linux/kernel/rcutree.c
<<
>>
Prefs
   1/*
   2 * Read-Copy Update mechanism for mutual exclusion
   3 *
   4 * This program is free software; you can redistribute it and/or modify
   5 * it under the terms of the GNU General Public License as published by
   6 * the Free Software Foundation; either version 2 of the License, or
   7 * (at your option) any later version.
   8 *
   9 * This program is distributed in the hope that it will be useful,
  10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12 * GNU General Public License for more details.
  13 *
  14 * You should have received a copy of the GNU General Public License
  15 * along with this program; if not, write to the Free Software
  16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  17 *
  18 * Copyright IBM Corporation, 2008
  19 *
  20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
  21 *          Manfred Spraul <manfred@colorfullife.com>
  22 *          Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
  23 *
  24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
  25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
  26 *
  27 * For detailed explanation of Read-Copy Update mechanism see -
  28 *      Documentation/RCU
  29 */
  30#include <linux/types.h>
  31#include <linux/kernel.h>
  32#include <linux/init.h>
  33#include <linux/spinlock.h>
  34#include <linux/smp.h>
  35#include <linux/rcupdate.h>
  36#include <linux/interrupt.h>
  37#include <linux/sched.h>
  38#include <linux/nmi.h>
  39#include <linux/atomic.h>
  40#include <linux/bitops.h>
  41#include <linux/export.h>
  42#include <linux/completion.h>
  43#include <linux/moduleparam.h>
  44#include <linux/percpu.h>
  45#include <linux/notifier.h>
  46#include <linux/cpu.h>
  47#include <linux/mutex.h>
  48#include <linux/time.h>
  49#include <linux/kernel_stat.h>
  50#include <linux/wait.h>
  51#include <linux/kthread.h>
  52#include <linux/prefetch.h>
  53#include <linux/delay.h>
  54#include <linux/stop_machine.h>
  55#include <linux/random.h>
  56
  57#include "rcutree.h"
  58#include <trace/events/rcu.h>
  59
  60#include "rcu.h"
  61
  62/* Data structures. */
  63
  64static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
  65static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
  66
  67#define RCU_STATE_INITIALIZER(sname, cr) { \
  68        .level = { &sname##_state.node[0] }, \
  69        .call = cr, \
  70        .fqs_state = RCU_GP_IDLE, \
  71        .gpnum = 0UL - 300UL, \
  72        .completed = 0UL - 300UL, \
  73        .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
  74        .orphan_nxttail = &sname##_state.orphan_nxtlist, \
  75        .orphan_donetail = &sname##_state.orphan_donelist, \
  76        .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
  77        .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
  78        .name = #sname, \
  79}
  80
  81struct rcu_state rcu_sched_state =
  82        RCU_STATE_INITIALIZER(rcu_sched, call_rcu_sched);
  83DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
  84
  85struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh, call_rcu_bh);
  86DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
  87
  88static struct rcu_state *rcu_state;
  89LIST_HEAD(rcu_struct_flavors);
  90
  91/* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
  92static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
  93module_param(rcu_fanout_leaf, int, 0444);
  94int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
  95static int num_rcu_lvl[] = {  /* Number of rcu_nodes at specified level. */
  96        NUM_RCU_LVL_0,
  97        NUM_RCU_LVL_1,
  98        NUM_RCU_LVL_2,
  99        NUM_RCU_LVL_3,
 100        NUM_RCU_LVL_4,
 101};
 102int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
 103
 104/*
 105 * The rcu_scheduler_active variable transitions from zero to one just
 106 * before the first task is spawned.  So when this variable is zero, RCU
 107 * can assume that there is but one task, allowing RCU to (for example)
 108 * optimize synchronize_sched() to a simple barrier().  When this variable
 109 * is one, RCU must actually do all the hard work required to detect real
 110 * grace periods.  This variable is also used to suppress boot-time false
 111 * positives from lockdep-RCU error checking.
 112 */
 113int rcu_scheduler_active __read_mostly;
 114EXPORT_SYMBOL_GPL(rcu_scheduler_active);
 115
 116/*
 117 * The rcu_scheduler_fully_active variable transitions from zero to one
 118 * during the early_initcall() processing, which is after the scheduler
 119 * is capable of creating new tasks.  So RCU processing (for example,
 120 * creating tasks for RCU priority boosting) must be delayed until after
 121 * rcu_scheduler_fully_active transitions from zero to one.  We also
 122 * currently delay invocation of any RCU callbacks until after this point.
 123 *
 124 * It might later prove better for people registering RCU callbacks during
 125 * early boot to take responsibility for these callbacks, but one step at
 126 * a time.
 127 */
 128static int rcu_scheduler_fully_active __read_mostly;
 129
 130#ifdef CONFIG_RCU_BOOST
 131
 132/*
 133 * Control variables for per-CPU and per-rcu_node kthreads.  These
 134 * handle all flavors of RCU.
 135 */
 136static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
 137DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
 138DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
 139DEFINE_PER_CPU(char, rcu_cpu_has_work);
 140
 141#endif /* #ifdef CONFIG_RCU_BOOST */
 142
 143static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
 144static void invoke_rcu_core(void);
 145static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
 146
 147/*
 148 * Track the rcutorture test sequence number and the update version
 149 * number within a given test.  The rcutorture_testseq is incremented
 150 * on every rcutorture module load and unload, so has an odd value
 151 * when a test is running.  The rcutorture_vernum is set to zero
 152 * when rcutorture starts and is incremented on each rcutorture update.
 153 * These variables enable correlating rcutorture output with the
 154 * RCU tracing information.
 155 */
 156unsigned long rcutorture_testseq;
 157unsigned long rcutorture_vernum;
 158
 159/*
 160 * Return true if an RCU grace period is in progress.  The ACCESS_ONCE()s
 161 * permit this function to be invoked without holding the root rcu_node
 162 * structure's ->lock, but of course results can be subject to change.
 163 */
 164static int rcu_gp_in_progress(struct rcu_state *rsp)
 165{
 166        return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
 167}
 168
 169/*
 170 * Note a quiescent state.  Because we do not need to know
 171 * how many quiescent states passed, just if there was at least
 172 * one since the start of the grace period, this just sets a flag.
 173 * The caller must have disabled preemption.
 174 */
 175void rcu_sched_qs(int cpu)
 176{
 177        struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
 178
 179        if (rdp->passed_quiesce == 0)
 180                trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
 181        rdp->passed_quiesce = 1;
 182}
 183
 184void rcu_bh_qs(int cpu)
 185{
 186        struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
 187
 188        if (rdp->passed_quiesce == 0)
 189                trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
 190        rdp->passed_quiesce = 1;
 191}
 192
 193/*
 194 * Note a context switch.  This is a quiescent state for RCU-sched,
 195 * and requires special handling for preemptible RCU.
 196 * The caller must have disabled preemption.
 197 */
 198void rcu_note_context_switch(int cpu)
 199{
 200        trace_rcu_utilization("Start context switch");
 201        rcu_sched_qs(cpu);
 202        rcu_preempt_note_context_switch(cpu);
 203        trace_rcu_utilization("End context switch");
 204}
 205EXPORT_SYMBOL_GPL(rcu_note_context_switch);
 206
 207DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
 208        .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
 209        .dynticks = ATOMIC_INIT(1),
 210};
 211
 212static long blimit = 10;        /* Maximum callbacks per rcu_do_batch. */
 213static long qhimark = 10000;    /* If this many pending, ignore blimit. */
 214static long qlowmark = 100;     /* Once only this many pending, use blimit. */
 215
 216module_param(blimit, long, 0444);
 217module_param(qhimark, long, 0444);
 218module_param(qlowmark, long, 0444);
 219
 220static ulong jiffies_till_first_fqs = RCU_JIFFIES_TILL_FORCE_QS;
 221static ulong jiffies_till_next_fqs = RCU_JIFFIES_TILL_FORCE_QS;
 222
 223module_param(jiffies_till_first_fqs, ulong, 0644);
 224module_param(jiffies_till_next_fqs, ulong, 0644);
 225
 226static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *));
 227static void force_quiescent_state(struct rcu_state *rsp);
 228static int rcu_pending(int cpu);
 229
 230/*
 231 * Return the number of RCU-sched batches processed thus far for debug & stats.
 232 */
 233long rcu_batches_completed_sched(void)
 234{
 235        return rcu_sched_state.completed;
 236}
 237EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
 238
 239/*
 240 * Return the number of RCU BH batches processed thus far for debug & stats.
 241 */
 242long rcu_batches_completed_bh(void)
 243{
 244        return rcu_bh_state.completed;
 245}
 246EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
 247
 248/*
 249 * Force a quiescent state for RCU BH.
 250 */
 251void rcu_bh_force_quiescent_state(void)
 252{
 253        force_quiescent_state(&rcu_bh_state);
 254}
 255EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
 256
 257/*
 258 * Record the number of times rcutorture tests have been initiated and
 259 * terminated.  This information allows the debugfs tracing stats to be
 260 * correlated to the rcutorture messages, even when the rcutorture module
 261 * is being repeatedly loaded and unloaded.  In other words, we cannot
 262 * store this state in rcutorture itself.
 263 */
 264void rcutorture_record_test_transition(void)
 265{
 266        rcutorture_testseq++;
 267        rcutorture_vernum = 0;
 268}
 269EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
 270
 271/*
 272 * Record the number of writer passes through the current rcutorture test.
 273 * This is also used to correlate debugfs tracing stats with the rcutorture
 274 * messages.
 275 */
 276void rcutorture_record_progress(unsigned long vernum)
 277{
 278        rcutorture_vernum++;
 279}
 280EXPORT_SYMBOL_GPL(rcutorture_record_progress);
 281
 282/*
 283 * Force a quiescent state for RCU-sched.
 284 */
 285void rcu_sched_force_quiescent_state(void)
 286{
 287        force_quiescent_state(&rcu_sched_state);
 288}
 289EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
 290
 291/*
 292 * Does the CPU have callbacks ready to be invoked?
 293 */
 294static int
 295cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
 296{
 297        return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
 298               rdp->nxttail[RCU_DONE_TAIL] != NULL;
 299}
 300
 301/*
 302 * Does the current CPU require a not-yet-started grace period?
 303 * The caller must have disabled interrupts to prevent races with
 304 * normal callback registry.
 305 */
 306static int
 307cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
 308{
 309        int i;
 310
 311        if (rcu_gp_in_progress(rsp))
 312                return 0;  /* No, a grace period is already in progress. */
 313        if (!rdp->nxttail[RCU_NEXT_TAIL])
 314                return 0;  /* No, this is a no-CBs (or offline) CPU. */
 315        if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
 316                return 1;  /* Yes, this CPU has newly registered callbacks. */
 317        for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
 318                if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
 319                    ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
 320                                 rdp->nxtcompleted[i]))
 321                        return 1;  /* Yes, CBs for future grace period. */
 322        return 0; /* No grace period needed. */
 323}
 324
 325/*
 326 * Return the root node of the specified rcu_state structure.
 327 */
 328static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
 329{
 330        return &rsp->node[0];
 331}
 332
 333/*
 334 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
 335 *
 336 * If the new value of the ->dynticks_nesting counter now is zero,
 337 * we really have entered idle, and must do the appropriate accounting.
 338 * The caller must have disabled interrupts.
 339 */
 340static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
 341                                bool user)
 342{
 343        trace_rcu_dyntick("Start", oldval, rdtp->dynticks_nesting);
 344        if (!user && !is_idle_task(current)) {
 345                struct task_struct *idle = idle_task(smp_processor_id());
 346
 347                trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
 348                ftrace_dump(DUMP_ORIG);
 349                WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
 350                          current->pid, current->comm,
 351                          idle->pid, idle->comm); /* must be idle task! */
 352        }
 353        rcu_prepare_for_idle(smp_processor_id());
 354        /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
 355        smp_mb__before_atomic_inc();  /* See above. */
 356        atomic_inc(&rdtp->dynticks);
 357        smp_mb__after_atomic_inc();  /* Force ordering with next sojourn. */
 358        WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
 359
 360        /*
 361         * It is illegal to enter an extended quiescent state while
 362         * in an RCU read-side critical section.
 363         */
 364        rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
 365                           "Illegal idle entry in RCU read-side critical section.");
 366        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
 367                           "Illegal idle entry in RCU-bh read-side critical section.");
 368        rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
 369                           "Illegal idle entry in RCU-sched read-side critical section.");
 370}
 371
 372/*
 373 * Enter an RCU extended quiescent state, which can be either the
 374 * idle loop or adaptive-tickless usermode execution.
 375 */
 376static void rcu_eqs_enter(bool user)
 377{
 378        long long oldval;
 379        struct rcu_dynticks *rdtp;
 380
 381        rdtp = &__get_cpu_var(rcu_dynticks);
 382        oldval = rdtp->dynticks_nesting;
 383        WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
 384        if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
 385                rdtp->dynticks_nesting = 0;
 386        else
 387                rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
 388        rcu_eqs_enter_common(rdtp, oldval, user);
 389}
 390
 391/**
 392 * rcu_idle_enter - inform RCU that current CPU is entering idle
 393 *
 394 * Enter idle mode, in other words, -leave- the mode in which RCU
 395 * read-side critical sections can occur.  (Though RCU read-side
 396 * critical sections can occur in irq handlers in idle, a possibility
 397 * handled by irq_enter() and irq_exit().)
 398 *
 399 * We crowbar the ->dynticks_nesting field to zero to allow for
 400 * the possibility of usermode upcalls having messed up our count
 401 * of interrupt nesting level during the prior busy period.
 402 */
 403void rcu_idle_enter(void)
 404{
 405        unsigned long flags;
 406
 407        local_irq_save(flags);
 408        rcu_eqs_enter(false);
 409        local_irq_restore(flags);
 410}
 411EXPORT_SYMBOL_GPL(rcu_idle_enter);
 412
 413#ifdef CONFIG_RCU_USER_QS
 414/**
 415 * rcu_user_enter - inform RCU that we are resuming userspace.
