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  * Return then value that expedited-grace-period counter will have
  33  * at the end of the current grace period.
  34  */
  35 static __maybe_unused unsigned long rcu_exp_gp_seq_endval(struct rcu_state *rsp)
  36 {
  37         return rcu_seq_endval(&rsp->expedited_sequence);
  38 }
  39
  40 /*
  41  * Record the end of an expedited grace period.
  42  */
  43 static void rcu_exp_gp_seq_end(struct rcu_state *rsp)
  44 {
  45         rcu_seq_end(&rsp->expedited_sequence);
  46         smp_mb(); /* Ensure that consecutive grace periods serialize. */
  47 }
  48
  49 /*
  50  * Take a snapshot of the expedited-grace-period counter.
  51  */
  52 static unsigned long rcu_exp_gp_seq_snap(struct rcu_state *rsp)
  53 {
  54         unsigned long s;
  55
  56         smp_mb(); /* Caller's modifications seen first by other CPUs. */
  57         s = rcu_seq_snap(&rsp->expedited_sequence);
  58         trace_rcu_exp_grace_period(rsp->name, s, TPS("snap"));
  59         return s;
  60 }
  61
  62 /*
  63  * Given a counter snapshot from rcu_exp_gp_seq_snap(), return true
  64  * if a full expedited grace period has elapsed since that snapshot
  65  * was taken.
  66  */
  67 static bool rcu_exp_gp_seq_done(struct rcu_state *rsp, unsigned long s)
  68 {
  69         return rcu_seq_done(&rsp->expedited_sequence, s);
  70 }
  71
  72 /*
  73  * Reset the ->expmaskinit values in the rcu_node tree to reflect any
  74  * recent CPU-online activity.  Note that these masks are not cleared
  75  * when CPUs go offline, so they reflect the union of all CPUs that have
  76  * ever been online.  This means that this function normally takes its
  77  * no-work-to-do fastpath.
  78  */
  79 static void sync_exp_reset_tree_hotplug(struct rcu_state *rsp)
  80 {
  81         bool done;
  82         unsigned long flags;
  83         unsigned long mask;
  84         unsigned long oldmask;
  85         int ncpus = smp_load_acquire(&rsp->ncpus); /* Order against locking. */
  86         struct rcu_node *rnp;
  87         struct rcu_node *rnp_up;
  88
  89         /* If no new CPUs onlined since last time, nothing to do. */
  90         if (likely(ncpus == rsp->ncpus_snap))
  91                 return;
  92         rsp->ncpus_snap = ncpus;
  93
  94         /*
  95          * Each pass through the following loop propagates newly onlined
  96          * CPUs for the current rcu_node structure up the rcu_node tree.
  97          */
  98         rcu_for_each_leaf_node(rsp, rnp) {
  99                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 100                 if (rnp->expmaskinit == rnp->expmaskinitnext) {
 101                         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 102                         continue;  /* No new CPUs, nothing to do. */
 103                 }
 104
 105                 /* Update this node's mask, track old value for propagation. */
 106                 oldmask = rnp->expmaskinit;
 107                 rnp->expmaskinit = rnp->expmaskinitnext;
 108                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 109
 110                 /* If was already nonzero, nothing to propagate. */
 111                 if (oldmask)
 112                         continue;
 113
 114                 /* Propagate the new CPU up the tree. */
 115                 mask = rnp->grpmask;
 116                 rnp_up = rnp->parent;
 117                 done = false;
 118                 while (rnp_up) {
 119                         raw_spin_lock_irqsave_rcu_node(rnp_up, flags);
 120                         if (rnp_up->expmaskinit)
 121                                 done = true;
 122                         rnp_up->expmaskinit |= mask;
 123                         raw_spin_unlock_irqrestore_rcu_node(rnp_up, flags);
 124                         if (done)
 125                                 break;
 126                         mask = rnp_up->grpmask;
 127                         rnp_up = rnp_up->parent;
 128                 }
 129         }
 130 }
 131
 132 /*
 133  * Reset the ->expmask values in the rcu_node tree in preparation for
 134  * a new expedited grace period.
