linux/kernel/rcutree_plugin.h
<<
>>
Prefs
   1/*
   2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
   3 * Internal non-public definitions that provide either classic
   4 * or preemptible semantics.
   5 *
   6 * This program is free software; you can redistribute it and/or modify
   7 * it under the terms of the GNU General Public License as published by
   8 * the Free Software Foundation; either version 2 of the License, or
   9 * (at your option) any later version.
  10 *
  11 * This program is distributed in the hope that it will be useful,
  12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14 * GNU General Public License for more details.
  15 *
  16 * You should have received a copy of the GNU General Public License
  17 * along with this program; if not, write to the Free Software
  18 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  19 *
  20 * Copyright Red Hat, 2009
  21 * Copyright IBM Corporation, 2009
  22 *
  23 * Author: Ingo Molnar <mingo@elte.hu>
  24 *         Paul E. McKenney <paulmck@linux.vnet.ibm.com>
  25 */
  26
  27#include <linux/delay.h>
  28
  29#define RCU_KTHREAD_PRIO 1
  30
  31#ifdef CONFIG_RCU_BOOST
  32#define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO
  33#else
  34#define RCU_BOOST_PRIO RCU_KTHREAD_PRIO
  35#endif
  36
  37/*
  38 * Check the RCU kernel configuration parameters and print informative
  39 * messages about anything out of the ordinary.  If you like #ifdef, you
  40 * will love this function.
  41 */
  42static void __init rcu_bootup_announce_oddness(void)
  43{
  44#ifdef CONFIG_RCU_TRACE
  45        printk(KERN_INFO "\tRCU debugfs-based tracing is enabled.\n");
  46#endif
  47#if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
  48        printk(KERN_INFO "\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
  49               CONFIG_RCU_FANOUT);
  50#endif
  51#ifdef CONFIG_RCU_FANOUT_EXACT
  52        printk(KERN_INFO "\tHierarchical RCU autobalancing is disabled.\n");
  53#endif
  54#ifdef CONFIG_RCU_FAST_NO_HZ
  55        printk(KERN_INFO
  56               "\tRCU dyntick-idle grace-period acceleration is enabled.\n");
  57#endif
  58#ifdef CONFIG_PROVE_RCU
  59        printk(KERN_INFO "\tRCU lockdep checking is enabled.\n");
  60#endif
  61#ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
  62        printk(KERN_INFO "\tRCU torture testing starts during boot.\n");
  63#endif
  64#if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE)
  65        printk(KERN_INFO "\tDump stacks of tasks blocking RCU-preempt GP.\n");
  66#endif
  67#if defined(CONFIG_RCU_CPU_STALL_INFO)
  68        printk(KERN_INFO "\tAdditional per-CPU info printed with stalls.\n");
  69#endif
  70#if NUM_RCU_LVL_4 != 0
  71        printk(KERN_INFO "\tFour-level hierarchy is enabled.\n");
  72#endif
  73        if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF)
  74                printk(KERN_INFO "\tExperimental boot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
  75        if (nr_cpu_ids != NR_CPUS)
  76                printk(KERN_INFO "\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
  77}
  78
  79#ifdef CONFIG_TREE_PREEMPT_RCU
  80
  81struct rcu_state rcu_preempt_state =
  82        RCU_STATE_INITIALIZER(rcu_preempt, call_rcu);
  83DEFINE_PER_CPU(struct rcu_data, rcu_preempt_data);
  84static struct rcu_state *rcu_state = &rcu_preempt_state;
  85
  86static int rcu_preempted_readers_exp(struct rcu_node *rnp);
  87
  88/*
  89 * Tell them what RCU they are running.
  90 */
  91static void __init rcu_bootup_announce(void)
  92{
  93        printk(KERN_INFO "Preemptible hierarchical RCU implementation.\n");
  94        rcu_bootup_announce_oddness();
  95}
  96
  97/*
  98 * Return the number of RCU-preempt batches processed thus far
  99 * for debug and statistics.
 100 */
 101long rcu_batches_completed_preempt(void)
 102{
 103        return rcu_preempt_state.completed;
 104}
 105EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
 106
 107/*
 108 * Return the number of RCU batches processed thus far for debug & stats.
 109 */
 110long rcu_batches_completed(void)
 111{
 112        return rcu_batches_completed_preempt();
 113}
 114EXPORT_SYMBOL_GPL(rcu_batches_completed);
 115
 116/*
 117 * Force a quiescent state for preemptible RCU.
 118 */
 119void rcu_force_quiescent_state(void)
 120{
 121        force_quiescent_state(&rcu_preempt_state, 0);
 122}
 123EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
 124
 125/*
 126 * Record a preemptible-RCU quiescent state for the specified CPU.  Note
 127 * that this just means that the task currently running on the CPU is
 128 * not in a quiescent state.  There might be any number of tasks blocked
 129 * while in an RCU read-side critical section.
 130 *
 131 * Unlike the other rcu_*_qs() functions, callers to this function
 132 * must disable irqs in order to protect the assignment to
 133 * ->rcu_read_unlock_special.
 134 */
 135static void rcu_preempt_qs(int cpu)
 136{
 137        struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
 138
 139        rdp->passed_quiesce_gpnum = rdp->gpnum;
 140        barrier();
 141        if (rdp->passed_quiesce == 0)
 142                trace_rcu_grace_period("rcu_preempt", rdp->gpnum, "cpuqs");
 143        rdp->passed_quiesce = 1;
 144        current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
 145}
 146
 147/*
 148 * We have entered the scheduler, and the current task might soon be
 149 * context-switched away from.  If this task is in an RCU read-side
 150 * critical section, we will no longer be able to rely on the CPU to
 151 * record that fact, so we enqueue the task on the blkd_tasks list.
 152 * The task will dequeue itself when it exits the outermost enclosing
 153 * RCU read-side critical section.  Therefore, the current grace period
 154 * cannot be permitted to complete until the blkd_tasks list entries
 155 * predating the current grace period drain, in other words, until
 156 * rnp->gp_tasks becomes NULL.
 157 *
 158 * Caller must disable preemption.
 159 */
 160static void rcu_preempt_note_context_switch(int cpu)
 161{
 162        struct task_struct *t = current;
 163        unsigned long flags;
 164        struct rcu_data *rdp;
 165        struct rcu_node *rnp;
 166
 167        if (t->rcu_read_lock_nesting > 0 &&
 168            (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
 169
 170                /* Possibly blocking in an RCU read-side critical section. */
 171                rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu);
 172                rnp = rdp->mynode;
 173                raw_spin_lock_irqsave(&rnp->lock, flags);
 174                t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
 175                t->rcu_blocked_node = rnp;
 176
 177                /*
 178                 * If this CPU has already checked in, then this task
 179                 * will hold up the next grace period rather than the
 180                 * current grace period.  Queue the task accordingly.
 181                 * If the task is queued for the current grace period
 182                 * (i.e., this CPU has not yet passed through a quiescent
 183                 * state for the current grace period), then as long
 184                 * as that task remains queued, the current grace period
 185                 * cannot end.  Note that there is some uncertainty as
 186                 * to exactly when the current grace period started.
 187                 * We take a conservative approach, which can result
 188                 * in unnecessarily waiting on tasks that started very
 189                 * slightly after the current grace period began.  C'est
 190                 * la vie!!!
 191                 *
 192                 * But first, note that the current CPU must still be
 193                 * on line!
 194                 */
 195                WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
 196                WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
 197                if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
 198                        list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
 199                        rnp->gp_tasks = &t->rcu_node_entry;
 200#ifdef CONFIG_RCU_BOOST
 201                        if (rnp->boost_tasks != NULL)
 202                                rnp->boost_tasks = rnp->gp_tasks;
 203#endif /* #ifdef CONFIG_RCU_BOOST */
 204                } else {
 205                        list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
 206                        if (rnp->qsmask & rdp->grpmask)
 207                                rnp->gp_tasks = &t->rcu_node_entry;
 208                }
 209                trace_rcu_preempt_task(rdp->rsp->name,
 210                                       t->pid,
 211                                       (rnp->qsmask & rdp->grpmask)
 212                                       ? rnp->gpnum
 213                                       : rnp->gpnum + 1);
 214                raw_spin_unlock_irqrestore(&rnp->lock, flags);
 215        } else if (t->rcu_read_lock_nesting < 0 &&
 216                   t->rcu_read_unlock_special) {
 217
 218                /*
 219                 * Complete exit from RCU read-side critical section on
 220                 * behalf of preempted instance of __rcu_read_unlock().
 221                 */
 222                rcu_read_unlock_special(t);
 223        }
 224
 225        /*
 226         * Either we were not in an RCU read-side critical section to
 227         * begin with, or we have now recorded that critical section
 228         * globally.  Either way, we can now note a quiescent state
 229         * for this CPU.  Again, if we were in an RCU read-side critical
 230         * section, and if that critical section was blocking the current
 231         * grace period, then the fact that the task has been enqueued
 232         * means that we continue to block the current grace period.
 233         */
 234        local_irq_save(flags);
 235        rcu_preempt_qs(cpu);
 236        local_irq_restore(flags);
 237}
 238
 239/*
 240 * Check for preempted RCU readers blocking the current grace period
 241 * for the specified rcu_node structure.  If the caller needs a reliable
 242 * answer, it must hold the rcu_node's ->lock.
 243 */
 244static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
 245{
 246        return rnp->gp_tasks != NULL;
 247}
 248
 249/*
 250 * Record a quiescent state for all tasks that were previously queued
 251 * on the specified rcu_node structure and that were blocking the current
 252 * RCU grace period.  The caller must hold the specified rnp->lock with
 253 * irqs disabled, and this lock is released upon return, but irqs remain
 254 * disabled.
 255 */
 256static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
 257        __releases(rnp->lock)
 258{
 259        unsigned long mask;
 260        struct rcu_node *rnp_p;
 261
 262        if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
 263                raw_spin_unlock_irqrestore(&rnp->lock, flags);
 264                return;  /* Still need more quiescent states! */
 265        }
 266
 267        rnp_p = rnp->parent;
 268        if (rnp_p == NULL) {
 269                /*
 270                 * Either there is only one rcu_node in the tree,
 271                 * or tasks were kicked up to root rcu_node due to
 272                 * CPUs going offline.
