kernel/rcu/tree_exp.h

   1 /*
   2  * RCU expedited grace periods
   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, you can access it online at
  16  * http://www.gnu.org/licenses/gpl-2.0.html.
  17  *
  18  * Copyright IBM Corporation, 2016
  19  *
  20  * Authors: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
  21  */
  22
  23 /*
  24  * Record the start of an expedited grace period.
  25  */
  26 static void rcu_exp_gp_seq_start(struct rcu_state *rsp)
  27 {
  28         rcu_seq_start(&rsp->expedited_sequence);
  29 }
  30
  31 /*
  32  * Record the end of an expedited grace period.
  33  */
  34 static void rcu_exp_gp_seq_end(struct rcu_state *rsp)
  35 {
  36         rcu_seq_end(&rsp->expedited_sequence);
  37         smp_mb(); /* Ensure that consecutive grace periods serialize. */
  38 }
  39
  40 /*
  41  * Take a snapshot of the expedited-grace-period counter.
  42  */
  43 static unsigned long rcu_exp_gp_seq_snap(struct rcu_state *rsp)
  44 {
  45         unsigned long s;
  46
  47         smp_mb(); /* Caller's modifications seen first by other CPUs. */
  48         s = rcu_seq_snap(&rsp->expedited_sequence);
  49         trace_rcu_exp_grace_period(rsp->name, s, TPS("snap"));
  50         return s;
  51 }
  52
  53 /*
  54  * Given a counter snapshot from rcu_exp_gp_seq_snap(), return true
  55  * if a full expedited grace period has elapsed since that snapshot
  56  * was taken.
  57  */
  58 static bool rcu_exp_gp_seq_done(struct rcu_state *rsp, unsigned long s)
  59 {
  60         return rcu_seq_done(&rsp->expedited_sequence, s);
  61 }
  62
  63 /*
  64  * Reset the ->expmaskinit values in the rcu_node tree to reflect any
  65  * recent CPU-online activity.  Note that these masks are not cleared
  66  * when CPUs go offline, so they reflect the union of all CPUs that have
  67  * ever been online.  This means that this function normally takes its
  68  * no-work-to-do fastpath.
  69  */
  70 static void sync_exp_reset_tree_hotplug(struct rcu_state *rsp)
  71 {
  72         bool done;
  73         unsigned long flags;
  74         unsigned long mask;
  75         unsigned long oldmask;
  76         int ncpus = smp_load_acquire(&rsp->ncpus); /* Order against locking. */
  77         struct rcu_node *rnp;
  78         struct rcu_node *rnp_up;
  79
  80         /* If no new CPUs onlined since last time, nothing to do. */
  81         if (likely(ncpus == rsp->ncpus_snap))
  82                 return;
  83         rsp->ncpus_snap = ncpus;
  84
  85         /*
  86          * Each pass through the following loop propagates newly onlined
  87          * CPUs for the current rcu_node structure up the rcu_node tree.
  88          */
  89         rcu_for_each_leaf_node(rsp, rnp) {
  90                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
  91                 if (rnp->expmaskinit == rnp->expmaskinitnext) {
  92                         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
  93                         continue;  /* No new CPUs, nothing to do. */
  94                 }
  95
  96                 /* Update this node's mask, track old value for propagation. */
  97                 oldmask = rnp->expmaskinit;
  98                 rnp->expmaskinit = rnp->expmaskinitnext;
  99                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 100
 101                 /* If was already nonzero, nothing to propagate. */
 102                 if (oldmask)
 103                         continue;
 104
 105                 /* Propagate the new CPU up the tree. */
 106                 mask = rnp->grpmask;
 107                 rnp_up = rnp->parent;
 108                 done = false;
 109                 while (rnp_up) {
 110                         raw_spin_lock_irqsave_rcu_node(rnp_up, flags);
 111                         if (rnp_up->expmaskinit)
 112                                 done = true;
 113                         rnp_up->expmaskinit |= mask;
 114                         raw_spin_unlock_irqrestore_rcu_node(rnp_up, flags);
 115                         if (done)
 116                                 break;
 117                         mask = rnp_up->grpmask;
 118                         rnp_up = rnp_up->parent;
 119                 }
 120         }
 121 }
 122
 123 /*
 124  * Reset the ->expmask values in the rcu_node tree in preparation for
 125  * a new expedited grace period.
