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, sabbr, 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        .abbr = sabbr, \
  80}
  81
  82struct rcu_state rcu_sched_state =
  83        RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
  84DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
  85
  86struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
  87DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
  88
  89static struct rcu_state *rcu_state;
  90LIST_HEAD(rcu_struct_flavors);
  91
  92/* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
  93static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
  94module_param(rcu_fanout_leaf, int, 0444);
  95int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
  96static int num_rcu_lvl[] = {  /* Number of rcu_nodes at specified level. */
  97        NUM_RCU_LVL_0,
  98        NUM_RCU_LVL_1,
  99        NUM_RCU_LVL_2,
 100        NUM_RCU_LVL_3,
 101        NUM_RCU_LVL_4,
 102};
 103int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
 104
 105/*
 106 * The rcu_scheduler_active variable transitions from zero to one just
 107 * before the first task is spawned.  So when this variable is zero, RCU
 108 * can assume that there is but one task, allowing RCU to (for example)
 109 * optimize synchronize_sched() to a simple barrier().  When this variable
 110 * is one, RCU must actually do all the hard work required to detect real
 111 * grace periods.  This variable is also used to suppress boot-time false
 112 * positives from lockdep-RCU error checking.
 113 */
 114int rcu_scheduler_active __read_mostly;
 115EXPORT_SYMBOL_GPL(rcu_scheduler_active);
 116
 117/*
 118 * The rcu_scheduler_fully_active variable transitions from zero to one
 119 * during the early_initcall() processing, which is after the scheduler
 120 * is capable of creating new tasks.  So RCU processing (for example,
 121 * creating tasks for RCU priority boosting) must be delayed until after
 122 * rcu_scheduler_fully_active transitions from zero to one.  We also
 123 * currently delay invocation of any RCU callbacks until after this point.
 124 *
 125 * It might later prove better for people registering RCU callbacks during
 126 * early boot to take responsibility for these callbacks, but one step at
 127 * a time.
 128 */
 129static int rcu_scheduler_fully_active __read_mostly;
 130
 131#ifdef CONFIG_RCU_BOOST
 132
 133/*
 134 * Control variables for per-CPU and per-rcu_node kthreads.  These
 135 * handle all flavors of RCU.
 136 */
 137static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
 138DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
 139DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
 140DEFINE_PER_CPU(char, rcu_cpu_has_work);
 141
 142#endif /* #ifdef CONFIG_RCU_BOOST */
 143
 144static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
 145static void invoke_rcu_core(void);
 146static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
 147
 148/*
 149 * Track the rcutorture test sequence number and the update version
 150 * number within a given test.  The rcutorture_testseq is incremented
 151 * on every rcutorture module load and unload, so has an odd value
 152 * when a test is running.  The rcutorture_vernum is set to zero
 153 * when rcutorture starts and is incremented on each rcutorture update.
 154 * These variables enable correlating rcutorture output with the
 155 * RCU tracing information.
 156 */
 157unsigned long rcutorture_testseq;
 158unsigned long rcutorture_vernum;
 159
 160/*
 161 * Return true if an RCU grace period is in progress.  The ACCESS_ONCE()s
 162 * permit this function to be invoked without holding the root rcu_node
 163 * structure's ->lock, but of course results can be subject to change.
 164 */
 165static int rcu_gp_in_progress(struct rcu_state *rsp)
 166{
 167        return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
 168}
 169
 170/*
 171 * Note a quiescent state.  Because we do not need to know
 172 * how many quiescent states passed, just if there was at least
 173 * one since the start of the grace period, this just sets a flag.
 174 * The caller must have disabled preemption.
 175 */
 176void rcu_sched_qs(int cpu)
 177{
 178        struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
 179
 180        if (rdp->passed_quiesce == 0)
 181                trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
 182        rdp->passed_quiesce = 1;
 183}
 184
 185void rcu_bh_qs(int cpu)
 186{
 187        struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
 188
 189        if (rdp->passed_quiesce == 0)
 190                trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
 191        rdp->passed_quiesce = 1;
 192}
 193
 194/*
 195 * Note a context switch.  This is a quiescent state for RCU-sched,
 196 * and requires special handling for preemptible RCU.
 197 * The caller must have disabled preemption.
 198 */
 199void rcu_note_context_switch(int cpu)
 200{
 201        trace_rcu_utilization("Start context switch");
 202        rcu_sched_qs(cpu);
 203        rcu_preempt_note_context_switch(cpu);
 204        trace_rcu_utilization("End context switch");
 205}
 206EXPORT_SYMBOL_GPL(rcu_note_context_switch);
 207
 208DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
 209        .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
 210        .dynticks = ATOMIC_INIT(1),
 211};
 212
 213static long blimit = 10;        /* Maximum callbacks per rcu_do_batch. */
 214static long qhimark = 10000;    /* If this many pending, ignore blimit. */
 215static long qlowmark = 100;     /* Once only this many pending, use blimit. */
 216
 217module_param(blimit, long, 0444);
 218module_param(qhimark, long, 0444);
 219module_param(qlowmark, long, 0444);
 220
 221static ulong jiffies_till_first_fqs = RCU_JIFFIES_TILL_FORCE_QS;
 222static ulong jiffies_till_next_fqs = RCU_JIFFIES_TILL_FORCE_QS;
 223
 224module_param(jiffies_till_first_fqs, ulong, 0644);
 225module_param(jiffies_till_next_fqs, ulong, 0644);
 226
 227static void rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
 228                                  struct rcu_data *rdp);
 229static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *));
 230static void force_quiescent_state(struct rcu_state *rsp);
 231static int rcu_pending(int cpu);
 232
 233/*
 234 * Return the number of RCU-sched batches processed thus far for debug & stats.
 235 */
 236long rcu_batches_completed_sched(void)
 237{
 238        return rcu_sched_state.completed;
 239}
 240EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
 241
 242/*
 243 * Return the number of RCU BH batches processed thus far for debug & stats.
 244 */
 245long rcu_batches_completed_bh(void)
 246{
 247        return rcu_bh_state.completed;
 248}
 249EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
 250
 251/*
 252 * Force a quiescent state for RCU BH.
 253 */
 254void rcu_bh_force_quiescent_state(void)
 255{
 256        force_quiescent_state(&rcu_bh_state);
 257}
 258EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
 259
 260/*
 261 * Record the number of times rcutorture tests have been initiated and
 262 * terminated.  This information allows the debugfs tracing stats to be
 263 * correlated to the rcutorture messages, even when the rcutorture module
 264 * is being repeatedly loaded and unloaded.  In other words, we cannot
 265 * store this state in rcutorture itself.
 266 */
 267void rcutorture_record_test_transition(void)
 268{
 269        rcutorture_testseq++;
 270        rcutorture_vernum = 0;
 271}
 272EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
 273
 274/*
 275 * Record the number of writer passes through the current rcutorture test.
 276 * This is also used to correlate debugfs tracing stats with the rcutorture
 277 * messages.
 278 */
 279void rcutorture_record_progress(unsigned long vernum)
 280{
 281        rcutorture_vernum++;
 282}
 283EXPORT_SYMBOL_GPL(rcutorture_record_progress);
 284
 285/*
 286 * Force a quiescent state for RCU-sched.
 287 */
 288void rcu_sched_force_quiescent_state(void)
 289{
 290        force_quiescent_state(&rcu_sched_state);
 291}
 292EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
 293
 294/*
 295 * Does the CPU have callbacks ready to be invoked?
 296 */
 297static int
 298cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
 299{
 300        return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
 301               rdp->nxttail[RCU_DONE_TAIL] != NULL;
 302}
 303
 304/*
 305 * Does the current CPU require a not-yet-started grace period?
 306 * The caller must have disabled interrupts to prevent races with
 307 * normal callback registry.
 308 */
 309static int
 310cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
 311{
 312        int i;
 313
 314        if (rcu_gp_in_progress(rsp))
 315                return 0;  /* No, a grace period is already in progress. */
 316        if (rcu_nocb_needs_gp(rsp))
 317                return 1;  /* Yes, a no-CBs CPU needs one. */
 318        if (!rdp->nxttail[RCU_NEXT_TAIL])
 319                return 0;  /* No, this is a no-CBs (or offline) CPU. */
 320        if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
 321                return 1;  /* Yes, this CPU has newly registered callbacks. */
 322        for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
 323                if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
 324                    ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
 325                                 rdp->nxtcompleted[i]))
 326                        return 1;  /* Yes, CBs for future grace period. */
 327        return 0; /* No grace period needed. */
 328}
 329
 330/*
 331 * Return the root node of the specified rcu_state structure.
 332 */
 333static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
 334{
 335        return &rsp->node[0];
 336}
 337
 338/*
 339 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
 340 *
 341 * If the new value of the ->dynticks_nesting counter now is zero,
 342 * we really have entered idle, and must do the appropriate accounting.
 343 * The caller must have disabled interrupts.
 344 */
 345static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
 346                                bool user)
 347{
 348        trace_rcu_dyntick("Start", oldval, rdtp->dynticks_nesting);
 349        if (!user && !is_idle_task(current)) {
 350                struct task_struct *idle = idle_task(smp_processor_id());
 351
 352                trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
 353                ftrace_dump(DUMP_ORIG);
 354                WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
 355                          current->pid, current->comm,
 356                          idle->pid, idle->comm); /* must be idle task! */
 357        }
 358        rcu_prepare_for_idle(smp_processor_id());
 359        /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
 360        smp_mb__before_atomic_inc();  /* See above. */
 361        atomic_inc(&rdtp->dynticks);
 362        smp_mb__after_atomic_inc();  /* Force ordering with next sojourn. */
 363        WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
 364
 365        /*
 366         * It is illegal to enter an extended quiescent state while
 367         * in an RCU read-side critical section.
 368         */
 369        rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
 370                           "Illegal idle entry in RCU read-side critical section.");
 371        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
 372                           "Illegal idle entry in RCU-bh read-side critical section.");
 373        rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
 374                           "Illegal idle entry in RCU-sched read-side critical section.");
 375}
 376
 377/*
 378 * Enter an RCU extended quiescent state, which can be either the
 379 * idle loop or adaptive-tickless usermode execution.
 380 */
 381static void rcu_eqs_enter(bool user)
 382{
 383        long long oldval;
 384        struct rcu_dynticks *rdtp;
 385
 386        rdtp = &__get_cpu_var(rcu_dynticks);
 387        oldval = rdtp->dynticks_nesting;
 388        WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
 389        if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
 390                rdtp->dynticks_nesting = 0;
 391        else
 392                rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
 393        rcu_eqs_enter_common(rdtp, oldval, user);
 394}
 395
 396/**
 397 * rcu_idle_enter - inform RCU that current CPU is entering idle
 398 *
 399 * Enter idle mode, in other words, -leave- the mode in which RCU
 400 * read-side critical sections can occur.  (Though RCU read-side
 401 * critical sections can occur in irq handlers in idle, a possibility
 402 * handled by irq_enter() and irq_exit().)
 403 *
 404 * We crowbar the ->dynticks_nesting field to zero to allow for
 405 * the possibility of usermode upcalls having messed up our count
 406 * of interrupt nesting level during the prior busy period.
 407 */
 408void rcu_idle_enter(void)
 409{
 410        unsigned long flags;
 411
 412        local_irq_save(flags);
 413        rcu_eqs_enter(false);
 414        local_irq_restore(flags);
 415}
 416EXPORT_SYMBOL_GPL(rcu_idle_enter);
 417
 418#ifdef CONFIG_RCU_USER_QS
 419/**
 420 * rcu_user_enter - inform RCU that we are resuming userspace.
 421 *
 422 * Enter RCU idle mode right before resuming userspace.  No use of RCU
 423 * is permitted between this call and rcu_user_exit(). This way the
 424 * CPU doesn't need to maintain the tick for RCU maintenance purposes
 425 * when the CPU runs in userspace.
 426 */
 427void rcu_user_enter(void)
 428{
 429        rcu_eqs_enter(1);
 430}
 431
 432/**
 433 * rcu_user_enter_after_irq - inform RCU that we are going to resume userspace
 434 * after the current irq returns.
 435 *
 436 * This is similar to rcu_user_enter() but in the context of a non-nesting
 437 * irq. After this call, RCU enters into idle mode when the interrupt
 438 * returns.
 439 */
 440void rcu_user_enter_after_irq(void)
 441{
 442        unsigned long flags;
 443        struct rcu_dynticks *rdtp;
 444
 445        local_irq_save(flags);
 446        rdtp = &__get_cpu_var(rcu_dynticks);
 447        /* Ensure this irq is interrupting a non-idle RCU state.  */
 448        WARN_ON_ONCE(!(rdtp->dynticks_nesting & DYNTICK_TASK_MASK));
 449        rdtp->dynticks_nesting = 1;
 450        local_irq_restore(flags);
 451}
 452#endif /* CONFIG_RCU_USER_QS */
 453
 454/**
 455 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
 456 *
 457 * Exit from an interrupt handler, which might possibly result in entering
 458 * idle mode, in other words, leaving the mode in which read-side critical
 459 * sections can occur.
 460 *
 461 * This code assumes that the idle loop never does anything that might
 462 * result in unbalanced calls to irq_enter() and irq_exit().  If your
 463 * architecture violates this assumption, RCU will give you what you
 464 * deserve, good and hard.  But very infrequently and irreproducibly.
 465 *
 466 * Use things like work queues to work around this limitation.
 467 *
 468 * You have been warned.
 469 */
 470void rcu_irq_exit(void)
 471{
 472        unsigned long flags;
 473        long long oldval;
 474        struct rcu_dynticks *rdtp;
 475
 476        local_irq_save(flags);
 477        rdtp = &__get_cpu_var(rcu_dynticks);
 478        oldval = rdtp->dynticks_nesting;
 479        rdtp->dynticks_nesting--;
 480        WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
 481        if (rdtp->dynticks_nesting)
 482                trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
 483        else
 484                rcu_eqs_enter_common(rdtp, oldval, true);
 485        local_irq_restore(flags);
 486}
 487
 488/*
 489 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
 490 *
 491 * If the new value of the ->dynticks_nesting counter was previously zero,
 492 * we really have exited idle, and must do the appropriate accounting.