 273                 */
 274                rcu_report_qs_rsp(&rcu_preempt_state, flags);
 275                return;
 276        }
 277
 278        /* Report up the rest of the hierarchy. */
 279        mask = rnp->grpmask;
 280        raw_spin_unlock(&rnp->lock);    /* irqs remain disabled. */
 281        raw_spin_lock(&rnp_p->lock);    /* irqs already disabled. */
 282        rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
 283}
 284
 285/*
 286 * Advance a ->blkd_tasks-list pointer to the next entry, instead
 287 * returning NULL if at the end of the list.
 288 */
 289static struct list_head *rcu_next_node_entry(struct task_struct *t,
 290                                             struct rcu_node *rnp)
 291{
 292        struct list_head *np;
 293
 294        np = t->rcu_node_entry.next;
 295        if (np == &rnp->blkd_tasks)
 296                np = NULL;
 297        return np;
 298}
 299
 300/*
 301 * Handle special cases during rcu_read_unlock(), such as needing to
 302 * notify RCU core processing or task having blocked during the RCU
 303 * read-side critical section.
 304 */
 305void rcu_read_unlock_special(struct task_struct *t)
 306{
 307        int empty;
 308        int empty_exp;
 309        int empty_exp_now;
 310        unsigned long flags;
 311        struct list_head *np;
 312#ifdef CONFIG_RCU_BOOST
 313        struct rt_mutex *rbmp = NULL;
 314#endif /* #ifdef CONFIG_RCU_BOOST */
 315        struct rcu_node *rnp;
 316        int special;
 317
 318        /* NMI handlers cannot block and cannot safely manipulate state. */
 319        if (in_nmi())
 320                return;
 321
 322        local_irq_save(flags);
 323
 324        /*
 325         * If RCU core is waiting for this CPU to exit critical section,
 326         * let it know that we have done so.
 327         */
 328        special = t->rcu_read_unlock_special;
 329        if (special & RCU_READ_UNLOCK_NEED_QS) {
 330                rcu_preempt_qs(smp_processor_id());
 331        }
 332
 333        /* Hardware IRQ handlers cannot block. */
 334        if (in_irq() || in_serving_softirq()) {
 335                local_irq_restore(flags);
 336                return;
 337        }
 338
 339        /* Clean up if blocked during RCU read-side critical section. */
 340        if (special & RCU_READ_UNLOCK_BLOCKED) {
 341                t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
 342
 343                /*
 344                 * Remove this task from the list it blocked on.  The
 345                 * task can migrate while we acquire the lock, but at
 346                 * most one time.  So at most two passes through loop.
 347                 */
 348                for (;;) {
 349                        rnp = t->rcu_blocked_node;
 350                        raw_spin_lock(&rnp->lock);  /* irqs already disabled. */
 351                        if (rnp == t->rcu_blocked_node)
 352                                break;
 353                        raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
 354                }
 355                empty = !rcu_preempt_blocked_readers_cgp(rnp);
 356                empty_exp = !rcu_preempted_readers_exp(rnp);
 357                smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
 358                np = rcu_next_node_entry(t, rnp);
 359                list_del_init(&t->rcu_node_entry);
 360                t->rcu_blocked_node = NULL;
 361                trace_rcu_unlock_preempted_task("rcu_preempt",
 362                                                rnp->gpnum, t->pid);
 363                if (&t->rcu_node_entry == rnp->gp_tasks)
 364                        rnp->gp_tasks = np;
 365                if (&t->rcu_node_entry == rnp->exp_tasks)
 366                        rnp->exp_tasks = np;
 367#ifdef CONFIG_RCU_BOOST
 368                if (&t->rcu_node_entry == rnp->boost_tasks)
 369                        rnp->boost_tasks = np;
 370                /* Snapshot/clear ->rcu_boost_mutex with rcu_node lock held. */
 371                if (t->rcu_boost_mutex) {
 372                        rbmp = t->rcu_boost_mutex;
 373                        t->rcu_boost_mutex = NULL;
 374                }
 375#endif /* #ifdef CONFIG_RCU_BOOST */
 376
 377                /*
 378                 * If this was the last task on the current list, and if
 379                 * we aren't waiting on any CPUs, report the quiescent state.
 380                 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
 381                 * so we must take a snapshot of the expedited state.
 382                 */
 383                empty_exp_now = !rcu_preempted_readers_exp(rnp);
 384                if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) {
 385                        trace_rcu_quiescent_state_report("preempt_rcu",
 386                                                         rnp->gpnum,
 387                                                         0, rnp->qsmask,
 388                                                         rnp->level,
 389                                                         rnp->grplo,
 390                                                         rnp->grphi,
 391                                                         !!rnp->gp_tasks);
 392                        rcu_report_unblock_qs_rnp(rnp, flags);
 393                } else {
 394                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 395                }
 396
 397#ifdef CONFIG_RCU_BOOST
 398                /* Unboost if we were boosted. */
 399                if (rbmp)
 400                        rt_mutex_unlock(rbmp);
 401#endif /* #ifdef CONFIG_RCU_BOOST */
 402
 403                /*
 404                 * If this was the last task on the expedited lists,
 405                 * then we need to report up the rcu_node hierarchy.
 406                 */
 407                if (!empty_exp && empty_exp_now)
 408                        rcu_report_exp_rnp(&rcu_preempt_state, rnp, true);
 409        } else {
 410                local_irq_restore(flags);
 411        }
 412}
 413
 414#ifdef CONFIG_RCU_CPU_STALL_VERBOSE
 415
 416/*
 417 * Dump detailed information for all tasks blocking the current RCU
 418 * grace period on the specified rcu_node structure.
 419 */
 420static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
 421{
 422        unsigned long flags;
 423        struct task_struct *t;
 424
 425        if (!rcu_preempt_blocked_readers_cgp(rnp))
 426                return;
 427        raw_spin_lock_irqsave(&rnp->lock, flags);
 428        t = list_entry(rnp->gp_tasks,
 429                       struct task_struct, rcu_node_entry);
 430        list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
 431                sched_show_task(t);
 432        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 433}
 434
 435/*
 436 * Dump detailed information for all tasks blocking the current RCU
 437 * grace period.
 438 */
 439static void rcu_print_detail_task_stall(struct rcu_state *rsp)
 440{
 441        struct rcu_node *rnp = rcu_get_root(rsp);
 442
 443        rcu_print_detail_task_stall_rnp(rnp);
 444        rcu_for_each_leaf_node(rsp, rnp)
 445                rcu_print_detail_task_stall_rnp(rnp);
 446}
 447
 448#else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
 449
 450static void rcu_print_detail_task_stall(struct rcu_state *rsp)
 451{
 452}
 453
 454#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
 455
 456#ifdef CONFIG_RCU_CPU_STALL_INFO
 457
 458static void rcu_print_task_stall_begin(struct rcu_node *rnp)
 459{
 460        printk(KERN_ERR "\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
 461               rnp->level, rnp->grplo, rnp->grphi);
 462}
 463
 464static void rcu_print_task_stall_end(void)
 465{
 466        printk(KERN_CONT "\n");
 467}
 468
 469#else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
 470
 471static void rcu_print_task_stall_begin(struct rcu_node *rnp)
 472{
 473}
 474
 475static void rcu_print_task_stall_end(void)
 476{
 477}
 478
 479#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
 480
 481/*
 482 * Scan the current list of tasks blocked within RCU read-side critical
 483 * sections, printing out the tid of each.
 484 */
 485static int rcu_print_task_stall(struct rcu_node *rnp)
 486{
 487        struct task_struct *t;
 488        int ndetected = 0;
 489
 490        if (!rcu_preempt_blocked_readers_cgp(rnp))
 491                return 0;
 492        rcu_print_task_stall_begin(rnp);
 493        t = list_entry(rnp->gp_tasks,
 494                       struct task_struct, rcu_node_entry);
 495        list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
 496                printk(KERN_CONT " P%d", t->pid);
 497                ndetected++;
 498        }
 499        rcu_print_task_stall_end();
 500        return ndetected;
 501}
 502
 503/*
 504 * Check that the list of blocked tasks for the newly completed grace
 505 * period is in fact empty.  It is a serious bug to complete a grace
 506 * period that still has RCU readers blocked!  This function must be
 507 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
 508 * must be held by the caller.
 509 *
 510 * Also, if there are blocked tasks on the list, they automatically
 511 * block the newly created grace period, so set up ->gp_tasks accordingly.
 512 */
 513static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
 514{
 515        WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
 516        if (!list_empty(&rnp->blkd_tasks))
 517                rnp->gp_tasks = rnp->blkd_tasks.next;
 518        WARN_ON_ONCE(rnp->qsmask);
 519}
 520
 521#ifdef CONFIG_HOTPLUG_CPU
 522
 523/*
 524 * Handle tasklist migration for case in which all CPUs covered by the
 525 * specified rcu_node have gone offline.  Move them up to the root
 526 * rcu_node.  The reason for not just moving them to the immediate
 527 * parent is to remove the need for rcu_read_unlock_special() to
 528 * make more than two attempts to acquire the target rcu_node's lock.
 529 * Returns true if there were tasks blocking the current RCU grace
 530 * period.
 531 *
 532 * Returns 1 if there was previously a task blocking the current grace
 533 * period on the specified rcu_node structure.
 534 *
 535 * The caller must hold rnp->lock with irqs disabled.
 536 */
 537static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
 538                                     struct rcu_node *rnp,
 539                                     struct rcu_data *rdp)
 540{
 541        struct list_head *lp;
 542        struct list_head *lp_root;
 543        int retval = 0;
 544        struct rcu_node *rnp_root = rcu_get_root(rsp);
 545        struct task_struct *t;
 546
 547        if (rnp == rnp_root) {
 548                WARN_ONCE(1, "Last CPU thought to be offlined?");
 549                return 0;  /* Shouldn't happen: at least one CPU online. */
 550        }
 551
 552        /* If we are on an internal node, complain bitterly. */
 553        WARN_ON_ONCE(rnp != rdp->mynode);
 554
 555        /*
 556         * Move tasks up to root rcu_node.  Don't try to get fancy for
 557         * this corner-case operation -- just put this node's tasks
 558         * at the head of the root node's list, and update the root node's
 559         * ->gp_tasks and ->exp_tasks pointers to those of this node's,
 560         * if non-NULL.  This might result in waiting for more tasks than
 561         * absolutely necessary, but this is a good performance/complexity
 562         * tradeoff.
