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         rcu_poll_gp_seq_start_unlocked(&rcu_state.gp_seq_polled_exp_snap);
  22 }
  23
  24 /*
  25  * Return the value that the expedited-grace-period counter will have
  26  * at the end of the current grace period.
  27  */
  28 static __maybe_unused unsigned long rcu_exp_gp_seq_endval(void)
  29 {
  30         return rcu_seq_endval(&rcu_state.expedited_sequence);
  31 }
  32
  33 /*
  34  * Record the end of an expedited grace period.
  35  */
  36 static void rcu_exp_gp_seq_end(void)
  37 {
  38         rcu_poll_gp_seq_end_unlocked(&rcu_state.gp_seq_polled_exp_snap);
  39         rcu_seq_end(&rcu_state.expedited_sequence);
  40         smp_mb(); /* Ensure that consecutive grace periods serialize. */
  41 }
  42
  43 /*
  44  * Take a snapshot of the expedited-grace-period counter, which is the
  45  * earliest value that will indicate that a full grace period has
  46  * elapsed since the current time.
  47  */
  48 static unsigned long rcu_exp_gp_seq_snap(void)
  49 {
  50         unsigned long s;
  51
  52         smp_mb(); /* Caller's modifications seen first by other CPUs. */
  53         s = rcu_seq_snap(&rcu_state.expedited_sequence);
  54         trace_rcu_exp_grace_period(rcu_state.name, s, TPS("snap"));
  55         return s;
  56 }
  57
  58 /*
  59  * Given a counter snapshot from rcu_exp_gp_seq_snap(), return true
  60  * if a full expedited grace period has elapsed since that snapshot
  61  * was taken.
  62  */
  63 static bool rcu_exp_gp_seq_done(unsigned long s)
  64 {
  65         return rcu_seq_done(&rcu_state.expedited_sequence, s);
  66 }
  67
  68 /*
  69  * Reset the ->expmaskinit values in the rcu_node tree to reflect any
  70  * recent CPU-online activity.  Note that these masks are not cleared
  71  * when CPUs go offline, so they reflect the union of all CPUs that have
  72  * ever been online.  This means that this function normally takes its
  73  * no-work-to-do fastpath.
  74  */
  75 static void sync_exp_reset_tree_hotplug(void)
  76 {
  77         bool done;
  78         unsigned long flags;
  79         unsigned long mask;
  80         unsigned long oldmask;
  81         int ncpus = smp_load_acquire(&rcu_state.ncpus); /* Order vs. locking. */
  82         struct rcu_node *rnp;
  83         struct rcu_node *rnp_up;
  84
  85         /* If no new CPUs onlined since last time, nothing to do. */
  86         if (likely(ncpus == rcu_state.ncpus_snap))
  87                 return;
  88         rcu_state.ncpus_snap = ncpus;
  89
  90         /*
  91          * Each pass through the following loop propagates newly onlined
  92          * CPUs for the current rcu_node structure up the rcu_node tree.
  93          */
  94         rcu_for_each_leaf_node(rnp) {
  95                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
  96                 if (rnp->expmaskinit == rnp->expmaskinitnext) {
  97                         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
  98                         continue;  /* No new CPUs, nothing to do. */
  99                 }
 100
 101                 /* Update this node's mask, track old value for propagation. */
 102                 oldmask = rnp->expmaskinit;
 103                 rnp->expmaskinit = rnp->expmaskinitnext;
 104                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 105
 106                 /* If was already nonzero, nothing to propagate. */
 107                 if (oldmask)
 108                         continue;
 109
 110                 /* Propagate the new CPU up the tree. */
 111                 mask = rnp->grpmask;
 112                 rnp_up = rnp->parent;
 113                 done = false;
 114                 while (rnp_up) {
 115                         raw_spin_lock_irqsave_rcu_node(rnp_up, flags);
 116                         if (rnp_up->expmaskinit)
 117                                 done = true;
 118                         rnp_up->expmaskinit |= mask;
 119                         raw_spin_unlock_irqrestore_rcu_node(rnp_up, flags);
 120                         if (done)
 121                                 break;
 122                         mask = rnp_up->grpmask;
 123                         rnp_up = rnp_up->parent;
 124                 }
 125         }
 126 }
 127
 128 /*
 129  * Reset the ->expmask values in the rcu_node tree in preparation for
 130  * a new expedited grace period.
 131  */
 132 static void __maybe_unused sync_exp_reset_tree(void)
 133 {
 134         unsigned long flags;
 135         struct rcu_node *rnp;
 136
 137         sync_exp_reset_tree_hotplug();
 138         rcu_for_each_node_breadth_first(rnp) {
 139                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 140                 WARN_ON_ONCE(rnp->expmask);
 141                 WRITE_ONCE(rnp->expmask, rnp->expmaskinit);
 142                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 143         }
 144 }
 145
 146 /*
 147  * Return non-zero if there is no RCU expedited grace period in progress
 148  * for the specified rcu_node structure, in other words, if all CPUs and
 149  * tasks covered by the specified rcu_node structure have done their bit
 150  * for the current expedited grace period.
 151  */
 152 static bool sync_rcu_exp_done(struct rcu_node *rnp)
 153 {
 154         raw_lockdep_assert_held_rcu_node(rnp);
 155         return READ_ONCE(rnp->exp_tasks) == NULL &&
 156                READ_ONCE(rnp->expmask) == 0;
 157 }
 158
 159 /*
 160  * Like sync_rcu_exp_done(), but where the caller does not hold the
 161  * rcu_node's ->lock.
 162  */
 163 static bool sync_rcu_exp_done_unlocked(struct rcu_node *rnp)
 164 {
 165         unsigned long flags;
 166         bool ret;
 167
 168         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 169         ret = sync_rcu_exp_done(rnp);
 170         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 171
 172         return ret;
 173 }
 174
 175
 176 /*
 177  * Report the exit from RCU read-side critical section for the last task
 178  * that queued itself during or before the current expedited preemptible-RCU
 179  * grace period.  This event is reported either to the rcu_node structure on
 180  * which the task was queued or to one of that rcu_node structure's ancestors,
 181  * recursively up the tree.  (Calm down, calm down, we do the recursion
 182  * iteratively!)
 183  */
 184 static void __rcu_report_exp_rnp(struct rcu_node *rnp,
 185                                  bool wake, unsigned long flags)
 186         __releases(rnp->lock)
 187 {
 188         unsigned long mask;
 189
 190         raw_lockdep_assert_held_rcu_node(rnp);
 191         for (;;) {
 192                 if (!sync_rcu_exp_done(rnp)) {
 193                         if (!rnp->expmask)
 194                                 rcu_initiate_boost(rnp, flags);
 195                         else
 196                                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 197                         break;
 198                 }
 199                 if (rnp->parent == NULL) {
 200                         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 201                         if (wake) {
 202                                 smp_mb(); /* EGP done before wake_up(). */
 203                                 swake_up_one(&rcu_state.expedited_wq);
 204                         }
 205                         break;
 206                 }
 207                 mask = rnp->grpmask;
 208                 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled */
 209                 rnp = rnp->parent;
 210                 raw_spin_lock_rcu_node(rnp); /* irqs already disabled */
 211                 WARN_ON_ONCE(!(rnp->expmask & mask));
 212                 WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
 213         }
 214 }
 215
 216 /*
 217  * Report expedited quiescent state for specified node.  This is a
 218  * lock-acquisition wrapper function for __rcu_report_exp_rnp().
 219  */
 220 static void __maybe_unused rcu_report_exp_rnp(struct rcu_node *rnp, bool wake)
 221 {
 222         unsigned long flags;
 223
 224         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 225         __rcu_report_exp_rnp(rnp, wake, flags);
 226 }
 227
 228 /*
 229  * Report expedited quiescent state for multiple CPUs, all covered by the
 230  * specified leaf rcu_node structure.
 231  */
 232 static void rcu_report_exp_cpu_mult(struct rcu_node *rnp,
 233                                     unsigned long mask, bool wake)
 234 {
 235         int cpu;
 236         unsigned long flags;
 237         struct rcu_data *rdp;
 238
 239         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 240         if (!(rnp->expmask & mask)) {
 241                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 242                 return;
 243         }
 244         WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
 245         for_each_leaf_node_cpu_mask(rnp, cpu, mask) {
 246                 rdp = per_cpu_ptr(&rcu_data, cpu);
 247                 if (!IS_ENABLED(CONFIG_NO_HZ_FULL) || !rdp->rcu_forced_tick_exp)
 248                         continue;
 249                 rdp->rcu_forced_tick_exp = false;
 250                 tick_dep_clear_cpu(cpu, TICK_DEP_BIT_RCU_EXP);
 251         }
 252         __rcu_report_exp_rnp(rnp, wake, flags); /* Releases rnp->lock. */
 253 }
 254
 255 /*
 256  * Report expedited quiescent state for specified rcu_data (CPU).
 257  */
 258 static void rcu_report_exp_rdp(struct rcu_data *rdp)
 259 {
 260         WRITE_ONCE(rdp->cpu_no_qs.b.exp, false);
 261         rcu_report_exp_cpu_mult(rdp->mynode, rdp->grpmask, true);
 262 }
 263
 264 /* Common code for work-done checking. */
 265 static bool sync_exp_work_done(unsigned long s)
 266 {
 267         if (rcu_exp_gp_seq_done(s)) {
 268                 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("done"));
 269                 smp_mb(); /* Ensure test happens before caller kfree(). */
 270                 return true;
 271         }
 272         return false;
 273 }
 274
 275 /*
 276  * Funnel-lock acquisition for expedited grace periods.  Returns true
 277  * if some other task completed an expedited grace period that this task
 278  * can piggy-back on, and with no mutex held.  Otherwise, returns false
 279  * with the mutex held, indicating that the caller must actually do the
 280  * expedited grace period.
 281  */
 282 static bool exp_funnel_lock(unsigned long s)
 283 {
 284         struct rcu_data *rdp = per_cpu_ptr(&rcu_data, raw_smp_processor_id());
 285         struct rcu_node *rnp = rdp->mynode;
 286         struct rcu_node *rnp_root = rcu_get_root();
 287
 288         /* Low-contention fastpath. */
 289         if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s) &&
 290             (rnp == rnp_root ||
 291              ULONG_CMP_LT(READ_ONCE(rnp_root->exp_seq_rq), s)) &&
 292             mutex_trylock(&rcu_state.exp_mutex))
 293                 goto fastpath;
 294
 295         /*
 296          * Each pass through the following loop works its way up
 297          * the rcu_node tree, returning if others have done the work or
 298          * otherwise falls through to acquire ->exp_mutex.  The mapping
 299          * from CPU to rcu_node structure can be inexact, as it is just
 300          * promoting locality and is not strictly needed for correctness.
 301          */
 302         for (; rnp != NULL; rnp = rnp->parent) {
 303                 if (sync_exp_work_done(s))
 304                         return true;
 305
 306                 /* Work not done, either wait here or go up. */
 307                 spin_lock(&rnp->exp_lock);
 308                 if (ULONG_CMP_GE(rnp->exp_seq_rq, s)) {
 309
 310                         /* Someone else doing GP, so wait for them. */
 311                         spin_unlock(&rnp->exp_lock);
 312                         trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
 313                                                   rnp->grplo, rnp->grphi,
 314                                                   TPS("wait"));
 315                         wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
 316                                    sync_exp_work_done(s));
 317                         return true;
 318                 }
 319                 WRITE_ONCE(rnp->exp_seq_rq, s); /* Followers can wait on us. */
 320                 spin_unlock(&rnp->exp_lock);
 321                 trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
 322                                           rnp->grplo, rnp->grphi, TPS("nxtlvl"));
 323         }
 324         mutex_lock(&rcu_state.exp_mutex);
 325 fastpath:
 326         if (sync_exp_work_done(s)) {
 327                 mutex_unlock(&rcu_state.exp_mutex);
 328                 return true;
 329         }
 330         rcu_exp_gp_seq_start();
 331         trace_rcu_exp_grace_period(rcu_state.name, s, TPS("start"));
 332         return false;
 333 }
 334
 335 /*
 336  * Select the CPUs within the specified rcu_node that the upcoming
 337  * expedited grace period needs to wait for.
 338  */
 339 static void __sync_rcu_exp_select_node_cpus(struct rcu_exp_work *rewp)
 340 {
 341         int cpu;
 342         unsigned long flags;
 343         unsigned long mask_ofl_test;
 344         unsigned long mask_ofl_ipi;
 345         int ret;
 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(cpu);
 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                 WRITE_ONCE(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                         mask_ofl_test |= mask;
 391                         put_cpu();
 392                         continue;
 393                 }
 394                 ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
 395                 put_cpu();
 396                 /* The CPU will report the QS in response to the IPI. */
 397                 if (!ret)
 398                         continue;
 399
 400                 /* Failed, raced with CPU hotplug operation. */
 401                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 402                 if ((rnp->qsmaskinitnext & mask) &&
 403                     (rnp->expmask & mask)) {
 404                         /* Online, so delay for a bit and try again. */
 405                         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 406                         trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("selectofl"));
 407                         schedule_timeout_idle(1);
 408                         goto retry_ipi;
 409                 }
 410                 /* CPU really is offline, so we must report its QS. */
 411                 if (rnp->expmask & mask)
 412                         mask_ofl_test |= mask;
 413                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 414         }
 415         /* Report quiescent states for those that went offline. */
 416         if (mask_ofl_test)
 417                 rcu_report_exp_cpu_mult(rnp, mask_ofl_test, false);
 418 }
 419
 420 static void rcu_exp_sel_wait_wake(unsigned long s);
 421
 422 #ifdef CONFIG_RCU_EXP_KTHREAD
 423 static void sync_rcu_exp_select_node_cpus(struct kthread_work *wp)
 424 {
 425         struct rcu_exp_work *rewp =
 426                 container_of(wp, struct rcu_exp_work, rew_work);
 427
 428         __sync_rcu_exp_select_node_cpus(rewp);
 429 }
 430
 431 static inline bool rcu_gp_par_worker_started(void)
 432 {
 433         return !!READ_ONCE(rcu_exp_par_gp_kworker);
 434 }
 435
 436 static inline void sync_rcu_exp_select_cpus_queue_work(struct rcu_node *rnp)
 437 {
 438         kthread_init_work(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
 439         /*
 440          * Use rcu_exp_par_gp_kworker, because flushing a work item from
 441          * another work item on the same kthread worker can result in
 442          * deadlock.
 443          */
 444         kthread_queue_work(rcu_exp_par_gp_kworker, &rnp->rew.rew_work);
 445 }
 446
 447 static inline void sync_rcu_exp_select_cpus_flush_work(struct rcu_node *rnp)
 448 {
 449         kthread_flush_work(&rnp->rew.rew_work);
 450 }
 451
 452 /*
 453  * Work-queue handler to drive an expedited grace period forward.
 454  */
 455 static void wait_rcu_exp_gp(struct kthread_work *wp)
 456 {
 457         struct rcu_exp_work *rewp;
 458
 459         rewp = container_of(wp, struct rcu_exp_work, rew_work);
 460         rcu_exp_sel_wait_wake(rewp->rew_s);
 461 }
 462
 463 static inline void synchronize_rcu_expedited_queue_work(struct rcu_exp_work *rew)
 464 {
 465         kthread_init_work(&rew->rew_work, wait_rcu_exp_gp);
 466         kthread_queue_work(rcu_exp_gp_kworker, &rew->rew_work);
 467 }
 468
 469 static inline void synchronize_rcu_expedited_destroy_work(struct rcu_exp_work *rew)
 470 {
 471 }
 472 #else /* !CONFIG_RCU_EXP_KTHREAD */
 473 static void sync_rcu_exp_select_node_cpus(struct work_struct *wp)
 474 {
 475         struct rcu_exp_work *rewp =
 476                 container_of(wp, struct rcu_exp_work, rew_work);
 477
 478         __sync_rcu_exp_select_node_cpus(rewp);
 479 }
 480
 481 static inline bool rcu_gp_par_worker_started(void)
 482 {
 483         return !!READ_ONCE(rcu_par_gp_wq);
 484 }
 485
 486 static inline void sync_rcu_exp_select_cpus_queue_work(struct rcu_node *rnp)
 487 {
 488         int cpu = find_next_bit(&rnp->ffmask, BITS_PER_LONG, -1);
 489
 490         INIT_WORK(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
 491         /* If all offline, queue the work on an unbound CPU. */
 492         if (unlikely(cpu > rnp->grphi - rnp->grplo))
 493                 cpu = WORK_CPU_UNBOUND;
 494         else
 495                 cpu += rnp->grplo;
 496         queue_work_on(cpu, rcu_par_gp_wq, &rnp->rew.rew_work);
 497 }
 498
 499 static inline void sync_rcu_exp_select_cpus_flush_work(struct rcu_node *rnp)
 500 {
 501         flush_work(&rnp->rew.rew_work);
 502 }
 503
 504 /*
 505  * Work-queue handler to drive an expedited grace period forward.
 506  */
 507 static void wait_rcu_exp_gp(struct work_struct *wp)
 508 {
 509         struct rcu_exp_work *rewp;
 510
 511         rewp = container_of(wp, struct rcu_exp_work, rew_work);
 512         rcu_exp_sel_wait_wake(rewp->rew_s);
 513 }
 514
 515 static inline void synchronize_rcu_expedited_queue_work(struct rcu_exp_work *rew)
 516 {
 517         INIT_WORK_ONSTACK(&rew->rew_work, wait_rcu_exp_gp);
 518         queue_work(rcu_gp_wq, &rew->rew_work);
 519 }
 520
 521 static inline void synchronize_rcu_expedited_destroy_work(struct rcu_exp_work *rew)
 522 {
 523         destroy_work_on_stack(&rew->rew_work);
 524 }
 525 #endif /* CONFIG_RCU_EXP_KTHREAD */
 526
 527 /*
 528  * Select the nodes that the upcoming expedited grace period needs
 529  * to wait for.
 530  */
 531 static void sync_rcu_exp_select_cpus(void)
 532 {
 533         struct rcu_node *rnp;
 534
 535         trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("reset"));
 536         sync_exp_reset_tree();
 537         trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("select"));
 538
 539         /* Schedule work for each leaf rcu_node structure. */
 540         rcu_for_each_leaf_node(rnp) {
 541                 rnp->exp_need_flush = false;
 542                 if (!READ_ONCE(rnp->expmask))
 543                         continue; /* Avoid early boot non-existent wq. */
 544                 if (!rcu_gp_par_worker_started() ||
 545                     rcu_scheduler_active != RCU_SCHEDULER_RUNNING ||
 546                     rcu_is_last_leaf_node(rnp)) {
 547                         /* No worker started yet or last leaf, do direct call. */
 548                         sync_rcu_exp_select_node_cpus(&rnp->rew.rew_work);
 549                         continue;
 550                 }
 551                 sync_rcu_exp_select_cpus_queue_work(rnp);
 552                 rnp->exp_need_flush = true;
 553         }
 554
 555         /* Wait for jobs (if any) to complete. */
 556         rcu_for_each_leaf_node(rnp)
 557                 if (rnp->exp_need_flush)
 558                         sync_rcu_exp_select_cpus_flush_work(rnp);
 559 }
 560
 561 /*
 562  * Wait for the expedited grace period to elapse, within time limit.
 563  * If the time limit is exceeded without the grace period elapsing,
 564  * return false, otherwise return true.
 565  */
 566 static bool synchronize_rcu_expedited_wait_once(long tlimit)
 567 {
 568         int t;
 569         struct rcu_node *rnp_root = rcu_get_root();
 570
 571         t = swait_event_timeout_exclusive(rcu_state.expedited_wq,
 572                                           sync_rcu_exp_done_unlocked(rnp_root),
 573                                           tlimit);
 574         // Workqueues should not be signaled.
 575         if (t > 0 || sync_rcu_exp_done_unlocked(rnp_root))
 576                 return true;
 577         WARN_ON(t < 0);  /* workqueues should not be signaled. */
 578         return false;
 579 }
 580
 581 /*
 582  * Wait for the expedited grace period to elapse, issuing any needed
 583  * RCU CPU stall warnings along the way.
 584  */
 585 static void synchronize_rcu_expedited_wait(void)
 586 {
 587         int cpu;
 588         unsigned long j;
 589         unsigned long jiffies_stall;
 590         unsigned long jiffies_start;
 591         unsigned long mask;
 592         int ndetected;
 593         struct rcu_data *rdp;
 594         struct rcu_node *rnp;
 595         struct rcu_node *rnp_root = rcu_get_root();
 596
 597         trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("startwait"));
 598         jiffies_stall = rcu_exp_jiffies_till_stall_check();
 599         jiffies_start = jiffies;
 600         if (tick_nohz_full_enabled() && rcu_inkernel_boot_has_ended()) {
 601                 if (synchronize_rcu_expedited_wait_once(1))
 602                         return;
 603                 rcu_for_each_leaf_node(rnp) {
 604                         mask = READ_ONCE(rnp->expmask);
 605                         for_each_leaf_node_cpu_mask(rnp, cpu, mask) {
 606                                 rdp = per_cpu_ptr(&rcu_data, cpu);
 607                                 if (rdp->rcu_forced_tick_exp)
 608                                         continue;
 609                                 rdp->rcu_forced_tick_exp = true;
 610                                 preempt_disable();
 611                                 if (cpu_online(cpu))
 612                                         tick_dep_set_cpu(cpu, TICK_DEP_BIT_RCU_EXP);
 613                                 preempt_enable();
 614                         }
 615                 }
 616                 j = READ_ONCE(jiffies_till_first_fqs);
 617                 if (synchronize_rcu_expedited_wait_once(j + HZ))
 618                         return;
 619         }
 620
 621         for (;;) {
 622                 if (synchronize_rcu_expedited_wait_once(jiffies_stall))
 623                         return;
 624                 if (rcu_stall_is_suppressed())
 625                         continue;
 626                 trace_rcu_stall_warning(rcu_state.name, TPS("ExpeditedStall"));
 627                 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
 628                        rcu_state.name);
 629                 ndetected = 0;
 630                 rcu_for_each_leaf_node(rnp) {
 631                         ndetected += rcu_print_task_exp_stall(rnp);
 632                         for_each_leaf_node_possible_cpu(rnp, cpu) {
 633                                 struct rcu_data *rdp;
 634
 635                                 mask = leaf_node_cpu_bit(rnp, cpu);
 636                                 if (!(READ_ONCE(rnp->expmask) & mask))
 637                                         continue;
 638                                 ndetected++;
 639                                 rdp = per_cpu_ptr(&rcu_data, cpu);
 640                                 pr_cont(" %d-%c%c%c%c", cpu,
 641                                         "O."[!!cpu_online(cpu)],
 642                                         "o."[!!(rdp->grpmask & rnp->expmaskinit)],
 643                                         "N."[!!(rdp->grpmask & rnp->expmaskinitnext)],
 644                                         "D."[!!(rdp->cpu_no_qs.b.exp)]);
 645                         }
 646                 }
 647                 pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
 648                         jiffies - jiffies_start, rcu_state.expedited_sequence,
 649                         data_race(rnp_root->expmask),
 650                         ".T"[!!data_race(rnp_root->exp_tasks)]);
 651                 if (ndetected) {
 652                         pr_err("blocking rcu_node structures (internal RCU debug):");
 653                         rcu_for_each_node_breadth_first(rnp) {
 654                                 if (rnp == rnp_root)
 655                                         continue; /* printed unconditionally */
 656                                 if (sync_rcu_exp_done_unlocked(rnp))
 657                                         continue;
 658                                 pr_cont(" l=%u:%d-%d:%#lx/%c",
 659                                         rnp->level, rnp->grplo, rnp->grphi,
 660                                         data_race(rnp->expmask),
 661                                         ".T"[!!data_race(rnp->exp_tasks)]);
 662                         }
 663                         pr_cont("\n");
 664                 }
 665                 rcu_for_each_leaf_node(rnp) {
 666                         for_each_leaf_node_possible_cpu(rnp, cpu) {
 667                                 mask = leaf_node_cpu_bit(rnp, cpu);
 668                                 if (!(READ_ONCE(rnp->expmask) & mask))
 669                                         continue;
 670                                 dump_cpu_task(cpu);
 671                         }
 672                 }
 673                 jiffies_stall = 3 * rcu_exp_jiffies_till_stall_check() + 3;
 674                 panic_on_rcu_stall();
 675         }
 676 }
 677
 678 /*
 679  * Wait for the current expedited grace period to complete, and then
 680  * wake up everyone who piggybacked on the just-completed expedited
 681  * grace period.  Also update all the ->exp_seq_rq counters as needed
 682  * in order to avoid counter-wrap problems.
 683  */
 684 static void rcu_exp_wait_wake(unsigned long s)
 685 {
 686         struct rcu_node *rnp;
 687
 688         synchronize_rcu_expedited_wait();
 689
 690         // Switch over to wakeup mode, allowing the next GP to proceed.
 691         // End the previous grace period only after acquiring the mutex
 692         // to ensure that only one GP runs concurrently with wakeups.
 693         mutex_lock(&rcu_state.exp_wake_mutex);
 694         rcu_exp_gp_seq_end();
 695         trace_rcu_exp_grace_period(rcu_state.name, s, TPS("end"));
 696
 697         rcu_for_each_node_breadth_first(rnp) {
 698                 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s)) {
 699                         spin_lock(&rnp->exp_lock);
 700                         /* Recheck, avoid hang in case someone just arrived. */
 701                         if (ULONG_CMP_LT(rnp->exp_seq_rq, s))
 702                                 WRITE_ONCE(rnp->exp_seq_rq, s);
 703                         spin_unlock(&rnp->exp_lock);
 704                 }
 705                 smp_mb(); /* All above changes before wakeup. */
 706                 wake_up_all(&rnp->exp_wq[rcu_seq_ctr(s) & 0x3]);
 707         }
 708         trace_rcu_exp_grace_period(rcu_state.name, s, TPS("endwake"));
 709         mutex_unlock(&rcu_state.exp_wake_mutex);
 710 }
 711
 712 /*
 713  * Common code to drive an expedited grace period forward, used by
 714  * workqueues and mid-boot-time tasks.
 715  */
 716 static void rcu_exp_sel_wait_wake(unsigned long s)
 717 {
 718         /* Initialize the rcu_node tree in preparation for the wait. */
 719         sync_rcu_exp_select_cpus();
 720
 721         /* Wait and clean up, including waking everyone. */
 722         rcu_exp_wait_wake(s);
 723 }
 724
 725 #ifdef CONFIG_PREEMPT_RCU
 726
 727 /*
 728  * Remote handler for smp_call_function_single().  If there is an
 729  * RCU read-side critical section in effect, request that the
 730  * next rcu_read_unlock() record the quiescent state up the
 731  * ->expmask fields in the rcu_node tree.  Otherwise, immediately
 732  * report the quiescent state.
 733  */
 734 static void rcu_exp_handler(void *unused)
 735 {
 736         int depth = rcu_preempt_depth();
 737         unsigned long flags;
 738         struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 739         struct rcu_node *rnp = rdp->mynode;
 740         struct task_struct *t = current;
 741
 742         /*
 743          * First, the common case of not being in an RCU read-side
 744          * critical section.  If also enabled or idle, immediately
 745          * report the quiescent state, otherwise defer.
 746          */
 747         if (!depth) {
 748                 if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
 749                     rcu_is_cpu_rrupt_from_idle()) {
 750                         rcu_report_exp_rdp(rdp);
 751                 } else {
 752                         WRITE_ONCE(rdp->cpu_no_qs.b.exp, true);
 753                         set_tsk_need_resched(t);
 754                         set_preempt_need_resched();
 755                 }
 756                 return;
 757         }
 758
 759         /*
 760          * Second, the less-common case of being in an RCU read-side
 761          * critical section.  In this case we can count on a future
 762          * rcu_read_unlock().  However, this rcu_read_unlock() might
 763          * execute on some other CPU, but in that case there will be
 764          * a future context switch.  Either way, if the expedited
 765          * grace period is still waiting on this CPU, set ->deferred_qs
 766          * so that the eventual quiescent state will be reported.
 767          * Note that there is a large group of race conditions that
 768          * can have caused this quiescent state to already have been
 769          * reported, so we really do need to check ->expmask.
 770          */
 771         if (depth > 0) {
 772                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 773                 if (rnp->expmask & rdp->grpmask) {
 774                         WRITE_ONCE(rdp->cpu_no_qs.b.exp, true);
 775                         t->rcu_read_unlock_special.b.exp_hint = true;
 776                 }
 777                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 778                 return;
 779         }
 780
 781         // Finally, negative nesting depth should not happen.
 782         WARN_ON_ONCE(1);
 783 }
 784
 785 /* PREEMPTION=y, so no PREEMPTION=n expedited grace period to clean up after. */
 786 static void sync_sched_exp_online_cleanup(int cpu)
 787 {
 788 }
 789
 790 /*
 791  * Scan the current list of tasks blocked within RCU read-side critical
 792  * sections, printing out the tid of each that is blocking the current
 793  * expedited grace period.
 794  */
 795 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
 796 {
 797         unsigned long flags;
 798         int ndetected = 0;
 799         struct task_struct *t;
 800
 801         if (!READ_ONCE(rnp->exp_tasks))
 802                 return 0;
 803         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 804         t = list_entry(rnp->exp_tasks->prev,
 805                        struct task_struct, rcu_node_entry);
 806         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
 807                 pr_cont(" P%d", t->pid);
 808                 ndetected++;
 809         }
 810         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 811         return ndetected;
 812 }
 813
 814 #else /* #ifdef CONFIG_PREEMPT_RCU */
 815
 816 /* Request an expedited quiescent state. */
 817 static void rcu_exp_need_qs(void)
 818 {
 819         __this_cpu_write(rcu_data.cpu_no_qs.b.exp, true);
 820         /* Store .exp before .rcu_urgent_qs. */
 821         smp_store_release(this_cpu_ptr(&rcu_data.rcu_urgent_qs), true);
 822         set_tsk_need_resched(current);
 823         set_preempt_need_resched();
 824 }
 825
 826 /* Invoked on each online non-idle CPU for expedited quiescent state. */
 827 static void rcu_exp_handler(void *unused)
 828 {
 829         struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 830         struct rcu_node *rnp = rdp->mynode;
 831         bool preempt_bh_enabled = !(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
 832
 833         if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
 834             __this_cpu_read(rcu_data.cpu_no_qs.b.exp))
 835                 return;
 836         if (rcu_is_cpu_rrupt_from_idle() ||
 837             (IS_ENABLED(CONFIG_PREEMPT_COUNT) && preempt_bh_enabled)) {
 838                 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
 839                 return;
 840         }
 841         rcu_exp_need_qs();
 842 }
 843
 844 /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
 845 static void sync_sched_exp_online_cleanup(int cpu)
 846 {
 847         unsigned long flags;
 848         int my_cpu;
 849         struct rcu_data *rdp;
 850         int ret;
 851         struct rcu_node *rnp;
 852
 853         rdp = per_cpu_ptr(&rcu_data, cpu);
 854         rnp = rdp->mynode;
 855         my_cpu = get_cpu();
 856         /* Quiescent state either not needed or already requested, leave. */
 857         if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
 858             READ_ONCE(rdp->cpu_no_qs.b.exp)) {
 859                 put_cpu();
 860                 return;
 861         }
 862         /* Quiescent state needed on current CPU, so set it up locally. */
 863         if (my_cpu == cpu) {
 864                 local_irq_save(flags);
 865                 rcu_exp_need_qs();
 866                 local_irq_restore(flags);
 867                 put_cpu();
 868                 return;
 869         }
 870         /* Quiescent state needed on some other CPU, send IPI. */
 871         ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
 872         put_cpu();
 873         WARN_ON_ONCE(ret);
 874 }
 875
 876 /*
 877  * Because preemptible RCU does not exist, we never have to check for
 878  * tasks blocked within RCU read-side critical sections that are
 879  * blocking the current expedited grace period.
 880  */
 881 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
 882 {
 883         return 0;
 884 }
 885
 886 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
 887
 888 /**
 889  * synchronize_rcu_expedited - Brute-force RCU grace period
 890  *
 891  * Wait for an RCU grace period, but expedite it.  The basic idea is to
 892  * IPI all non-idle non-nohz online CPUs.  The IPI handler checks whether
 893  * the CPU is in an RCU critical section, and if so, it sets a flag that
 894  * causes the outermost rcu_read_unlock() to report the quiescent state
 895  * for RCU-preempt or asks the scheduler for help for RCU-sched.  On the
 896  * other hand, if the CPU is not in an RCU read-side critical section,
 897  * the IPI handler reports the quiescent state immediately.
 898  *
 899  * Although this is a great improvement over previous expedited
 900  * implementations, it is still unfriendly to real-time workloads, so is
 901  * thus not recommended for any sort of common-case code.  In fact, if
 902  * you are using synchronize_rcu_expedited() in a loop, please restructure
 903  * your code to batch your updates, and then use a single synchronize_rcu()
 904  * instead.
 905  *
 906  * This has the same semantics as (but is more brutal than) synchronize_rcu().
 907  */
 908 void synchronize_rcu_expedited(void)
 909 {
 910         bool boottime = (rcu_scheduler_active == RCU_SCHEDULER_INIT);
 911         unsigned long flags;
 912         struct rcu_exp_work rew;
 913         struct rcu_node *rnp;
 914         unsigned long s;
 915
 916         RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
 917                          lock_is_held(&rcu_lock_map) ||
 918                          lock_is_held(&rcu_sched_lock_map),
 919                          "Illegal synchronize_rcu_expedited() in RCU read-side critical section");
 920
 921         /* Is the state is such that the call is a grace period? */
 922         if (rcu_blocking_is_gp()) {
 923                 // Note well that this code runs with !PREEMPT && !SMP.
 924                 // In addition, all code that advances grace periods runs
 925                 // at process level.  Therefore, this expedited GP overlaps
 926                 // with other expedited GPs only by being fully nested within
 927                 // them, which allows reuse of ->gp_seq_polled_exp_snap.
 928                 rcu_poll_gp_seq_start_unlocked(&rcu_state.gp_seq_polled_exp_snap);
 929                 rcu_poll_gp_seq_end_unlocked(&rcu_state.gp_seq_polled_exp_snap);
 930
 931                 local_irq_save(flags);
 932                 WARN_ON_ONCE(num_online_cpus() > 1);
 933                 rcu_state.expedited_sequence += (1 << RCU_SEQ_CTR_SHIFT);
 934                 local_irq_restore(flags);
 935                 return;  // Context allows vacuous grace periods.
 936         }
 937
 938         /* If expedited grace periods are prohibited, fall back to normal. */
 939         if (rcu_gp_is_normal()) {
 940                 wait_rcu_gp(call_rcu);
 941                 return;
 942         }
 943
 944         /* Take a snapshot of the sequence number.  */
 945         s = rcu_exp_gp_seq_snap();
 946         if (exp_funnel_lock(s))
 947                 return;  /* Someone else did our work for us. */
 948
 949         /* Ensure that load happens before action based on it. */
 950         if (unlikely(boottime)) {
 951                 /* Direct call during scheduler init and early_initcalls(). */
 952                 rcu_exp_sel_wait_wake(s);
 953         } else {
 954                 /* Marshall arguments & schedule the expedited grace period. */
 955                 rew.rew_s = s;
 956                 synchronize_rcu_expedited_queue_work(&rew);
 957         }
 958
 959         /* Wait for expedited grace period to complete. */
 960         rnp = rcu_get_root();
 961         wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
 962                    sync_exp_work_done(s));
 963         smp_mb(); /* Work actions happen before return. */
 964
 965         /* Let the next expedited grace period start. */
 966         mutex_unlock(&rcu_state.exp_mutex);
 967
 968         if (likely(!boottime))
 969                 synchronize_rcu_expedited_destroy_work(&rew);
 970 }
 971 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
 972
 973 /*
 974  * Ensure that start_poll_synchronize_rcu_expedited() has the expedited
 975  * RCU grace periods that it needs.
 976  */
 977 static void sync_rcu_do_polled_gp(struct work_struct *wp)
 978 {
 979         unsigned long flags;
 980         int i = 0;
 981         struct rcu_node *rnp = container_of(wp, struct rcu_node, exp_poll_wq);
 982         unsigned long s;
 983
 984         raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
 985         s = rnp->exp_seq_poll_rq;
 986         rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED;
 987         raw_spin_unlock_irqrestore(&rnp->exp_poll_lock, flags);
 988         if (s == RCU_GET_STATE_COMPLETED)
 989                 return;
 990         while (!poll_state_synchronize_rcu(s)) {
 991                 synchronize_rcu_expedited();
 992                 if (i == 10 || i == 20)
 993                         pr_info("%s: i = %d s = %lx gp_seq_polled = %lx\n", __func__, i, s, READ_ONCE(rcu_state.gp_seq_polled));
 994                 i++;
 995         }
 996         raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
 997         s = rnp->exp_seq_poll_rq;
 998         if (poll_state_synchronize_rcu(s))
 999                 rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED;
1000         raw_spin_unlock_irqrestore(&rnp->exp_poll_lock, flags);
1001 }
1002
1003 /**
1004  * start_poll_synchronize_rcu_expedited - Snapshot current RCU state and start expedited grace period
1005  *
1006  * Returns a cookie to pass to a call to cond_synchronize_rcu(),
1007  * cond_synchronize_rcu_expedited(), or poll_state_synchronize_rcu(),
1008  * allowing them to determine whether or not any sort of grace period has
1009  * elapsed in the meantime.  If the needed expedited grace period is not
1010  * already slated to start, initiates that grace period.
1011  */
1012 unsigned long start_poll_synchronize_rcu_expedited(void)
1013 {
1014         unsigned long flags;
1015         struct rcu_data *rdp;
1016         struct rcu_node *rnp;
1017         unsigned long s;
1018
1019         s = get_state_synchronize_rcu();
1020         rdp = per_cpu_ptr(&rcu_data, raw_smp_processor_id());
1021         rnp = rdp->mynode;
1022         if (rcu_init_invoked())
1023                 raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
1024         if (!poll_state_synchronize_rcu(s)) {
1025                 rnp->exp_seq_poll_rq = s;
1026                 if (rcu_init_invoked())
1027                         queue_work(rcu_gp_wq, &rnp->exp_poll_wq);
1028         }
1029         if (rcu_init_invoked())
1030                 raw_spin_unlock_irqrestore(&rnp->exp_poll_lock, flags);
1031
1032         return s;
1033 }
1034 EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_expedited);
1035
1036 /**
1037  * start_poll_synchronize_rcu_expedited_full - Take a full snapshot and start expedited grace period
1038  * @rgosp: Place to put snapshot of grace-period state
1039  *
1040  * Places the normal and expedited grace-period states in rgosp.  This
1041  * state value can be passed to a later call to cond_synchronize_rcu_full()
1042  * or poll_state_synchronize_rcu_full() to determine whether or not a
1043  * grace period (whether normal or expedited) has elapsed in the meantime.
1044  * If the needed expedited grace period is not already slated to start,
1045  * initiates that grace period.
1046  */
1047 void start_poll_synchronize_rcu_expedited_full(struct rcu_gp_oldstate *rgosp)
1048 {
1049         get_state_synchronize_rcu_full(rgosp);
1050         (void)start_poll_synchronize_rcu_expedited();
1051 }
1052 EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_expedited_full);
1053
1054 /**
1055  * cond_synchronize_rcu_expedited - Conditionally wait for an expedited RCU grace period
1056  *
1057  * @oldstate: value from get_state_synchronize_rcu(), start_poll_synchronize_rcu(), or start_poll_synchronize_rcu_expedited()
1058  *
1059  * If any type of full RCU grace period has elapsed since the earlier
1060  * call to get_state_synchronize_rcu(), start_poll_synchronize_rcu(),
1061  * or start_poll_synchronize_rcu_expedited(), just return.  Otherwise,
1062  * invoke synchronize_rcu_expedited() to wait for a full grace period.
1063  *
1064  * Yes, this function does not take counter wrap into account.
1065  * But counter wrap is harmless.  If the counter wraps, we have waited for
1066  * more than 2 billion grace periods (and way more on a 64-bit system!),
1067  * so waiting for a couple of additional grace periods should be just fine.
1068  *
1069  * This function provides the same memory-ordering guarantees that
1070  * would be provided by a synchronize_rcu() that was invoked at the call
1071  * to the function that provided @oldstate and that returned at the end
1072  * of this function.
1073  */
1074 void cond_synchronize_rcu_expedited(unsigned long oldstate)
1075 {
1076         if (!poll_state_synchronize_rcu(oldstate))
1077                 synchronize_rcu_expedited();
1078 }
1079 EXPORT_SYMBOL_GPL(cond_synchronize_rcu_expedited);
1080
1081 /**
1082  * cond_synchronize_rcu_expedited_full - Conditionally wait for an expedited RCU grace period
1083  * @rgosp: value from get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(), or start_poll_synchronize_rcu_expedited_full()
1084  *
1085  * If a full RCU grace period has elapsed since the call to
1086  * get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(),
1087  * or start_poll_synchronize_rcu_expedited_full() from which @rgosp was
1088  * obtained, just return.  Otherwise, invoke synchronize_rcu_expedited()
1089  * to wait for a full grace period.
1090  *
1091  * Yes, this function does not take counter wrap into account.
1092  * But counter wrap is harmless.  If the counter wraps, we have waited for
1093  * more than 2 billion grace periods (and way more on a 64-bit system!),
1094  * so waiting for a couple of additional grace periods should be just fine.
1095  *
1096  * This function provides the same memory-ordering guarantees that
1097  * would be provided by a synchronize_rcu() that was invoked at the call
1098  * to the function that provided @rgosp and that returned at the end of
1099  * this function.
1100  */
1101 void cond_synchronize_rcu_expedited_full(struct rcu_gp_oldstate *rgosp)
1102 {
1103         if (!poll_state_synchronize_rcu_full(rgosp))
1104                 synchronize_rcu_expedited();
1105 }
1106 EXPORT_SYMBOL_GPL(cond_synchronize_rcu_expedited_full);