting counter now is zero,
 354 * we really have entered idle, and must do the appropriate accounting.
 355 * The caller must have disabled interrupts.
 356 */
 357static void rcu_idle_enter_common(struct rcu_dynticks *rdtp, long long oldval)
 358{
 359        trace_rcu_dyntick("Start", oldval, 0);
 360        if (!is_idle_task(current)) {
 361                struct task_struct *idle = idle_task(smp_processor_id());
 362
 363                trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
 364                ftrace_dump(DUMP_ORIG);
 365                WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
 366                          current->pid, current->comm,
 367                          idle->pid, idle->comm); /* must be idle task! */
 368        }
 369        rcu_prepare_for_idle(smp_processor_id());
 370        /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
 371        smp_mb__before_atomic_inc();  /* See above. */
 372        atomic_inc(&rdtp->dynticks);
 373        smp_mb__after_atomic_inc();  /* Force ordering with next sojourn. */
 374        WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
 375
 376        /*
 377         * The idle task is not permitted to enter the idle loop while
 378         * in an RCU read-side critical section.
 379         */
 380        rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
 381                           "Illegal idle entry in RCU read-side critical section.");
 382        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
 383                           "Illegal idle entry in RCU-bh read-side critical section.");
 384        rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
 385                           "Illegal idle entry in RCU-sched read-side critical section.");
 386}
 387
 388/**
 389 * rcu_idle_enter - inform RCU that current CPU is entering idle
 390 *
 391 * Enter idle mode, in other words, -leave- the mode in which RCU
 392 * read-side critical sections can occur.  (Though RCU read-side
 393 * critical sections can occur in irq handlers in idle, a possibility
 394 * handled by irq_enter() and irq_exit().)
 395 *
 396 * We crowbar the ->dynticks_nesting field to zero to allow for
 397 * the possibility of usermode upcalls having messed up our count
 398 * of interrupt nesting level during the prior busy period.
 399 */
 400void rcu_idle_enter(void)
 401{
 402        unsigned long flags;
 403        long long oldval;
 404        struct rcu_dynticks *rdtp;
 405
 406        local_irq_save(flags);
 407        rdtp = &__get_cpu_var(rcu_dynticks);
 408        oldval = rdtp->dynticks_nesting;
 409        WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
 410        if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
 411                rdtp->dynticks_nesting = 0;
 412        else
 413                rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
 414        rcu_idle_enter_common(rdtp, oldval);
 415        local_irq_restore(flags);
 416}
 417EXPORT_SYMBOL_GPL(rcu_idle_enter);
 418
 419/**
 420 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
 421 *
 422 * Exit from an interrupt handler, which might possibly result in entering
 423 * idle mode, in other words, leaving the mode in which read-side critical
 424 * sections can occur.
 425 *
 426 * This code assumes that the idle loop never does anything that might
 427 * result in unbalanced calls to irq_enter() and irq_exit().  If your
 428 * architecture violates this assumption, RCU will give you what you
 429 * deserve, good and hard.  But very infrequently and irreproducibly.
 430 *
 431 * Use things like work queues to work around this limitation.
 432 *
 433 * You have been warned.
 434 */
 435void rcu_irq_exit(void)
 436{
 437        unsigned long flags;
 438        long long oldval;
 439        struct rcu_dynticks *rdtp;
 440
 441        local_irq_save(flags);
 442        rdtp = &__get_cpu_var(rcu_dynticks);
 443        oldval = rdtp->dynticks_nesting;
 444        rdtp->dynticks_nesting--;
 445        WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
 446        if (rdtp->dynticks_nesting)
 447                trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
 448        else
 449                rcu_idle_enter_common(rdtp, oldval);
 450        local_irq_restore(flags);
 451}
 452
 453/*
 454 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
 455 *
 456 * If the new value of the ->dynticks_nesting counter was previously zero,
 457 * we really have exited idle, and must do the appropriate accounting.
 458 * The caller must have disabled interrupts.
 459 */
 460static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval)
 461{
 462        smp_mb__before_atomic_inc();  /* Force ordering w/previous sojourn. */
 463        atomic_inc(&rdtp->dynticks);
 464        /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
 465        smp_mb__after_atomic_inc();  /* See above. */
 466        WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
 467        rcu_cleanup_after_idle(smp_processor_id());
 468        trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
 469        if (!is_idle_task(current)) {
 470                struct task_struct *idle = idle_task(smp_processor_id());
 471
 472                trace_rcu_dyntick("Error on exit: not idle task",
 473                                  oldval, rdtp->dynticks_nesting);
 474                ftrace_dump(DUMP_ORIG);
 475                WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
 476                          current->pid, current->comm,
 477                          idle->pid, idle->comm); /* must be idle task! */
 478        }
 479}
 480
 481/**
 482 * rcu_idle_exit - inform RCU that current CPU is leaving idle
 483 *
 484 * Exit idle mode, in other words, -enter- the mode in which RCU
 485 * read-side critical sections can occur.
 486 *
 487 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
 488 * allow for the possibility of usermode upcalls messing up our count
 489 * of interrupt nesting level during the busy period that is just
 490 * now starting.
 491 */
 492void rcu_idle_exit(void)
 493{
 494        unsigned long flags;
 495        struct rcu_dynticks *rdtp;
 496        long long oldval;
 497
 498        local_irq_save(flags);
 499        rdtp = &__get_cpu_var(rcu_dynticks);
 500        oldval = rdtp->dynticks_nesting;
 501        WARN_ON_ONCE(oldval < 0);
 502        if (oldval & DYNTICK_TASK_NEST_MASK)
 503                rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
 504        else
 505                rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
 506        rcu_idle_exit_common(rdtp, oldval);
 507        local_irq_restore(flags);
 508}
 509EXPORT_SYMBOL_GPL(rcu_idle_exit);
 510
 511/**
 512 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
 513 *
 514 * Enter an interrupt handler, which might possibly result in exiting
 515 * idle mode, in other words, entering the mode in which read-side critical
 516 * sections can occur.
 517 *
 518 * Note that the Linux kernel is fully capable of entering an interrupt
 519 * handler that it never exits, for example when doing upcalls to
 520 * user mode!  This code assumes that the idle loop never does upcalls to
 521 * user mode.  If your architecture does do upcalls from the idle loop (or
 522 * does anything else that results in unbalanced calls to the irq_enter()
 523 * and irq_exit() functions), RCU will give you what you deserve, good
 524 * and hard.  But very infrequently and irreproducibly.
 525 *
 526 * Use things like work queues to work around this limitation.
 527 *
 528 * You have been warned.
 529 */
 530void rcu_irq_enter(void)
 531{
 532        unsigned long flags;
 533        struct rcu_dynticks *rdtp;
 534        long long oldval;
 535
 536        local_irq_save(flags);
 537        rdtp = &__get_cpu_var(rcu_dynticks);
 538        oldval = rdtp->dynticks_nesting;
 539        rdtp->dynticks_nesting++;
 540        WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
 541        if (oldval)
 542                trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
 543        else
 544                rcu_idle_exit_common(rdtp, oldval);
 545        local_irq_restore(flags);
 546}
 547
 548/**
 549 * rcu_nmi_enter - inform RCU of entry to NMI context
 550 *
 551 * If the CPU was idle with dynamic ticks active, and there is no
 552 * irq handler running, this updates rdtp->dynticks_nmi to let the
 553 * RCU grace-period handling know that the CPU is active.
 554 */
 555void rcu_nmi_enter(void)
 556{
 557        struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
 558
 559        if (rdtp->dynticks_nmi_nesting == 0 &&
 560            (atomic_read(&rdtp->dynticks) & 0x1))
 561                return;
 562        rdtp->dynticks_nmi_nesting++;
 563        smp_mb__before_atomic_inc();  /* Force delay from prior write. */
 564        atomic_inc(&rdtp->dynticks);
 565        /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
 566        smp_mb__after_atomic_inc();  /* See above. */
 567        WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
 568}
 569
 570/**
 571 * rcu_nmi_exit - inform RCU of exit from NMI context
 572 *
 573 * If the CPU was idle with dynamic ticks active, and there is no
 574 * irq handler running, this updates rdtp->dynticks_nmi to let the
 575 * RCU grace-period handling know that the CPU is no longer active.
 576 */
 577void rcu_nmi_exit(void)
 578{
 579        struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
 580
 581        if (rdtp->dynticks_nmi_nesting == 0 ||
 582            --rdtp->dynticks_nmi_nesting != 0)
 583                return;
 584        /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
 585        smp_mb__before_atomic_inc();  /* See above. */
 586        atomic_inc(&rdtp->dynticks);
 587        smp_mb__after_atomic_inc();  /* Force delay to next write. */
 588        WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
 589}
 590
 591/**
 592 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
 593 *
 594 * If the current CPU is in its idle loop and is neither in an interrupt
 595 * or NMI handler, return true.
 596 */
 597int rcu_is_cpu_idle(void)
 598{
 599        int ret;
 600
 601        preempt_disable();
 602        ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
 603        preempt_enable();
 604        return ret;
 605}
 606EXPORT_SYMBOL(rcu_is_cpu_idle);
 607
 608#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
 609
 610/*
 611 * Is the current CPU online?  Disable preemption to avoid false positives
 612 * that could otherwise happen due to the current CPU number being sampled,
 613 * this task being preempted, its old CPU being taken offline, resuming
 614 * on some other CPU, then determining that its old CPU is now offline.
 615 * It is OK to use RCU on an offline processor during initial boot, hence
 616 * the check for rcu_scheduler_fully_active.  Note also that it is OK
 617 * for a CPU coming online to use RCU for one jiffy prior to marking itself
 618 * online in the cpu_online_mask.  Similarly, it is OK for a CPU going
 619 * offline to continue to use RCU for one jiffy after marking itself
 620 * offline in the cpu_online_mask.  This leniency is necessary given the
 621 * non-atomic nature of the online and offline processing, for example,
 622 * the fact that a CPU enters the scheduler after completing the CPU_DYING
 623 * notifiers.
 624 *
 625 * This is also why RCU internally marks CPUs online during the
 626 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
 627 *
 628 * Disable checking if in an NMI handler because we cannot safely report
 629 * errors from NMI handlers anyway.
 630 */
 631bool rcu_lockdep_current_cpu_online(void)
 632{
 633        struct rcu_data *rdp;
 634        struct rcu_node *rnp;
 635        bool ret;
 636
 637        if (in_nmi())
 638                return 1;
 639        preempt_disable();
 640        rdp = &__get_cpu_var(rcu_sched_data);
 641        rnp = rdp->mynode;
 642        ret = (rdp->grpmask & rnp->qsmaskinit) ||
 643              !rcu_scheduler_fully_active;
 644        preempt_enable();
 645        return ret;
 646}
 647EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
 648
 649#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
 650
 651/**
 652 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
 653 *
 654 * If the current CPU is idle or running at a first-level (not nested)
 655 * interrupt from idle, return true.  The caller must have at least
 656 * disabled preemption.