 563         */
 564        if (rcu_preempt_blocked_readers_cgp(rnp) && rnp->qsmask == 0)
 565                retval |= RCU_OFL_TASKS_NORM_GP;
 566        if (rcu_preempted_readers_exp(rnp))
 567                retval |= RCU_OFL_TASKS_EXP_GP;
 568        lp = &rnp->blkd_tasks;
 569        lp_root = &rnp_root->blkd_tasks;
 570        while (!list_empty(lp)) {
 571                t = list_entry(lp->next, typeof(*t), rcu_node_entry);
 572                raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
 573                list_del(&t->rcu_node_entry);
 574                t->rcu_blocked_node = rnp_root;
 575                list_add(&t->rcu_node_entry, lp_root);
 576                if (&t->rcu_node_entry == rnp->gp_tasks)
 577                        rnp_root->gp_tasks = rnp->gp_tasks;
 578                if (&t->rcu_node_entry == rnp->exp_tasks)
 579                        rnp_root->exp_tasks = rnp->exp_tasks;
 580#ifdef CONFIG_RCU_BOOST
 581                if (&t->rcu_node_entry == rnp->boost_tasks)
 582                        rnp_root->boost_tasks = rnp->boost_tasks;
 583#endif /* #ifdef CONFIG_RCU_BOOST */
 584                raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
 585        }
 586
 587#ifdef CONFIG_RCU_BOOST
 588        /* In case root is being boosted and leaf is not. */
 589        raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
 590        if (rnp_root->boost_tasks != NULL &&
 591            rnp_root->boost_tasks != rnp_root->gp_tasks)
 592                rnp_root->boost_tasks = rnp_root->gp_tasks;
 593        raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
 594#endif /* #ifdef CONFIG_RCU_BOOST */
 595
 596        rnp->gp_tasks = NULL;
 597        rnp->exp_tasks = NULL;
 598        return retval;
 599}
 600
 601#endif /* #ifdef CONFIG_HOTPLUG_CPU */
 602
 603/*
 604 * Check for a quiescent state from the current CPU.  When a task blocks,
 605 * the task is recorded in the corresponding CPU's rcu_node structure,
 606 * which is checked elsewhere.
 607 *
 608 * Caller must disable hard irqs.
 609 */
 610static void rcu_preempt_check_callbacks(int cpu)
 611{
 612        struct task_struct *t = current;
 613
 614        if (t->rcu_read_lock_nesting == 0) {
 615                rcu_preempt_qs(cpu);
 616                return;
 617        }
 618        if (t->rcu_read_lock_nesting > 0 &&
 619            per_cpu(rcu_preempt_data, cpu).qs_pending)
 620                t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
 621}
 622
 623#ifdef CONFIG_RCU_BOOST
 624
 625static void rcu_preempt_do_callbacks(void)
 626{
 627        rcu_do_batch(&rcu_preempt_state, &__get_cpu_var(rcu_preempt_data));
 628}
 629
 630#endif /* #ifdef CONFIG_RCU_BOOST */
 631
 632/*
 633 * Queue a preemptible-RCU callback for invocation after a grace period.
 634 */
 635void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
 636{
 637        __call_rcu(head, func, &rcu_preempt_state, 0);
 638}
 639EXPORT_SYMBOL_GPL(call_rcu);
 640
 641/*
 642 * Queue an RCU callback for lazy invocation after a grace period.
 643 * This will likely be later named something like "call_rcu_lazy()",
 644 * but this change will require some way of tagging the lazy RCU
 645 * callbacks in the list of pending callbacks.  Until then, this
 646 * function may only be called from __kfree_rcu().
 647 */
 648void kfree_call_rcu(struct rcu_head *head,
 649                    void (*func)(struct rcu_head *rcu))
 650{
 651        __call_rcu(head, func, &rcu_preempt_state, 1);
 652}
 653EXPORT_SYMBOL_GPL(kfree_call_rcu);
 654
 655/**
 656 * synchronize_rcu - wait until a grace period has elapsed.
 657 *
 658 * Control will return to the caller some time after a full grace
 659 * period has elapsed, in other words after all currently executing RCU
 660 * read-side critical sections have completed.  Note, however, that
 661 * upon return from synchronize_rcu(), the caller might well be executing
 662 * concurrently with new RCU read-side critical sections that began while
 663 * synchronize_rcu() was waiting.  RCU read-side critical sections are
 664 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
 665 */
 666void synchronize_rcu(void)
 667{
 668        rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
 669                           !lock_is_held(&rcu_lock_map) &&
 670                           !lock_is_held(&rcu_sched_lock_map),
 671                           "Illegal synchronize_rcu() in RCU read-side critical section");
 672        if (!rcu_scheduler_active)
 673                return;
 674        wait_rcu_gp(call_rcu);
 675}
 676EXPORT_SYMBOL_GPL(synchronize_rcu);
 677
 678static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
 679static long sync_rcu_preempt_exp_count;
 680static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
 681
 682/*
 683 * Return non-zero if there are any tasks in RCU read-side critical
 684 * sections blocking the current preemptible-RCU expedited grace period.
 685 * If there is no preemptible-RCU expedited grace period currently in
 686 * progress, returns zero unconditionally.
 687 */
 688static int rcu_preempted_readers_exp(struct rcu_node *rnp)
 689{
 690        return rnp->exp_tasks != NULL;
 691}
 692
 693/*
 694 * return non-zero if there is no RCU expedited grace period in progress
 695 * for the specified rcu_node structure, in other words, if all CPUs and
 696 * tasks covered by the specified rcu_node structure have done their bit
 697 * for the current expedited grace period.  Works only for preemptible
 698 * RCU -- other RCU implementation use other means.
 699 *
 700 * Caller must hold sync_rcu_preempt_exp_mutex.
 701 */
 702static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
 703{
 704        return !rcu_preempted_readers_exp(rnp) &&
 705               ACCESS_ONCE(rnp->expmask) == 0;
 706}
 707
 708/*
 709 * Report the exit from RCU read-side critical section for the last task
 710 * that queued itself during or before the current expedited preemptible-RCU
 711 * grace period.  This event is reported either to the rcu_node structure on
 712 * which the task was queued or to one of that rcu_node structure's ancestors,
 713 * recursively up the tree.  (Calm down, calm down, we do the recursion
 714 * iteratively!)
 715 *
 716 * Most callers will set the "wake" flag, but the task initiating the
 717 * expedited grace period need not wake itself.
 718 *
 719 * Caller must hold sync_rcu_preempt_exp_mutex.
 720 */
 721static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
 722                               bool wake)
 723{
 724        unsigned long flags;
 725        unsigned long mask;
 726
 727        raw_spin_lock_irqsave(&rnp->lock, flags);
 728        for (;;) {
 729                if (!sync_rcu_preempt_exp_done(rnp)) {
 730                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 731                        break;
 732                }
 733                if (rnp->parent == NULL) {
 734                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 735                        if (wake)
 736                                wake_up(&sync_rcu_preempt_exp_wq);
 737                        break;
 738                }
 739                mask = rnp->grpmask;
 740                raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
 741                rnp = rnp->parent;
 742                raw_spin_lock(&rnp->lock); /* irqs already disabled */
 743                rnp->expmask &= ~mask;
 744        }
 745}
 746
 747/*
 748 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
 749 * grace period for the specified rcu_node structure.  If there are no such
 750 * tasks, report it up the rcu_node hierarchy.
 751 *
 752 * Caller must hold sync_rcu_preempt_exp_mutex and rsp->onofflock.
 753 */
 754static void
 755sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
 756{
 757        unsigned long flags;
 758        int must_wait = 0;
 759
 760        raw_spin_lock_irqsave(&rnp->lock, flags);
 761        if (list_empty(&rnp->blkd_tasks)) {
 762                raw_spin_unlock_irqrestore(&rnp->lock, flags);
 763        } else {
 764                rnp->exp_tasks = rnp->blkd_tasks.next;
 765                rcu_initiate_boost(rnp, flags);  /* releases rnp->lock */
 766                must_wait = 1;
 767        }
 768        if (!must_wait)
 769                rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */
 770}
 771
 772/**
 773 * synchronize_rcu_expedited - Brute-force RCU grace period
 774 *
 775 * Wait for an RCU-preempt grace period, but expedite it.  The basic
 776 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
 777 * the ->blkd_tasks lists and wait for this list to drain.  This consumes
 778 * significant time on all CPUs and is unfriendly to real-time workloads,
 779 * so is thus not recommended for any sort of common-case code.
 780 * In fact, if you are using synchronize_rcu_expedited() in a loop,
 781 * please restructure your code to batch your updates, and then Use a
 782 * single synchronize_rcu() instead.
 783 *
 784 * Note that it is illegal to call this function while holding any lock
 785 * that is acquired by a CPU-hotplug notifier.  And yes, it is also illegal
 786 * to call this function from a CPU-hotplug notifier.  Failing to observe
 787 * these restriction will result in deadlock.
 788 */
 789void synchronize_rcu_expedited(void)
 790{
 791        unsigned long flags;
 792        struct rcu_node *rnp;
 793        struct rcu_state *rsp = &rcu_preempt_state;
 794        long snap;
 795        int trycount = 0;
 796
 797        smp_mb(); /* Caller's modifications seen first by other CPUs. */
 798        snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
 799        smp_mb(); /* Above access cannot bleed into critical section. */
 800
 801        /*
 802         * Acquire lock, falling back to synchronize_rcu() if too many
 803         * lock-acquisition failures.  Of course, if someone does the
 804         * expedited grace period for us, just leave.