 135  */
 136 static void __maybe_unused sync_exp_reset_tree(struct rcu_state *rsp)
 137 {
 138         unsigned long flags;
 139         struct rcu_node *rnp;
 140
 141         sync_exp_reset_tree_hotplug(rsp);
 142         rcu_for_each_node_breadth_first(rsp, rnp) {
 143                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 144                 WARN_ON_ONCE(rnp->expmask);
 145                 rnp->expmask = rnp->expmaskinit;
 146                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 147         }
 148 }
 149
 150 /*
 151  * Return non-zero if there is no RCU expedited grace period in progress
 152  * for the specified rcu_node structure, in other words, if all CPUs and
 153  * tasks covered by the specified rcu_node structure have done their bit
 154  * for the current expedited grace period.  Works only for preemptible
 155  * RCU -- other RCU implementation use other means.
 156  *
 157  * Caller must hold the rcu_state's exp_mutex.
 158  */
 159 static bool sync_rcu_preempt_exp_done(struct rcu_node *rnp)
 160 {
 161         return rnp->exp_tasks == NULL &&
 162                READ_ONCE(rnp->expmask) == 0;
 163 }
 164
 165 /*
 166  * Report the exit from RCU read-side critical section for the last task
 167  * that queued itself during or before the current expedited preemptible-RCU
 168  * grace period.  This event is reported either to the rcu_node structure on
 169  * which the task was queued or to one of that rcu_node structure's ancestors,
 170  * recursively up the tree.  (Calm down, calm down, we do the recursion
 171  * iteratively!)
 172  *
 173  * Caller must hold the rcu_state's exp_mutex and the specified rcu_node
 174  * structure's ->lock.
 175  */
 176 static void __rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
 177                                  bool wake, unsigned long flags)
 178         __releases(rnp->lock)
 179 {
 180         unsigned long mask;
 181
 182         for (;;) {
 183                 if (!sync_rcu_preempt_exp_done(rnp)) {
 184                         if (!rnp->expmask)
 185                                 rcu_initiate_boost(rnp, flags);
 186                         else
 187                                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 188                         break;
 189                 }
 190                 if (rnp->parent == NULL) {
 191                         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 192                         if (wake) {
 193                                 smp_mb(); /* EGP done before wake_up(). */
 194                                 swake_up(&rsp->expedited_wq);
 195                         }
 196                         break;
 197                 }
 198                 mask = rnp->grpmask;
 199                 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled */
 200                 rnp = rnp->parent;
 201                 raw_spin_lock_rcu_node(rnp); /* irqs already disabled */
 202                 WARN_ON_ONCE(!(rnp->expmask & mask));
 203                 rnp->expmask &= ~mask;
 204         }
 205 }
 206
 207 /*
 208  * Report expedited quiescent state for specified node.  This is a
 209  * lock-acquisition wrapper function for __rcu_report_exp_rnp().
 210  *
 211  * Caller must hold the rcu_state's exp_mutex.
 212  */
 213 static void __maybe_unused rcu_report_exp_rnp(struct rcu_state *rsp,
 214                                               struct rcu_node *rnp, bool wake)
 215 {
 216         unsigned long flags;
 217
 218         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 219         __rcu_report_exp_rnp(rsp, rnp, wake, flags);
 220 }
 221
 222 /*
 223  * Report expedited quiescent state for multiple CPUs, all covered by the
 224  * specified leaf rcu_node structure.  Caller must hold the rcu_state's
 225  * exp_mutex.
 226  */
 227 static void rcu_report_exp_cpu_mult(struct rcu_state *rsp, struct rcu_node *rnp,
 228                                     unsigned long mask, bool wake)
 229 {
 230         unsigned long flags;
 231
 232         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 233         if (!(rnp->expmask & mask)) {
 234                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 235                 return;
 236         }
 237         rnp->expmask &= ~mask;
 238         __rcu_report_exp_rnp(rsp, rnp, wake, flags); /* Releases rnp->lock. */
 239 }
 240
 241 /*
 242  * Report expedited quiescent state for specified rcu_data (CPU).
 243  */
 244 static void rcu_report_exp_rdp(struct rcu_state *rsp, struct rcu_data *rdp,
 245                                bool wake)
 246 {
 247         rcu_report_exp_cpu_mult(rsp, rdp->mynode, rdp->grpmask, wake);
 248 }
 249
 250 /* Common code for synchronize_{rcu,sched}_expedited() work-done checking. */
 251 static bool sync_exp_work_done(struct rcu_state *rsp, atomic_long_t *stat,
 252                                unsigned long s)
 253 {
 254         if (rcu_exp_gp_seq_done(rsp, s)) {
 255                 trace_rcu_exp_grace_period(rsp->name, s, TPS("done"));
 256                 /* Ensure test happens before caller kfree(). */
 257                 smp_mb__before_atomic(); /* ^^^ */
 258                 atomic_long_inc(stat);
 259                 return true;
 260         }
 261         return false;
 262 }
 263
 264 /*
 265  * Funnel-lock acquisition for expedited grace periods.  Returns true
 266  * if some other task completed an expedited grace period that this task
 267  * can piggy-back on, and with no mutex held.  Otherwise, returns false
 268  * with the mutex held, indicating that the caller must actually do the
 269  * expedited grace period.
 270  */
 271 static bool exp_funnel_lock(struct rcu_state *rsp, unsigned long s)
 272 {
 273         struct rcu_data *rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id());
 274         struct rcu_node *rnp = rdp->mynode;
 275         struct rcu_node *rnp_root = rcu_get_root(rsp);
 276
 277         /* Low-contention fastpath. */
 278         if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s) &&
 279             (rnp == rnp_root ||
 280              ULONG_CMP_LT(READ_ONCE(rnp_root->exp_seq_rq), s)) &&
 281             mutex_trylock(&rsp->exp_mutex))
 282                 goto fastpath;
 283
 284         /*
 285          * Each pass through the following loop works its way up
 286          * the rcu_node tree, returning if others have done the work or
 287          * otherwise falls through to acquire rsp->exp_mutex.  The mapping
 288          * from CPU to rcu_node structure can be inexact, as it is just
 289          * promoting locality and is not strictly needed for correctness.
 290          */
 291         for (; rnp != NULL; rnp = rnp->parent) {
 292                 if (sync_exp_work_done(rsp, &rdp->exp_workdone1, s))
 293                         return true;
 294
 295                 /* Work not done, either wait here or go up. */
 296                 spin_lock(&rnp->exp_lock);
 297                 if (ULONG_CMP_GE(rnp->exp_seq_rq, s)) {
 298
 299                         /* Someone else doing GP, so wait for them. */
 300                         spin_unlock(&rnp->exp_lock);
 301                         trace_rcu_exp_funnel_lock(rsp->name, rnp->level,
 302                                                   rnp->grplo, rnp->grphi,
 303                                                   TPS("wait"));
 304                         wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
 305                                    sync_exp_work_done(rsp,
 306                                                       &rdp->exp_workdone2, s));
 307                         return true;
 308                 }
 309                 rnp->exp_seq_rq = s; /* Followers can wait on us. */
 310                 spin_unlock(&rnp->exp_lock);
 311                 trace_rcu_exp_funnel_lock(rsp->name, rnp->level, rnp->grplo,
 312                                           rnp->grphi, TPS("nxtlvl"));
 313         }
 314         mutex_lock(&rsp->exp_mutex);
 315 fastpath:
 316         if (sync_exp_work_done(rsp, &rdp->exp_workdone3, s)) {
 317                 mutex_unlock(&rsp->exp_mutex);
 318                 return true;
 319         }
 320         rcu_exp_gp_seq_start(rsp);
 321         trace_rcu_exp_grace_period(rsp->name, s, TPS("start"));
 322         return false;
 323 }
 324
 325 /* Invoked on each online non-idle CPU for expedited quiescent state. */
 326 static void sync_sched_exp_handler(void *data)
 327 {
 328         struct rcu_data *rdp;
 329         struct rcu_node *rnp;
 330         struct rcu_state *rsp = data;
 331
 332         rdp = this_cpu_ptr(rsp->rda);
 333         rnp = rdp->mynode;
 334         if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
 335             __this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
 336                 return;
 337         if (rcu_is_cpu_rrupt_from_idle()) {
 338                 rcu_report_exp_rdp(&rcu_sched_state,
 339                                    this_cpu_ptr(&rcu_sched_data), true);
 340                 return;
 341         }
 342         __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, true);
 343         /* Store .exp before .rcu_urgent_qs. */
 344         smp_store_release(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs), true);
 345         resched_cpu(smp_processor_id());
 346 }
 347
 348 /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
 349 static void sync_sched_exp_online_cleanup(int cpu)
 350 {
 351         struct rcu_data *rdp;
 352         int ret;
 353         struct rcu_node *rnp;
 354         struct rcu_state *rsp = &rcu_sched_state;
 355
 356         rdp = per_cpu_ptr(rsp->rda, cpu);
 357         rnp = rdp->mynode;
 358         if (!(READ_ONCE(rnp->expmask) & rdp->grpmask))
 359                 return;
 360         ret = smp_call_function_single(cpu, sync_sched_exp_handler, rsp, 0);
 361         WARN_ON_ONCE(ret);
 362 }
 363
 364 /*
 365  * Select the nodes that the upcoming expedited grace period needs
 366  * to wait for.
 367  */
 368 static void sync_rcu_exp_select_cpus(struct rcu_state *rsp,
 369                                      smp_call_func_t func)
 370 {
 371         int cpu;
 372         unsigned long flags;
 373         unsigned long mask_ofl_test;
 374         unsigned long mask_ofl_ipi;
 375         int ret;
 376         struct rcu_node *rnp;
 377
 378         trace_rcu_exp_grace_period(rsp->name, rcu_exp_gp_seq_endval(rsp), TPS("reset"));
 379         sync_exp_reset_tree(rsp);
 380         trace_rcu_exp_grace_period(rsp->name, rcu_exp_gp_seq_endval(rsp), TPS("select"));
 381         rcu_for_each_leaf_node(rsp, rnp) {
 382                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 383
 384                 /* Each pass checks a CPU for identity, offline, and idle. */
 385                 mask_ofl_test = 0;
 386                 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
 387                         unsigned long mask = leaf_node_cpu_bit(rnp, cpu);
 388                         struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
 389                         struct rcu_dynticks *rdtp = per_cpu_ptr(&rcu_dynticks, cpu);
 390                         int snap;
 391
 392                         if (raw_smp_processor_id() == cpu ||
 393                             !(rnp->qsmaskinitnext & mask)) {
 394                                 mask_ofl_test |= mask;
 395                         } else {
 396                                 snap = rcu_dynticks_snap(rdtp);
 397                                 if (rcu_dynticks_in_eqs(snap))
 398                                         mask_ofl_test |= mask;
 399                                 else
 400                                         rdp->exp_dynticks_snap = snap;
 401                         }
 402                 }
 403                 mask_ofl_ipi = rnp->expmask & ~mask_ofl_test;
 404
 405                 /*
 406                  * Need to wait for any blocked tasks as well.  Note that
 407                  * additional blocking tasks will also block the expedited
 408                  * GP until such time as the ->expmask bits are cleared.
 409                  */
 410                 if (rcu_preempt_has_tasks(rnp))
 411                         rnp->exp_tasks = rnp->blkd_tasks.next;
 412                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 413
 414                 /* IPI the remaining CPUs for expedited quiescent state. */
 415                 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
 416                         unsigned long mask = leaf_node_cpu_bit(rnp, cpu);
 417                         struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
 418
 419                         if (!(mask_ofl_ipi & mask))
 420                                 continue;
 421 retry_ipi:
 422                         if (rcu_dynticks_in_eqs_since(rdp->dynticks,
 423                                                       rdp->exp_dynticks_snap)) {
 424                                 mask_ofl_test |= mask;
 425                                 continue;
 426                         }
 427                         ret = smp_call_function_single(cpu, func, rsp, 0);
 428                         if (!ret) {
 429                                 mask_ofl_ipi &= ~mask;
 430                                 continue;
 431                         }
 432                         /* Failed, raced with CPU hotplug operation. */
 433                         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 434                         if ((rnp->qsmaskinitnext & mask) &&
 435                             (rnp->expmask & mask)) {
 436                                 /* Online, so delay for a bit and try again. */
 437                                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 438                                 trace_rcu_exp_grace_period(rsp->name, rcu_exp_gp_seq_endval(rsp), TPS("selectofl"));
 439                                 schedule_timeout_uninterruptible(1);
 440                                 goto retry_ipi;
 441                         }
 442                         /* CPU really is offline, so we can ignore it. */
 443                         if (!(rnp->expmask & mask))
 444                                 mask_ofl_ipi &= ~mask;
 445                         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 446                 }
 447                 /* Report quiescent states for those that went offline. */
 448                 mask_ofl_test |= mask_ofl_ipi;
 449                 if (mask_ofl_test)
 450                         rcu_report_exp_cpu_mult(rsp, rnp, mask_ofl_test, false);
