kernel/rcu/tree_stall.h

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * RCU CPU stall warnings for normal RCU grace periods
   4  *
   5  * Copyright IBM Corporation, 2019
   6  *
   7  * Author: Paul E. McKenney <paulmck@linux.ibm.com>
   8  */
   9
  10 //////////////////////////////////////////////////////////////////////////////
  11 //
  12 // Controlling CPU stall warnings, including delay calculation.
  13
  14 /* panic() on RCU Stall sysctl. */
  15 int sysctl_panic_on_rcu_stall __read_mostly;
  16
  17 #ifdef CONFIG_PROVE_RCU
  18 #define RCU_STALL_DELAY_DELTA           (5 * HZ)
  19 #else
  20 #define RCU_STALL_DELAY_DELTA           0
  21 #endif
  22 #define RCU_STALL_MIGHT_DIV             8
  23 #define RCU_STALL_MIGHT_MIN             (2 * HZ)
  24
  25 /* Limit-check stall timeouts specified at boottime and runtime. */
  26 int rcu_jiffies_till_stall_check(void)
  27 {
  28         int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
  29
  30         /*
  31          * Limit check must be consistent with the Kconfig limits
  32          * for CONFIG_RCU_CPU_STALL_TIMEOUT.
  33          */
  34         if (till_stall_check < 3) {
  35                 WRITE_ONCE(rcu_cpu_stall_timeout, 3);
  36                 till_stall_check = 3;
  37         } else if (till_stall_check > 300) {
  38                 WRITE_ONCE(rcu_cpu_stall_timeout, 300);
  39                 till_stall_check = 300;
  40         }
  41         return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
  42 }
  43 EXPORT_SYMBOL_GPL(rcu_jiffies_till_stall_check);
  44
  45 /**
  46  * rcu_gp_might_be_stalled - Is it likely that the grace period is stalled?
  47  *
  48  * Returns @true if the current grace period is sufficiently old that
  49  * it is reasonable to assume that it might be stalled.  This can be
  50  * useful when deciding whether to allocate memory to enable RCU-mediated
  51  * freeing on the one hand or just invoking synchronize_rcu() on the other.
  52  * The latter is preferable when the grace period is stalled.
  53  *
  54  * Note that sampling of the .gp_start and .gp_seq fields must be done
  55  * carefully to avoid false positives at the beginnings and ends of
  56  * grace periods.
  57  */
  58 bool rcu_gp_might_be_stalled(void)
  59 {
  60         unsigned long d = rcu_jiffies_till_stall_check() / RCU_STALL_MIGHT_DIV;
  61         unsigned long j = jiffies;
  62
  63         if (d < RCU_STALL_MIGHT_MIN)
  64                 d = RCU_STALL_MIGHT_MIN;
  65         smp_mb(); // jiffies before .gp_seq to avoid false positives.
  66         if (!rcu_gp_in_progress())
  67                 return false;
  68         // Long delays at this point avoids false positive, but a delay
  69         // of ULONG_MAX/4 jiffies voids your no-false-positive warranty.
  70         smp_mb(); // .gp_seq before second .gp_start
  71         // And ditto here.
  72         return !time_before(j, READ_ONCE(rcu_state.gp_start) + d);
  73 }
  74
  75 /* Don't do RCU CPU stall warnings during long sysrq printouts. */
  76 void rcu_sysrq_start(void)
  77 {
  78         if (!rcu_cpu_stall_suppress)
  79                 rcu_cpu_stall_suppress = 2;
  80 }
  81
  82 void rcu_sysrq_end(void)
  83 {
  84         if (rcu_cpu_stall_suppress == 2)
  85                 rcu_cpu_stall_suppress = 0;
  86 }
  87
  88 /* Don't print RCU CPU stall warnings during a kernel panic. */
  89 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
  90 {
  91         rcu_cpu_stall_suppress = 1;
  92         return NOTIFY_DONE;
  93 }
  94
  95 static struct notifier_block rcu_panic_block = {
  96         .notifier_call = rcu_panic,
  97 };
  98
  99 static int __init check_cpu_stall_init(void)
 100 {
 101         atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
 102         return 0;
 103 }
 104 early_initcall(check_cpu_stall_init);
 105
 106 /* If so specified via sysctl, panic, yielding cleaner stall-warning output. */
 107 static void panic_on_rcu_stall(void)
 108 {
 109         if (sysctl_panic_on_rcu_stall)
 110                 panic("RCU Stall\n");
 111 }
 112
 113 /**
 114  * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
 115  *
 116  * Set the stall-warning timeout way off into the future, thus preventing
 117  * any RCU CPU stall-warning messages from appearing in the current set of
 118  * RCU grace periods.
 119  *
 120  * The caller must disable hard irqs.
 121  */
 122 void rcu_cpu_stall_reset(void)
 123 {
 124         WRITE_ONCE(rcu_state.jiffies_stall, jiffies + ULONG_MAX / 2);
 125 }
 126
 127 //////////////////////////////////////////////////////////////////////////////
 128 //
 129 // Interaction with RCU grace periods
 130
 131 /* Start of new grace period, so record stall time (and forcing times). */
 132 static void record_gp_stall_check_time(void)
 133 {
 134         unsigned long j = jiffies;
 135         unsigned long j1;
 136
 137         WRITE_ONCE(rcu_state.gp_start, j);
 138         j1 = rcu_jiffies_till_stall_check();
 139         smp_mb(); // ->gp_start before ->jiffies_stall and caller's ->gp_seq.
 140         WRITE_ONCE(rcu_state.jiffies_stall, j + j1);
 141         rcu_state.jiffies_resched = j + j1 / 2;
 142         rcu_state.n_force_qs_gpstart = READ_ONCE(rcu_state.n_force_qs);
 143 }
 144
 145 /* Zero ->ticks_this_gp and snapshot the number of RCU softirq handlers. */
 146 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
 147 {
 148         rdp->ticks_this_gp = 0;
 149         rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
 150         WRITE_ONCE(rdp->last_fqs_resched, jiffies);
 151 }
 152
 153 /*
 154  * If too much time has passed in the current grace period, and if
 155  * so configured, go kick the relevant kthreads.
 156  */
 157 static void rcu_stall_kick_kthreads(void)
 158 {
 159         unsigned long j;
 160
 161         if (!rcu_kick_kthreads)
 162                 return;
 163         j = READ_ONCE(rcu_state.jiffies_kick_kthreads);
 164         if (time_after(jiffies, j) && rcu_state.gp_kthread &&
 165             (rcu_gp_in_progress() || READ_ONCE(rcu_state.gp_flags))) {
 166                 WARN_ONCE(1, "Kicking %s grace-period kthread\n",
 167                           rcu_state.name);
 168                 rcu_ftrace_dump(DUMP_ALL);
 169                 wake_up_process(rcu_state.gp_kthread);
 170                 WRITE_ONCE(rcu_state.jiffies_kick_kthreads, j + HZ);
 171         }
 172 }
 173
 174 /*
 175  * Handler for the irq_work request posted about halfway into the RCU CPU
 176  * stall timeout, and used to detect excessive irq disabling.  Set state
 177  * appropriately, but just complain if there is unexpected state on entry.
 178  */
 179 static void rcu_iw_handler(struct irq_work *iwp)
 180 {
 181         struct rcu_data *rdp;
 182         struct rcu_node *rnp;
 183
 184         rdp = container_of(iwp, struct rcu_data, rcu_iw);
 185         rnp = rdp->mynode;
 186         raw_spin_lock_rcu_node(rnp);
 187         if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
 188                 rdp->rcu_iw_gp_seq = rnp->gp_seq;
 189                 rdp->rcu_iw_pending = false;
 190         }
 191         raw_spin_unlock_rcu_node(rnp);
 192 }
 193
 194 //////////////////////////////////////////////////////////////////////////////
 195 //
 196 // Printing RCU CPU stall warnings
 197
 198 #ifdef CONFIG_PREEMPT_RCU
 199
 200 /*
 201  * Dump detailed information for all tasks blocking the current RCU
 202  * grace period on the specified rcu_node structure.
 203  */
 204 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
 205 {
 206         unsigned long flags;
 207         struct task_struct *t;
 208
 209         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 210         if (!rcu_preempt_blocked_readers_cgp(rnp)) {
 211                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 212                 return;
 213         }
 214         t = list_entry(rnp->gp_tasks->prev,
 215                        struct task_struct, rcu_node_entry);
 216         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
 217                 /*
 218                  * We could be printing a lot while holding a spinlock.
 219                  * Avoid triggering hard lockup.
 220                  */
 221                 touch_nmi_watchdog();
 222                 sched_show_task(t);
 223         }
 224         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 225 }
 226
 227 // Communicate task state back to the RCU CPU stall warning request.
 228 struct rcu_stall_chk_rdr {
 229         int nesting;
 230         union rcu_special rs;
 231         bool on_blkd_list;
 232 };
 233
 234 /*
 235  * Report out the state of a not-running task that is stalling the
 236  * current RCU grace period.
 237  */
 238 static bool check_slow_task(struct task_struct *t, void *arg)
 239 {
 240         struct rcu_node *rnp;
 241         struct rcu_stall_chk_rdr *rscrp = arg;
 242
 243         if (task_curr(t))
 244                 return false; // It is running, so decline to inspect it.
 245         rscrp->nesting = t->rcu_read_lock_nesting;
 246         rscrp->rs = t->rcu_read_unlock_special;
 247         rnp = t->rcu_blocked_node;
 248         rscrp->on_blkd_list = !list_empty(&t->rcu_node_entry);
 249         return true;
 250 }
 251
 252 /*
 253  * Scan the current list of tasks blocked within RCU read-side critical
 254  * sections, printing out the tid of each.
 255  */
 256 static int rcu_print_task_stall(struct rcu_node *rnp)
 257 {
 258         int ndetected = 0;
 259         struct rcu_stall_chk_rdr rscr;
 260         struct task_struct *t;
 261
 262         if (!rcu_preempt_blocked_readers_cgp(rnp))
 263                 return 0;
 264         pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
 265                rnp->level, rnp->grplo, rnp->grphi);
 266         t = list_entry(rnp->gp_tasks->prev,
 267                        struct task_struct, rcu_node_entry);
 268         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
 269                 if (!try_invoke_on_locked_down_task(t, check_slow_task, &rscr))
 270                         pr_cont(" P%d", t->pid);
 271                 else
 272                         pr_cont(" P%d/%d:%c%c%c%c",
 273                                 t->pid, rscr.nesting,
 274                                 ".b"[rscr.rs.b.blocked],
 275                                 ".q"[rscr.rs.b.need_qs],
 276                                 ".e"[rscr.rs.b.exp_hint],
 277                                 ".l"[rscr.on_blkd_list]);
 278                 ndetected++;
 279         }
 280         pr_cont("\n");
 281         return ndetected;
 282 }
 283
 284 #else /* #ifdef CONFIG_PREEMPT_RCU */
 285
 286 /*
 287  * Because preemptible RCU does not exist, we never have to check for
 288  * tasks blocked within RCU read-side critical sections.
 289  */
 290 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
 291 {
 292 }
 293
 294 /*
 295  * Because preemptible RCU does not exist, we never have to check for
 296  * tasks blocked within RCU read-side critical sections.
 297  */
 298 static int rcu_print_task_stall(struct rcu_node *rnp)
 299 {
 300         return 0;
 301 }
 302 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
 303
 304 /*
 305  * Dump stacks of all tasks running on stalled CPUs.  First try using
 306  * NMIs, but fall back to manual remote stack tracing on architectures
 307  * that don't support NMI-based stack dumps.  The NMI-triggered stack
 308  * traces are more accurate because they are printed by the target CPU.
 309  */
 310 static void rcu_dump_cpu_stacks(void)
 311 {
 312         int cpu;
 313         unsigned long flags;
 314         struct rcu_node *rnp;
 315
 316         rcu_for_each_leaf_node(rnp) {
 317                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 318                 for_each_leaf_node_possible_cpu(rnp, cpu)
 319                         if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
 320                                 if (!trigger_single_cpu_backtrace(cpu))
 321                                         dump_cpu_task(cpu);
 322                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 323         }
 324 }
 325
 326 #ifdef CONFIG_RCU_FAST_NO_HZ
 327
 328 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
 329 {
 330         struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
 331
 332         sprintf(cp, "last_accelerate: %04lx/%04lx dyntick_enabled: %d",
 333                 rdp->last_accelerate & 0xffff, jiffies & 0xffff,
 334                 !!rdp->tick_nohz_enabled_snap);
 335 }
 336
 337 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
 338
 339 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
 340 {
 341         *cp = '\0';
 342 }
 343
 344 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
 345
 346 static const char * const gp_state_names[] = {
 347         [RCU_GP_IDLE] = "RCU_GP_IDLE",
 348         [RCU_GP_WAIT_GPS] = "RCU_GP_WAIT_GPS",
 349         [RCU_GP_DONE_GPS] = "RCU_GP_DONE_GPS",
 350         [RCU_GP_ONOFF] = "RCU_GP_ONOFF",
 351         [RCU_GP_INIT] = "RCU_GP_INIT",
 352         [RCU_GP_WAIT_FQS] = "RCU_GP_WAIT_FQS",
 353         [RCU_GP_DOING_FQS] = "RCU_GP_DOING_FQS",
 354         [RCU_GP_CLEANUP] = "RCU_GP_CLEANUP",
 355         [RCU_GP_CLEANED] = "RCU_GP_CLEANED",
 356 };
 357
 358 /*
 359  * Convert a ->gp_state value to a character string.
 360  */
 361 static const char *gp_state_getname(short gs)
 362 {
 363         if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
 364                 return "???";
 365         return gp_state_names[gs];
 366 }
 367
 368 /* Is the RCU grace-period kthread being starved of CPU time? */
 369 static bool rcu_is_gp_kthread_starving(unsigned long *jp)
 370 {
 371         unsigned long j = jiffies - READ_ONCE(rcu_state.gp_activity);
 372
 373         if (jp)
 374                 *jp = j;
 375         return j > 2 * HZ;
 376 }
 377
 378 /*
 379  * Print out diagnostic information for the specified stalled CPU.
 380  *
 381  * If the specified CPU is aware of the current RCU grace period, then
 382  * print the number of scheduling clock interrupts the CPU has taken
 383  * during the time that it has been aware.  Otherwise, print the number
 384  * of RCU grace periods that this CPU is ignorant of, for example, "1"
 385  * if the CPU was aware of the previous grace period.
 386  *
 387  * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
 388  */
 389 static void print_cpu_stall_info(int cpu)
 390 {
 391         unsigned long delta;
 392         bool falsepositive;
 393         char fast_no_hz[72];
 394         struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 395         char *ticks_title;
 396         unsigned long ticks_value;
 397
 398         /*
 399          * We could be printing a lot while holding a spinlock.  Avoid
 400          * triggering hard lockup.
 401          */
 402         touch_nmi_watchdog();
 403
 404         ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
 405         if (ticks_value) {
 406                 ticks_title = "GPs behind";
 407         } else {
 408                 ticks_title = "ticks this GP";
 409                 ticks_value = rdp->ticks_this_gp;
 410         }
 411         print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
 412         delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
 413         falsepositive = rcu_is_gp_kthread_starving(NULL) &&
 414                         rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp));
 415         pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%03x/%ld/%#lx softirq=%u/%u fqs=%ld %s%s\n",
 416                cpu,
 417                "O."[!!cpu_online(cpu)],
 418                "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
 419                "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
 420                !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
 421                         rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
 422                                 "!."[!delta],
 423                ticks_value, ticks_title,
 424                rcu_dynticks_snap(rdp) & 0xfff,
 425                rdp->dynticks_nesting, rdp->dynticks_nmi_nesting,
 426                rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
 427                data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
 428                fast_no_hz,
 429                falsepositive ? " (false positive?)" : "");
 430 }
 431
 432 /* Complain about starvation of grace-period kthread.  */
 433 static void rcu_check_gp_kthread_starvation(void)
 434 {
 435         struct task_struct *gpk = rcu_state.gp_kthread;
 436         unsigned long j;
 437
 438         if (rcu_is_gp_kthread_starving(&j)) {
 439                 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#lx ->cpu=%d\n",
 440                        rcu_state.name, j,
 441                        (long)rcu_seq_current(&rcu_state.gp_seq),
 442                        data_race(rcu_state.gp_flags),
 443                        gp_state_getname(rcu_state.gp_state), rcu_state.gp_state,
 444                        gpk ? gpk->state : ~0, gpk ? task_cpu(gpk) : -1);
 445                 if (gpk) {
 446                         pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name);
 447                         pr_err("RCU grace-period kthread stack dump:\n");
 448                         sched_show_task(gpk);
 449                         wake_up_process(gpk);
 450                 }
 451         }
 452 }
 453
 454 static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
 455 {
 456         int cpu;
 457         unsigned long flags;
 458         unsigned long gpa;
 459         unsigned long j;
 460         int ndetected = 0;
 461         struct rcu_node *rnp;
 462         long totqlen = 0;
 463
 464         /* Kick and suppress, if so configured. */
 465         rcu_stall_kick_kthreads();
 466         if (rcu_stall_is_suppressed())
 467                 return;
 468
 469         /*
 470          * OK, time to rat on our buddy...
 471          * See Documentation/RCU/stallwarn.txt for info on how to debug
 472          * RCU CPU stall warnings.
 473          */
 474         pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
 475         rcu_for_each_leaf_node(rnp) {
 476                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 477                 ndetected += rcu_print_task_stall(rnp);
 478                 if (rnp->qsmask != 0) {
 479                         for_each_leaf_node_possible_cpu(rnp, cpu)
 480                                 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
 481                                         print_cpu_stall_info(cpu);
 482                                         ndetected++;
 483                                 }
 484                 }
 485                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 486         }
 487
 488         for_each_possible_cpu(cpu)
 489                 totqlen += rcu_get_n_cbs_cpu(cpu);
 490         pr_cont("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu)\n",
 491                smp_processor_id(), (long)(jiffies - gps),
 492                (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
 493         if (ndetected) {
 494                 rcu_dump_cpu_stacks();
 495
 496                 /* Complain about tasks blocking the grace period. */
 497                 rcu_for_each_leaf_node(rnp)
 498                         rcu_print_detail_task_stall_rnp(rnp);
 499         } else {
 500                 if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
 501                         pr_err("INFO: Stall ended before state dump start\n");
 502                 } else {
 503                         j = jiffies;
 504                         gpa = data_race(rcu_state.gp_activity);
 505                         pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
 506                                rcu_state.name, j - gpa, j, gpa,
 507                                data_race(jiffies_till_next_fqs),
 508                                rcu_get_root()->qsmask);
 509                 }
 510         }
 511         /* Rewrite if needed in case of slow consoles. */
 512         if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
 513                 WRITE_ONCE(rcu_state.jiffies_stall,
 514                            jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
 515
 516         rcu_check_gp_kthread_starvation();
 517
 518         panic_on_rcu_stall();
 519
 520         rcu_force_quiescent_state();  /* Kick them all. */
 521 }
 522
 523 static void print_cpu_stall(unsigned long gps)
 524 {
 525         int cpu;
 526         unsigned long flags;
 527         struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 528         struct rcu_node *rnp = rcu_get_root();
 529         long totqlen = 0;
 530
 531         /* Kick and suppress, if so configured. */
 532         rcu_stall_kick_kthreads();
 533         if (rcu_stall_is_suppressed())
 534                 return;
 535
 536         /*
 537          * OK, time to rat on ourselves...
 538          * See Documentation/RCU/stallwarn.txt for info on how to debug
 539          * RCU CPU stall warnings.
 540          */
 541         pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
 542         raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
 543         print_cpu_stall_info(smp_processor_id());
 544         raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
 545         for_each_possible_cpu(cpu)
 546                 totqlen += rcu_get_n_cbs_cpu(cpu);
 547         pr_cont("\t(t=%lu jiffies g=%ld q=%lu)\n",
 548                 jiffies - gps,
 549                 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
 550
 551         rcu_check_gp_kthread_starvation();
 552
 553         rcu_dump_cpu_stacks();
 554
 555         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 556         /* Rewrite if needed in case of slow consoles. */
 557         if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
 558                 WRITE_ONCE(rcu_state.jiffies_stall,
 559                            jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
 560         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 561
 562         panic_on_rcu_stall();
 563
 564         /*
 565          * Attempt to revive the RCU machinery by forcing a context switch.
 566          *
 567          * A context switch would normally allow the RCU state machine to make
 568          * progress and it could be we're stuck in kernel space without context
 569          * switches for an entirely unreasonable amount of time.
 570          */
 571         set_tsk_need_resched(current);
 572         set_preempt_need_resched();
 573 }
 574
 575 static void check_cpu_stall(struct rcu_data *rdp)
 576 {
 577         unsigned long gs1;
 578         unsigned long gs2;
 579         unsigned long gps;
 580         unsigned long j;
 581         unsigned long jn;
 582         unsigned long js;
 583         struct rcu_node *rnp;
 584
 585         if ((rcu_stall_is_suppressed() && !rcu_kick_kthreads) ||
 586             !rcu_gp_in_progress())
 587                 return;
 588         rcu_stall_kick_kthreads();
 589         j = jiffies;
 590
 591         /*
 592          * Lots of memory barriers to reject false positives.
 593          *
 594          * The idea is to pick up rcu_state.gp_seq, then
 595          * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
 596          * another copy of rcu_state.gp_seq.  These values are updated in
 597          * the opposite order with memory barriers (or equivalent) during
 598          * grace-period initialization and cleanup.  Now, a false positive
 599          * can occur if we get an new value of rcu_state.gp_start and a old
 600          * value of rcu_state.jiffies_stall.  But given the memory barriers,
 601          * the only way that this can happen is if one grace period ends
 602          * and another starts between these two fetches.  This is detected
 603          * by comparing the second fetch of rcu_state.gp_seq with the
 604          * previous fetch from rcu_state.gp_seq.
 605          *
 606          * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
 607          * and rcu_state.gp_start suffice to forestall false positives.
 608          */
 609         gs1 = READ_ONCE(rcu_state.gp_seq);
 610         smp_rmb(); /* Pick up ->gp_seq first... */
 611         js = READ_ONCE(rcu_state.jiffies_stall);
 612         smp_rmb(); /* ...then ->jiffies_stall before the rest... */
 613         gps = READ_ONCE(rcu_state.gp_start);
 614         smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
 615         gs2 = READ_ONCE(rcu_state.gp_seq);
 616         if (gs1 != gs2 ||
 617             ULONG_CMP_LT(j, js) ||
 618             ULONG_CMP_GE(gps, js))
 619                 return; /* No stall or GP completed since entering function. */
 620         rnp = rdp->mynode;
 621         jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
 622         if (rcu_gp_in_progress() &&
 623             (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
 624             cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
 625
 626                 /* We haven't checked in, so go dump stack. */
 627                 print_cpu_stall(gps);
 628                 if (rcu_cpu_stall_ftrace_dump)
 629                         rcu_ftrace_dump(DUMP_ALL);
 630
 631         } else if (rcu_gp_in_progress() &&
 632                    ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
 633                    cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
 634
 635                 /* They had a few time units to dump stack, so complain. */
 636                 print_other_cpu_stall(gs2, gps);
 637                 if (rcu_cpu_stall_ftrace_dump)
 638                         rcu_ftrace_dump(DUMP_ALL);
 639         }
 640 }
 641
 642 //////////////////////////////////////////////////////////////////////////////
 643 //
 644 // RCU forward-progress mechanisms, including of callback invocation.
 645
 646
 647 /*
 648  * Show the state of the grace-period kthreads.
 649  */
 650 void show_rcu_gp_kthreads(void)
 651 {
 652         int cpu;
 653         unsigned long j;
 654         unsigned long ja;
 655         unsigned long jr;
 656         unsigned long jw;
 657         struct rcu_data *rdp;
 658         struct rcu_node *rnp;
 659         struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
 660
 661         j = jiffies;
 662         ja = j - data_race(rcu_state.gp_activity);
 663         jr = j - data_race(rcu_state.gp_req_activity);
 664         jw = j - data_race(rcu_state.gp_wake_time);
 665         pr_info("%s: wait state: %s(%d) ->state: %#lx delta ->gp_activity %lu ->gp_req_activity %lu ->gp_wake_time %lu ->gp_wake_seq %ld ->gp_seq %ld ->gp_seq_needed %ld ->gp_flags %#x\n",
 666                 rcu_state.name, gp_state_getname(rcu_state.gp_state),
 667                 rcu_state.gp_state, t ? t->state : 0x1ffffL,
 668                 ja, jr, jw, (long)data_race(rcu_state.gp_wake_seq),
 669                 (long)data_race(rcu_state.gp_seq),
 670                 (long)data_race(rcu_get_root()->gp_seq_needed),
 671                 data_race(rcu_state.gp_flags));
 672         rcu_for_each_node_breadth_first(rnp) {
 673                 if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
 674                                  READ_ONCE(rnp->gp_seq_needed)))
 675                         continue;
 676                 pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld\n",
 677                         rnp->grplo, rnp->grphi, (long)data_race(rnp->gp_seq),
 678                         (long)data_race(rnp->gp_seq_needed));
 679                 if (!rcu_is_leaf_node(rnp))
 680                         continue;
 681                 for_each_leaf_node_possible_cpu(rnp, cpu) {
 682                         rdp = per_cpu_ptr(&rcu_data, cpu);
 683                         if (READ_ONCE(rdp->gpwrap) ||
 684                             ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
 685                                          READ_ONCE(rdp->gp_seq_needed)))
 686                                 continue;
 687                         pr_info("\tcpu %d ->gp_seq_needed %ld\n",
 688                                 cpu, (long)data_race(rdp->gp_seq_needed));
 689                 }
 690         }
 691         for_each_possible_cpu(cpu) {
 692                 rdp = per_cpu_ptr(&rcu_data, cpu);
 693                 if (rcu_segcblist_is_offloaded(&rdp->cblist))
 694                         show_rcu_nocb_state(rdp);
 695         }
 696         show_rcu_tasks_gp_kthreads();
 697 }
 698 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
 699
 700 /*
 701  * This function checks for grace-period requests that fail to motivate
 702  * RCU to come out of its idle mode.
 703  */
 704 static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
 705                                      const unsigned long gpssdelay)
 706 {
 707         unsigned long flags;
 708         unsigned long j;
 709         struct rcu_node *rnp_root = rcu_get_root();
 710         static atomic_t warned = ATOMIC_INIT(0);
 711
 712         if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
 713             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 714                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 715             !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread.
 716                 return;
 717         j = jiffies; /* Expensive access, and in common case don't get here. */
 718         if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 719             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 720             atomic_read(&warned))
 721                 return;
 722
 723         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 724         j = jiffies;
 725         if (rcu_gp_in_progress() ||
 726             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 727                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 728             time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 729             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 730             atomic_read(&warned)) {
 731                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 732                 return;
 733         }
 734         /* Hold onto the leaf lock to make others see warned==1. */
 735
 736         if (rnp_root != rnp)
 737                 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
 738         j = jiffies;
 739         if (rcu_gp_in_progress() ||
 740             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 741                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 742             time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 743             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 744             atomic_xchg(&warned, 1)) {
 745                 if (rnp_root != rnp)
 746                         /* irqs remain disabled. */
 747                         raw_spin_unlock_rcu_node(rnp_root);
 748                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 749                 return;
 750         }
 751         WARN_ON(1);
 752         if (rnp_root != rnp)
 753                 raw_spin_unlock_rcu_node(rnp_root);
 754         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 755         show_rcu_gp_kthreads();
 756 }
 757
 758 /*
 759  * Do a forward-progress check for rcutorture.  This is normally invoked
 760  * due to an OOM event.  The argument "j" gives the time period during
 761  * which rcutorture would like progress to have been made.
 762  */
 763 void rcu_fwd_progress_check(unsigned long j)
 764 {
 765         unsigned long cbs;
 766         int cpu;
 767         unsigned long max_cbs = 0;
 768         int max_cpu = -1;
 769         struct rcu_data *rdp;
 770
 771         if (rcu_gp_in_progress()) {
 772                 pr_info("%s: GP age %lu jiffies\n",
 773                         __func__, jiffies - rcu_state.gp_start);
 774                 show_rcu_gp_kthreads();
 775         } else {
 776                 pr_info("%s: Last GP end %lu jiffies ago\n",
 777                         __func__, jiffies - rcu_state.gp_end);
 778                 preempt_disable();
 779                 rdp = this_cpu_ptr(&rcu_data);
 780                 rcu_check_gp_start_stall(rdp->mynode, rdp, j);
 781                 preempt_enable();
 782         }
 783         for_each_possible_cpu(cpu) {
 784                 cbs = rcu_get_n_cbs_cpu(cpu);
 785                 if (!cbs)
 786                         continue;
 787                 if (max_cpu < 0)
 788                         pr_info("%s: callbacks", __func__);
 789                 pr_cont(" %d: %lu", cpu, cbs);
 790                 if (cbs <= max_cbs)
 791                         continue;
 792                 max_cbs = cbs;
 793                 max_cpu = cpu;
 794         }
 795         if (max_cpu >= 0)
 796                 pr_cont("\n");
 797 }
 798 EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
 799
 800 /* Commandeer a sysrq key to dump RCU's tree. */
 801 static bool sysrq_rcu;
 802 module_param(sysrq_rcu, bool, 0444);
 803
 804 /* Dump grace-period-request information due to commandeered sysrq. */
 805 static void sysrq_show_rcu(int key)
 806 {
 807         show_rcu_gp_kthreads();
 808 }
 809
 810 static const struct sysrq_key_op sysrq_rcudump_op = {
 811         .handler = sysrq_show_rcu,
 812         .help_msg = "show-rcu(y)",
 813         .action_msg = "Show RCU tree",
 814         .enable_mask = SYSRQ_ENABLE_DUMP,
 815 };
 816
 817 static int __init rcu_sysrq_init(void)
 818 {
 819         if (sysrq_rcu)
 820                 return register_sysrq_key('y', &sysrq_rcudump_op);
 821         return 0;
 822 }
 823 early_initcall(rcu_sysrq_init);