 805         */
 806        while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
 807                if (trycount++ < 10) {
 808                        udelay(trycount * num_online_cpus());
 809                } else {
 810                        synchronize_rcu();
 811                        return;
 812                }
 813                if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0)
 814                        goto mb_ret; /* Others did our work for us. */
 815        }
 816        if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0)
 817                goto unlock_mb_ret; /* Others did our work for us. */
 818
 819        /* force all RCU readers onto ->blkd_tasks lists. */
 820        synchronize_sched_expedited();
 821
 822        raw_spin_lock_irqsave(&rsp->onofflock, flags);
 823
 824        /* Initialize ->expmask for all non-leaf rcu_node structures. */
 825        rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
 826                raw_spin_lock(&rnp->lock); /* irqs already disabled. */
 827                rnp->expmask = rnp->qsmaskinit;
 828                raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
 829        }
 830
 831        /* Snapshot current state of ->blkd_tasks lists. */
 832        rcu_for_each_leaf_node(rsp, rnp)
 833                sync_rcu_preempt_exp_init(rsp, rnp);
 834        if (NUM_RCU_NODES > 1)
 835                sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
 836
 837        raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
 838
 839        /* Wait for snapshotted ->blkd_tasks lists to drain. */
 840        rnp = rcu_get_root(rsp);
 841        wait_event(sync_rcu_preempt_exp_wq,
 842                   sync_rcu_preempt_exp_done(rnp));
 843
 844        /* Clean up and exit. */
 845        smp_mb(); /* ensure expedited GP seen before counter increment. */
 846        ACCESS_ONCE(sync_rcu_preempt_exp_count)++;
 847unlock_mb_ret:
 848        mutex_unlock(&sync_rcu_preempt_exp_mutex);
 849mb_ret:
 850        smp_mb(); /* ensure subsequent action seen after grace period. */
 851}
 852EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
 853
 854/**
 855 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
 856 */
 857void rcu_barrier(void)
 858{
 859        _rcu_barrier(&rcu_preempt_state);
 860}
 861EXPORT_SYMBOL_GPL(rcu_barrier);
 862
 863/*
 864 * Initialize preemptible RCU's state structures.
 865 */
 866static void __init __rcu_init_preempt(void)
 867{
 868        rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
 869}
 870
 871#else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
 872
 873static struct rcu_state *rcu_state = &rcu_sched_state;
 874
 875/*
 876 * Tell them what RCU they are running.
 877 */
 878static void __init rcu_bootup_announce(void)
 879{
 880        printk(KERN_INFO "Hierarchical RCU implementation.\n");
 881        rcu_bootup_announce_oddness();
 882}
 883
 884/*
 885 * Return the number of RCU batches processed thus far for debug & stats.
 886 */
 887long rcu_batches_completed(void)
 888{
 889        return rcu_batches_completed_sched();
 890}
 891EXPORT_SYMBOL_GPL(rcu_batches_completed);
 892
 893/*
 894 * Force a quiescent state for RCU, which, because there is no preemptible
 895 * RCU, becomes the same as rcu-sched.
 896 */
 897void rcu_force_quiescent_state(void)
 898{
 899        rcu_sched_force_quiescent_state();
 900}
 901EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
 902
 903/*
 904 * Because preemptible RCU does not exist, we never have to check for
 905 * CPUs being in quiescent states.
 906 */
 907static void rcu_preempt_note_context_switch(int cpu)
 908{
 909}
 910
 911/*
 912 * Because preemptible RCU does not exist, there are never any preempted
 913 * RCU readers.
 914 */
 915static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
 916{
 917        return 0;
 918}
 919
 920#ifdef CONFIG_HOTPLUG_CPU
 921
 922/* Because preemptible RCU does not exist, no quieting of tasks. */
 923static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
 924{
 925        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 926}
 927
 928#endif /* #ifdef CONFIG_HOTPLUG_CPU */
 929
 930/*
 931 * Because preemptible RCU does not exist, we never have to check for
 932 * tasks blocked within RCU read-side critical sections.
 933 */
 934static void rcu_print_detail_task_stall(struct rcu_state *rsp)
 935{
 936}
 937
 938/*
 939 * Because preemptible RCU does not exist, we never have to check for
 940 * tasks blocked within RCU read-side critical sections.
 941 */
 942static int rcu_print_task_stall(struct rcu_node *rnp)
 943{
 944        return 0;
 945}
 946
 947/*
 948 * Because there is no preemptible RCU, there can be no readers blocked,
 949 * so there is no need to check for blocked tasks.  So check only for
 950 * bogus qsmask values.
 951 */
 952static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
 953{
 954        WARN_ON_ONCE(rnp->qsmask);
 955}
 956
 957#ifdef CONFIG_HOTPLUG_CPU
 958
 959/*
 960 * Because preemptible RCU does not exist, it never needs to migrate
 961 * tasks that were blocked within RCU read-side critical sections, and
 962 * such non-existent tasks cannot possibly have been blocking the current
 963 * grace period.
 964 */
 965static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
 966                                     struct rcu_node *rnp,
 967                                     struct rcu_data *rdp)
 968{
 969        return 0;
 970}
 971
 972#endif /* #ifdef CONFIG_HOTPLUG_CPU */
 973
 974/*
 975 * Because preemptible RCU does not exist, it never has any callbacks
 976 * to check.
 977 */
 978static void rcu_preempt_check_callbacks(int cpu)
 979{
 980}
 981
 982/*
 983 * Queue an RCU callback for lazy invocation after a grace period.
 984 * This will likely be later named something like "call_rcu_lazy()",
 985 * but this change will require some way of tagging the lazy RCU
 986 * callbacks in the list of pending callbacks.  Until then, this
 987 * function may only be called from __kfree_rcu().
 988 *
 989 * Because there is no preemptible RCU, we use RCU-sched instead.
 990 */
 991void kfree_call_rcu(struct rcu_head *head,
 992                    void (*func)(struct rcu_head *rcu))
 993{
 994        __call_rcu(head, func, &rcu_sched_state, 1);
 995}
 996EXPORT_SYMBOL_GPL(kfree_call_rcu);
 997
 998/*
 999 * Wait for an rcu-preempt grace period, but make it happen quickly.
1000 * But because preemptible RCU does not exist, map to rcu-sched.
1001 */
1002void synchronize_rcu_expedited(void)
1003{
1004        synchronize_sched_expedited();
1005}
1006EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1007
1008#ifdef CONFIG_HOTPLUG_CPU
1009
1010/*
1011 * Because preemptible RCU does not exist, there is never any need to
1012 * report on tasks preempted in RCU read-side critical sections during
1013 * expedited RCU grace periods.
1014 */
1015static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
1016                               bool wake)
1017{
1018}
1019
1020#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1021
1022/*
1023 * Because preemptible RCU does not exist, rcu_barrier() is just
1024 * another name for rcu_barrier_sched().
1025 */
1026void rcu_barrier(void)
1027{
1028        rcu_barrier_sched();
1029}
1030EXPORT_SYMBOL_GPL(rcu_barrier);
1031
1032/*
1033 * Because preemptible RCU does not exist, it need not be initialized.
1034 */
1035static void __init __rcu_init_preempt(void)
1036{
1037}
1038
1039#endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1040
1041#ifdef CONFIG_RCU_BOOST
1042
1043#include "rtmutex_common.h"
1044
1045#ifdef CONFIG_RCU_TRACE
1046
1047static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1048{
1049        if (list_empty(&rnp->blkd_tasks))
1050                rnp->n_balk_blkd_tasks++;
1051        else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
1052                rnp->n_balk_exp_gp_tasks++;
1053        else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
1054                rnp->n_balk_boost_tasks++;
1055        else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
1056                rnp->n_balk_notblocked++;
1057        else if (rnp->gp_tasks != NULL &&
1058                 ULONG_CMP_LT(jiffies, rnp->boost_time))
1059                rnp->n_balk_notyet++;
1060        else
1061                rnp->n_balk_nos++;
1062}
1063
1064#else /* #ifdef CONFIG_RCU_TRACE */
1065
1066static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1067{
1068}
1069
1070#endif /* #else #ifdef CONFIG_RCU_TRACE */
1071
1072/*
1073 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1074 * or ->boost_tasks, advancing the pointer to the next task in the
1075 * ->blkd_tasks list.
1076 *
1077 * Note that irqs must be enabled: boosting the task can block.
1078 * Returns 1 if there are more tasks needing to be boosted.
1079 */
1080static int rcu_boost(struct rcu_node *rnp)
1081{
1082        unsigned long flags;
1083        struct rt_mutex mtx;
1084        struct task_struct *t;
1085        struct list_head *tb;
1086
1087        if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL)
1088                return 0;  /* Nothing left to boost. */
1089
1090        raw_spin_lock_irqsave(&rnp->lock, flags);
1091
1092        /*
1093         * Recheck under the lock: all tasks in need of boosting
1094         * might exit their RCU read-side critical sections on their own.
1095         */
1096        if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1097                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1098                return 0;
1099        }
1100
1101        /*
1102         * Preferentially boost tasks blocking expedited grace periods.
1103         * This cannot starve the normal grace periods because a second
1104         * expedited grace period must boost all blocked tasks, including
1105         * those blocking the pre-existing normal grace period.
1106         */
1107        if (rnp->exp_tasks != NULL) {
1108                tb = rnp->exp_tasks;
1109                rnp->n_exp_boosts++;
1110        } else {
1111                tb = rnp->boost_tasks;
1112                rnp->n_normal_boosts++;
1113        }
1114        rnp->n_tasks_boosted++;
1115
1116        /*
1117         * We boost task t by manufacturing an rt_mutex that appears to
1118         * be held by task t.  We leave a pointer to that rt_mutex where
1119         * task t can find it, and task t will release the mutex when it
1120         * exits its outermost RCU read-side critical section.  Then
1121         * simply acquiring this artificial rt_mutex will boost task
1122         * t's priority.  (Thanks to tglx for suggesting this approach!)
1123         *
1124         * Note that task t must acquire rnp->lock to remove itself from
1125         * the ->blkd_tasks list, which it will do from exit() if from
1126         * nowhere else.  We therefore are guaranteed that task t will
1127         * stay around at least until we drop rnp->lock.  Note that
1128         * rnp->lock also resolves races between our priority boosting
1129         * and task t's exiting its outermost RCU read-side critical
1130         * section.
1131         */
1132        t = container_of(tb, struct task_struct, rcu_node_entry);
1133        rt_mutex_init_proxy_locked(&mtx, t);
1134        t->rcu_boost_mutex = &mtx;
1135        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1136        rt_mutex_lock(&mtx);  /* Side effect: boosts task t's priority. */
1137        rt_mutex_unlock(&mtx);  /* Keep lockdep happy. */
1138
1139        return ACCESS_ONCE(rnp->exp_tasks) != NULL ||
1140               ACCESS_ONCE(rnp->boost_tasks) != NULL;
1141}
1142
1143/*
1144 * Timer handler to initiate waking up of boost kthreads that
1145 * have yielded the CPU due to excessive numbers of tasks to
1146 * boost.  We wake up the per-rcu_node kthread, which in turn
1147 * will wake up the booster kthread.