 451         }
 452 }
 453
 454 static void synchronize_sched_expedited_wait(struct rcu_state *rsp)
 455 {
 456         int cpu;
 457         unsigned long jiffies_stall;
 458         unsigned long jiffies_start;
 459         unsigned long mask;
 460         int ndetected;
 461         struct rcu_node *rnp;
 462         struct rcu_node *rnp_root = rcu_get_root(rsp);
 463         int ret;
 464
 465         trace_rcu_exp_grace_period(rsp->name, rcu_exp_gp_seq_endval(rsp), TPS("startwait"));
 466         jiffies_stall = rcu_jiffies_till_stall_check();
 467         jiffies_start = jiffies;
 468
 469         for (;;) {
 470                 ret = swait_event_timeout(
 471                                 rsp->expedited_wq,
 472                                 sync_rcu_preempt_exp_done(rnp_root),
 473                                 jiffies_stall);
 474                 if (ret > 0 || sync_rcu_preempt_exp_done(rnp_root))
 475                         return;
 476                 WARN_ON(ret < 0);  /* workqueues should not be signaled. */
 477                 if (rcu_cpu_stall_suppress)
 478                         continue;
 479                 panic_on_rcu_stall();
 480                 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
 481                        rsp->name);
 482                 ndetected = 0;
 483                 rcu_for_each_leaf_node(rsp, rnp) {
 484                         ndetected += rcu_print_task_exp_stall(rnp);
 485                         for_each_leaf_node_possible_cpu(rnp, cpu) {
 486                                 struct rcu_data *rdp;
 487
 488                                 mask = leaf_node_cpu_bit(rnp, cpu);
 489                                 if (!(rnp->expmask & mask))
 490                                         continue;
 491                                 ndetected++;
 492                                 rdp = per_cpu_ptr(rsp->rda, cpu);
 493                                 pr_cont(" %d-%c%c%c", cpu,
 494                                         "O."[!!cpu_online(cpu)],
 495                                         "o."[!!(rdp->grpmask & rnp->expmaskinit)],
 496                                         "N."[!!(rdp->grpmask & rnp->expmaskinitnext)]);
 497                         }
 498                 }
 499                 pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
 500                         jiffies - jiffies_start, rsp->expedited_sequence,
 501                         rnp_root->expmask, ".T"[!!rnp_root->exp_tasks]);
 502                 if (ndetected) {
 503                         pr_err("blocking rcu_node structures:");
 504                         rcu_for_each_node_breadth_first(rsp, rnp) {
 505                                 if (rnp == rnp_root)
 506                                         continue; /* printed unconditionally */
 507                                 if (sync_rcu_preempt_exp_done(rnp))
 508                                         continue;
 509                                 pr_cont(" l=%u:%d-%d:%#lx/%c",
 510                                         rnp->level, rnp->grplo, rnp->grphi,
 511                                         rnp->expmask,
 512                                         ".T"[!!rnp->exp_tasks]);
 513                         }
 514                         pr_cont("\n");
 515                 }
 516                 rcu_for_each_leaf_node(rsp, rnp) {
 517                         for_each_leaf_node_possible_cpu(rnp, cpu) {
 518                                 mask = leaf_node_cpu_bit(rnp, cpu);
 519                                 if (!(rnp->expmask & mask))
 520                                         continue;
 521                                 dump_cpu_task(cpu);
 522                         }
 523                 }
 524                 jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3;
 525         }
 526 }
 527
 528 /*
 529  * Wait for the current expedited grace period to complete, and then
 530  * wake up everyone who piggybacked on the just-completed expedited
 531  * grace period.  Also update all the ->exp_seq_rq counters as needed
 532  * in order to avoid counter-wrap problems.
 533  */
 534 static void rcu_exp_wait_wake(struct rcu_state *rsp, unsigned long s)
 535 {
 536         struct rcu_node *rnp;
 537
 538         synchronize_sched_expedited_wait(rsp);
 539         rcu_exp_gp_seq_end(rsp);
 540         trace_rcu_exp_grace_period(rsp->name, s, TPS("end"));
 541
 542         /*
 543          * Switch over to wakeup mode, allowing the next GP, but -only- the
 544          * next GP, to proceed.
 545          */
 546         mutex_lock(&rsp->exp_wake_mutex);
 547
 548         rcu_for_each_node_breadth_first(rsp, rnp) {
 549                 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s)) {
 550                         spin_lock(&rnp->exp_lock);
 551                         /* Recheck, avoid hang in case someone just arrived. */
 552                         if (ULONG_CMP_LT(rnp->exp_seq_rq, s))
 553                                 rnp->exp_seq_rq = s;
 554                         spin_unlock(&rnp->exp_lock);
 555                 }
 556                 smp_mb(); /* All above changes before wakeup. */
 557                 wake_up_all(&rnp->exp_wq[rcu_seq_ctr(rsp->expedited_sequence) & 0x3]);
 558         }
 559         trace_rcu_exp_grace_period(rsp->name, s, TPS("endwake"));
 560         mutex_unlock(&rsp->exp_wake_mutex);
 561 }
 562
 563 /* Let the workqueue handler know what it is supposed to do. */
 564 struct rcu_exp_work {
 565         smp_call_func_t rew_func;
 566         struct rcu_state *rew_rsp;
 567         unsigned long rew_s;
 568         struct work_struct rew_work;
 569 };
 570
 571 /*
 572  * Common code to drive an expedited grace period forward, used by
 573  * workqueues and mid-boot-time tasks.
 574  */
 575 static void rcu_exp_sel_wait_wake(struct rcu_state *rsp,
 576                                   smp_call_func_t func, unsigned long s)
 577 {
 578         /* Initialize the rcu_node tree in preparation for the wait. */
 579         sync_rcu_exp_select_cpus(rsp, func);
 580
 581         /* Wait and clean up, including waking everyone. */
 582         rcu_exp_wait_wake(rsp, s);
 583 }
 584
 585 /*
 586  * Work-queue handler to drive an expedited grace period forward.
 587  */
 588 static void wait_rcu_exp_gp(struct work_struct *wp)
 589 {
 590         struct rcu_exp_work *rewp;
 591
 592         rewp = container_of(wp, struct rcu_exp_work, rew_work);
 593         rcu_exp_sel_wait_wake(rewp->rew_rsp, rewp->rew_func, rewp->rew_s);
 594 }
 595
 596 /*
 597  * Given an rcu_state pointer and a smp_call_function() handler, kick
 598  * off the specified flavor of expedited grace period.
 599  */
 600 static void _synchronize_rcu_expedited(struct rcu_state *rsp,
 601                                        smp_call_func_t func)
 602 {
 603         struct rcu_data *rdp;
 604         struct rcu_exp_work rew;
 605         struct rcu_node *rnp;
 606         unsigned long s;
 607
 608         /* If expedited grace periods are prohibited, fall back to normal. */
 609         if (rcu_gp_is_normal()) {
 610                 wait_rcu_gp(rsp->call);
 611                 return;
 612         }
 613
 614         /* Take a snapshot of the sequence number.  */
 615         s = rcu_exp_gp_seq_snap(rsp);
 616         if (exp_funnel_lock(rsp, s))
 617                 return;  /* Someone else did our work for us. */
 618
 619         /* Ensure that load happens before action based on it. */
 620         if (unlikely(rcu_scheduler_active == RCU_SCHEDULER_INIT)) {
 621                 /* Direct call during scheduler init and early_initcalls(). */
 622                 rcu_exp_sel_wait_wake(rsp, func, s);
 623         } else {
 624                 /* Marshall arguments & schedule the expedited grace period. */
 625                 rew.rew_func = func;
 626                 rew.rew_rsp = rsp;
 627                 rew.rew_s = s;
 628                 INIT_WORK_ONSTACK(&rew.rew_work, wait_rcu_exp_gp);
 629                 queue_work(rcu_gp_wq, &rew.rew_work);
 630         }
 631
 632         /* Wait for expedited grace period to complete. */
 633         rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id());
 634         rnp = rcu_get_root(rsp);
 635         wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
 636                    sync_exp_work_done(rsp, &rdp->exp_workdone0, s));
 637         smp_mb(); /* Workqueue actions happen before return. */
 638
 639         /* Let the next expedited grace period start. */
 640         mutex_unlock(&rsp->exp_mutex);
 641 }
 642
 643 /**
 644  * synchronize_sched_expedited - Brute-force RCU-sched grace period
 645  *
 646  * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
 647  * approach to force the grace period to end quickly.  This consumes
 648  * significant time on all CPUs and is unfriendly to real-time workloads,
 649  * so is thus not recommended for any sort of common-case code.  In fact,
 650  * if you are using synchronize_sched_expedited() in a loop, please
 651  * restructure your code to batch your updates, and then use a single
 652  * synchronize_sched() instead.
