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