kernel/rcu/tree_exp.h

   1 /* SPDX-License-Identifier: GPL-2.0+ */
   2 /*
   3  * RCU expedited grace periods
   4  *
   5  * Copyright IBM Corporation, 2016
   6  *
   7  * Authors: Paul E. McKenney <paulmck@linux.ibm.com>
   8  */
   9
  10 #include <linux/lockdep.h>
  11
  12 static void rcu_exp_handler(void *unused);
  13 static int rcu_print_task_exp_stall(struct rcu_node *rnp);
  14
  15 /*
  16  * Record the start of an expedited grace period.
  17  */
  18 static void rcu_exp_gp_seq_start(void)
  19 {
  20         rcu_seq_start(&rcu_state.expedited_sequence);
  21 }
  22
  23 /*
  24  * Return the value that the expedited-grace-period counter will have
  25  * at the end of the current grace period.
  26  */
  27 static __maybe_unused unsigned long rcu_exp_gp_seq_endval(void)
  28 {
  29         return rcu_seq_endval(&rcu_state.expedited_sequence);
  30 }
  31
  32 /*
  33  * Record the end of an expedited grace period.
  34  */
  35 static void rcu_exp_gp_seq_end(void)
  36 {
  37         rcu_seq_end(&rcu_state.expedited_sequence);
  38         smp_mb(); /* Ensure that consecutive grace periods serialize. */
  39 }
  40
  41 /*
  42  * Take a snapshot of the expedited-grace-period counter, which is the
  43  * earliest value that will indicate that a full grace period has
  44  * elapsed since the current time.
  45  */
  46 static unsigned long rcu_exp_gp_seq_snap(void)
  47 {
  48         unsigned long s;
  49
  50         smp_mb(); /* Caller's modifications seen first by other CPUs. */
  51         s = rcu_seq_snap(&rcu_state.expedited_sequence);
  52         trace_rcu_exp_grace_period(rcu_state.name, s, TPS("snap"));
  53         return s;
  54 }
  55
  56 /*
  57  * Given a counter snapshot from rcu_exp_gp_seq_snap(), return true
  58  * if a full expedited grace period has elapsed since that snapshot
  59  * was taken.
  60  */
  61 static bool rcu_exp_gp_seq_done(unsigned long s)
  62 {
  63         return rcu_seq_done(&rcu_state.expedited_sequence, s);
  64 }
  65
  66 /*
  67  * Reset the ->expmaskinit values in the rcu_node tree to reflect any
  68  * recent CPU-online activity.  Note that these masks are not cleared
  69  * when CPUs go offline, so they reflect the union of all CPUs that have
  70  * ever been online.  This means that this function normally takes its
  71  * no-work-to-do fastpath.
  72  */
  73 static void sync_exp_reset_tree_hotplug(void)
  74 {
  75         bool done;
  76         unsigned long flags;
  77         unsigned long mask;
  78         unsigned long oldmask;
  79         int ncpus = smp_load_acquire(&rcu_state.ncpus); /* Order vs. locking. */
  80         struct rcu_node *rnp;
  81         struct rcu_node *rnp_up;
  82
  83         /* If no new CPUs onlined since last time, nothing to do. */
  84         if (likely(ncpus == rcu_state.ncpus_snap))
  85                 return;
  86         rcu_state.ncpus_snap = ncpus;
  87
  88         /*
  89          * Each pass through the following loop propagates newly onlined
  90          * CPUs for the current rcu_node structure up the rcu_node tree.
  91          */
  92         rcu_for_each_leaf_node(rnp) {
  93                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
  94                 if (rnp->expmaskinit == rnp->expmaskinitnext) {
  95                         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
  96                         continue;  /* No new CPUs, nothing to do. */
  97                 }
  98
  99                 /* Update this node's mask, track old value for propagation. */
 100                 oldmask = rnp->expmaskinit;
 101                 rnp->expmaskinit = rnp->expmaskinitnext;
 102                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 103
 104                 /* If was already nonzero, nothing to propagate. */
 105                 if (oldmask)
 106                         continue;
 107
 108                 /* Propagate the new CPU up the tree. */
 109                 mask = rnp->grpmask;
 110                 rnp_up = rnp->parent;
 111                 done = false;
 112                 while (rnp_up) {
 113                         raw_spin_lock_irqsave_rcu_node(rnp_up, flags);
 114                         if (rnp_up->expmaskinit)
 115                                 done = true;
 116                         rnp_up->expmaskinit |= mask;
 117                         raw_spin_unlock_irqrestore_rcu_node(rnp_up, flags);
 118                         if (done)
 119                                 break;
 120                         mask = rnp_up->grpmask;
 121                         rnp_up = rnp_up->parent;
 122                 }
 123         }
 124 }
 125
 126 /*
 127  * Reset the ->expmask values in the rcu_node tree in preparation for
 128  * a new expedited grace period.
 129  */
 130 static void __maybe_unused sync_exp_reset_tree(void)
 131 {
 132         unsigned long flags;
 133         struct rcu_node *rnp;
 134
 135         sync_exp_reset_tree_hotplug();
 136         rcu_for_each_node_breadth_first(rnp) {
 137                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 138                 WARN_ON_ONCE(rnp->expmask);
 139                 WRITE_ONCE(rnp->expmask, rnp->expmaskinit);
 140                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 141         }
 142 }
 143
 144 /*
 145  * Return non-zero if there is no RCU expedited grace period in progress
 146  * for the specified rcu_node structure, in other words, if all CPUs and
 147  * tasks covered by the specified rcu_node structure have done their bit
 148  * for the current expedited grace period.
 149  */
 150 static bool sync_rcu_exp_done(struct rcu_node *rnp)
 151 {
 152         raw_lockdep_assert_held_rcu_node(rnp);
 153         return rnp->exp_tasks == NULL &&
 154                READ_ONCE(rnp->expmask) == 0;
 155 }
 156
 157 /*
 158  * Like sync_rcu_exp_done(), but where the caller does not hold the
 159  * rcu_node's ->lock.
 160  */
 161 static bool sync_rcu_exp_done_unlocked(struct rcu_node *rnp)
 162 {
 163         unsigned long flags;
 164         bool ret;
 165
 166         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 167         ret = sync_rcu_exp_done(rnp);
 168         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 169
 170         return ret;
 171 }
 172
 173
 174 /*
 175  * Report the exit from RCU read-side critical section for the last task
 176  * that queued itself during or before the current expedited preemptible-RCU
 177  * grace period.  This event is reported either to the rcu_node structure on
 178  * which the task was queued or to one of that rcu_node structure's ancestors,
 179  * recursively up the tree.  (Calm down, calm down, we do the recursion
 180  * iteratively!)
 181  */
 182 static void __rcu_report_exp_rnp(struct rcu_node *rnp,
 183                                  bool wake, unsigned long flags)
 184         __releases(rnp->lock)
 185 {
 186         unsigned long mask;
 187
 188         raw_lockdep_assert_held_rcu_node(rnp);
 189         for (;;) {
 190                 if (!sync_rcu_exp_done(rnp)) {
 191                         if (!rnp->expmask)
 192                                 rcu_initiate_boost(rnp, flags);
 193                         else
 194                                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 195                         break;
 196                 }
 197                 if (rnp->parent == NULL) {
 198                         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 199                         if (wake) {
 200                                 smp_mb(); /* EGP done before wake_up(). */
 201                                 swake_up_one(&rcu_state.expedited_wq);
 202                         }
 203                         break;
 204                 }
 205                 mask = rnp->grpmask;
 206                 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled */
 207                 rnp = rnp->parent;
 208                 raw_spin_lock_rcu_node(rnp); /* irqs already disabled */
 209                 WARN_ON_ONCE(!(rnp->expmask & mask));
 210                 WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
 211         }
 212 }
 213
 214 /*
 215  * Report expedited quiescent state for specified node.  This is a
 216  * lock-acquisition wrapper function for __rcu_report_exp_rnp().
 217  */
 218 static void __maybe_unused rcu_report_exp_rnp(struct rcu_node *rnp, bool wake)
 219 {
 220         unsigned long flags;
 221
 222         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 223         __rcu_report_exp_rnp(rnp, wake, flags);
 224 }
 225
 226 /*
 227  * Report expedited quiescent state for multiple CPUs, all covered by the
 228  * specified leaf rcu_node structure.
 229  */
 230 static void rcu_report_exp_cpu_mult(struct rcu_node *rnp,
 231                                     unsigned long mask, bool wake)
 232 {
 233         int cpu;
 234         unsigned long flags;
 235         struct rcu_data *rdp;
 236
 237         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 238         if (!(rnp->expmask & mask)) {
 239                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 240                 return;
 241         }
 242         WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
 243         for_each_leaf_node_cpu_mask(rnp, cpu, mask) {
 244                 rdp = per_cpu_ptr(&rcu_data, cpu);
 245                 if (!IS_ENABLED(CONFIG_NO_HZ_FULL) || !rdp->rcu_forced_tick_exp)
 246                         continue;
 247                 rdp->rcu_forced_tick_exp = false;
 248                 tick_dep_clear_cpu(cpu, TICK_DEP_BIT_RCU_EXP);
 249         }
 250         __rcu_report_exp_rnp(rnp, wake, flags); /* Releases rnp->lock. */
 251 }
 252
 253 /*
 254  * Report expedited quiescent state for specified rcu_data (CPU).
 255  */
 256 static void rcu_report_exp_rdp(struct rcu_data *rdp)
 257 {
 258         WRITE_ONCE(rdp->exp_deferred_qs, false);
 259         rcu_report_exp_cpu_mult(rdp->mynode, rdp->grpmask, true);
 260 }
 261
 262 /* Common code for work-done checking. */
 263 static bool sync_exp_work_done(unsigned long s)
 264 {
 265         if (rcu_exp_gp_seq_done(s)) {
 266                 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("done"));
 267                 smp_mb(); /* Ensure test happens before caller kfree(). */
 268                 return true;
 269         }
 270         return false;
 271 }
 272
 273 /*
 274  * Funnel-lock acquisition for expedited grace periods.  Returns true
 275  * if some other task completed an expedited grace period that this task
 276  * can piggy-back on, and with no mutex held.  Otherwise, returns false
 277  * with the mutex held, indicating that the caller must actually do the
 278  * expedited grace period.
 279  */
 280 static bool exp_funnel_lock(unsigned long s)
 281 {
 282         struct rcu_data *rdp = per_cpu_ptr(&rcu_data, raw_smp_processor_id());
 283         struct rcu_node *rnp = rdp->mynode;
 284         struct rcu_node *rnp_root = rcu_get_root();
 285
 286         /* Low-contention fastpath. */
 287         if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s) &&
 288             (rnp == rnp_root ||
 289              ULONG_CMP_LT(READ_ONCE(rnp_root->exp_seq_rq), s)) &&
 290             mutex_trylock(&rcu_state.exp_mutex))
 291                 goto fastpath;
 292
 293         /*
 294          * Each pass through the following loop works its way up
 295          * the rcu_node tree, returning if others have done the work or
 296          * otherwise falls through to acquire ->exp_mutex.  The mapping
 297          * from CPU to rcu_node structure can be inexact, as it is just
 298          * promoting locality and is not strictly needed for correctness.
 299          */
 300         for (; rnp != NULL; rnp = rnp->parent) {
 301                 if (sync_exp_work_done(s))
 302                         return true;
 303
 304                 /* Work not done, either wait here or go up. */
 305                 spin_lock(&rnp->exp_lock);
 306                 if (ULONG_CMP_GE(rnp->exp_seq_rq, s)) {
 307
 308                         /* Someone else doing GP, so wait for them. */
 309                         spin_unlock(&rnp->exp_lock);
 310                         trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
 311                                                   rnp->grplo, rnp->grphi,
 312                                                   TPS("wait"));
 313                         wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
 314                                    sync_exp_work_done(s));
 315                         return true;
 316                 }
 317                 rnp->exp_seq_rq = s; /* Followers can wait on us. */
 318                 spin_unlock(&rnp->exp_lock);
 319                 trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
 320                                           rnp->grplo, rnp->grphi, TPS("nxtlvl"));
 321         }
 322         mutex_lock(&rcu_state.exp_mutex);
 323 fastpath:
 324         if (sync_exp_work_done(s)) {
 325                 mutex_unlock(&rcu_state.exp_mutex);
 326                 return true;
 327         }
 328         rcu_exp_gp_seq_start();
 329         trace_rcu_exp_grace_period(rcu_state.name, s, TPS("start"));
 330         return false;
 331 }
 332
 333 /*
 334  * Select the CPUs within the specified rcu_node that the upcoming
 335  * expedited grace period needs to wait for.
