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 #include <linux/kvm_para.h>
  11
  12 //////////////////////////////////////////////////////////////////////////////
  13 //
  14 // Controlling CPU stall warnings, including delay calculation.
  15
  16 /* panic() on RCU Stall sysctl. */
  17 int sysctl_panic_on_rcu_stall __read_mostly;
  18
  19 #ifdef CONFIG_PROVE_RCU
  20 #define RCU_STALL_DELAY_DELTA           (5 * HZ)
  21 #else
  22 #define RCU_STALL_DELAY_DELTA           0
  23 #endif
  24 #define RCU_STALL_MIGHT_DIV             8
  25 #define RCU_STALL_MIGHT_MIN             (2 * HZ)
  26
  27 /* Limit-check stall timeouts specified at boottime and runtime. */
  28 int rcu_jiffies_till_stall_check(void)
  29 {
  30         int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
  31
  32         /*
  33          * Limit check must be consistent with the Kconfig limits
  34          * for CONFIG_RCU_CPU_STALL_TIMEOUT.
  35          */
  36         if (till_stall_check < 3) {
  37                 WRITE_ONCE(rcu_cpu_stall_timeout, 3);
  38                 till_stall_check = 3;
  39         } else if (till_stall_check > 300) {
  40                 WRITE_ONCE(rcu_cpu_stall_timeout, 300);
  41                 till_stall_check = 300;
  42         }
  43         return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
  44 }
  45 EXPORT_SYMBOL_GPL(rcu_jiffies_till_stall_check);
  46
  47 /**
  48  * rcu_gp_might_be_stalled - Is it likely that the grace period is stalled?
  49  *
  50  * Returns @true if the current grace period is sufficiently old that
  51  * it is reasonable to assume that it might be stalled.  This can be
  52  * useful when deciding whether to allocate memory to enable RCU-mediated
  53  * freeing on the one hand or just invoking synchronize_rcu() on the other.
  54  * The latter is preferable when the grace period is stalled.
  55  *
  56  * Note that sampling of the .gp_start and .gp_seq fields must be done
  57  * carefully to avoid false positives at the beginnings and ends of
  58  * grace periods.
  59  */
  60 bool rcu_gp_might_be_stalled(void)
  61 {
  62         unsigned long d = rcu_jiffies_till_stall_check() / RCU_STALL_MIGHT_DIV;
  63         unsigned long j = jiffies;
  64
  65         if (d < RCU_STALL_MIGHT_MIN)
  66                 d = RCU_STALL_MIGHT_MIN;
  67         smp_mb(); // jiffies before .gp_seq to avoid false positives.
  68         if (!rcu_gp_in_progress())
  69                 return false;
  70         // Long delays at this point avoids false positive, but a delay
  71         // of ULONG_MAX/4 jiffies voids your no-false-positive warranty.
  72         smp_mb(); // .gp_seq before second .gp_start
  73         // And ditto here.
  74         return !time_before(j, READ_ONCE(rcu_state.gp_start) + d);
  75 }
  76
  77 /* Don't do RCU CPU stall warnings during long sysrq printouts. */
  78 void rcu_sysrq_start(void)
  79 {
  80         if (!rcu_cpu_stall_suppress)
  81                 rcu_cpu_stall_suppress = 2;
  82 }
  83
  84 void rcu_sysrq_end(void)
  85 {
  86         if (rcu_cpu_stall_suppress == 2)
  87                 rcu_cpu_stall_suppress = 0;
  88 }
  89
  90 /* Don't print RCU CPU stall warnings during a kernel panic. */
  91 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
  92 {
  93         rcu_cpu_stall_suppress = 1;
  94         return NOTIFY_DONE;
  95 }
  96
  97 static struct notifier_block rcu_panic_block = {
  98         .notifier_call = rcu_panic,
  99 };
 100
 101 static int __init check_cpu_stall_init(void)
 102 {
 103         atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
 104         return 0;
 105 }
 106 early_initcall(check_cpu_stall_init);
 107
 108 /* If so specified via sysctl, panic, yielding cleaner stall-warning output. */
 109 static void panic_on_rcu_stall(void)
 110 {
 111         if (sysctl_panic_on_rcu_stall)
 112                 panic("RCU Stall\n");
 113 }
 114
 115 /**
 116  * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
 117  *
 118  * Set the stall-warning timeout way off into the future, thus preventing
 119  * any RCU CPU stall-warning messages from appearing in the current set of
 120  * RCU grace periods.
 121  *
 122  * The caller must disable hard irqs.
 123  */
 124 void rcu_cpu_stall_reset(void)
 125 {
 126         WRITE_ONCE(rcu_state.jiffies_stall, jiffies + ULONG_MAX / 2);
 127 }
 128
 129 //////////////////////////////////////////////////////////////////////////////
 130 //
 131 // Interaction with RCU grace periods
 132
 133 /* Start of new grace period, so record stall time (and forcing times). */
 134 static void record_gp_stall_check_time(void)
 135 {
 136         unsigned long j = jiffies;
 137         unsigned long j1;
 138
 139         WRITE_ONCE(rcu_state.gp_start, j);
 140         j1 = rcu_jiffies_till_stall_check();
 141         smp_mb(); // ->gp_start before ->jiffies_stall and caller's ->gp_seq.
 142         WRITE_ONCE(rcu_state.jiffies_stall, j + j1);
 143         rcu_state.jiffies_resched = j + j1 / 2;
 144         rcu_state.n_force_qs_gpstart = READ_ONCE(rcu_state.n_force_qs);
 145 }
 146
 147 /* Zero ->ticks_this_gp and snapshot the number of RCU softirq handlers. */
 148 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
 149 {
 150         rdp->ticks_this_gp = 0;
 151         rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
 152         WRITE_ONCE(rdp->last_fqs_resched, jiffies);
 153 }
 154
 155 /*
 156  * If too much time has passed in the current grace period, and if
 157  * so configured, go kick the relevant kthreads.
 158  */
 159 static void rcu_stall_kick_kthreads(void)
 160 {
 161         unsigned long j;
 162
 163         if (!READ_ONCE(rcu_kick_kthreads))
 164                 return;
 165         j = READ_ONCE(rcu_state.jiffies_kick_kthreads);
 166         if (time_after(jiffies, j) && rcu_state.gp_kthread &&
 167             (rcu_gp_in_progress() || READ_ONCE(rcu_state.gp_flags))) {
 168                 WARN_ONCE(1, "Kicking %s grace-period kthread\n",
 169                           rcu_state.name);
 170                 rcu_ftrace_dump(DUMP_ALL);
 171                 wake_up_process(rcu_state.gp_kthread);
 172                 WRITE_ONCE(rcu_state.jiffies_kick_kthreads, j + HZ);
 173         }
 174 }
 175
 176 /*
 177  * Handler for the irq_work request posted about halfway into the RCU CPU
 178  * stall timeout, and used to detect excessive irq disabling.  Set state
 179  * appropriately, but just complain if there is unexpected state on entry.
 180  */
 181 static void rcu_iw_handler(struct irq_work *iwp)
 182 {
 183         struct rcu_data *rdp;
 184         struct rcu_node *rnp;
 185
 186         rdp = container_of(iwp, struct rcu_data, rcu_iw);
 187         rnp = rdp->mynode;
 188         raw_spin_lock_rcu_node(rnp);
 189         if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
 190                 rdp->rcu_iw_gp_seq = rnp->gp_seq;
 191                 rdp->rcu_iw_pending = false;
 192         }
 193         raw_spin_unlock_rcu_node(rnp);
 194 }
 195
 196 //////////////////////////////////////////////////////////////////////////////
 197 //
 198 // Printing RCU CPU stall warnings
 199
 200 #ifdef CONFIG_PREEMPT_RCU
 201
 202 /*
 203  * Dump detailed information for all tasks blocking the current RCU
 204  * grace period on the specified rcu_node structure.
 205  */
 206 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
 207 {
 208         unsigned long flags;
 209         struct task_struct *t;
 210
 211         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 212         if (!rcu_preempt_blocked_readers_cgp(rnp)) {
 213                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 214                 return;
 215         }
 216         t = list_entry(rnp->gp_tasks->prev,
 217                        struct task_struct, rcu_node_entry);
 218         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
 219                 /*
 220                  * We could be printing a lot while holding a spinlock.
 221                  * Avoid triggering hard lockup.
 222                  */
 223                 touch_nmi_watchdog();
 224                 sched_show_task(t);
 225         }
 226         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 227 }
 228
 229 // Communicate task state back to the RCU CPU stall warning request.
 230 struct rcu_stall_chk_rdr {
 231         int nesting;
 232         union rcu_special rs;
 233         bool on_blkd_list;
 234 };
 235
 236 /*
 237  * Report out the state of a not-running task that is stalling the
 238  * current RCU grace period.
 239  */
 240 static bool check_slow_task(struct task_struct *t, void *arg)
 241 {
 242         struct rcu_stall_chk_rdr *rscrp = arg;
 243
 244         if (task_curr(t))
 245                 return false; // It is running, so decline to inspect it.
 246         rscrp->nesting = t->rcu_read_lock_nesting;
 247         rscrp->rs = t->rcu_read_unlock_special;
 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 of the first few of them.
 255  */
 256 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
 257         __releases(rnp->lock)
 258 {
 259         int i = 0;
 260         int ndetected = 0;
 261         struct rcu_stall_chk_rdr rscr;
 262         struct task_struct *t;
 263         struct task_struct *ts[8];
 264
 265         lockdep_assert_irqs_disabled();
 266         if (!rcu_preempt_blocked_readers_cgp(rnp)) {
 267                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 268                 return 0;
 269         }
 270         pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
 271                rnp->level, rnp->grplo, rnp->grphi);
 272         t = list_entry(rnp->gp_tasks->prev,
 273                        struct task_struct, rcu_node_entry);
 274         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
 275                 get_task_struct(t);
 276                 ts[i++] = t;
 277                 if (i >= ARRAY_SIZE(ts))
 278                         break;
 279         }
 280         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 281         while (i) {
 282                 t = ts[--i];
 283                 if (!try_invoke_on_locked_down_task(t, check_slow_task, &rscr))
 284                         pr_cont(" P%d", t->pid);
 285                 else
 286                         pr_cont(" P%d/%d:%c%c%c%c",
 287                                 t->pid, rscr.nesting,
 288                                 ".b"[rscr.rs.b.blocked],
 289                                 ".q"[rscr.rs.b.need_qs],
 290                                 ".e"[rscr.rs.b.exp_hint],
 291                                 ".l"[rscr.on_blkd_list]);
 292                 lockdep_assert_irqs_disabled();
 293                 put_task_struct(t);
 294                 ndetected++;
 295         }
 296         pr_cont("\n");
 297         return ndetected;
 298 }
 299
 300 #else /* #ifdef CONFIG_PREEMPT_RCU */
 301
 302 /*
 303  * Because preemptible RCU does not exist, we never have to check for
 304  * tasks blocked within RCU read-side critical sections.
 305  */
 306 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
 307 {
 308 }
 309
 310 /*
 311  * Because preemptible RCU does not exist, we never have to check for
 312  * tasks blocked within RCU read-side critical sections.
 313  */
 314 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
 315 {
 316         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 317         return 0;
 318 }
 319 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
 320
 321 /*
 322  * Dump stacks of all tasks running on stalled CPUs.  First try using
 323  * NMIs, but fall back to manual remote stack tracing on architectures
 324  * that don't support NMI-based stack dumps.  The NMI-triggered stack
 325  * traces are more accurate because they are printed by the target CPU.
 326  */
 327 static void rcu_dump_cpu_stacks(void)
 328 {
 329         int cpu;
 330         unsigned long flags;
 331         struct rcu_node *rnp;
 332
 333         rcu_for_each_leaf_node(rnp) {
 334                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 335                 for_each_leaf_node_possible_cpu(rnp, cpu)
 336                         if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
 337                                 if (!trigger_single_cpu_backtrace(cpu))
 338                                         dump_cpu_task(cpu);
 339                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 340         }
 341 }
 342
 343 #ifdef CONFIG_RCU_FAST_NO_HZ
 344
 345 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
 346 {
 347         struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
 348
 349         sprintf(cp, "last_accelerate: %04lx/%04lx dyntick_enabled: %d",
 350                 rdp->last_accelerate & 0xffff, jiffies & 0xffff,
 351                 !!rdp->tick_nohz_enabled_snap);
 352 }
 353
 354 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
 355
 356 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
 357 {
 358         *cp = '\0';
 359 }
 360
 361 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
 362
 363 static const char * const gp_state_names[] = {
 364         [RCU_GP_IDLE] = "RCU_GP_IDLE",
 365         [RCU_GP_WAIT_GPS] = "RCU_GP_WAIT_GPS",
 366         [RCU_GP_DONE_GPS] = "RCU_GP_DONE_GPS",
 367         [RCU_GP_ONOFF] = "RCU_GP_ONOFF",
 368         [RCU_GP_INIT] = "RCU_GP_INIT",
 369         [RCU_GP_WAIT_FQS] = "RCU_GP_WAIT_FQS",
 370         [RCU_GP_DOING_FQS] = "RCU_GP_DOING_FQS",
 371         [RCU_GP_CLEANUP] = "RCU_GP_CLEANUP",
 372         [RCU_GP_CLEANED] = "RCU_GP_CLEANED",
 373 };
 374
 375 /*
 376  * Convert a ->gp_state value to a character string.
 377  */
 378 static const char *gp_state_getname(short gs)
 379 {
 380         if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
 381                 return "???";
 382         return gp_state_names[gs];
 383 }
 384
 385 /* Is the RCU grace-period kthread being starved of CPU time? */
 386 static bool rcu_is_gp_kthread_starving(unsigned long *jp)
 387 {
 388         unsigned long j = jiffies - READ_ONCE(rcu_state.gp_activity);
 389
 390         if (jp)
 391                 *jp = j;
 392         return j > 2 * HZ;
 393 }
 394
 395 /*
 396  * Print out diagnostic information for the specified stalled CPU.
 397  *
 398  * If the specified CPU is aware of the current RCU grace period, then
 399  * print the number of scheduling clock interrupts the CPU has taken
 400  * during the time that it has been aware.  Otherwise, print the number
 401  * of RCU grace periods that this CPU is ignorant of, for example, "1"
 402  * if the CPU was aware of the previous grace period.
 403  *
 404  * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
 405  */
 406 static void print_cpu_stall_info(int cpu)
 407 {
 408         unsigned long delta;
 409         bool falsepositive;
 410         char fast_no_hz[72];
 411         struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 412         char *ticks_title;
 413         unsigned long ticks_value;
 414
 415         /*
 416          * We could be printing a lot while holding a spinlock.  Avoid
 417          * triggering hard lockup.
 418          */
 419         touch_nmi_watchdog();
 420
 421         ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
 422         if (ticks_value) {
 423                 ticks_title = "GPs behind";
 424         } else {
 425                 ticks_title = "ticks this GP";
 426                 ticks_value = rdp->ticks_this_gp;
 427         }
 428         print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
 429         delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
 430         falsepositive = rcu_is_gp_kthread_starving(NULL) &&
 431                         rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp));
 432         pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%03x/%ld/%#lx softirq=%u/%u fqs=%ld %s%s\n",
 433                cpu,
 434                "O."[!!cpu_online(cpu)],
 435                "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
 436                "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
 437                !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
 438                         rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
 439                                 "!."[!delta],
 440                ticks_value, ticks_title,
 441                rcu_dynticks_snap(rdp) & 0xfff,
 442                rdp->dynticks_nesting, rdp->dynticks_nmi_nesting,
 443                rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
 444                data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
 445                fast_no_hz,
 446                falsepositive ? " (false positive?)" : "");
 447 }
 448
 449 /* Complain about starvation of grace-period kthread.  */
 450 static void rcu_check_gp_kthread_starvation(void)
 451 {
 452         struct task_struct *gpk = rcu_state.gp_kthread;
 453         unsigned long j;
 454
 455         if (rcu_is_gp_kthread_starving(&j)) {
 456                 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#lx ->cpu=%d\n",
 457                        rcu_state.name, j,
 458                        (long)rcu_seq_current(&rcu_state.gp_seq),
 459                        data_race(rcu_state.gp_flags),
 460                        gp_state_getname(rcu_state.gp_state), rcu_state.gp_state,
 461                        gpk ? gpk->state : ~0, gpk ? task_cpu(gpk) : -1);
 462                 if (gpk) {
 463                         pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name);
 464                         pr_err("RCU grace-period kthread stack dump:\n");
 465                         sched_show_task(gpk);
 466                         wake_up_process(gpk);
 467                 }
 468         }
 469 }
 470
 471 static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
 472 {
 473         int cpu;
 474         unsigned long flags;
 475         unsigned long gpa;
 476         unsigned long j;
 477         int ndetected = 0;
 478         struct rcu_node *rnp;
 479         long totqlen = 0;
 480
 481         lockdep_assert_irqs_disabled();
 482
 483         /* Kick and suppress, if so configured. */
 484         rcu_stall_kick_kthreads();
 485         if (rcu_stall_is_suppressed())
 486                 return;
 487
 488         /*
 489          * OK, time to rat on our buddy...
 490          * See Documentation/RCU/stallwarn.rst for info on how to debug
 491          * RCU CPU stall warnings.
 492          */
 493         pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
 494         rcu_for_each_leaf_node(rnp) {
 495                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 496                 if (rnp->qsmask != 0) {
 497                         for_each_leaf_node_possible_cpu(rnp, cpu)
 498                                 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
 499                                         print_cpu_stall_info(cpu);
 500                                         ndetected++;
 501                                 }
 502                 }
 503                 ndetected += rcu_print_task_stall(rnp, flags); // Releases rnp->lock.
 504                 lockdep_assert_irqs_disabled();
 505         }
 506
 507         for_each_possible_cpu(cpu)
 508                 totqlen += rcu_get_n_cbs_cpu(cpu);
 509         pr_cont("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu)\n",
 510                smp_processor_id(), (long)(jiffies - gps),
 511                (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
 512         if (ndetected) {
 513                 rcu_dump_cpu_stacks();
 514
 515                 /* Complain about tasks blocking the grace period. */
 516                 rcu_for_each_leaf_node(rnp)
 517                         rcu_print_detail_task_stall_rnp(rnp);
 518         } else {
 519                 if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
 520                         pr_err("INFO: Stall ended before state dump start\n");
 521                 } else {
 522                         j = jiffies;
 523                         gpa = data_race(rcu_state.gp_activity);
 524                         pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
 525                                rcu_state.name, j - gpa, j, gpa,
 526                                data_race(jiffies_till_next_fqs),
 527                                rcu_get_root()->qsmask);
 528                 }
 529         }
 530         /* Rewrite if needed in case of slow consoles. */
 531         if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
 532                 WRITE_ONCE(rcu_state.jiffies_stall,
 533                            jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
 534
 535         rcu_check_gp_kthread_starvation();
 536
 537         panic_on_rcu_stall();
 538
 539         rcu_force_quiescent_state();  /* Kick them all. */
 540 }
 541
 542 static void print_cpu_stall(unsigned long gps)
 543 {
 544         int cpu;
 545         unsigned long flags;
 546         struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 547         struct rcu_node *rnp = rcu_get_root();
 548         long totqlen = 0;
 549
 550         lockdep_assert_irqs_disabled();
 551
 552         /* Kick and suppress, if so configured. */
 553         rcu_stall_kick_kthreads();
 554         if (rcu_stall_is_suppressed())
 555                 return;
 556
 557         /*
 558          * OK, time to rat on ourselves...
 559          * See Documentation/RCU/stallwarn.rst for info on how to debug
 560          * RCU CPU stall warnings.
 561          */
 562         pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
 563         raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
 564         print_cpu_stall_info(smp_processor_id());
 565         raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
 566         for_each_possible_cpu(cpu)
 567                 totqlen += rcu_get_n_cbs_cpu(cpu);
 568         pr_cont("\t(t=%lu jiffies g=%ld q=%lu)\n",
 569                 jiffies - gps,
 570                 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
 571
 572         rcu_check_gp_kthread_starvation();
 573
 574         rcu_dump_cpu_stacks();
 575
 576         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 577         /* Rewrite if needed in case of slow consoles. */
 578         if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
 579                 WRITE_ONCE(rcu_state.jiffies_stall,
 580                            jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
 581         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 582
 583         panic_on_rcu_stall();
 584
 585         /*
 586          * Attempt to revive the RCU machinery by forcing a context switch.
 587          *
 588          * A context switch would normally allow the RCU state machine to make
 589          * progress and it could be we're stuck in kernel space without context
 590          * switches for an entirely unreasonable amount of time.
 591          */
 592         set_tsk_need_resched(current);
 593         set_preempt_need_resched();
 594 }
 595
 596 static void check_cpu_stall(struct rcu_data *rdp)
 597 {
 598         unsigned long gs1;
 599         unsigned long gs2;
 600         unsigned long gps;
 601         unsigned long j;
 602         unsigned long jn;
 603         unsigned long js;
 604         struct rcu_node *rnp;
 605
 606         lockdep_assert_irqs_disabled();
 607         if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
 608             !rcu_gp_in_progress())
 609                 return;
 610         rcu_stall_kick_kthreads();
 611         j = jiffies;
 612
 613         /*
 614          * Lots of memory barriers to reject false positives.
 615          *
 616          * The idea is to pick up rcu_state.gp_seq, then
 617          * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
 618          * another copy of rcu_state.gp_seq.  These values are updated in
 619          * the opposite order with memory barriers (or equivalent) during
 620          * grace-period initialization and cleanup.  Now, a false positive
 621          * can occur if we get an new value of rcu_state.gp_start and a old
 622          * value of rcu_state.jiffies_stall.  But given the memory barriers,
 623          * the only way that this can happen is if one grace period ends
 624          * and another starts between these two fetches.  This is detected
 625          * by comparing the second fetch of rcu_state.gp_seq with the
 626          * previous fetch from rcu_state.gp_seq.
 627          *
 628          * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
 629          * and rcu_state.gp_start suffice to forestall false positives.
 630          */
 631         gs1 = READ_ONCE(rcu_state.gp_seq);
 632         smp_rmb(); /* Pick up ->gp_seq first... */
 633         js = READ_ONCE(rcu_state.jiffies_stall);
 634         smp_rmb(); /* ...then ->jiffies_stall before the rest... */
 635         gps = READ_ONCE(rcu_state.gp_start);
 636         smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
 637         gs2 = READ_ONCE(rcu_state.gp_seq);
 638         if (gs1 != gs2 ||
 639             ULONG_CMP_LT(j, js) ||
 640             ULONG_CMP_GE(gps, js))
 641                 return; /* No stall or GP completed since entering function. */
 642         rnp = rdp->mynode;
 643         jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
 644         if (rcu_gp_in_progress() &&
 645             (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
 646             cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
 647
 648                 /*
 649                  * If a virtual machine is stopped by the host it can look to
 650                  * the watchdog like an RCU stall. Check to see if the host
 651                  * stopped the vm.
 652                  */
 653                 if (kvm_check_and_clear_guest_paused())
 654                         return;
 655
 656                 /* We haven't checked in, so go dump stack. */
 657                 print_cpu_stall(gps);
 658                 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
 659                         rcu_ftrace_dump(DUMP_ALL);
 660
 661         } else if (rcu_gp_in_progress() &&
 662                    ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
 663                    cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
 664
 665                 /*
 666                  * If a virtual machine is stopped by the host it can look to
 667                  * the watchdog like an RCU stall. Check to see if the host
 668                  * stopped the vm.
 669                  */
 670                 if (kvm_check_and_clear_guest_paused())
 671                         return;
 672
 673                 /* They had a few time units to dump stack, so complain. */
 674                 print_other_cpu_stall(gs2, gps);
 675                 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
 676                         rcu_ftrace_dump(DUMP_ALL);
 677         }
 678 }
 679
 680 //////////////////////////////////////////////////////////////////////////////
 681 //
 682 // RCU forward-progress mechanisms, including of callback invocation.
 683
 684
 685 /*
 686  * Show the state of the grace-period kthreads.
 687  */
 688 void show_rcu_gp_kthreads(void)
 689 {
 690         unsigned long cbs = 0;
 691         int cpu;
 692         unsigned long j;
 693         unsigned long ja;
 694         unsigned long jr;
 695         unsigned long jw;
 696         struct rcu_data *rdp;
 697         struct rcu_node *rnp;
 698         struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
 699
 700         j = jiffies;
 701         ja = j - data_race(rcu_state.gp_activity);
 702         jr = j - data_race(rcu_state.gp_req_activity);
 703         jw = j - data_race(rcu_state.gp_wake_time);
 704         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",
 705                 rcu_state.name, gp_state_getname(rcu_state.gp_state),
 706                 rcu_state.gp_state, t ? t->state : 0x1ffffL,
 707                 ja, jr, jw, (long)data_race(rcu_state.gp_wake_seq),
 708                 (long)data_race(rcu_state.gp_seq),
 709                 (long)data_race(rcu_get_root()->gp_seq_needed),
 710                 data_race(rcu_state.gp_flags));
 711         rcu_for_each_node_breadth_first(rnp) {
 712                 if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
 713                                  READ_ONCE(rnp->gp_seq_needed)))
 714                         continue;
 715                 pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld\n",
 716                         rnp->grplo, rnp->grphi, (long)data_race(rnp->gp_seq),
 717                         (long)data_race(rnp->gp_seq_needed));
 718                 if (!rcu_is_leaf_node(rnp))
 719                         continue;
 720                 for_each_leaf_node_possible_cpu(rnp, cpu) {
 721                         rdp = per_cpu_ptr(&rcu_data, cpu);
 722                         if (READ_ONCE(rdp->gpwrap) ||
 723                             ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
 724                                          READ_ONCE(rdp->gp_seq_needed)))
 725                                 continue;
 726                         pr_info("\tcpu %d ->gp_seq_needed %ld\n",
 727                                 cpu, (long)data_race(rdp->gp_seq_needed));
 728                 }
 729         }
 730         for_each_possible_cpu(cpu) {
 731                 rdp = per_cpu_ptr(&rcu_data, cpu);
 732                 cbs += data_race(rdp->n_cbs_invoked);
 733                 if (rcu_segcblist_is_offloaded(&rdp->cblist))
 734                         show_rcu_nocb_state(rdp);
 735         }
 736         pr_info("RCU callbacks invoked since boot: %lu\n", cbs);
 737         show_rcu_tasks_gp_kthreads();
 738 }
 739 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
 740
 741 /*
 742  * This function checks for grace-period requests that fail to motivate
 743  * RCU to come out of its idle mode.
 744  */
 745 static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
 746                                      const unsigned long gpssdelay)
 747 {
 748         unsigned long flags;
 749         unsigned long j;
 750         struct rcu_node *rnp_root = rcu_get_root();
 751         static atomic_t warned = ATOMIC_INIT(0);
 752
 753         if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
 754             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 755                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 756             !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread.
 757                 return;
 758         j = jiffies; /* Expensive access, and in common case don't get here. */
 759         if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 760             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 761             atomic_read(&warned))
 762                 return;
 763
 764         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 765         j = jiffies;
 766         if (rcu_gp_in_progress() ||
 767             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 768                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 769             time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 770             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 771             atomic_read(&warned)) {
 772                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 773                 return;
 774         }
 775         /* Hold onto the leaf lock to make others see warned==1. */
 776
 777         if (rnp_root != rnp)
 778                 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
 779         j = jiffies;
 780         if (rcu_gp_in_progress() ||
 781             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 782                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 783             time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 784             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 785             atomic_xchg(&warned, 1)) {
 786                 if (rnp_root != rnp)
 787                         /* irqs remain disabled. */
 788                         raw_spin_unlock_rcu_node(rnp_root);
 789                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 790                 return;
 791         }
 792         WARN_ON(1);
 793         if (rnp_root != rnp)
 794                 raw_spin_unlock_rcu_node(rnp_root);
 795         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 796         show_rcu_gp_kthreads();
 797 }
 798
 799 /*
 800  * Do a forward-progress check for rcutorture.  This is normally invoked
 801  * due to an OOM event.  The argument "j" gives the time period during
 802  * which rcutorture would like progress to have been made.
 803  */
 804 void rcu_fwd_progress_check(unsigned long j)
 805 {
 806         unsigned long cbs;
 807         int cpu;
 808         unsigned long max_cbs = 0;
 809         int max_cpu = -1;
 810         struct rcu_data *rdp;
 811
 812         if (rcu_gp_in_progress()) {
 813                 pr_info("%s: GP age %lu jiffies\n",
 814                         __func__, jiffies - rcu_state.gp_start);
 815                 show_rcu_gp_kthreads();
 816         } else {
 817                 pr_info("%s: Last GP end %lu jiffies ago\n",
 818                         __func__, jiffies - rcu_state.gp_end);
 819                 preempt_disable();
 820                 rdp = this_cpu_ptr(&rcu_data);
 821                 rcu_check_gp_start_stall(rdp->mynode, rdp, j);
 822                 preempt_enable();
 823         }
 824         for_each_possible_cpu(cpu) {
 825                 cbs = rcu_get_n_cbs_cpu(cpu);
 826                 if (!cbs)
 827                         continue;
 828                 if (max_cpu < 0)
 829                         pr_info("%s: callbacks", __func__);
 830                 pr_cont(" %d: %lu", cpu, cbs);
 831                 if (cbs <= max_cbs)
 832                         continue;
 833                 max_cbs = cbs;
 834                 max_cpu = cpu;
 835         }
 836         if (max_cpu >= 0)
 837                 pr_cont("\n");
 838 }
 839 EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
 840
 841 /* Commandeer a sysrq key to dump RCU's tree. */
 842 static bool sysrq_rcu;
 843 module_param(sysrq_rcu, bool, 0444);
 844
 845 /* Dump grace-period-request information due to commandeered sysrq. */
 846 static void sysrq_show_rcu(int key)
 847 {
 848         show_rcu_gp_kthreads();
 849 }
 850
 851 static const struct sysrq_key_op sysrq_rcudump_op = {
 852         .handler = sysrq_show_rcu,
 853         .help_msg = "show-rcu(y)",
 854         .action_msg = "Show RCU tree",
 855         .enable_mask = SYSRQ_ENABLE_DUMP,
 856 };
 857
 858 static int __init rcu_sysrq_init(void)
 859 {
 860         if (sysrq_rcu)
 861                 return register_sysrq_key('y', &sysrq_rcudump_op);
 862         return 0;
 863 }
 864 early_initcall(rcu_sysrq_init);