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 = ULONG_MAX;
 222static ulong jiffies_till_next_fqs = ULONG_MAX;
 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        pr_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                pr_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        pr_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 * Initialize the specified rcu_data structure's callback list to empty.
 989 */
 990static void init_callback_list(struct rcu_data *rdp)
 991{
 992        int i;
 993
 994        if (init_nocb_callback_list(rdp))
 995                return;
 996        rdp->nxtlist = NULL;
 997        for (i = 0; i < RCU_NEXT_SIZE; i++)
 998                rdp->nxttail[i] = &rdp->nxtlist;
 999}
1000
1001/*
1002 * Determine the value that ->completed will have at the end of the
1003 * next subsequent grace period.  This is used to tag callbacks so that
1004 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1005 * been dyntick-idle for an extended period with callbacks under the
1006 * influence of RCU_FAST_NO_HZ.
1007 *
1008 * The caller must hold rnp->lock with interrupts disabled.
1009 */
1010static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1011                                       struct rcu_node *rnp)
1012{
1013        /*
1014         * If RCU is idle, we just wait for the next grace period.
1015         * But we can only be sure that RCU is idle if we are looking
1016         * at the root rcu_node structure -- otherwise, a new grace
1017         * period might have started, but just not yet gotten around
1018         * to initializing the current non-root rcu_node structure.
1019         */
1020        if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1021                return rnp->completed + 1;
1022
1023        /*
1024         * Otherwise, wait for a possible partial grace period and
1025         * then the subsequent full grace period.
1026         */
1027        return rnp->completed + 2;
1028}
1029
1030/*
1031 * Trace-event helper function for rcu_start_future_gp() and
1032 * rcu_nocb_wait_gp().
1033 */
1034static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1035                                unsigned long c, char *s)
1036{
1037        trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1038                                      rnp->completed, c, rnp->level,
1039                                      rnp->grplo, rnp->grphi, s);
1040}
1041
1042/*
1043 * Start some future grace period, as needed to handle newly arrived
1044 * callbacks.  The required future grace periods are recorded in each
1045 * rcu_node structure's ->need_future_gp field.
1046 *
1047 * The caller must hold the specified rcu_node structure's ->lock.
1048 */
1049static unsigned long __maybe_unused
1050rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
1051{
1052        unsigned long c;
1053        int i;
1054        struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1055
1056        /*
1057         * Pick up grace-period number for new callbacks.  If this
1058         * grace period is already marked as needed, return to the caller.
1059         */
1060        c = rcu_cbs_completed(rdp->rsp, rnp);
1061        trace_rcu_future_gp(rnp, rdp, c, "Startleaf");
1062        if (rnp->need_future_gp[c & 0x1]) {
1063                trace_rcu_future_gp(rnp, rdp, c, "Prestartleaf");
1064                return c;
1065        }
1066
1067        /*
1068         * If either this rcu_node structure or the root rcu_node structure
1069         * believe that a grace period is in progress, then we must wait
1070         * for the one following, which is in "c".  Because our request
1071         * will be noticed at the end of the current grace period, we don't
1072         * need to explicitly start one.
1073         */
1074        if (rnp->gpnum != rnp->completed ||
1075            ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
1076                rnp->need_future_gp[c & 0x1]++;
1077                trace_rcu_future_gp(rnp, rdp, c, "Startedleaf");
1078                return c;
1079        }
1080
1081        /*
1082         * There might be no grace period in progress.  If we don't already
1083         * hold it, acquire the root rcu_node structure's lock in order to
1084         * start one (if needed).
1085         */
1086        if (rnp != rnp_root)
1087                raw_spin_lock(&rnp_root->lock);
1088
1089        /*
1090         * Get a new grace-period number.  If there really is no grace
1091         * period in progress, it will be smaller than the one we obtained
1092         * earlier.  Adjust callbacks as needed.  Note that even no-CBs
1093         * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1094         */
1095        c = rcu_cbs_completed(rdp->rsp, rnp_root);
1096        for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1097                if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1098                        rdp->nxtcompleted[i] = c;
1099
1100        /*
1101         * If the needed for the required grace period is already
1102         * recorded, trace and leave.
1103         */
1104        if (rnp_root->need_future_gp[c & 0x1]) {
1105                trace_rcu_future_gp(rnp, rdp, c, "Prestartedroot");
1106                goto unlock_out;
1107        }
1108
1109        /* Record the need for the future grace period. */
1110        rnp_root->need_future_gp[c & 0x1]++;
1111
1112        /* If a grace period is not already in progress, start one. */
1113        if (rnp_root->gpnum != rnp_root->completed) {
1114                trace_rcu_future_gp(rnp, rdp, c, "Startedleafroot");
1115        } else {
1116                trace_rcu_future_gp(rnp, rdp, c, "Startedroot");
1117                rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1118        }
1119unlock_out:
1120        if (rnp != rnp_root)
1121                raw_spin_unlock(&rnp_root->lock);
1122        return c;
1123}
1124
1125/*
1126 * Clean up any old requests for the just-ended grace period.  Also return
1127 * whether any additional grace periods have been requested.  Also invoke
1128 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1129 * waiting for this grace period to complete.
1130 */
1131static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1132{
1133        int c = rnp->completed;
1134        int needmore;
1135        struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1136
1137        rcu_nocb_gp_cleanup(rsp, rnp);
1138        rnp->need_future_gp[c & 0x1] = 0;
1139        needmore = rnp->need_future_gp[(c + 1) & 0x1];
1140        trace_rcu_future_gp(rnp, rdp, c, needmore ? "CleanupMore" : "Cleanup");
1141        return needmore;
1142}
1143
1144/*
1145 * If there is room, assign a ->completed number to any callbacks on
1146 * this CPU that have not already been assigned.  Also accelerate any
1147 * callbacks that were previously assigned a ->completed number that has
1148 * since proven to be too conservative, which can happen if callbacks get
1149 * assigned a ->completed number while RCU is idle, but with reference to
1150 * a non-root rcu_node structure.  This function is idempotent, so it does
1151 * not hurt to call it repeatedly.
1152 *
1153 * The caller must hold rnp->lock with interrupts disabled.
1154 */
1155static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1156                               struct rcu_data *rdp)
1157{
1158        unsigned long c;
1159        int i;
1160
1161        /* If the CPU has no callbacks, nothing to do. */
1162        if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1163                return;
1164
1165        /*
1166         * Starting from the sublist containing the callbacks most
1167         * recently assigned a ->completed number and working down, find the
1168         * first sublist that is not assignable to an upcoming grace period.
1169         * Such a sublist has something in it (first two tests) and has
1170         * a ->completed number assigned that will complete sooner than
1171         * the ->completed number for newly arrived callbacks (last test).
1172         *
1173         * The key point is that any later sublist can be assigned the
1174         * same ->completed number as the newly arrived callbacks, which
1175         * means that the callbacks in any of these later sublist can be
1176         * grouped into a single sublist, whether or not they have already
1177         * been assigned a ->completed number.
1178         */
1179        c = rcu_cbs_completed(rsp, rnp);
1180        for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1181                if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1182                    !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1183                        break;
1184
1185        /*
1186         * If there are no sublist for unassigned callbacks, leave.
1187         * At the same time, advance "i" one sublist, so that "i" will
1188         * index into the sublist where all the remaining callbacks should
1189         * be grouped into.
1190         */
1191        if (++i >= RCU_NEXT_TAIL)
1192                return;
1193
1194        /*
1195         * Assign all subsequent callbacks' ->completed number to the next
1196         * full grace period and group them all in the sublist initially
1197         * indexed by "i".
1198         */
1199        for (; i <= RCU_NEXT_TAIL; i++) {
1200                rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1201                rdp->nxtcompleted[i] = c;
1202        }
1203        /* Record any needed additional grace periods. */
1204        rcu_start_future_gp(rnp, rdp);
1205
1206        /* Trace depending on how much we were able to accelerate. */
1207        if (!*rdp->nxttail[RCU_WAIT_TAIL])
1208                trace_rcu_grace_period(rsp->name, rdp->gpnum, "AccWaitCB");
1209        else
1210                trace_rcu_grace_period(rsp->name, rdp->gpnum, "AccReadyCB");
1211}
1212
1213/*
1214 * Move any callbacks whose grace period has completed to the
1215 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1216 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1217 * sublist.  This function is idempotent, so it does not hurt to
1218 * invoke it repeatedly.  As long as it is not invoked -too- often...
1219 *
1220 * The caller must hold rnp->lock with interrupts disabled.
1221 */
1222static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1223                            struct rcu_data *rdp)
1224{
1225        int i, j;
1226
1227        /* If the CPU has no callbacks, nothing to do. */
1228        if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1229                return;
1230
1231        /*
1232         * Find all callbacks whose ->completed numbers indicate that they
1233         * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1234         */
1235        for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1236                if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1237                        break;
1238                rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1239        }
1240        /* Clean up any sublist tail pointers that were misordered above. */
1241        for (j = RCU_WAIT_TAIL; j < i; j++)
1242                rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1243
1244        /* Copy down callbacks to fill in empty sublists. */
1245        for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1246                if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1247                        break;
1248                rdp->nxttail[j] = rdp->nxttail[i];
1249                rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1250        }
1251
1252        /* Classify any remaining callbacks. */
1253        rcu_accelerate_cbs(rsp, rnp, rdp);
1254}
1255
1256/*
1257 * Update CPU-local rcu_data state to record the beginnings and ends of
1258 * grace periods.  The caller must hold the ->lock of the leaf rcu_node
1259 * structure corresponding to the current CPU, and must have irqs disabled.
1260 */
1261static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1262{
1263        /* Handle the ends of any preceding grace periods first. */
1264        if (rdp->completed == rnp->completed) {
1265
1266                /* No grace period end, so just accelerate recent callbacks. */
1267                rcu_accelerate_cbs(rsp, rnp, rdp);
1268
1269        } else {
1270
1271                /* Advance callbacks. */
1272                rcu_advance_cbs(rsp, rnp, rdp);
1273
1274                /* Remember that we saw this grace-period completion. */
1275                rdp->completed = rnp->completed;
1276                trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
1277        }
1278
1279        if (rdp->gpnum != rnp->gpnum) {
1280                /*
1281                 * If the current grace period is waiting for this CPU,
1282                 * set up to detect a quiescent state, otherwise don't
1283                 * go looking for one.
1284                 */
1285                rdp->gpnum = rnp->gpnum;
1286                trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
1287                rdp->passed_quiesce = 0;
1288                rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1289                zero_cpu_stall_ticks(rdp);
1290        }
1291}
1292
1293static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1294{
1295        unsigned long flags;
1296        struct rcu_node *rnp;
1297
1298        local_irq_save(flags);
1299        rnp = rdp->mynode;
1300        if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1301             rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */
1302            !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1303                local_irq_restore(flags);
1304                return;
1305        }
1306        __note_gp_changes(rsp, rnp, rdp);
1307        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1308}
1309
1310/*
1311 * Initialize a new grace period.
1312 */
1313static int rcu_gp_init(struct rcu_state *rsp)
1314{
1315        struct rcu_data *rdp;
1316        struct rcu_node *rnp = rcu_get_root(rsp);
1317
1318        raw_spin_lock_irq(&rnp->lock);
1319        rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1320
1321        if (rcu_gp_in_progress(rsp)) {
1322                /* Grace period already in progress, don't start another.  */
1323                raw_spin_unlock_irq(&rnp->lock);
1324                return 0;
1325        }
1326
1327        /* Advance to a new grace period and initialize state. */
1328        rsp->gpnum++;
1329        trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
1330        record_gp_stall_check_time(rsp);
1331        raw_spin_unlock_irq(&rnp->lock);
1332
1333        /* Exclude any concurrent CPU-hotplug operations. */
1334        mutex_lock(&rsp->onoff_mutex);
1335
1336        /*
1337         * Set the quiescent-state-needed bits in all the rcu_node
1338         * structures for all currently online CPUs in breadth-first order,
1339         * starting from the root rcu_node structure, relying on the layout
1340         * of the tree within the rsp->node[] array.  Note that other CPUs
1341         * will access only the leaves of the hierarchy, thus seeing that no
1342         * grace period is in progress, at least until the corresponding
1343         * leaf node has been initialized.  In addition, we have excluded
1344         * CPU-hotplug operations.
1345         *
1346         * The grace period cannot complete until the initialization
1347         * process finishes, because this kthread handles both.
1348         */
1349        rcu_for_each_node_breadth_first(rsp, rnp) {
1350                raw_spin_lock_irq(&rnp->lock);
1351                rdp = this_cpu_ptr(rsp->rda);
1352                rcu_preempt_check_blocked_tasks(rnp);
1353                rnp->qsmask = rnp->qsmaskinit;
1354                ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1355                WARN_ON_ONCE(rnp->completed != rsp->completed);
1356                ACCESS_ONCE(rnp->completed) = rsp->completed;
1357                if (rnp == rdp->mynode)
1358                        __note_gp_changes(rsp, rnp, rdp);
1359                rcu_preempt_boost_start_gp(rnp);
1360                trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1361                                            rnp->level, rnp->grplo,
1362                                            rnp->grphi, rnp->qsmask);
1363                raw_spin_unlock_irq(&rnp->lock);
1364#ifdef CONFIG_PROVE_RCU_DELAY
1365                if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 &&
1366                    system_state == SYSTEM_RUNNING)
1367                        udelay(200);
1368#endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1369                cond_resched();
1370        }
1371
1372        mutex_unlock(&rsp->onoff_mutex);
1373        return 1;
1374}
1375
1376/*
1377 * Do one round of quiescent-state forcing.
1378 */
1379int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1380{
1381        int fqs_state = fqs_state_in;
1382        struct rcu_node *rnp = rcu_get_root(rsp);
1383
1384        rsp->n_force_qs++;
1385        if (fqs_state == RCU_SAVE_DYNTICK) {
1386                /* Collect dyntick-idle snapshots. */
1387                force_qs_rnp(rsp, dyntick_save_progress_counter);
1388                fqs_state = RCU_FORCE_QS;
1389        } else {
1390                /* Handle dyntick-idle and offline CPUs. */
1391                force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1392        }
1393        /* Clear flag to prevent immediate re-entry. */
1394        if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1395                raw_spin_lock_irq(&rnp->lock);
1396                rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1397                raw_spin_unlock_irq(&rnp->lock);
1398        }
1399        return fqs_state;
1400}
1401
1402/*
1403 * Clean up after the old grace period.
1404 */
1405static void rcu_gp_cleanup(struct rcu_state *rsp)
1406{
1407        unsigned long gp_duration;
1408        int nocb = 0;
1409        struct rcu_data *rdp;
1410        struct rcu_node *rnp = rcu_get_root(rsp);
1411
1412        raw_spin_lock_irq(&rnp->lock);
1413        gp_duration = jiffies - rsp->gp_start;
1414        if (gp_duration > rsp->gp_max)
1415                rsp->gp_max = gp_duration;
1416
1417        /*
1418         * We know the grace period is complete, but to everyone else
1419         * it appears to still be ongoing.  But it is also the case
1420         * that to everyone else it looks like there is nothing that
1421         * they can do to advance the grace period.  It is therefore
1422         * safe for us to drop the lock in order to mark the grace
1423         * period as completed in all of the rcu_node structures.
1424         */
1425        raw_spin_unlock_irq(&rnp->lock);
1426
1427        /*
1428         * Propagate new ->completed value to rcu_node structures so
1429         * that other CPUs don't have to wait until the start of the next
1430         * grace period to process their callbacks.  This also avoids
1431         * some nasty RCU grace-period initialization races by forcing
1432         * the end of the current grace period to be completely recorded in
1433         * all of the rcu_node structures before the beginning of the next
1434         * grace period is recorded in any of the rcu_node structures.
1435         */
1436        rcu_for_each_node_breadth_first(rsp, rnp) {
1437                raw_spin_lock_irq(&rnp->lock);
1438                ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1439                rdp = this_cpu_ptr(rsp->rda);
1440                if (rnp == rdp->mynode)
1441                        __note_gp_changes(rsp, rnp, rdp);
1442                nocb += rcu_future_gp_cleanup(rsp, rnp);
1443                raw_spin_unlock_irq(&rnp->lock);
1444                cond_resched();
1445        }
1446        rnp = rcu_get_root(rsp);
1447        raw_spin_lock_irq(&rnp->lock);
1448        rcu_nocb_gp_set(rnp, nocb);
1449
1450        rsp->completed = rsp->gpnum; /* Declare grace period done. */
1451        trace_rcu_grace_period(rsp->name, rsp->completed, "end");
1452        rsp->fqs_state = RCU_GP_IDLE;
1453        rdp = this_cpu_ptr(rsp->rda);
1454        rcu_advance_cbs(rsp, rnp, rdp);  /* Reduce false positives below. */
1455        if (cpu_needs_another_gp(rsp, rdp))
1456                rsp->gp_flags = 1;
1457        raw_spin_unlock_irq(&rnp->lock);
1458}
1459
1460/*
1461 * Body of kthread that handles grace periods.
1462 */
1463static int __noreturn rcu_gp_kthread(void *arg)
1464{
1465        int fqs_state;
1466        unsigned long j;
1467        int ret;
1468        struct rcu_state *rsp = arg;
1469        struct rcu_node *rnp = rcu_get_root(rsp);
1470
1471        for (;;) {
1472
1473                /* Handle grace-period start. */
1474                for (;;) {
1475                        wait_event_interruptible(rsp->gp_wq,
1476                                                 rsp->gp_flags &
1477                                                 RCU_GP_FLAG_INIT);
1478                        if ((rsp->gp_flags & RCU_GP_FLAG_INIT) &&
1479                            rcu_gp_init(rsp))
1480                                break;
1481                        cond_resched();
1482                        flush_signals(current);
1483                }
1484
1485                /* Handle quiescent-state forcing. */
1486                fqs_state = RCU_SAVE_DYNTICK;
1487                j = jiffies_till_first_fqs;
1488                if (j > HZ) {
1489                        j = HZ;
1490                        jiffies_till_first_fqs = HZ;
1491                }
1492                for (;;) {
1493                        rsp->jiffies_force_qs = jiffies + j;
1494                        ret = wait_event_interruptible_timeout(rsp->gp_wq,
1495                                        (rsp->gp_flags & RCU_GP_FLAG_FQS) ||
1496                                        (!ACCESS_ONCE(rnp->qsmask) &&
1497                                         !rcu_preempt_blocked_readers_cgp(rnp)),
1498                                        j);
1499                        /* If grace period done, leave loop. */
1500                        if (!ACCESS_ONCE(rnp->qsmask) &&
1501                            !rcu_preempt_blocked_readers_cgp(rnp))
1502                                break;
1503                        /* If time for quiescent-state forcing, do it. */
1504                        if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) {
1505                                fqs_state = rcu_gp_fqs(rsp, fqs_state);
1506                                cond_resched();
1507                        } else {
1508                                /* Deal with stray signal. */
1509                                cond_resched();
1510                                flush_signals(current);
1511                        }
1512                        j = jiffies_till_next_fqs;
1513                        if (j > HZ) {
1514                                j = HZ;
1515                                jiffies_till_next_fqs = HZ;
1516                        } else if (j < 1) {
1517                                j = 1;
1518                                jiffies_till_next_fqs = 1;
1519                        }
1520                }
1521
1522                /* Handle grace-period end. */
1523                rcu_gp_cleanup(rsp);
1524        }
1525}
1526
1527static void rsp_wakeup(struct irq_work *work)
1528{
1529        struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work);
1530
1531        /* Wake up rcu_gp_kthread() to start the grace period. */
1532        wake_up(&rsp->gp_wq);
1533}
1534
1535/*
1536 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1537 * in preparation for detecting the next grace period.  The caller must hold
1538 * the root node's ->lock and hard irqs must be disabled.
1539 *
1540 * Note that it is legal for a dying CPU (which is marked as offline) to
1541 * invoke this function.  This can happen when the dying CPU reports its
1542 * quiescent state.
1543 */
1544static void
1545rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1546                      struct rcu_data *rdp)
1547{
1548        if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1549                /*
1550                 * Either we have not yet spawned the grace-period
1551                 * task, this CPU does not need another grace period,
1552                 * or a grace period is already in progress.
1553                 * Either way, don't start a new grace period.
1554                 */
1555                return;
1556        }
1557        rsp->gp_flags = RCU_GP_FLAG_INIT;
1558
1559        /*
1560         * We can't do wakeups while holding the rnp->lock, as that
1561         * could cause possible deadlocks with the rq->lock. Deter
1562         * the wakeup to interrupt context.
1563         */
1564        irq_work_queue(&rsp->wakeup_work);
1565}
1566
1567/*
1568 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1569 * callbacks.  Note that rcu_start_gp_advanced() cannot do this because it
1570 * is invoked indirectly from rcu_advance_cbs(), which would result in
1571 * endless recursion -- or would do so if it wasn't for the self-deadlock
1572 * that is encountered beforehand.
1573 */
1574static void
1575rcu_start_gp(struct rcu_state *rsp)
1576{
1577        struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1578        struct rcu_node *rnp = rcu_get_root(rsp);
1579
1580        /*
1581         * If there is no grace period in progress right now, any
1582         * callbacks we have up to this point will be satisfied by the
1583         * next grace period.  Also, advancing the callbacks reduces the
1584         * probability of false positives from cpu_needs_another_gp()
1585         * resulting in pointless grace periods.  So, advance callbacks
1586         * then start the grace period!
1587         */
1588        rcu_advance_cbs(rsp, rnp, rdp);
1589        rcu_start_gp_advanced(rsp, rnp, rdp);
1590}
1591
1592/*
1593 * Report a full set of quiescent states to the specified rcu_state
1594 * data structure.  This involves cleaning up after the prior grace
1595 * period and letting rcu_start_gp() start up the next grace period
1596 * if one is needed.  Note that the caller must hold rnp->lock, which
1597 * is released before return.
1598 */
1599static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1600        __releases(rcu_get_root(rsp)->lock)
1601{
1602        WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1603        raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1604        wake_up(&rsp->gp_wq);  /* Memory barrier implied by wake_up() path. */
1605}
1606
1607/*
1608 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1609 * Allows quiescent states for a group of CPUs to be reported at one go
1610 * to the specified rcu_node structure, though all the CPUs in the group
1611 * must be represented by the same rcu_node structure (which need not be
1612 * a leaf rcu_node structure, though it often will be).  That structure's
1613 * lock must be held upon entry, and it is released before return.
1614 */
1615static void
1616rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1617                  struct rcu_node *rnp, unsigned long flags)
1618        __releases(rnp->lock)
1619{
1620        struct rcu_node *rnp_c;
1621
1622        /* Walk up the rcu_node hierarchy. */
1623        for (;;) {
1624                if (!(rnp->qsmask & mask)) {
1625
1626                        /* Our bit has already been cleared, so done. */
1627                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1628                        return;
1629                }
1630                rnp->qsmask &= ~mask;
1631                trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1632                                                 mask, rnp->qsmask, rnp->level,
1633                                                 rnp->grplo, rnp->grphi,
1634                                                 !!rnp->gp_tasks);
1635                if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1636
1637                        /* Other bits still set at this level, so done. */
1638                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1639                        return;
1640                }
1641                mask = rnp->grpmask;
1642                if (rnp->parent == NULL) {
1643
1644                        /* No more levels.  Exit loop holding root lock. */
1645
1646                        break;
1647                }
1648                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1649                rnp_c = rnp;
1650                rnp = rnp->parent;
1651                raw_spin_lock_irqsave(&rnp->lock, flags);
1652                WARN_ON_ONCE(rnp_c->qsmask);
1653        }
1654
1655        /*
1656         * Get here if we are the last CPU to pass through a quiescent
1657         * state for this grace period.  Invoke rcu_report_qs_rsp()
1658         * to clean up and start the next grace period if one is needed.
1659         */
1660        rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1661}
1662
1663/*
1664 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1665 * structure.  This must be either called from the specified CPU, or
1666 * called when the specified CPU is known to be offline (and when it is
1667 * also known that no other CPU is concurrently trying to help the offline
1668 * CPU).  The lastcomp argument is used to make sure we are still in the
1669 * grace period of interest.  We don't want to end the current grace period
1670 * based on quiescent states detected in an earlier grace period!
1671 */
1672static void
1673rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1674{
1675        unsigned long flags;
1676        unsigned long mask;
1677        struct rcu_node *rnp;
1678
1679        rnp = rdp->mynode;
1680        raw_spin_lock_irqsave(&rnp->lock, flags);
1681        if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1682            rnp->completed == rnp->gpnum) {
1683
1684                /*
1685                 * The grace period in which this quiescent state was
1686                 * recorded has ended, so don't report it upwards.
1687                 * We will instead need a new quiescent state that lies
1688                 * within the current grace period.
1689                 */
1690                rdp->passed_quiesce = 0;        /* need qs for new gp. */
1691                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1692                return;
1693        }
1694        mask = rdp->grpmask;
1695        if ((rnp->qsmask & mask) == 0) {
1696                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1697        } else {
1698                rdp->qs_pending = 0;
1699
1700                /*
1701                 * This GP can't end until cpu checks in, so all of our
1702                 * callbacks can be processed during the next GP.
1703                 */
1704                rcu_accelerate_cbs(rsp, rnp, rdp);
1705
1706                rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1707        }
1708}
1709
1710/*
1711 * Check to see if there is a new grace period of which this CPU
1712 * is not yet aware, and if so, set up local rcu_data state for it.
1713 * Otherwise, see if this CPU has just passed through its first
1714 * quiescent state for this grace period, and record that fact if so.
1715 */
1716static void
1717rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1718{
1719        /* Check for grace-period ends and beginnings. */
1720        note_gp_changes(rsp, rdp);
1721
1722        /*
1723         * Does this CPU still need to do its part for current grace period?
1724         * If no, return and let the other CPUs do their part as well.
1725         */
1726        if (!rdp->qs_pending)
1727                return;
1728
1729        /*
1730         * Was there a quiescent state since the beginning of the grace
1731         * period? If no, then exit and wait for the next call.
1732         */
1733        if (!rdp->passed_quiesce)
1734                return;
1735
1736        /*
1737         * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1738         * judge of that).
1739         */
1740        rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1741}
1742
1743#ifdef CONFIG_HOTPLUG_CPU
1744
1745/*
1746 * Send the specified CPU's RCU callbacks to the orphanage.  The
1747 * specified CPU must be offline, and the caller must hold the
1748 * ->orphan_lock.
1749 */
1750static void
1751rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1752                          struct rcu_node *rnp, struct rcu_data *rdp)
1753{
1754        /* No-CBs CPUs do not have orphanable callbacks. */
1755        if (rcu_is_nocb_cpu(rdp->cpu))
1756                return;
1757
1758        /*
1759         * Orphan the callbacks.  First adjust the counts.  This is safe
1760         * because _rcu_barrier() excludes CPU-hotplug operations, so it
1761         * cannot be running now.  Thus no memory barrier is required.
1762         */
1763        if (rdp->nxtlist != NULL) {
1764                rsp->qlen_lazy += rdp->qlen_lazy;
1765                rsp->qlen += rdp->qlen;
1766                rdp->n_cbs_orphaned += rdp->qlen;
1767                rdp->qlen_lazy = 0;
1768                ACCESS_ONCE(rdp->qlen) = 0;
1769        }
1770
1771        /*
1772         * Next, move those callbacks still needing a grace period to
1773         * the orphanage, where some other CPU will pick them up.
1774         * Some of the callbacks might have gone partway through a grace
1775         * period, but that is too bad.  They get to start over because we
1776         * cannot assume that grace periods are synchronized across CPUs.
1777         * We don't bother updating the ->nxttail[] array yet, instead
1778         * we just reset the whole thing later on.
1779         */
1780        if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1781                *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1782                rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1783                *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1784        }
1785
1786        /*
1787         * Then move the ready-to-invoke callbacks to the orphanage,
1788         * where some other CPU will pick them up.  These will not be
1789         * required to pass though another grace period: They are done.
1790         */
1791        if (rdp->nxtlist != NULL) {
1792                *rsp->orphan_donetail = rdp->nxtlist;
1793                rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1794        }
1795
1796        /* Finally, initialize the rcu_data structure's list to empty.  */
1797        init_callback_list(rdp);
1798}
1799
1800/*
1801 * Adopt the RCU callbacks from the specified rcu_state structure's
1802 * orphanage.  The caller must hold the ->orphan_lock.
1803 */
1804static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1805{
1806        int i;
1807        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1808
1809        /* No-CBs CPUs are handled specially. */
1810        if (rcu_nocb_adopt_orphan_cbs(rsp, rdp))
1811                return;
1812
1813        /* Do the accounting first. */
1814        rdp->qlen_lazy += rsp->qlen_lazy;
1815        rdp->qlen += rsp->qlen;
1816        rdp->n_cbs_adopted += rsp->qlen;
1817        if (rsp->qlen_lazy != rsp->qlen)
1818                rcu_idle_count_callbacks_posted();
1819        rsp->qlen_lazy = 0;
1820        rsp->qlen = 0;
1821
1822        /*
1823         * We do not need a memory barrier here because the only way we
1824         * can get here if there is an rcu_barrier() in flight is if
1825         * we are the task doing the rcu_barrier().
1826         */
1827
1828        /* First adopt the ready-to-invoke callbacks. */
1829        if (rsp->orphan_donelist != NULL) {
1830                *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1831                *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1832                for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1833                        if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1834                                rdp->nxttail[i] = rsp->orphan_donetail;
1835                rsp->orphan_donelist = NULL;
1836                rsp->orphan_donetail = &rsp->orphan_donelist;
1837        }
1838
1839        /* And then adopt the callbacks that still need a grace period. */
1840        if (rsp->orphan_nxtlist != NULL) {
1841                *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1842                rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1843                rsp->orphan_nxtlist = NULL;
1844                rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1845        }
1846}
1847
1848/*
1849 * Trace the fact that this CPU is going offline.
1850 */
1851static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1852{
1853        RCU_TRACE(unsigned long mask);
1854        RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1855        RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1856
1857        RCU_TRACE(mask = rdp->grpmask);
1858        trace_rcu_grace_period(rsp->name,
1859                               rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1860                               "cpuofl");
1861}
1862
1863/*
1864 * The CPU has been completely removed, and some other CPU is reporting
1865 * this fact from process context.  Do the remainder of the cleanup,
1866 * including orphaning the outgoing CPU's RCU callbacks, and also
1867 * adopting them.  There can only be one CPU hotplug operation at a time,
1868 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1869 */
1870static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1871{
1872        unsigned long flags;
1873        unsigned long mask;
1874        int need_report = 0;
1875        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1876        struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
1877
1878        /* Adjust any no-longer-needed kthreads. */
1879        rcu_boost_kthread_setaffinity(rnp, -1);
1880
1881        /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1882
1883        /* Exclude any attempts to start a new grace period. */
1884        mutex_lock(&rsp->onoff_mutex);
1885        raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
1886
1887        /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1888        rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1889        rcu_adopt_orphan_cbs(rsp);
1890
1891        /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1892        mask = rdp->grpmask;    /* rnp->grplo is constant. */
1893        do {
1894                raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
1895                rnp->qsmaskinit &= ~mask;
1896                if (rnp->qsmaskinit != 0) {
1897                        if (rnp != rdp->mynode)
1898                                raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1899                        break;
1900                }
1901                if (rnp == rdp->mynode)
1902                        need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1903                else
1904                        raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1905                mask = rnp->grpmask;
1906                rnp = rnp->parent;
1907        } while (rnp != NULL);
1908
1909        /*
1910         * We still hold the leaf rcu_node structure lock here, and
1911         * irqs are still disabled.  The reason for this subterfuge is
1912         * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
1913         * held leads to deadlock.
1914         */
1915        raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
1916        rnp = rdp->mynode;
1917        if (need_report & RCU_OFL_TASKS_NORM_GP)
1918                rcu_report_unblock_qs_rnp(rnp, flags);
1919        else
1920                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1921        if (need_report & RCU_OFL_TASKS_EXP_GP)
1922                rcu_report_exp_rnp(rsp, rnp, true);
1923        WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1924                  "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1925                  cpu, rdp->qlen, rdp->nxtlist);
1926        init_callback_list(rdp);
1927        /* Disallow further callbacks on this CPU. */
1928        rdp->nxttail[RCU_NEXT_TAIL] = NULL;
1929        mutex_unlock(&rsp->onoff_mutex);
1930}
1931
1932#else /* #ifdef CONFIG_HOTPLUG_CPU */
1933
1934static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1935{
1936}
1937
1938static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1939{
1940}
1941
1942#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1943
1944/*
1945 * Invoke any RCU callbacks that have made it to the end of their grace
1946 * period.  Thottle as specified by rdp->blimit.
1947 */
1948static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1949{
1950        unsigned long flags;
1951        struct rcu_head *next, *list, **tail;
1952        long bl, count, count_lazy;
1953        int i;
1954
1955        /* If no callbacks are ready, just return. */
1956        if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
1957                trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
1958                trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1959                                    need_resched(), is_idle_task(current),
1960                                    rcu_is_callbacks_kthread());
1961                return;
1962        }
1963
1964        /*
1965         * Extract the list of ready callbacks, disabling to prevent
1966         * races with call_rcu() from interrupt handlers.
1967         */
1968        local_irq_save(flags);
1969        WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
1970        bl = rdp->blimit;
1971        trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
1972        list = rdp->nxtlist;
1973        rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1974        *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1975        tail = rdp->nxttail[RCU_DONE_TAIL];
1976        for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1977                if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1978                        rdp->nxttail[i] = &rdp->nxtlist;
1979        local_irq_restore(flags);
1980
1981        /* Invoke callbacks. */
1982        count = count_lazy = 0;
1983        while (list) {
1984                next = list->next;
1985                prefetch(next);
1986                debug_rcu_head_unqueue(list);
1987                if (__rcu_reclaim(rsp->name, list))
1988                        count_lazy++;
1989                list = next;
1990                /* Stop only if limit reached and CPU has something to do. */
1991                if (++count >= bl &&
1992                    (need_resched() ||
1993                     (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
1994                        break;
1995        }
1996
1997        local_irq_save(flags);
1998        trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1999                            is_idle_task(current),
2000                            rcu_is_callbacks_kthread());
2001
2002        /* Update count, and requeue any remaining callbacks. */
2003        if (list != NULL) {
2004                *tail = rdp->nxtlist;
2005                rdp->nxtlist = list;
2006                for (i = 0; i < RCU_NEXT_SIZE; i++)
2007                        if (&rdp->nxtlist == rdp->nxttail[i])
2008                                rdp->nxttail[i] = tail;
2009                        else
2010                                break;
2011        }
2012        smp_mb(); /* List handling before counting for rcu_barrier(). */
2013        rdp->qlen_lazy -= count_lazy;
2014        ACCESS_ONCE(rdp->qlen) -= count;
2015        rdp->n_cbs_invoked += count;
2016
2017        /* Reinstate batch limit if we have worked down the excess. */
2018        if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2019                rdp->blimit = blimit;
2020
2021        /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2022        if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2023                rdp->qlen_last_fqs_check = 0;
2024                rdp->n_force_qs_snap = rsp->n_force_qs;
2025        } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2026                rdp->qlen_last_fqs_check = rdp->qlen;
2027        WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2028
2029        local_irq_restore(flags);
2030
2031        /* Re-invoke RCU core processing if there are callbacks remaining. */
2032        if (cpu_has_callbacks_ready_to_invoke(rdp))
2033                invoke_rcu_core();
2034}
2035
2036/*
2037 * Check to see if this CPU is in a non-context-switch quiescent state
2038 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2039 * Also schedule RCU core processing.
2040 *
2041 * This function must be called from hardirq context.  It is normally
2042 * invoked from the scheduling-clock interrupt.  If rcu_pending returns
2043 * false, there is no point in invoking rcu_check_callbacks().
2044 */
2045void rcu_check_callbacks(int cpu, int user)
2046{
2047        trace_rcu_utilization("Start scheduler-tick");
2048        increment_cpu_stall_ticks();
2049        if (user || rcu_is_cpu_rrupt_from_idle()) {
2050
2051                /*
2052                 * Get here if this CPU took its interrupt from user
2053                 * mode or from the idle loop, and if this is not a
2054                 * nested interrupt.  In this case, the CPU is in
2055                 * a quiescent state, so note it.
2056                 *
2057                 * No memory barrier is required here because both
2058                 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2059                 * variables that other CPUs neither access nor modify,
2060                 * at least not while the corresponding CPU is online.
2061                 */
2062
2063                rcu_sched_qs(cpu);
2064                rcu_bh_qs(cpu);
2065
2066        } else if (!in_softirq()) {
2067
2068                /*
2069                 * Get here if this CPU did not take its interrupt from
2070                 * softirq, in other words, if it is not interrupting
2071                 * a rcu_bh read-side critical section.  This is an _bh
2072                 * critical section, so note it.
2073                 */
2074
2075                rcu_bh_qs(cpu);
2076        }
2077        rcu_preempt_check_callbacks(cpu);
2078        if (rcu_pending(cpu))
2079                invoke_rcu_core();
2080        trace_rcu_utilization("End scheduler-tick");
2081}
2082
2083/*
2084 * Scan the leaf rcu_node structures, processing dyntick state for any that
2085 * have not yet encountered a quiescent state, using the function specified.
2086 * Also initiate boosting for any threads blocked on the root rcu_node.
2087 *
2088 * The caller must have suppressed start of new grace periods.
2089 */
2090static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
2091{
2092        unsigned long bit;
2093        int cpu;
2094        unsigned long flags;
2095        unsigned long mask;
2096        struct rcu_node *rnp;
2097
2098        rcu_for_each_leaf_node(rsp, rnp) {
2099                cond_resched();
2100                mask = 0;
2101                raw_spin_lock_irqsave(&rnp->lock, flags);
2102                if (!rcu_gp_in_progress(rsp)) {
2103                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
2104                        return;
2105                }
2106                if (rnp->qsmask == 0) {
2107                        rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2108                        continue;
2109                }
2110                cpu = rnp->grplo;
2111                bit = 1;
2112                for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2113                        if ((rnp->qsmask & bit) != 0 &&
2114                            f(per_cpu_ptr(rsp->rda, cpu)))
2115                                mask |= bit;
2116                }
2117                if (mask != 0) {
2118
2119                        /* rcu_report_qs_rnp() releases rnp->lock. */
2120                        rcu_report_qs_rnp(mask, rsp, rnp, flags);
2121                        continue;
2122                }
2123                raw_spin_unlock_irqrestore(&rnp->lock, flags);
2124        }
2125        rnp = rcu_get_root(rsp);
2126        if (rnp->qsmask == 0) {
2127                raw_spin_lock_irqsave(&rnp->lock, flags);
2128                rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2129        }
2130}
2131
2132/*
2133 * Force quiescent states on reluctant CPUs, and also detect which
2134 * CPUs are in dyntick-idle mode.
2135 */
2136static void force_quiescent_state(struct rcu_state *rsp)
2137{
2138        unsigned long flags;
2139        bool ret;
2140        struct rcu_node *rnp;
2141        struct rcu_node *rnp_old = NULL;
2142
2143        /* Funnel through hierarchy to reduce memory contention. */
2144        rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2145        for (; rnp != NULL; rnp = rnp->parent) {
2146                ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2147                      !raw_spin_trylock(&rnp->fqslock);
2148                if (rnp_old != NULL)
2149                        raw_spin_unlock(&rnp_old->fqslock);
2150                if (ret) {
2151                        rsp->n_force_qs_lh++;
2152                        return;
2153                }
2154                rnp_old = rnp;
2155        }
2156        /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2157
2158        /* Reached the root of the rcu_node tree, acquire lock. */
2159        raw_spin_lock_irqsave(&rnp_old->lock, flags);
2160        raw_spin_unlock(&rnp_old->fqslock);
2161        if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2162                rsp->n_force_qs_lh++;
2163                raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2164                return;  /* Someone beat us to it. */
2165        }
2166        rsp->gp_flags |= RCU_GP_FLAG_FQS;
2167        raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2168        wake_up(&rsp->gp_wq);  /* Memory barrier implied by wake_up() path. */
2169}
2170
2171/*
2172 * This does the RCU core processing work for the specified rcu_state
2173 * and rcu_data structures.  This may be called only from the CPU to
2174 * whom the rdp belongs.
2175 */
2176static void
2177__rcu_process_callbacks(struct rcu_state *rsp)
2178{
2179        unsigned long flags;
2180        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2181
2182        WARN_ON_ONCE(rdp->beenonline == 0);
2183
2184        /* Update RCU state based on any recent quiescent states. */
2185        rcu_check_quiescent_state(rsp, rdp);
2186
2187        /* Does this CPU require a not-yet-started grace period? */
2188        local_irq_save(flags);
2189        if (cpu_needs_another_gp(rsp, rdp)) {
2190                raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2191                rcu_start_gp(rsp);
2192                raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2193        } else {
2194                local_irq_restore(flags);
2195        }
2196
2197        /* If there are callbacks ready, invoke them. */
2198        if (cpu_has_callbacks_ready_to_invoke(rdp))
2199                invoke_rcu_callbacks(rsp, rdp);
2200}
2201
2202/*
2203 * Do RCU core processing for the current CPU.
2204 */
2205static void rcu_process_callbacks(struct softirq_action *unused)
2206{
2207        struct rcu_state *rsp;
2208
2209        if (cpu_is_offline(smp_processor_id()))
2210                return;
2211        trace_rcu_utilization("Start RCU core");
2212        for_each_rcu_flavor(rsp)
2213                __rcu_process_callbacks(rsp);
2214        trace_rcu_utilization("End RCU core");
2215}
2216
2217/*
2218 * Schedule RCU callback invocation.  If the specified type of RCU
2219 * does not support RCU priority boosting, just do a direct call,
2220 * otherwise wake up the per-CPU kernel kthread.  Note that because we
2221 * are running on the current CPU with interrupts disabled, the
2222 * rcu_cpu_kthread_task cannot disappear out from under us.
2223 */
2224static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2225{
2226        if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2227                return;
2228        if (likely(!rsp->boost)) {
2229                rcu_do_batch(rsp, rdp);
2230                return;
2231        }
2232        invoke_rcu_callbacks_kthread();
2233}
2234
2235static void invoke_rcu_core(void)
2236{
2237        if (cpu_online(smp_processor_id()))
2238                raise_softirq(RCU_SOFTIRQ);
2239}
2240
2241/*
2242 * Handle any core-RCU processing required by a call_rcu() invocation.
2243 */
2244static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2245                            struct rcu_head *head, unsigned long flags)
2246{
2247        /*
2248         * If called from an extended quiescent state, invoke the RCU
2249         * core in order to force a re-evaluation of RCU's idleness.
2250         */
2251        if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2252                invoke_rcu_core();
2253
2254        /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2255        if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2256                return;
2257
2258        /*
2259         * Force the grace period if too many callbacks or too long waiting.
2260         * Enforce hysteresis, and don't invoke force_quiescent_state()
2261         * if some other CPU has recently done so.  Also, don't bother
2262         * invoking force_quiescent_state() if the newly enqueued callback
2263         * is the only one waiting for a grace period to complete.
2264         */
2265        if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2266
2267                /* Are we ignoring a completed grace period? */
2268                note_gp_changes(rsp, rdp);
2269
2270                /* Start a new grace period if one not already started. */
2271                if (!rcu_gp_in_progress(rsp)) {
2272                        struct rcu_node *rnp_root = rcu_get_root(rsp);
2273
2274                        raw_spin_lock(&rnp_root->lock);
2275                        rcu_start_gp(rsp);
2276                        raw_spin_unlock(&rnp_root->lock);
2277                } else {
2278                        /* Give the grace period a kick. */
2279                        rdp->blimit = LONG_MAX;
2280                        if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2281                            *rdp->nxttail[RCU_DONE_TAIL] != head)
2282                                force_quiescent_state(rsp);
2283                        rdp->n_force_qs_snap = rsp->n_force_qs;
2284                        rdp->qlen_last_fqs_check = rdp->qlen;
2285                }
2286        }
2287}
2288
2289/*
2290 * Helper function for call_rcu() and friends.  The cpu argument will
2291 * normally be -1, indicating "currently running CPU".  It may specify
2292 * a CPU only if that CPU is a no-CBs CPU.  Currently, only _rcu_barrier()
2293 * is expected to specify a CPU.
2294 */
2295static void
2296__call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2297           struct rcu_state *rsp, int cpu, bool lazy)
2298{
2299        unsigned long flags;
2300        struct rcu_data *rdp;
2301
2302        WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2303        debug_rcu_head_queue(head);
2304        head->func = func;
2305        head->next = NULL;
2306
2307        /*
2308         * Opportunistically note grace-period endings and beginnings.
2309         * Note that we might see a beginning right after we see an
2310         * end, but never vice versa, since this CPU has to pass through
2311         * a quiescent state betweentimes.
2312         */
2313        local_irq_save(flags);
2314        rdp = this_cpu_ptr(rsp->rda);
2315
2316        /* Add the callback to our list. */
2317        if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2318                int offline;
2319
2320                if (cpu != -1)
2321                        rdp = per_cpu_ptr(rsp->rda, cpu);
2322                offline = !__call_rcu_nocb(rdp, head, lazy);
2323                WARN_ON_ONCE(offline);
2324                /* _call_rcu() is illegal on offline CPU; leak the callback. */
2325                local_irq_restore(flags);
2326                return;
2327        }
2328        ACCESS_ONCE(rdp->qlen)++;
2329        if (lazy)
2330                rdp->qlen_lazy++;
2331        else
2332                rcu_idle_count_callbacks_posted();
2333        smp_mb();  /* Count before adding callback for rcu_barrier(). */
2334        *rdp->nxttail[RCU_NEXT_TAIL] = head;
2335        rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2336
2337        if (__is_kfree_rcu_offset((unsigned long)func))
2338                trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2339                                         rdp->qlen_lazy, rdp->qlen);
2340        else
2341                trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2342
2343        /* Go handle any RCU core processing required. */
2344        __call_rcu_core(rsp, rdp, head, flags);
2345        local_irq_restore(flags);
2346}
2347
2348/*
2349 * Queue an RCU-sched callback for invocation after a grace period.
2350 */
2351void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2352{
2353        __call_rcu(head, func, &rcu_sched_state, -1, 0);
2354}
2355EXPORT_SYMBOL_GPL(call_rcu_sched);
2356
2357/*
2358 * Queue an RCU callback for invocation after a quicker grace period.
2359 */
2360void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2361{
2362        __call_rcu(head, func, &rcu_bh_state, -1, 0);
2363}
2364EXPORT_SYMBOL_GPL(call_rcu_bh);
2365
2366/*
2367 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2368 * any blocking grace-period wait automatically implies a grace period
2369 * if there is only one CPU online at any point time during execution
2370 * of either synchronize_sched() or synchronize_rcu_bh().  It is OK to
2371 * occasionally incorrectly indicate that there are multiple CPUs online
2372 * when there was in fact only one the whole time, as this just adds
2373 * some overhead: RCU still operates correctly.
2374 */
2375static inline int rcu_blocking_is_gp(void)
2376{
2377        int ret;
2378
2379        might_sleep();  /* Check for RCU read-side critical section. */
2380        preempt_disable();
2381        ret = num_online_cpus() <= 1;
2382        preempt_enable();
2383        return ret;
2384}
2385
2386/**
2387 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2388 *
2389 * Control will return to the caller some time after a full rcu-sched
2390 * grace period has elapsed, in other words after all currently executing
2391 * rcu-sched read-side critical sections have completed.   These read-side
2392 * critical sections are delimited by rcu_read_lock_sched() and
2393 * rcu_read_unlock_sched(), and may be nested.  Note that preempt_disable(),
2394 * local_irq_disable(), and so on may be used in place of
2395 * rcu_read_lock_sched().
2396 *
2397 * This means that all preempt_disable code sequences, including NMI and
2398 * non-threaded hardware-interrupt handlers, in progress on entry will
2399 * have completed before this primitive returns.  However, this does not
2400 * guarantee that softirq handlers will have completed, since in some
2401 * kernels, these handlers can run in process context, and can block.
2402 *
2403 * Note that this guarantee implies further memory-ordering guarantees.
2404 * On systems with more than one CPU, when synchronize_sched() returns,
2405 * each CPU is guaranteed to have executed a full memory barrier since the
2406 * end of its last RCU-sched read-side critical section whose beginning
2407 * preceded the call to synchronize_sched().  In addition, each CPU having
2408 * an RCU read-side critical section that extends beyond the return from
2409 * synchronize_sched() is guaranteed to have executed a full memory barrier
2410 * after the beginning of synchronize_sched() and before the beginning of
2411 * that RCU read-side critical section.  Note that these guarantees include
2412 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2413 * that are executing in the kernel.
2414 *
2415 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2416 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2417 * to have executed a full memory barrier during the execution of
2418 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2419 * again only if the system has more than one CPU).
2420 *
2421 * This primitive provides the guarantees made by the (now removed)
2422 * synchronize_kernel() API.  In contrast, synchronize_rcu() only
2423 * guarantees that rcu_read_lock() sections will have completed.
2424 * In "classic RCU", these two guarantees happen to be one and
2425 * the same, but can differ in realtime RCU implementations.
2426 */
2427void synchronize_sched(void)
2428{
2429        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2430                           !lock_is_held(&rcu_lock_map) &&
2431                           !lock_is_held(&rcu_sched_lock_map),
2432                           "Illegal synchronize_sched() in RCU-sched read-side critical section");
2433        if (rcu_blocking_is_gp())
2434                return;
2435        if (rcu_expedited)
2436                synchronize_sched_expedited();
2437        else
2438                wait_rcu_gp(call_rcu_sched);
2439}
2440EXPORT_SYMBOL_GPL(synchronize_sched);
2441
2442/**
2443 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2444 *
2445 * Control will return to the caller some time after a full rcu_bh grace
2446 * period has elapsed, in other words after all currently executing rcu_bh
2447 * read-side critical sections have completed.  RCU read-side critical
2448 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2449 * and may be nested.
2450 *
2451 * See the description of synchronize_sched() for more detailed information
2452 * on memory ordering guarantees.
2453 */
2454void synchronize_rcu_bh(void)
2455{
2456        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2457                           !lock_is_held(&rcu_lock_map) &&
2458                           !lock_is_held(&rcu_sched_lock_map),
2459                           "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2460        if (rcu_blocking_is_gp())
2461                return;
2462        if (rcu_expedited)
2463                synchronize_rcu_bh_expedited();
2464        else
2465                wait_rcu_gp(call_rcu_bh);
2466}
2467EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2468
2469static int synchronize_sched_expedited_cpu_stop(void *data)
2470{
2471        /*
2472         * There must be a full memory barrier on each affected CPU
2473         * between the time that try_stop_cpus() is called and the
2474         * time that it returns.
2475         *
2476         * In the current initial implementation of cpu_stop, the
2477         * above condition is already met when the control reaches
2478         * this point and the following smp_mb() is not strictly
2479         * necessary.  Do smp_mb() anyway for documentation and
2480         * robustness against future implementation changes.
2481         */
2482        smp_mb(); /* See above comment block. */
2483        return 0;
2484}
2485
2486/**
2487 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2488 *
2489 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2490 * approach to force the grace period to end quickly.  This consumes
2491 * significant time on all CPUs and is unfriendly to real-time workloads,
2492 * so is thus not recommended for any sort of common-case code.  In fact,
2493 * if you are using synchronize_sched_expedited() in a loop, please
2494 * restructure your code to batch your updates, and then use a single
2495 * synchronize_sched() instead.
2496 *
2497 * Note that it is illegal to call this function while holding any lock
2498 * that is acquired by a CPU-hotplug notifier.  And yes, it is also illegal
2499 * to call this function from a CPU-hotplug notifier.  Failing to observe
2500 * these restriction will result in deadlock.
2501 *
2502 * This implementation can be thought of as an application of ticket
2503 * locking to RCU, with sync_sched_expedited_started and
2504 * sync_sched_expedited_done taking on the roles of the halves
2505 * of the ticket-lock word.  Each task atomically increments
2506 * sync_sched_expedited_started upon entry, snapshotting the old value,
2507 * then attempts to stop all the CPUs.  If this succeeds, then each
2508 * CPU will have executed a context switch, resulting in an RCU-sched
2509 * grace period.  We are then done, so we use atomic_cmpxchg() to
2510 * update sync_sched_expedited_done to match our snapshot -- but
2511 * only if someone else has not already advanced past our snapshot.
2512 *
2513 * On the other hand, if try_stop_cpus() fails, we check the value
2514 * of sync_sched_expedited_done.  If it has advanced past our
2515 * initial snapshot, then someone else must have forced a grace period
2516 * some time after we took our snapshot.  In this case, our work is
2517 * done for us, and we can simply return.  Otherwise, we try again,
2518 * but keep our initial snapshot for purposes of checking for someone
2519 * doing our work for us.
2520 *
2521 * If we fail too many times in a row, we fall back to synchronize_sched().
2522 */
2523void synchronize_sched_expedited(void)
2524{
2525        long firstsnap, s, snap;
2526        int trycount = 0;
2527        struct rcu_state *rsp = &rcu_sched_state;
2528
2529        /*
2530         * If we are in danger of counter wrap, just do synchronize_sched().
2531         * By allowing sync_sched_expedited_started to advance no more than
2532         * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2533         * that more than 3.5 billion CPUs would be required to force a
2534         * counter wrap on a 32-bit system.  Quite a few more CPUs would of
2535         * course be required on a 64-bit system.
2536         */
2537        if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2538                         (ulong)atomic_long_read(&rsp->expedited_done) +
2539                         ULONG_MAX / 8)) {
2540                synchronize_sched();
2541                atomic_long_inc(&rsp->expedited_wrap);
2542                return;
2543        }
2544
2545        /*
2546         * Take a ticket.  Note that atomic_inc_return() implies a
2547         * full memory barrier.
2548         */
2549        snap = atomic_long_inc_return(&rsp->expedited_start);
2550        firstsnap = snap;
2551        get_online_cpus();
2552        WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2553
2554        /*
2555         * Each pass through the following loop attempts to force a
2556         * context switch on each CPU.
2557         */
2558        while (try_stop_cpus(cpu_online_mask,
2559                             synchronize_sched_expedited_cpu_stop,
2560                             NULL) == -EAGAIN) {
2561                put_online_cpus();
2562                atomic_long_inc(&rsp->expedited_tryfail);
2563
2564                /* Check to see if someone else did our work for us. */
2565                s = atomic_long_read(&rsp->expedited_done);
2566                if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2567                        /* ensure test happens before caller kfree */
2568                        smp_mb__before_atomic_inc(); /* ^^^ */
2569                        atomic_long_inc(&rsp->expedited_workdone1);
2570                        return;
2571                }
2572
2573                /* No joy, try again later.  Or just synchronize_sched(). */
2574                if (trycount++ < 10) {
2575                        udelay(trycount * num_online_cpus());
2576                } else {
2577                        wait_rcu_gp(call_rcu_sched);
2578                        atomic_long_inc(&rsp->expedited_normal);
2579                        return;
2580                }
2581
2582                /* Recheck to see if someone else did our work for us. */
2583                s = atomic_long_read(&rsp->expedited_done);
2584                if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2585                        /* ensure test happens before caller kfree */
2586                        smp_mb__before_atomic_inc(); /* ^^^ */
2587                        atomic_long_inc(&rsp->expedited_workdone2);
2588                        return;
2589                }
2590
2591                /*
2592                 * Refetching sync_sched_expedited_started allows later
2593                 * callers to piggyback on our grace period.  We retry
2594                 * after they started, so our grace period works for them,
2595                 * and they started after our first try, so their grace
2596                 * period works for us.
2597                 */
2598                get_online_cpus();
2599                snap = atomic_long_read(&rsp->expedited_start);
2600                smp_mb(); /* ensure read is before try_stop_cpus(). */
2601        }
2602        atomic_long_inc(&rsp->expedited_stoppedcpus);
2603
2604        /*
2605         * Everyone up to our most recent fetch is covered by our grace
2606         * period.  Update the counter, but only if our work is still
2607         * relevant -- which it won't be if someone who started later
2608         * than we did already did their update.
2609         */
2610        do {
2611                atomic_long_inc(&rsp->expedited_done_tries);
2612                s = atomic_long_read(&rsp->expedited_done);
2613                if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2614                        /* ensure test happens before caller kfree */
2615                        smp_mb__before_atomic_inc(); /* ^^^ */
2616                        atomic_long_inc(&rsp->expedited_done_lost);
2617                        break;
2618                }
2619        } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2620        atomic_long_inc(&rsp->expedited_done_exit);
2621
2622        put_online_cpus();
2623}
2624EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2625
2626/*
2627 * Check to see if there is any immediate RCU-related work to be done
2628 * by the current CPU, for the specified type of RCU, returning 1 if so.
2629 * The checks are in order of increasing expense: checks that can be
2630 * carried out against CPU-local state are performed first.  However,
2631 * we must check for CPU stalls first, else we might not get a chance.
2632 */
2633static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2634{
2635        struct rcu_node *rnp = rdp->mynode;
2636
2637        rdp->n_rcu_pending++;
2638
2639        /* Check for CPU stalls, if enabled. */
2640        check_cpu_stall(rsp, rdp);
2641
2642        /* Is the RCU core waiting for a quiescent state from this CPU? */
2643        if (rcu_scheduler_fully_active &&
2644            rdp->qs_pending && !rdp->passed_quiesce) {
2645                rdp->n_rp_qs_pending++;
2646        } else if (rdp->qs_pending && rdp->passed_quiesce) {
2647                rdp->n_rp_report_qs++;
2648                return 1;
2649        }
2650
2651        /* Does this CPU have callbacks ready to invoke? */
2652        if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2653                rdp->n_rp_cb_ready++;
2654                return 1;
2655        }
2656
2657        /* Has RCU gone idle with this CPU needing another grace period? */
2658        if (cpu_needs_another_gp(rsp, rdp)) {
2659                rdp->n_rp_cpu_needs_gp++;
2660                return 1;
2661        }
2662
2663        /* Has another RCU grace period completed?  */
2664        if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2665                rdp->n_rp_gp_completed++;
2666                return 1;
2667        }
2668
2669        /* Has a new RCU grace period started? */
2670        if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2671                rdp->n_rp_gp_started++;
2672                return 1;
2673        }
2674
2675        /* nothing to do */
2676        rdp->n_rp_need_nothing++;
2677        return 0;
2678}
2679
2680/*
2681 * Check to see if there is any immediate RCU-related work to be done
2682 * by the current CPU, returning 1 if so.  This function is part of the
2683 * RCU implementation; it is -not- an exported member of the RCU API.
2684 */
2685static int rcu_pending(int cpu)
2686{
2687        struct rcu_state *rsp;
2688
2689        for_each_rcu_flavor(rsp)
2690                if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2691                        return 1;
2692        return 0;
2693}
2694
2695/*
2696 * Return true if the specified CPU has any callback.  If all_lazy is
2697 * non-NULL, store an indication of whether all callbacks are lazy.
2698 * (If there are no callbacks, all of them are deemed to be lazy.)
2699 */
2700static int rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
2701{
2702        bool al = true;
2703        bool hc = false;
2704        struct rcu_data *rdp;
2705        struct rcu_state *rsp;
2706
2707        for_each_rcu_flavor(rsp) {
2708                rdp = per_cpu_ptr(rsp->rda, cpu);
2709                if (rdp->qlen != rdp->qlen_lazy)
2710                        al = false;
2711                if (rdp->nxtlist)
2712                        hc = true;
2713        }
2714        if (all_lazy)
2715                *all_lazy = al;
2716        return hc;
2717}
2718
2719/*
2720 * Helper function for _rcu_barrier() tracing.  If tracing is disabled,
2721 * the compiler is expected to optimize this away.
2722 */
2723static void _rcu_barrier_trace(struct rcu_state *rsp, char *s,
2724                               int cpu, unsigned long done)
2725{
2726        trace_rcu_barrier(rsp->name, s, cpu,
2727                          atomic_read(&rsp->barrier_cpu_count), done);
2728}
2729
2730/*
2731 * RCU callback function for _rcu_barrier().  If we are last, wake
2732 * up the task executing _rcu_barrier().
2733 */
2734static void rcu_barrier_callback(struct rcu_head *rhp)
2735{
2736        struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2737        struct rcu_state *rsp = rdp->rsp;
2738
2739        if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2740                _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2741                complete(&rsp->barrier_completion);
2742        } else {
2743                _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2744        }
2745}
2746
2747/*
2748 * Called with preemption disabled, and from cross-cpu IRQ context.
2749 */
2750static void rcu_barrier_func(void *type)
2751{
2752        struct rcu_state *rsp = type;
2753        struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2754
2755        _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2756        atomic_inc(&rsp->barrier_cpu_count);
2757        rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2758}
2759
2760/*
2761 * Orchestrate the specified type of RCU barrier, waiting for all
2762 * RCU callbacks of the specified type to complete.
2763 */
2764static void _rcu_barrier(struct rcu_state *rsp)
2765{
2766        int cpu;
2767        struct rcu_data *rdp;
2768        unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2769        unsigned long snap_done;
2770
2771        _rcu_barrier_trace(rsp, "Begin", -1, snap);
2772
2773        /* Take mutex to serialize concurrent rcu_barrier() requests. */
2774        mutex_lock(&rsp->barrier_mutex);
2775
2776        /*
2777         * Ensure that all prior references, including to ->n_barrier_done,
2778         * are ordered before the _rcu_barrier() machinery.
2779         */
2780        smp_mb();  /* See above block comment. */
2781
2782        /*
2783         * Recheck ->n_barrier_done to see if others did our work for us.
2784         * This means checking ->n_barrier_done for an even-to-odd-to-even
2785         * transition.  The "if" expression below therefore rounds the old
2786         * value up to the next even number and adds two before comparing.
2787         */
2788        snap_done = ACCESS_ONCE(rsp->n_barrier_done);
2789        _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2790        if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) {
2791                _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2792                smp_mb(); /* caller's subsequent code after above check. */
2793                mutex_unlock(&rsp->barrier_mutex);
2794                return;
2795        }
2796
2797        /*
2798         * Increment ->n_barrier_done to avoid duplicate work.  Use
2799         * ACCESS_ONCE() to prevent the compiler from speculating
2800         * the increment to precede the early-exit check.
2801         */
2802        ACCESS_ONCE(rsp->n_barrier_done)++;
2803        WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2804        _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2805        smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2806
2807        /*
2808         * Initialize the count to one rather than to zero in order to
2809         * avoid a too-soon return to zero in case of a short grace period
2810         * (or preemption of this task).  Exclude CPU-hotplug operations
2811         * to ensure that no offline CPU has callbacks queued.
2812         */
2813        init_completion(&rsp->barrier_completion);
2814        atomic_set(&rsp->barrier_cpu_count, 1);
2815        get_online_cpus();
2816
2817        /*
2818         * Force each CPU with callbacks to register a new callback.
2819         * When that callback is invoked, we will know that all of the
2820         * corresponding CPU's preceding callbacks have been invoked.
2821         */
2822        for_each_possible_cpu(cpu) {
2823                if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
2824                        continue;
2825                rdp = per_cpu_ptr(rsp->rda, cpu);
2826                if (rcu_is_nocb_cpu(cpu)) {
2827                        _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
2828                                           rsp->n_barrier_done);
2829                        atomic_inc(&rsp->barrier_cpu_count);
2830                        __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
2831                                   rsp, cpu, 0);
2832                } else if (ACCESS_ONCE(rdp->qlen)) {
2833                        _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2834                                           rsp->n_barrier_done);
2835                        smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2836                } else {
2837                        _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2838                                           rsp->n_barrier_done);
2839                }
2840        }
2841        put_online_cpus();
2842
2843        /*
2844         * Now that we have an rcu_barrier_callback() callback on each
2845         * CPU, and thus each counted, remove the initial count.
2846         */
2847        if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2848                complete(&rsp->barrier_completion);
2849
2850        /* Increment ->n_barrier_done to prevent duplicate work. */
2851        smp_mb(); /* Keep increment after above mechanism. */
2852        ACCESS_ONCE(rsp->n_barrier_done)++;
2853        WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2854        _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2855        smp_mb(); /* Keep increment before caller's subsequent code. */
2856
2857        /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2858        wait_for_completion(&rsp->barrier_completion);
2859
2860        /* Other rcu_barrier() invocations can now safely proceed. */
2861        mutex_unlock(&rsp->barrier_mutex);
2862}
2863
2864/**
2865 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2866 */
2867void rcu_barrier_bh(void)
2868{
2869        _rcu_barrier(&rcu_bh_state);
2870}
2871EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2872
2873/**
2874 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2875 */
2876void rcu_barrier_sched(void)
2877{
2878        _rcu_barrier(&rcu_sched_state);
2879}
2880EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2881
2882/*
2883 * Do boot-time initialization of a CPU's per-CPU RCU data.
2884 */
2885static void __init
2886rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2887{
2888        unsigned long flags;
2889        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2890        struct rcu_node *rnp = rcu_get_root(rsp);
2891
2892        /* Set up local state, ensuring consistent view of global state. */
2893        raw_spin_lock_irqsave(&rnp->lock, flags);
2894        rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2895        init_callback_list(rdp);
2896        rdp->qlen_lazy = 0;
2897        ACCESS_ONCE(rdp->qlen) = 0;
2898        rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2899        WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
2900        WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
2901        rdp->cpu = cpu;
2902        rdp->rsp = rsp;
2903        rcu_boot_init_nocb_percpu_data(rdp);
2904        raw_spin_unlock_irqrestore(&rnp->lock, flags);
2905}
2906
2907/*
2908 * Initialize a CPU's per-CPU RCU data.  Note that only one online or
2909 * offline event can be happening at a given time.  Note also that we
2910 * can accept some slop in the rsp->completed access due to the fact
2911 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2912 */
2913static void
2914rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2915{
2916        unsigned long flags;
2917        unsigned long mask;
2918        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2919        struct rcu_node *rnp = rcu_get_root(rsp);
2920
2921        /* Exclude new grace periods. */
2922        mutex_lock(&rsp->onoff_mutex);
2923
2924        /* Set up local state, ensuring consistent view of global state. */
2925        raw_spin_lock_irqsave(&rnp->lock, flags);
2926        rdp->beenonline = 1;     /* We have now been online. */
2927        rdp->preemptible = preemptible;
2928        rdp->qlen_last_fqs_check = 0;
2929        rdp->n_force_qs_snap = rsp->n_force_qs;
2930        rdp->blimit = blimit;
2931        init_callback_list(rdp);  /* Re-enable callbacks on this CPU. */
2932        rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
2933        atomic_set(&rdp->dynticks->dynticks,
2934                   (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
2935        raw_spin_unlock(&rnp->lock);            /* irqs remain disabled. */
2936
2937        /* Add CPU to rcu_node bitmasks. */
2938        rnp = rdp->mynode;
2939        mask = rdp->grpmask;
2940        do {
2941                /* Exclude any attempts to start a new GP on small systems. */
2942                raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
2943                rnp->qsmaskinit |= mask;
2944                mask = rnp->grpmask;
2945                if (rnp == rdp->mynode) {
2946                        /*
2947                         * If there is a grace period in progress, we will
2948                         * set up to wait for it next time we run the
2949                         * RCU core code.
2950                         */
2951                        rdp->gpnum = rnp->completed;
2952                        rdp->completed = rnp->completed;
2953                        rdp->passed_quiesce = 0;
2954                        rdp->qs_pending = 0;
2955                        trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
2956                }
2957                raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2958                rnp = rnp->parent;
2959        } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2960        local_irq_restore(flags);
2961
2962        mutex_unlock(&rsp->onoff_mutex);
2963}
2964
2965static void rcu_prepare_cpu(int cpu)
2966{
2967        struct rcu_state *rsp;
2968
2969        for_each_rcu_flavor(rsp)
2970                rcu_init_percpu_data(cpu, rsp,
2971                                     strcmp(rsp->name, "rcu_preempt") == 0);
2972}
2973
2974/*
2975 * Handle CPU online/offline notification events.
2976 */
2977static int rcu_cpu_notify(struct notifier_block *self,
2978                                    unsigned long action, void *hcpu)
2979{
2980        long cpu = (long)hcpu;
2981        struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2982        struct rcu_node *rnp = rdp->mynode;
2983        struct rcu_state *rsp;
2984
2985        trace_rcu_utilization("Start CPU hotplug");
2986        switch (action) {
2987        case CPU_UP_PREPARE:
2988        case CPU_UP_PREPARE_FROZEN:
2989                rcu_prepare_cpu(cpu);
2990                rcu_prepare_kthreads(cpu);
2991                break;
2992        case CPU_ONLINE:
2993        case CPU_DOWN_FAILED:
2994                rcu_boost_kthread_setaffinity(rnp, -1);
2995                break;
2996        case CPU_DOWN_PREPARE:
2997                rcu_boost_kthread_setaffinity(rnp, cpu);
2998                break;
2999        case CPU_DYING:
3000        case CPU_DYING_FROZEN:
3001                for_each_rcu_flavor(rsp)
3002                        rcu_cleanup_dying_cpu(rsp);
3003                break;
3004        case CPU_DEAD:
3005        case CPU_DEAD_FROZEN:
3006        case CPU_UP_CANCELED:
3007        case CPU_UP_CANCELED_FROZEN:
3008                for_each_rcu_flavor(rsp)
3009                        rcu_cleanup_dead_cpu(cpu, rsp);
3010                break;
3011        default:
3012                break;
3013        }
3014        trace_rcu_utilization("End CPU hotplug");
3015        return NOTIFY_OK;
3016}
3017
3018/*
3019 * Spawn the kthread that handles this RCU flavor's grace periods.
3020 */
3021static int __init rcu_spawn_gp_kthread(void)
3022{
3023        unsigned long flags;
3024        struct rcu_node *rnp;
3025        struct rcu_state *rsp;
3026        struct task_struct *t;
3027
3028        for_each_rcu_flavor(rsp) {
3029                t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
3030                BUG_ON(IS_ERR(t));
3031                rnp = rcu_get_root(rsp);
3032                raw_spin_lock_irqsave(&rnp->lock, flags);
3033                rsp->gp_kthread = t;
3034                raw_spin_unlock_irqrestore(&rnp->lock, flags);
3035                rcu_spawn_nocb_kthreads(rsp);
3036        }
3037        return 0;
3038}
3039early_initcall(rcu_spawn_gp_kthread);
3040
3041/*
3042 * This function is invoked towards the end of the scheduler's initialization
3043 * process.  Before this is called, the idle task might contain
3044 * RCU read-side critical sections (during which time, this idle
3045 * task is booting the system).  After this function is called, the
3046 * idle tasks are prohibited from containing RCU read-side critical
3047 * sections.  This function also enables RCU lockdep checking.
3048 */
3049void rcu_scheduler_starting(void)
3050{
3051        WARN_ON(num_online_cpus() != 1);
3052        WARN_ON(nr_context_switches() > 0);
3053        rcu_scheduler_active = 1;
3054}
3055
3056/*
3057 * Compute the per-level fanout, either using the exact fanout specified
3058 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3059 */
3060#ifdef CONFIG_RCU_FANOUT_EXACT
3061static void __init rcu_init_levelspread(struct rcu_state *rsp)
3062{
3063        int i;
3064
3065        for (i = rcu_num_lvls - 1; i > 0; i--)
3066                rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3067        rsp->levelspread[0] = rcu_fanout_leaf;
3068}
3069#else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3070static void __init rcu_init_levelspread(struct rcu_state *rsp)
3071{
3072        int ccur;
3073        int cprv;
3074        int i;
3075
3076        cprv = nr_cpu_ids;
3077        for (i = rcu_num_lvls - 1; i >= 0; i--) {
3078                ccur = rsp->levelcnt[i];
3079                rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3080                cprv = ccur;
3081        }
3082}
3083#endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3084
3085/*
3086 * Helper function for rcu_init() that initializes one rcu_state structure.
3087 */
3088static void __init rcu_init_one(struct rcu_state *rsp,
3089                struct rcu_data __percpu *rda)
3090{
3091        static char *buf[] = { "rcu_node_0",
3092                               "rcu_node_1",
3093                               "rcu_node_2",
3094                               "rcu_node_3" };  /* Match MAX_RCU_LVLS */
3095        static char *fqs[] = { "rcu_node_fqs_0",
3096                               "rcu_node_fqs_1",
3097                               "rcu_node_fqs_2",
3098                               "rcu_node_fqs_3" };  /* Match MAX_RCU_LVLS */
3099        int cpustride = 1;
3100        int i;
3101        int j;
3102        struct rcu_node *rnp;
3103
3104        BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf));  /* Fix buf[] init! */
3105
3106        /* Silence gcc 4.8 warning about array index out of range. */
3107        if (rcu_num_lvls > RCU_NUM_LVLS)
3108                panic("rcu_init_one: rcu_num_lvls overflow");
3109
3110        /* Initialize the level-tracking arrays. */
3111
3112        for (i = 0; i < rcu_num_lvls; i++)
3113                rsp->levelcnt[i] = num_rcu_lvl[i];
3114        for (i = 1; i < rcu_num_lvls; i++)
3115                rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3116        rcu_init_levelspread(rsp);
3117
3118        /* Initialize the elements themselves, starting from the leaves. */
3119
3120        for (i = rcu_num_lvls - 1; i >= 0; i--) {
3121                cpustride *= rsp->levelspread[i];
3122                rnp = rsp->level[i];
3123                for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3124                        raw_spin_lock_init(&rnp->lock);
3125                        lockdep_set_class_and_name(&rnp->lock,
3126                                                   &rcu_node_class[i], buf[i]);
3127                        raw_spin_lock_init(&rnp->fqslock);
3128                        lockdep_set_class_and_name(&rnp->fqslock,
3129                                                   &rcu_fqs_class[i], fqs[i]);
3130                        rnp->gpnum = rsp->gpnum;
3131                        rnp->completed = rsp->completed;
3132                        rnp->qsmask = 0;
3133                        rnp->qsmaskinit = 0;
3134                        rnp->grplo = j * cpustride;
3135                        rnp->grphi = (j + 1) * cpustride - 1;
3136                        if (rnp->grphi >= NR_CPUS)
3137                                rnp->grphi = NR_CPUS - 1;
3138                        if (i == 0) {
3139                                rnp->grpnum = 0;
3140                                rnp->grpmask = 0;
3141                                rnp->parent = NULL;
3142                        } else {
3143                                rnp->grpnum = j % rsp->levelspread[i - 1];
3144                                rnp->grpmask = 1UL << rnp->grpnum;
3145                                rnp->parent = rsp->level[i - 1] +
3146                                              j / rsp->levelspread[i - 1];
3147                        }
3148                        rnp->level = i;
3149                        INIT_LIST_HEAD(&rnp->blkd_tasks);
3150                        rcu_init_one_nocb(rnp);
3151                }
3152        }
3153
3154        rsp->rda = rda;
3155        init_waitqueue_head(&rsp->gp_wq);
3156        init_irq_work(&rsp->wakeup_work, rsp_wakeup);
3157        rnp = rsp->level[rcu_num_lvls - 1];
3158        for_each_possible_cpu(i) {
3159                while (i > rnp->grphi)
3160                        rnp++;
3161                per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3162                rcu_boot_init_percpu_data(i, rsp);
3163        }
3164        list_add(&rsp->flavors, &rcu_struct_flavors);
3165}
3166
3167/*
3168 * Compute the rcu_node tree geometry from kernel parameters.  This cannot
3169 * replace the definitions in rcutree.h because those are needed to size
3170 * the ->node array in the rcu_state structure.
3171 */
3172static void __init rcu_init_geometry(void)
3173{
3174        ulong d;
3175        int i;
3176        int j;
3177        int n = nr_cpu_ids;
3178        int rcu_capacity[MAX_RCU_LVLS + 1];
3179
3180        /*
3181         * Initialize any unspecified boot parameters.
3182         * The default values of jiffies_till_first_fqs and
3183         * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3184         * value, which is a function of HZ, then adding one for each
3185         * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3186         */
3187        d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3188        if (jiffies_till_first_fqs == ULONG_MAX)
3189                jiffies_till_first_fqs = d;
3190        if (jiffies_till_next_fqs == ULONG_MAX)
3191                jiffies_till_next_fqs = d;
3192
3193        /* If the compile-time values are accurate, just leave. */
3194        if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3195            nr_cpu_ids == NR_CPUS)
3196                return;
3197
3198        /*
3199         * Compute number of nodes that can be handled an rcu_node tree
3200         * with the given number of levels.  Setting rcu_capacity[0] makes
3201         * some of the arithmetic easier.
3202         */
3203        rcu_capacity[0] = 1;
3204        rcu_capacity[1] = rcu_fanout_leaf;
3205        for (i = 2; i <= MAX_RCU_LVLS; i++)
3206                rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3207
3208        /*
3209         * The boot-time rcu_fanout_leaf parameter is only permitted
3210         * to increase the leaf-level fanout, not decrease it.  Of course,
3211         * the leaf-level fanout cannot exceed the number of bits in
3212         * the rcu_node masks.  Finally, the tree must be able to accommodate
3213         * the configured number of CPUs.  Complain and fall back to the
3214         * compile-time values if these limits are exceeded.
3215         */
3216        if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3217            rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3218            n > rcu_capacity[MAX_RCU_LVLS]) {
3219                WARN_ON(1);
3220                return;
3221        }
3222
3223        /* Calculate the number of rcu_nodes at each level of the tree. */
3224        for (i = 1; i <= MAX_RCU_LVLS; i++)
3225                if (n <= rcu_capacity[i]) {
3226                        for (j = 0; j <= i; j++)
3227                                num_rcu_lvl[j] =
3228                                        DIV_ROUND_UP(n, rcu_capacity[i - j]);
3229                        rcu_num_lvls = i;
3230                        for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3231                                num_rcu_lvl[j] = 0;
3232                        break;
3233                }
3234
3235        /* Calculate the total number of rcu_node structures. */
3236        rcu_num_nodes = 0;
3237        for (i = 0; i <= MAX_RCU_LVLS; i++)
3238                rcu_num_nodes += num_rcu_lvl[i];
3239        rcu_num_nodes -= n;
3240}
3241
3242void __init rcu_init(void)
3243{
3244        int cpu;
3245
3246        rcu_bootup_announce();
3247        rcu_init_geometry();
3248        rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3249        rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3250        __rcu_init_preempt();
3251        open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3252
3253        /*
3254         * We don't need protection against CPU-hotplug here because
3255         * this is called early in boot, before either interrupts
3256         * or the scheduler are operational.
3257         */
3258        cpu_notifier(rcu_cpu_notify, 0);
3259        for_each_online_cpu(cpu)
3260                rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3261}
3262
3263#include "rcutree_plugin.h"
3264
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.