 657 */
 658int rcu_is_cpu_rrupt_from_idle(void)
 659{
 660        return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
 661}
 662
 663/*
 664 * Snapshot the specified CPU's dynticks counter so that we can later
 665 * credit them with an implicit quiescent state.  Return 1 if this CPU
 666 * is in dynticks idle mode, which is an extended quiescent state.
 667 */
 668static int dyntick_save_progress_counter(struct rcu_data *rdp)
 669{
 670        rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
 671        return (rdp->dynticks_snap & 0x1) == 0;
 672}
 673
 674/*
 675 * Return true if the specified CPU has passed through a quiescent
 676 * state by virtue of being in or having passed through an dynticks
 677 * idle state since the last call to dyntick_save_progress_counter()
 678 * for this same CPU.
 679 */
 680static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
 681{
 682        unsigned int curr;
 683        unsigned int snap;
 684
 685        curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
 686        snap = (unsigned int)rdp->dynticks_snap;
 687
 688        /*
 689         * If the CPU passed through or entered a dynticks idle phase with
 690         * no active irq/NMI handlers, then we can safely pretend that the CPU
 691         * already acknowledged the request to pass through a quiescent
 692         * state.  Either way, that CPU cannot possibly be in an RCU
 693         * read-side critical section that started before the beginning
 694         * of the current RCU grace period.
 695         */
 696        if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
 697                trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
 698                rdp->dynticks_fqs++;
 699                return 1;
 700        }
 701
 702        /* Go check for the CPU being offline. */
 703        return rcu_implicit_offline_qs(rdp);
 704}
 705
 706static int jiffies_till_stall_check(void)
 707{
 708        int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
 709
 710        /*
 711         * Limit check must be consistent with the Kconfig limits
 712         * for CONFIG_RCU_CPU_STALL_TIMEOUT.
 713         */
 714        if (till_stall_check < 3) {
 715                ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
 716                till_stall_check = 3;
 717        } else if (till_stall_check > 300) {
 718                ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
 719                till_stall_check = 300;
 720        }
 721        return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
 722}
 723
 724static void record_gp_stall_check_time(struct rcu_state *rsp)
 725{
 726        rsp->gp_start = jiffies;
 727        rsp->jiffies_stall = jiffies + jiffies_till_stall_check();
 728}
 729
 730static void print_other_cpu_stall(struct rcu_state *rsp)
 731{
 732        int cpu;
 733        long delta;
 734        unsigned long flags;
 735        int ndetected = 0;
 736        struct rcu_node *rnp = rcu_get_root(rsp);
 737
 738        /* Only let one CPU complain about others per time interval. */
 739
 740        raw_spin_lock_irqsave(&rnp->lock, flags);
 741        delta = jiffies - rsp->jiffies_stall;
 742        if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
 743                raw_spin_unlock_irqrestore(&rnp->lock, flags);
 744                return;
 745        }
 746        rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3;
 747        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 748
 749        /*
 750         * OK, time to rat on our buddy...
 751         * See Documentation/RCU/stallwarn.txt for info on how to debug
 752         * RCU CPU stall warnings.
 753         */
 754        printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
 755               rsp->name);
 756        print_cpu_stall_info_begin();
 757        rcu_for_each_leaf_node(rsp, rnp) {
 758                raw_spin_lock_irqsave(&rnp->lock, flags);
 759                ndetected += rcu_print_task_stall(rnp);
 760                raw_spin_unlock_irqrestore(&rnp->lock, flags);
 761                if (rnp->qsmask == 0)
 762                        continue;
 763                for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
 764                        if (rnp->qsmask & (1UL << cpu)) {
 765                                print_cpu_stall_info(rsp, rnp->grplo + cpu);
 766                                ndetected++;
 767                        }
 768        }
 769
 770        /*
 771         * Now rat on any tasks that got kicked up to the root rcu_node
 772         * due to CPU offlining.
 773         */
 774        rnp = rcu_get_root(rsp);
 775        raw_spin_lock_irqsave(&rnp->lock, flags);
 776        ndetected += rcu_print_task_stall(rnp);
 777        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 778
 779        print_cpu_stall_info_end();
 780        printk(KERN_CONT "(detected by %d, t=%ld jiffies)\n",
 781               smp_processor_id(), (long)(jiffies - rsp->gp_start));
 782        if (ndetected == 0)
 783                printk(KERN_ERR "INFO: Stall ended before state dump start\n");
 784        else if (!trigger_all_cpu_backtrace())
 785                dump_stack();
 786
 787        /* If so configured, complain about tasks blocking the grace period. */
 788
 789        rcu_print_detail_task_stall(rsp);
 790
 791        force_quiescent_state(rsp, 0);  /* Kick them all. */
 792}
 793
 794static void print_cpu_stall(struct rcu_state *rsp)
 795{
 796        unsigned long flags;
 797        struct rcu_node *rnp = rcu_get_root(rsp);
 798
 799        /*
 800         * OK, time to rat on ourselves...
 801         * See Documentation/RCU/stallwarn.txt for info on how to debug
 802         * RCU CPU stall warnings.
 803         */
 804        printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
 805        print_cpu_stall_info_begin();
 806        print_cpu_stall_info(rsp, smp_processor_id());
 807        print_cpu_stall_info_end();
 808        printk(KERN_CONT " (t=%lu jiffies)\n", jiffies - rsp->gp_start);
 809        if (!trigger_all_cpu_backtrace())
 810                dump_stack();
 811
 812        raw_spin_lock_irqsave(&rnp->lock, flags);
 813        if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
 814                rsp->jiffies_stall = jiffies +
 815                                     3 * jiffies_till_stall_check() + 3;
 816        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 817
 818        set_need_resched();  /* kick ourselves to get things going. */
 819}
 820
 821static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
 822{
 823        unsigned long j;
 824        unsigned long js;
 825        struct rcu_node *rnp;
 826
 827        if (rcu_cpu_stall_suppress)
 828                return;
 829        j = ACCESS_ONCE(jiffies);
 830        js = ACCESS_ONCE(rsp->jiffies_stall);
 831        rnp = rdp->mynode;
 832        if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
 833
 834                /* We haven't checked in, so go dump stack. */
 835                print_cpu_stall(rsp);
 836
 837        } else if (rcu_gp_in_progress(rsp) &&
 838                   ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
 839
 840                /* They had a few time units to dump stack, so complain. */
 841                print_other_cpu_stall(rsp);
 842        }
 843}
 844
 845static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
 846{
 847        rcu_cpu_stall_suppress = 1;
 848        return NOTIFY_DONE;
 849}
 850
 851/**
 852 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
 853 *
 854 * Set the stall-warning timeout way off into the future, thus preventing
 855 * any RCU CPU stall-warning messages from appearing in the current set of
 856 * RCU grace periods.
 857 *
 858 * The caller must disable hard irqs.
 859 */
 860void rcu_cpu_stall_reset(void)
 861{
 862        struct rcu_state *rsp;
 863
 864        for_each_rcu_flavor(rsp)
 865                rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
 866}
 867
 868static struct notifier_block rcu_panic_block = {
 869        .notifier_call = rcu_panic,
 870};
 871
 872static void __init check_cpu_stall_init(void)
 873{
 874        atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
 875}
 876
 877/*
 878 * Update CPU-local rcu_data state to record the newly noticed grace period.
 879 * This is used both when we started the grace period and when we notice
 880 * that someone else started the grace period.  The caller must hold the
 881 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
 882 *  and must have irqs disabled.
 883 */
 884static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
 885{
 886        if (rdp->gpnum != rnp->gpnum) {
 887                /*
 888                 * If the current grace period is waiting for this CPU,
 889                 * set up to detect a quiescent state, otherwise don't
 890                 * go looking for one.
 891                 */
 892                rdp->gpnum = rnp->gpnum;
 893                trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
 894                if (rnp->qsmask & rdp->grpmask) {
 895                        rdp->qs_pending = 1;
 896                        rdp->passed_quiesce = 0;
 897                } else {
 898                        rdp->qs_pending = 0;
 899                }
 900                zero_cpu_stall_ticks(rdp);
 901        }
 902}
 903
 904static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
 905{
 906        unsigned long flags;
 907        struct rcu_node *rnp;
 908
 909        local_irq_save(flags);
 910        rnp = rdp->mynode;
 911        if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
 912            !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
 913                local_irq_restore(flags);
 914                return;
 915        }
 916        __note_new_gpnum(rsp, rnp, rdp);
 917        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 918}
 919
 920/*
 921 * Did someone else start a new RCU grace period start since we last
 922 * checked?  Update local state appropriately if so.  Must be called
 923 * on the CPU corresponding to rdp.
 924 */
 925static int
 926check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
 927{
 928        unsigned long flags;
 929        int ret = 0;
 930
 931        local_irq_save(flags);
 932        if (rdp->gpnum != rsp->gpnum) {
 933                note_new_gpnum(rsp, rdp);
 934                ret = 1;
 935        }
 936        local_irq_restore(flags);
 937        return ret;
 938}
 939
 940/*
 941 * Initialize the specified rcu_data structure's callback list to empty.
 942 */
 943static void init_callback_list(struct rcu_data *rdp)
 944{
 945        int i;
 946
 947        rdp->nxtlist = NULL;
 948        for (i = 0; i < RCU_NEXT_SIZE; i++)
 949                rdp->nxttail[i] = &rdp->nxtlist;
 950}
 951
 952/*
 953 * Advance this CPU's callbacks, but only if the current grace period
 954 * has ended.  This may be called only from the CPU to whom the rdp
 955 * belongs.  In addition, the corresponding leaf rcu_node structure's
 956 * ->lock must be held by the caller, with irqs disabled.
 957 */
 958static void
 959__rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
 960{
 961        /* Did another grace period end? */
 962        if (rdp->completed != rnp->completed) {
 963
 964                /* Advance callbacks.  No harm if list empty. */
 965                rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
 966                rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
 967                rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
 968
 969                /* Remember that we saw this grace-period completion. */
 970                rdp->completed = rnp->completed;
 971                trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
 972
 973                /*
 974                 * If we were in an extended quiescent state, we may have
 975                 * missed some grace periods that others CPUs handled on
 976                 * our behalf. Catch up with this state to avoid noting
 977                 * spurious new grace periods.  If another grace period
 978                 * has started, then rnp->gpnum will have advanced, so
 979                 * we will detect this later on.
 980                 */
 981                if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
 982                        rdp->gpnum = rdp->completed;
 983
 984                /*
 985                 * If RCU does not need a quiescent state from this CPU,
 986                 * then make sure that this CPU doesn't go looking for one.