 653  *
 654  * This implementation can be thought of as an application of sequence
 655  * locking to expedited grace periods, but using the sequence counter to
 656  * determine when someone else has already done the work instead of for
 657  * retrying readers.
 658  */
 659 void synchronize_sched_expedited(void)
 660 {
 661         struct rcu_state *rsp = &rcu_sched_state;
 662
 663         RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
 664                          lock_is_held(&rcu_lock_map) ||
 665                          lock_is_held(&rcu_sched_lock_map),
 666                          "Illegal synchronize_sched_expedited() in RCU read-side critical section");
 667
 668         /* If only one CPU, this is automatically a grace period. */
 669         if (rcu_blocking_is_gp())
 670                 return;
 671
 672         _synchronize_rcu_expedited(rsp, sync_sched_exp_handler);
 673 }
 674 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
 675
 676 #ifdef CONFIG_PREEMPT_RCU
 677
 678 /*
 679  * Remote handler for smp_call_function_single().  If there is an
 680  * RCU read-side critical section in effect, request that the
 681  * next rcu_read_unlock() record the quiescent state up the
 682  * ->expmask fields in the rcu_node tree.  Otherwise, immediately
 683  * report the quiescent state.
 684  */
 685 static void sync_rcu_exp_handler(void *info)
 686 {
 687         struct rcu_data *rdp;
 688         struct rcu_state *rsp = info;
 689         struct task_struct *t = current;
 690
 691         /*
 692          * Within an RCU read-side critical section, request that the next
 693          * rcu_read_unlock() report.  Unless this RCU read-side critical
 694          * section has already blocked, in which case it is already set
 695          * up for the expedited grace period to wait on it.
 696          */
 697         if (t->rcu_read_lock_nesting > 0 &&
 698             !t->rcu_read_unlock_special.b.blocked) {
 699                 t->rcu_read_unlock_special.b.exp_need_qs = true;
 700                 return;
 701         }
 702
 703         /*
 704          * We are either exiting an RCU read-side critical section (negative
 705          * values of t->rcu_read_lock_nesting) or are not in one at all
 706          * (zero value of t->rcu_read_lock_nesting).  Or we are in an RCU
 707          * read-side critical section that blocked before this expedited
 708          * grace period started.  Either way, we can immediately report
 709          * the quiescent state.
 710          */
 711         rdp = this_cpu_ptr(rsp->rda);
 712         rcu_report_exp_rdp(rsp, rdp, true);
 713 }
 714
 715 /**
 716  * synchronize_rcu_expedited - Brute-force RCU grace period
 717  *
 718  * Wait for an RCU-preempt grace period, but expedite it.  The basic
 719  * idea is to IPI all non-idle non-nohz online CPUs.  The IPI handler
 720  * checks whether the CPU is in an RCU-preempt critical section, and
 721  * if so, it sets a flag that causes the outermost rcu_read_unlock()
 722  * to report the quiescent state.  On the other hand, if the CPU is
 723  * not in an RCU read-side critical section, the IPI handler reports
 724  * the quiescent state immediately.
 725  *
 726  * Although this is a greate improvement over previous expedited
 727  * implementations, it is still unfriendly to real-time workloads, so is
 728  * thus not recommended for any sort of common-case code.  In fact, if
 729  * you are using synchronize_rcu_expedited() in a loop, please restructure
 730  * your code to batch your updates, and then Use a single synchronize_rcu()
 731  * instead.
 732  */
 733 void synchronize_rcu_expedited(void)
 734 {
 735         struct rcu_state *rsp = rcu_state_p;
 736
 737         RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
 738                          lock_is_held(&rcu_lock_map) ||
 739                          lock_is_held(&rcu_sched_lock_map),
 740                          "Illegal synchronize_rcu_expedited() in RCU read-side critical section");
 741
 742         if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
 743                 return;
 744         _synchronize_rcu_expedited(rsp, sync_rcu_exp_handler);
 745 }
 746 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
 747
 748 #else /* #ifdef CONFIG_PREEMPT_RCU */
 749
 750 /*
 751  * Wait for an rcu-preempt grace period, but make it happen quickly.
 752  * But because preemptible RCU does not exist, map to rcu-sched.
 753  */
 754 void synchronize_rcu_expedited(void)
 755 {
 756         synchronize_sched_expedited();
 757 }
 758 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
 759
 760 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */