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