 416 *
 417 * Enter RCU idle mode right before resuming userspace.  No use of RCU
 418 * is permitted between this call and rcu_user_exit(). This way the
 419 * CPU doesn't need to maintain the tick for RCU maintenance purposes
 420 * when the CPU runs in userspace.
 421 */
 422void rcu_user_enter(void)
 423{
 424        rcu_eqs_enter(1);
 425}
 426
 427/**
 428 * rcu_user_enter_after_irq - inform RCU that we are going to resume userspace
 429 * after the current irq returns.
 430 *
 431 * This is similar to rcu_user_enter() but in the context of a non-nesting
 432 * irq. After this call, RCU enters into idle mode when the interrupt
 433 * returns.
 434 */
 435void rcu_user_enter_after_irq(void)
 436{
 437        unsigned long flags;
 438        struct rcu_dynticks *rdtp;
 439
 440        local_irq_save(flags);
 441        rdtp = &__get_cpu_var(rcu_dynticks);
 442        /* Ensure this irq is interrupting a non-idle RCU state.  */
 443        WARN_ON_ONCE(!(rdtp->dynticks_nesting & DYNTICK_TASK_MASK));
 444        rdtp->dynticks_nesting = 1;
 445        local_irq_restore(flags);
 446}
 447#endif /* CONFIG_RCU_USER_QS */
 448
 449/**
 450 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
 451 *
 452 * Exit from an interrupt handler, which might possibly result in entering
 453 * idle mode, in other words, leaving the mode in which read-side critical
 454 * sections can occur.
 455 *
 456 * This code assumes that the idle loop never does anything that might
 457 * result in unbalanced calls to irq_enter() and irq_exit().  If your
 458 * architecture violates this assumption, RCU will give you what you
 459 * deserve, good and hard.  But very infrequently and irreproducibly.
 460 *
 461 * Use things like work queues to work around this limitation.
 462 *
 463 * You have been warned.
 464 */
 465void rcu_irq_exit(void)
 466{
 467        unsigned long flags;
 468        long long oldval;
 469        struct rcu_dynticks *rdtp;
 470
 471        local_irq_save(flags);
 472        rdtp = &__get_cpu_var(rcu_dynticks);
 473        oldval = rdtp->dynticks_nesting;
 474        rdtp->dynticks_nesting--;
 475        WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
 476        if (rdtp->dynticks_nesting)
 477                trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
 478        else
 479                rcu_eqs_enter_common(rdtp, oldval, true);
 480        local_irq_restore(flags);
 481}
 482
 483/*
 484 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
 485 *
 486 * If the new value of the ->dynticks_nesting counter was previously zero,
 487 * we really have exited idle, and must do the appropriate accounting.
 488 * The caller must have disabled interrupts.
 489 */
 490static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
 491                               int user)
 492{
 493        smp_mb__before_atomic_inc();  /* Force ordering w/previous sojourn. */
 494        atomic_inc(&rdtp->dynticks);
 495        /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
 496        smp_mb__after_atomic_inc();  /* See above. */
 497        WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
 498        rcu_cleanup_after_idle(smp_processor_id());
 499        trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
 500        if (!user && !is_idle_task(current)) {
 501                struct task_struct *idle = idle_task(smp_processor_id());
 502
 503                trace_rcu_dyntick("Error on exit: not idle task",
 504                                  oldval, rdtp->dynticks_nesting);
 505                ftrace_dump(DUMP_ORIG);
 506                WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
 507                          current->pid, current->comm,
 508                          idle->pid, idle->comm); /* must be idle task! */
 509        }
 510}
 511
 512/*
 513 * Exit an RCU extended quiescent state, which can be either the
 514 * idle loop or adaptive-tickless usermode execution.
 515 */
 516static void rcu_eqs_exit(bool user)
 517{
 518        struct rcu_dynticks *rdtp;
 519        long long oldval;
 520
 521        rdtp = &__get_cpu_var(rcu_dynticks);
 522        oldval = rdtp->dynticks_nesting;
 523        WARN_ON_ONCE(oldval < 0);
 524        if (oldval & DYNTICK_TASK_NEST_MASK)
 525                rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
 526        else
 527                rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
 528        rcu_eqs_exit_common(rdtp, oldval, user);
 529}
 530
 531/**
 532 * rcu_idle_exit - inform RCU that current CPU is leaving idle
 533 *
 534 * Exit idle mode, in other words, -enter- the mode in which RCU
 535 * read-side critical sections can occur.
 536 *
 537 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
 538 * allow for the possibility of usermode upcalls messing up our count
 539 * of interrupt nesting level during the busy period that is just
 540 * now starting.
 541 */
 542void rcu_idle_exit(void)
 543{
 544        unsigned long flags;
 545
 546        local_irq_save(flags);
 547        rcu_eqs_exit(false);
 548        local_irq_restore(flags);
 549}
 550EXPORT_SYMBOL_GPL(rcu_idle_exit);
 551
 552#ifdef CONFIG_RCU_USER_QS
 553/**
 554 * rcu_user_exit - inform RCU that we are exiting userspace.
 555 *
 556 * Exit RCU idle mode while entering the kernel because it can
 557 * run a RCU read side critical section anytime.
 558 */
 559void rcu_user_exit(void)
 560{
 561        rcu_eqs_exit(1);
 562}
 563
 564/**
 565 * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace
 566 * idle mode after the current non-nesting irq returns.
 567 *
 568 * This is similar to rcu_user_exit() but in the context of an irq.
 569 * This is called when the irq has interrupted a userspace RCU idle mode
 570 * context. When the current non-nesting interrupt returns after this call,
 571 * the CPU won't restore the RCU idle mode.
 572 */
 573void rcu_user_exit_after_irq(void)
 574{
 575        unsigned long flags;
 576        struct rcu_dynticks *rdtp;
 577
 578        local_irq_save(flags);
 579        rdtp = &__get_cpu_var(rcu_dynticks);
 580        /* Ensure we are interrupting an RCU idle mode. */
 581        WARN_ON_ONCE(rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK);
 582        rdtp->dynticks_nesting += DYNTICK_TASK_EXIT_IDLE;
 583        local_irq_restore(flags);
 584}
 585#endif /* CONFIG_RCU_USER_QS */
 586
 587/**
 588 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
 589 *
 590 * Enter an interrupt handler, which might possibly result in exiting
 591 * idle mode, in other words, entering the mode in which read-side critical
 592 * sections can occur.
 593 *
 594 * Note that the Linux kernel is fully capable of entering an interrupt
 595 * handler that it never exits, for example when doing upcalls to
 596 * user mode!  This code assumes that the idle loop never does upcalls to
 597 * user mode.  If your architecture does do upcalls from the idle loop (or
 598 * does anything else that results in unbalanced calls to the irq_enter()
 599 * and irq_exit() functions), RCU will give you what you deserve, good
 600 * and hard.  But very infrequently and irreproducibly.
 601 *
 602 * Use things like work queues to work around this limitation.
 603 *
 604 * You have been warned.
 605 */
 606void rcu_irq_enter(void)
 607{
 608        unsigned long flags;
 609        struct rcu_dynticks *rdtp;
 610        long long oldval;
 611
 612        local_irq_save(flags);
 613        rdtp = &__get_cpu_var(rcu_dynticks);
 614        oldval = rdtp->dynticks_nesting;
 615        rdtp->dynticks_nesting++;
 616        WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
 617        if (oldval)
 618                trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
 619        else
 620                rcu_eqs_exit_common(rdtp, oldval, true);
 621        local_irq_restore(flags);
 622}
 623
 624/**
 625 * rcu_nmi_enter - inform RCU of entry to NMI context
 626 *
 627 * If the CPU was idle with dynamic ticks active, and there is no
 628 * irq handler running, this updates rdtp->dynticks_nmi to let the
 629 * RCU grace-period handling know that the CPU is active.
 630 */
 631void rcu_nmi_enter(void)
 632{
 633        struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
 634
 635        if (rdtp->dynticks_nmi_nesting == 0 &&
 636            (atomic_read(&rdtp->dynticks) & 0x1))
 637                return;
 638        rdtp->dynticks_nmi_nesting++;
 639        smp_mb__before_atomic_inc();  /* Force delay from prior write. */
 640        atomic_inc(&rdtp->dynticks);
 641        /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
 642        smp_mb__after_atomic_inc();  /* See above. */
 643        WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
 644}
 645
 646/**
 647 * rcu_nmi_exit - inform RCU of exit from NMI context
 648 *
 649 * If the CPU was idle with dynamic ticks active, and there is no
 650 * irq handler running, this updates rdtp->dynticks_nmi to let the
 651 * RCU grace-period handling know that the CPU is no longer active.
 652 */
 653void rcu_nmi_exit(void)
 654{
 655        struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
 656
 657        if (rdtp->dynticks_nmi_nesting == 0 ||
 658            --rdtp->dynticks_nmi_nesting != 0)
 659                return;
 660        /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
 661        smp_mb__before_atomic_inc();  /* See above. */
 662        atomic_inc(&rdtp->dynticks);
 663        smp_mb__after_atomic_inc();  /* Force delay to next write. */
 664        WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
 665}
 666
 667/**
 668 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
 669 *
 670 * If the current CPU is in its idle loop and is neither in an interrupt
 671 * or NMI handler, return true.
 672 */
 673int rcu_is_cpu_idle(void)
 674{
 675        int ret;
 676
 677        preempt_disable();
 678        ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
 679        preempt_enable();
 680        return ret;
 681}
 682EXPORT_SYMBOL(rcu_is_cpu_idle);
 683
 684#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
 685
 686/*
 687 * Is the current CPU online?  Disable preemption to avoid false positives
 688 * that could otherwise happen due to the current CPU number being sampled,
 689 * this task being preempted, its old CPU being taken offline, resuming
 690 * on some other CPU, then determining that its old CPU is now offline.
 691 * It is OK to use RCU on an offline processor during initial boot, hence
 692 * the check for rcu_scheduler_fully_active.  Note also that it is OK
 693 * for a CPU coming online to use RCU for one jiffy prior to marking itself
 694 * online in the cpu_online_mask.  Similarly, it is OK for a CPU going
 695 * offline to continue to use RCU for one jiffy after marking itself
 696 * offline in the cpu_online_mask.  This leniency is necessary given the
 697 * non-atomic nature of the online and offline processing, for example,
 698 * the fact that a CPU enters the scheduler after completing the CPU_DYING
 699 * notifiers.
 700 *
 701 * This is also why RCU internally marks CPUs online during the
 702 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
 703 *
 704 * Disable checking if in an NMI handler because we cannot safely report
 705 * errors from NMI handlers anyway.
 706 */
 707bool rcu_lockdep_current_cpu_online(void)
 708{
 709        struct rcu_data *rdp;
 710        struct rcu_node *rnp;
 711        bool ret;
 712
 713        if (in_nmi())
 714                return 1;
 715        preempt_disable();
 716        rdp = &__get_cpu_var(rcu_sched_data);
 717        rnp = rdp->mynode;
 718        ret = (rdp->grpmask & rnp->qsmaskinit) ||
 719              !rcu_scheduler_fully_active;
 720        preempt_enable();
 721        return ret;
 722}
 723EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
 724
 725#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
 726
 727/**
 728 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
 729 *
 730 * If the current CPU is idle or running at a first-level (not nested)
 731 * interrupt from idle, return true.  The caller must have at least
 732 * disabled preemption.
 733 */
 734static int rcu_is_cpu_rrupt_from_idle(void)
 735{
 736        return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
 737}
 738
 739/*
 740 * Snapshot the specified CPU's dynticks counter so that we can later
 741 * credit them with an implicit quiescent state.  Return 1 if this CPU
 742 * is in dynticks idle mode, which is an extended quiescent state.
 743 */
 744static int dyntick_save_progress_counter(struct rcu_data *rdp)
 745{
 746        rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
 747        return (rdp->dynticks_snap & 0x1) == 0;
 748}
 749
 750/*
 751 * Return true if the specified CPU has passed through a quiescent
 752 * state by virtue of being in or having passed through an dynticks
 753 * idle state since the last call to dyntick_save_progress_counter()
 754 * for this same CPU, or by virtue of having been offline.
 755 */
 756static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
 757{
 758        unsigned int curr;
 759        unsigned int snap;
 760
 761        curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
 762        snap = (unsigned int)rdp->dynticks_snap;
 763
 764        /*
 765         * If the CPU passed through or entered a dynticks idle phase with
 766         * no active irq/NMI handlers, then we can safely pretend that the CPU
 767         * already acknowledged the request to pass through a quiescent
 768         * state.  Either way, that CPU cannot possibly be in an RCU
 769         * read-side critical section that started before the beginning
 770         * of the current RCU grace period.
 771         */
 772        if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
 773                trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
 774                rdp->dynticks_fqs++;
 775                return 1;
 776        }
 777
 778        /*
 779         * Check for the CPU being offline, but only if the grace period
 780         * is old enough.  We don't need to worry about the CPU changing
 781         * state: If we see it offline even once, it has been through a
 782         * quiescent state.
 783         *
 784         * The reason for insisting that the grace period be at least
 785         * one jiffy old is that CPUs that are not quite online and that
 786         * have just gone offline can still execute RCU read-side critical
 787         * sections.
 788         */
 789        if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
 790                return 0;  /* Grace period is not old enough. */
 791        barrier();
 792        if (cpu_is_offline(rdp->cpu)) {
 793                trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
 794                rdp->offline_fqs++;
 795                return 1;
 796        }
 797        return 0;
 798}
 799
 800static void record_gp_stall_check_time(struct rcu_state *rsp)
 801{
 802        rsp->gp_start = jiffies;
 803        rsp->jiffies_stall = jiffies + rcu_jiffies_till_stall_check();
 804}
 805
 806/*
 807 * Dump stacks of all tasks running on stalled CPUs.  This is a fallback
 808 * for architectures that do not implement trigger_all_cpu_backtrace().