 336  */
 337 static void sync_rcu_exp_select_node_cpus(struct work_struct *wp)
 338 {
 339         int cpu;
 340         unsigned long flags;
 341         unsigned long mask_ofl_test;
 342         unsigned long mask_ofl_ipi;
 343         int ret;
 344         struct rcu_exp_work *rewp =
 345                 container_of(wp, struct rcu_exp_work, rew_work);
 346         struct rcu_node *rnp = container_of(rewp, struct rcu_node, rew);
 347
 348         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 349
 350         /* Each pass checks a CPU for identity, offline, and idle. */
 351         mask_ofl_test = 0;
 352         for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
 353                 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 354                 unsigned long mask = rdp->grpmask;
 355                 int snap;
 356
 357                 if (raw_smp_processor_id() == cpu ||
 358                     !(rnp->qsmaskinitnext & mask)) {
 359                         mask_ofl_test |= mask;
 360                 } else {
 361                         snap = rcu_dynticks_snap(rdp);
 362                         if (rcu_dynticks_in_eqs(snap))
 363                                 mask_ofl_test |= mask;
 364                         else
 365                                 rdp->exp_dynticks_snap = snap;
 366                 }
 367         }
 368         mask_ofl_ipi = rnp->expmask & ~mask_ofl_test;
 369
 370         /*
 371          * Need to wait for any blocked tasks as well.  Note that
 372          * additional blocking tasks will also block the expedited GP
 373          * until such time as the ->expmask bits are cleared.
 374          */
 375         if (rcu_preempt_has_tasks(rnp))
 376                 rnp->exp_tasks = rnp->blkd_tasks.next;
 377         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 378
 379         /* IPI the remaining CPUs for expedited quiescent state. */
 380         for_each_leaf_node_cpu_mask(rnp, cpu, mask_ofl_ipi) {
 381                 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 382                 unsigned long mask = rdp->grpmask;
 383
 384 retry_ipi:
 385                 if (rcu_dynticks_in_eqs_since(rdp, rdp->exp_dynticks_snap)) {
 386                         mask_ofl_test |= mask;
 387                         continue;
 388                 }
 389                 if (get_cpu() == cpu) {
 390                         put_cpu();
 391                         continue;
 392                 }
 393                 ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
 394                 put_cpu();
 395                 /* The CPU will report the QS in response to the IPI. */
 396                 if (!ret)
 397                         continue;
 398
 399                 /* Failed, raced with CPU hotplug operation. */
 400                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 401                 if ((rnp->qsmaskinitnext & mask) &&
 402                     (rnp->expmask & mask)) {
 403                         /* Online, so delay for a bit and try again. */
 404                         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 405                         trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("selectofl"));
 406                         schedule_timeout_uninterruptible(1);
 407                         goto retry_ipi;
 408                 }
 409                 /* CPU really is offline, so we must report its QS. */
 410                 if (rnp->expmask & mask)
 411                         mask_ofl_test |= mask;
 412                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 413         }
 414         /* Report quiescent states for those that went offline. */
 415         if (mask_ofl_test)
 416                 rcu_report_exp_cpu_mult(rnp, mask_ofl_test, false);
 417 }
 418
 419 /*
 420  * Select the nodes that the upcoming expedited grace period needs
 421  * to wait for.
 422  */
 423 static void sync_rcu_exp_select_cpus(void)
 424 {
 425         int cpu;
 426         struct rcu_node *rnp;
 427
 428         trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("reset"));
 429         sync_exp_reset_tree();
 430         trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("select"));
 431
 432         /* Schedule work for each leaf rcu_node structure. */
 433         rcu_for_each_leaf_node(rnp) {
 434                 rnp->exp_need_flush = false;
 435                 if (!READ_ONCE(rnp->expmask))
 436                         continue; /* Avoid early boot non-existent wq. */
 437                 if (!READ_ONCE(rcu_par_gp_wq) ||
 438                     rcu_scheduler_active != RCU_SCHEDULER_RUNNING ||
 439                     rcu_is_last_leaf_node(rnp)) {
 440                         /* No workqueues yet or last leaf, do direct call. */
 441                         sync_rcu_exp_select_node_cpus(&rnp->rew.rew_work);
 442                         continue;
 443                 }
 444                 INIT_WORK(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
 445                 cpu = find_next_bit(&rnp->ffmask, BITS_PER_LONG, -1);
 446                 /* If all offline, queue the work on an unbound CPU. */
 447                 if (unlikely(cpu > rnp->grphi - rnp->grplo))
 448                         cpu = WORK_CPU_UNBOUND;
 449                 else
 450                         cpu += rnp->grplo;
 451                 queue_work_on(cpu, rcu_par_gp_wq, &rnp->rew.rew_work);
 452                 rnp->exp_need_flush = true;
 453         }
 454
 455         /* Wait for workqueue jobs (if any) to complete. */
 456         rcu_for_each_leaf_node(rnp)
 457                 if (rnp->exp_need_flush)
 458                         flush_work(&rnp->rew.rew_work);
 459 }
 460
 461 /*
 462  * Wait for the expedited grace period to elapse, within time limit.
 463  * If the time limit is exceeded without the grace period elapsing,
 464  * return false, otherwise return true.
 465  */
 466 static bool synchronize_rcu_expedited_wait_once(long tlimit)
 467 {
 468         int t;
 469         struct rcu_node *rnp_root = rcu_get_root();
 470
 471         t = swait_event_timeout_exclusive(rcu_state.expedited_wq,
 472                                           sync_rcu_exp_done_unlocked(rnp_root),
 473                                           tlimit);
 474         // Workqueues should not be signaled.
 475         if (t > 0 || sync_rcu_exp_done_unlocked(rnp_root))
 476                 return true;
 477         WARN_ON(t < 0);  /* workqueues should not be signaled. */
 478         return false;
 479 }
 480
 481 /*
 482  * Wait for the expedited grace period to elapse, issuing any needed
 483  * RCU CPU stall warnings along the way.
 484  */
 485 static void synchronize_rcu_expedited_wait(void)
 486 {
 487         int cpu;
 488         unsigned long jiffies_stall;
 489         unsigned long jiffies_start;
 490         unsigned long mask;
 491         int ndetected;
 492         struct rcu_data *rdp;
 493         struct rcu_node *rnp;
 494         struct rcu_node *rnp_root = rcu_get_root();
 495
 496         trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("startwait"));
 497         jiffies_stall = rcu_jiffies_till_stall_check();
 498         jiffies_start = jiffies;
 499         if (IS_ENABLED(CONFIG_NO_HZ_FULL)) {
 500                 if (synchronize_rcu_expedited_wait_once(1))
 501                         return;
 502                 rcu_for_each_leaf_node(rnp) {
 503                         for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
 504                                 rdp = per_cpu_ptr(&rcu_data, cpu);
 505                                 if (rdp->rcu_forced_tick_exp)
 506                                         continue;
 507                                 rdp->rcu_forced_tick_exp = true;
 508                                 tick_dep_set_cpu(cpu, TICK_DEP_BIT_RCU_EXP);
 509                         }
 510                 }
 511         }
 512
 513         for (;;) {
 514                 if (synchronize_rcu_expedited_wait_once(jiffies_stall))
 515                         return;
 516                 if (rcu_cpu_stall_suppress)
 517                         continue;
 518                 panic_on_rcu_stall();
 519                 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
 520                        rcu_state.name);
 521                 ndetected = 0;
 522                 rcu_for_each_leaf_node(rnp) {
 523                         ndetected += rcu_print_task_exp_stall(rnp);
 524                         for_each_leaf_node_possible_cpu(rnp, cpu) {
 525                                 struct rcu_data *rdp;
 526
 527                                 mask = leaf_node_cpu_bit(rnp, cpu);
 528                                 if (!(READ_ONCE(rnp->expmask) & mask))
 529                                         continue;
 530                                 ndetected++;
 531                                 rdp = per_cpu_ptr(&rcu_data, cpu);
 532                                 pr_cont(" %d-%c%c%c", cpu,
 533                                         "O."[!!cpu_online(cpu)],
 534                                         "o."[!!(rdp->grpmask & rnp->expmaskinit)],
 535                                         "N."[!!(rdp->grpmask & rnp->expmaskinitnext)]);
 536                         }
 537                 }
 538                 pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
 539                         jiffies - jiffies_start, rcu_state.expedited_sequence,
 540                         READ_ONCE(rnp_root->expmask),
 541                         ".T"[!!rnp_root->exp_tasks]);
 542                 if (ndetected) {
 543                         pr_err("blocking rcu_node structures:");
 544                         rcu_for_each_node_breadth_first(rnp) {
 545                                 if (rnp == rnp_root)
 546                                         continue; /* printed unconditionally */
 547                                 if (sync_rcu_exp_done_unlocked(rnp))
 548                                         continue;
 549                                 pr_cont(" l=%u:%d-%d:%#lx/%c",
 550                                         rnp->level, rnp->grplo, rnp->grphi,
 551                                         READ_ONCE(rnp->expmask),
 552                                         ".T"[!!rnp->exp_tasks]);
 553                         }
 554                         pr_cont("\n");
 555                 }
 556                 rcu_for_each_leaf_node(rnp) {
 557                         for_each_leaf_node_possible_cpu(rnp, cpu) {
 558                                 mask = leaf_node_cpu_bit(rnp, cpu);
 559                                 if (!(READ_ONCE(rnp->expmask) & mask))
 560                                         continue;
 561                                 dump_cpu_task(cpu);
 562                         }
 563                 }
 564                 jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3;
 565         }
 566 }
 567
 568 /*
 569  * Wait for the current expedited grace period to complete, and then
 570  * wake up everyone who piggybacked on the just-completed expedited
 571  * grace period.  Also update all the ->exp_seq_rq counters as needed
 572  * in order to avoid counter-wrap problems.
 573  */
 574 static void rcu_exp_wait_wake(unsigned long s)
 575 {
 576         struct rcu_node *rnp;
 577
 578         synchronize_rcu_expedited_wait();
 579
 580         // Switch over to wakeup mode, allowing the next GP to proceed.
 581         // End the previous grace period only after acquiring the mutex
 582         // to ensure that only one GP runs concurrently with wakeups.
 583         mutex_lock(&rcu_state.exp_wake_mutex);
 584         rcu_exp_gp_seq_end();
 585         trace_rcu_exp_grace_period(rcu_state.name, s, TPS("end"));
 586
 587         rcu_for_each_node_breadth_first(rnp) {
 588                 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s)) {
 589                         spin_lock(&rnp->exp_lock);
 590                         /* Recheck, avoid hang in case someone just arrived. */
 591                         if (ULONG_CMP_LT(rnp->exp_seq_rq, s))
 592                                 rnp->exp_seq_rq = s;
 593                         spin_unlock(&rnp->exp_lock);
 594                 }
 595                 smp_mb(); /* All above changes before wakeup. */
 596                 wake_up_all(&rnp->exp_wq[rcu_seq_ctr(s) & 0x3]);
 597         }
 598         trace_rcu_exp_grace_period(rcu_state.name, s, TPS("endwake"));
 599         mutex_unlock(&rcu_state.exp_wake_mutex);
 600 }
 601
 602 /*
 603  * Common code to drive an expedited grace period forward, used by
 604  * workqueues and mid-boot-time tasks.
 605  */
 606 static void rcu_exp_sel_wait_wake(unsigned long s)
 607 {
 608         /* Initialize the rcu_node tree in preparation for the wait. */
 609         sync_rcu_exp_select_cpus();
 610
 611         /* Wait and clean up, including waking everyone. */
 612         rcu_exp_wait_wake(s);
 613 }
 614
 615 /*
 616  * Work-queue handler to drive an expedited grace period forward.
 617  */
 618 static void wait_rcu_exp_gp(struct work_struct *wp)
 619 {
 620         struct rcu_exp_work *rewp;
 621
 622         rewp = container_of(wp, struct rcu_exp_work, rew_work);
 623         rcu_exp_sel_wait_wake(rewp->rew_s);
 624 }
 625
 626 #ifdef CONFIG_PREEMPT_RCU
 627
 628 /*
 629  * Remote handler for smp_call_function_single().  If there is an
 630  * RCU read-side critical section in effect, request that the
 631  * next rcu_read_unlock() record the quiescent state up the
 632  * ->expmask fields in the rcu_node tree.  Otherwise, immediately
 633  * report the quiescent state.
 634  */
 635 static void rcu_exp_handler(void *unused)
 636 {
 637         unsigned long flags;
 638         struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 639         struct rcu_node *rnp = rdp->mynode;
 640         struct task_struct *t = current;
 641
 642         /*
 643          * First, the common case of not being in an RCU read-side
 644          * critical section.  If also enabled or idle, immediately
 645          * report the quiescent state, otherwise defer.
 646          */
 647         if (!rcu_preempt_depth()) {
 648                 if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
 649                     rcu_dynticks_curr_cpu_in_eqs()) {
 650                         rcu_report_exp_rdp(rdp);
 651                 } else {
 652                         rdp->exp_deferred_qs = true;
 653                         set_tsk_need_resched(t);
 654                         set_preempt_need_resched();
 655                 }
 656                 return;
 657         }
 658
 659         /*
 660          * Second, the less-common case of being in an RCU read-side
 661          * critical section.  In this case we can count on a future
 662          * rcu_read_unlock().  However, this rcu_read_unlock() might
 663          * execute on some other CPU, but in that case there will be
 664          * a future context switch.  Either way, if the expedited
 665          * grace period is still waiting on this CPU, set ->deferred_qs
 666          * so that the eventual quiescent state will be reported.
 667          * Note that there is a large group of race conditions that
 668          * can have caused this quiescent state to already have been
 669          * reported, so we really do need to check ->expmask.
 670          */
 671         if (rcu_preempt_depth() > 0) {
 672                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 673                 if (rnp->expmask & rdp->grpmask) {
 674                         rdp->exp_deferred_qs = true;
 675                         t->rcu_read_unlock_special.b.exp_hint = true;
 676                 }
 677                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 678                 return;
 679         }
 680
 681         /*
 682          * The final and least likely case is where the interrupted
 683          * code was just about to or just finished exiting the RCU-preempt
 684          * read-side critical section, and no, we can't tell which.
 685          * So either way, set ->deferred_qs to flag later code that
 686          * a quiescent state is required.
 687          *
 688          * If the CPU is fully enabled (or if some buggy RCU-preempt
 689          * read-side critical section is being used from idle), just
 690          * invoke rcu_preempt_deferred_qs() to immediately report the
 691          * quiescent state.  We cannot use rcu_read_unlock_special()
 692          * because we are in an interrupt handler, which will cause that
 693          * function to take an early exit without doing anything.
 694          *
 695          * Otherwise, force a context switch after the CPU enables everything.
 696          */
 697         rdp->exp_deferred_qs = true;
 698         if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
 699             WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs())) {
 700                 rcu_preempt_deferred_qs(t);
 701         } else {
 702                 set_tsk_need_resched(t);
 703                 set_preempt_need_resched();
 704         }
 705 }
 706
 707 /* PREEMPTION=y, so no PREEMPTION=n expedited grace period to clean up after. */
 708 static void sync_sched_exp_online_cleanup(int cpu)
 709 {
 710 }
 711
 712 /*
 713  * Scan the current list of tasks blocked within RCU read-side critical
 714  * sections, printing out the tid of each that is blocking the current
 715  * expedited grace period.
 716  */
 717 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
 718 {
 719         struct task_struct *t;
 720         int ndetected = 0;
 721
 722         if (!rnp->exp_tasks)
 723                 return 0;
 724         t = list_entry(rnp->exp_tasks->prev,
 725                        struct task_struct, rcu_node_entry);
 726         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
 727                 pr_cont(" P%d", t->pid);
 728                 ndetected++;
 729         }
 730         return ndetected;
 731 }
 732
 733 #else /* #ifdef CONFIG_PREEMPT_RCU */
 734
 735 /* Request an expedited quiescent state. */
 736 static void rcu_exp_need_qs(void)
 737 {
 738         __this_cpu_write(rcu_data.cpu_no_qs.b.exp, true);
 739         /* Store .exp before .rcu_urgent_qs. */
 740         smp_store_release(this_cpu_ptr(&rcu_data.rcu_urgent_qs), true);
 741         set_tsk_need_resched(current);
 742         set_preempt_need_resched();
 743 }
 744
 745 /* Invoked on each online non-idle CPU for expedited quiescent state. */
 746 static void rcu_exp_handler(void *unused)
 747 {
 748         struct rcu_data *rdp;
 749         struct rcu_node *rnp;
 750
 751         rdp = this_cpu_ptr(&rcu_data);
 752         rnp = rdp->mynode;
 753         if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
 754             __this_cpu_read(rcu_data.cpu_no_qs.b.exp))
 755                 return;
 756         if (rcu_is_cpu_rrupt_from_idle()) {
 757                 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
 758                 return;
 759         }
 760         rcu_exp_need_qs();
 761 }
 762
 763 /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
 764 static void sync_sched_exp_online_cleanup(int cpu)
 765 {
 766         unsigned long flags;
 767         int my_cpu;
 768         struct rcu_data *rdp;
 769         int ret;
 770         struct rcu_node *rnp;
 771
 772         rdp = per_cpu_ptr(&rcu_data, cpu);
 773         rnp = rdp->mynode;
 774         my_cpu = get_cpu();
 775         /* Quiescent state either not needed or already requested, leave. */
 776         if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
 777             __this_cpu_read(rcu_data.cpu_no_qs.b.exp)) {
 778                 put_cpu();
 779                 return;
 780         }
 781         /* Quiescent state needed on current CPU, so set it up locally. */
 782         if (my_cpu == cpu) {
 783                 local_irq_save(flags);
 784                 rcu_exp_need_qs();
 785                 local_irq_restore(flags);
 786                 put_cpu();
 787                 return;
 788         }
 789         /* Quiescent state needed on some other CPU, send IPI. */
 790         ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
 791         put_cpu();
 792         WARN_ON_ONCE(ret);
 793 }
 794
 795 /*
 796  * Because preemptible RCU does not exist, we never have to check for
 797  * tasks blocked within RCU read-side critical sections that are
 798  * blocking the current expedited grace period.
 799  */
 800 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
 801 {
 802         return 0;
 803 }
 804
 805 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
 806
 807 /**
 808  * synchronize_rcu_expedited - Brute-force RCU grace period
 809  *
 810  * Wait for an RCU grace period, but expedite it.  The basic idea is to
 811  * IPI all non-idle non-nohz online CPUs.  The IPI handler checks whether
 812  * the CPU is in an RCU critical section, and if so, it sets a flag that
 813  * causes the outermost rcu_read_unlock() to report the quiescent state
 814  * for RCU-preempt or asks the scheduler for help for RCU-sched.  On the
 815  * other hand, if the CPU is not in an RCU read-side critical section,
 816  * the IPI handler reports the quiescent state immediately.
 817  *
 818  * Although this is a great improvement over previous expedited
 819  * implementations, it is still unfriendly to real-time workloads, so is
 820  * thus not recommended for any sort of common-case code.  In fact, if
 821  * you are using synchronize_rcu_expedited() in a loop, please restructure
 822  * your code to batch your updates, and then use a single synchronize_rcu()
 823  * instead.
 824  *
 825  * This has the same semantics as (but is more brutal than) synchronize_rcu().
 826  */
 827 void synchronize_rcu_expedited(void)
 828 {
 829         bool boottime = (rcu_scheduler_active == RCU_SCHEDULER_INIT);
 830         struct rcu_exp_work rew;
 831         struct rcu_node *rnp;
 832         unsigned long s;
 833
 834         RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
 835                          lock_is_held(&rcu_lock_map) ||
 836                          lock_is_held(&rcu_sched_lock_map),
 837                          "Illegal synchronize_rcu_expedited() in RCU read-side critical section");
 838
 839         /* Is the state is such that the call is a grace period? */
 840         if (rcu_blocking_is_gp())
 841                 return;
 842
 843         /* If expedited grace periods are prohibited, fall back to normal. */
 844         if (rcu_gp_is_normal()) {
 845                 wait_rcu_gp(call_rcu);
 846                 return;
 847         }
 848
 849         /* Take a snapshot of the sequence number.  */
 850         s = rcu_exp_gp_seq_snap();
 851         if (exp_funnel_lock(s))
 852                 return;  /* Someone else did our work for us. */
 853
 854         /* Ensure that load happens before action based on it. */
 855         if (unlikely(boottime)) {
 856                 /* Direct call during scheduler init and early_initcalls(). */
 857                 rcu_exp_sel_wait_wake(s);
 858         } else {
 859                 /* Marshall arguments & schedule the expedited grace period. */
 860                 rew.rew_s = s;
 861                 INIT_WORK_ONSTACK(&rew.rew_work, wait_rcu_exp_gp);
 862                 queue_work(rcu_gp_wq, &rew.rew_work);
 863         }
 864
 865         /* Wait for expedited grace period to complete. */
 866         rnp = rcu_get_root();
 867         wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
 868                    sync_exp_work_done(s));
 869         smp_mb(); /* Workqueue actions happen before return. */
 870
 871         /* Let the next expedited grace period start. */
 872         mutex_unlock(&rcu_state.exp_mutex);
 873
 874         if (likely(!boottime))
 875                 destroy_work_on_stack(&rew.rew_work);
 876 }
 877 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);