 126  */
 127 static void __maybe_unused sync_exp_reset_tree(struct rcu_state *rsp)
 128 {
 129         unsigned long flags;
 130         struct rcu_node *rnp;
 131
 132         sync_exp_reset_tree_hotplug(rsp);
 133         rcu_for_each_node_breadth_first(rsp, rnp) {
 134                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 135                 WARN_ON_ONCE(rnp->expmask);
 136                 rnp->expmask = rnp->expmaskinit;
 137                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 138         }
 139 }
 140
 141 /*
 142  * Return non-zero if there is no RCU expedited grace period in progress
 143  * for the specified rcu_node structure, in other words, if all CPUs and
 144  * tasks covered by the specified rcu_node structure have done their bit
 145  * for the current expedited grace period.  Works only for preemptible
 146  * RCU -- other RCU implementation use other means.
 147  *
 148  * Caller must hold the rcu_state's exp_mutex.
 149  */
 150 static bool sync_rcu_preempt_exp_done(struct rcu_node *rnp)
 151 {
 152         return rnp->exp_tasks == NULL &&
 153                READ_ONCE(rnp->expmask) == 0;
 154 }
 155
 156 /*
 157  * Report the exit from RCU read-side critical section for the last task
 158  * that queued itself during or before the current expedited preemptible-RCU
 159  * grace period.  This event is reported either to the rcu_node structure on
 160  * which the task was queued or to one of that rcu_node structure's ancestors,
 161  * recursively up the tree.  (Calm down, calm down, we do the recursion
 162  * iteratively!)
 163  *
 164  * Caller must hold the rcu_state's exp_mutex and the specified rcu_node
 165  * structure's ->lock.
 166  */
 167 static void __rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
 168                                  bool wake, unsigned long flags)
 169         __releases(rnp->lock)
 170 {
 171         unsigned long mask;
 172
 173         for (;;) {
 174                 if (!sync_rcu_preempt_exp_done(rnp)) {
 175                         if (!rnp->expmask)
 176                                 rcu_initiate_boost(rnp, flags);
 177                         else
 178                                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 179                         break;
 180                 }
 181                 if (rnp->parent == NULL) {
 182                         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 183                         if (wake) {
 184                                 smp_mb(); /* EGP done before wake_up(). */
 185                                 swake_up(&rsp->expedited_wq);
 186                         }
 187                         break;
 188                 }
 189                 mask = rnp->grpmask;
 190                 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled */
 191                 rnp = rnp->parent;
 192                 raw_spin_lock_rcu_node(rnp); /* irqs already disabled */
 193                 WARN_ON_ONCE(!(rnp->expmask & mask));
 194                 rnp->expmask &= ~mask;
 195         }
 196 }
 197
 198 /*
 199  * Report expedited quiescent state for specified node.  This is a
 200  * lock-acquisition wrapper function for __rcu_report_exp_rnp().
 201  *
 202  * Caller must hold the rcu_state's exp_mutex.
 203  */
 204 static void __maybe_unused rcu_report_exp_rnp(struct rcu_state *rsp,
 205                                               struct rcu_node *rnp, bool wake)
 206 {
 207         unsigned long flags;
 208
 209         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 210         __rcu_report_exp_rnp(rsp, rnp, wake, flags);
 211 }
 212
 213 /*
 214  * Report expedited quiescent state for multiple CPUs, all covered by the
 215  * specified leaf rcu_node structure.  Caller must hold the rcu_state's
 216  * exp_mutex.
 217  */
 218 static void rcu_report_exp_cpu_mult(struct rcu_state *rsp, struct rcu_node *rnp,
 219                                     unsigned long mask, bool wake)
 220 {
 221         unsigned long flags;
 222
 223         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 224         if (!(rnp->expmask & mask)) {
 225                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 226                 return;
 227         }
 228         rnp->expmask &= ~mask;
 229         __rcu_report_exp_rnp(rsp, rnp, wake, flags); /* Releases rnp->lock. */
 230 }
 231
 232 /*
 233  * Report expedited quiescent state for specified rcu_data (CPU).
 234  */
 235 static void rcu_report_exp_rdp(struct rcu_state *rsp, struct rcu_data *rdp,
 236                                bool wake)
 237 {
 238         rcu_report_exp_cpu_mult(rsp, rdp->mynode, rdp->grpmask, wake);
 239 }
 240
 241 /* Common code for synchronize_{rcu,sched}_expedited() work-done checking. */
 242 static bool sync_exp_work_done(struct rcu_state *rsp, atomic_long_t *stat,
 243                                unsigned long s)
 244 {
 245         if (rcu_exp_gp_seq_done(rsp, s)) {
 246                 trace_rcu_exp_grace_period(rsp->name, s, TPS("done"));
 247                 /* Ensure test happens before caller kfree(). */
 248                 smp_mb__before_atomic(); /* ^^^ */
 249                 atomic_long_inc(stat);
 250                 return true;
 251         }
 252         return false;
 253 }
 254
 255 /*
 256  * Funnel-lock acquisition for expedited grace periods.  Returns true
 257  * if some other task completed an expedited grace period that this task
 258  * can piggy-back on, and with no mutex held.  Otherwise, returns false
 259  * with the mutex held, indicating that the caller must actually do the
 260  * expedited grace period.
 261  */
 262 static bool exp_funnel_lock(struct rcu_state *rsp, unsigned long s)
 263 {
 264         struct rcu_data *rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id());
 265         struct rcu_node *rnp = rdp->mynode;
 266         struct rcu_node *rnp_root = rcu_get_root(rsp);
 267
 268         /* Low-contention fastpath. */
 269         if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s) &&
 270             (rnp == rnp_root ||
 271              ULONG_CMP_LT(READ_ONCE(rnp_root->exp_seq_rq), s)) &&
 272             mutex_trylock(&rsp->exp_mutex))
 273                 goto fastpath;
 274
 275         /*
 276          * Each pass through the following loop works its way up
 277          * the rcu_node tree, returning if others have done the work or
 278          * otherwise falls through to acquire rsp->exp_mutex.  The mapping
 279          * from CPU to rcu_node structure can be inexact, as it is just
 280          * promoting locality and is not strictly needed for correctness.
 281          */
 282         for (; rnp != NULL; rnp = rnp->parent) {
 283                 if (sync_exp_work_done(rsp, &rdp->exp_workdone1, s))
 284                         return true;
 285
 286                 /* Work not done, either wait here or go up. */
 287                 spin_lock(&rnp->exp_lock);
 288                 if (ULONG_CMP_GE(rnp->exp_seq_rq, s)) {
 289
 290                         /* Someone else doing GP, so wait for them. */
 291                         spin_unlock(&rnp->exp_lock);
 292                         trace_rcu_exp_funnel_lock(rsp->name, rnp->level,
 293                                                   rnp->grplo, rnp->grphi,
 294                                                   TPS("wait"));
 295                         wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
 296                                    sync_exp_work_done(rsp,
 297                                                       &rdp->exp_workdone2, s));
 298                         return true;
 299                 }
 300                 rnp->exp_seq_rq = s; /* Followers can wait on us. */
 301                 spin_unlock(&rnp->exp_lock);
 302                 trace_rcu_exp_funnel_lock(rsp->name, rnp->level, rnp->grplo,
 303                                           rnp->grphi, TPS("nxtlvl"));
 304         }
 305         mutex_lock(&rsp->exp_mutex);
 306 fastpath:
 307         if (sync_exp_work_done(rsp, &rdp->exp_workdone3, s)) {
 308                 mutex_unlock(&rsp->exp_mutex);
 309                 return true;
 310         }
 311         rcu_exp_gp_seq_start(rsp);
 312         trace_rcu_exp_grace_period(rsp->name, s, TPS("start"));
 313         return false;
 314 }
 315
 316 /* Invoked on each online non-idle CPU for expedited quiescent state. */
 317 static void sync_sched_exp_handler(void *data)
 318 {
 319         struct rcu_data *rdp;
 320         struct rcu_node *rnp;
 321         struct rcu_state *rsp = data;
 322
 323         rdp = this_cpu_ptr(rsp->rda);
 324         rnp = rdp->mynode;
 325         if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
 326             __this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
 327                 return;
 328         if (rcu_is_cpu_rrupt_from_idle()) {
 329                 rcu_report_exp_rdp(&rcu_sched_state,
 330                                    this_cpu_ptr(&rcu_sched_data), true);
 331                 return;
 332         }
 333         __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, true);
 334         /* Store .exp before .rcu_urgent_qs. */
 335         smp_store_release(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs), true);
 336         resched_cpu(smp_processor_id());
 337 }
 338
 339 /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
 340 static void sync_sched_exp_online_cleanup(int cpu)
 341 {
 342         struct rcu_data *rdp;
 343         int ret;
 344         struct rcu_node *rnp;
 345         struct rcu_state *rsp = &rcu_sched_state;
 346
 347         rdp = per_cpu_ptr(rsp->rda, cpu);
 348         rnp = rdp->mynode;
 349         if (!(READ_ONCE(rnp->expmask) & rdp->grpmask))
 350                 return;
 351         ret = smp_call_function_single(cpu, sync_sched_exp_handler, rsp, 0);
 352         WARN_ON_ONCE(ret);
 353 }
 354
 355 /*
 356  * Select the nodes that the upcoming expedited grace period needs
 357  * to wait for.