 809 * The NMI-triggered stack traces are more accurate because they are
 810 * printed by the target CPU.
 811 */
 812static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
 813{
 814        int cpu;
 815        unsigned long flags;
 816        struct rcu_node *rnp;
 817
 818        rcu_for_each_leaf_node(rsp, rnp) {
 819                raw_spin_lock_irqsave(&rnp->lock, flags);
 820                if (rnp->qsmask != 0) {
 821                        for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
 822                                if (rnp->qsmask & (1UL << cpu))
 823                                        dump_cpu_task(rnp->grplo + cpu);
 824                }
 825                raw_spin_unlock_irqrestore(&rnp->lock, flags);
 826        }
 827}
 828
 829static void print_other_cpu_stall(struct rcu_state *rsp)
 830{
 831        int cpu;
 832        long delta;
 833        unsigned long flags;
 834        int ndetected = 0;
 835        struct rcu_node *rnp = rcu_get_root(rsp);
 836        long totqlen = 0;
 837
 838        /* Only let one CPU complain about others per time interval. */
 839
 840        raw_spin_lock_irqsave(&rnp->lock, flags);
 841        delta = jiffies - rsp->jiffies_stall;
 842        if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
 843                raw_spin_unlock_irqrestore(&rnp->lock, flags);
 844                return;
 845        }
 846        rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
 847        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 848
 849        /*
 850         * OK, time to rat on our buddy...
 851         * See Documentation/RCU/stallwarn.txt for info on how to debug
 852         * RCU CPU stall warnings.
 853         */
 854        printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
 855               rsp->name);
 856        print_cpu_stall_info_begin();
 857        rcu_for_each_leaf_node(rsp, rnp) {
 858                raw_spin_lock_irqsave(&rnp->lock, flags);
 859                ndetected += rcu_print_task_stall(rnp);
 860                if (rnp->qsmask != 0) {
 861                        for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
 862                                if (rnp->qsmask & (1UL << cpu)) {
 863                                        print_cpu_stall_info(rsp,
 864                                                             rnp->grplo + cpu);
 865                                        ndetected++;
 866                                }
 867                }
 868                raw_spin_unlock_irqrestore(&rnp->lock, flags);
 869        }
 870
 871        /*
 872         * Now rat on any tasks that got kicked up to the root rcu_node
 873         * due to CPU offlining.
 874         */
 875        rnp = rcu_get_root(rsp);
 876        raw_spin_lock_irqsave(&rnp->lock, flags);
 877        ndetected += rcu_print_task_stall(rnp);
 878        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 879
 880        print_cpu_stall_info_end();
 881        for_each_possible_cpu(cpu)
 882                totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
 883        pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
 884               smp_processor_id(), (long)(jiffies - rsp->gp_start),
 885               rsp->gpnum, rsp->completed, totqlen);
 886        if (ndetected == 0)
 887                printk(KERN_ERR "INFO: Stall ended before state dump start\n");
 888        else if (!trigger_all_cpu_backtrace())
 889                rcu_dump_cpu_stacks(rsp);
 890
 891        /* Complain about tasks blocking the grace period. */
 892
 893        rcu_print_detail_task_stall(rsp);
 894
 895        force_quiescent_state(rsp);  /* Kick them all. */
 896}
 897
 898static void print_cpu_stall(struct rcu_state *rsp)
 899{
 900        int cpu;
 901        unsigned long flags;
 902        struct rcu_node *rnp = rcu_get_root(rsp);
 903        long totqlen = 0;
 904
 905        /*
 906         * OK, time to rat on ourselves...
 907         * See Documentation/RCU/stallwarn.txt for info on how to debug
 908         * RCU CPU stall warnings.
 909         */
 910        printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
 911        print_cpu_stall_info_begin();
 912        print_cpu_stall_info(rsp, smp_processor_id());
 913        print_cpu_stall_info_end();
 914        for_each_possible_cpu(cpu)
 915                totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
 916        pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
 917                jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
 918        if (!trigger_all_cpu_backtrace())
 919                dump_stack();
 920
 921        raw_spin_lock_irqsave(&rnp->lock, flags);
 922        if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
 923                rsp->jiffies_stall = jiffies +
 924                                     3 * rcu_jiffies_till_stall_check() + 3;
 925        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 926
 927        set_need_resched();  /* kick ourselves to get things going. */
 928}
 929
 930static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
 931{
 932        unsigned long j;
 933        unsigned long js;
 934        struct rcu_node *rnp;
 935
 936        if (rcu_cpu_stall_suppress)
 937                return;
 938        j = ACCESS_ONCE(jiffies);
 939        js = ACCESS_ONCE(rsp->jiffies_stall);
 940        rnp = rdp->mynode;
 941        if (rcu_gp_in_progress(rsp) &&
 942            (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
 943
 944                /* We haven't checked in, so go dump stack. */
 945                print_cpu_stall(rsp);
 946
 947        } else if (rcu_gp_in_progress(rsp) &&
 948                   ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
 949
 950                /* They had a few time units to dump stack, so complain. */
 951                print_other_cpu_stall(rsp);
 952        }
 953}
 954
 955/**
 956 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
 957 *
 958 * Set the stall-warning timeout way off into the future, thus preventing
 959 * any RCU CPU stall-warning messages from appearing in the current set of
 960 * RCU grace periods.
 961 *
 962 * The caller must disable hard irqs.
 963 */
 964void rcu_cpu_stall_reset(void)
 965{
 966        struct rcu_state *rsp;
 967
 968        for_each_rcu_flavor(rsp)
 969                rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
 970}
 971
 972/*
 973 * Update CPU-local rcu_data state to record the newly noticed grace period.
 974 * This is used both when we started the grace period and when we notice
 975 * that someone else started the grace period.  The caller must hold the
 976 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
 977 *  and must have irqs disabled.
 978 */
 979static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
 980{
 981        if (rdp->gpnum != rnp->gpnum) {
 982                /*
 983                 * If the current grace period is waiting for this CPU,
 984                 * set up to detect a quiescent state, otherwise don't
 985                 * go looking for one.
 986                 */
 987                rdp->gpnum = rnp->gpnum;
 988                trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
 989                rdp->passed_quiesce = 0;
 990                rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
 991                zero_cpu_stall_ticks(rdp);
 992        }
 993}
 994
 995static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
 996{
 997        unsigned long flags;
 998        struct rcu_node *rnp;
 999
1000        local_irq_save(flags);
1001        rnp = rdp->mynode;
1002        if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
1003            !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1004                local_irq_restore(flags);
1005                return;
1006        }
1007        __note_new_gpnum(rsp, rnp, rdp);
1008        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1009}
1010
1011/*
1012 * Did someone else start a new RCU grace period start since we last
1013 * checked?  Update local state appropriately if so.  Must be called
1014 * on the CPU corresponding to rdp.
1015 */
1016static int
1017check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
1018{
1019        unsigned long flags;
1020        int ret = 0;
1021
1022        local_irq_save(flags);
1023        if (rdp->gpnum != rsp->gpnum) {
1024                note_new_gpnum(rsp, rdp);
1025                ret = 1;
1026        }
1027        local_irq_restore(flags);
1028        return ret;
1029}
1030
1031/*
1032 * Initialize the specified rcu_data structure's callback list to empty.
1033 */
1034static void init_callback_list(struct rcu_data *rdp)
1035{
1036        int i;
1037
1038        rdp->nxtlist = NULL;
1039        for (i = 0; i < RCU_NEXT_SIZE; i++)
1040                rdp->nxttail[i] = &rdp->nxtlist;
1041        init_nocb_callback_list(rdp);
1042}
1043
1044/*
1045 * Determine the value that ->completed will have at the end of the
1046 * next subsequent grace period.  This is used to tag callbacks so that
1047 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1048 * been dyntick-idle for an extended period with callbacks under the
1049 * influence of RCU_FAST_NO_HZ.
1050 *
1051 * The caller must hold rnp->lock with interrupts disabled.
1052 */
1053static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1054                                       struct rcu_node *rnp)
1055{
1056        /*
1057         * If RCU is idle, we just wait for the next grace period.
1058         * But we can only be sure that RCU is idle if we are looking
1059         * at the root rcu_node structure -- otherwise, a new grace
1060         * period might have started, but just not yet gotten around
1061         * to initializing the current non-root rcu_node structure.
1062         */
1063        if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1064                return rnp->completed + 1;
1065
1066        /*
1067         * Otherwise, wait for a possible partial grace period and
1068         * then the subsequent full grace period.
1069         */
1070        return rnp->completed + 2;
1071}
1072
1073/*
1074 * If there is room, assign a ->completed number to any callbacks on
1075 * this CPU that have not already been assigned.  Also accelerate any
1076 * callbacks that were previously assigned a ->completed number that has
1077 * since proven to be too conservative, which can happen if callbacks get
1078 * assigned a ->completed number while RCU is idle, but with reference to
1079 * a non-root rcu_node structure.  This function is idempotent, so it does
1080 * not hurt to call it repeatedly.
1081 *
1082 * The caller must hold rnp->lock with interrupts disabled.
1083 */
1084static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1085                               struct rcu_data *rdp)
1086{
1087        unsigned long c;
1088        int i;
1089
1090        /* If the CPU has no callbacks, nothing to do. */
1091        if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1092                return;
1093
1094        /*
1095         * Starting from the sublist containing the callbacks most
1096         * recently assigned a ->completed number and working down, find the
1097         * first sublist that is not assignable to an upcoming grace period.
1098         * Such a sublist has something in it (first two tests) and has
1099         * a ->completed number assigned that will complete sooner than
1100         * the ->completed number for newly arrived callbacks (last test).
1101         *
1102         * The key point is that any later sublist can be assigned the
1103         * same ->completed number as the newly arrived callbacks, which
1104         * means that the callbacks in any of these later sublist can be
1105         * grouped into a single sublist, whether or not they have already
1106         * been assigned a ->completed number.
1107         */
1108        c = rcu_cbs_completed(rsp, rnp);
1109        for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1110                if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1111                    !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1112                        break;
1113
1114        /*
1115         * If there are no sublist for unassigned callbacks, leave.
1116         * At the same time, advance "i" one sublist, so that "i" will
1117         * index into the sublist where all the remaining callbacks should
1118         * be grouped into.
1119         */
1120        if (++i >= RCU_NEXT_TAIL)
1121                return;
1122
1123        /*
1124         * Assign all subsequent callbacks' ->completed number to the next
1125         * full grace period and group them all in the sublist initially
1126         * indexed by "i".
1127         */
1128        for (; i <= RCU_NEXT_TAIL; i++) {
1129                rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1130                rdp->nxtcompleted[i] = c;
1131        }
1132
1133        /* Trace depending on how much we were able to accelerate. */
1134        if (!*rdp->nxttail[RCU_WAIT_TAIL])
1135                trace_rcu_grace_period(rsp->name, rdp->gpnum, "AccWaitCB");
1136        else
1137                trace_rcu_grace_period(rsp->name, rdp->gpnum, "AccReadyCB");
1138}
1139
1140/*
1141 * Move any callbacks whose grace period has completed to the
1142 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1143 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1144 * sublist.  This function is idempotent, so it does not hurt to
1145 * invoke it repeatedly.  As long as it is not invoked -too- often...
1146 *
1147 * The caller must hold rnp->lock with interrupts disabled.
1148 */
1149static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1150                            struct rcu_data *rdp)
1151{
1152        int i, j;
1153
1154        /* If the CPU has no callbacks, nothing to do. */
1155        if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1156                return;
1157
1158        /*
1159         * Find all callbacks whose ->completed numbers indicate that they
1160         * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1161         */
1162        for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1163                if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1164                        break;
1165                rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1166        }
1167        /* Clean up any sublist tail pointers that were misordered above. */
1168        for (j = RCU_WAIT_TAIL; j < i; j++)
1169                rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1170
1171        /* Copy down callbacks to fill in empty sublists. */
1172        for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1173                if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1174                        break;
1175                rdp->nxttail[j] = rdp->nxttail[i];
1176                rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1177        }
1178
1179        /* Classify any remaining callbacks. */
1180        rcu_accelerate_cbs(rsp, rnp, rdp);
1181}
1182
1183/*
1184 * Advance this CPU's callbacks, but only if the current grace period
1185 * has ended.  This may be called only from the CPU to whom the rdp
1186 * belongs.  In addition, the corresponding leaf rcu_node structure's
1187 * ->lock must be held by the caller, with irqs disabled.
1188 */
1189static void
1190__rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1191{
1192        /* Did another grace period end? */
1193        if (rdp->completed == rnp->completed) {
1194
1195                /* No, so just accelerate recent callbacks. */
1196                rcu_accelerate_cbs(rsp, rnp, rdp);
1197
1198        } else {
1199
1200                /* Advance callbacks. */
1201                rcu_advance_cbs(rsp, rnp, rdp);
1202
1203                /* Remember that we saw this grace-period completion. */
1204                rdp->completed = rnp->completed;
1205                trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
1206
1207                /*
1208                 * If we were in an extended quiescent state, we may have
1209                 * missed some grace periods that others CPUs handled on
1210                 * our behalf. Catch up with this state to avoid noting
1211                 * spurious new grace periods.  If another grace period
1212                 * has started, then rnp->gpnum will have advanced, so
1213                 * we will detect this later on.  Of course, any quiescent
1214                 * states we found for the old GP are now invalid.
1215                 */
1216                if (ULONG_CMP_LT(rdp->gpnum, rdp->completed)) {
1217                        rdp->gpnum = rdp->completed;
1218                        rdp->passed_quiesce = 0;
1219                }
1220
1221                /*
1222                 * If RCU does not need a quiescent state from this CPU,
1223                 * then make sure that this CPU doesn't go looking for one.
1224                 */
1225                if ((rnp->qsmask & rdp->grpmask) == 0)
1226                        rdp->qs_pending = 0;
1227        }
1228}
1229
1230/*
1231 * Advance this CPU's callbacks, but only if the current grace period
1232 * has ended.  This may be called only from the CPU to whom the rdp
1233 * belongs.
1234 */
1235static void
1236rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
1237{
1238        unsigned long flags;
1239        struct rcu_node *rnp;
1240
1241        local_irq_save(flags);
1242        rnp = rdp->mynode;
1243        if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
1244            !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1245                local_irq_restore(flags);
1246                return;
1247        }
1248        __rcu_process_gp_end(rsp, rnp, rdp);
1249        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1250}
1251
1252/*
1253 * Do per-CPU grace-period initialization for running CPU.  The caller
1254 * must hold the lock of the leaf rcu_node structure corresponding to
1255 * this CPU.
1256 */
1257static void
1258rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1259{
1260        /* Prior grace period ended, so advance callbacks for current CPU. */
1261        __rcu_process_gp_end(rsp, rnp, rdp);
1262
1263        /* Set state so that this CPU will detect the next quiescent state. */
1264        __note_new_gpnum(rsp, rnp, rdp);
1265}
1266
1267/*
1268 * Initialize a new grace period.
1269 */
1270static int rcu_gp_init(struct rcu_state *rsp)
1271{
1272        struct rcu_data *rdp;
1273        struct rcu_node *rnp = rcu_get_root(rsp);
1274
1275        raw_spin_lock_irq(&rnp->lock);
1276        rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1277
1278        if (rcu_gp_in_progress(rsp)) {
1279                /* Grace period already in progress, don't start another.  */
1280                raw_spin_unlock_irq(&rnp->lock);
1281                return 0;
1282        }
1283
1284        /* Advance to a new grace period and initialize state. */
1285        rsp->gpnum++;
1286        trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1287        record_gp_stall_check_time(rsp);
1288        raw_spin_unlock_irq(&rnp->lock);
1289
1290        /* Exclude any concurrent CPU-hotplug operations. */
1291        mutex_lock(&rsp->onoff_mutex);
1292
1293        /*
1294         * Set the quiescent-state-needed bits in all the rcu_node
1295         * structures for all currently online CPUs in breadth-first order,
1296         * starting from the root rcu_node structure, relying on the layout
1297         * of the tree within the rsp->node[] array.  Note that other CPUs
1298         * will access only the leaves of the hierarchy, thus seeing that no
1299         * grace period is in progress, at least until the corresponding
1300         * leaf node has been initialized.  In addition, we have excluded
1301         * CPU-hotplug operations.
1302         *
1303         * The grace period cannot complete until the initialization
1304         * process finishes, because this kthread handles both.
1305         */
1306        rcu_for_each_node_breadth_first(rsp, rnp) {
1307                raw_spin_lock_irq(&rnp->lock);
1308                rdp = this_cpu_ptr(rsp->rda);
1309                rcu_preempt_check_blocked_tasks(rnp);
1310                rnp->qsmask = rnp->qsmaskinit;
1311                rnp->gpnum = rsp->gpnum;
1312                WARN_ON_ONCE(rnp->completed != rsp->completed);
1313                rnp->completed = rsp->completed;
1314                if (rnp == rdp->mynode)
1315                        rcu_start_gp_per_cpu(rsp, rnp, rdp);
1316                rcu_preempt_boost_start_gp(rnp);
1317                trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1318                                            rnp->level, rnp->grplo,
1319                                            rnp->grphi, rnp->qsmask);
1320                raw_spin_unlock_irq(&rnp->lock);
1321#ifdef CONFIG_PROVE_RCU_DELAY
1322                if ((random32() % (rcu_num_nodes * 8)) == 0)
1323                        schedule_timeout_uninterruptible(2);
1324#endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1325                cond_resched();
1326        }
1327
1328        mutex_unlock(&rsp->onoff_mutex);
1329        return 1;
1330}
1331
1332/*
1333 * Do one round of quiescent-state forcing.
1334 */
1335int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1336{
1337        int fqs_state = fqs_state_in;
1338        struct rcu_node *rnp = rcu_get_root(rsp);
1339
1340        rsp->n_force_qs++;
1341        if (fqs_state == RCU_SAVE_DYNTICK) {
1342                /* Collect dyntick-idle snapshots. */
1343                force_qs_rnp(rsp, dyntick_save_progress_counter);
1344                fqs_state = RCU_FORCE_QS;
1345        } else {
1346                /* Handle dyntick-idle and offline CPUs. */
1347                force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1348        }
1349        /* Clear flag to prevent immediate re-entry. */
1350        if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1351                raw_spin_lock_irq(&rnp->lock);
1352                rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1353                raw_spin_unlock_irq(&rnp->lock);
1354        }
1355        return fqs_state;
1356}
1357
1358/*
1359 * Clean up after the old grace period.
1360 */
1361static void rcu_gp_cleanup(struct rcu_state *rsp)
1362{
1363        unsigned long gp_duration;
1364        struct rcu_data *rdp;
1365        struct rcu_node *rnp = rcu_get_root(rsp);
1366
1367        raw_spin_lock_irq(&rnp->lock);
1368        gp_duration = jiffies - rsp->gp_start;
1369        if (gp_duration > rsp->gp_max)
1370                rsp->gp_max = gp_duration;
1371
1372        /*
1373         * We know the grace period is complete, but to everyone else
1374         * it appears to still be ongoing.  But it is also the case
1375         * that to everyone else it looks like there is nothing that
1376         * they can do to advance the grace period.  It is therefore
1377         * safe for us to drop the lock in order to mark the grace
1378         * period as completed in all of the rcu_node structures.
1379         */
1380        raw_spin_unlock_irq(&rnp->lock);
1381
1382        /*
1383         * Propagate new ->completed value to rcu_node structures so
1384         * that other CPUs don't have to wait until the start of the next
1385         * grace period to process their callbacks.  This also avoids
1386         * some nasty RCU grace-period initialization races by forcing
1387         * the end of the current grace period to be completely recorded in
1388         * all of the rcu_node structures before the beginning of the next
1389         * grace period is recorded in any of the rcu_node structures.
1390         */
1391        rcu_for_each_node_breadth_first(rsp, rnp) {
1392                raw_spin_lock_irq(&rnp->lock);
1393                rnp->completed = rsp->gpnum;
1394                raw_spin_unlock_irq(&rnp->lock);
1395                cond_resched();
1396        }
1397        rnp = rcu_get_root(rsp);
1398        raw_spin_lock_irq(&rnp->lock);
1399
1400        rsp->completed = rsp->gpnum; /* Declare grace period done. */
1401        trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1402        rsp->fqs_state = RCU_GP_IDLE;
1403        rdp = this_cpu_ptr(rsp->rda);
1404        if (cpu_needs_another_gp(rsp, rdp))
1405                rsp->gp_flags = 1;
1406        raw_spin_unlock_irq(&rnp->lock);
1407}
1408
1409/*
1410 * Body of kthread that handles grace periods.
1411 */
1412static int __noreturn rcu_gp_kthread(void *arg)
1413{
1414        int fqs_state;
1415        unsigned long j;
1416        int ret;
1417        struct rcu_state *rsp = arg;
1418        struct rcu_node *rnp = rcu_get_root(rsp);
1419
1420        for (;;) {
1421
1422                /* Handle grace-period start. */
1423                for (;;) {
1424                        wait_event_interruptible(rsp->gp_wq,
1425                                                 rsp->gp_flags &
1426                                                 RCU_GP_FLAG_INIT);
1427                        if ((rsp->gp_flags & RCU_GP_FLAG_INIT) &&
1428                            rcu_gp_init(rsp))
1429                                break;
1430                        cond_resched();
1431                        flush_signals(current);
1432                }
1433
1434                /* Handle quiescent-state forcing. */
1435                fqs_state = RCU_SAVE_DYNTICK;
1436                j = jiffies_till_first_fqs;
1437                if (j > HZ) {
1438                        j = HZ;
1439                        jiffies_till_first_fqs = HZ;
1440                }
1441                for (;;) {
1442                        rsp->jiffies_force_qs = jiffies + j;
1443                        ret = wait_event_interruptible_timeout(rsp->gp_wq,
1444                                        (rsp->gp_flags & RCU_GP_FLAG_FQS) ||
1445                                        (!ACCESS_ONCE(rnp->qsmask) &&
1446                                         !rcu_preempt_blocked_readers_cgp(rnp)),
1447                                        j);
1448                        /* If grace period done, leave loop. */
1449                        if (!ACCESS_ONCE(rnp->qsmask) &&
1450                            !rcu_preempt_blocked_readers_cgp(rnp))
1451                                break;
1452                        /* If time for quiescent-state forcing, do it. */
1453                        if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) {
1454                                fqs_state = rcu_gp_fqs(rsp, fqs_state);
1455                                cond_resched();
1456                        } else {
1457                                /* Deal with stray signal. */
1458                                cond_resched();
1459                                flush_signals(current);
1460                        }
1461                        j = jiffies_till_next_fqs;
1462                        if (j > HZ) {
1463                                j = HZ;
1464                                jiffies_till_next_fqs = HZ;
1465                        } else if (j < 1) {
1466                                j = 1;
1467                                jiffies_till_next_fqs = 1;
1468                        }
1469                }
1470
1471                /* Handle grace-period end. */
1472                rcu_gp_cleanup(rsp);
1473        }
1474}
1475
1476/*
1477 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1478 * in preparation for detecting the next grace period.  The caller must hold
1479 * the root node's ->lock, which is released before return.  Hard irqs must
1480 * be disabled.
1481 *
1482 * Note that it is legal for a dying CPU (which is marked as offline) to
1483 * invoke this function.  This can happen when the dying CPU reports its
1484 * quiescent state.
1485 */
1486static void
1487rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1488        __releases(rcu_get_root(rsp)->lock)
1489{
1490        struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1491        struct rcu_node *rnp = rcu_get_root(rsp);
1492
1493        if (!rsp->gp_kthread ||
1494            !cpu_needs_another_gp(rsp, rdp)) {
1495                /*
1496                 * Either we have not yet spawned the grace-period
1497                 * task, this CPU does not need another grace period,
1498                 * or a grace period is already in progress.
1499                 * Either way, don't start a new grace period.
1500                 */
1501                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1502                return;
1503        }
1504
1505        /*
1506         * Because there is no grace period in progress right now,
1507         * any callbacks we have up to this point will be satisfied
1508         * by the next grace period.  So this is a good place to
1509         * assign a grace period number to recently posted callbacks.
1510         */
1511        rcu_accelerate_cbs(rsp, rnp, rdp);
1512
1513        rsp->gp_flags = RCU_GP_FLAG_INIT;
1514        raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */
1515
1516        /* Ensure that CPU is aware of completion of last grace period. */
1517        rcu_process_gp_end(rsp, rdp);
1518        local_irq_restore(flags);
1519
1520        /* Wake up rcu_gp_kthread() to start the grace period. */
1521        wake_up(&rsp->gp_wq);
1522}
1523
1524/*
1525 * Report a full set of quiescent states to the specified rcu_state
1526 * data structure.  This involves cleaning up after the prior grace
1527 * period and letting rcu_start_gp() start up the next grace period
1528 * if one is needed.  Note that the caller must hold rnp->lock, as
1529 * required by rcu_start_gp(), which will release it.
1530 */
1531static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1532        __releases(rcu_get_root(rsp)->lock)
1533{
1534        WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1535        raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1536        wake_up(&rsp->gp_wq);  /* Memory barrier implied by wake_up() path. */
1537}
1538
1539/*
1540 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1541 * Allows quiescent states for a group of CPUs to be reported at one go
1542 * to the specified rcu_node structure, though all the CPUs in the group
1543 * must be represented by the same rcu_node structure (which need not be
1544 * a leaf rcu_node structure, though it often will be).  That structure's
1545 * lock must be held upon entry, and it is released before return.
1546 */
1547static void
1548rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1549                  struct rcu_node *rnp, unsigned long flags)
1550        __releases(rnp->lock)
1551{
1552        struct rcu_node *rnp_c;
1553
1554        /* Walk up the rcu_node hierarchy. */
1555        for (;;) {
1556                if (!(rnp->qsmask & mask)) {
1557
1558                        /* Our bit has already been cleared, so done. */
1559                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1560                        return;
1561                }
1562                rnp->qsmask &= ~mask;
1563                trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1564                                                 mask, rnp->qsmask, rnp->level,
1565                                                 rnp->grplo, rnp->grphi,
1566                                                 !!rnp->gp_tasks);
1567                if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1568
1569                        /* Other bits still set at this level, so done. */
1570                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1571                        return;
1572                }
1573                mask = rnp->grpmask;
1574                if (rnp->parent == NULL) {
1575
1576                        /* No more levels.  Exit loop holding root lock. */
1577
1578                        break;
1579                }
1580                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1581                rnp_c = rnp;
1582                rnp = rnp->parent;
1583                raw_spin_lock_irqsave(&rnp->lock, flags);
1584                WARN_ON_ONCE(rnp_c->qsmask);
1585        }
1586
1587        /*
1588         * Get here if we are the last CPU to pass through a quiescent
1589         * state for this grace period.  Invoke rcu_report_qs_rsp()
1590         * to clean up and start the next grace period if one is needed.
1591         */
1592        rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1593}
1594
1595/*
1596 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1597 * structure.  This must be either called from the specified CPU, or
1598 * called when the specified CPU is known to be offline (and when it is
1599 * also known that no other CPU is concurrently trying to help the offline
1600 * CPU).  The lastcomp argument is used to make sure we are still in the
1601 * grace period of interest.  We don't want to end the current grace period
1602 * based on quiescent states detected in an earlier grace period!
1603 */
1604static void
1605rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1606{
1607        unsigned long flags;
1608        unsigned long mask;
1609        struct rcu_node *rnp;
1610
1611        rnp = rdp->mynode;
1612        raw_spin_lock_irqsave(&rnp->lock, flags);
1613        if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1614            rnp->completed == rnp->gpnum) {
1615
1616                /*
1617                 * The grace period in which this quiescent state was
1618                 * recorded has ended, so don't report it upwards.
1619                 * We will instead need a new quiescent state that lies
1620                 * within the current grace period.
1621                 */
1622                rdp->passed_quiesce = 0;        /* need qs for new gp. */
1623                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1624                return;
1625        }
1626        mask = rdp->grpmask;
1627        if ((rnp->qsmask & mask) == 0) {
1628                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1629        } else {
1630                rdp->qs_pending = 0;
1631
1632                /*
1633                 * This GP can't end until cpu checks in, so all of our
1634                 * callbacks can be processed during the next GP.
1635                 */
1636                rcu_accelerate_cbs(rsp, rnp, rdp);
1637
1638                rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1639        }
1640}
1641
1642/*
1643 * Check to see if there is a new grace period of which this CPU
1644 * is not yet aware, and if so, set up local rcu_data state for it.
1645 * Otherwise, see if this CPU has just passed through its first
1646 * quiescent state for this grace period, and record that fact if so.
1647 */
1648static void
1649rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1650{
1651        /* If there is now a new grace period, record and return. */
1652        if (check_for_new_grace_period(rsp, rdp))
1653                return;
1654
1655        /*
1656         * Does this CPU still need to do its part for current grace period?
1657         * If no, return and let the other CPUs do their part as well.
1658         */
1659        if (!rdp->qs_pending)
1660                return;
1661
1662        /*
1663         * Was there a quiescent state since the beginning of the grace
1664         * period? If no, then exit and wait for the next call.
1665         */
1666        if (!rdp->passed_quiesce)
1667                return;
1668
1669        /*
1670         * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1671         * judge of that).
1672         */
1673        rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1674}
1675
1676#ifdef CONFIG_HOTPLUG_CPU
1677
1678/*
1679 * Send the specified CPU's RCU callbacks to the orphanage.  The
1680 * specified CPU must be offline, and the caller must hold the
1681 * ->orphan_lock.
1682 */
1683static void
1684rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1685                          struct rcu_node *rnp, struct rcu_data *rdp)
1686{
1687        /* No-CBs CPUs do not have orphanable callbacks. */
1688        if (is_nocb_cpu(rdp->cpu))
1689                return;
1690
1691        /*
1692         * Orphan the callbacks.  First adjust the counts.  This is safe
1693         * because _rcu_barrier() excludes CPU-hotplug operations, so it
1694         * cannot be running now.  Thus no memory barrier is required.
1695         */
1696        if (rdp->nxtlist != NULL) {
1697                rsp->qlen_lazy += rdp->qlen_lazy;
1698                rsp->qlen += rdp->qlen;
1699                rdp->n_cbs_orphaned += rdp->qlen;
1700                rdp->qlen_lazy = 0;
1701                ACCESS_ONCE(rdp->qlen) = 0;
1702        }
1703
1704        /*
1705         * Next, move those callbacks still needing a grace period to
1706         * the orphanage, where some other CPU will pick them up.
1707         * Some of the callbacks might have gone partway through a grace
1708         * period, but that is too bad.  They get to start over because we
1709         * cannot assume that grace periods are synchronized across CPUs.
1710         * We don't bother updating the ->nxttail[] array yet, instead
1711         * we just reset the whole thing later on.
1712         */
1713        if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1714                *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1715                rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1716                *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1717        }
1718
1719        /*
1720         * Then move the ready-to-invoke callbacks to the orphanage,
1721         * where some other CPU will pick them up.  These will not be
1722         * required to pass though another grace period: They are done.
1723         */
1724        if (rdp->nxtlist != NULL) {
1725                *rsp->orphan_donetail = rdp->nxtlist;
1726                rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1727        }
1728
1729        /* Finally, initialize the rcu_data structure's list to empty.  */
1730        init_callback_list(rdp);
1731}
1732
1733/*
1734 * Adopt the RCU callbacks from the specified rcu_state structure's
1735 * orphanage.  The caller must hold the ->orphan_lock.
1736 */
1737static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1738{
1739        int i;
1740        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1741
1742        /* No-CBs CPUs are handled specially. */
1743        if (rcu_nocb_adopt_orphan_cbs(rsp, rdp))
1744                return;
1745
1746        /* Do the accounting first. */
1747        rdp->qlen_lazy += rsp->qlen_lazy;
1748        rdp->qlen += rsp->qlen;
1749        rdp->n_cbs_adopted += rsp->qlen;
1750        if (rsp->qlen_lazy != rsp->qlen)
1751                rcu_idle_count_callbacks_posted();
1752        rsp->qlen_lazy = 0;
1753        rsp->qlen = 0;
1754
1755        /*
1756         * We do not need a memory barrier here because the only way we
1757         * can get here if there is an rcu_barrier() in flight is if
1758         * we are the task doing the rcu_barrier().
1759         */
1760
1761        /* First adopt the ready-to-invoke callbacks. */
1762        if (rsp->orphan_donelist != NULL) {
1763                *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1764                *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1765                for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1766                        if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1767                                rdp->nxttail[i] = rsp->orphan_donetail;
1768                rsp->orphan_donelist = NULL;
1769                rsp->orphan_donetail = &rsp->orphan_donelist;
1770        }
1771
1772        /* And then adopt the callbacks that still need a grace period. */
1773        if (rsp->orphan_nxtlist != NULL) {
1774                *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1775                rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1776                rsp->orphan_nxtlist = NULL;
1777                rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1778        }
1779}
1780
1781/*
1782 * Trace the fact that this CPU is going offline.
1783 */
1784static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1785{
1786        RCU_TRACE(unsigned long mask);
1787        RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1788        RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1789
1790        RCU_TRACE(mask = rdp->grpmask);
1791        trace_rcu_grace_period(rsp->name,
1792                               rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1793                               "cpuofl");
1794}
1795
1796/*
1797 * The CPU has been completely removed, and some other CPU is reporting
1798 * this fact from process context.  Do the remainder of the cleanup,
1799 * including orphaning the outgoing CPU's RCU callbacks, and also
1800 * adopting them.  There can only be one CPU hotplug operation at a time,
1801 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1802 */
1803static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1804{
1805        unsigned long flags;
1806        unsigned long mask;
1807        int need_report = 0;
1808        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1809        struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
1810
1811        /* Adjust any no-longer-needed kthreads. */
1812        rcu_boost_kthread_setaffinity(rnp, -1);
1813
1814        /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1815
1816        /* Exclude any attempts to start a new grace period. */
1817        mutex_lock(&rsp->onoff_mutex);
1818        raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
1819
1820        /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1821        rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1822        rcu_adopt_orphan_cbs(rsp);
1823
1824        /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1825        mask = rdp->grpmask;    /* rnp->grplo is constant. */
1826        do {
1827                raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
1828                rnp->qsmaskinit &= ~mask;
1829                if (rnp->qsmaskinit != 0) {
1830                        if (rnp != rdp->mynode)
1831                                raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1832                        break;
1833                }
1834                if (rnp == rdp->mynode)
1835                        need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1836                else
1837                        raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1838                mask = rnp->grpmask;
1839                rnp = rnp->parent;
1840        } while (rnp != NULL);
1841
1842        /*
1843         * We still hold the leaf rcu_node structure lock here, and
1844         * irqs are still disabled.  The reason for this subterfuge is
1845         * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
1846         * held leads to deadlock.
1847         */
1848        raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
1849        rnp = rdp->mynode;
1850        if (need_report & RCU_OFL_TASKS_NORM_GP)
1851                rcu_report_unblock_qs_rnp(rnp, flags);
1852        else
1853                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1854        if (need_report & RCU_OFL_TASKS_EXP_GP)
1855                rcu_report_exp_rnp(rsp, rnp, true);
1856        WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1857                  "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1858                  cpu, rdp->qlen, rdp->nxtlist);
1859        init_callback_list(rdp);
1860        /* Disallow further callbacks on this CPU. */
1861        rdp->nxttail[RCU_NEXT_TAIL] = NULL;
1862        mutex_unlock(&rsp->onoff_mutex);
1863}
1864
1865#else /* #ifdef CONFIG_HOTPLUG_CPU */
1866
1867static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1868{
1869}
1870
1871static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1872{
1873}
1874
1875#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1876
1877/*
1878 * Invoke any RCU callbacks that have made it to the end of their grace
1879 * period.  Thottle as specified by rdp->blimit.
1880 */
1881static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1882{
1883        unsigned long flags;
1884        struct rcu_head *next, *list, **tail;
1885        long bl, count, count_lazy;
1886        int i;
1887
1888        /* If no callbacks are ready, just return. */
1889        if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1890                trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1891                trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1892                                    need_resched(), is_idle_task(current),
1893                                    rcu_is_callbacks_kthread());
1894                return;
1895        }
1896
1897        /*
1898         * Extract the list of ready callbacks, disabling to prevent
1899         * races with call_rcu() from interrupt handlers.
1900         */
1901        local_irq_save(flags);
1902        WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1903        bl = rdp->blimit;
1904        trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1905        list = rdp->nxtlist;
1906        rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1907        *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1908        tail = rdp->nxttail[RCU_DONE_TAIL];
1909        for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1910                if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1911                        rdp->nxttail[i] = &rdp->nxtlist;
1912        local_irq_restore(flags);
1913
1914        /* Invoke callbacks. */
1915        count = count_lazy = 0;
1916        while (list) {
1917                next = list->next;
1918                prefetch(next);
1919                debug_rcu_head_unqueue(list);
1920                if (__rcu_reclaim(rsp->name, list))
1921                        count_lazy++;
1922                list = next;
1923                /* Stop only if limit reached and CPU has something to do. */
1924                if (++count >= bl &&
1925                    (need_resched() ||
1926                     (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1927                        break;
1928        }
1929
1930        local_irq_save(flags);
1931        trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1932                            is_idle_task(current),
1933                            rcu_is_callbacks_kthread());
1934
1935        /* Update count, and requeue any remaining callbacks. */
1936        if (list != NULL) {
1937                *tail = rdp->nxtlist;
1938                rdp->nxtlist = list;
1939                for (i = 0; i < RCU_NEXT_SIZE; i++)
1940                        if (&rdp->nxtlist == rdp->nxttail[i])
1941                                rdp->nxttail[i] = tail;
1942                        else
1943                                break;
1944        }
1945        smp_mb(); /* List handling before counting for rcu_barrier(). */
1946        rdp->qlen_lazy -= count_lazy;
1947        ACCESS_ONCE(rdp->qlen) -= count;
1948        rdp->n_cbs_invoked += count;
1949
1950        /* Reinstate batch limit if we have worked down the excess. */
1951        if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1952                rdp->blimit = blimit;
1953
1954        /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1955        if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1956                rdp->qlen_last_fqs_check = 0;
1957                rdp->n_force_qs_snap = rsp->n_force_qs;
1958        } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1959                rdp->qlen_last_fqs_check = rdp->qlen;
1960        WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
1961
1962        local_irq_restore(flags);
1963
1964        /* Re-invoke RCU core processing if there are callbacks remaining. */
1965        if (cpu_has_callbacks_ready_to_invoke(rdp))
1966                invoke_rcu_core();
1967}
1968
1969/*
1970 * Check to see if this CPU is in a non-context-switch quiescent state
1971 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1972 * Also schedule RCU core processing.
1973 *
1974 * This function must be called from hardirq context.  It is normally
1975 * invoked from the scheduling-clock interrupt.  If rcu_pending returns
1976 * false, there is no point in invoking rcu_check_callbacks().
1977 */
1978void rcu_check_callbacks(int cpu, int user)
1979{
1980        trace_rcu_utilization("Start scheduler-tick");
1981        increment_cpu_stall_ticks();
1982        if (user || rcu_is_cpu_rrupt_from_idle()) {
1983
1984                /*
1985                 * Get here if this CPU took its interrupt from user
1986                 * mode or from the idle loop, and if this is not a
1987                 * nested interrupt.  In this case, the CPU is in
1988                 * a quiescent state, so note it.
1989                 *
1990                 * No memory barrier is required here because both
1991                 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1992                 * variables that other CPUs neither access nor modify,
1993                 * at least not while the corresponding CPU is online.
1994                 */
1995
1996                rcu_sched_qs(cpu);
1997                rcu_bh_qs(cpu);
1998
1999        } else if (!in_softirq()) {
2000
2001                /*
2002                 * Get here if this CPU did not take its interrupt from
2003                 * softirq, in other words, if it is not interrupting
2004                 * a rcu_bh read-side critical section.  This is an _bh
2005                 * critical section, so note it.
2006                 */
2007
2008                rcu_bh_qs(cpu);
2009        }
2010        rcu_preempt_check_callbacks(cpu);
2011        if (rcu_pending(cpu))
2012                invoke_rcu_core();
2013        trace_rcu_utilization("End scheduler-tick");
2014}
2015
2016/*
2017 * Scan the leaf rcu_node structures, processing dyntick state for any that
2018 * have not yet encountered a quiescent state, using the function specified.
2019 * Also initiate boosting for any threads blocked on the root rcu_node.
2020 *
2021 * The caller must have suppressed start of new grace periods.
2022 */
2023static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
2024{
2025        unsigned long bit;
2026        int cpu;
2027        unsigned long flags;
2028        unsigned long mask;
2029        struct rcu_node *rnp;
2030
2031        rcu_for_each_leaf_node(rsp, rnp) {
2032                cond_resched();
2033                mask = 0;
2034                raw_spin_lock_irqsave(&rnp->lock, flags);
2035                if (!rcu_gp_in_progress(rsp)) {
2036                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
2037                        return;
2038                }
2039                if (rnp->qsmask == 0) {
2040                        rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2041                        continue;
2042                }
2043                cpu = rnp->grplo;
2044                bit = 1;
2045                for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2046                        if ((rnp->qsmask & bit) != 0 &&
2047                            f(per_cpu_ptr(rsp->rda, cpu)))
2048                                mask |= bit;
2049                }
2050                if (mask != 0) {
2051
2052                        /* rcu_report_qs_rnp() releases rnp->lock. */
2053                        rcu_report_qs_rnp(mask, rsp, rnp, flags);
2054                        continue;
2055                }
2056                raw_spin_unlock_irqrestore(&rnp->lock, flags);
2057        }
2058        rnp = rcu_get_root(rsp);
2059        if (rnp->qsmask == 0) {
2060                raw_spin_lock_irqsave(&rnp->lock, flags);
2061                rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2062        }
2063}
2064
2065/*
2066 * Force quiescent states on reluctant CPUs, and also detect which
2067 * CPUs are in dyntick-idle mode.
2068 */
2069static void force_quiescent_state(struct rcu_state *rsp)
2070{
2071        unsigned long flags;
2072        bool ret;
2073        struct rcu_node *rnp;
2074        struct rcu_node *rnp_old = NULL;
2075
2076        /* Funnel through hierarchy to reduce memory contention. */
2077        rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2078        for (; rnp != NULL; rnp = rnp->parent) {
2079                ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2080                      !raw_spin_trylock(&rnp->fqslock);
2081                if (rnp_old != NULL)
2082                        raw_spin_unlock(&rnp_old->fqslock);
2083                if (ret) {
2084                        rsp->n_force_qs_lh++;
2085                        return;
2086                }
2087                rnp_old = rnp;
2088        }
2089        /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2090
2091        /* Reached the root of the rcu_node tree, acquire lock. */
2092        raw_spin_lock_irqsave(&rnp_old->lock, flags);
2093        raw_spin_unlock(&rnp_old->fqslock);
2094        if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2095                rsp->n_force_qs_lh++;
2096                raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2097                return;  /* Someone beat us to it. */
2098        }
2099        rsp->gp_flags |= RCU_GP_FLAG_FQS;
2100        raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2101        wake_up(&rsp->gp_wq);  /* Memory barrier implied by wake_up() path. */
2102}
2103
2104/*
2105 * This does the RCU core processing work for the specified rcu_state
2106 * and rcu_data structures.  This may be called only from the CPU to
2107 * whom the rdp belongs.
2108 */
2109static void
2110__rcu_process_callbacks(struct rcu_state *rsp)
2111{
2112        unsigned long flags;
2113        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2114
2115        WARN_ON_ONCE(rdp->beenonline == 0);
2116
2117        /* Handle the end of a grace period that some other CPU ended.  */
2118        rcu_process_gp_end(rsp, rdp);
2119
2120        /* Update RCU state based on any recent quiescent states. */
2121        rcu_check_quiescent_state(rsp, rdp);
2122
2123        /* Does this CPU require a not-yet-started grace period? */
2124        local_irq_save(flags);
2125        if (cpu_needs_another_gp(rsp, rdp)) {
2126                raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2127                rcu_start_gp(rsp, flags);  /* releases above lock */
2128        } else {
2129                local_irq_restore(flags);
2130        }
2131
2132        /* If there are callbacks ready, invoke them. */
2133        if (cpu_has_callbacks_ready_to_invoke(rdp))
2134                invoke_rcu_callbacks(rsp, rdp);
2135}
2136
2137/*
2138 * Do RCU core processing for the current CPU.
2139 */
2140static void rcu_process_callbacks(struct softirq_action *unused)
2141{
2142        struct rcu_state *rsp;
2143
2144        if (cpu_is_offline(smp_processor_id()))
2145                return;
2146        trace_rcu_utilization("Start RCU core");
2147        for_each_rcu_flavor(rsp)
2148                __rcu_process_callbacks(rsp);
2149        trace_rcu_utilization("End RCU core");
2150}
2151
2152/*
2153 * Schedule RCU callback invocation.  If the specified type of RCU
2154 * does not support RCU priority boosting, just do a direct call,
2155 * otherwise wake up the per-CPU kernel kthread.  Note that because we
2156 * are running on the current CPU with interrupts disabled, the
2157 * rcu_cpu_kthread_task cannot disappear out from under us.
2158 */
2159static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2160{
2161        if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2162                return;
2163        if (likely(!rsp->boost)) {
2164                rcu_do_batch(rsp, rdp);
2165                return;
2166        }
2167        invoke_rcu_callbacks_kthread();
2168}
2169
2170static void invoke_rcu_core(void)
2171{
2172        raise_softirq(RCU_SOFTIRQ);
2173}
2174
2175/*
2176 * Handle any core-RCU processing required by a call_rcu() invocation.
2177 */
2178static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2179                            struct rcu_head *head, unsigned long flags)
2180{
2181        /*
2182         * If called from an extended quiescent state, invoke the RCU
2183         * core in order to force a re-evaluation of RCU's idleness.
2184         */
2185        if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2186                invoke_rcu_core();
2187
2188        /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2189        if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2190                return;
2191
2192        /*
2193         * Force the grace period if too many callbacks or too long waiting.
2194         * Enforce hysteresis, and don't invoke force_quiescent_state()
2195         * if some other CPU has recently done so.  Also, don't bother
2196         * invoking force_quiescent_state() if the newly enqueued callback
2197         * is the only one waiting for a grace period to complete.
2198         */
2199        if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2200
2201                /* Are we ignoring a completed grace period? */
2202                rcu_process_gp_end(rsp, rdp);
2203                check_for_new_grace_period(rsp, rdp);
2204
2205                /* Start a new grace period if one not already started. */
2206                if (!rcu_gp_in_progress(rsp)) {
2207                        unsigned long nestflag;
2208                        struct rcu_node *rnp_root = rcu_get_root(rsp);
2209
2210                        raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
2211                        rcu_start_gp(rsp, nestflag);  /* rlses rnp_root->lock */
2212                } else {
2213                        /* Give the grace period a kick. */
2214                        rdp->blimit = LONG_MAX;
2215                        if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2216                            *rdp->nxttail[RCU_DONE_TAIL] != head)
2217                                force_quiescent_state(rsp);
2218                        rdp->n_force_qs_snap = rsp->n_force_qs;
2219                        rdp->qlen_last_fqs_check = rdp->qlen;
2220                }
2221        }
2222}
2223
2224/*
2225 * Helper function for call_rcu() and friends.  The cpu argument will
2226 * normally be -1, indicating "currently running CPU".  It may specify
2227 * a CPU only if that CPU is a no-CBs CPU.  Currently, only _rcu_barrier()
2228 * is expected to specify a CPU.
2229 */
2230static void
2231__call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2232           struct rcu_state *rsp, int cpu, bool lazy)
2233{
2234        unsigned long flags;
2235        struct rcu_data *rdp;
2236
2237        WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2238        debug_rcu_head_queue(head);
2239        head->func = func;
2240        head->next = NULL;
2241
2242        /*
2243         * Opportunistically note grace-period endings and beginnings.
2244         * Note that we might see a beginning right after we see an
2245         * end, but never vice versa, since this CPU has to pass through
2246         * a quiescent state betweentimes.
2247         */
2248        local_irq_save(flags);
2249        rdp = this_cpu_ptr(rsp->rda);
2250
2251        /* Add the callback to our list. */
2252        if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2253                int offline;
2254
2255                if (cpu != -1)
2256                        rdp = per_cpu_ptr(rsp->rda, cpu);
2257                offline = !__call_rcu_nocb(rdp, head, lazy);
2258                WARN_ON_ONCE(offline);
2259                /* _call_rcu() is illegal on offline CPU; leak the callback. */
2260                local_irq_restore(flags);
2261                return;
2262        }
2263        ACCESS_ONCE(rdp->qlen)++;
2264        if (lazy)
2265                rdp->qlen_lazy++;
2266        else
2267                rcu_idle_count_callbacks_posted();
2268        smp_mb();  /* Count before adding callback for rcu_barrier(). */
2269        *rdp->nxttail[RCU_NEXT_TAIL] = head;
2270        rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2271
2272        if (__is_kfree_rcu_offset((unsigned long)func))
2273                trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2274                                         rdp->qlen_lazy, rdp->qlen);
2275        else
2276                trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2277
2278        /* Go handle any RCU core processing required. */
2279        __call_rcu_core(rsp, rdp, head, flags);
2280        local_irq_restore(flags);
2281}
2282
2283/*
2284 * Queue an RCU-sched callback for invocation after a grace period.
2285 */
2286void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2287{
2288        __call_rcu(head, func, &rcu_sched_state, -1, 0);
2289}
2290EXPORT_SYMBOL_GPL(call_rcu_sched);
2291
2292/*
2293 * Queue an RCU callback for invocation after a quicker grace period.
2294 */
2295void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2296{
2297        __call_rcu(head, func, &rcu_bh_state, -1, 0);
2298}
2299EXPORT_SYMBOL_GPL(call_rcu_bh);
2300
2301/*
2302 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2303 * any blocking grace-period wait automatically implies a grace period
2304 * if there is only one CPU online at any point time during execution
2305 * of either synchronize_sched() or synchronize_rcu_bh().  It is OK to
2306 * occasionally incorrectly indicate that there are multiple CPUs online
2307 * when there was in fact only one the whole time, as this just adds
2308 * some overhead: RCU still operates correctly.
2309 */
2310static inline int rcu_blocking_is_gp(void)
2311{
2312        int ret;
2313
2314        might_sleep();  /* Check for RCU read-side critical section. */
2315        preempt_disable();
2316        ret = num_online_cpus() <= 1;
2317        preempt_enable();
2318        return ret;
2319}
2320
2321/**
2322 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2323 *
2324 * Control will return to the caller some time after a full rcu-sched
2325 * grace period has elapsed, in other words after all currently executing
2326 * rcu-sched read-side critical sections have completed.   These read-side
2327 * critical sections are delimited by rcu_read_lock_sched() and
2328 * rcu_read_unlock_sched(), and may be nested.  Note that preempt_disable(),
2329 * local_irq_disable(), and so on may be used in place of
2330 * rcu_read_lock_sched().
2331 *
2332 * This means that all preempt_disable code sequences, including NMI and
2333 * non-threaded hardware-interrupt handlers, in progress on entry will
2334 * have completed before this primitive returns.  However, this does not
2335 * guarantee that softirq handlers will have completed, since in some
2336 * kernels, these handlers can run in process context, and can block.
2337 *
2338 * Note that this guarantee implies further memory-ordering guarantees.
2339 * On systems with more than one CPU, when synchronize_sched() returns,
2340 * each CPU is guaranteed to have executed a full memory barrier since the
2341 * end of its last RCU-sched read-side critical section whose beginning
2342 * preceded the call to synchronize_sched().  In addition, each CPU having
2343 * an RCU read-side critical section that extends beyond the return from
2344 * synchronize_sched() is guaranteed to have executed a full memory barrier
2345 * after the beginning of synchronize_sched() and before the beginning of
2346 * that RCU read-side critical section.  Note that these guarantees include
2347 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2348 * that are executing in the kernel.
2349 *
2350 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2351 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2352 * to have executed a full memory barrier during the execution of
2353 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2354 * again only if the system has more than one CPU).
2355 *
2356 * This primitive provides the guarantees made by the (now removed)
2357 * synchronize_kernel() API.  In contrast, synchronize_rcu() only
2358 * guarantees that rcu_read_lock() sections will have completed.
2359 * In "classic RCU", these two guarantees happen to be one and
2360 * the same, but can differ in realtime RCU implementations.
2361 */
2362void synchronize_sched(void)
2363{
2364        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2365                           !lock_is_held(&rcu_lock_map) &&
2366                           !lock_is_held(&rcu_sched_lock_map),
2367                           "Illegal synchronize_sched() in RCU-sched read-side critical section");
2368        if (rcu_blocking_is_gp())
2369                return;
2370        if (rcu_expedited)
2371                synchronize_sched_expedited();
2372        else
2373                wait_rcu_gp(call_rcu_sched);
2374}
2375EXPORT_SYMBOL_GPL(synchronize_sched);
2376
2377/**
2378 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2379 *
2380 * Control will return to the caller some time after a full rcu_bh grace
2381 * period has elapsed, in other words after all currently executing rcu_bh
2382 * read-side critical sections have completed.  RCU read-side critical
2383 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2384 * and may be nested.
2385 *
2386 * See the description of synchronize_sched() for more detailed information
2387 * on memory ordering guarantees.
2388 */
2389void synchronize_rcu_bh(void)
2390{
2391        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2392                           !lock_is_held(&rcu_lock_map) &&
2393                           !lock_is_held(&rcu_sched_lock_map),
2394                           "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2395        if (rcu_blocking_is_gp())
2396                return;
2397        if (rcu_expedited)
2398                synchronize_rcu_bh_expedited();
2399        else
2400                wait_rcu_gp(call_rcu_bh);
2401}
2402EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2403
2404static int synchronize_sched_expedited_cpu_stop(void *data)
2405{
2406        /*
2407         * There must be a full memory barrier on each affected CPU
2408         * between the time that try_stop_cpus() is called and the
2409         * time that it returns.
2410         *
2411         * In the current initial implementation of cpu_stop, the
2412         * above condition is already met when the control reaches
2413         * this point and the following smp_mb() is not strictly
2414         * necessary.  Do smp_mb() anyway for documentation and
2415         * robustness against future implementation changes.
2416         */
2417        smp_mb(); /* See above comment block. */
2418        return 0;
2419}
2420
2421/**
2422 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2423 *
2424 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2425 * approach to force the grace period to end quickly.  This consumes
2426 * significant time on all CPUs and is unfriendly to real-time workloads,
2427 * so is thus not recommended for any sort of common-case code.  In fact,
2428 * if you are using synchronize_sched_expedited() in a loop, please
2429 * restructure your code to batch your updates, and then use a single
2430 * synchronize_sched() instead.
2431 *
2432 * Note that it is illegal to call this function while holding any lock
2433 * that is acquired by a CPU-hotplug notifier.  And yes, it is also illegal
2434 * to call this function from a CPU-hotplug notifier.  Failing to observe
2435 * these restriction will result in deadlock.
2436 *
2437 * This implementation can be thought of as an application of ticket
2438 * locking to RCU, with sync_sched_expedited_started and
2439 * sync_sched_expedited_done taking on the roles of the halves
2440 * of the ticket-lock word.  Each task atomically increments
2441 * sync_sched_expedited_started upon entry, snapshotting the old value,
2442 * then attempts to stop all the CPUs.  If this succeeds, then each
2443 * CPU will have executed a context switch, resulting in an RCU-sched
2444 * grace period.  We are then done, so we use atomic_cmpxchg() to
2445 * update sync_sched_expedited_done to match our snapshot -- but
2446 * only if someone else has not already advanced past our snapshot.
2447 *
2448 * On the other hand, if try_stop_cpus() fails, we check the value
2449 * of sync_sched_expedited_done.  If it has advanced past our
2450 * initial snapshot, then someone else must have forced a grace period
2451 * some time after we took our snapshot.  In this case, our work is
2452 * done for us, and we can simply return.  Otherwise, we try again,
2453 * but keep our initial snapshot for purposes of checking for someone
2454 * doing our work for us.
2455 *
2456 * If we fail too many times in a row, we fall back to synchronize_sched().
2457 */
2458void synchronize_sched_expedited(void)
2459{
2460        long firstsnap, s, snap;
2461        int trycount = 0;
2462        struct rcu_state *rsp = &rcu_sched_state;
2463
2464        /*
2465         * If we are in danger of counter wrap, just do synchronize_sched().
2466         * By allowing sync_sched_expedited_started to advance no more than
2467         * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2468         * that more than 3.5 billion CPUs would be required to force a
2469         * counter wrap on a 32-bit system.  Quite a few more CPUs would of
2470         * course be required on a 64-bit system.
2471         */
2472        if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2473                         (ulong)atomic_long_read(&rsp->expedited_done) +
2474                         ULONG_MAX / 8)) {
2475                synchronize_sched();
2476                atomic_long_inc(&rsp->expedited_wrap);
2477                return;
2478        }
2479
2480        /*
2481         * Take a ticket.  Note that atomic_inc_return() implies a
2482         * full memory barrier.
2483         */
2484        snap = atomic_long_inc_return(&rsp->expedited_start);
2485        firstsnap = snap;
2486        get_online_cpus();
2487        WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2488
2489        /*
2490         * Each pass through the following loop attempts to force a
2491         * context switch on each CPU.
2492         */
2493        while (try_stop_cpus(cpu_online_mask,
2494                             synchronize_sched_expedited_cpu_stop,
2495                             NULL) == -EAGAIN) {
2496                put_online_cpus();
2497                atomic_long_inc(&rsp->expedited_tryfail);
2498
2499                /* Check to see if someone else did our work for us. */
2500                s = atomic_long_read(&rsp->expedited_done);
2501                if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2502                        /* ensure test happens before caller kfree */
2503                        smp_mb__before_atomic_inc(); /* ^^^ */
2504                        atomic_long_inc(&rsp->expedited_workdone1);
2505                        return;
2506                }
2507
2508                /* No joy, try again later.  Or just synchronize_sched(). */
2509                if (trycount++ < 10) {
2510                        udelay(trycount * num_online_cpus());
2511                } else {
2512                        wait_rcu_gp(call_rcu_sched);
2513                        atomic_long_inc(&rsp->expedited_normal);
2514                        return;
2515                }
2516
2517                /* Recheck to see if someone else did our work for us. */
2518                s = atomic_long_read(&rsp->expedited_done);
2519                if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2520                        /* ensure test happens before caller kfree */
2521                        smp_mb__before_atomic_inc(); /* ^^^ */
2522                        atomic_long_inc(&rsp->expedited_workdone2);
2523                        return;
2524                }
2525
2526                /*
2527                 * Refetching sync_sched_expedited_started allows later
2528                 * callers to piggyback on our grace period.  We retry
2529                 * after they started, so our grace period works for them,
2530                 * and they started after our first try, so their grace
2531                 * period works for us.
2532                 */
2533                get_online_cpus();
2534                snap = atomic_long_read(&rsp->expedited_start);
2535                smp_mb(); /* ensure read is before try_stop_cpus(). */
2536        }
2537        atomic_long_inc(&rsp->expedited_stoppedcpus);
2538
2539        /*
2540         * Everyone up to our most recent fetch is covered by our grace
2541         * period.  Update the counter, but only if our work is still
2542         * relevant -- which it won't be if someone who started later
2543         * than we did already did their update.
2544         */
2545        do {
2546                atomic_long_inc(&rsp->expedited_done_tries);
2547                s = atomic_long_read(&rsp->expedited_done);
2548                if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2549                        /* ensure test happens before caller kfree */
2550                        smp_mb__before_atomic_inc(); /* ^^^ */
2551                        atomic_long_inc(&rsp->expedited_done_lost);
2552                        break;
2553                }
2554        } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2555        atomic_long_inc(&rsp->expedited_done_exit);
2556
2557        put_online_cpus();
2558}
2559EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2560
2561/*
2562 * Check to see if there is any immediate RCU-related work to be done
2563 * by the current CPU, for the specified type of RCU, returning 1 if so.
2564 * The checks are in order of increasing expense: checks that can be
2565 * carried out against CPU-local state are performed first.  However,
2566 * we must check for CPU stalls first, else we might not get a chance.
2567 */
2568static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2569{
2570        struct rcu_node *rnp = rdp->mynode;
2571
2572        rdp->n_rcu_pending++;
2573
2574        /* Check for CPU stalls, if enabled. */
2575        check_cpu_stall(rsp, rdp);
2576
2577        /* Is the RCU core waiting for a quiescent state from this CPU? */
2578        if (rcu_scheduler_fully_active &&
2579            rdp->qs_pending && !rdp->passed_quiesce) {
2580                rdp->n_rp_qs_pending++;
2581        } else if (rdp->qs_pending && rdp->passed_quiesce) {
2582                rdp->n_rp_report_qs++;
2583                return 1;
2584        }
2585
2586        /* Does this CPU have callbacks ready to invoke? */
2587        if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2588                rdp->n_rp_cb_ready++;
2589                return 1;
2590        }
2591
2592        /* Has RCU gone idle with this CPU needing another grace period? */
2593        if (cpu_needs_another_gp(rsp, rdp)) {
2594                rdp->n_rp_cpu_needs_gp++;
2595                return 1;
2596        }
2597
2598        /* Has another RCU grace period completed?  */
2599        if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2600                rdp->n_rp_gp_completed++;
2601                return 1;
2602        }
2603
2604        /* Has a new RCU grace period started? */
2605        if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2606                rdp->n_rp_gp_started++;
2607                return 1;
2608        }
2609
2610        /* nothing to do */
2611        rdp->n_rp_need_nothing++;
2612        return 0;
2613}
2614
2615/*
2616 * Check to see if there is any immediate RCU-related work to be done
2617 * by the current CPU, returning 1 if so.  This function is part of the
2618 * RCU implementation; it is -not- an exported member of the RCU API.
2619 */
2620static int rcu_pending(int cpu)
2621{
2622        struct rcu_state *rsp;
2623
2624        for_each_rcu_flavor(rsp)
2625                if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2626                        return 1;
2627        return 0;
2628}
2629
2630/*
2631 * Check to see if any future RCU-related work will need to be done
2632 * by the current CPU, even if none need be done immediately, returning
2633 * 1 if so.
2634 */
2635static int rcu_cpu_has_callbacks(int cpu)
2636{
2637        struct rcu_state *rsp;
2638
2639        /* RCU callbacks either ready or pending? */
2640        for_each_rcu_flavor(rsp)
2641                if (per_cpu_ptr(rsp->rda, cpu)->nxtlist)
2642                        return 1;
2643        return 0;
2644}
2645
2646/*
2647 * Helper function for _rcu_barrier() tracing.  If tracing is disabled,
2648 * the compiler is expected to optimize this away.
2649 */
2650static void _rcu_barrier_trace(struct rcu_state *rsp, char *s,
2651                               int cpu, unsigned long done)
2652{
2653        trace_rcu_barrier(rsp->name, s, cpu,
2654                          atomic_read(&rsp->barrier_cpu_count), done);
2655}
2656
2657/*
2658 * RCU callback function for _rcu_barrier().  If we are last, wake
2659 * up the task executing _rcu_barrier().
2660 */
2661static void rcu_barrier_callback(struct rcu_head *rhp)
2662{
2663        struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2664        struct rcu_state *rsp = rdp->rsp;
2665
2666        if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2667                _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2668                complete(&rsp->barrier_completion);
2669        } else {
2670                _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2671        }
2672}
2673
2674/*
2675 * Called with preemption disabled, and from cross-cpu IRQ context.
2676 */
2677static void rcu_barrier_func(void *type)
2678{
2679        struct rcu_state *rsp = type;
2680        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2681
2682        _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2683        atomic_inc(&rsp->barrier_cpu_count);
2684        rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2685}
2686
2687/*
2688 * Orchestrate the specified type of RCU barrier, waiting for all
2689 * RCU callbacks of the specified type to complete.
2690 */
2691static void _rcu_barrier(struct rcu_state *rsp)
2692{
2693        int cpu;
2694        struct rcu_data *rdp;
2695        unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2696        unsigned long snap_done;
2697
2698        _rcu_barrier_trace(rsp, "Begin", -1, snap);
2699
2700        /* Take mutex to serialize concurrent rcu_barrier() requests. */
2701        mutex_lock(&rsp->barrier_mutex);
2702
2703        /*
2704         * Ensure that all prior references, including to ->n_barrier_done,
2705         * are ordered before the _rcu_barrier() machinery.
2706         */
2707        smp_mb();  /* See above block comment. */
2708
2709        /*
2710         * Recheck ->n_barrier_done to see if others did our work for us.
2711         * This means checking ->n_barrier_done for an even-to-odd-to-even
2712         * transition.  The "if" expression below therefore rounds the old
2713         * value up to the next even number and adds two before comparing.
2714         */
2715        snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2716        _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2717        if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2718                _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2719                smp_mb(); /* caller's subsequent code after above check. */
2720                mutex_unlock(&rsp->barrier_mutex);
2721                return;
2722        }
2723
2724        /*
2725         * Increment ->n_barrier_done to avoid duplicate work.  Use
2726         * ACCESS_ONCE() to prevent the compiler from speculating
2727         * the increment to precede the early-exit check.
2728         */
2729        ACCESS_ONCE(rsp->n_barrier_done)++;
2730        WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2731        _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2732        smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2733
2734        /*
2735         * Initialize the count to one rather than to zero in order to
2736         * avoid a too-soon return to zero in case of a short grace period
2737         * (or preemption of this task).  Exclude CPU-hotplug operations
2738         * to ensure that no offline CPU has callbacks queued.
2739         */
2740        init_completion(&rsp->barrier_completion);
2741        atomic_set(&rsp->barrier_cpu_count, 1);
2742        get_online_cpus();
2743
2744        /*
2745         * Force each CPU with callbacks to register a new callback.
2746         * When that callback is invoked, we will know that all of the
2747         * corresponding CPU's preceding callbacks have been invoked.
2748         */
2749        for_each_possible_cpu(cpu) {
2750                if (!cpu_online(cpu) && !is_nocb_cpu(cpu))
2751                        continue;
2752                rdp = per_cpu_ptr(rsp->rda, cpu);
2753                if (is_nocb_cpu(cpu)) {
2754                        _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
2755                                           rsp->n_barrier_done);
2756                        atomic_inc(&rsp->barrier_cpu_count);
2757                        __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
2758                                   rsp, cpu, 0);
2759                } else if (ACCESS_ONCE(rdp->qlen)) {
2760                        _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2761                                           rsp->n_barrier_done);
2762                        smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2763                } else {
2764                        _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2765                                           rsp->n_barrier_done);
2766                }
2767        }
2768        put_online_cpus();
2769
2770        /*
2771         * Now that we have an rcu_barrier_callback() callback on each
2772         * CPU, and thus each counted, remove the initial count.
2773         */
2774        if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2775                complete(&rsp->barrier_completion);
2776
2777        /* Increment ->n_barrier_done to prevent duplicate work. */
2778        smp_mb(); /* Keep increment after above mechanism. */
2779        ACCESS_ONCE(rsp->n_barrier_done)++;
2780        WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2781        _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2782        smp_mb(); /* Keep increment before caller's subsequent code. */
2783
2784        /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2785        wait_for_completion(&rsp->barrier_completion);
2786
2787        /* Other rcu_barrier() invocations can now safely proceed. */
2788        mutex_unlock(&rsp->barrier_mutex);
2789}
2790
2791/**
2792 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2793 */
2794void rcu_barrier_bh(void)
2795{
2796        _rcu_barrier(&rcu_bh_state);
2797}
2798EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2799
2800/**
2801 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2802 */
2803void rcu_barrier_sched(void)
2804{
2805        _rcu_barrier(&rcu_sched_state);
2806}
2807EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2808
2809/*
2810 * Do boot-time initialization of a CPU's per-CPU RCU data.
2811 */
2812static void __init
2813rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2814{
2815        unsigned long flags;
2816        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2817        struct rcu_node *rnp = rcu_get_root(rsp);
2818
2819        /* Set up local state, ensuring consistent view of global state. */
2820        raw_spin_lock_irqsave(&rnp->lock, flags);
2821        rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2822        init_callback_list(rdp);
2823        rdp->qlen_lazy = 0;
2824        ACCESS_ONCE(rdp->qlen) = 0;
2825        rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2826        WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2827        WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2828        rdp->cpu = cpu;
2829        rdp->rsp = rsp;
2830        rcu_boot_init_nocb_percpu_data(rdp);
2831        raw_spin_unlock_irqrestore(&rnp->lock, flags);
2832}
2833
2834/*
2835 * Initialize a CPU's per-CPU RCU data.  Note that only one online or
2836 * offline event can be happening at a given time.  Note also that we
2837 * can accept some slop in the rsp->completed access due to the fact
2838 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2839 */
2840static void __cpuinit
2841rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2842{
2843        unsigned long flags;
2844        unsigned long mask;
2845        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2846        struct rcu_node *rnp = rcu_get_root(rsp);
2847
2848        /* Exclude new grace periods. */
2849        mutex_lock(&rsp->onoff_mutex);
2850
2851        /* Set up local state, ensuring consistent view of global state. */
2852        raw_spin_lock_irqsave(&rnp->lock, flags);
2853        rdp->beenonline = 1;     /* We have now been online. */
2854        rdp->preemptible = preemptible;
2855        rdp->qlen_last_fqs_check = 0;
2856        rdp->n_force_qs_snap = rsp->n_force_qs;
2857        rdp->blimit = blimit;
2858        init_callback_list(rdp);  /* Re-enable callbacks on this CPU. */
2859        rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2860        atomic_set(&rdp->dynticks->dynticks,
2861                   (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2862        rcu_prepare_for_idle_init(cpu);
2863        raw_spin_unlock(&rnp->lock);            /* irqs remain disabled. */
2864
2865        /* Add CPU to rcu_node bitmasks. */
2866        rnp = rdp->mynode;
2867        mask = rdp->grpmask;
2868        do {
2869                /* Exclude any attempts to start a new GP on small systems. */
2870                raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
2871                rnp->qsmaskinit |= mask;
2872                mask = rnp->grpmask;
2873                if (rnp == rdp->mynode) {
2874                        /*
2875                         * If there is a grace period in progress, we will
2876                         * set up to wait for it next time we run the
2877                         * RCU core code.
2878                         */
2879                        rdp->gpnum = rnp->completed;
2880                        rdp->completed = rnp->completed;
2881                        rdp->passed_quiesce = 0;
2882                        rdp->qs_pending = 0;
2883                        trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2884                }
2885                raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2886                rnp = rnp->parent;
2887        } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2888        local_irq_restore(flags);
2889
2890        mutex_unlock(&rsp->onoff_mutex);
2891}
2892
2893static void __cpuinit rcu_prepare_cpu(int cpu)
2894{
2895        struct rcu_state *rsp;
2896
2897        for_each_rcu_flavor(rsp)
2898                rcu_init_percpu_data(cpu, rsp,
2899                                     strcmp(rsp->name, "rcu_preempt") == 0);
2900}
2901
2902/*
2903 * Handle CPU online/offline notification events.
2904 */
2905static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2906                                    unsigned long action, void *hcpu)
2907{
2908        long cpu = (long)hcpu;
2909        struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2910        struct rcu_node *rnp = rdp->mynode;
2911        struct rcu_state *rsp;
2912        int ret = NOTIFY_OK;
2913
2914        trace_rcu_utilization("Start CPU hotplug");
2915        switch (action) {
2916        case CPU_UP_PREPARE:
2917        case CPU_UP_PREPARE_FROZEN:
2918                rcu_prepare_cpu(cpu);
2919                rcu_prepare_kthreads(cpu);
2920                break;
2921        case CPU_ONLINE:
2922        case CPU_DOWN_FAILED:
2923                rcu_boost_kthread_setaffinity(rnp, -1);
2924                break;
2925        case CPU_DOWN_PREPARE:
2926                if (nocb_cpu_expendable(cpu))
2927                        rcu_boost_kthread_setaffinity(rnp, cpu);
2928                else
2929                        ret = NOTIFY_BAD;
2930                break;
2931        case CPU_DYING:
2932        case CPU_DYING_FROZEN:
2933                /*
2934                 * The whole machine is "stopped" except this CPU, so we can
2935                 * touch any data without introducing corruption. We send the
2936                 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2937                 */
2938                for_each_rcu_flavor(rsp)
2939                        rcu_cleanup_dying_cpu(rsp);
2940                rcu_cleanup_after_idle(cpu);
2941                break;
2942        case CPU_DEAD:
2943        case CPU_DEAD_FROZEN:
2944        case CPU_UP_CANCELED:
2945        case CPU_UP_CANCELED_FROZEN:
2946                for_each_rcu_flavor(rsp)
2947                        rcu_cleanup_dead_cpu(cpu, rsp);
2948                break;
2949        default:
2950                break;
2951        }
2952        trace_rcu_utilization("End CPU hotplug");
2953        return ret;
2954}
2955
2956/*
2957 * Spawn the kthread that handles this RCU flavor's grace periods.
2958 */
2959static int __init rcu_spawn_gp_kthread(void)
2960{
2961        unsigned long flags;
2962        struct rcu_node *rnp;
2963        struct rcu_state *rsp;
2964        struct task_struct *t;
2965
2966        for_each_rcu_flavor(rsp) {
2967                t = kthread_run(rcu_gp_kthread, rsp, rsp->name);
2968                BUG_ON(IS_ERR(t));
2969                rnp = rcu_get_root(rsp);
2970                raw_spin_lock_irqsave(&rnp->lock, flags);
2971                rsp->gp_kthread = t;
2972                raw_spin_unlock_irqrestore(&rnp->lock, flags);
2973                rcu_spawn_nocb_kthreads(rsp);
2974        }
2975        return 0;
2976}
2977early_initcall(rcu_spawn_gp_kthread);
2978
2979/*
2980 * This function is invoked towards the end of the scheduler's initialization
2981 * process.  Before this is called, the idle task might contain
2982 * RCU read-side critical sections (during which time, this idle
2983 * task is booting the system).  After this function is called, the
2984 * idle tasks are prohibited from containing RCU read-side critical
2985 * sections.  This function also enables RCU lockdep checking.
2986 */
2987void rcu_scheduler_starting(void)
2988{
2989        WARN_ON(num_online_cpus() != 1);
2990        WARN_ON(nr_context_switches() > 0);
2991        rcu_scheduler_active = 1;
2992}
2993
2994/*
2995 * Compute the per-level fanout, either using the exact fanout specified
2996 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2997 */
2998#ifdef CONFIG_RCU_FANOUT_EXACT
2999static void __init rcu_init_levelspread(struct rcu_state *rsp)
3000{
3001        int i;
3002
3003        for (i = rcu_num_lvls - 1; i > 0; i--)
3004                rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3005        rsp->levelspread[0] = rcu_fanout_leaf;
3006}
3007#else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3008static void __init rcu_init_levelspread(struct rcu_state *rsp)
3009{
3010        int ccur;
3011        int cprv;
3012        int i;
3013
3014        cprv = nr_cpu_ids;
3015        for (i = rcu_num_lvls - 1; i >= 0; i--) {
3016                ccur = rsp->levelcnt[i];
3017                rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3018                cprv = ccur;
3019        }
3020}
3021#endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3022
3023/*
3024 * Helper function for rcu_init() that initializes one rcu_state structure.
3025 */
3026static void __init rcu_init_one(struct rcu_state *rsp,
3027                struct rcu_data __percpu *rda)
3028{
3029        static char *buf[] = { "rcu_node_0",
3030                               "rcu_node_1",
3031                               "rcu_node_2",
3032                               "rcu_node_3" };  /* Match MAX_RCU_LVLS */
3033        static char *fqs[] = { "rcu_node_fqs_0",
3034                               "rcu_node_fqs_1",
3035                               "rcu_node_fqs_2",
3036                               "rcu_node_fqs_3" };  /* Match MAX_RCU_LVLS */
3037        int cpustride = 1;
3038        int i;
3039        int j;
3040        struct rcu_node *rnp;
3041
3042        BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf));  /* Fix buf[] init! */
3043
3044        /* Silence gcc 4.8 warning about array index out of range. */
3045        if (rcu_num_lvls > RCU_NUM_LVLS)
3046                panic("rcu_init_one: rcu_num_lvls overflow");
3047
3048        /* Initialize the level-tracking arrays. */
3049
3050        for (i = 0; i < rcu_num_lvls; i++)
3051                rsp->levelcnt[i] = num_rcu_lvl[i];
3052        for (i = 1; i < rcu_num_lvls; i++)
3053                rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3054        rcu_init_levelspread(rsp);
3055
3056        /* Initialize the elements themselves, starting from the leaves. */
3057
3058        for (i = rcu_num_lvls - 1; i >= 0; i--) {
3059                cpustride *= rsp->levelspread[i];
3060                rnp = rsp->level[i];
3061                for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3062                        raw_spin_lock_init(&rnp->lock);
3063                        lockdep_set_class_and_name(&rnp->lock,
3064                                                   &rcu_node_class[i], buf[i]);
3065                        raw_spin_lock_init(&rnp->fqslock);
3066                        lockdep_set_class_and_name(&rnp->fqslock,
3067                                                   &rcu_fqs_class[i], fqs[i]);
3068                        rnp->gpnum = rsp->gpnum;
3069                        rnp->completed = rsp->completed;
3070                        rnp->qsmask = 0;
3071                        rnp->qsmaskinit = 0;
3072                        rnp->grplo = j * cpustride;
3073                        rnp->grphi = (j + 1) * cpustride - 1;
3074                        if (rnp->grphi >= NR_CPUS)
3075                                rnp->grphi = NR_CPUS - 1;
3076                        if (i == 0) {
3077                                rnp->grpnum = 0;
3078                                rnp->grpmask = 0;
3079                                rnp->parent = NULL;
3080                        } else {
3081                                rnp->grpnum = j % rsp->levelspread[i - 1];
3082                                rnp->grpmask = 1UL << rnp->grpnum;
3083                                rnp->parent = rsp->level[i - 1] +
3084                                              j / rsp->levelspread[i - 1];
3085                        }
3086                        rnp->level = i;
3087                        INIT_LIST_HEAD(&rnp->blkd_tasks);
3088                }
3089        }
3090
3091        rsp->rda = rda;
3092        init_waitqueue_head(&rsp->gp_wq);
3093        rnp = rsp->level[rcu_num_lvls - 1];
3094        for_each_possible_cpu(i) {
3095                while (i > rnp->grphi)
3096                        rnp++;
3097                per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3098                rcu_boot_init_percpu_data(i, rsp);
3099        }
3100        list_add(&rsp->flavors, &rcu_struct_flavors);
3101}
3102
3103/*
3104 * Compute the rcu_node tree geometry from kernel parameters.  This cannot
3105 * replace the definitions in rcutree.h because those are needed to size
3106 * the ->node array in the rcu_state structure.
3107 */
3108static void __init rcu_init_geometry(void)
3109{
3110        int i;
3111        int j;
3112        int n = nr_cpu_ids;
3113        int rcu_capacity[MAX_RCU_LVLS + 1];
3114
3115        /* If the compile-time values are accurate, just leave. */
3116        if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3117            nr_cpu_ids == NR_CPUS)
3118                return;
3119
3120        /*
3121         * Compute number of nodes that can be handled an rcu_node tree
3122         * with the given number of levels.  Setting rcu_capacity[0] makes
3123         * some of the arithmetic easier.
3124         */
3125        rcu_capacity[0] = 1;
3126        rcu_capacity[1] = rcu_fanout_leaf;
3127        for (i = 2; i <= MAX_RCU_LVLS; i++)
3128                rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3129
3130        /*
3131         * The boot-time rcu_fanout_leaf parameter is only permitted
3132         * to increase the leaf-level fanout, not decrease it.  Of course,
3133         * the leaf-level fanout cannot exceed the number of bits in
3134         * the rcu_node masks.  Finally, the tree must be able to accommodate
3135         * the configured number of CPUs.  Complain and fall back to the
3136         * compile-time values if these limits are exceeded.
3137         */
3138        if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3139            rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3140            n > rcu_capacity[MAX_RCU_LVLS]) {
3141                WARN_ON(1);
3142                return;
3143        }
3144
3145        /* Calculate the number of rcu_nodes at each level of the tree. */
3146        for (i = 1; i <= MAX_RCU_LVLS; i++)
3147                if (n <= rcu_capacity[i]) {
3148                        for (j = 0; j <= i; j++)
3149                                num_rcu_lvl[j] =
3150                                        DIV_ROUND_UP(n, rcu_capacity[i - j]);
3151                        rcu_num_lvls = i;
3152                        for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3153                                num_rcu_lvl[j] = 0;
3154                        break;
3155                }
3156
3157        /* Calculate the total number of rcu_node structures. */
3158        rcu_num_nodes = 0;
3159        for (i = 0; i <= MAX_RCU_LVLS; i++)
3160                rcu_num_nodes += num_rcu_lvl[i];
3161        rcu_num_nodes -= n;
3162}
3163
3164void __init rcu_init(void)
3165{
3166        int cpu;
3167
3168        rcu_bootup_announce();
3169        rcu_init_geometry();
3170        rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3171        rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3172        __rcu_init_preempt();
3173        rcu_init_nocb();
3174         open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3175
3176        /*
3177         * We don't need protection against CPU-hotplug here because
3178         * this is called early in boot, before either interrupts
3179         * or the scheduler are operational.
3180         */
3181        cpu_notifier(rcu_cpu_notify, 0);
3182        for_each_online_cpu(cpu)
3183                rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3184}
3185
3186#include "rcutree_plugin.h"
3187
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.