1148 */
1149static void rcu_boost_kthread_timer(unsigned long arg)
1150{
1151        invoke_rcu_node_kthread((struct rcu_node *)arg);
1152}
1153
1154/*
1155 * Priority-boosting kthread.  One per leaf rcu_node and one for the
1156 * root rcu_node.
1157 */
1158static int rcu_boost_kthread(void *arg)
1159{
1160        struct rcu_node *rnp = (struct rcu_node *)arg;
1161        int spincnt = 0;
1162        int more2boost;
1163
1164        trace_rcu_utilization("Start boost kthread@init");
1165        for (;;) {
1166                rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1167                trace_rcu_utilization("End boost kthread@rcu_wait");
1168                rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1169                trace_rcu_utilization("Start boost kthread@rcu_wait");
1170                rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1171                more2boost = rcu_boost(rnp);
1172                if (more2boost)
1173                        spincnt++;
1174                else
1175                        spincnt = 0;
1176                if (spincnt > 10) {
1177                        trace_rcu_utilization("End boost kthread@rcu_yield");
1178                        rcu_yield(rcu_boost_kthread_timer, (unsigned long)rnp);
1179                        trace_rcu_utilization("Start boost kthread@rcu_yield");
1180                        spincnt = 0;
1181                }
1182        }
1183        /* NOTREACHED */
1184        trace_rcu_utilization("End boost kthread@notreached");
1185        return 0;
1186}
1187
1188/*
1189 * Check to see if it is time to start boosting RCU readers that are
1190 * blocking the current grace period, and, if so, tell the per-rcu_node
1191 * kthread to start boosting them.  If there is an expedited grace
1192 * period in progress, it is always time to boost.
1193 *
1194 * The caller must hold rnp->lock, which this function releases,
1195 * but irqs remain disabled.  The ->boost_kthread_task is immortal,
1196 * so we don't need to worry about it going away.
1197 */
1198static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1199{
1200        struct task_struct *t;
1201
1202        if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1203                rnp->n_balk_exp_gp_tasks++;
1204                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1205                return;
1206        }
1207        if (rnp->exp_tasks != NULL ||
1208            (rnp->gp_tasks != NULL &&
1209             rnp->boost_tasks == NULL &&
1210             rnp->qsmask == 0 &&
1211             ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1212                if (rnp->exp_tasks == NULL)
1213                        rnp->boost_tasks = rnp->gp_tasks;
1214                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1215                t = rnp->boost_kthread_task;
1216                if (t != NULL)
1217                        wake_up_process(t);
1218        } else {
1219                rcu_initiate_boost_trace(rnp);
1220                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1221        }
1222}
1223
1224/*
1225 * Wake up the per-CPU kthread to invoke RCU callbacks.
1226 */
1227static void invoke_rcu_callbacks_kthread(void)
1228{
1229        unsigned long flags;
1230
1231        local_irq_save(flags);
1232        __this_cpu_write(rcu_cpu_has_work, 1);
1233        if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1234            current != __this_cpu_read(rcu_cpu_kthread_task))
1235                wake_up_process(__this_cpu_read(rcu_cpu_kthread_task));
1236        local_irq_restore(flags);
1237}
1238
1239/*
1240 * Is the current CPU running the RCU-callbacks kthread?
1241 * Caller must have preemption disabled.
1242 */
1243static bool rcu_is_callbacks_kthread(void)
1244{
1245        return __get_cpu_var(rcu_cpu_kthread_task) == current;
1246}
1247
1248/*
1249 * Set the affinity of the boost kthread.  The CPU-hotplug locks are
1250 * held, so no one should be messing with the existence of the boost
1251 * kthread.
1252 */
1253static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp,
1254                                          cpumask_var_t cm)
1255{
1256        struct task_struct *t;
1257
1258        t = rnp->boost_kthread_task;
1259        if (t != NULL)
1260                set_cpus_allowed_ptr(rnp->boost_kthread_task, cm);
1261}
1262
1263#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1264
1265/*
1266 * Do priority-boost accounting for the start of a new grace period.
1267 */
1268static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1269{
1270        rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1271}
1272
1273/*
1274 * Create an RCU-boost kthread for the specified node if one does not
1275 * already exist.  We only create this kthread for preemptible RCU.
1276 * Returns zero if all is well, a negated errno otherwise.
1277 */
1278static int __cpuinit rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1279                                                 struct rcu_node *rnp,
1280                                                 int rnp_index)
1281{
1282        unsigned long flags;
1283        struct sched_param sp;
1284        struct task_struct *t;
1285
1286        if (&rcu_preempt_state != rsp)
1287                return 0;
1288        rsp->boost = 1;
1289        if (rnp->boost_kthread_task != NULL)
1290                return 0;
1291        t = kthread_create(rcu_boost_kthread, (void *)rnp,
1292                           "rcub/%d", rnp_index);
1293        if (IS_ERR(t))
1294                return PTR_ERR(t);
1295        raw_spin_lock_irqsave(&rnp->lock, flags);
1296        rnp->boost_kthread_task = t;
1297        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1298        sp.sched_priority = RCU_BOOST_PRIO;
1299        sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1300        wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1301        return 0;
1302}
1303
1304#ifdef CONFIG_HOTPLUG_CPU
1305
1306/*
1307 * Stop the RCU's per-CPU kthread when its CPU goes offline,.
1308 */
1309static void rcu_stop_cpu_kthread(int cpu)
1310{
1311        struct task_struct *t;
1312
1313        /* Stop the CPU's kthread. */
1314        t = per_cpu(rcu_cpu_kthread_task, cpu);
1315        if (t != NULL) {
1316                per_cpu(rcu_cpu_kthread_task, cpu) = NULL;
1317                kthread_stop(t);
1318        }
1319}
1320
1321#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1322
1323static void rcu_kthread_do_work(void)
1324{
1325        rcu_do_batch(&rcu_sched_state, &__get_cpu_var(rcu_sched_data));
1326        rcu_do_batch(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1327        rcu_preempt_do_callbacks();
1328}
1329
1330/*
1331 * Wake up the specified per-rcu_node-structure kthread.
1332 * Because the per-rcu_node kthreads are immortal, we don't need
1333 * to do anything to keep them alive.
1334 */
1335static void invoke_rcu_node_kthread(struct rcu_node *rnp)
1336{
1337        struct task_struct *t;
1338
1339        t = rnp->node_kthread_task;
1340        if (t != NULL)
1341                wake_up_process(t);
1342}
1343
1344/*
1345 * Set the specified CPU's kthread to run RT or not, as specified by
1346 * the to_rt argument.  The CPU-hotplug locks are held, so the task
1347 * is not going away.
1348 */
1349static void rcu_cpu_kthread_setrt(int cpu, int to_rt)
1350{
1351        int policy;
1352        struct sched_param sp;
1353        struct task_struct *t;
1354
1355        t = per_cpu(rcu_cpu_kthread_task, cpu);
1356        if (t == NULL)
1357                return;
1358        if (to_rt) {
1359                policy = SCHED_FIFO;
1360                sp.sched_priority = RCU_KTHREAD_PRIO;
1361        } else {
1362                policy = SCHED_NORMAL;
1363                sp.sched_priority = 0;
1364        }
1365        sched_setscheduler_nocheck(t, policy, &sp);
1366}
1367
1368/*
1369 * Timer handler to initiate the waking up of per-CPU kthreads that
1370 * have yielded the CPU due to excess numbers of RCU callbacks.
1371 * We wake up the per-rcu_node kthread, which in turn will wake up
1372 * the booster kthread.
1373 */
1374static void rcu_cpu_kthread_timer(unsigned long arg)
1375{
1376        struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, arg);
1377        struct rcu_node *rnp = rdp->mynode;
1378
1379        atomic_or(rdp->grpmask, &rnp->wakemask);
1380        invoke_rcu_node_kthread(rnp);
1381}
1382
1383/*
1384 * Drop to non-real-time priority and yield, but only after posting a
1385 * timer that will cause us to regain our real-time priority if we
1386 * remain preempted.  Either way, we restore our real-time priority
1387 * before returning.
1388 */
1389static void rcu_yield(void (*f)(unsigned long), unsigned long arg)
1390{
1391        struct sched_param sp;
1392        struct timer_list yield_timer;
1393        int prio = current->rt_priority;
1394
1395        setup_timer_on_stack(&yield_timer, f, arg);
1396        mod_timer(&yield_timer, jiffies + 2);
1397        sp.sched_priority = 0;
1398        sched_setscheduler_nocheck(current, SCHED_NORMAL, &sp);
1399        set_user_nice(current, 19);
1400        schedule();
1401        set_user_nice(current, 0);
1402        sp.sched_priority = prio;
1403        sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1404        del_timer(&yield_timer);
1405}
1406
1407/*
1408 * Handle cases where the rcu_cpu_kthread() ends up on the wrong CPU.
1409 * This can happen while the corresponding CPU is either coming online
1410 * or going offline.  We cannot wait until the CPU is fully online
1411 * before starting the kthread, because the various notifier functions
1412 * can wait for RCU grace periods.  So we park rcu_cpu_kthread() until
1413 * the corresponding CPU is online.
1414 *
1415 * Return 1 if the kthread needs to stop, 0 otherwise.
1416 *
1417 * Caller must disable bh.  This function can momentarily enable it.
1418 */
1419static int rcu_cpu_kthread_should_stop(int cpu)
1420{
1421        while (cpu_is_offline(cpu) ||
1422               !cpumask_equal(&current->cpus_allowed, cpumask_of(cpu)) ||
1423               smp_processor_id() != cpu) {
1424                if (kthread_should_stop())
1425                        return 1;
1426                per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1427                per_cpu(rcu_cpu_kthread_cpu, cpu) = raw_smp_processor_id();
1428                local_bh_enable();
1429                schedule_timeout_uninterruptible(1);
1430                if (!cpumask_equal(&current->cpus_allowed, cpumask_of(cpu)))
1431                        set_cpus_allowed_ptr(current, cpumask_of(cpu));
1432                local_bh_disable();
1433        }
1434        per_cpu(rcu_cpu_kthread_cpu, cpu) = cpu;
1435        return 0;
1436}
1437
1438/*
1439 * Per-CPU kernel thread that invokes RCU callbacks.  This replaces the
1440 * RCU softirq used in flavors and configurations of RCU that do not
1441 * support RCU priority boosting.
1442 */
1443static int rcu_cpu_kthread(void *arg)
1444{
1445        int cpu = (int)(long)arg;
1446        unsigned long flags;
1447        int spincnt = 0;
1448        unsigned int *statusp = &per_cpu(rcu_cpu_kthread_status, cpu);
1449        char work;
1450        char *workp = &per_cpu(rcu_cpu_has_work, cpu);
1451
1452        trace_rcu_utilization("Start CPU kthread@init");
1453        for (;;) {
1454                *statusp = RCU_KTHREAD_WAITING;
1455                trace_rcu_utilization("End CPU kthread@rcu_wait");
1456                rcu_wait(*workp != 0 || kthread_should_stop());
1457                trace_rcu_utilization("Start CPU kthread@rcu_wait");
1458                local_bh_disable();
1459                if (rcu_cpu_kthread_should_stop(cpu)) {
1460                        local_bh_enable();
1461                        break;
1462                }
1463                *statusp = RCU_KTHREAD_RUNNING;
1464                per_cpu(rcu_cpu_kthread_loops, cpu)++;
1465                local_irq_save(flags);
1466                work = *workp;
1467                *workp = 0;
1468                local_irq_restore(flags);
1469                if (work)
1470                        rcu_kthread_do_work();
1471                local_bh_enable();
1472                if (*workp != 0)
1473                        spincnt++;
1474                else
1475                        spincnt = 0;
1476                if (spincnt > 10) {
1477                        *statusp = RCU_KTHREAD_YIELDING;
1478                        trace_rcu_utilization("End CPU kthread@rcu_yield");
1479                        rcu_yield(rcu_cpu_kthread_timer, (unsigned long)cpu);
1480                        trace_rcu_utilization("Start CPU kthread@rcu_yield");
1481                        spincnt = 0;
1482                }
1483        }
1484        *statusp = RCU_KTHREAD_STOPPED;
1485        trace_rcu_utilization("End CPU kthread@term");
1486        return 0;
1487}
1488
1489/*
1490 * Spawn a per-CPU kthread, setting up affinity and priority.
1491 * Because the CPU hotplug lock is held, no other CPU will be attempting
1492 * to manipulate rcu_cpu_kthread_task.  There might be another CPU
1493 * attempting to access it during boot, but the locking in kthread_bind()
1494 * will enforce sufficient ordering.
1495 *
1496 * Please note that we cannot simply refuse to wake up the per-CPU
1497 * kthread because kthreads are created in TASK_UNINTERRUPTIBLE state,
1498 * which can result in softlockup complaints if the task ends up being
1499 * idle for more than a couple of minutes.
1500 *
1501 * However, please note also that we cannot bind the per-CPU kthread to its
1502 * CPU until that CPU is fully online.  We also cannot wait until the
1503 * CPU is fully online before we create its per-CPU kthread, as this would
1504 * deadlock the system when CPU notifiers tried waiting for grace
1505 * periods.  So we bind the per-CPU kthread to its CPU only if the CPU
1506 * is online.  If its CPU is not yet fully online, then the code in
1507 * rcu_cpu_kthread() will wait until it is fully online, and then do
1508 * the binding.
1509 */
1510static int __cpuinit rcu_spawn_one_cpu_kthread(int cpu)
1511{
1512        struct sched_param sp;
1513        struct task_struct *t;
1514
1515        if (!rcu_scheduler_fully_active ||
1516            per_cpu(rcu_cpu_kthread_task, cpu) != NULL)
1517                return 0;
1518        t = kthread_create_on_node(rcu_cpu_kthread,
1519                                   (void *)(long)cpu,
1520                                   cpu_to_node(cpu),
1521                                   "rcuc/%d", cpu);
1522        if (IS_ERR(t))
1523                return PTR_ERR(t);
1524        if (cpu_online(cpu))
1525                kthread_bind(t, cpu);
1526        per_cpu(rcu_cpu_kthread_cpu, cpu) = cpu;
1527        WARN_ON_ONCE(per_cpu(rcu_cpu_kthread_task, cpu) != NULL);
1528        sp.sched_priority = RCU_KTHREAD_PRIO;
1529        sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1530        per_cpu(rcu_cpu_kthread_task, cpu) = t;
1531        wake_up_process(t); /* Get to TASK_INTERRUPTIBLE quickly. */
1532        return 0;
1533}
1534
1535/*
1536 * Per-rcu_node kthread, which is in charge of waking up the per-CPU
1537 * kthreads when needed.  We ignore requests to wake up kthreads
1538 * for offline CPUs, which is OK because force_quiescent_state()
1539 * takes care of this case.
1540 */
1541static int rcu_node_kthread(void *arg)
1542{
1543        int cpu;
1544        unsigned long flags;
1545        unsigned long mask;
1546        struct rcu_node *rnp = (struct rcu_node *)arg;
1547        struct sched_param sp;
1548        struct task_struct *t;
1549
1550        for (;;) {
1551                rnp->node_kthread_status = RCU_KTHREAD_WAITING;
1552                rcu_wait(atomic_read(&rnp->wakemask) != 0);
1553                rnp->node_kthread_status = RCU_KTHREAD_RUNNING;
1554                raw_spin_lock_irqsave(&rnp->lock, flags);
1555                mask = atomic_xchg(&rnp->wakemask, 0);
1556                rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1557                for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1) {
1558                        if ((mask & 0x1) == 0)
1559                                continue;
1560                        preempt_disable();
1561                        t = per_cpu(rcu_cpu_kthread_task, cpu);
1562                        if (!cpu_online(cpu) || t == NULL) {
1563                                preempt_enable();
1564                                continue;
1565                        }
1566                        per_cpu(rcu_cpu_has_work, cpu) = 1;
1567                        sp.sched_priority = RCU_KTHREAD_PRIO;
1568                        sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1569                        preempt_enable();
1570                }
1571        }
1572        /* NOTREACHED */
1573        rnp->node_kthread_status = RCU_KTHREAD_STOPPED;
1574        return 0;
1575}
1576
1577/*
1578 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1579 * served by the rcu_node in question.  The CPU hotplug lock is still
1580 * held, so the value of rnp->qsmaskinit will be stable.
1581 *
1582 * We don't include outgoingcpu in the affinity set, use -1 if there is
1583 * no outgoing CPU.  If there are no CPUs left in the affinity set,
1584 * this function allows the kthread to execute on any CPU.
1585 */
1586static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1587{
1588        cpumask_var_t cm;
1589        int cpu;
1590        unsigned long mask = rnp->qsmaskinit;
1591
1592        if (rnp->node_kthread_task == NULL)
1593                return;
1594        if (!alloc_cpumask_var(&cm, GFP_KERNEL))
1595                return;
1596        cpumask_clear(cm);
1597        for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1598                if ((mask & 0x1) && cpu != outgoingcpu)
1599                        cpumask_set_cpu(cpu, cm);
1600        if (cpumask_weight(cm) == 0) {
1601                cpumask_setall(cm);
1602                for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1603                        cpumask_clear_cpu(cpu, cm);
1604                WARN_ON_ONCE(cpumask_weight(cm) == 0);
1605        }
1606        set_cpus_allowed_ptr(rnp->node_kthread_task, cm);
1607        rcu_boost_kthread_setaffinity(rnp, cm);
1608        free_cpumask_var(cm);
1609}
1610
1611/*
1612 * Spawn a per-rcu_node kthread, setting priority and affinity.
1613 * Called during boot before online/offline can happen, or, if
1614 * during runtime, with the main CPU-hotplug locks held.  So only
1615 * one of these can be executing at a time.
1616 */
1617static int __cpuinit rcu_spawn_one_node_kthread(struct rcu_state *rsp,
1618                                                struct rcu_node *rnp)
1619{
1620        unsigned long flags;
1621        int rnp_index = rnp - &rsp->node[0];
1622        struct sched_param sp;
1623        struct task_struct *t;
1624
1625        if (!rcu_scheduler_fully_active ||
1626            rnp->qsmaskinit == 0)
1627                return 0;
1628        if (rnp->node_kthread_task == NULL) {
1629                t = kthread_create(rcu_node_kthread, (void *)rnp,
1630                                   "rcun/%d", rnp_index);
1631                if (IS_ERR(t))
1632                        return PTR_ERR(t);
1633                raw_spin_lock_irqsave(&rnp->lock, flags);
1634                rnp->node_kthread_task = t;
1635                raw_spin_unlock_irqrestore(&rnp->lock, flags);
1636                sp.sched_priority = 99;
1637                sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1638                wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1639        }
1640        return rcu_spawn_one_boost_kthread(rsp, rnp, rnp_index);
1641}
1642
1643/*
1644 * Spawn all kthreads -- called as soon as the scheduler is running.
1645 */
1646static int __init rcu_spawn_kthreads(void)
1647{
1648        int cpu;
1649        struct rcu_node *rnp;
1650
1651        rcu_scheduler_fully_active = 1;
1652        for_each_possible_cpu(cpu) {
1653                per_cpu(rcu_cpu_has_work, cpu) = 0;
1654                if (cpu_online(cpu))
1655                        (void)rcu_spawn_one_cpu_kthread(cpu);
1656        }
1657        rnp = rcu_get_root(rcu_state);
1658        (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1659        if (NUM_RCU_NODES > 1) {
1660                rcu_for_each_leaf_node(rcu_state, rnp)
1661                        (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1662        }
1663        return 0;
1664}
1665early_initcall(rcu_spawn_kthreads);
1666
1667static void __cpuinit rcu_prepare_kthreads(int cpu)
1668{
1669        struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
1670        struct rcu_node *rnp = rdp->mynode;
1671
1672        /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1673        if (rcu_scheduler_fully_active) {
1674                (void)rcu_spawn_one_cpu_kthread(cpu);
1675                if (rnp->node_kthread_task == NULL)
1676                        (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1677        }
1678}
1679
1680#else /* #ifdef CONFIG_RCU_BOOST */
1681
1682static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1683{
1684        raw_spin_unlock_irqrestore(&rnp->lock, flags);
1685}
1686
1687static void invoke_rcu_callbacks_kthread(void)
1688{
1689        WARN_ON_ONCE(1);
1690}
1691
1692static bool rcu_is_callbacks_kthread(void)
1693{
1694        return false;
1695}
1696
1697static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1698{
1699}
1700
1701#ifdef CONFIG_HOTPLUG_CPU
1702
1703static void rcu_stop_cpu_kthread(int cpu)
1704{
1705}
1706
1707#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1708
1709static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1710{
1711}
1712
1713static void rcu_cpu_kthread_setrt(int cpu, int to_rt)
1714{
1715}
1716
1717static int __init rcu_scheduler_really_started(void)
1718{
1719        rcu_scheduler_fully_active = 1;
1720        return 0;
1721}
1722early_initcall(rcu_scheduler_really_started);
1723
1724static void __cpuinit rcu_prepare_kthreads(int cpu)
1725{
1726}
1727
1728#endif /* #else #ifdef CONFIG_RCU_BOOST */
1729
1730#if !defined(CONFIG_RCU_FAST_NO_HZ)
1731
1732/*
1733 * Check to see if any future RCU-related work will need to be done
1734 * by the current CPU, even if none need be done immediately, returning
1735 * 1 if so.  This function is part of the RCU implementation; it is -not-
1736 * an exported member of the RCU API.
1737 *
1738 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1739 * any flavor of RCU.
1740 */
1741int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies)
1742{
1743        *delta_jiffies = ULONG_MAX;
1744        return rcu_cpu_has_callbacks(cpu);
1745}
1746
1747/*
1748 * Because we do not have RCU_FAST_NO_HZ, don't bother initializing for it.
1749 */
1750static void rcu_prepare_for_idle_init(int cpu)
1751{
1752}
1753
1754/*
1755 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1756 * after it.
1757 */
1758static void rcu_cleanup_after_idle(int cpu)
1759{
1760}
1761
1762/*
1763 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1764 * is nothing.
1765 */
1766static void rcu_prepare_for_idle(int cpu)
1767{
1768}
1769
1770/*
1771 * Don't bother keeping a running count of the number of RCU callbacks
1772 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1773 */
1774static void rcu_idle_count_callbacks_posted(void)
1775{
1776}
1777
1778#else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1779
1780/*
1781 * This code is invoked when a CPU goes idle, at which point we want
1782 * to have the CPU do everything required for RCU so that it can enter
1783 * the energy-efficient dyntick-idle mode.  This is handled by a
1784 * state machine implemented by rcu_prepare_for_idle() below.
1785 *
1786 * The following three proprocessor symbols control this state machine:
1787 *
1788 * RCU_IDLE_FLUSHES gives the maximum number of times that we will attempt
1789 *      to satisfy RCU.  Beyond this point, it is better to incur a periodic
1790 *      scheduling-clock interrupt than to loop through the state machine
1791 *      at full power.
1792 * RCU_IDLE_OPT_FLUSHES gives the number of RCU_IDLE_FLUSHES that are
1793 *      optional if RCU does not need anything immediately from this
1794 *      CPU, even if this CPU still has RCU callbacks queued.  The first
1795 *      times through the state machine are mandatory: we need to give
1796 *      the state machine a chance to communicate a quiescent state
1797 *      to the RCU core.
1798 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1799 *      to sleep in dyntick-idle mode with RCU callbacks pending.  This
1800 *      is sized to be roughly one RCU grace period.  Those energy-efficiency
1801 *      benchmarkers who might otherwise be tempted to set this to a large
1802 *      number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1803 *      system.  And if you are -that- concerned about energy efficiency,
1804 *      just power the system down and be done with it!
1805 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1806 *      permitted to sleep in dyntick-idle mode with only lazy RCU
1807 *      callbacks pending.  Setting this too high can OOM your system.
1808 *
1809 * The values below work well in practice.  If future workloads require
1810 * adjustment, they can be converted into kernel config parameters, though
1811 * making the state machine smarter might be a better option.
1812 */
1813#define RCU_IDLE_FLUSHES 5              /* Number of dyntick-idle tries. */
1814#define RCU_IDLE_OPT_FLUSHES 3          /* Optional dyntick-idle tries. */
1815#define RCU_IDLE_GP_DELAY 4             /* Roughly one grace period. */
1816#define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1817
1818extern int tick_nohz_enabled;
1819
1820/*
1821 * Does the specified flavor of RCU have non-lazy callbacks pending on
1822 * the specified CPU?  Both RCU flavor and CPU are specified by the
1823 * rcu_data structure.
1824 */
1825static bool __rcu_cpu_has_nonlazy_callbacks(struct rcu_data *rdp)
1826{
1827        return rdp->qlen != rdp->qlen_lazy;
1828}
1829
1830#ifdef CONFIG_TREE_PREEMPT_RCU
1831
1832/*
1833 * Are there non-lazy RCU-preempt callbacks?  (There cannot be if there
1834 * is no RCU-preempt in the kernel.)
1835 */
1836static bool rcu_preempt_cpu_has_nonlazy_callbacks(int cpu)
1837{
1838        struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
1839
1840        return __rcu_cpu_has_nonlazy_callbacks(rdp);
1841}
1842
1843#else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
1844
1845static bool rcu_preempt_cpu_has_nonlazy_callbacks(int cpu)
1846{
1847        return 0;
1848}
1849
1850#endif /* else #ifdef CONFIG_TREE_PREEMPT_RCU */
1851
1852/*
1853 * Does any flavor of RCU have non-lazy callbacks on the specified CPU?
1854 */
1855static bool rcu_cpu_has_nonlazy_callbacks(int cpu)
1856{
1857        return __rcu_cpu_has_nonlazy_callbacks(&per_cpu(rcu_sched_data, cpu)) ||
1858               __rcu_cpu_has_nonlazy_callbacks(&per_cpu(rcu_bh_data, cpu)) ||
1859               rcu_preempt_cpu_has_nonlazy_callbacks(cpu);
1860}
1861
1862/*
1863 * Allow the CPU to enter dyntick-idle mode if either: (1) There are no
1864 * callbacks on this CPU, (2) this CPU has not yet attempted to enter
1865 * dyntick-idle mode, or (3) this CPU is in the process of attempting to
1866 * enter dyntick-idle mode.  Otherwise, if we have recently tried and failed
1867 * to enter dyntick-idle mode, we refuse to try to enter it.  After all,
1868 * it is better to incur scheduling-clock interrupts than to spin
1869 * continuously for the same time duration!
1870 *
1871 * The delta_jiffies argument is used to store the time when RCU is
1872 * going to need the CPU again if it still has callbacks.  The reason
1873 * for this is that rcu_prepare_for_idle() might need to post a timer,
1874 * but if so, it will do so after tick_nohz_stop_sched_tick() has set
1875 * the wakeup time for this CPU.  This means that RCU's timer can be
1876 * delayed until the wakeup time, which defeats the purpose of posting
1877 * a timer.
1878 */
1879int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies)
1880{
1881        struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1882
1883        /* Flag a new idle sojourn to the idle-entry state machine. */
1884        rdtp->idle_first_pass = 1;
1885        /* If no callbacks, RCU doesn't need the CPU. */
1886        if (!rcu_cpu_has_callbacks(cpu)) {
1887                *delta_jiffies = ULONG_MAX;
1888                return 0;
1889        }
1890        if (rdtp->dyntick_holdoff == jiffies) {
1891                /* RCU recently tried and failed, so don't try again. */
1892                *delta_jiffies = 1;
1893                return 1;
1894        }
1895        /* Set up for the possibility that RCU will post a timer. */
1896        if (rcu_cpu_has_nonlazy_callbacks(cpu)) {
1897                *delta_jiffies = round_up(RCU_IDLE_GP_DELAY + jiffies,
1898                                          RCU_IDLE_GP_DELAY) - jiffies;
1899        } else {
1900                *delta_jiffies = jiffies + RCU_IDLE_LAZY_GP_DELAY;
1901                *delta_jiffies = round_jiffies(*delta_jiffies) - jiffies;
1902        }
1903        return 0;
1904}
1905
1906/*
1907 * Handler for smp_call_function_single().  The only point of this
1908 * handler is to wake the CPU up, so the handler does only tracing.
1909 */
1910void rcu_idle_demigrate(void *unused)
1911{
1912        trace_rcu_prep_idle("Demigrate");
1913}
1914
1915/*
1916 * Timer handler used to force CPU to start pushing its remaining RCU
1917 * callbacks in the case where it entered dyntick-idle mode with callbacks
1918 * pending.  The hander doesn't really need to do anything because the
1919 * real work is done upon re-entry to idle, or by the next scheduling-clock
1920 * interrupt should idle not be re-entered.
1921 *
1922 * One special case: the timer gets migrated without awakening the CPU
1923 * on which the timer was scheduled on.  In this case, we must wake up
1924 * that CPU.  We do so with smp_call_function_single().
1925 */
1926static void rcu_idle_gp_timer_func(unsigned long cpu_in)
1927{
1928        int cpu = (int)cpu_in;
1929
1930        trace_rcu_prep_idle("Timer");
1931        if (cpu != smp_processor_id())
1932                smp_call_function_single(cpu, rcu_idle_demigrate, NULL, 0);
1933        else
1934                WARN_ON_ONCE(1); /* Getting here can hang the system... */
1935}
1936
1937/*
1938 * Initialize the timer used to pull CPUs out of dyntick-idle mode.
1939 */
1940static void rcu_prepare_for_idle_init(int cpu)
1941{
1942        struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1943
1944        rdtp->dyntick_holdoff = jiffies - 1;
1945        setup_timer(&rdtp->idle_gp_timer, rcu_idle_gp_timer_func, cpu);
1946        rdtp->idle_gp_timer_expires = jiffies - 1;
1947        rdtp->idle_first_pass = 1;
1948}
1949
1950/*
1951 * Clean up for exit from idle.  Because we are exiting from idle, there
1952 * is no longer any point to ->idle_gp_timer, so cancel it.  This will
1953 * do nothing if this timer is not active, so just cancel it unconditionally.
1954 */
1955static void rcu_cleanup_after_idle(int cpu)
1956{
1957        struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1958
1959        del_timer(&rdtp->idle_gp_timer);
1960        trace_rcu_prep_idle("Cleanup after idle");
1961        rdtp->tick_nohz_enabled_snap = ACCESS_ONCE(tick_nohz_enabled);
1962}
1963
1964/*
1965 * Check to see if any RCU-related work can be done by the current CPU,
1966 * and if so, schedule a softirq to get it done.  This function is part
1967 * of the RCU implementation; it is -not- an exported member of the RCU API.
1968 *
1969 * The idea is for the current CPU to clear out all work required by the
1970 * RCU core for the current grace period, so that this CPU can be permitted
1971 * to enter dyntick-idle mode.  In some cases, it will need to be awakened
1972 * at the end of the grace period by whatever CPU ends the grace period.
1973 * This allows CPUs to go dyntick-idle more quickly, and to reduce the
1974 * number of wakeups by a modest integer factor.
1975 *
1976 * Because it is not legal to invoke rcu_process_callbacks() with irqs
1977 * disabled, we do one pass of force_quiescent_state(), then do a
1978 * invoke_rcu_core() to cause rcu_process_callbacks() to be invoked
1979 * later.  The ->dyntick_drain field controls the sequencing.
1980 *
1981 * The caller must have disabled interrupts.
1982 */
1983static void rcu_prepare_for_idle(int cpu)
1984{
1985        struct timer_list *tp;
1986        struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1987        int tne;
1988
1989        /* Handle nohz enablement switches conservatively. */
1990        tne = ACCESS_ONCE(tick_nohz_enabled);
1991        if (tne != rdtp->tick_nohz_enabled_snap) {
1992                if (rcu_cpu_has_callbacks(cpu))
1993                        invoke_rcu_core(); /* force nohz to see update. */
1994                rdtp->tick_nohz_enabled_snap = tne;
1995                return;
1996        }
1997        if (!tne)
1998                return;
1999
2000        /*
2001         * If this is an idle re-entry, for example, due to use of
2002         * RCU_NONIDLE() or the new idle-loop tracing API within the idle
2003         * loop, then don't take any state-machine actions, unless the
2004         * momentary exit from idle queued additional non-lazy callbacks.
2005         * Instead, repost the ->idle_gp_timer if this CPU has callbacks
2006         * pending.
2007         */
2008        if (!rdtp->idle_first_pass &&
2009            (rdtp->nonlazy_posted == rdtp->nonlazy_posted_snap)) {
2010                if (rcu_cpu_has_callbacks(cpu)) {
2011                        tp = &rdtp->idle_gp_timer;
2012                        mod_timer_pinned(tp, rdtp->idle_gp_timer_expires);
2013                }
2014                return;
2015        }
2016        rdtp->idle_first_pass = 0;
2017        rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted - 1;
2018
2019        /*
2020         * If there are no callbacks on this CPU, enter dyntick-idle mode.
2021         * Also reset state to avoid prejudicing later attempts.
2022         */
2023        if (!rcu_cpu_has_callbacks(cpu)) {
2024                rdtp->dyntick_holdoff = jiffies - 1;
2025                rdtp->dyntick_drain = 0;
2026                trace_rcu_prep_idle("No callbacks");
2027                return;
2028        }
2029
2030        /*
2031         * If in holdoff mode, just return.  We will presumably have
2032         * refrained from disabling the scheduling-clock tick.
2033         */
2034        if (rdtp->dyntick_holdoff == jiffies) {
2035                trace_rcu_prep_idle("In holdoff");
2036                return;
2037        }
2038
2039        /* Check and update the ->dyntick_drain sequencing. */
2040        if (rdtp->dyntick_drain <= 0) {
2041                /* First time through, initialize the counter. */
2042                rdtp->dyntick_drain = RCU_IDLE_FLUSHES;
2043        } else if (rdtp->dyntick_drain <= RCU_IDLE_OPT_FLUSHES &&
2044                   !rcu_pending(cpu) &&
2045                   !local_softirq_pending()) {
2046                /* Can we go dyntick-idle despite still having callbacks? */
2047                rdtp->dyntick_drain = 0;
2048                rdtp->dyntick_holdoff = jiffies;
2049                if (rcu_cpu_has_nonlazy_callbacks(cpu)) {
2050                        trace_rcu_prep_idle("Dyntick with callbacks");
2051                        rdtp->idle_gp_timer_expires =
2052                                round_up(jiffies + RCU_IDLE_GP_DELAY,
2053                                         RCU_IDLE_GP_DELAY);
2054                } else {
2055                        rdtp->idle_gp_timer_expires =
2056                                round_jiffies(jiffies + RCU_IDLE_LAZY_GP_DELAY);
2057                        trace_rcu_prep_idle("Dyntick with lazy callbacks");
2058                }
2059                tp = &rdtp->idle_gp_timer;
2060                mod_timer_pinned(tp, rdtp->idle_gp_timer_expires);
2061                rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
2062                return; /* Nothing more to do immediately. */
2063        } else if (--(rdtp->dyntick_drain) <= 0) {
2064                /* We have hit the limit, so time to give up. */
2065                rdtp->dyntick_holdoff = jiffies;
2066                trace_rcu_prep_idle("Begin holdoff");
2067                invoke_rcu_core();  /* Force the CPU out of dyntick-idle. */
2068                return;
2069        }
2070
2071        /*
2072         * Do one step of pushing the remaining RCU callbacks through
2073         * the RCU core state machine.
2074         */
2075#ifdef CONFIG_TREE_PREEMPT_RCU
2076        if (per_cpu(rcu_preempt_data, cpu).nxtlist) {
2077                rcu_preempt_qs(cpu);
2078                force_quiescent_state(&rcu_preempt_state, 0);
2079        }
2080#endif /* #ifdef CONFIG_TREE_PREEMPT_RCU */
2081        if (per_cpu(rcu_sched_data, cpu).nxtlist) {
2082                rcu_sched_qs(cpu);
2083                force_quiescent_state(&rcu_sched_state, 0);
2084        }
2085        if (per_cpu(rcu_bh_data, cpu).nxtlist) {
2086                rcu_bh_qs(cpu);
2087                force_quiescent_state(&rcu_bh_state, 0);
2088        }
2089
2090        /*
2091         * If RCU callbacks are still pending, RCU still needs this CPU.
2092         * So try forcing the callbacks through the grace period.
2093         */
2094        if (rcu_cpu_has_callbacks(cpu)) {
2095                trace_rcu_prep_idle("More callbacks");
2096                invoke_rcu_core();
2097        } else {
2098                trace_rcu_prep_idle("Callbacks drained");
2099        }
2100}
2101
2102/*
2103 * Keep a running count of the number of non-lazy callbacks posted
2104 * on this CPU.  This running counter (which is never decremented) allows
2105 * rcu_prepare_for_idle() to detect when something out of the idle loop
2106 * posts a callback, even if an equal number of callbacks are invoked.
2107 * Of course, callbacks should only be posted from within a trace event
2108 * designed to be called from idle or from within RCU_NONIDLE().
2109 */
2110static void rcu_idle_count_callbacks_posted(void)
2111{
2112        __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
2113}
2114
2115#endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
2116
2117#ifdef CONFIG_RCU_CPU_STALL_INFO
2118
2119#ifdef CONFIG_RCU_FAST_NO_HZ
2120
2121static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
2122{
2123        struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
2124        struct timer_list *tltp = &rdtp->idle_gp_timer;
2125
2126        sprintf(cp, "drain=%d %c timer=%lu",
2127                rdtp->dyntick_drain,
2128                rdtp->dyntick_holdoff == jiffies ? 'H' : '.',
2129                timer_pending(tltp) ? tltp->expires - jiffies : -1);
2130}
2131
2132#else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
2133
2134static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
2135{
2136        *cp = '\0';
2137}
2138
2139#endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
2140
2141/* Initiate the stall-info list. */
2142static void print_cpu_stall_info_begin(void)
2143{
2144        printk(KERN_CONT "\n");
2145}
2146
2147/*
2148 * Print out diagnostic information for the specified stalled CPU.
2149 *
2150 * If the specified CPU is aware of the current RCU grace period
2151 * (flavor specified by rsp), then print the number of scheduling
2152 * clock interrupts the CPU has taken during the time that it has
2153 * been aware.  Otherwise, print the number of RCU grace periods
2154 * that this CPU is ignorant of, for example, "1" if the CPU was
2155 * aware of the previous grace period.
2156 *
2157 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
2158 */
2159static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
2160{
2161        char fast_no_hz[72];
2162        struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2163        struct rcu_dynticks *rdtp = rdp->dynticks;
2164        char *ticks_title;
2165        unsigned long ticks_value;
2166
2167        if (rsp->gpnum == rdp->gpnum) {
2168                ticks_title = "ticks this GP";
2169                ticks_value = rdp->ticks_this_gp;
2170        } else {
2171                ticks_title = "GPs behind";
2172                ticks_value = rsp->gpnum - rdp->gpnum;
2173        }
2174        print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
2175        printk(KERN_ERR "\t%d: (%lu %s) idle=%03x/%llx/%d %s\n",
2176               cpu, ticks_value, ticks_title,
2177               atomic_read(&rdtp->dynticks) & 0xfff,
2178               rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
2179               fast_no_hz);
2180}
2181
2182/* Terminate the stall-info list. */
2183static void print_cpu_stall_info_end(void)
2184{
2185        printk(KERN_ERR "\t");
2186}
2187
2188/* Zero ->ticks_this_gp for all flavors of RCU. */
2189static void zero_cpu_stall_ticks(struct rcu_data *rdp)
2190{
2191        rdp->ticks_this_gp = 0;
2192}
2193
2194/* Increment ->ticks_this_gp for all flavors of RCU. */
2195static void increment_cpu_stall_ticks(void)
2196{
2197        __get_cpu_var(rcu_sched_data).ticks_this_gp++;
2198        __get_cpu_var(rcu_bh_data).ticks_this_gp++;
2199#ifdef CONFIG_TREE_PREEMPT_RCU
2200        __get_cpu_var(rcu_preempt_data).ticks_this_gp++;
2201#endif /* #ifdef CONFIG_TREE_PREEMPT_RCU */
2202}
2203
2204#else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
2205
2206static void print_cpu_stall_info_begin(void)
2207{
2208        printk(KERN_CONT " {");
2209}
2210
2211static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
2212{
2213        printk(KERN_CONT " %d", cpu);
2214}
2215
2216static void print_cpu_stall_info_end(void)
2217{
2218        printk(KERN_CONT "} ");
2219}
2220
2221static void zero_cpu_stall_ticks(struct rcu_data *rdp)
2222{
2223}
2224
2225static void increment_cpu_stall_ticks(void)
2226{
2227}
2228
2229#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
2230
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.