 358  */
 359 static void sync_rcu_exp_select_cpus(struct rcu_state *rsp,
 360                                      smp_call_func_t func)
 361 {
 362         int cpu;
 363         unsigned long flags;
 364         unsigned long mask_ofl_test;
 365         unsigned long mask_ofl_ipi;
 366         int ret;
 367         struct rcu_node *rnp;
 368
 369         sync_exp_reset_tree(rsp);
 370         rcu_for_each_leaf_node(rsp, rnp) {
 371                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 372
 373                 /* Each pass checks a CPU for identity, offline, and idle. */
 374                 mask_ofl_test = 0;
 375                 for_each_leaf_node_possible_cpu(rnp, cpu) {
 376                         struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
 377
 378                         rdp->exp_dynticks_snap =
 379                                 rcu_dynticks_snap(rdp->dynticks);
 380                         if (raw_smp_processor_id() == cpu ||
 381                             rcu_dynticks_in_eqs(rdp->exp_dynticks_snap) ||
 382                             !(rnp->qsmaskinitnext & rdp->grpmask))
 383                                 mask_ofl_test |= rdp->grpmask;
 384                 }
 385                 mask_ofl_ipi = rnp->expmask & ~mask_ofl_test;
 386
 387                 /*
 388                  * Need to wait for any blocked tasks as well.  Note that
 389                  * additional blocking tasks will also block the expedited
 390                  * GP until such time as the ->expmask bits are cleared.
 391                  */
 392                 if (rcu_preempt_has_tasks(rnp))
 393                         rnp->exp_tasks = rnp->blkd_tasks.next;
 394                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 395
 396                 /* IPI the remaining CPUs for expedited quiescent state. */
 397                 for_each_leaf_node_possible_cpu(rnp, cpu) {
 398                         unsigned long mask = leaf_node_cpu_bit(rnp, cpu);
 399                         struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
 400
 401                         if (!(mask_ofl_ipi & mask))
 402                                 continue;
 403 retry_ipi:
 404                         if (rcu_dynticks_in_eqs_since(rdp->dynticks,
 405                                                       rdp->exp_dynticks_snap)) {
 406                                 mask_ofl_test |= mask;
 407                                 continue;
 408                         }
 409                         ret = smp_call_function_single(cpu, func, rsp, 0);
 410                         if (!ret) {
 411                                 mask_ofl_ipi &= ~mask;
 412                                 continue;
 413                         }
 414                         /* Failed, raced with CPU hotplug operation. */
 415                         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 416                         if ((rnp->qsmaskinitnext & mask) &&
 417                             (rnp->expmask & mask)) {
 418                                 /* Online, so delay for a bit and try again. */
 419                                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 420                                 schedule_timeout_uninterruptible(1);
 421                                 goto retry_ipi;
 422                         }
 423                         /* CPU really is offline, so we can ignore it. */
 424                         if (!(rnp->expmask & mask))
 425                                 mask_ofl_ipi &= ~mask;
 426                         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 427                 }
 428                 /* Report quiescent states for those that went offline. */
 429                 mask_ofl_test |= mask_ofl_ipi;
 430                 if (mask_ofl_test)
 431                         rcu_report_exp_cpu_mult(rsp, rnp, mask_ofl_test, false);
 432         }
 433 }
 434
 435 static void synchronize_sched_expedited_wait(struct rcu_state *rsp)
 436 {
 437         int cpu;
 438         unsigned long jiffies_stall;
 439         unsigned long jiffies_start;
 440         unsigned long mask;
 441         int ndetected;
 442         struct rcu_node *rnp;
 443         struct rcu_node *rnp_root = rcu_get_root(rsp);
 444         int ret;
 445
 446         jiffies_stall = rcu_jiffies_till_stall_check();
 447         jiffies_start = jiffies;
 448
 449         for (;;) {
 450                 ret = swait_event_timeout(
 451                                 rsp->expedited_wq,
 452                                 sync_rcu_preempt_exp_done(rnp_root),
 453                                 jiffies_stall);
 454                 if (ret > 0 || sync_rcu_preempt_exp_done(rnp_root))
 455                         return;
 456                 WARN_ON(ret < 0);  /* workqueues should not be signaled. */
 457                 if (rcu_cpu_stall_suppress)
 458                         continue;
 459                 panic_on_rcu_stall();
 460                 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
 461                        rsp->name);
 462                 ndetected = 0;
 463                 rcu_for_each_leaf_node(rsp, rnp) {
 464                         ndetected += rcu_print_task_exp_stall(rnp);
 465                         for_each_leaf_node_possible_cpu(rnp, cpu) {
 466                                 struct rcu_data *rdp;
 467
 468                                 mask = leaf_node_cpu_bit(rnp, cpu);
 469                                 if (!(rnp->expmask & mask))
 470                                         continue;
 471                                 ndetected++;
 472                                 rdp = per_cpu_ptr(rsp->rda, cpu);
 473                                 pr_cont(" %d-%c%c%c", cpu,
 474                                         "O."[!!cpu_online(cpu)],
 475                                         "o."[!!(rdp->grpmask & rnp->expmaskinit)],
 476                                         "N."[!!(rdp->grpmask & rnp->expmaskinitnext)]);
 477                         }
 478                 }
 479                 pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
 480                         jiffies - jiffies_start, rsp->expedited_sequence,
 481                         rnp_root->expmask, ".T"[!!rnp_root->exp_tasks]);
 482                 if (ndetected) {
 483                         pr_err("blocking rcu_node structures:");
 484                         rcu_for_each_node_breadth_first(rsp, rnp) {
 485                                 if (rnp == rnp_root)
 486                                         continue; /* printed unconditionally */
 487                                 if (sync_rcu_preempt_exp_done(rnp))
 488                                         continue;
 489                                 pr_cont(" l=%u:%d-%d:%#lx/%c",
 490                                         rnp->level, rnp->grplo, rnp->grphi,
 491                                         rnp->expmask,
 492                                         ".T"[!!rnp->exp_tasks]);
 493                         }
 494                         pr_cont("\n");
 495                 }
 496                 rcu_for_each_leaf_node(rsp, rnp) {
 497                         for_each_leaf_node_possible_cpu(rnp, cpu) {
 498                                 mask = leaf_node_cpu_bit(rnp, cpu);
 499                                 if (!(rnp->expmask & mask))
 500                                         continue;
 501                                 dump_cpu_task(cpu);
 502                         }
 503                 }
 504                 jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3;
 505         }
 506 }
 507
 508 /*
 509  * Wait for the current expedited grace period to complete, and then
 510  * wake up everyone who piggybacked on the just-completed expedited
 511  * grace period.  Also update all the ->exp_seq_rq counters as needed
 512  * in order to avoid counter-wrap problems.
 513  */
 514 static void rcu_exp_wait_wake(struct rcu_state *rsp, unsigned long s)
 515 {
 516         struct rcu_node *rnp;
 517
 518         synchronize_sched_expedited_wait(rsp);
 519         rcu_exp_gp_seq_end(rsp);
 520         trace_rcu_exp_grace_period(rsp->name, s, TPS("end"));
 521
 522         /*
 523          * Switch over to wakeup mode, allowing the next GP, but -only- the
 524          * next GP, to proceed.