 493 * The caller must have disabled interrupts.
 494 */
 495static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
 496                               int user)
 497{
 498        smp_mb__before_atomic_inc();  /* Force ordering w/previous sojourn. */
 499        atomic_inc(&rdtp->dynticks);
 500        /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
 501        smp_mb__after_atomic_inc();  /* See above. */
 502        WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
 503        rcu_cleanup_after_idle(smp_processor_id());
 504        trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
 505        if (!user && !is_idle_task(current)) {
 506                struct task_struct *idle = idle_task(smp_processor_id());
 507
 508                trace_rcu_dyntick("Error on exit: not idle task",
 509                                  oldval, rdtp->dynticks_nesting);
 510                ftrace_dump(DUMP_ORIG);
 511                WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
 512                          current->pid, current->comm,
 513                          idle->pid, idle->comm); /* must be idle task! */
 514        }
 515}
 516
 517/*
 518 * Exit an RCU extended quiescent state, which can be either the
 519 * idle loop or adaptive-tickless usermode execution.
 520 */
 521static void rcu_eqs_exit(bool user)
 522{
 523        struct rcu_dynticks *rdtp;
 524        long long oldval;
 525
 526        rdtp = &__get_cpu_var(rcu_dynticks);
 527        oldval = rdtp->dynticks_nesting;
 528        WARN_ON_ONCE(oldval < 0);
 529        if (oldval & DYNTICK_TASK_NEST_MASK)
 530                rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
 531        else
 532                rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
 533        rcu_eqs_exit_common(rdtp, oldval, user);
 534}
 535
 536/**
 537 * rcu_idle_exit - inform RCU that current CPU is leaving idle
 538 *
 539 * Exit idle mode, in other words, -enter- the mode in which RCU
 540 * read-side critical sections can occur.
 541 *
 542 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
 543 * allow for the possibility of usermode upcalls messing up our count
 544 * of interrupt nesting level during the busy period that is just
 545 * now starting.
 546 */
 547void rcu_idle_exit(void)
 548{
 549        unsigned long flags;
 550
 551        local_irq_save(flags);
 552        rcu_eqs_exit(false);
 553        local_irq_restore(flags);
 554}
 555EXPORT_SYMBOL_GPL(rcu_idle_exit);
 556
 557#ifdef CONFIG_RCU_USER_QS
 558/**
 559 * rcu_user_exit - inform RCU that we are exiting userspace.
 560 *
 561 * Exit RCU idle mode while entering the kernel because it can
 562 * run a RCU read side critical section anytime.
 563 */
 564void rcu_user_exit(void)
 565{
 566        rcu_eqs_exit(1);
 567}
 568
 569/**
 570 * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace
 571 * idle mode after the current non-nesting irq returns.
 572 *
 573 * This is similar to rcu_user_exit() but in the context of an irq.
 574 * This is called when the irq has interrupted a userspace RCU idle mode
 575 * context. When the current non-nesting interrupt returns after this call,
 576 * the CPU won't restore the RCU idle mode.
 577 */
 578void rcu_user_exit_after_irq(void)
 579{
 580        unsigned long flags;
 581        struct rcu_dynticks *rdtp;
 582
 583        local_irq_save(flags);
 584        rdtp = &__get_cpu_var(rcu_dynticks);
 585        /* Ensure we are interrupting an RCU idle mode. */
 586        WARN_ON_ONCE(rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK);
 587        rdtp->dynticks_nesting += DYNTICK_TASK_EXIT_IDLE;
 588        local_irq_restore(flags);
 589}
 590#endif /* CONFIG_RCU_USER_QS */
 591
 592/**
 593 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
 594 *
 595 * Enter an interrupt handler, which might possibly result in exiting
 596 * idle mode, in other words, entering the mode in which read-side critical
 597 * sections can occur.
 598 *
 599 * Note that the Linux kernel is fully capable of entering an interrupt
 600 * handler that it never exits, for example when doing upcalls to
 601 * user mode!  This code assumes that the idle loop never does upcalls to
 602 * user mode.  If your architecture does do upcalls from the idle loop (or
 603 * does anything else that results in unbalanced calls to the irq_enter()
 604 * and irq_exit() functions), RCU will give you what you deserve, good
 605 * and hard.  But very infrequently and irreproducibly.
 606 *
 607 * Use things like work queues to work around this limitation.
 608 *
 609 * You have been warned.
 610 */
 611void rcu_irq_enter(void)
 612{
 613        unsigned long flags;
 614        struct rcu_dynticks *rdtp;
 615        long long oldval;
 616
 617        local_irq_save(flags);
 618        rdtp = &__get_cpu_var(rcu_dynticks);
 619        oldval = rdtp->dynticks_nesting;
 620        rdtp->dynticks_nesting++;
 621        WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
 622        if (oldval)
 623                trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
 624        else
 625                rcu_eqs_exit_common(rdtp, oldval, true);
 626        local_irq_restore(flags);
 627}
 628
 629/**
 630 * rcu_nmi_enter - inform RCU of entry to NMI context
 631 *
 632 * If the CPU was idle with dynamic ticks active, and there is no
 633 * irq handler running, this updates rdtp->dynticks_nmi to let the
 634 * RCU grace-period handling know that the CPU is active.
 635 */
 636void rcu_nmi_enter(void)
 637{
 638        struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
 639
 640        if (rdtp->dynticks_nmi_nesting == 0 &&
 641            (atomic_read(&rdtp->dynticks) & 0x1))
 642                return;
 643        rdtp->dynticks_nmi_nesting++;
 644        smp_mb__before_atomic_inc();  /* Force delay from prior write. */
 645        atomic_inc(&rdtp->dynticks);
 646        /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
 647        smp_mb__after_atomic_inc();  /* See above. */
 648        WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
 649}
 650
 651/**
 652 * rcu_nmi_exit - inform RCU of exit from NMI context
 653 *
 654 * If the CPU was idle with dynamic ticks active, and there is no
 655 * irq handler running, this updates rdtp->dynticks_nmi to let the
 656 * RCU grace-period handling know that the CPU is no longer active.
 657 */
 658void rcu_nmi_exit(void)
 659{
 660        struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
 661
 662        if (rdtp->dynticks_nmi_nesting == 0 ||
 663            --rdtp->dynticks_nmi_nesting != 0)
 664                return;
 665        /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
 666        smp_mb__before_atomic_inc();  /* See above. */
 667        atomic_inc(&rdtp->dynticks);
 668        smp_mb__after_atomic_inc();  /* Force delay to next write. */
 669        WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
 670}
 671
 672/**
 673 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
 674 *
 675 * If the current CPU is in its idle loop and is neither in an interrupt
 676 * or NMI handler, return true.
 677 */
 678int rcu_is_cpu_idle(void)
 679{
 680        int ret;
 681
 682        preempt_disable();
 683        ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
 684        preempt_enable();
 685        return ret;
 686}
 687EXPORT_SYMBOL(rcu_is_cpu_idle);
 688
 689#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
 690
 691/*
 692 * Is the current CPU online?  Disable preemption to avoid false positives
 693 * that could otherwise happen due to the current CPU number being sampled,
 694 * this task being preempted, its old CPU being taken offline, resuming
 695 * on some other CPU, then determining that its old CPU is now offline.
 696 * It is OK to use RCU on an offline processor during initial boot, hence
 697 * the check for rcu_scheduler_fully_active.  Note also that it is OK
 698 * for a CPU coming online to use RCU for one jiffy prior to marking itself
 699 * online in the cpu_online_mask.  Similarly, it is OK for a CPU going
 700 * offline to continue to use RCU for one jiffy after marking itself
 701 * offline in the cpu_online_mask.  This leniency is necessary given the
 702 * non-atomic nature of the online and offline processing, for example,
 703 * the fact that a CPU enters the scheduler after completing the CPU_DYING
 704 * notifiers.
 705 *
 706 * This is also why RCU internally marks CPUs online during the
 707 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
 708 *
 709 * Disable checking if in an NMI handler because we cannot safely report
 710 * errors from NMI handlers anyway.
 711 */
 712bool rcu_lockdep_current_cpu_online(void)
 713{
 714        struct rcu_data *rdp;
 715        struct rcu_node *rnp;
 716        bool ret;
 717
 718        if (in_nmi())
 719                return 1;
 720        preempt_disable();
 721        rdp = &__get_cpu_var(rcu_sched_data);
 722        rnp = rdp->mynode;
 723        ret = (rdp->grpmask & rnp->qsmaskinit) ||
 724              !rcu_scheduler_fully_active;
 725        preempt_enable();
 726        return ret;
 727}
 728EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
 729
 730#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
 731
 732/**
 733 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
 734 *
 735 * If the current CPU is idle or running at a first-level (not nested)
 736 * interrupt from idle, return true.  The caller must have at least
 737 * disabled preemption.
 738 */
 739static int rcu_is_cpu_rrupt_from_idle(void)
 740{
 741        return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
 742}
 743
 744/*
 745 * Snapshot the specified CPU's dynticks counter so that we can later
 746 * credit them with an implicit quiescent state.  Return 1 if this CPU
 747 * is in dynticks idle mode, which is an extended quiescent state.
 748 */
 749static int dyntick_save_progress_counter(struct rcu_data *rdp)
 750{
 751        rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
 752        return (rdp->dynticks_snap & 0x1) == 0;
 753}
 754
 755/*
 756 * Return true if the specified CPU has passed through a quiescent
 757 * state by virtue of being in or having passed through an dynticks
 758 * idle state since the last call to dyntick_save_progress_counter()
 759 * for this same CPU, or by virtue of having been offline.
 760 */
 761static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
 762{
 763        unsigned int curr;
 764        unsigned int snap;
 765
 766        curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
 767        snap = (unsigned int)rdp->dynticks_snap;
 768
 769        /*
 770         * If the CPU passed through or entered a dynticks idle phase with
 771         * no active irq/NMI handlers, then we can safely pretend that the CPU
 772         * already acknowledged the request to pass through a quiescent
 773         * state.  Either way, that CPU cannot possibly be in an RCU
 774         * read-side critical section that started before the beginning
 775         * of the current RCU grace period.
 776         */
 777        if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
 778                trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
 779                rdp->dynticks_fqs++;
 780                return 1;
 781        }
 782
 783        /*
 784         * Check for the CPU being offline, but only if the grace period
 785         * is old enough.  We don't need to worry about the CPU changing
 786         * state: If we see it offline even once, it has been through a
 787         * quiescent state.
 788         *
 789         * The reason for insisting that the grace period be at least
 790         * one jiffy old is that CPUs that are not quite online and that
 791         * have just gone offline can still execute RCU read-side critical
 792         * sections.
 793         */
 794        if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
 795                return 0;  /* Grace period is not old enough. */
 796        barrier();
 797        if (cpu_is_offline(rdp->cpu)) {
 798                trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
 799                rdp->offline_fqs++;
 800                return 1;
 801        }
 802
 803        /*
 804         * There is a possibility that a CPU in adaptive-ticks state
 805         * might run in the kernel with the scheduling-clock tick disabled
 806         * for an extended time period.  Invoke rcu_kick_nohz_cpu() to
 807         * force the CPU to restart the scheduling-clock tick in this
 808         * CPU is in this state.
 809         */
 810        rcu_kick_nohz_cpu(rdp->cpu);
 811
 812        return 0;
 813}
 814
 815static void record_gp_stall_check_time(struct rcu_state *rsp)
 816{
 817        rsp->gp_start = jiffies;
 818        rsp->jiffies_stall = jiffies + rcu_jiffies_till_stall_check();
 819}
 820
 821/*
 822 * Dump stacks of all tasks running on stalled CPUs.  This is a fallback
 823 * for architectures that do not implement trigger_all_cpu_backtrace().
 824 * The NMI-triggered stack traces are more accurate because they are
 825 * printed by the target CPU.
 826 */
 827static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
 828{
 829        int cpu;
 830        unsigned long flags;
 831        struct rcu_node *rnp;
 832
 833        rcu_for_each_leaf_node(rsp, rnp) {
 834                raw_spin_lock_irqsave(&rnp->lock, flags);
 835                if (rnp->qsmask != 0) {
 836                        for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
 837                                if (rnp->qsmask & (1UL << cpu))
 838                                        dump_cpu_task(rnp->grplo + cpu);
 839                }
 840                raw_spin_unlock_irqrestore(&rnp->lock, flags);
 841        }
 842}
 843
 844static void print_other_cpu_stall(struct rcu_state *rsp)
 845{
 846        int cpu;
 847        long delta;
 848        unsigned long flags;
 849        int ndetected = 0;
 850        struct rcu_node *rnp = rcu_get_root(rsp);
 851        long totqlen = 0;
 852
 853        /* Only let one CPU complain about others per time interval. */
 854
 855        raw_spin_lock_irqsave(&rnp->lock, flags);
 856        delta = jiffies - rsp->jiffies_stall;
 857        if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
 858                raw_spin_unlock_irqrestore(&rnp->lock, flags);
 859                return;
 860        }
 861        rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
 862        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 863
 864        /*
 865         * OK, time to rat on our buddy...
 866         * See Documentation/RCU/stallwarn.txt for info on how to debug
 867         * RCU CPU stall warnings.
 868         */
 869        printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
 870               rsp->name);
 871        print_cpu_stall_info_begin();
 872        rcu_for_each_leaf_node(rsp, rnp) {
 873                raw_spin_lock_irqsave(&rnp->lock, flags);
 874                ndetected += rcu_print_task_stall(rnp);
 875                if (rnp->qsmask != 0) {
 876                        for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
 877                                if (rnp->qsmask & (1UL << cpu)) {
 878                                        print_cpu_stall_info(rsp,
 879                                                             rnp->grplo + cpu);
 880                                        ndetected++;
 881                                }
 882                }
 883                raw_spin_unlock_irqrestore(&rnp->lock, flags);
 884        }
 885
 886        /*
 887         * Now rat on any tasks that got kicked up to the root rcu_node
 888         * due to CPU offlining.
 889         */
 890        rnp = rcu_get_root(rsp);
 891        raw_spin_lock_irqsave(&rnp->lock, flags);
 892        ndetected += rcu_print_task_stall(rnp);
 893        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 894
 895        print_cpu_stall_info_end();
 896        for_each_possible_cpu(cpu)
 897                totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
 898        pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
 899               smp_processor_id(), (long)(jiffies - rsp->gp_start),
 900               rsp->gpnum, rsp->completed, totqlen);
 901        if (ndetected == 0)
 902                printk(KERN_ERR "INFO: Stall ended before state dump start\n");
 903        else if (!trigger_all_cpu_backtrace())
 904                rcu_dump_cpu_stacks(rsp);
 905
 906        /* Complain about tasks blocking the grace period. */
 907
 908        rcu_print_detail_task_stall(rsp);
 909
 910        force_quiescent_state(rsp);  /* Kick them all. */
 911}
 912
 913static void print_cpu_stall(struct rcu_state *rsp)
 914{
 915        int cpu;
 916        unsigned long flags;
 917        struct rcu_node *rnp = rcu_get_root(rsp);
 918        long totqlen = 0;
 919
 920        /*
 921         * OK, time to rat on ourselves...
 922         * See Documentation/RCU/stallwarn.txt for info on how to debug
 923         * RCU CPU stall warnings.
 924         */
 925        printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
 926        print_cpu_stall_info_begin();
 927        print_cpu_stall_info(rsp, smp_processor_id());
 928        print_cpu_stall_info_end();
 929        for_each_possible_cpu(cpu)
 930                totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
 931        pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
 932                jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
 933        if (!trigger_all_cpu_backtrace())
 934                dump_stack();
 935
 936        raw_spin_lock_irqsave(&rnp->lock, flags);
 937        if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
 938                rsp->jiffies_stall = jiffies +
 939                                     3 * rcu_jiffies_till_stall_check() + 3;
 940        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 941
 942        set_need_resched();  /* kick ourselves to get things going. */
 943}
 944
 945static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
 946{
 947        unsigned long j;
 948        unsigned long js;
 949        struct rcu_node *rnp;
 950
 951        if (rcu_cpu_stall_suppress)
 952                return;
 953        j = ACCESS_ONCE(jiffies);
 954        js = ACCESS_ONCE(rsp->jiffies_stall);
 955        rnp = rdp->mynode;
 956        if (rcu_gp_in_progress(rsp) &&
 957            (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
 958
 959                /* We haven't checked in, so go dump stack. */
 960                print_cpu_stall(rsp);
 961
 962        } else if (rcu_gp_in_progress(rsp) &&
 963                   ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
 964
 965                /* They had a few time units to dump stack, so complain. */
 966                print_other_cpu_stall(rsp);
 967        }
 968}
 969
 970/**
 971 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
 972 *
 973 * Set the stall-warning timeout way off into the future, thus preventing
 974 * any RCU CPU stall-warning messages from appearing in the current set of
 975 * RCU grace periods.
 976 *
 977 * The caller must disable hard irqs.
 978 */
 979void rcu_cpu_stall_reset(void)
 980{
 981        struct rcu_state *rsp;
 982
 983        for_each_rcu_flavor(rsp)
 984                rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
 985}
 986
 987/*
 988 * Update CPU-local rcu_data state to record the newly noticed grace period.
 989 * This is used both when we started the grace period and when we notice
 990 * that someone else started the grace period.  The caller must hold the
 991 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
 992 *  and must have irqs disabled.
 993 */
 994static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
 995{
 996        if (rdp->gpnum != rnp->gpnum) {
 997                /*
 998                 * If the current grace period is waiting for this CPU,
 999                 * set up to detect a quiescent state, otherwise don't
1000                 * go looking for one.
1001                 */
1002                rdp->gpnum = rnp->gpnum;
1003                trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
1004                rdp->passed_quiesce = 0;
1005                rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1006                zero_cpu_stall_ticks(rdp);
1007        }
1008}
1009
1010static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
1011{
1012        unsigned long flags;
1013        struct rcu_node *rnp;
1014
1015        local_irq_save(flags);
1016        rnp = rdp->mynode;
1017        if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
1018            !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1019                local_irq_restore(flags);
1020                return;
1021        }
1022        __note_new_gpnum(rsp, rnp, rdp);
1023        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1024}
1025
1026/*
1027 * Did someone else start a new RCU grace period start since we last
1028 * checked?  Update local state appropriately if so.  Must be called
1029 * on the CPU corresponding to rdp.
1030 */
1031static int
1032check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
1033{
1034        unsigned long flags;
1035        int ret = 0;
1036
1037        local_irq_save(flags);
1038        if (rdp->gpnum != rsp->gpnum) {
1039                note_new_gpnum(rsp, rdp);
1040                ret = 1;
1041        }
1042        local_irq_restore(flags);
1043        return ret;
1044}
1045
1046/*
1047 * Initialize the specified rcu_data structure's callback list to empty.
1048 */
1049static void init_callback_list(struct rcu_data *rdp)
1050{
1051        int i;
1052
1053        if (init_nocb_callback_list(rdp))
1054                return;
1055        rdp->nxtlist = NULL;
1056        for (i = 0; i < RCU_NEXT_SIZE; i++)
1057                rdp->nxttail[i] = &rdp->nxtlist;
1058}
1059
1060/*
1061 * Determine the value that ->completed will have at the end of the
1062 * next subsequent grace period.  This is used to tag callbacks so that
1063 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1064 * been dyntick-idle for an extended period with callbacks under the
1065 * influence of RCU_FAST_NO_HZ.
1066 *
1067 * The caller must hold rnp->lock with interrupts disabled.
1068 */
1069static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1070                                       struct rcu_node *rnp)
1071{
1072        /*
1073         * If RCU is idle, we just wait for the next grace period.
1074         * But we can only be sure that RCU is idle if we are looking
1075         * at the root rcu_node structure -- otherwise, a new grace
1076         * period might have started, but just not yet gotten around
1077         * to initializing the current non-root rcu_node structure.
1078         */
1079        if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1080                return rnp->completed + 1;
1081
1082        /*
1083         * Otherwise, wait for a possible partial grace period and
1084         * then the subsequent full grace period.
1085         */
1086        return rnp->completed + 2;
1087}
1088
1089/*
1090 * Trace-event helper function for rcu_start_future_gp() and
1091 * rcu_nocb_wait_gp().
1092 */
1093static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1094                                unsigned long c, char *s)
1095{
1096        trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1097                                      rnp->completed, c, rnp->level,
1098                                      rnp->grplo, rnp->grphi, s);
1099}
1100
1101/*
1102 * Start some future grace period, as needed to handle newly arrived
1103 * callbacks.  The required future grace periods are recorded in each
1104 * rcu_node structure's ->need_future_gp field.
1105 *
1106 * The caller must hold the specified rcu_node structure's ->lock.
1107 */
1108static unsigned long __maybe_unused
1109rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
1110{
1111        unsigned long c;
1112        int i;
1113        struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1114
1115        /*
1116         * Pick up grace-period number for new callbacks.  If this
1117         * grace period is already marked as needed, return to the caller.
1118         */
1119        c = rcu_cbs_completed(rdp->rsp, rnp);
1120        trace_rcu_future_gp(rnp, rdp, c, "Startleaf");
1121        if (rnp->need_future_gp[c & 0x1]) {
1122                trace_rcu_future_gp(rnp, rdp, c, "Prestartleaf");
1123                return c;
1124        }
1125
1126        /*
1127         * If either this rcu_node structure or the root rcu_node structure
1128         * believe that a grace period is in progress, then we must wait
1129         * for the one following, which is in "c".  Because our request
1130         * will be noticed at the end of the current grace period, we don't
1131         * need to explicitly start one.
1132         */
1133        if (rnp->gpnum != rnp->completed ||
1134            ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
1135                rnp->need_future_gp[c & 0x1]++;
1136                trace_rcu_future_gp(rnp, rdp, c, "Startedleaf");
1137                return c;
1138        }
1139
1140        /*
1141         * There might be no grace period in progress.  If we don't already
1142         * hold it, acquire the root rcu_node structure's lock in order to
1143         * start one (if needed).
1144         */
1145        if (rnp != rnp_root)
1146                raw_spin_lock(&rnp_root->lock);
1147
1148        /*
1149         * Get a new grace-period number.  If there really is no grace
1150         * period in progress, it will be smaller than the one we obtained
1151         * earlier.  Adjust callbacks as needed.  Note that even no-CBs
1152         * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1153         */
1154        c = rcu_cbs_completed(rdp->rsp, rnp_root);
1155        for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1156                if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1157                        rdp->nxtcompleted[i] = c;
1158
1159        /*
1160         * If the needed for the required grace period is already
1161         * recorded, trace and leave.
1162         */
1163        if (rnp_root->need_future_gp[c & 0x1]) {
1164                trace_rcu_future_gp(rnp, rdp, c, "Prestartedroot");
1165                goto unlock_out;
1166        }
1167
1168        /* Record the need for the future grace period. */
1169        rnp_root->need_future_gp[c & 0x1]++;
1170
1171        /* If a grace period is not already in progress, start one. */
1172        if (rnp_root->gpnum != rnp_root->completed) {
1173                trace_rcu_future_gp(rnp, rdp, c, "Startedleafroot");
1174        } else {
1175                trace_rcu_future_gp(rnp, rdp, c, "Startedroot");
1176                rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1177        }
1178unlock_out:
1179        if (rnp != rnp_root)
1180                raw_spin_unlock(&rnp_root->lock);
1181        return c;
1182}
1183
1184/*
1185 * Clean up any old requests for the just-ended grace period.  Also return
1186 * whether any additional grace periods have been requested.  Also invoke
1187 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1188 * waiting for this grace period to complete.
1189 */
1190static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1191{
1192        int c = rnp->completed;
1193        int needmore;
1194        struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1195
1196        rcu_nocb_gp_cleanup(rsp, rnp);
1197        rnp->need_future_gp[c & 0x1] = 0;
1198        needmore = rnp->need_future_gp[(c + 1) & 0x1];
1199        trace_rcu_future_gp(rnp, rdp, c, needmore ? "CleanupMore" : "Cleanup");
1200        return needmore;
1201}
1202
1203/*
1204 * If there is room, assign a ->completed number to any callbacks on
1205 * this CPU that have not already been assigned.  Also accelerate any
1206 * callbacks that were previously assigned a ->completed number that has
1207 * since proven to be too conservative, which can happen if callbacks get
1208 * assigned a ->completed number while RCU is idle, but with reference to
1209 * a non-root rcu_node structure.  This function is idempotent, so it does
1210 * not hurt to call it repeatedly.
1211 *
1212 * The caller must hold rnp->lock with interrupts disabled.
1213 */
1214static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1215                               struct rcu_data *rdp)
1216{
1217        unsigned long c;
1218        int i;
1219
1220        /* If the CPU has no callbacks, nothing to do. */
1221        if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1222                return;
1223
1224        /*
1225         * Starting from the sublist containing the callbacks most
1226         * recently assigned a ->completed number and working down, find the
1227         * first sublist that is not assignable to an upcoming grace period.
1228         * Such a sublist has something in it (first two tests) and has
1229         * a ->completed number assigned that will complete sooner than
1230         * the ->completed number for newly arrived callbacks (last test).
1231         *
1232         * The key point is that any later sublist can be assigned the
1233         * same ->completed number as the newly arrived callbacks, which
1234         * means that the callbacks in any of these later sublist can be
1235         * grouped into a single sublist, whether or not they have already
1236         * been assigned a ->completed number.
1237         */
1238        c = rcu_cbs_completed(rsp, rnp);
1239        for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1240                if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1241                    !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1242                        break;
1243
1244        /*
1245         * If there are no sublist for unassigned callbacks, leave.
1246         * At the same time, advance "i" one sublist, so that "i" will
1247         * index into the sublist where all the remaining callbacks should
1248         * be grouped into.
1249         */
1250        if (++i >= RCU_NEXT_TAIL)
1251                return;
1252
1253        /*
1254         * Assign all subsequent callbacks' ->completed number to the next
1255         * full grace period and group them all in the sublist initially
1256         * indexed by "i".
1257         */
1258        for (; i <= RCU_NEXT_TAIL; i++) {
1259                rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1260                rdp->nxtcompleted[i] = c;
1261        }
1262        /* Record any needed additional grace periods. */
1263        rcu_start_future_gp(rnp, rdp);
1264
1265        /* Trace depending on how much we were able to accelerate. */
1266        if (!*rdp->nxttail[RCU_WAIT_TAIL])
1267                trace_rcu_grace_period(rsp->name, rdp->gpnum, "AccWaitCB");
1268        else
1269                trace_rcu_grace_period(rsp->name, rdp->gpnum, "AccReadyCB");
1270}
1271
1272/*
1273 * Move any callbacks whose grace period has completed to the
1274 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1275 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1276 * sublist.  This function is idempotent, so it does not hurt to
1277 * invoke it repeatedly.  As long as it is not invoked -too- often...
1278 *
1279 * The caller must hold rnp->lock with interrupts disabled.
1280 */
1281static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1282                            struct rcu_data *rdp)
1283{
1284        int i, j;
1285
1286        /* If the CPU has no callbacks, nothing to do. */
1287        if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1288                return;
1289
1290        /*
1291         * Find all callbacks whose ->completed numbers indicate that they
1292         * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1293         */
1294        for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1295                if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1296                        break;
1297                rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1298        }
1299        /* Clean up any sublist tail pointers that were misordered above. */
1300        for (j = RCU_WAIT_TAIL; j < i; j++)
1301                rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1302
1303        /* Copy down callbacks to fill in empty sublists. */
1304        for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1305                if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1306                        break;
1307                rdp->nxttail[j] = rdp->nxttail[i];
1308                rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1309        }
1310
1311        /* Classify any remaining callbacks. */
1312        rcu_accelerate_cbs(rsp, rnp, rdp);
1313}
1314
1315/*
1316 * Advance this CPU's callbacks, but only if the current grace period
1317 * has ended.  This may be called only from the CPU to whom the rdp
1318 * belongs.  In addition, the corresponding leaf rcu_node structure's
1319 * ->lock must be held by the caller, with irqs disabled.
1320 */
1321static void
1322__rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1323{
1324        /* Did another grace period end? */
1325        if (rdp->completed == rnp->completed) {
1326
1327                /* No, so just accelerate recent callbacks. */
1328                rcu_accelerate_cbs(rsp, rnp, rdp);
1329
1330        } else {
1331
1332                /* Advance callbacks. */
1333                rcu_advance_cbs(rsp, rnp, rdp);
1334
1335                /* Remember that we saw this grace-period completion. */
1336                rdp->completed = rnp->completed;
1337                trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
1338
1339                /*
1340                 * If we were in an extended quiescent state, we may have
1341                 * missed some grace periods that others CPUs handled on
1342                 * our behalf. Catch up with this state to avoid noting
1343                 * spurious new grace periods.  If another grace period
1344                 * has started, then rnp->gpnum will have advanced, so
1345                 * we will detect this later on.  Of course, any quiescent
1346                 * states we found for the old GP are now invalid.
1347                 */
1348                if (ULONG_CMP_LT(rdp->gpnum, rdp->completed)) {
1349                        rdp->gpnum = rdp->completed;
1350                        rdp->passed_quiesce = 0;
1351                }
1352
1353                /*
1354                 * If RCU does not need a quiescent state from this CPU,
1355                 * then make sure that this CPU doesn't go looking for one.
1356                 */
1357                if ((rnp->qsmask & rdp->grpmask) == 0)
1358                        rdp->qs_pending = 0;
1359        }
1360}
1361
1362/*
1363 * Advance this CPU's callbacks, but only if the current grace period
1364 * has ended.  This may be called only from the CPU to whom the rdp
1365 * belongs.
1366 */
1367static void
1368rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
1369{
1370        unsigned long flags;
1371        struct rcu_node *rnp;
1372
1373        local_irq_save(flags);
1374        rnp = rdp->mynode;
1375        if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
1376            !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1377                local_irq_restore(flags);
1378                return;
1379        }
1380        __rcu_process_gp_end(rsp, rnp, rdp);
1381        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1382}
1383
1384/*
1385 * Do per-CPU grace-period initialization for running CPU.  The caller
1386 * must hold the lock of the leaf rcu_node structure corresponding to
1387 * this CPU.
1388 */
1389static void
1390rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1391{
1392        /* Prior grace period ended, so advance callbacks for current CPU. */
1393        __rcu_process_gp_end(rsp, rnp, rdp);
1394
1395        /* Set state so that this CPU will detect the next quiescent state. */
1396        __note_new_gpnum(rsp, rnp, rdp);
1397}
1398
1399/*
1400 * Initialize a new grace period.
1401 */
1402static int rcu_gp_init(struct rcu_state *rsp)
1403{
1404        struct rcu_data *rdp;
1405        struct rcu_node *rnp = rcu_get_root(rsp);
1406
1407        raw_spin_lock_irq(&rnp->lock);
1408        rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1409
1410        if (rcu_gp_in_progress(rsp)) {
1411                /* Grace period already in progress, don't start another.  */
1412                raw_spin_unlock_irq(&rnp->lock);
1413                return 0;
1414        }
1415
1416        /* Advance to a new grace period and initialize state. */
1417        rsp->gpnum++;
1418        trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1419        record_gp_stall_check_time(rsp);
1420        raw_spin_unlock_irq(&rnp->lock);
1421
1422        /* Exclude any concurrent CPU-hotplug operations. */
1423        mutex_lock(&rsp->onoff_mutex);
1424
1425        /*
1426         * Set the quiescent-state-needed bits in all the rcu_node
1427         * structures for all currently online CPUs in breadth-first order,
1428         * starting from the root rcu_node structure, relying on the layout
1429         * of the tree within the rsp->node[] array.  Note that other CPUs
1430         * will access only the leaves of the hierarchy, thus seeing that no
1431         * grace period is in progress, at least until the corresponding
1432         * leaf node has been initialized.  In addition, we have excluded
1433         * CPU-hotplug operations.
1434         *
1435         * The grace period cannot complete until the initialization
1436         * process finishes, because this kthread handles both.
1437         */
1438        rcu_for_each_node_breadth_first(rsp, rnp) {
1439                raw_spin_lock_irq(&rnp->lock);
1440                rdp = this_cpu_ptr(rsp->rda);
1441                rcu_preempt_check_blocked_tasks(rnp);
1442                rnp->qsmask = rnp->qsmaskinit;
1443                ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1444                WARN_ON_ONCE(rnp->completed != rsp->completed);
1445                ACCESS_ONCE(rnp->completed) = rsp->completed;
1446                if (rnp == rdp->mynode)
1447                        rcu_start_gp_per_cpu(rsp, rnp, rdp);
1448                rcu_preempt_boost_start_gp(rnp);
1449                trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1450                                            rnp->level, rnp->grplo,
1451                                            rnp->grphi, rnp->qsmask);
1452                raw_spin_unlock_irq(&rnp->lock);
1453#ifdef CONFIG_PROVE_RCU_DELAY
1454                if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 &&
1455                    system_state == SYSTEM_RUNNING)
1456                        udelay(200);
1457#endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1458                cond_resched();
1459        }
1460
1461        mutex_unlock(&rsp->onoff_mutex);
1462        return 1;
1463}
1464
1465/*
1466 * Do one round of quiescent-state forcing.
1467 */
1468int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1469{
1470        int fqs_state = fqs_state_in;
1471        struct rcu_node *rnp = rcu_get_root(rsp);
1472
1473        rsp->n_force_qs++;
1474        if (fqs_state == RCU_SAVE_DYNTICK) {
1475                /* Collect dyntick-idle snapshots. */
1476                force_qs_rnp(rsp, dyntick_save_progress_counter);
1477                fqs_state = RCU_FORCE_QS;
1478        } else {
1479                /* Handle dyntick-idle and offline CPUs. */
1480                force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1481        }
1482        /* Clear flag to prevent immediate re-entry. */
1483        if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1484                raw_spin_lock_irq(&rnp->lock);
1485                rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1486                raw_spin_unlock_irq(&rnp->lock);
1487        }
1488        return fqs_state;
1489}
1490
1491/*
1492 * Clean up after the old grace period.
1493 */
1494static void rcu_gp_cleanup(struct rcu_state *rsp)
1495{
1496        unsigned long gp_duration;
1497        int nocb = 0;
1498        struct rcu_data *rdp;
1499        struct rcu_node *rnp = rcu_get_root(rsp);
1500
1501        raw_spin_lock_irq(&rnp->lock);
1502        gp_duration = jiffies - rsp->gp_start;
1503        if (gp_duration > rsp->gp_max)
1504                rsp->gp_max = gp_duration;
1505
1506        /*
1507         * We know the grace period is complete, but to everyone else
1508         * it appears to still be ongoing.  But it is also the case
1509         * that to everyone else it looks like there is nothing that
1510         * they can do to advance the grace period.  It is therefore
1511         * safe for us to drop the lock in order to mark the grace
1512         * period as completed in all of the rcu_node structures.
1513         */
1514        raw_spin_unlock_irq(&rnp->lock);
1515
1516        /*
1517         * Propagate new ->completed value to rcu_node structures so
1518         * that other CPUs don't have to wait until the start of the next
1519         * grace period to process their callbacks.  This also avoids
1520         * some nasty RCU grace-period initialization races by forcing
1521         * the end of the current grace period to be completely recorded in
1522         * all of the rcu_node structures before the beginning of the next
1523         * grace period is recorded in any of the rcu_node structures.
1524         */
1525        rcu_for_each_node_breadth_first(rsp, rnp) {
1526                raw_spin_lock_irq(&rnp->lock);
1527                ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1528                rdp = this_cpu_ptr(rsp->rda);
1529                if (rnp == rdp->mynode)
1530                        __rcu_process_gp_end(rsp, rnp, rdp);
1531                nocb += rcu_future_gp_cleanup(rsp, rnp);
1532                raw_spin_unlock_irq(&rnp->lock);
1533                cond_resched();
1534        }
1535        rnp = rcu_get_root(rsp);
1536        raw_spin_lock_irq(&rnp->lock);
1537        rcu_nocb_gp_set(rnp, nocb);
1538
1539        rsp->completed = rsp->gpnum; /* Declare grace period done. */
1540        trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1541        rsp->fqs_state = RCU_GP_IDLE;
1542        rdp = this_cpu_ptr(rsp->rda);
1543        rcu_advance_cbs(rsp, rnp, rdp);  /* Reduce false positives below. */
1544        if (cpu_needs_another_gp(rsp, rdp))
1545                rsp->gp_flags = 1;
1546        raw_spin_unlock_irq(&rnp->lock);
1547}
1548
1549/*
1550 * Body of kthread that handles grace periods.
1551 */
1552static int __noreturn rcu_gp_kthread(void *arg)
1553{
1554        int fqs_state;
1555        unsigned long j;
1556        int ret;
1557        struct rcu_state *rsp = arg;
1558        struct rcu_node *rnp = rcu_get_root(rsp);
1559
1560        for (;;) {
1561
1562                /* Handle grace-period start. */
1563                for (;;) {
1564                        wait_event_interruptible(rsp->gp_wq,
1565                                                 rsp->gp_flags &
1566                                                 RCU_GP_FLAG_INIT);
1567                        if ((rsp->gp_flags & RCU_GP_FLAG_INIT) &&
1568                            rcu_gp_init(rsp))
1569                                break;
1570                        cond_resched();
1571                        flush_signals(current);
1572                }
1573
1574                /* Handle quiescent-state forcing. */
1575                fqs_state = RCU_SAVE_DYNTICK;
1576                j = jiffies_till_first_fqs;
1577                if (j > HZ) {
1578                        j = HZ;
1579                        jiffies_till_first_fqs = HZ;
1580                }
1581                for (;;) {
1582                        rsp->jiffies_force_qs = jiffies + j;
1583                        ret = wait_event_interruptible_timeout(rsp->gp_wq,
1584                                        (rsp->gp_flags & RCU_GP_FLAG_FQS) ||
1585                                        (!ACCESS_ONCE(rnp->qsmask) &&
1586                                         !rcu_preempt_blocked_readers_cgp(rnp)),
1587                                        j);
1588                        /* If grace period done, leave loop. */
1589                        if (!ACCESS_ONCE(rnp->qsmask) &&
1590                            !rcu_preempt_blocked_readers_cgp(rnp))
1591                                break;
1592                        /* If time for quiescent-state forcing, do it. */
1593                        if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) {
1594                                fqs_state = rcu_gp_fqs(rsp, fqs_state);
1595                                cond_resched();
1596                        } else {
1597                                /* Deal with stray signal. */
1598                                cond_resched();
1599                                flush_signals(current);
1600                        }
1601                        j = jiffies_till_next_fqs;
1602                        if (j > HZ) {
1603                                j = HZ;
1604                                jiffies_till_next_fqs = HZ;
1605                        } else if (j < 1) {
1606                                j = 1;
1607                                jiffies_till_next_fqs = 1;
1608                        }
1609                }
1610
1611                /* Handle grace-period end. */
1612                rcu_gp_cleanup(rsp);
1613        }
1614}
1615
1616static void rsp_wakeup(struct irq_work *work)
1617{
1618        struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work);
1619
1620        /* Wake up rcu_gp_kthread() to start the grace period. */
1621        wake_up(&rsp->gp_wq);
1622}
1623
1624/*
1625 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1626 * in preparation for detecting the next grace period.  The caller must hold
1627 * the root node's ->lock and hard irqs must be disabled.
1628 *
1629 * Note that it is legal for a dying CPU (which is marked as offline) to
1630 * invoke this function.  This can happen when the dying CPU reports its
1631 * quiescent state.
1632 */
1633static void
1634rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1635                      struct rcu_data *rdp)
1636{
1637        if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1638                /*
1639                 * Either we have not yet spawned the grace-period
1640                 * task, this CPU does not need another grace period,
1641                 * or a grace period is already in progress.
1642                 * Either way, don't start a new grace period.
1643                 */
1644                return;
1645        }
1646        rsp->gp_flags = RCU_GP_FLAG_INIT;
1647
1648        /*
1649         * We can't do wakeups while holding the rnp->lock, as that
1650         * could cause possible deadlocks with the rq->lock. Deter
1651         * the wakeup to interrupt context.
1652         */
1653        irq_work_queue(&rsp->wakeup_work);
1654}
1655
1656/*
1657 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1658 * callbacks.  Note that rcu_start_gp_advanced() cannot do this because it
1659 * is invoked indirectly from rcu_advance_cbs(), which would result in
1660 * endless recursion -- or would do so if it wasn't for the self-deadlock
1661 * that is encountered beforehand.
1662 */
1663static void
1664rcu_start_gp(struct rcu_state *rsp)
1665{
1666        struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1667        struct rcu_node *rnp = rcu_get_root(rsp);
1668
1669        /*
1670         * If there is no grace period in progress right now, any
1671         * callbacks we have up to this point will be satisfied by the
1672         * next grace period.  Also, advancing the callbacks reduces the
1673         * probability of false positives from cpu_needs_another_gp()
1674         * resulting in pointless grace periods.  So, advance callbacks
1675         * then start the grace period!
1676         */
1677        rcu_advance_cbs(rsp, rnp, rdp);
1678        rcu_start_gp_advanced(rsp, rnp, rdp);
1679}
1680
1681/*
1682 * Report a full set of quiescent states to the specified rcu_state
1683 * data structure.  This involves cleaning up after the prior grace
1684 * period and letting rcu_start_gp() start up the next grace period
1685 * if one is needed.  Note that the caller must hold rnp->lock, which
1686 * is released before return.
1687 */
1688static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1689        __releases(rcu_get_root(rsp)->lock)
1690{
1691        WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1692        raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1693        wake_up(&rsp->gp_wq);  /* Memory barrier implied by wake_up() path. */
1694}
1695
1696/*
1697 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1698 * Allows quiescent states for a group of CPUs to be reported at one go
1699 * to the specified rcu_node structure, though all the CPUs in the group
1700 * must be represented by the same rcu_node structure (which need not be
1701 * a leaf rcu_node structure, though it often will be).  That structure's
1702 * lock must be held upon entry, and it is released before return.
1703 */
1704static void
1705rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1706                  struct rcu_node *rnp, unsigned long flags)
1707        __releases(rnp->lock)
1708{
1709        struct rcu_node *rnp_c;
1710
1711        /* Walk up the rcu_node hierarchy. */
1712        for (;;) {
1713                if (!(rnp->qsmask & mask)) {
1714
1715                        /* Our bit has already been cleared, so done. */
1716                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1717                        return;
1718                }
1719                rnp->qsmask &= ~mask;
1720                trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1721                                                 mask, rnp->qsmask, rnp->level,
1722                                                 rnp->grplo, rnp->grphi,
1723                                                 !!rnp->gp_tasks);
1724                if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1725
1726                        /* Other bits still set at this level, so done. */
1727                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1728                        return;
1729                }
1730                mask = rnp->grpmask;
1731                if (rnp->parent == NULL) {
1732
1733                        /* No more levels.  Exit loop holding root lock. */
1734
1735                        break;
1736                }
1737                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1738                rnp_c = rnp;
1739                rnp = rnp->parent;
1740                raw_spin_lock_irqsave(&rnp->lock, flags);
1741                WARN_ON_ONCE(rnp_c->qsmask);
1742        }
1743
1744        /*
1745         * Get here if we are the last CPU to pass through a quiescent
1746         * state for this grace period.  Invoke rcu_report_qs_rsp()
1747         * to clean up and start the next grace period if one is needed.
1748         */
1749        rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1750}
1751
1752/*
1753 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1754 * structure.  This must be either called from the specified CPU, or
1755 * called when the specified CPU is known to be offline (and when it is
1756 * also known that no other CPU is concurrently trying to help the offline
1757 * CPU).  The lastcomp argument is used to make sure we are still in the
1758 * grace period of interest.  We don't want to end the current grace period
1759 * based on quiescent states detected in an earlier grace period!
1760 */
1761static void
1762rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1763{
1764        unsigned long flags;
1765        unsigned long mask;
1766        struct rcu_node *rnp;
1767
1768        rnp = rdp->mynode;
1769        raw_spin_lock_irqsave(&rnp->lock, flags);
1770        if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1771            rnp->completed == rnp->gpnum) {
1772
1773                /*
1774                 * The grace period in which this quiescent state was
1775                 * recorded has ended, so don't report it upwards.
1776                 * We will instead need a new quiescent state that lies
1777                 * within the current grace period.
1778                 */
1779                rdp->passed_quiesce = 0;        /* need qs for new gp. */
1780                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1781                return;
1782        }
1783        mask = rdp->grpmask;
1784        if ((rnp->qsmask & mask) == 0) {
1785                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1786        } else {
1787                rdp->qs_pending = 0;
1788
1789                /*
1790                 * This GP can't end until cpu checks in, so all of our
1791                 * callbacks can be processed during the next GP.
1792                 */
1793                rcu_accelerate_cbs(rsp, rnp, rdp);
1794
1795                rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1796        }
1797}
1798
1799/*
1800 * Check to see if there is a new grace period of which this CPU
1801 * is not yet aware, and if so, set up local rcu_data state for it.
1802 * Otherwise, see if this CPU has just passed through its first
1803 * quiescent state for this grace period, and record that fact if so.
1804 */
1805static void
1806rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1807{
1808        /* If there is now a new grace period, record and return. */
1809        if (check_for_new_grace_period(rsp, rdp))
1810                return;
1811
1812        /*
1813         * Does this CPU still need to do its part for current grace period?
1814         * If no, return and let the other CPUs do their part as well.
1815         */
1816        if (!rdp->qs_pending)
1817                return;
1818
1819        /*
1820         * Was there a quiescent state since the beginning of the grace
1821         * period? If no, then exit and wait for the next call.
1822         */
1823        if (!rdp->passed_quiesce)
1824                return;
1825
1826        /*
1827         * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1828         * judge of that).
1829         */
1830        rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1831}
1832
1833#ifdef CONFIG_HOTPLUG_CPU
1834
1835/*
1836 * Send the specified CPU's RCU callbacks to the orphanage.  The
1837 * specified CPU must be offline, and the caller must hold the
1838 * ->orphan_lock.
1839 */
1840static void
1841rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1842                          struct rcu_node *rnp, struct rcu_data *rdp)
1843{
1844        /* No-CBs CPUs do not have orphanable callbacks. */
1845        if (rcu_is_nocb_cpu(rdp->cpu))
1846                return;
1847
1848        /*
1849         * Orphan the callbacks.  First adjust the counts.  This is safe
1850         * because _rcu_barrier() excludes CPU-hotplug operations, so it
1851         * cannot be running now.  Thus no memory barrier is required.
1852         */
1853        if (rdp->nxtlist != NULL) {
1854                rsp->qlen_lazy += rdp->qlen_lazy;
1855                rsp->qlen += rdp->qlen;
1856                rdp->n_cbs_orphaned += rdp->qlen;
1857                rdp->qlen_lazy = 0;
1858                ACCESS_ONCE(rdp->qlen) = 0;
1859        }
1860
1861        /*
1862         * Next, move those callbacks still needing a grace period to
1863         * the orphanage, where some other CPU will pick them up.
1864         * Some of the callbacks might have gone partway through a grace
1865         * period, but that is too bad.  They get to start over because we
1866         * cannot assume that grace periods are synchronized across CPUs.
1867         * We don't bother updating the ->nxttail[] array yet, instead
1868         * we just reset the whole thing later on.
1869         */
1870        if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1871                *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1872                rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1873                *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1874        }
1875
1876        /*
1877         * Then move the ready-to-invoke callbacks to the orphanage,
1878         * where some other CPU will pick them up.  These will not be
1879         * required to pass though another grace period: They are done.
1880         */
1881        if (rdp->nxtlist != NULL) {
1882                *rsp->orphan_donetail = rdp->nxtlist;
1883                rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1884        }
1885
1886        /* Finally, initialize the rcu_data structure's list to empty.  */
1887        init_callback_list(rdp);
1888}
1889
1890/*
1891 * Adopt the RCU callbacks from the specified rcu_state structure's
1892 * orphanage.  The caller must hold the ->orphan_lock.
1893 */
1894static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1895{
1896        int i;
1897        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1898
1899        /* No-CBs CPUs are handled specially. */
1900        if (rcu_nocb_adopt_orphan_cbs(rsp, rdp))
1901                return;
1902
1903        /* Do the accounting first. */
1904        rdp->qlen_lazy += rsp->qlen_lazy;
1905        rdp->qlen += rsp->qlen;
1906        rdp->n_cbs_adopted += rsp->qlen;
1907        if (rsp->qlen_lazy != rsp->qlen)
1908                rcu_idle_count_callbacks_posted();
1909        rsp->qlen_lazy = 0;
1910        rsp->qlen = 0;
1911
1912        /*
1913         * We do not need a memory barrier here because the only way we
1914         * can get here if there is an rcu_barrier() in flight is if
1915         * we are the task doing the rcu_barrier().
1916         */
1917
1918        /* First adopt the ready-to-invoke callbacks. */
1919        if (rsp->orphan_donelist != NULL) {
1920                *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1921                *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1922                for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1923                        if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1924                                rdp->nxttail[i] = rsp->orphan_donetail;
1925                rsp->orphan_donelist = NULL;
1926                rsp->orphan_donetail = &rsp->orphan_donelist;
1927        }
1928
1929        /* And then adopt the callbacks that still need a grace period. */
1930        if (rsp->orphan_nxtlist != NULL) {
1931                *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1932                rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1933                rsp->orphan_nxtlist = NULL;
1934                rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1935        }
1936}
1937
1938/*
1939 * Trace the fact that this CPU is going offline.
1940 */
1941static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1942{
1943        RCU_TRACE(unsigned long mask);
1944        RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1945        RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1946
1947        RCU_TRACE(mask = rdp->grpmask);
1948        trace_rcu_grace_period(rsp->name,
1949                               rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1950                               "cpuofl");
1951}
1952
1953/*
1954 * The CPU has been completely removed, and some other CPU is reporting
1955 * this fact from process context.  Do the remainder of the cleanup,
1956 * including orphaning the outgoing CPU's RCU callbacks, and also
1957 * adopting them.  There can only be one CPU hotplug operation at a time,
1958 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1959 */
1960static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1961{
1962        unsigned long flags;
1963        unsigned long mask;
1964        int need_report = 0;
1965        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1966        struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
1967
1968        /* Adjust any no-longer-needed kthreads. */
1969        rcu_boost_kthread_setaffinity(rnp, -1);
1970
1971        /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1972
1973        /* Exclude any attempts to start a new grace period. */
1974        mutex_lock(&rsp->onoff_mutex);
1975        raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
1976
1977        /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1978        rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1979        rcu_adopt_orphan_cbs(rsp);
1980
1981        /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1982        mask = rdp->grpmask;    /* rnp->grplo is constant. */
1983        do {
1984                raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
1985                rnp->qsmaskinit &= ~mask;
1986                if (rnp->qsmaskinit != 0) {
1987                        if (rnp != rdp->mynode)
1988                                raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1989                        break;
1990                }
1991                if (rnp == rdp->mynode)
1992                        need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1993                else
1994                        raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1995                mask = rnp->grpmask;
1996                rnp = rnp->parent;
1997        } while (rnp != NULL);
1998
1999        /*
2000         * We still hold the leaf rcu_node structure lock here, and
2001         * irqs are still disabled.  The reason for this subterfuge is
2002         * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
2003         * held leads to deadlock.
2004         */
2005        raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
2006        rnp = rdp->mynode;
2007        if (need_report & RCU_OFL_TASKS_NORM_GP)
2008                rcu_report_unblock_qs_rnp(rnp, flags);
2009        else
2010                raw_spin_unlock_irqrestore(&rnp->lock, flags);
2011        if (need_report & RCU_OFL_TASKS_EXP_GP)
2012                rcu_report_exp_rnp(rsp, rnp, true);
2013        WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2014                  "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2015                  cpu, rdp->qlen, rdp->nxtlist);
2016        init_callback_list(rdp);
2017        /* Disallow further callbacks on this CPU. */
2018        rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2019        mutex_unlock(&rsp->onoff_mutex);
2020}
2021
2022#else /* #ifdef CONFIG_HOTPLUG_CPU */
2023
2024static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2025{
2026}
2027
2028static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2029{
2030}
2031
2032#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2033
2034/*
2035 * Invoke any RCU callbacks that have made it to the end of their grace
2036 * period.  Thottle as specified by rdp->blimit.
2037 */
2038static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2039{
2040        unsigned long flags;
2041        struct rcu_head *next, *list, **tail;
2042        long bl, count, count_lazy;
2043        int i;
2044
2045        /* If no callbacks are ready, just return. */
2046        if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2047                trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2048                trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
2049                                    need_resched(), is_idle_task(current),
2050                                    rcu_is_callbacks_kthread());
2051                return;
2052        }
2053
2054        /*
2055         * Extract the list of ready callbacks, disabling to prevent
2056         * races with call_rcu() from interrupt handlers.
2057         */
2058        local_irq_save(flags);
2059        WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2060        bl = rdp->blimit;
2061        trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2062        list = rdp->nxtlist;
2063        rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2064        *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2065        tail = rdp->nxttail[RCU_DONE_TAIL];
2066        for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2067                if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2068                        rdp->nxttail[i] = &rdp->nxtlist;
2069        local_irq_restore(flags);
2070
2071        /* Invoke callbacks. */
2072        count = count_lazy = 0;
2073        while (list) {
2074                next = list->next;
2075                prefetch(next);
2076                debug_rcu_head_unqueue(list);
2077                if (__rcu_reclaim(rsp->name, list))
2078                        count_lazy++;
2079                list = next;
2080                /* Stop only if limit reached and CPU has something to do. */
2081                if (++count >= bl &&
2082                    (need_resched() ||
2083                     (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2084                        break;
2085        }
2086
2087        local_irq_save(flags);
2088        trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2089                            is_idle_task(current),
2090                            rcu_is_callbacks_kthread());
2091
2092        /* Update count, and requeue any remaining callbacks. */
2093        if (list != NULL) {
2094                *tail = rdp->nxtlist;
2095                rdp->nxtlist = list;
2096                for (i = 0; i < RCU_NEXT_SIZE; i++)
2097                        if (&rdp->nxtlist == rdp->nxttail[i])
2098                                rdp->nxttail[i] = tail;
2099                        else
2100                                break;
2101        }
2102        smp_mb(); /* List handling before counting for rcu_barrier(). */
2103        rdp->qlen_lazy -= count_lazy;
2104        ACCESS_ONCE(rdp->qlen) -= count;
2105        rdp->n_cbs_invoked += count;
2106
2107        /* Reinstate batch limit if we have worked down the excess. */
2108        if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2109                rdp->blimit = blimit;
2110
2111        /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2112        if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2113                rdp->qlen_last_fqs_check = 0;
2114                rdp->n_force_qs_snap = rsp->n_force_qs;
2115        } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2116                rdp->qlen_last_fqs_check = rdp->qlen;
2117        WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2118
2119        local_irq_restore(flags);
2120
2121        /* Re-invoke RCU core processing if there are callbacks remaining. */
2122        if (cpu_has_callbacks_ready_to_invoke(rdp))
2123                invoke_rcu_core();
2124}
2125
2126/*
2127 * Check to see if this CPU is in a non-context-switch quiescent state
2128 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2129 * Also schedule RCU core processing.
2130 *
2131 * This function must be called from hardirq context.  It is normally
2132 * invoked from the scheduling-clock interrupt.  If rcu_pending returns
2133 * false, there is no point in invoking rcu_check_callbacks().
2134 */
2135void rcu_check_callbacks(int cpu, int user)
2136{
2137        trace_rcu_utilization("Start scheduler-tick");
2138        increment_cpu_stall_ticks();
2139        if (user || rcu_is_cpu_rrupt_from_idle()) {
2140
2141                /*
2142                 * Get here if this CPU took its interrupt from user
2143                 * mode or from the idle loop, and if this is not a
2144                 * nested interrupt.  In this case, the CPU is in
2145                 * a quiescent state, so note it.
2146                 *
2147                 * No memory barrier is required here because both
2148                 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2149                 * variables that other CPUs neither access nor modify,
2150                 * at least not while the corresponding CPU is online.
2151                 */
2152
2153                rcu_sched_qs(cpu);
2154                rcu_bh_qs(cpu);
2155
2156        } else if (!in_softirq()) {
2157
2158                /*
2159                 * Get here if this CPU did not take its interrupt from
2160                 * softirq, in other words, if it is not interrupting
2161                 * a rcu_bh read-side critical section.  This is an _bh
2162                 * critical section, so note it.
2163                 */
2164
2165                rcu_bh_qs(cpu);
2166        }
2167        rcu_preempt_check_callbacks(cpu);
2168        if (rcu_pending(cpu))
2169                invoke_rcu_core();
2170        trace_rcu_utilization("End scheduler-tick");
2171}
2172
2173/*
2174 * Scan the leaf rcu_node structures, processing dyntick state for any that
2175 * have not yet encountered a quiescent state, using the function specified.
2176 * Also initiate boosting for any threads blocked on the root rcu_node.
2177 *
2178 * The caller must have suppressed start of new grace periods.
2179 */
2180static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
2181{
2182        unsigned long bit;
2183        int cpu;
2184        unsigned long flags;
2185        unsigned long mask;
2186        struct rcu_node *rnp;
2187
2188        rcu_for_each_leaf_node(rsp, rnp) {
2189                cond_resched();
2190                mask = 0;
2191                raw_spin_lock_irqsave(&rnp->lock, flags);
2192                if (!rcu_gp_in_progress(rsp)) {
2193                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
2194                        return;
2195                }
2196                if (rnp->qsmask == 0) {
2197                        rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2198                        continue;
2199                }
2200                cpu = rnp->grplo;
2201                bit = 1;
2202                for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2203                        if ((rnp->qsmask & bit) != 0 &&
2204                            f(per_cpu_ptr(rsp->rda, cpu)))
2205                                mask |= bit;
2206                }
2207                if (mask != 0) {
2208
2209                        /* rcu_report_qs_rnp() releases rnp->lock. */
2210                        rcu_report_qs_rnp(mask, rsp, rnp, flags);
2211                        continue;
2212                }
2213                raw_spin_unlock_irqrestore(&rnp->lock, flags);
2214        }
2215        rnp = rcu_get_root(rsp);
2216        if (rnp->qsmask == 0) {
2217                raw_spin_lock_irqsave(&rnp->lock, flags);
2218                rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2219        }
2220}
2221
2222/*
2223 * Force quiescent states on reluctant CPUs, and also detect which
2224 * CPUs are in dyntick-idle mode.
2225 */
2226static void force_quiescent_state(struct rcu_state *rsp)
2227{
2228        unsigned long flags;
2229        bool ret;
2230        struct rcu_node *rnp;
2231        struct rcu_node *rnp_old = NULL;
2232
2233        /* Funnel through hierarchy to reduce memory contention. */
2234        rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2235        for (; rnp != NULL; rnp = rnp->parent) {
2236                ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2237                      !raw_spin_trylock(&rnp->fqslock);
2238                if (rnp_old != NULL)
2239                        raw_spin_unlock(&rnp_old->fqslock);
2240                if (ret) {
2241                        rsp->n_force_qs_lh++;
2242                        return;
2243                }
2244                rnp_old = rnp;
2245        }
2246        /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2247
2248        /* Reached the root of the rcu_node tree, acquire lock. */
2249        raw_spin_lock_irqsave(&rnp_old->lock, flags);
2250        raw_spin_unlock(&rnp_old->fqslock);
2251        if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2252                rsp->n_force_qs_lh++;
2253                raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2254                return;  /* Someone beat us to it. */
2255        }
2256        rsp->gp_flags |= RCU_GP_FLAG_FQS;
2257        raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2258        wake_up(&rsp->gp_wq);  /* Memory barrier implied by wake_up() path. */
2259}
2260
2261/*
2262 * This does the RCU core processing work for the specified rcu_state
2263 * and rcu_data structures.  This may be called only from the CPU to
2264 * whom the rdp belongs.
2265 */
2266static void
2267__rcu_process_callbacks(struct rcu_state *rsp)
2268{
2269        unsigned long flags;
2270        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2271
2272        WARN_ON_ONCE(rdp->beenonline == 0);
2273
2274        /* Handle the end of a grace period that some other CPU ended.  */
2275        rcu_process_gp_end(rsp, rdp);
2276
2277        /* Update RCU state based on any recent quiescent states. */
2278        rcu_check_quiescent_state(rsp, rdp);
2279
2280        /* Does this CPU require a not-yet-started grace period? */
2281        local_irq_save(flags);
2282        if (cpu_needs_another_gp(rsp, rdp)) {
2283                raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2284                rcu_start_gp(rsp);
2285                raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2286        } else {
2287                local_irq_restore(flags);
2288        }
2289
2290        /* If there are callbacks ready, invoke them. */
2291        if (cpu_has_callbacks_ready_to_invoke(rdp))
2292                invoke_rcu_callbacks(rsp, rdp);
2293}
2294
2295/*
2296 * Do RCU core processing for the current CPU.
2297 */
2298static void rcu_process_callbacks(struct softirq_action *unused)
2299{
2300        struct rcu_state *rsp;
2301
2302        if (cpu_is_offline(smp_processor_id()))
2303                return;
2304        trace_rcu_utilization("Start RCU core");
2305        for_each_rcu_flavor(rsp)
2306                __rcu_process_callbacks(rsp);
2307        trace_rcu_utilization("End RCU core");
2308}
2309
2310/*
2311 * Schedule RCU callback invocation.  If the specified type of RCU
2312 * does not support RCU priority boosting, just do a direct call,
2313 * otherwise wake up the per-CPU kernel kthread.  Note that because we
2314 * are running on the current CPU with interrupts disabled, the
2315 * rcu_cpu_kthread_task cannot disappear out from under us.
2316 */
2317static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2318{
2319        if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2320                return;
2321        if (likely(!rsp->boost)) {
2322                rcu_do_batch(rsp, rdp);
2323                return;
2324        }
2325        invoke_rcu_callbacks_kthread();
2326}
2327
2328static void invoke_rcu_core(void)
2329{
2330        if (cpu_online(smp_processor_id()))
2331                raise_softirq(RCU_SOFTIRQ);
2332}
2333
2334/*
2335 * Handle any core-RCU processing required by a call_rcu() invocation.
2336 */
2337static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2338                            struct rcu_head *head, unsigned long flags)
2339{
2340        /*
2341         * If called from an extended quiescent state, invoke the RCU
2342         * core in order to force a re-evaluation of RCU's idleness.
2343         */
2344        if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2345                invoke_rcu_core();
2346
2347        /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2348        if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2349                return;
2350
2351        /*
2352         * Force the grace period if too many callbacks or too long waiting.
2353         * Enforce hysteresis, and don't invoke force_quiescent_state()
2354         * if some other CPU has recently done so.  Also, don't bother
2355         * invoking force_quiescent_state() if the newly enqueued callback
2356         * is the only one waiting for a grace period to complete.
2357         */
2358        if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2359
2360                /* Are we ignoring a completed grace period? */
2361                rcu_process_gp_end(rsp, rdp);
2362                check_for_new_grace_period(rsp, rdp);
2363
2364                /* Start a new grace period if one not already started. */
2365                if (!rcu_gp_in_progress(rsp)) {
2366                        struct rcu_node *rnp_root = rcu_get_root(rsp);
2367
2368                        raw_spin_lock(&rnp_root->lock);
2369                        rcu_start_gp(rsp);
2370                        raw_spin_unlock(&rnp_root->lock);
2371                } else {
2372                        /* Give the grace period a kick. */
2373                        rdp->blimit = LONG_MAX;
2374                        if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2375                            *rdp->nxttail[RCU_DONE_TAIL] != head)
2376                                force_quiescent_state(rsp);
2377                        rdp->n_force_qs_snap = rsp->n_force_qs;
2378                        rdp->qlen_last_fqs_check = rdp->qlen;
2379                }
2380        }
2381}
2382
2383/*
2384 * Helper function for call_rcu() and friends.  The cpu argument will
2385 * normally be -1, indicating "currently running CPU".  It may specify
2386 * a CPU only if that CPU is a no-CBs CPU.  Currently, only _rcu_barrier()
2387 * is expected to specify a CPU.
2388 */
2389static void
2390__call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2391           struct rcu_state *rsp, int cpu, bool lazy)
2392{
2393        unsigned long flags;
2394        struct rcu_data *rdp;
2395
2396        WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2397        debug_rcu_head_queue(head);
2398        head->func = func;
2399        head->next = NULL;
2400
2401        /*
2402         * Opportunistically note grace-period endings and beginnings.
2403         * Note that we might see a beginning right after we see an
2404         * end, but never vice versa, since this CPU has to pass through
2405         * a quiescent state betweentimes.
2406         */
2407        local_irq_save(flags);
2408        rdp = this_cpu_ptr(rsp->rda);
2409
2410        /* Add the callback to our list. */
2411        if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2412                int offline;
2413
2414                if (cpu != -1)
2415                        rdp = per_cpu_ptr(rsp->rda, cpu);
2416                offline = !__call_rcu_nocb(rdp, head, lazy);
2417                WARN_ON_ONCE(offline);
2418                /* _call_rcu() is illegal on offline CPU; leak the callback. */
2419                local_irq_restore(flags);
2420                return;
2421        }
2422        ACCESS_ONCE(rdp->qlen)++;
2423        if (lazy)
2424                rdp->qlen_lazy++;
2425        else
2426                rcu_idle_count_callbacks_posted();
2427        smp_mb();  /* Count before adding callback for rcu_barrier(). */
2428        *rdp->nxttail[RCU_NEXT_TAIL] = head;
2429        rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2430
2431        if (__is_kfree_rcu_offset((unsigned long)func))
2432                trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2433                                         rdp->qlen_lazy, rdp->qlen);
2434        else
2435                trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2436
2437        /* Go handle any RCU core processing required. */
2438        __call_rcu_core(rsp, rdp, head, flags);
2439        local_irq_restore(flags);
2440}
2441
2442/*
2443 * Queue an RCU-sched callback for invocation after a grace period.
2444 */
2445void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2446{
2447        __call_rcu(head, func, &rcu_sched_state, -1, 0);
2448}
2449EXPORT_SYMBOL_GPL(call_rcu_sched);
2450
2451/*
2452 * Queue an RCU callback for invocation after a quicker grace period.
2453 */
2454void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2455{
2456        __call_rcu(head, func, &rcu_bh_state, -1, 0);
2457}
2458EXPORT_SYMBOL_GPL(call_rcu_bh);
2459
2460/*
2461 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2462 * any blocking grace-period wait automatically implies a grace period
2463 * if there is only one CPU online at any point time during execution
2464 * of either synchronize_sched() or synchronize_rcu_bh().  It is OK to
2465 * occasionally incorrectly indicate that there are multiple CPUs online
2466 * when there was in fact only one the whole time, as this just adds
2467 * some overhead: RCU still operates correctly.
2468 */
2469static inline int rcu_blocking_is_gp(void)
2470{
2471        int ret;
2472
2473        might_sleep();  /* Check for RCU read-side critical section. */
2474        preempt_disable();
2475        ret = num_online_cpus() <= 1;
2476        preempt_enable();
2477        return ret;
2478}
2479
2480/**
2481 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2482 *
2483 * Control will return to the caller some time after a full rcu-sched
2484 * grace period has elapsed, in other words after all currently executing
2485 * rcu-sched read-side critical sections have completed.   These read-side
2486 * critical sections are delimited by rcu_read_lock_sched() and
2487 * rcu_read_unlock_sched(), and may be nested.  Note that preempt_disable(),
2488 * local_irq_disable(), and so on may be used in place of
2489 * rcu_read_lock_sched().
2490 *
2491 * This means that all preempt_disable code sequences, including NMI and
2492 * non-threaded hardware-interrupt handlers, in progress on entry will
2493 * have completed before this primitive returns.  However, this does not
2494 * guarantee that softirq handlers will have completed, since in some
2495 * kernels, these handlers can run in process context, and can block.
2496 *
2497 * Note that this guarantee implies further memory-ordering guarantees.
2498 * On systems with more than one CPU, when synchronize_sched() returns,
2499 * each CPU is guaranteed to have executed a full memory barrier since the
2500 * end of its last RCU-sched read-side critical section whose beginning
2501 * preceded the call to synchronize_sched().  In addition, each CPU having
2502 * an RCU read-side critical section that extends beyond the return from
2503 * synchronize_sched() is guaranteed to have executed a full memory barrier
2504 * after the beginning of synchronize_sched() and before the beginning of
2505 * that RCU read-side critical section.  Note that these guarantees include
2506 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2507 * that are executing in the kernel.
2508 *
2509 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2510 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2511 * to have executed a full memory barrier during the execution of
2512 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2513 * again only if the system has more than one CPU).
2514 *
2515 * This primitive provides the guarantees made by the (now removed)
2516 * synchronize_kernel() API.  In contrast, synchronize_rcu() only
2517 * guarantees that rcu_read_lock() sections will have completed.
2518 * In "classic RCU", these two guarantees happen to be one and
2519 * the same, but can differ in realtime RCU implementations.
2520 */
2521void synchronize_sched(void)
2522{
2523        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2524                           !lock_is_held(&rcu_lock_map) &&
2525                           !lock_is_held(&rcu_sched_lock_map),
2526                           "Illegal synchronize_sched() in RCU-sched read-side critical section");
2527        if (rcu_blocking_is_gp())
2528                return;
2529        if (rcu_expedited)
2530                synchronize_sched_expedited();
2531        else
2532                wait_rcu_gp(call_rcu_sched);
2533}
2534EXPORT_SYMBOL_GPL(synchronize_sched);
2535
2536/**
2537 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2538 *
2539 * Control will return to the caller some time after a full rcu_bh grace
2540 * period has elapsed, in other words after all currently executing rcu_bh
2541 * read-side critical sections have completed.  RCU read-side critical
2542 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2543 * and may be nested.
2544 *
2545 * See the description of synchronize_sched() for more detailed information
2546 * on memory ordering guarantees.
2547 */
2548void synchronize_rcu_bh(void)
2549{
2550        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2551                           !lock_is_held(&rcu_lock_map) &&
2552                           !lock_is_held(&rcu_sched_lock_map),
2553                           "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2554        if (rcu_blocking_is_gp())
2555                return;
2556        if (rcu_expedited)
2557                synchronize_rcu_bh_expedited();
2558        else
2559                wait_rcu_gp(call_rcu_bh);
2560}
2561EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2562
2563static int synchronize_sched_expedited_cpu_stop(void *data)
2564{
2565        /*
2566         * There must be a full memory barrier on each affected CPU
2567         * between the time that try_stop_cpus() is called and the
2568         * time that it returns.
2569         *
2570         * In the current initial implementation of cpu_stop, the
2571         * above condition is already met when the control reaches
2572         * this point and the following smp_mb() is not strictly
2573         * necessary.  Do smp_mb() anyway for documentation and
2574         * robustness against future implementation changes.
2575         */
2576        smp_mb(); /* See above comment block. */
2577        return 0;
2578}
2579
2580/**
2581 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2582 *
2583 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2584 * approach to force the grace period to end quickly.  This consumes
2585 * significant time on all CPUs and is unfriendly to real-time workloads,
2586 * so is thus not recommended for any sort of common-case code.  In fact,
2587 * if you are using synchronize_sched_expedited() in a loop, please
2588 * restructure your code to batch your updates, and then use a single
2589 * synchronize_sched() instead.
2590 *
2591 * Note that it is illegal to call this function while holding any lock
2592 * that is acquired by a CPU-hotplug notifier.  And yes, it is also illegal
2593 * to call this function from a CPU-hotplug notifier.  Failing to observe
2594 * these restriction will result in deadlock.
2595 *
2596 * This implementation can be thought of as an application of ticket
2597 * locking to RCU, with sync_sched_expedited_started and
2598 * sync_sched_expedited_done taking on the roles of the halves
2599 * of the ticket-lock word.  Each task atomically increments
2600 * sync_sched_expedited_started upon entry, snapshotting the old value,
2601 * then attempts to stop all the CPUs.  If this succeeds, then each
2602 * CPU will have executed a context switch, resulting in an RCU-sched
2603 * grace period.  We are then done, so we use atomic_cmpxchg() to
2604 * update sync_sched_expedited_done to match our snapshot -- but
2605 * only if someone else has not already advanced past our snapshot.
2606 *
2607 * On the other hand, if try_stop_cpus() fails, we check the value
2608 * of sync_sched_expedited_done.  If it has advanced past our
2609 * initial snapshot, then someone else must have forced a grace period
2610 * some time after we took our snapshot.  In this case, our work is
2611 * done for us, and we can simply return.  Otherwise, we try again,
2612 * but keep our initial snapshot for purposes of checking for someone
2613 * doing our work for us.
2614 *
2615 * If we fail too many times in a row, we fall back to synchronize_sched().
2616 */
2617void synchronize_sched_expedited(void)
2618{
2619        long firstsnap, s, snap;
2620        int trycount = 0;
2621        struct rcu_state *rsp = &rcu_sched_state;
2622
2623        /*
2624         * If we are in danger of counter wrap, just do synchronize_sched().
2625         * By allowing sync_sched_expedited_started to advance no more than
2626         * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2627         * that more than 3.5 billion CPUs would be required to force a
2628         * counter wrap on a 32-bit system.  Quite a few more CPUs would of
2629         * course be required on a 64-bit system.
2630         */
2631        if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2632                         (ulong)atomic_long_read(&rsp->expedited_done) +
2633                         ULONG_MAX / 8)) {
2634                synchronize_sched();
2635                atomic_long_inc(&rsp->expedited_wrap);
2636                return;
2637        }
2638
2639        /*
2640         * Take a ticket.  Note that atomic_inc_return() implies a
2641         * full memory barrier.
2642         */
2643        snap = atomic_long_inc_return(&rsp->expedited_start);
2644        firstsnap = snap;
2645        get_online_cpus();
2646        WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2647
2648        /*
2649         * Each pass through the following loop attempts to force a
2650         * context switch on each CPU.
2651         */
2652        while (try_stop_cpus(cpu_online_mask,
2653                             synchronize_sched_expedited_cpu_stop,
2654                             NULL) == -EAGAIN) {
2655                put_online_cpus();
2656                atomic_long_inc(&rsp->expedited_tryfail);
2657
2658                /* Check to see if someone else did our work for us. */
2659                s = atomic_long_read(&rsp->expedited_done);
2660                if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2661                        /* ensure test happens before caller kfree */
2662                        smp_mb__before_atomic_inc(); /* ^^^ */
2663                        atomic_long_inc(&rsp->expedited_workdone1);
2664                        return;
2665                }
2666
2667                /* No joy, try again later.  Or just synchronize_sched(). */
2668                if (trycount++ < 10) {
2669                        udelay(trycount * num_online_cpus());
2670                } else {
2671                        wait_rcu_gp(call_rcu_sched);
2672                        atomic_long_inc(&rsp->expedited_normal);
2673                        return;
2674                }
2675
2676                /* Recheck to see if someone else did our work for us. */
2677                s = atomic_long_read(&rsp->expedited_done);
2678                if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2679                        /* ensure test happens before caller kfree */
2680                        smp_mb__before_atomic_inc(); /* ^^^ */
2681                        atomic_long_inc(&rsp->expedited_workdone2);
2682                        return;
2683                }
2684
2685                /*
2686                 * Refetching sync_sched_expedited_started allows later
2687                 * callers to piggyback on our grace period.  We retry
2688                 * after they started, so our grace period works for them,
2689                 * and they started after our first try, so their grace
2690                 * period works for us.
2691                 */
2692                get_online_cpus();
2693                snap = atomic_long_read(&rsp->expedited_start);
2694                smp_mb(); /* ensure read is before try_stop_cpus(). */
2695        }
2696        atomic_long_inc(&rsp->expedited_stoppedcpus);
2697
2698        /*
2699         * Everyone up to our most recent fetch is covered by our grace
2700         * period.  Update the counter, but only if our work is still
2701         * relevant -- which it won't be if someone who started later
2702         * than we did already did their update.
2703         */
2704        do {
2705                atomic_long_inc(&rsp->expedited_done_tries);
2706                s = atomic_long_read(&rsp->expedited_done);
2707                if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2708                        /* ensure test happens before caller kfree */
2709                        smp_mb__before_atomic_inc(); /* ^^^ */
2710                        atomic_long_inc(&rsp->expedited_done_lost);
2711                        break;
2712                }
2713        } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2714        atomic_long_inc(&rsp->expedited_done_exit);
2715
2716        put_online_cpus();
2717}
2718EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2719
2720/*
2721 * Check to see if there is any immediate RCU-related work to be done
2722 * by the current CPU, for the specified type of RCU, returning 1 if so.
2723 * The checks are in order of increasing expense: checks that can be
2724 * carried out against CPU-local state are performed first.  However,
2725 * we must check for CPU stalls first, else we might not get a chance.
2726 */
2727static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2728{
2729        struct rcu_node *rnp = rdp->mynode;
2730
2731        rdp->n_rcu_pending++;
2732
2733        /* Check for CPU stalls, if enabled. */
2734        check_cpu_stall(rsp, rdp);
2735
2736        /* Is the RCU core waiting for a quiescent state from this CPU? */
2737        if (rcu_scheduler_fully_active &&
2738            rdp->qs_pending && !rdp->passed_quiesce) {
2739                rdp->n_rp_qs_pending++;
2740        } else if (rdp->qs_pending && rdp->passed_quiesce) {
2741                rdp->n_rp_report_qs++;
2742                return 1;
2743        }
2744
2745        /* Does this CPU have callbacks ready to invoke? */
2746        if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2747                rdp->n_rp_cb_ready++;
2748                return 1;
2749        }
2750
2751        /* Has RCU gone idle with this CPU needing another grace period? */
2752        if (cpu_needs_another_gp(rsp, rdp)) {
2753                rdp->n_rp_cpu_needs_gp++;
2754                return 1;
2755        }
2756
2757        /* Has another RCU grace period completed?  */
2758        if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2759                rdp->n_rp_gp_completed++;
2760                return 1;
2761        }
2762
2763        /* Has a new RCU grace period started? */
2764        if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2765                rdp->n_rp_gp_started++;
2766                return 1;
2767        }
2768
2769        /* nothing to do */
2770        rdp->n_rp_need_nothing++;
2771        return 0;
2772}
2773
2774/*
2775 * Check to see if there is any immediate RCU-related work to be done
2776 * by the current CPU, returning 1 if so.  This function is part of the
2777 * RCU implementation; it is -not- an exported member of the RCU API.
2778 */
2779static int rcu_pending(int cpu)
2780{
2781        struct rcu_state *rsp;
2782
2783        for_each_rcu_flavor(rsp)
2784                if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2785                        return 1;
2786        return 0;
2787}
2788
2789/*
2790 * Return true if the specified CPU has any callback.  If all_lazy is
2791 * non-NULL, store an indication of whether all callbacks are lazy.
2792 * (If there are no callbacks, all of them are deemed to be lazy.)
2793 */
2794static int rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
2795{
2796        bool al = true;
2797        bool hc = false;
2798        struct rcu_data *rdp;
2799        struct rcu_state *rsp;
2800
2801        for_each_rcu_flavor(rsp) {
2802                rdp = per_cpu_ptr(rsp->rda, cpu);
2803                if (rdp->qlen != rdp->qlen_lazy)
2804                        al = false;
2805                if (rdp->nxtlist)
2806                        hc = true;
2807        }
2808        if (all_lazy)
2809                *all_lazy = al;
2810        return hc;
2811}
2812
2813/*
2814 * Helper function for _rcu_barrier() tracing.  If tracing is disabled,
2815 * the compiler is expected to optimize this away.
2816 */
2817static void _rcu_barrier_trace(struct rcu_state *rsp, char *s,
2818                               int cpu, unsigned long done)
2819{
2820        trace_rcu_barrier(rsp->name, s, cpu,
2821                          atomic_read(&rsp->barrier_cpu_count), done);
2822}
2823
2824/*
2825 * RCU callback function for _rcu_barrier().  If we are last, wake
2826 * up the task executing _rcu_barrier().
2827 */
2828static void rcu_barrier_callback(struct rcu_head *rhp)
2829{
2830        struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2831        struct rcu_state *rsp = rdp->rsp;
2832
2833        if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2834                _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2835                complete(&rsp->barrier_completion);
2836        } else {
2837                _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2838        }
2839}
2840
2841/*
2842 * Called with preemption disabled, and from cross-cpu IRQ context.
2843 */
2844static void rcu_barrier_func(void *type)
2845{
2846        struct rcu_state *rsp = type;
2847        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2848
2849        _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2850        atomic_inc(&rsp->barrier_cpu_count);
2851        rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2852}
2853
2854/*
2855 * Orchestrate the specified type of RCU barrier, waiting for all
2856 * RCU callbacks of the specified type to complete.
2857 */
2858static void _rcu_barrier(struct rcu_state *rsp)
2859{
2860        int cpu;
2861        struct rcu_data *rdp;
2862        unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2863        unsigned long snap_done;
2864
2865        _rcu_barrier_trace(rsp, "Begin", -1, snap);
2866
2867        /* Take mutex to serialize concurrent rcu_barrier() requests. */
2868        mutex_lock(&rsp->barrier_mutex);
2869
2870        /*
2871         * Ensure that all prior references, including to ->n_barrier_done,
2872         * are ordered before the _rcu_barrier() machinery.
2873         */
2874        smp_mb();  /* See above block comment. */
2875
2876        /*
2877         * Recheck ->n_barrier_done to see if others did our work for us.
2878         * This means checking ->n_barrier_done for an even-to-odd-to-even
2879         * transition.  The "if" expression below therefore rounds the old
2880         * value up to the next even number and adds two before comparing.
2881         */
2882        snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2883        _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2884        if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2885                _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2886                smp_mb(); /* caller's subsequent code after above check. */
2887                mutex_unlock(&rsp->barrier_mutex);
2888                return;
2889        }
2890
2891        /*
2892         * Increment ->n_barrier_done to avoid duplicate work.  Use
2893         * ACCESS_ONCE() to prevent the compiler from speculating
2894         * the increment to precede the early-exit check.
2895         */
2896        ACCESS_ONCE(rsp->n_barrier_done)++;
2897        WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2898        _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2899        smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2900
2901        /*
2902         * Initialize the count to one rather than to zero in order to
2903         * avoid a too-soon return to zero in case of a short grace period
2904         * (or preemption of this task).  Exclude CPU-hotplug operations
2905         * to ensure that no offline CPU has callbacks queued.
2906         */
2907        init_completion(&rsp->barrier_completion);
2908        atomic_set(&rsp->barrier_cpu_count, 1);
2909        get_online_cpus();
2910
2911        /*
2912         * Force each CPU with callbacks to register a new callback.
2913         * When that callback is invoked, we will know that all of the
2914         * corresponding CPU's preceding callbacks have been invoked.
2915         */
2916        for_each_possible_cpu(cpu) {
2917                if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
2918                        continue;
2919                rdp = per_cpu_ptr(rsp->rda, cpu);
2920                if (rcu_is_nocb_cpu(cpu)) {
2921                        _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
2922                                           rsp->n_barrier_done);
2923                        atomic_inc(&rsp->barrier_cpu_count);
2924                        __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
2925                                   rsp, cpu, 0);
2926                } else if (ACCESS_ONCE(rdp->qlen)) {
2927                        _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2928                                           rsp->n_barrier_done);
2929                        smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2930                } else {
2931                        _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2932                                           rsp->n_barrier_done);
2933                }
2934        }
2935        put_online_cpus();
2936
2937        /*
2938         * Now that we have an rcu_barrier_callback() callback on each
2939         * CPU, and thus each counted, remove the initial count.
2940         */
2941        if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2942                complete(&rsp->barrier_completion);
2943
2944        /* Increment ->n_barrier_done to prevent duplicate work. */
2945        smp_mb(); /* Keep increment after above mechanism. */
2946        ACCESS_ONCE(rsp->n_barrier_done)++;
2947        WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2948        _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2949        smp_mb(); /* Keep increment before caller's subsequent code. */
2950
2951        /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2952        wait_for_completion(&rsp->barrier_completion);
2953
2954        /* Other rcu_barrier() invocations can now safely proceed. */
2955        mutex_unlock(&rsp->barrier_mutex);
2956}
2957
2958/**
2959 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2960 */
2961void rcu_barrier_bh(void)
2962{
2963        _rcu_barrier(&rcu_bh_state);
2964}
2965EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2966
2967/**
2968 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2969 */
2970void rcu_barrier_sched(void)
2971{
2972        _rcu_barrier(&rcu_sched_state);
2973}
2974EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2975
2976/*
2977 * Do boot-time initialization of a CPU's per-CPU RCU data.
2978 */
2979static void __init
2980rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2981{
2982        unsigned long flags;
2983        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2984        struct rcu_node *rnp = rcu_get_root(rsp);
2985
2986        /* Set up local state, ensuring consistent view of global state. */
2987        raw_spin_lock_irqsave(&rnp->lock, flags);
2988        rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2989        init_callback_list(rdp);
2990        rdp->qlen_lazy = 0;
2991        ACCESS_ONCE(rdp->qlen) = 0;
2992        rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2993        WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2994        WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2995        rdp->cpu = cpu;
2996        rdp->rsp = rsp;
2997        rcu_boot_init_nocb_percpu_data(rdp);
2998        raw_spin_unlock_irqrestore(&rnp->lock, flags);
2999}
3000
3001/*
3002 * Initialize a CPU's per-CPU RCU data.  Note that only one online or
3003 * offline event can be happening at a given time.  Note also that we
3004 * can accept some slop in the rsp->completed access due to the fact
3005 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3006 */
3007static void __cpuinit
3008rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
3009{
3010        unsigned long flags;
3011        unsigned long mask;
3012        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3013        struct rcu_node *rnp = rcu_get_root(rsp);
3014
3015        /* Exclude new grace periods. */
3016        mutex_lock(&rsp->onoff_mutex);
3017
3018        /* Set up local state, ensuring consistent view of global state. */
3019        raw_spin_lock_irqsave(&rnp->lock, flags);
3020        rdp->beenonline = 1;     /* We have now been online. */
3021        rdp->preemptible = preemptible;
3022        rdp->qlen_last_fqs_check = 0;
3023        rdp->n_force_qs_snap = rsp->n_force_qs;
3024        rdp->blimit = blimit;
3025        init_callback_list(rdp);  /* Re-enable callbacks on this CPU. */
3026        rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3027        atomic_set(&rdp->dynticks->dynticks,
3028                   (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
3029        raw_spin_unlock(&rnp->lock);            /* irqs remain disabled. */
3030
3031        /* Add CPU to rcu_node bitmasks. */
3032        rnp = rdp->mynode;
3033        mask = rdp->grpmask;
3034        do {
3035                /* Exclude any attempts to start a new GP on small systems. */
3036                raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
3037                rnp->qsmaskinit |= mask;
3038                mask = rnp->grpmask;
3039                if (rnp == rdp->mynode) {
3040                        /*
3041                         * If there is a grace period in progress, we will
3042                         * set up to wait for it next time we run the
3043                         * RCU core code.
3044                         */
3045                        rdp->gpnum = rnp->completed;
3046                        rdp->completed = rnp->completed;
3047                        rdp->passed_quiesce = 0;
3048                        rdp->qs_pending = 0;
3049                        trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
3050                }
3051                raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
3052                rnp = rnp->parent;
3053        } while (rnp != NULL && !(rnp->qsmaskinit & mask));
3054        local_irq_restore(flags);
3055
3056        mutex_unlock(&rsp->onoff_mutex);
3057}
3058
3059static void __cpuinit rcu_prepare_cpu(int cpu)
3060{
3061        struct rcu_state *rsp;
3062
3063        for_each_rcu_flavor(rsp)
3064                rcu_init_percpu_data(cpu, rsp,
3065                                     strcmp(rsp->name, "rcu_preempt") == 0);
3066}
3067
3068/*
3069 * Handle CPU online/offline notification events.
3070 */
3071static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
3072                                    unsigned long action, void *hcpu)
3073{
3074        long cpu = (long)hcpu;
3075        struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
3076        struct rcu_node *rnp = rdp->mynode;
3077        struct rcu_state *rsp;
3078
3079        trace_rcu_utilization("Start CPU hotplug");
3080        switch (action) {
3081        case CPU_UP_PREPARE:
3082        case CPU_UP_PREPARE_FROZEN:
3083                rcu_prepare_cpu(cpu);
3084                rcu_prepare_kthreads(cpu);
3085                break;
3086        case CPU_ONLINE:
3087        case CPU_DOWN_FAILED:
3088                rcu_boost_kthread_setaffinity(rnp, -1);
3089                break;
3090        case CPU_DOWN_PREPARE:
3091                rcu_boost_kthread_setaffinity(rnp, cpu);
3092                break;
3093        case CPU_DYING:
3094        case CPU_DYING_FROZEN:
3095                for_each_rcu_flavor(rsp)
3096                        rcu_cleanup_dying_cpu(rsp);
3097                break;
3098        case CPU_DEAD:
3099        case CPU_DEAD_FROZEN:
3100        case CPU_UP_CANCELED:
3101        case CPU_UP_CANCELED_FROZEN:
3102                for_each_rcu_flavor(rsp)
3103                        rcu_cleanup_dead_cpu(cpu, rsp);
3104                break;
3105        default:
3106                break;
3107        }
3108        trace_rcu_utilization("End CPU hotplug");
3109        return NOTIFY_OK;
3110}
3111
3112/*
3113 * Spawn the kthread that handles this RCU flavor's grace periods.
3114 */
3115static int __init rcu_spawn_gp_kthread(void)
3116{
3117        unsigned long flags;
3118        struct rcu_node *rnp;
3119        struct rcu_state *rsp;
3120        struct task_struct *t;
3121
3122        for_each_rcu_flavor(rsp) {
3123                t = kthread_run(rcu_gp_kthread, rsp, rsp->name);
3124                BUG_ON(IS_ERR(t));
3125                rnp = rcu_get_root(rsp);
3126                raw_spin_lock_irqsave(&rnp->lock, flags);
3127                rsp->gp_kthread = t;
3128                raw_spin_unlock_irqrestore(&rnp->lock, flags);
3129                rcu_spawn_nocb_kthreads(rsp);
3130        }
3131        return 0;
3132}
3133early_initcall(rcu_spawn_gp_kthread);
3134
3135/*
3136 * This function is invoked towards the end of the scheduler's initialization
3137 * process.  Before this is called, the idle task might contain
3138 * RCU read-side critical sections (during which time, this idle
3139 * task is booting the system).  After this function is called, the
3140 * idle tasks are prohibited from containing RCU read-side critical
3141 * sections.  This function also enables RCU lockdep checking.
3142 */
3143void rcu_scheduler_starting(void)
3144{
3145        WARN_ON(num_online_cpus() != 1);
3146        WARN_ON(nr_context_switches() > 0);
3147        rcu_scheduler_active = 1;
3148}
3149
3150/*
3151 * Compute the per-level fanout, either using the exact fanout specified
3152 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3153 */
3154#ifdef CONFIG_RCU_FANOUT_EXACT
3155static void __init rcu_init_levelspread(struct rcu_state *rsp)
3156{
3157        int i;
3158
3159        for (i = rcu_num_lvls - 1; i > 0; i--)
3160                rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3161        rsp->levelspread[0] = rcu_fanout_leaf;
3162}
3163#else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3164static void __init rcu_init_levelspread(struct rcu_state *rsp)
3165{
3166        int ccur;
3167        int cprv;
3168        int i;
3169
3170        cprv = nr_cpu_ids;
3171        for (i = rcu_num_lvls - 1; i >= 0; i--) {
3172                ccur = rsp->levelcnt[i];
3173                rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3174                cprv = ccur;
3175        }
3176}
3177#endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3178
3179/*
3180 * Helper function for rcu_init() that initializes one rcu_state structure.
3181 */
3182static void __init rcu_init_one(struct rcu_state *rsp,
3183                struct rcu_data __percpu *rda)
3184{
3185        static char *buf[] = { "rcu_node_0",
3186                               "rcu_node_1",
3187                               "rcu_node_2",
3188                               "rcu_node_3" };  /* Match MAX_RCU_LVLS */
3189        static char *fqs[] = { "rcu_node_fqs_0",
3190                               "rcu_node_fqs_1",
3191                               "rcu_node_fqs_2",
3192                               "rcu_node_fqs_3" };  /* Match MAX_RCU_LVLS */
3193        int cpustride = 1;
3194        int i;
3195        int j;
3196        struct rcu_node *rnp;
3197
3198        BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf));  /* Fix buf[] init! */
3199
3200        /* Silence gcc 4.8 warning about array index out of range. */
3201        if (rcu_num_lvls > RCU_NUM_LVLS)
3202                panic("rcu_init_one: rcu_num_lvls overflow");
3203
3204        /* Initialize the level-tracking arrays. */
3205
3206        for (i = 0; i < rcu_num_lvls; i++)
3207                rsp->levelcnt[i] = num_rcu_lvl[i];
3208        for (i = 1; i < rcu_num_lvls; i++)
3209                rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3210        rcu_init_levelspread(rsp);
3211
3212        /* Initialize the elements themselves, starting from the leaves. */
3213
3214        for (i = rcu_num_lvls - 1; i >= 0; i--) {
3215                cpustride *= rsp->levelspread[i];
3216                rnp = rsp->level[i];
3217                for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3218                        raw_spin_lock_init(&rnp->lock);
3219                        lockdep_set_class_and_name(&rnp->lock,
3220                                                   &rcu_node_class[i], buf[i]);
3221                        raw_spin_lock_init(&rnp->fqslock);
3222                        lockdep_set_class_and_name(&rnp->fqslock,
3223                                                   &rcu_fqs_class[i], fqs[i]);
3224                        rnp->gpnum = rsp->gpnum;
3225                        rnp->completed = rsp->completed;
3226                        rnp->qsmask = 0;
3227                        rnp->qsmaskinit = 0;
3228                        rnp->grplo = j * cpustride;
3229                        rnp->grphi = (j + 1) * cpustride - 1;
3230                        if (rnp->grphi >= NR_CPUS)
3231                                rnp->grphi = NR_CPUS - 1;
3232                        if (i == 0) {
3233                                rnp->grpnum = 0;
3234                                rnp->grpmask = 0;
3235                                rnp->parent = NULL;
3236                        } else {
3237                                rnp->grpnum = j % rsp->levelspread[i - 1];
3238                                rnp->grpmask = 1UL << rnp->grpnum;
3239                                rnp->parent = rsp->level[i - 1] +
3240                                              j / rsp->levelspread[i - 1];
3241                        }
3242                        rnp->level = i;
3243                        INIT_LIST_HEAD(&rnp->blkd_tasks);
3244                        rcu_init_one_nocb(rnp);
3245                }
3246        }
3247
3248        rsp->rda = rda;
3249        init_waitqueue_head(&rsp->gp_wq);
3250        init_irq_work(&rsp->wakeup_work, rsp_wakeup);
3251        rnp = rsp->level[rcu_num_lvls - 1];
3252        for_each_possible_cpu(i) {
3253                while (i > rnp->grphi)
3254                        rnp++;
3255                per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3256                rcu_boot_init_percpu_data(i, rsp);
3257        }
3258        list_add(&rsp->flavors, &rcu_struct_flavors);
3259}
3260
3261/*
3262 * Compute the rcu_node tree geometry from kernel parameters.  This cannot
3263 * replace the definitions in rcutree.h because those are needed to size
3264 * the ->node array in the rcu_state structure.
3265 */
3266static void __init rcu_init_geometry(void)
3267{
3268        int i;
3269        int j;
3270        int n = nr_cpu_ids;
3271        int rcu_capacity[MAX_RCU_LVLS + 1];
3272
3273        /* If the compile-time values are accurate, just leave. */
3274        if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3275            nr_cpu_ids == NR_CPUS)
3276                return;
3277
3278        /*
3279         * Compute number of nodes that can be handled an rcu_node tree
3280         * with the given number of levels.  Setting rcu_capacity[0] makes
3281         * some of the arithmetic easier.
3282         */
3283        rcu_capacity[0] = 1;
3284        rcu_capacity[1] = rcu_fanout_leaf;
3285        for (i = 2; i <= MAX_RCU_LVLS; i++)
3286                rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3287
3288        /*
3289         * The boot-time rcu_fanout_leaf parameter is only permitted
3290         * to increase the leaf-level fanout, not decrease it.  Of course,
3291         * the leaf-level fanout cannot exceed the number of bits in
3292         * the rcu_node masks.  Finally, the tree must be able to accommodate
3293         * the configured number of CPUs.  Complain and fall back to the
3294         * compile-time values if these limits are exceeded.
3295         */
3296        if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3297            rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3298            n > rcu_capacity[MAX_RCU_LVLS]) {
3299                WARN_ON(1);
3300                return;
3301        }
3302
3303        /* Calculate the number of rcu_nodes at each level of the tree. */
3304        for (i = 1; i <= MAX_RCU_LVLS; i++)
3305                if (n <= rcu_capacity[i]) {
3306                        for (j = 0; j <= i; j++)
3307                                num_rcu_lvl[j] =
3308                                        DIV_ROUND_UP(n, rcu_capacity[i - j]);
3309                        rcu_num_lvls = i;
3310                        for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3311                                num_rcu_lvl[j] = 0;
3312                        break;
3313                }
3314
3315        /* Calculate the total number of rcu_node structures. */
3316        rcu_num_nodes = 0;
3317        for (i = 0; i <= MAX_RCU_LVLS; i++)
3318                rcu_num_nodes += num_rcu_lvl[i];
3319        rcu_num_nodes -= n;
3320}
3321
3322void __init rcu_init(void)
3323{
3324        int cpu;
3325
3326        rcu_bootup_announce();
3327        rcu_init_geometry();
3328        rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3329        rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3330        __rcu_init_preempt();
3331        open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3332
3333        /*
3334         * We don't need protection against CPU-hotplug here because
3335         * this is called early in boot, before either interrupts
3336         * or the scheduler are operational.
3337         */
3338        cpu_notifier(rcu_cpu_notify, 0);
3339        for_each_online_cpu(cpu)
3340                rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3341}
3342
3343#include "rcutree_plugin.h"
3344
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.