 987                 */
 988                if ((rnp->qsmask & rdp->grpmask) == 0)
 989                        rdp->qs_pending = 0;
 990        }
 991}
 992
 993/*
 994 * Advance this CPU's callbacks, but only if the current grace period
 995 * has ended.  This may be called only from the CPU to whom the rdp
 996 * belongs.
 997 */
 998static void
 999rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
1000{
1001        unsigned long flags;
1002        struct rcu_node *rnp;
1003
1004        local_irq_save(flags);
1005        rnp = rdp->mynode;
1006        if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
1007            !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1008                local_irq_restore(flags);
1009                return;
1010        }
1011        __rcu_process_gp_end(rsp, rnp, rdp);
1012        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1013}
1014
1015/*
1016 * Do per-CPU grace-period initialization for running CPU.  The caller
1017 * must hold the lock of the leaf rcu_node structure corresponding to
1018 * this CPU.
1019 */
1020static void
1021rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1022{
1023        /* Prior grace period ended, so advance callbacks for current CPU. */
1024        __rcu_process_gp_end(rsp, rnp, rdp);
1025
1026        /*
1027         * Because this CPU just now started the new grace period, we know
1028         * that all of its callbacks will be covered by this upcoming grace
1029         * period, even the ones that were registered arbitrarily recently.
1030         * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
1031         *
1032         * Other CPUs cannot be sure exactly when the grace period started.
1033         * Therefore, their recently registered callbacks must pass through
1034         * an additional RCU_NEXT_READY stage, so that they will be handled
1035         * by the next RCU grace period.
1036         */
1037        rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1038        rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1039
1040        /* Set state so that this CPU will detect the next quiescent state. */
1041        __note_new_gpnum(rsp, rnp, rdp);
1042}
1043
1044/*
1045 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1046 * in preparation for detecting the next grace period.  The caller must hold
1047 * the root node's ->lock, which is released before return.  Hard irqs must
1048 * be disabled.
1049 *
1050 * Note that it is legal for a dying CPU (which is marked as offline) to
1051 * invoke this function.  This can happen when the dying CPU reports its
1052 * quiescent state.
1053 */
1054static void
1055rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1056        __releases(rcu_get_root(rsp)->lock)
1057{
1058        struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1059        struct rcu_node *rnp = rcu_get_root(rsp);
1060
1061        if (!rcu_scheduler_fully_active ||
1062            !cpu_needs_another_gp(rsp, rdp)) {
1063                /*
1064                 * Either the scheduler hasn't yet spawned the first
1065                 * non-idle task or this CPU does not need another
1066                 * grace period.  Either way, don't start a new grace
1067                 * period.
1068                 */
1069                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1070                return;
1071        }
1072
1073        if (rsp->fqs_active) {
1074                /*
1075                 * This CPU needs a grace period, but force_quiescent_state()
1076                 * is running.  Tell it to start one on this CPU's behalf.
1077                 */
1078                rsp->fqs_need_gp = 1;
1079                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1080                return;
1081        }
1082
1083        /* Advance to a new grace period and initialize state. */
1084        rsp->gpnum++;
1085        trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1086        WARN_ON_ONCE(rsp->fqs_state == RCU_GP_INIT);
1087        rsp->fqs_state = RCU_GP_INIT; /* Hold off force_quiescent_state. */
1088        rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1089        record_gp_stall_check_time(rsp);
1090        raw_spin_unlock(&rnp->lock);  /* leave irqs disabled. */
1091
1092        /* Exclude any concurrent CPU-hotplug operations. */
1093        raw_spin_lock(&rsp->onofflock);  /* irqs already disabled. */
1094
1095        /*
1096         * Set the quiescent-state-needed bits in all the rcu_node
1097         * structures for all currently online CPUs in breadth-first
1098         * order, starting from the root rcu_node structure.  This
1099         * operation relies on the layout of the hierarchy within the
1100         * rsp->node[] array.  Note that other CPUs will access only
1101         * the leaves of the hierarchy, which still indicate that no
1102         * grace period is in progress, at least until the corresponding
1103         * leaf node has been initialized.  In addition, we have excluded
1104         * CPU-hotplug operations.
1105         *
1106         * Note that the grace period cannot complete until we finish
1107         * the initialization process, as there will be at least one
1108         * qsmask bit set in the root node until that time, namely the
1109         * one corresponding to this CPU, due to the fact that we have
1110         * irqs disabled.
1111         */
1112        rcu_for_each_node_breadth_first(rsp, rnp) {
1113                raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
1114                rcu_preempt_check_blocked_tasks(rnp);
1115                rnp->qsmask = rnp->qsmaskinit;
1116                rnp->gpnum = rsp->gpnum;
1117                rnp->completed = rsp->completed;
1118                if (rnp == rdp->mynode)
1119                        rcu_start_gp_per_cpu(rsp, rnp, rdp);
1120                rcu_preempt_boost_start_gp(rnp);
1121                trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1122                                            rnp->level, rnp->grplo,
1123                                            rnp->grphi, rnp->qsmask);
1124                raw_spin_unlock(&rnp->lock);    /* irqs remain disabled. */
1125        }
1126
1127        rnp = rcu_get_root(rsp);
1128        raw_spin_lock(&rnp->lock);              /* irqs already disabled. */
1129        rsp->fqs_state = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
1130        raw_spin_unlock(&rnp->lock);            /* irqs remain disabled. */
1131        raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1132}
1133
1134/*
1135 * Report a full set of quiescent states to the specified rcu_state
1136 * data structure.  This involves cleaning up after the prior grace
1137 * period and letting rcu_start_gp() start up the next grace period
1138 * if one is needed.  Note that the caller must hold rnp->lock, as
1139 * required by rcu_start_gp(), which will release it.
1140 */
1141static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1142        __releases(rcu_get_root(rsp)->lock)
1143{
1144        unsigned long gp_duration;
1145        struct rcu_node *rnp = rcu_get_root(rsp);
1146        struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1147
1148        WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1149
1150        /*
1151         * Ensure that all grace-period and pre-grace-period activity
1152         * is seen before the assignment to rsp->completed.
1153         */
1154        smp_mb(); /* See above block comment. */
1155        gp_duration = jiffies - rsp->gp_start;
1156        if (gp_duration > rsp->gp_max)
1157                rsp->gp_max = gp_duration;
1158
1159        /*
1160         * We know the grace period is complete, but to everyone else
1161         * it appears to still be ongoing.  But it is also the case
1162         * that to everyone else it looks like there is nothing that
1163         * they can do to advance the grace period.  It is therefore
1164         * safe for us to drop the lock in order to mark the grace
1165         * period as completed in all of the rcu_node structures.
1166         *
1167         * But if this CPU needs another grace period, it will take
1168         * care of this while initializing the next grace period.
1169         * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1170         * because the callbacks have not yet been advanced: Those
1171         * callbacks are waiting on the grace period that just now
1172         * completed.
1173         */
1174        if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) {
1175                raw_spin_unlock(&rnp->lock);     /* irqs remain disabled. */
1176
1177                /*
1178                 * Propagate new ->completed value to rcu_node structures
1179                 * so that other CPUs don't have to wait until the start
1180                 * of the next grace period to process their callbacks.
1181                 */
1182                rcu_for_each_node_breadth_first(rsp, rnp) {
1183                        raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1184                        rnp->completed = rsp->gpnum;
1185                        raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1186                }
1187                rnp = rcu_get_root(rsp);
1188                raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1189        }
1190
1191        rsp->completed = rsp->gpnum;  /* Declare the grace period complete. */
1192        trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1193        rsp->fqs_state = RCU_GP_IDLE;
1194        rcu_start_gp(rsp, flags);  /* releases root node's rnp->lock. */
1195}
1196
1197/*
1198 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1199 * Allows quiescent states for a group of CPUs to be reported at one go
1200 * to the specified rcu_node structure, though all the CPUs in the group
1201 * must be represented by the same rcu_node structure (which need not be
1202 * a leaf rcu_node structure, though it often will be).  That structure's
1203 * lock must be held upon entry, and it is released before return.
1204 */
1205static void
1206rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1207                  struct rcu_node *rnp, unsigned long flags)
1208        __releases(rnp->lock)
1209{
1210        struct rcu_node *rnp_c;
1211
1212        /* Walk up the rcu_node hierarchy. */
1213        for (;;) {
1214                if (!(rnp->qsmask & mask)) {
1215
1216                        /* Our bit has already been cleared, so done. */
1217                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1218                        return;
1219                }
1220                rnp->qsmask &= ~mask;
1221                trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1222                                                 mask, rnp->qsmask, rnp->level,
1223                                                 rnp->grplo, rnp->grphi,
1224                                                 !!rnp->gp_tasks);
1225                if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1226
1227                        /* Other bits still set at this level, so done. */
1228                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1229                        return;
1230                }
1231                mask = rnp->grpmask;
1232                if (rnp->parent == NULL) {
1233
1234                        /* No more levels.  Exit loop holding root lock. */
1235
1236                        break;
1237                }
1238                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1239                rnp_c = rnp;
1240                rnp = rnp->parent;
1241                raw_spin_lock_irqsave(&rnp->lock, flags);
1242                WARN_ON_ONCE(rnp_c->qsmask);
1243        }
1244
1245        /*
1246         * Get here if we are the last CPU to pass through a quiescent
1247         * state for this grace period.  Invoke rcu_report_qs_rsp()
1248         * to clean up and start the next grace period if one is needed.
1249         */
1250        rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1251}
1252
1253/*
1254 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1255 * structure.  This must be either called from the specified CPU, or
1256 * called when the specified CPU is known to be offline (and when it is
1257 * also known that no other CPU is concurrently trying to help the offline
1258 * CPU).  The lastcomp argument is used to make sure we are still in the
1259 * grace period of interest.  We don't want to end the current grace period
1260 * based on quiescent states detected in an earlier grace period!
1261 */
1262static void
1263rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastgp)
1264{
1265        unsigned long flags;
1266        unsigned long mask;
1267        struct rcu_node *rnp;
1268
1269        rnp = rdp->mynode;
1270        raw_spin_lock_irqsave(&rnp->lock, flags);
1271        if (lastgp != rnp->gpnum || rnp->completed == rnp->gpnum) {
1272
1273                /*
1274                 * The grace period in which this quiescent state was
1275                 * recorded has ended, so don't report it upwards.
1276                 * We will instead need a new quiescent state that lies
1277                 * within the current grace period.
1278                 */
1279                rdp->passed_quiesce = 0;        /* need qs for new gp. */
1280                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1281                return;
1282        }
1283        mask = rdp->grpmask;
1284        if ((rnp->qsmask & mask) == 0) {
1285                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1286        } else {
1287                rdp->qs_pending = 0;
1288
1289                /*
1290                 * This GP can't end until cpu checks in, so all of our
1291                 * callbacks can be processed during the next GP.
1292                 */
1293                rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1294
1295                rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1296        }
1297}
1298
1299/*
1300 * Check to see if there is a new grace period of which this CPU
1301 * is not yet aware, and if so, set up local rcu_data state for it.
1302 * Otherwise, see if this CPU has just passed through its first
1303 * quiescent state for this grace period, and record that fact if so.
1304 */
1305static void
1306rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1307{
1308        /* If there is now a new grace period, record and return. */
1309        if (check_for_new_grace_period(rsp, rdp))
1310                return;
1311
1312        /*
1313         * Does this CPU still need to do its part for current grace period?
1314         * If no, return and let the other CPUs do their part as well.
1315         */
1316        if (!rdp->qs_pending)
1317                return;
1318
1319        /*
1320         * Was there a quiescent state since the beginning of the grace
1321         * period? If no, then exit and wait for the next call.
1322         */
1323        if (!rdp->passed_quiesce)
1324                return;
1325
1326        /*
1327         * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1328         * judge of that).
1329         */
1330        rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesce_gpnum);
1331}
1332
1333#ifdef CONFIG_HOTPLUG_CPU
1334
1335/*
1336 * Send the specified CPU's RCU callbacks to the orphanage.  The
1337 * specified CPU must be offline, and the caller must hold the
1338 * ->onofflock.
1339 */
1340static void
1341rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1342                          struct rcu_node *rnp, struct rcu_data *rdp)
1343{
1344        /*
1345         * Orphan the callbacks.  First adjust the counts.  This is safe
1346         * because ->onofflock excludes _rcu_barrier()'s adoption of
1347         * the callbacks, thus no memory barrier is required.
1348         */
1349        if (rdp->nxtlist != NULL) {
1350                rsp->qlen_lazy += rdp->qlen_lazy;
1351                rsp->qlen += rdp->qlen;
1352                rdp->n_cbs_orphaned += rdp->qlen;
1353                rdp->qlen_lazy = 0;
1354                ACCESS_ONCE(rdp->qlen) = 0;
1355        }
1356
1357        /*
1358         * Next, move those callbacks still needing a grace period to
1359         * the orphanage, where some other CPU will pick them up.
1360         * Some of the callbacks might have gone partway through a grace
1361         * period, but that is too bad.  They get to start over because we
1362         * cannot assume that grace periods are synchronized across CPUs.
1363         * We don't bother updating the ->nxttail[] array yet, instead
1364         * we just reset the whole thing later on.
1365         */
1366        if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1367                *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1368                rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1369                *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1370        }
1371
1372        /*
1373         * Then move the ready-to-invoke callbacks to the orphanage,
1374         * where some other CPU will pick them up.  These will not be
1375         * required to pass though another grace period: They are done.
1376         */
1377        if (rdp->nxtlist != NULL) {
1378                *rsp->orphan_donetail = rdp->nxtlist;
1379                rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1380        }
1381
1382        /* Finally, initialize the rcu_data structure's list to empty.  */
1383        init_callback_list(rdp);
1384}
1385
1386/*
1387 * Adopt the RCU callbacks from the specified rcu_state structure's
1388 * orphanage.  The caller must hold the ->onofflock.
1389 */
1390static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1391{
1392        int i;
1393        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1394
1395        /*
1396         * If there is an rcu_barrier() operation in progress, then
1397         * only the task doing that operation is permitted to adopt
1398         * callbacks.  To do otherwise breaks rcu_barrier() and friends
1399         * by causing them to fail to wait for the callbacks in the
1400         * orphanage.
1401         */
1402        if (rsp->rcu_barrier_in_progress &&
1403            rsp->rcu_barrier_in_progress != current)
1404                return;
1405
1406        /* Do the accounting first. */
1407        rdp->qlen_lazy += rsp->qlen_lazy;
1408        rdp->qlen += rsp->qlen;
1409        rdp->n_cbs_adopted += rsp->qlen;
1410        if (rsp->qlen_lazy != rsp->qlen)
1411                rcu_idle_count_callbacks_posted();
1412        rsp->qlen_lazy = 0;
1413        rsp->qlen = 0;
1414
1415        /*
1416         * We do not need a memory barrier here because the only way we
1417         * can get here if there is an rcu_barrier() in flight is if
1418         * we are the task doing the rcu_barrier().
1419         */
1420
1421        /* First adopt the ready-to-invoke callbacks. */
1422        if (rsp->orphan_donelist != NULL) {
1423                *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1424                *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1425                for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1426                        if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1427                                rdp->nxttail[i] = rsp->orphan_donetail;
1428                rsp->orphan_donelist = NULL;
1429                rsp->orphan_donetail = &rsp->orphan_donelist;
1430        }
1431
1432        /* And then adopt the callbacks that still need a grace period. */
1433        if (rsp->orphan_nxtlist != NULL) {
1434                *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1435                rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1436                rsp->orphan_nxtlist = NULL;
1437                rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1438        }
1439}
1440
1441/*
1442 * Trace the fact that this CPU is going offline.
1443 */
1444static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1445{
1446        RCU_TRACE(unsigned long mask);
1447        RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1448        RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1449
1450        RCU_TRACE(mask = rdp->grpmask);
1451        trace_rcu_grace_period(rsp->name,
1452                               rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1453                               "cpuofl");
1454}
1455
1456/*
1457 * The CPU has been completely removed, and some other CPU is reporting
1458 * this fact from process context.  Do the remainder of the cleanup,
1459 * including orphaning the outgoing CPU's RCU callbacks, and also
1460 * adopting them, if there is no _rcu_barrier() instance running.
1461 * There can only be one CPU hotplug operation at a time, so no other
1462 * CPU can be attempting to update rcu_cpu_kthread_task.
1463 */
1464static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1465{
1466        unsigned long flags;
1467        unsigned long mask;
1468        int need_report = 0;
1469        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1470        struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
1471
1472        /* Adjust any no-longer-needed kthreads. */
1473        rcu_stop_cpu_kthread(cpu);
1474        rcu_node_kthread_setaffinity(rnp, -1);
1475
1476        /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1477
1478        /* Exclude any attempts to start a new grace period. */
1479        raw_spin_lock_irqsave(&rsp->onofflock, flags);
1480
1481        /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1482        rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1483        rcu_adopt_orphan_cbs(rsp);
1484
1485        /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1486        mask = rdp->grpmask;    /* rnp->grplo is constant. */
1487        do {
1488                raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
1489                rnp->qsmaskinit &= ~mask;
1490                if (rnp->qsmaskinit != 0) {
1491                        if (rnp != rdp->mynode)
1492                                raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1493                        break;
1494                }
1495                if (rnp == rdp->mynode)
1496                        need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1497                else
1498                        raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1499                mask = rnp->grpmask;
1500                rnp = rnp->parent;
1501        } while (rnp != NULL);
1502
1503        /*
1504         * We still hold the leaf rcu_node structure lock here, and
1505         * irqs are still disabled.  The reason for this subterfuge is
1506         * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1507         * held leads to deadlock.
1508         */
1509        raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1510        rnp = rdp->mynode;
1511        if (need_report & RCU_OFL_TASKS_NORM_GP)
1512                rcu_report_unblock_qs_rnp(rnp, flags);
1513        else
1514                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1515        if (need_report & RCU_OFL_TASKS_EXP_GP)
1516                rcu_report_exp_rnp(rsp, rnp, true);
1517        WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1518                  "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1519                  cpu, rdp->qlen, rdp->nxtlist);
1520}
1521
1522#else /* #ifdef CONFIG_HOTPLUG_CPU */
1523
1524static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1525{
1526}
1527
1528static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1529{
1530}
1531
1532static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1533{
1534}
1535
1536#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1537
1538/*
1539 * Invoke any RCU callbacks that have made it to the end of their grace
1540 * period.  Thottle as specified by rdp->blimit.
1541 */
1542static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1543{
1544        unsigned long flags;
1545        struct rcu_head *next, *list, **tail;
1546        int bl, count, count_lazy, i;
1547
1548        /* If no callbacks are ready, just return.*/
1549        if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1550                trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1551                trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1552                                    need_resched(), is_idle_task(current),
1553                                    rcu_is_callbacks_kthread());
1554                return;
1555        }
1556
1557        /*
1558         * Extract the list of ready callbacks, disabling to prevent
1559         * races with call_rcu() from interrupt handlers.
1560         */
1561        local_irq_save(flags);
1562        WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1563        bl = rdp->blimit;
1564        trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1565        list = rdp->nxtlist;
1566        rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1567        *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1568        tail = rdp->nxttail[RCU_DONE_TAIL];
1569        for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1570                if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1571                        rdp->nxttail[i] = &rdp->nxtlist;
1572        local_irq_restore(flags);
1573
1574        /* Invoke callbacks. */
1575        count = count_lazy = 0;
1576        while (list) {
1577                next = list->next;
1578                prefetch(next);
1579                debug_rcu_head_unqueue(list);
1580                if (__rcu_reclaim(rsp->name, list))
1581                        count_lazy++;
1582                list = next;
1583                /* Stop only if limit reached and CPU has something to do. */
1584                if (++count >= bl &&
1585                    (need_resched() ||
1586                     (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1587                        break;
1588        }
1589
1590        local_irq_save(flags);
1591        trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1592                            is_idle_task(current),
1593                            rcu_is_callbacks_kthread());
1594
1595        /* Update count, and requeue any remaining callbacks. */
1596        if (list != NULL) {
1597                *tail = rdp->nxtlist;
1598                rdp->nxtlist = list;
1599                for (i = 0; i < RCU_NEXT_SIZE; i++)
1600                        if (&rdp->nxtlist == rdp->nxttail[i])
1601                                rdp->nxttail[i] = tail;
1602                        else
1603                                break;
1604        }
1605        smp_mb(); /* List handling before counting for rcu_barrier(). */
1606        rdp->qlen_lazy -= count_lazy;
1607        ACCESS_ONCE(rdp->qlen) -= count;
1608        rdp->n_cbs_invoked += count;
1609
1610        /* Reinstate batch limit if we have worked down the excess. */
1611        if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1612                rdp->blimit = blimit;
1613
1614        /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1615        if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1616                rdp->qlen_last_fqs_check = 0;
1617                rdp->n_force_qs_snap = rsp->n_force_qs;
1618        } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1619                rdp->qlen_last_fqs_check = rdp->qlen;
1620        WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
1621
1622        local_irq_restore(flags);
1623
1624        /* Re-invoke RCU core processing if there are callbacks remaining. */
1625        if (cpu_has_callbacks_ready_to_invoke(rdp))
1626                invoke_rcu_core();
1627}
1628
1629/*
1630 * Check to see if this CPU is in a non-context-switch quiescent state
1631 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1632 * Also schedule RCU core processing.
1633 *
1634 * This function must be called from hardirq context.  It is normally
1635 * invoked from the scheduling-clock interrupt.  If rcu_pending returns
1636 * false, there is no point in invoking rcu_check_callbacks().
1637 */
1638void rcu_check_callbacks(int cpu, int user)
1639{
1640        trace_rcu_utilization("Start scheduler-tick");
1641        increment_cpu_stall_ticks();
1642        if (user || rcu_is_cpu_rrupt_from_idle()) {
1643
1644                /*
1645                 * Get here if this CPU took its interrupt from user
1646                 * mode or from the idle loop, and if this is not a
1647                 * nested interrupt.  In this case, the CPU is in
1648                 * a quiescent state, so note it.
1649                 *
1650                 * No memory barrier is required here because both
1651                 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1652                 * variables that other CPUs neither access nor modify,
1653                 * at least not while the corresponding CPU is online.
1654                 */
1655
1656                rcu_sched_qs(cpu);
1657                rcu_bh_qs(cpu);
1658
1659        } else if (!in_softirq()) {
1660
1661                /*
1662                 * Get here if this CPU did not take its interrupt from
1663                 * softirq, in other words, if it is not interrupting
1664                 * a rcu_bh read-side critical section.  This is an _bh
1665                 * critical section, so note it.
1666                 */
1667
1668                rcu_bh_qs(cpu);
1669        }
1670        rcu_preempt_check_callbacks(cpu);
1671        if (rcu_pending(cpu))
1672                invoke_rcu_core();
1673        trace_rcu_utilization("End scheduler-tick");
1674}
1675
1676/*
1677 * Scan the leaf rcu_node structures, processing dyntick state for any that
1678 * have not yet encountered a quiescent state, using the function specified.
1679 * Also initiate boosting for any threads blocked on the root rcu_node.
1680 *
1681 * The caller must have suppressed start of new grace periods.
1682 */
1683static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1684{
1685        unsigned long bit;
1686        int cpu;
1687        unsigned long flags;
1688        unsigned long mask;
1689        struct rcu_node *rnp;
1690
1691        rcu_for_each_leaf_node(rsp, rnp) {
1692                mask = 0;
1693                raw_spin_lock_irqsave(&rnp->lock, flags);
1694                if (!rcu_gp_in_progress(rsp)) {
1695                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1696                        return;
1697                }
1698                if (rnp->qsmask == 0) {
1699                        rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1700                        continue;
1701                }
1702                cpu = rnp->grplo;
1703                bit = 1;
1704                for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1705                        if ((rnp->qsmask & bit) != 0 &&
1706                            f(per_cpu_ptr(rsp->rda, cpu)))
1707                                mask |= bit;
1708                }
1709                if (mask != 0) {
1710
1711                        /* rcu_report_qs_rnp() releases rnp->lock. */
1712                        rcu_report_qs_rnp(mask, rsp, rnp, flags);
1713                        continue;
1714                }
1715                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1716        }
1717        rnp = rcu_get_root(rsp);
1718        if (rnp->qsmask == 0) {
1719                raw_spin_lock_irqsave(&rnp->lock, flags);
1720                rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1721        }
1722}
1723
1724/*
1725 * Force quiescent states on reluctant CPUs, and also detect which
1726 * CPUs are in dyntick-idle mode.
1727 */
1728static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1729{
1730        unsigned long flags;
1731        struct rcu_node *rnp = rcu_get_root(rsp);
1732
1733        trace_rcu_utilization("Start fqs");
1734        if (!rcu_gp_in_progress(rsp)) {
1735                trace_rcu_utilization("End fqs");
1736                return;  /* No grace period in progress, nothing to force. */
1737        }
1738        if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1739                rsp->n_force_qs_lh++; /* Inexact, can lose counts.  Tough! */
1740                trace_rcu_utilization("End fqs");
1741                return; /* Someone else is already on the job. */
1742        }
1743        if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1744                goto unlock_fqs_ret; /* no emergency and done recently. */
1745        rsp->n_force_qs++;
1746        raw_spin_lock(&rnp->lock);  /* irqs already disabled */
1747        rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1748        if(!rcu_gp_in_progress(rsp)) {
1749                rsp->n_force_qs_ngp++;
1750                raw_spin_unlock(&rnp->lock);  /* irqs remain disabled */
1751                goto unlock_fqs_ret;  /* no GP in progress, time updated. */
1752        }
1753        rsp->fqs_active = 1;
1754        switch (rsp->fqs_state) {
1755        case RCU_GP_IDLE:
1756        case RCU_GP_INIT:
1757
1758                break; /* grace period idle or initializing, ignore. */
1759
1760        case RCU_SAVE_DYNTICK:
1761
1762                raw_spin_unlock(&rnp->lock);  /* irqs remain disabled */
1763
1764                /* Record dyntick-idle state. */
1765                force_qs_rnp(rsp, dyntick_save_progress_counter);
1766                raw_spin_lock(&rnp->lock);  /* irqs already disabled */
1767                if (rcu_gp_in_progress(rsp))
1768                        rsp->fqs_state = RCU_FORCE_QS;
1769                break;
1770
1771        case RCU_FORCE_QS:
1772
1773                /* Check dyntick-idle state, send IPI to laggarts. */
1774                raw_spin_unlock(&rnp->lock);  /* irqs remain disabled */
1775                force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1776
1777                /* Leave state in case more forcing is required. */
1778
1779                raw_spin_lock(&rnp->lock);  /* irqs already disabled */
1780                break;
1781        }
1782        rsp->fqs_active = 0;
1783        if (rsp->fqs_need_gp) {
1784                raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1785                rsp->fqs_need_gp = 0;
1786                rcu_start_gp(rsp, flags); /* releases rnp->lock */
1787                trace_rcu_utilization("End fqs");
1788                return;
1789        }
1790        raw_spin_unlock(&rnp->lock);  /* irqs remain disabled */
1791unlock_fqs_ret:
1792        raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1793        trace_rcu_utilization("End fqs");
1794}
1795
1796/*
1797 * This does the RCU core processing work for the specified rcu_state
1798 * and rcu_data structures.  This may be called only from the CPU to
1799 * whom the rdp belongs.
1800 */
1801static void
1802__rcu_process_callbacks(struct rcu_state *rsp)
1803{
1804        unsigned long flags;
1805        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1806
1807        WARN_ON_ONCE(rdp->beenonline == 0);
1808
1809        /*
1810         * If an RCU GP has gone long enough, go check for dyntick
1811         * idle CPUs and, if needed, send resched IPIs.
1812         */
1813        if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1814                force_quiescent_state(rsp, 1);
1815
1816        /*
1817         * Advance callbacks in response to end of earlier grace
1818         * period that some other CPU ended.
1819         */
1820        rcu_process_gp_end(rsp, rdp);
1821
1822        /* Update RCU state based on any recent quiescent states. */
1823        rcu_check_quiescent_state(rsp, rdp);
1824
1825        /* Does this CPU require a not-yet-started grace period? */
1826        if (cpu_needs_another_gp(rsp, rdp)) {
1827                raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1828                rcu_start_gp(rsp, flags);  /* releases above lock */
1829        }
1830
1831        /* If there are callbacks ready, invoke them. */
1832        if (cpu_has_callbacks_ready_to_invoke(rdp))
1833                invoke_rcu_callbacks(rsp, rdp);
1834}
1835
1836/*
1837 * Do RCU core processing for the current CPU.
1838 */
1839static void rcu_process_callbacks(struct softirq_action *unused)
1840{
1841        struct rcu_state *rsp;
1842
1843        trace_rcu_utilization("Start RCU core");
1844        for_each_rcu_flavor(rsp)
1845                __rcu_process_callbacks(rsp);
1846        trace_rcu_utilization("End RCU core");
1847}
1848
1849/*
1850 * Schedule RCU callback invocation.  If the specified type of RCU
1851 * does not support RCU priority boosting, just do a direct call,
1852 * otherwise wake up the per-CPU kernel kthread.  Note that because we
1853 * are running on the current CPU with interrupts disabled, the
1854 * rcu_cpu_kthread_task cannot disappear out from under us.
1855 */
1856static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1857{
1858        if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1859                return;
1860        if (likely(!rsp->boost)) {
1861                rcu_do_batch(rsp, rdp);
1862                return;
1863        }
1864        invoke_rcu_callbacks_kthread();
1865}
1866
1867static void invoke_rcu_core(void)
1868{
1869        raise_softirq(RCU_SOFTIRQ);
1870}
1871
1872/*
1873 * Handle any core-RCU processing required by a call_rcu() invocation.
1874 */
1875static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
1876                            struct rcu_head *head, unsigned long flags)
1877{
1878        /*
1879         * If called from an extended quiescent state, invoke the RCU
1880         * core in order to force a re-evaluation of RCU's idleness.
1881         */
1882        if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
1883                invoke_rcu_core();
1884
1885        /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
1886        if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
1887                return;
1888
1889        /*
1890         * Force the grace period if too many callbacks or too long waiting.
1891         * Enforce hysteresis, and don't invoke force_quiescent_state()
1892         * if some other CPU has recently done so.  Also, don't bother
1893         * invoking force_quiescent_state() if the newly enqueued callback
1894         * is the only one waiting for a grace period to complete.
1895         */
1896        if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1897
1898                /* Are we ignoring a completed grace period? */
1899                rcu_process_gp_end(rsp, rdp);
1900                check_for_new_grace_period(rsp, rdp);
1901
1902                /* Start a new grace period if one not already started. */
1903                if (!rcu_gp_in_progress(rsp)) {
1904                        unsigned long nestflag;
1905                        struct rcu_node *rnp_root = rcu_get_root(rsp);
1906
1907                        raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1908                        rcu_start_gp(rsp, nestflag);  /* rlses rnp_root->lock */
1909                } else {
1910                        /* Give the grace period a kick. */
1911                        rdp->blimit = LONG_MAX;
1912                        if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1913                            *rdp->nxttail[RCU_DONE_TAIL] != head)
1914                                force_quiescent_state(rsp, 0);
1915                        rdp->n_force_qs_snap = rsp->n_force_qs;
1916                        rdp->qlen_last_fqs_check = rdp->qlen;
1917                }
1918        } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1919                force_quiescent_state(rsp, 1);
1920}
1921
1922static void
1923__call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1924           struct rcu_state *rsp, bool lazy)
1925{
1926        unsigned long flags;
1927        struct rcu_data *rdp;
1928
1929        WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
1930        debug_rcu_head_queue(head);
1931        head->func = func;
1932        head->next = NULL;
1933
1934        smp_mb(); /* Ensure RCU update seen before callback registry. */
1935
1936        /*
1937         * Opportunistically note grace-period endings and beginnings.
1938         * Note that we might see a beginning right after we see an
1939         * end, but never vice versa, since this CPU has to pass through
1940         * a quiescent state betweentimes.
1941         */
1942        local_irq_save(flags);
1943        rdp = this_cpu_ptr(rsp->rda);
1944
1945        /* Add the callback to our list. */
1946        ACCESS_ONCE(rdp->qlen)++;
1947        if (lazy)
1948                rdp->qlen_lazy++;
1949        else
1950                rcu_idle_count_callbacks_posted();
1951        smp_mb();  /* Count before adding callback for rcu_barrier(). */
1952        *rdp->nxttail[RCU_NEXT_TAIL] = head;
1953        rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1954
1955        if (__is_kfree_rcu_offset((unsigned long)func))
1956                trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
1957                                         rdp->qlen_lazy, rdp->qlen);
1958        else
1959                trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
1960
1961        /* Go handle any RCU core processing required. */
1962        __call_rcu_core(rsp, rdp, head, flags);
1963        local_irq_restore(flags);
1964}
1965
1966/*
1967 * Queue an RCU-sched callback for invocation after a grace period.
1968 */
1969void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1970{
1971        __call_rcu(head, func, &rcu_sched_state, 0);
1972}
1973EXPORT_SYMBOL_GPL(call_rcu_sched);
1974
1975/*
1976 * Queue an RCU callback for invocation after a quicker grace period.
1977 */
1978void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1979{
1980        __call_rcu(head, func, &rcu_bh_state, 0);
1981}
1982EXPORT_SYMBOL_GPL(call_rcu_bh);
1983
1984/*
1985 * Because a context switch is a grace period for RCU-sched and RCU-bh,
1986 * any blocking grace-period wait automatically implies a grace period
1987 * if there is only one CPU online at any point time during execution
1988 * of either synchronize_sched() or synchronize_rcu_bh().  It is OK to
1989 * occasionally incorrectly indicate that there are multiple CPUs online
1990 * when there was in fact only one the whole time, as this just adds
1991 * some overhead: RCU still operates correctly.
1992 */
1993static inline int rcu_blocking_is_gp(void)
1994{
1995        int ret;
1996
1997        might_sleep();  /* Check for RCU read-side critical section. */
1998        preempt_disable();
1999        ret = num_online_cpus() <= 1;
2000        preempt_enable();
2001        return ret;
2002}
2003
2004/**
2005 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2006 *
2007 * Control will return to the caller some time after a full rcu-sched
2008 * grace period has elapsed, in other words after all currently executing
2009 * rcu-sched read-side critical sections have completed.   These read-side
2010 * critical sections are delimited by rcu_read_lock_sched() and
2011 * rcu_read_unlock_sched(), and may be nested.  Note that preempt_disable(),
2012 * local_irq_disable(), and so on may be used in place of
2013 * rcu_read_lock_sched().
2014 *
2015 * This means that all preempt_disable code sequences, including NMI and
2016 * hardware-interrupt handlers, in progress on entry will have completed
2017 * before this primitive returns.  However, this does not guarantee that
2018 * softirq handlers will have completed, since in some kernels, these
2019 * handlers can run in process context, and can block.
2020 *
2021 * This primitive provides the guarantees made by the (now removed)
2022 * synchronize_kernel() API.  In contrast, synchronize_rcu() only
2023 * guarantees that rcu_read_lock() sections will have completed.
2024 * In "classic RCU", these two guarantees happen to be one and
2025 * the same, but can differ in realtime RCU implementations.
2026 */
2027void synchronize_sched(void)
2028{
2029        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2030                           !lock_is_held(&rcu_lock_map) &&
2031                           !lock_is_held(&rcu_sched_lock_map),
2032                           "Illegal synchronize_sched() in RCU-sched read-side critical section");
2033        if (rcu_blocking_is_gp())
2034                return;
2035        wait_rcu_gp(call_rcu_sched);
2036}
2037EXPORT_SYMBOL_GPL(synchronize_sched);
2038
2039/**
2040 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2041 *
2042 * Control will return to the caller some time after a full rcu_bh grace
2043 * period has elapsed, in other words after all currently executing rcu_bh
2044 * read-side critical sections have completed.  RCU read-side critical
2045 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2046 * and may be nested.
2047 */
2048void synchronize_rcu_bh(void)
2049{
2050        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2051                           !lock_is_held(&rcu_lock_map) &&
2052                           !lock_is_held(&rcu_sched_lock_map),
2053                           "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2054        if (rcu_blocking_is_gp())
2055                return;
2056        wait_rcu_gp(call_rcu_bh);
2057}
2058EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2059
2060static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
2061static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
2062
2063static int synchronize_sched_expedited_cpu_stop(void *data)
2064{
2065        /*
2066         * There must be a full memory barrier on each affected CPU
2067         * between the time that try_stop_cpus() is called and the
2068         * time that it returns.
2069         *
2070         * In the current initial implementation of cpu_stop, the
2071         * above condition is already met when the control reaches
2072         * this point and the following smp_mb() is not strictly
2073         * necessary.  Do smp_mb() anyway for documentation and
2074         * robustness against future implementation changes.
2075         */
2076        smp_mb(); /* See above comment block. */
2077        return 0;
2078}
2079
2080/**
2081 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2082 *
2083 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2084 * approach to force the grace period to end quickly.  This consumes
2085 * significant time on all CPUs and is unfriendly to real-time workloads,
2086 * so is thus not recommended for any sort of common-case code.  In fact,
2087 * if you are using synchronize_sched_expedited() in a loop, please
2088 * restructure your code to batch your updates, and then use a single
2089 * synchronize_sched() instead.
2090 *
2091 * Note that it is illegal to call this function while holding any lock
2092 * that is acquired by a CPU-hotplug notifier.  And yes, it is also illegal
2093 * to call this function from a CPU-hotplug notifier.  Failing to observe
2094 * these restriction will result in deadlock.
2095 *
2096 * This implementation can be thought of as an application of ticket
2097 * locking to RCU, with sync_sched_expedited_started and
2098 * sync_sched_expedited_done taking on the roles of the halves
2099 * of the ticket-lock word.  Each task atomically increments
2100 * sync_sched_expedited_started upon entry, snapshotting the old value,
2101 * then attempts to stop all the CPUs.  If this succeeds, then each
2102 * CPU will have executed a context switch, resulting in an RCU-sched
2103 * grace period.  We are then done, so we use atomic_cmpxchg() to
2104 * update sync_sched_expedited_done to match our snapshot -- but
2105 * only if someone else has not already advanced past our snapshot.
2106 *
2107 * On the other hand, if try_stop_cpus() fails, we check the value
2108 * of sync_sched_expedited_done.  If it has advanced past our
2109 * initial snapshot, then someone else must have forced a grace period
2110 * some time after we took our snapshot.  In this case, our work is
2111 * done for us, and we can simply return.  Otherwise, we try again,
2112 * but keep our initial snapshot for purposes of checking for someone
2113 * doing our work for us.
2114 *
2115 * If we fail too many times in a row, we fall back to synchronize_sched().
2116 */
2117void synchronize_sched_expedited(void)
2118{
2119        int firstsnap, s, snap, trycount = 0;
2120
2121        /* Note that atomic_inc_return() implies full memory barrier. */
2122        firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
2123        get_online_cpus();
2124        WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2125
2126        /*
2127         * Each pass through the following loop attempts to force a
2128         * context switch on each CPU.
2129         */
2130        while (try_stop_cpus(cpu_online_mask,
2131                             synchronize_sched_expedited_cpu_stop,
2132                             NULL) == -EAGAIN) {
2133                put_online_cpus();
2134
2135                /* No joy, try again later.  Or just synchronize_sched(). */
2136                if (trycount++ < 10) {
2137                        udelay(trycount * num_online_cpus());
2138                } else {
2139                        synchronize_sched();
2140                        return;
2141                }
2142
2143                /* Check to see if someone else did our work for us. */
2144                s = atomic_read(&sync_sched_expedited_done);
2145                if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
2146                        smp_mb(); /* ensure test happens before caller kfree */
2147                        return;
2148                }
2149
2150                /*
2151                 * Refetching sync_sched_expedited_started allows later
2152                 * callers to piggyback on our grace period.  We subtract
2153                 * 1 to get the same token that the last incrementer got.
2154                 * We retry after they started, so our grace period works
2155                 * for them, and they started after our first try, so their
2156                 * grace period works for us.
2157                 */
2158                get_online_cpus();
2159                snap = atomic_read(&sync_sched_expedited_started);
2160                smp_mb(); /* ensure read is before try_stop_cpus(). */
2161        }
2162
2163        /*
2164         * Everyone up to our most recent fetch is covered by our grace
2165         * period.  Update the counter, but only if our work is still
2166         * relevant -- which it won't be if someone who started later
2167         * than we did beat us to the punch.
2168         */
2169        do {
2170                s = atomic_read(&sync_sched_expedited_done);
2171                if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
2172                        smp_mb(); /* ensure test happens before caller kfree */
2173                        break;
2174                }
2175        } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
2176
2177        put_online_cpus();
2178}
2179EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2180
2181/*
2182 * Check to see if there is any immediate RCU-related work to be done
2183 * by the current CPU, for the specified type of RCU, returning 1 if so.
2184 * The checks are in order of increasing expense: checks that can be
2185 * carried out against CPU-local state are performed first.  However,
2186 * we must check for CPU stalls first, else we might not get a chance.
2187 */
2188static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2189{
2190        struct rcu_node *rnp = rdp->mynode;
2191
2192        rdp->n_rcu_pending++;
2193
2194        /* Check for CPU stalls, if enabled. */
2195        check_cpu_stall(rsp, rdp);
2196
2197        /* Is the RCU core waiting for a quiescent state from this CPU? */
2198        if (rcu_scheduler_fully_active &&
2199            rdp->qs_pending && !rdp->passed_quiesce) {
2200
2201                /*
2202                 * If force_quiescent_state() coming soon and this CPU
2203                 * needs a quiescent state, and this is either RCU-sched
2204                 * or RCU-bh, force a local reschedule.
2205                 */
2206                rdp->n_rp_qs_pending++;
2207                if (!rdp->preemptible &&
2208                    ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
2209                                 jiffies))
2210                        set_need_resched();
2211        } else if (rdp->qs_pending && rdp->passed_quiesce) {
2212                rdp->n_rp_report_qs++;
2213                return 1;
2214        }
2215
2216        /* Does this CPU have callbacks ready to invoke? */
2217        if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2218                rdp->n_rp_cb_ready++;
2219                return 1;
2220        }
2221
2222        /* Has RCU gone idle with this CPU needing another grace period? */
2223        if (cpu_needs_another_gp(rsp, rdp)) {
2224                rdp->n_rp_cpu_needs_gp++;
2225                return 1;
2226        }
2227
2228        /* Has another RCU grace period completed?  */
2229        if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2230                rdp->n_rp_gp_completed++;
2231                return 1;
2232        }
2233
2234        /* Has a new RCU grace period started? */
2235        if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2236                rdp->n_rp_gp_started++;
2237                return 1;
2238        }
2239
2240        /* Has an RCU GP gone long enough to send resched IPIs &c? */
2241        if (rcu_gp_in_progress(rsp) &&
2242            ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
2243                rdp->n_rp_need_fqs++;
2244                return 1;
2245        }
2246
2247        /* nothing to do */
2248        rdp->n_rp_need_nothing++;
2249        return 0;
2250}
2251
2252/*
2253 * Check to see if there is any immediate RCU-related work to be done
2254 * by the current CPU, returning 1 if so.  This function is part of the
2255 * RCU implementation; it is -not- an exported member of the RCU API.
2256 */
2257static int rcu_pending(int cpu)
2258{
2259        struct rcu_state *rsp;
2260
2261        for_each_rcu_flavor(rsp)
2262                if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2263                        return 1;
2264        return 0;
2265}
2266
2267/*
2268 * Check to see if any future RCU-related work will need to be done
2269 * by the current CPU, even if none need be done immediately, returning
2270 * 1 if so.
2271 */
2272static int rcu_cpu_has_callbacks(int cpu)
2273{
2274        struct rcu_state *rsp;
2275
2276        /* RCU callbacks either ready or pending? */
2277        for_each_rcu_flavor(rsp)
2278                if (per_cpu_ptr(rsp->rda, cpu)->nxtlist)
2279                        return 1;
2280        return 0;
2281}
2282
2283/*
2284 * Helper function for _rcu_barrier() tracing.  If tracing is disabled,
2285 * the compiler is expected to optimize this away.
2286 */
2287static void _rcu_barrier_trace(struct rcu_state *rsp, char *s,
2288                               int cpu, unsigned long done)
2289{
2290        trace_rcu_barrier(rsp->name, s, cpu,
2291                          atomic_read(&rsp->barrier_cpu_count), done);
2292}
2293
2294/*
2295 * RCU callback function for _rcu_barrier().  If we are last, wake
2296 * up the task executing _rcu_barrier().
2297 */
2298static void rcu_barrier_callback(struct rcu_head *rhp)
2299{
2300        struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2301        struct rcu_state *rsp = rdp->rsp;
2302
2303        if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2304                _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2305                complete(&rsp->barrier_completion);
2306        } else {
2307                _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2308        }
2309}
2310
2311/*
2312 * Called with preemption disabled, and from cross-cpu IRQ context.
2313 */
2314static void rcu_barrier_func(void *type)
2315{
2316        struct rcu_state *rsp = type;
2317        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2318
2319        _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2320        atomic_inc(&rsp->barrier_cpu_count);
2321        rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2322}
2323
2324/*
2325 * Orchestrate the specified type of RCU barrier, waiting for all
2326 * RCU callbacks of the specified type to complete.
2327 */
2328static void _rcu_barrier(struct rcu_state *rsp)
2329{
2330        int cpu;
2331        unsigned long flags;
2332        struct rcu_data *rdp;
2333        struct rcu_data rd;
2334        unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2335        unsigned long snap_done;
2336
2337        init_rcu_head_on_stack(&rd.barrier_head);
2338        _rcu_barrier_trace(rsp, "Begin", -1, snap);
2339
2340        /* Take mutex to serialize concurrent rcu_barrier() requests. */
2341        mutex_lock(&rsp->barrier_mutex);
2342
2343        /*
2344         * Ensure that all prior references, including to ->n_barrier_done,
2345         * are ordered before the _rcu_barrier() machinery.
2346         */
2347        smp_mb();  /* See above block comment. */
2348
2349        /*
2350         * Recheck ->n_barrier_done to see if others did our work for us.
2351         * This means checking ->n_barrier_done for an even-to-odd-to-even
2352         * transition.  The "if" expression below therefore rounds the old
2353         * value up to the next even number and adds two before comparing.
2354         */
2355        snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2356        _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2357        if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2358                _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2359                smp_mb(); /* caller's subsequent code after above check. */
2360                mutex_unlock(&rsp->barrier_mutex);
2361                return;
2362        }
2363
2364        /*
2365         * Increment ->n_barrier_done to avoid duplicate work.  Use
2366         * ACCESS_ONCE() to prevent the compiler from speculating
2367         * the increment to precede the early-exit check.
2368         */
2369        ACCESS_ONCE(rsp->n_barrier_done)++;
2370        WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2371        _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2372        smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2373
2374        /*
2375         * Initialize the count to one rather than to zero in order to
2376         * avoid a too-soon return to zero in case of a short grace period
2377         * (or preemption of this task).  Also flag this task as doing
2378         * an rcu_barrier().  This will prevent anyone else from adopting
2379         * orphaned callbacks, which could cause otherwise failure if a
2380         * CPU went offline and quickly came back online.  To see this,
2381         * consider the following sequence of events:
2382         *
2383         * 1.   We cause CPU 0 to post an rcu_barrier_callback() callback.
2384         * 2.   CPU 1 goes offline, orphaning its callbacks.
2385         * 3.   CPU 0 adopts CPU 1's orphaned callbacks.
2386         * 4.   CPU 1 comes back online.
2387         * 5.   We cause CPU 1 to post an rcu_barrier_callback() callback.
2388         * 6.   Both rcu_barrier_callback() callbacks are invoked, awakening
2389         *      us -- but before CPU 1's orphaned callbacks are invoked!!!
2390         */
2391        init_completion(&rsp->barrier_completion);
2392        atomic_set(&rsp->barrier_cpu_count, 1);
2393        raw_spin_lock_irqsave(&rsp->onofflock, flags);
2394        rsp->rcu_barrier_in_progress = current;
2395        raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2396
2397        /*
2398         * Force every CPU with callbacks to register a new callback
2399         * that will tell us when all the preceding callbacks have
2400         * been invoked.  If an offline CPU has callbacks, wait for
2401         * it to either come back online or to finish orphaning those
2402         * callbacks.
2403         */
2404        for_each_possible_cpu(cpu) {
2405                preempt_disable();
2406                rdp = per_cpu_ptr(rsp->rda, cpu);
2407                if (cpu_is_offline(cpu)) {
2408                        _rcu_barrier_trace(rsp, "Offline", cpu,
2409                                           rsp->n_barrier_done);
2410                        preempt_enable();
2411                        while (cpu_is_offline(cpu) && ACCESS_ONCE(rdp->qlen))
2412                                schedule_timeout_interruptible(1);
2413                } else if (ACCESS_ONCE(rdp->qlen)) {
2414                        _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2415                                           rsp->n_barrier_done);
2416                        smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2417                        preempt_enable();
2418                } else {
2419                        _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2420                                           rsp->n_barrier_done);
2421                        preempt_enable();
2422                }
2423        }
2424
2425        /*
2426         * Now that all online CPUs have rcu_barrier_callback() callbacks
2427         * posted, we can adopt all of the orphaned callbacks and place
2428         * an rcu_barrier_callback() callback after them.  When that is done,
2429         * we are guaranteed to have an rcu_barrier_callback() callback
2430         * following every callback that could possibly have been
2431         * registered before _rcu_barrier() was called.
2432         */
2433        raw_spin_lock_irqsave(&rsp->onofflock, flags);
2434        rcu_adopt_orphan_cbs(rsp);
2435        rsp->rcu_barrier_in_progress = NULL;
2436        raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2437        atomic_inc(&rsp->barrier_cpu_count);
2438        smp_mb__after_atomic_inc(); /* Ensure atomic_inc() before callback. */
2439        rd.rsp = rsp;
2440        rsp->call(&rd.barrier_head, rcu_barrier_callback);
2441
2442        /*
2443         * Now that we have an rcu_barrier_callback() callback on each
2444         * CPU, and thus each counted, remove the initial count.
2445         */
2446        if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2447                complete(&rsp->barrier_completion);
2448
2449        /* Increment ->n_barrier_done to prevent duplicate work. */
2450        smp_mb(); /* Keep increment after above mechanism. */
2451        ACCESS_ONCE(rsp->n_barrier_done)++;
2452        WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2453        _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2454        smp_mb(); /* Keep increment before caller's subsequent code. */
2455
2456        /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2457        wait_for_completion(&rsp->barrier_completion);
2458
2459        /* Other rcu_barrier() invocations can now safely proceed. */
2460        mutex_unlock(&rsp->barrier_mutex);
2461
2462        destroy_rcu_head_on_stack(&rd.barrier_head);
2463}
2464
2465/**
2466 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2467 */
2468void rcu_barrier_bh(void)
2469{
2470        _rcu_barrier(&rcu_bh_state);
2471}
2472EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2473
2474/**
2475 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2476 */
2477void rcu_barrier_sched(void)
2478{
2479        _rcu_barrier(&rcu_sched_state);
2480}
2481EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2482
2483/*
2484 * Do boot-time initialization of a CPU's per-CPU RCU data.
2485 */
2486static void __init
2487rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2488{
2489        unsigned long flags;
2490        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2491        struct rcu_node *rnp = rcu_get_root(rsp);
2492
2493        /* Set up local state, ensuring consistent view of global state. */
2494        raw_spin_lock_irqsave(&rnp->lock, flags);
2495        rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2496        init_callback_list(rdp);
2497        rdp->qlen_lazy = 0;
2498        ACCESS_ONCE(rdp->qlen) = 0;
2499        rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2500        WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2501        WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2502        rdp->cpu = cpu;
2503        rdp->rsp = rsp;
2504        raw_spin_unlock_irqrestore(&rnp->lock, flags);
2505}
2506
2507/*
2508 * Initialize a CPU's per-CPU RCU data.  Note that only one online or
2509 * offline event can be happening at a given time.  Note also that we
2510 * can accept some slop in the rsp->completed access due to the fact
2511 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2512 */
2513static void __cpuinit
2514rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2515{
2516        unsigned long flags;
2517        unsigned long mask;
2518        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2519        struct rcu_node *rnp = rcu_get_root(rsp);
2520
2521        /* Set up local state, ensuring consistent view of global state. */
2522        raw_spin_lock_irqsave(&rnp->lock, flags);
2523        rdp->beenonline = 1;     /* We have now been online. */
2524        rdp->preemptible = preemptible;
2525        rdp->qlen_last_fqs_check = 0;
2526        rdp->n_force_qs_snap = rsp->n_force_qs;
2527        rdp->blimit = blimit;
2528        rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2529        atomic_set(&rdp->dynticks->dynticks,
2530                   (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2531        rcu_prepare_for_idle_init(cpu);
2532        raw_spin_unlock(&rnp->lock);            /* irqs remain disabled. */
2533
2534        /*
2535         * A new grace period might start here.  If so, we won't be part
2536         * of it, but that is OK, as we are currently in a quiescent state.
2537         */
2538
2539        /* Exclude any attempts to start a new GP on large systems. */
2540        raw_spin_lock(&rsp->onofflock);         /* irqs already disabled. */
2541
2542        /* Add CPU to rcu_node bitmasks. */
2543        rnp = rdp->mynode;
2544        mask = rdp->grpmask;
2545        do {
2546                /* Exclude any attempts to start a new GP on small systems. */
2547                raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
2548                rnp->qsmaskinit |= mask;
2549                mask = rnp->grpmask;
2550                if (rnp == rdp->mynode) {
2551                        /*
2552                         * If there is a grace period in progress, we will
2553                         * set up to wait for it next time we run the
2554                         * RCU core code.
2555                         */
2556                        rdp->gpnum = rnp->completed;
2557                        rdp->completed = rnp->completed;
2558                        rdp->passed_quiesce = 0;
2559                        rdp->qs_pending = 0;
2560                        rdp->passed_quiesce_gpnum = rnp->gpnum - 1;
2561                        trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2562                }
2563                raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2564                rnp = rnp->parent;
2565        } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2566
2567        raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2568}
2569
2570static void __cpuinit rcu_prepare_cpu(int cpu)
2571{
2572        struct rcu_state *rsp;
2573
2574        for_each_rcu_flavor(rsp)
2575                rcu_init_percpu_data(cpu, rsp,
2576                                     strcmp(rsp->name, "rcu_preempt") == 0);
2577}
2578
2579/*
2580 * Handle CPU online/offline notification events.
2581 */
2582static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2583                                    unsigned long action, void *hcpu)
2584{
2585        long cpu = (long)hcpu;
2586        struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2587        struct rcu_node *rnp = rdp->mynode;
2588        struct rcu_state *rsp;
2589
2590        trace_rcu_utilization("Start CPU hotplug");
2591        switch (action) {
2592        case CPU_UP_PREPARE:
2593        case CPU_UP_PREPARE_FROZEN:
2594                rcu_prepare_cpu(cpu);
2595                rcu_prepare_kthreads(cpu);
2596                break;
2597        case CPU_ONLINE:
2598        case CPU_DOWN_FAILED:
2599                rcu_node_kthread_setaffinity(rnp, -1);
2600                rcu_cpu_kthread_setrt(cpu, 1);
2601                break;
2602        case CPU_DOWN_PREPARE:
2603                rcu_node_kthread_setaffinity(rnp, cpu);
2604                rcu_cpu_kthread_setrt(cpu, 0);
2605                break;
2606        case CPU_DYING:
2607        case CPU_DYING_FROZEN:
2608                /*
2609                 * The whole machine is "stopped" except this CPU, so we can
2610                 * touch any data without introducing corruption. We send the
2611                 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2612                 */
2613                for_each_rcu_flavor(rsp)
2614                        rcu_cleanup_dying_cpu(rsp);
2615                rcu_cleanup_after_idle(cpu);
2616                break;
2617        case CPU_DEAD:
2618        case CPU_DEAD_FROZEN:
2619        case CPU_UP_CANCELED:
2620        case CPU_UP_CANCELED_FROZEN:
2621                for_each_rcu_flavor(rsp)
2622                        rcu_cleanup_dead_cpu(cpu, rsp);
2623                break;
2624        default:
2625                break;
2626        }
2627        trace_rcu_utilization("End CPU hotplug");
2628        return NOTIFY_OK;
2629}
2630
2631/*
2632 * This function is invoked towards the end of the scheduler's initialization
2633 * process.  Before this is called, the idle task might contain
2634 * RCU read-side critical sections (during which time, this idle
2635 * task is booting the system).  After this function is called, the
2636 * idle tasks are prohibited from containing RCU read-side critical
2637 * sections.  This function also enables RCU lockdep checking.
2638 */
2639void rcu_scheduler_starting(void)
2640{
2641        WARN_ON(num_online_cpus() != 1);
2642        WARN_ON(nr_context_switches() > 0);
2643        rcu_scheduler_active = 1;
2644}
2645
2646/*
2647 * Compute the per-level fanout, either using the exact fanout specified
2648 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2649 */
2650#ifdef CONFIG_RCU_FANOUT_EXACT
2651static void __init rcu_init_levelspread(struct rcu_state *rsp)
2652{
2653        int i;
2654
2655        for (i = rcu_num_lvls - 1; i > 0; i--)
2656                rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2657        rsp->levelspread[0] = rcu_fanout_leaf;
2658}
2659#else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2660static void __init rcu_init_levelspread(struct rcu_state *rsp)
2661{
2662        int ccur;
2663        int cprv;
2664        int i;
2665
2666        cprv = NR_CPUS;
2667        for (i = rcu_num_lvls - 1; i >= 0; i--) {
2668                ccur = rsp->levelcnt[i];
2669                rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2670                cprv = ccur;
2671        }
2672}
2673#endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2674
2675/*
2676 * Helper function for rcu_init() that initializes one rcu_state structure.
2677 */
2678static void __init rcu_init_one(struct rcu_state *rsp,
2679                struct rcu_data __percpu *rda)
2680{
2681        static char *buf[] = { "rcu_node_level_0",
2682                               "rcu_node_level_1",
2683                               "rcu_node_level_2",
2684                               "rcu_node_level_3" };  /* Match MAX_RCU_LVLS */
2685        int cpustride = 1;
2686        int i;
2687        int j;
2688        struct rcu_node *rnp;
2689
2690        BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf));  /* Fix buf[] init! */
2691
2692        /* Initialize the level-tracking arrays. */
2693
2694        for (i = 0; i < rcu_num_lvls; i++)
2695                rsp->levelcnt[i] = num_rcu_lvl[i];
2696        for (i = 1; i < rcu_num_lvls; i++)
2697                rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2698        rcu_init_levelspread(rsp);
2699
2700        /* Initialize the elements themselves, starting from the leaves. */
2701
2702        for (i = rcu_num_lvls - 1; i >= 0; i--) {
2703                cpustride *= rsp->levelspread[i];
2704                rnp = rsp->level[i];
2705                for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2706                        raw_spin_lock_init(&rnp->lock);
2707                        lockdep_set_class_and_name(&rnp->lock,
2708                                                   &rcu_node_class[i], buf[i]);
2709                        rnp->gpnum = 0;
2710                        rnp->qsmask = 0;
2711                        rnp->qsmaskinit = 0;
2712                        rnp->grplo = j * cpustride;
2713                        rnp->grphi = (j + 1) * cpustride - 1;
2714                        if (rnp->grphi >= NR_CPUS)
2715                                rnp->grphi = NR_CPUS - 1;
2716                        if (i == 0) {
2717                                rnp->grpnum = 0;
2718                                rnp->grpmask = 0;
2719                                rnp->parent = NULL;
2720                        } else {
2721                                rnp->grpnum = j % rsp->levelspread[i - 1];
2722                                rnp->grpmask = 1UL << rnp->grpnum;
2723                                rnp->parent = rsp->level[i - 1] +
2724                                              j / rsp->levelspread[i - 1];
2725                        }
2726                        rnp->level = i;
2727                        INIT_LIST_HEAD(&rnp->blkd_tasks);
2728                }
2729        }
2730
2731        rsp->rda = rda;
2732        rnp = rsp->level[rcu_num_lvls - 1];
2733        for_each_possible_cpu(i) {
2734                while (i > rnp->grphi)
2735                        rnp++;
2736                per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2737                rcu_boot_init_percpu_data(i, rsp);
2738        }
2739        list_add(&rsp->flavors, &rcu_struct_flavors);
2740}
2741
2742/*
2743 * Compute the rcu_node tree geometry from kernel parameters.  This cannot
2744 * replace the definitions in rcutree.h because those are needed to size
2745 * the ->node array in the rcu_state structure.
2746 */
2747static void __init rcu_init_geometry(void)
2748{
2749        int i;
2750        int j;
2751        int n = nr_cpu_ids;
2752        int rcu_capacity[MAX_RCU_LVLS + 1];
2753
2754        /* If the compile-time values are accurate, just leave. */
2755        if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF)
2756                return;
2757
2758        /*
2759         * Compute number of nodes that can be handled an rcu_node tree
2760         * with the given number of levels.  Setting rcu_capacity[0] makes
2761         * some of the arithmetic easier.
2762         */
2763        rcu_capacity[0] = 1;
2764        rcu_capacity[1] = rcu_fanout_leaf;
2765        for (i = 2; i <= MAX_RCU_LVLS; i++)
2766                rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
2767
2768        /*
2769         * The boot-time rcu_fanout_leaf parameter is only permitted
2770         * to increase the leaf-level fanout, not decrease it.  Of course,
2771         * the leaf-level fanout cannot exceed the number of bits in
2772         * the rcu_node masks.  Finally, the tree must be able to accommodate
2773         * the configured number of CPUs.  Complain and fall back to the
2774         * compile-time values if these limits are exceeded.
2775         */
2776        if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
2777            rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
2778            n > rcu_capacity[MAX_RCU_LVLS]) {
2779                WARN_ON(1);
2780                return;
2781        }
2782
2783        /* Calculate the number of rcu_nodes at each level of the tree. */
2784        for (i = 1; i <= MAX_RCU_LVLS; i++)
2785                if (n <= rcu_capacity[i]) {
2786                        for (j = 0; j <= i; j++)
2787                                num_rcu_lvl[j] =
2788                                        DIV_ROUND_UP(n, rcu_capacity[i - j]);
2789                        rcu_num_lvls = i;
2790                        for (j = i + 1; j <= MAX_RCU_LVLS; j++)
2791                                num_rcu_lvl[j] = 0;
2792                        break;
2793                }
2794
2795        /* Calculate the total number of rcu_node structures. */
2796        rcu_num_nodes = 0;
2797        for (i = 0; i <= MAX_RCU_LVLS; i++)
2798                rcu_num_nodes += num_rcu_lvl[i];
2799        rcu_num_nodes -= n;
2800}
2801
2802void __init rcu_init(void)
2803{
2804        int cpu;
2805
2806        rcu_bootup_announce();
2807        rcu_init_geometry();
2808        rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2809        rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2810        __rcu_init_preempt();
2811         open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2812
2813        /*
2814         * We don't need protection against CPU-hotplug here because
2815         * this is called early in boot, before either interrupts
2816         * or the scheduler are operational.
2817         */
2818        cpu_notifier(rcu_cpu_notify, 0);
2819        for_each_online_cpu(cpu)
2820                rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2821        check_cpu_stall_init();
2822}
2823
2824#include "rcutree_plugin.h"
2825
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.