 525          */
 526         mutex_lock(&rsp->exp_wake_mutex);
 527
 528         rcu_for_each_node_breadth_first(rsp, rnp) {
 529                 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s)) {
 530                         spin_lock(&rnp->exp_lock);
 531                         /* Recheck, avoid hang in case someone just arrived. */
 532                         if (ULONG_CMP_LT(rnp->exp_seq_rq, s))
 533                                 rnp->exp_seq_rq = s;
 534                         spin_unlock(&rnp->exp_lock);
 535                 }
 536                 smp_mb(); /* All above changes before wakeup. */
 537                 wake_up_all(&rnp->exp_wq[rcu_seq_ctr(rsp->expedited_sequence) & 0x3]);
 538         }
 539         trace_rcu_exp_grace_period(rsp->name, s, TPS("endwake"));
 540         mutex_unlock(&rsp->exp_wake_mutex);
 541 }
 542
 543 /* Let the workqueue handler know what it is supposed to do. */
 544 struct rcu_exp_work {
 545         smp_call_func_t rew_func;
 546         struct rcu_state *rew_rsp;
 547         unsigned long rew_s;
 548         struct work_struct rew_work;
 549 };
 550
 551 /*
 552  * Common code to drive an expedited grace period forward, used by
 553  * workqueues and mid-boot-time tasks.
 554  */
 555 static void rcu_exp_sel_wait_wake(struct rcu_state *rsp,
 556                                   smp_call_func_t func, unsigned long s)
 557 {
 558         /* Initialize the rcu_node tree in preparation for the wait. */
 559         sync_rcu_exp_select_cpus(rsp, func);
 560
 561         /* Wait and clean up, including waking everyone. */
 562         rcu_exp_wait_wake(rsp, s);
 563 }
 564
 565 /*
 566  * Work-queue handler to drive an expedited grace period forward.
 567  */
 568 static void wait_rcu_exp_gp(struct work_struct *wp)
 569 {
 570         struct rcu_exp_work *rewp;
 571
 572         rewp = container_of(wp, struct rcu_exp_work, rew_work);
 573         rcu_exp_sel_wait_wake(rewp->rew_rsp, rewp->rew_func, rewp->rew_s);
 574 }
 575
 576 /*
 577  * Given an rcu_state pointer and a smp_call_function() handler, kick
 578  * off the specified flavor of expedited grace period.
 579  */
 580 static void _synchronize_rcu_expedited(struct rcu_state *rsp,
 581                                        smp_call_func_t func)
 582 {
 583         struct rcu_data *rdp;
 584         struct rcu_exp_work rew;
 585         struct rcu_node *rnp;
 586         unsigned long s;
 587
 588         /* If expedited grace periods are prohibited, fall back to normal. */
 589         if (rcu_gp_is_normal()) {
 590                 wait_rcu_gp(rsp->call);
 591                 return;
 592         }
 593
 594         /* Take a snapshot of the sequence number.  */
 595         s = rcu_exp_gp_seq_snap(rsp);
 596         if (exp_funnel_lock(rsp, s))
 597                 return;  /* Someone else did our work for us. */
 598
 599         /* Ensure that load happens before action based on it. */
 600         if (unlikely(rcu_scheduler_active == RCU_SCHEDULER_INIT)) {
 601                 /* Direct call during scheduler init and early_initcalls(). */
 602                 rcu_exp_sel_wait_wake(rsp, func, s);
 603         } else {
 604                 /* Marshall arguments & schedule the expedited grace period. */
 605                 rew.rew_func = func;
 606                 rew.rew_rsp = rsp;
 607                 rew.rew_s = s;
 608                 INIT_WORK_ONSTACK(&rew.rew_work, wait_rcu_exp_gp);
 609                 schedule_work(&rew.rew_work);
 610         }
 611
 612         /* Wait for expedited grace period to complete. */
 613         rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id());
 614         rnp = rcu_get_root(rsp);
 615         wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
 616                    sync_exp_work_done(rsp, &rdp->exp_workdone0, s));
 617         smp_mb(); /* Workqueue actions happen before return. */
 618
 619         /* Let the next expedited grace period start. */
 620         mutex_unlock(&rsp->exp_mutex);
 621 }
 622
 623 /**
 624  * synchronize_sched_expedited - Brute-force RCU-sched grace period
 625  *
 626  * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
 627  * approach to force the grace period to end quickly.  This consumes
 628  * significant time on all CPUs and is unfriendly to real-time workloads,
 629  * so is thus not recommended for any sort of common-case code.  In fact,
 630  * if you are using synchronize_sched_expedited() in a loop, please
 631  * restructure your code to batch your updates, and then use a single
 632  * synchronize_sched() instead.
 633  *
 634  * This implementation can be thought of as an application of sequence
 635  * locking to expedited grace periods, but using the sequence counter to
 636  * determine when someone else has already done the work instead of for
 637  * retrying readers.
 638  */
 639 void synchronize_sched_expedited(void)
 640 {
 641         struct rcu_state *rsp = &rcu_sched_state;
 642
 643         RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
 644                          lock_is_held(&rcu_lock_map) ||
 645                          lock_is_held(&rcu_sched_lock_map),
 646                          "Illegal synchronize_sched_expedited() in RCU read-side critical section");
 647
 648         /* If only one CPU, this is automatically a grace period. */
 649         if (rcu_blocking_is_gp())
 650                 return;
 651
 652         _synchronize_rcu_expedited(rsp, sync_sched_exp_handler);
 653 }
 654 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
 655
 656 #ifdef CONFIG_PREEMPT_RCU
 657
 658 /*
 659  * Remote handler for smp_call_function_single().  If there is an
 660  * RCU read-side critical section in effect, request that the
 661  * next rcu_read_unlock() record the quiescent state up the
 662  * ->expmask fields in the rcu_node tree.  Otherwise, immediately
 663  * report the quiescent state.
 664  */
 665 static void sync_rcu_exp_handler(void *info)
 666 {
 667         struct rcu_data *rdp;
 668         struct rcu_state *rsp = info;
 669         struct task_struct *t = current;
 670
 671         /*
 672          * Within an RCU read-side critical section, request that the next
 673          * rcu_read_unlock() report.  Unless this RCU read-side critical
 674          * section has already blocked, in which case it is already set
 675          * up for the expedited grace period to wait on it.
 676          */
 677         if (t->rcu_read_lock_nesting > 0 &&
 678             !t->rcu_read_unlock_special.b.blocked) {
 679                 t->rcu_read_unlock_special.b.exp_need_qs = true;
 680                 return;
 681         }
 682
 683         /*
 684          * We are either exiting an RCU read-side critical section (negative
 685          * values of t->rcu_read_lock_nesting) or are not in one at all
 686          * (zero value of t->rcu_read_lock_nesting).  Or we are in an RCU
 687          * read-side critical section that blocked before this expedited
 688          * grace period started.  Either way, we can immediately report
 689          * the quiescent state.
 690          */
 691         rdp = this_cpu_ptr(rsp->rda);
 692         rcu_report_exp_rdp(rsp, rdp, true);
 693 }
 694
 695 /**
 696  * synchronize_rcu_expedited - Brute-force RCU grace period
 697  *
 698  * Wait for an RCU-preempt grace period, but expedite it.  The basic
 699  * idea is to IPI all non-idle non-nohz online CPUs.  The IPI handler
 700  * checks whether the CPU is in an RCU-preempt critical section, and
 701  * if so, it sets a flag that causes the outermost rcu_read_unlock()
 702  * to report the quiescent state.  On the other hand, if the CPU is
 703  * not in an RCU read-side critical section, the IPI handler reports
 704  * the quiescent state immediately.
 705  *
 706  * Although this is a greate improvement over previous expedited
 707  * implementations, it is still unfriendly to real-time workloads, so is
 708  * thus not recommended for any sort of common-case code.  In fact, if
 709  * you are using synchronize_rcu_expedited() in a loop, please restructure
 710  * your code to batch your updates, and then Use a single synchronize_rcu()
 711  * instead.
 712  */
 713 void synchronize_rcu_expedited(void)
 714 {
 715         struct rcu_state *rsp = rcu_state_p;
 716
 717         RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
 718                          lock_is_held(&rcu_lock_map) ||
 719                          lock_is_held(&rcu_sched_lock_map),
 720                          "Illegal synchronize_rcu_expedited() in RCU read-side critical section");
 721
 722         if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
 723                 return;
 724         _synchronize_rcu_expedited(rsp, sync_rcu_exp_handler);
 725 }
 726 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
 727
 728 #else /* #ifdef CONFIG_PREEMPT_RCU */
 729
 730 /*
 731  * Wait for an rcu-preempt grace period, but make it happen quickly.
 732  * But because preemptible RCU does not exist, map to rcu-sched.
 733  */
 734 void synchronize_rcu_expedited(void)
 735 {
 736         synchronize_sched_expedited();
 737 }
 738 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
 739
 740 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */