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 int sysctl_max_rcu_stall_to_panic __read_mostly;
  17
  18 #ifdef CONFIG_PROVE_RCU
  19 #define RCU_STALL_DELAY_DELTA           (5 * HZ)
  20 #else
  21 #define RCU_STALL_DELAY_DELTA           0
  22 #endif
  23 #define RCU_STALL_MIGHT_DIV             8
  24 #define RCU_STALL_MIGHT_MIN             (2 * HZ)
  25
  26 /* Limit-check stall timeouts specified at boottime and runtime. */
  27 int rcu_jiffies_till_stall_check(void)
  28 {
  29         int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
  30
  31         /*
  32          * Limit check must be consistent with the Kconfig limits
  33          * for CONFIG_RCU_CPU_STALL_TIMEOUT.
  34          */
  35         if (till_stall_check < 3) {
  36                 WRITE_ONCE(rcu_cpu_stall_timeout, 3);
  37                 till_stall_check = 3;
  38         } else if (till_stall_check > 300) {
  39                 WRITE_ONCE(rcu_cpu_stall_timeout, 300);
  40                 till_stall_check = 300;
  41         }
  42         return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
  43 }
  44 EXPORT_SYMBOL_GPL(rcu_jiffies_till_stall_check);
  45
  46 /**
  47  * rcu_gp_might_be_stalled - Is it likely that the grace period is stalled?
  48  *
  49  * Returns @true if the current grace period is sufficiently old that
  50  * it is reasonable to assume that it might be stalled.  This can be
  51  * useful when deciding whether to allocate memory to enable RCU-mediated
  52  * freeing on the one hand or just invoking synchronize_rcu() on the other.
  53  * The latter is preferable when the grace period is stalled.
  54  *
  55  * Note that sampling of the .gp_start and .gp_seq fields must be done
  56  * carefully to avoid false positives at the beginnings and ends of
  57  * grace periods.
  58  */
  59 bool rcu_gp_might_be_stalled(void)
  60 {
  61         unsigned long d = rcu_jiffies_till_stall_check() / RCU_STALL_MIGHT_DIV;
  62         unsigned long j = jiffies;
  63
  64         if (d < RCU_STALL_MIGHT_MIN)
  65                 d = RCU_STALL_MIGHT_MIN;
  66         smp_mb(); // jiffies before .gp_seq to avoid false positives.
  67         if (!rcu_gp_in_progress())
  68                 return false;
  69         // Long delays at this point avoids false positive, but a delay
  70         // of ULONG_MAX/4 jiffies voids your no-false-positive warranty.
  71         smp_mb(); // .gp_seq before second .gp_start
  72         // And ditto here.
  73         return !time_before(j, READ_ONCE(rcu_state.gp_start) + d);
  74 }
  75
  76 /* Don't do RCU CPU stall warnings during long sysrq printouts. */
  77 void rcu_sysrq_start(void)
  78 {
  79         if (!rcu_cpu_stall_suppress)
  80                 rcu_cpu_stall_suppress = 2;
  81 }
  82
  83 void rcu_sysrq_end(void)
  84 {
  85         if (rcu_cpu_stall_suppress == 2)
  86                 rcu_cpu_stall_suppress = 0;
  87 }
  88
  89 /* Don't print RCU CPU stall warnings during a kernel panic. */
  90 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
  91 {
  92         rcu_cpu_stall_suppress = 1;
  93         return NOTIFY_DONE;
  94 }
  95
  96 static struct notifier_block rcu_panic_block = {
  97         .notifier_call = rcu_panic,
  98 };
  99
 100 static int __init check_cpu_stall_init(void)
 101 {
 102         atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
 103         return 0;
 104 }
 105 early_initcall(check_cpu_stall_init);
 106
 107 /* If so specified via sysctl, panic, yielding cleaner stall-warning output. */
 108 static void panic_on_rcu_stall(void)
 109 {
 110         static int cpu_stall;
 111
 112         if (++cpu_stall < sysctl_max_rcu_stall_to_panic)
 113                 return;
 114
 115         if (sysctl_panic_on_rcu_stall)
 116                 panic("RCU Stall\n");
 117 }
 118
 119 /**
 120  * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
 121  *
 122  * Set the stall-warning timeout way off into the future, thus preventing
 123  * any RCU CPU stall-warning messages from appearing in the current set of
 124  * RCU grace periods.
 125  *
 126  * The caller must disable hard irqs.
 127  */
 128 void rcu_cpu_stall_reset(void)
 129 {
 130         WRITE_ONCE(rcu_state.jiffies_stall, jiffies + ULONG_MAX / 2);
 131 }
 132
 133 //////////////////////////////////////////////////////////////////////////////
 134 //
 135 // Interaction with RCU grace periods
 136
 137 /* Start of new grace period, so record stall time (and forcing times). */
 138 static void record_gp_stall_check_time(void)
 139 {
 140         unsigned long j = jiffies;
 141         unsigned long j1;
 142
 143         WRITE_ONCE(rcu_state.gp_start, j);
 144         j1 = rcu_jiffies_till_stall_check();
 145         smp_mb(); // ->gp_start before ->jiffies_stall and caller's ->gp_seq.
 146         WRITE_ONCE(rcu_state.jiffies_stall, j + j1);
 147         rcu_state.jiffies_resched = j + j1 / 2;
 148         rcu_state.n_force_qs_gpstart = READ_ONCE(rcu_state.n_force_qs);
 149 }
 150
 151 /* Zero ->ticks_this_gp and snapshot the number of RCU softirq handlers. */
 152 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
 153 {
 154         rdp->ticks_this_gp = 0;
 155         rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
 156         WRITE_ONCE(rdp->last_fqs_resched, jiffies);
 157 }
 158
 159 /*
 160  * If too much time has passed in the current grace period, and if
 161  * so configured, go kick the relevant kthreads.
 162  */
 163 static void rcu_stall_kick_kthreads(void)
 164 {
 165         unsigned long j;
 166
 167         if (!READ_ONCE(rcu_kick_kthreads))
 168                 return;
 169         j = READ_ONCE(rcu_state.jiffies_kick_kthreads);
 170         if (time_after(jiffies, j) && rcu_state.gp_kthread &&
 171             (rcu_gp_in_progress() || READ_ONCE(rcu_state.gp_flags))) {
 172                 WARN_ONCE(1, "Kicking %s grace-period kthread\n",
 173                           rcu_state.name);
 174                 rcu_ftrace_dump(DUMP_ALL);
 175                 wake_up_process(rcu_state.gp_kthread);
 176                 WRITE_ONCE(rcu_state.jiffies_kick_kthreads, j + HZ);
 177         }
 178 }
 179
 180 /*
 181  * Handler for the irq_work request posted about halfway into the RCU CPU
 182  * stall timeout, and used to detect excessive irq disabling.  Set state
 183  * appropriately, but just complain if there is unexpected state on entry.
 184  */
 185 static void rcu_iw_handler(struct irq_work *iwp)
 186 {
 187         struct rcu_data *rdp;
 188         struct rcu_node *rnp;
 189
 190         rdp = container_of(iwp, struct rcu_data, rcu_iw);
 191         rnp = rdp->mynode;
 192         raw_spin_lock_rcu_node(rnp);
 193         if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
 194                 rdp->rcu_iw_gp_seq = rnp->gp_seq;
 195                 rdp->rcu_iw_pending = false;
 196         }
 197         raw_spin_unlock_rcu_node(rnp);
 198 }
 199
 200 //////////////////////////////////////////////////////////////////////////////
 201 //
 202 // Printing RCU CPU stall warnings
 203
 204 #ifdef CONFIG_PREEMPT_RCU
 205
 206 /*
 207  * Dump detailed information for all tasks blocking the current RCU
 208  * grace period on the specified rcu_node structure.
 209  */
 210 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
 211 {
 212         unsigned long flags;
 213         struct task_struct *t;
 214
 215         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 216         if (!rcu_preempt_blocked_readers_cgp(rnp)) {
 217                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 218                 return;
 219         }
 220         t = list_entry(rnp->gp_tasks->prev,
 221                        struct task_struct, rcu_node_entry);
 222         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
 223                 /*
 224                  * We could be printing a lot while holding a spinlock.
 225                  * Avoid triggering hard lockup.
 226                  */
 227                 touch_nmi_watchdog();
 228                 sched_show_task(t);
 229         }
 230         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 231 }
 232
 233 // Communicate task state back to the RCU CPU stall warning request.
 234 struct rcu_stall_chk_rdr {
 235         int nesting;
 236         union rcu_special rs;
 237         bool on_blkd_list;
 238 };
 239
 240 /*
 241  * Report out the state of a not-running task that is stalling the
 242  * current RCU grace period.
 243  */
 244 static bool check_slow_task(struct task_struct *t, void *arg)
 245 {
 246         struct rcu_stall_chk_rdr *rscrp = arg;
 247
 248         if (task_curr(t))
 249                 return false; // It is running, so decline to inspect it.
 250         rscrp->nesting = t->rcu_read_lock_nesting;
 251         rscrp->rs = t->rcu_read_unlock_special;
 252         rscrp->on_blkd_list = !list_empty(&t->rcu_node_entry);
 253         return true;
 254 }
 255
 256 /*
 257  * Scan the current list of tasks blocked within RCU read-side critical
 258  * sections, printing out the tid of each of the first few of them.
 259  */
 260 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
 261         __releases(rnp->lock)
 262 {
 263         int i = 0;
 264         int ndetected = 0;
 265         struct rcu_stall_chk_rdr rscr;
 266         struct task_struct *t;
 267         struct task_struct *ts[8];
 268
 269         if (!rcu_preempt_blocked_readers_cgp(rnp))
 270                 return 0;
 271         pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
 272                rnp->level, rnp->grplo, rnp->grphi);
 273         t = list_entry(rnp->gp_tasks->prev,
 274                        struct task_struct, rcu_node_entry);
 275         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
 276                 get_task_struct(t);
 277                 ts[i++] = t;
 278                 if (i >= ARRAY_SIZE(ts))
 279                         break;
 280         }
 281         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 282         for (i--; i; i--) {
 283                 t = ts[i];
 284                 if (!try_invoke_on_locked_down_task(t, check_slow_task, &rscr))
 285                         pr_cont(" P%d", t->pid);
 286                 else
 287                         pr_cont(" P%d/%d:%c%c%c%c",
 288                                 t->pid, rscr.nesting,
 289                                 ".b"[rscr.rs.b.blocked],
 290                                 ".q"[rscr.rs.b.need_qs],
 291                                 ".e"[rscr.rs.b.exp_hint],
 292                                 ".l"[rscr.on_blkd_list]);
 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         /* Kick and suppress, if so configured. */
 482         rcu_stall_kick_kthreads();
 483         if (rcu_stall_is_suppressed())
 484                 return;
 485
 486         /*
 487          * OK, time to rat on our buddy...
 488          * See Documentation/RCU/stallwarn.rst for info on how to debug
 489          * RCU CPU stall warnings.
 490          */
 491         pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
 492         rcu_for_each_leaf_node(rnp) {
 493                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 494                 if (rnp->qsmask != 0) {
 495                         for_each_leaf_node_possible_cpu(rnp, cpu)
 496                                 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
 497                                         print_cpu_stall_info(cpu);
 498                                         ndetected++;
 499                                 }
 500                 }
 501                 ndetected += rcu_print_task_stall(rnp, flags); // Releases rnp->lock.
 502         }
 503
 504         for_each_possible_cpu(cpu)
 505                 totqlen += rcu_get_n_cbs_cpu(cpu);
 506         pr_cont("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu)\n",
 507                smp_processor_id(), (long)(jiffies - gps),
 508                (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
 509         if (ndetected) {
 510                 rcu_dump_cpu_stacks();
 511
 512                 /* Complain about tasks blocking the grace period. */
 513                 rcu_for_each_leaf_node(rnp)
 514                         rcu_print_detail_task_stall_rnp(rnp);
 515         } else {
 516                 if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
 517                         pr_err("INFO: Stall ended before state dump start\n");
 518                 } else {
 519                         j = jiffies;
 520                         gpa = data_race(rcu_state.gp_activity);
 521                         pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
 522                                rcu_state.name, j - gpa, j, gpa,
 523                                data_race(jiffies_till_next_fqs),
 524                                rcu_get_root()->qsmask);
 525                 }
 526         }
 527         /* Rewrite if needed in case of slow consoles. */
 528         if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
 529                 WRITE_ONCE(rcu_state.jiffies_stall,
 530                            jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
 531
 532         rcu_check_gp_kthread_starvation();
 533
 534         panic_on_rcu_stall();
 535
 536         rcu_force_quiescent_state();  /* Kick them all. */
 537 }
 538
 539 static void print_cpu_stall(unsigned long gps)
 540 {
 541         int cpu;
 542         unsigned long flags;
 543         struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 544         struct rcu_node *rnp = rcu_get_root();
 545         long totqlen = 0;
 546
 547         /* Kick and suppress, if so configured. */
 548         rcu_stall_kick_kthreads();
 549         if (rcu_stall_is_suppressed())
 550                 return;
 551
 552         /*
 553          * OK, time to rat on ourselves...
 554          * See Documentation/RCU/stallwarn.rst for info on how to debug
 555          * RCU CPU stall warnings.
 556          */
 557         pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
 558         raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
 559         print_cpu_stall_info(smp_processor_id());
 560         raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
 561         for_each_possible_cpu(cpu)
 562                 totqlen += rcu_get_n_cbs_cpu(cpu);
 563         pr_cont("\t(t=%lu jiffies g=%ld q=%lu)\n",
 564                 jiffies - gps,
 565                 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
 566
 567         rcu_check_gp_kthread_starvation();
 568
 569         rcu_dump_cpu_stacks();
 570
 571         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 572         /* Rewrite if needed in case of slow consoles. */
 573         if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
 574                 WRITE_ONCE(rcu_state.jiffies_stall,
 575                            jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
 576         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 577
 578         panic_on_rcu_stall();
 579
 580         /*
 581          * Attempt to revive the RCU machinery by forcing a context switch.
 582          *
 583          * A context switch would normally allow the RCU state machine to make
 584          * progress and it could be we're stuck in kernel space without context
 585          * switches for an entirely unreasonable amount of time.
 586          */
 587         set_tsk_need_resched(current);
 588         set_preempt_need_resched();
 589 }
 590
 591 static void check_cpu_stall(struct rcu_data *rdp)
 592 {
 593         unsigned long gs1;
 594         unsigned long gs2;
 595         unsigned long gps;
 596         unsigned long j;
 597         unsigned long jn;
 598         unsigned long js;
 599         struct rcu_node *rnp;
 600
 601         if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
 602             !rcu_gp_in_progress())
 603                 return;
 604         rcu_stall_kick_kthreads();
 605         j = jiffies;
 606
 607         /*
 608          * Lots of memory barriers to reject false positives.
 609          *
 610          * The idea is to pick up rcu_state.gp_seq, then
 611          * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
 612          * another copy of rcu_state.gp_seq.  These values are updated in
 613          * the opposite order with memory barriers (or equivalent) during
 614          * grace-period initialization and cleanup.  Now, a false positive
 615          * can occur if we get an new value of rcu_state.gp_start and a old
 616          * value of rcu_state.jiffies_stall.  But given the memory barriers,
 617          * the only way that this can happen is if one grace period ends
 618          * and another starts between these two fetches.  This is detected
 619          * by comparing the second fetch of rcu_state.gp_seq with the
 620          * previous fetch from rcu_state.gp_seq.
 621          *
 622          * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
 623          * and rcu_state.gp_start suffice to forestall false positives.
 624          */
 625         gs1 = READ_ONCE(rcu_state.gp_seq);
 626         smp_rmb(); /* Pick up ->gp_seq first... */
 627         js = READ_ONCE(rcu_state.jiffies_stall);
 628         smp_rmb(); /* ...then ->jiffies_stall before the rest... */
 629         gps = READ_ONCE(rcu_state.gp_start);
 630         smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
 631         gs2 = READ_ONCE(rcu_state.gp_seq);
 632         if (gs1 != gs2 ||
 633             ULONG_CMP_LT(j, js) ||
 634             ULONG_CMP_GE(gps, js))
 635                 return; /* No stall or GP completed since entering function. */
 636         rnp = rdp->mynode;
 637         jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
 638         if (rcu_gp_in_progress() &&
 639             (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
 640             cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
 641
 642                 /* We haven't checked in, so go dump stack. */
 643                 print_cpu_stall(gps);
 644                 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
 645                         rcu_ftrace_dump(DUMP_ALL);
 646
 647         } else if (rcu_gp_in_progress() &&
 648                    ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
 649                    cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
 650
 651                 /* They had a few time units to dump stack, so complain. */
 652                 print_other_cpu_stall(gs2, gps);
 653                 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
 654                         rcu_ftrace_dump(DUMP_ALL);
 655         }
 656 }
 657
 658 //////////////////////////////////////////////////////////////////////////////
 659 //
 660 // RCU forward-progress mechanisms, including of callback invocation.
 661
 662
 663 /*
 664  * Show the state of the grace-period kthreads.
 665  */
 666 void show_rcu_gp_kthreads(void)
 667 {
 668         unsigned long cbs = 0;
 669         int cpu;
 670         unsigned long j;
 671         unsigned long ja;
 672         unsigned long jr;
 673         unsigned long jw;
 674         struct rcu_data *rdp;
 675         struct rcu_node *rnp;
 676         struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
 677
 678         j = jiffies;
 679         ja = j - data_race(rcu_state.gp_activity);
 680         jr = j - data_race(rcu_state.gp_req_activity);
 681         jw = j - data_race(rcu_state.gp_wake_time);
 682         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",
 683                 rcu_state.name, gp_state_getname(rcu_state.gp_state),
 684                 rcu_state.gp_state, t ? t->state : 0x1ffffL,
 685                 ja, jr, jw, (long)data_race(rcu_state.gp_wake_seq),
 686                 (long)data_race(rcu_state.gp_seq),
 687                 (long)data_race(rcu_get_root()->gp_seq_needed),
 688                 data_race(rcu_state.gp_flags));
 689         rcu_for_each_node_breadth_first(rnp) {
 690                 if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
 691                                  READ_ONCE(rnp->gp_seq_needed)))
 692                         continue;
 693                 pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld\n",
 694                         rnp->grplo, rnp->grphi, (long)data_race(rnp->gp_seq),
 695                         (long)data_race(rnp->gp_seq_needed));
 696                 if (!rcu_is_leaf_node(rnp))
 697                         continue;
 698                 for_each_leaf_node_possible_cpu(rnp, cpu) {
 699                         rdp = per_cpu_ptr(&rcu_data, cpu);
 700                         if (READ_ONCE(rdp->gpwrap) ||
 701                             ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
 702                                          READ_ONCE(rdp->gp_seq_needed)))
 703                                 continue;
 704                         pr_info("\tcpu %d ->gp_seq_needed %ld\n",
 705                                 cpu, (long)data_race(rdp->gp_seq_needed));
 706                 }
 707         }
 708         for_each_possible_cpu(cpu) {
 709                 rdp = per_cpu_ptr(&rcu_data, cpu);
 710                 cbs += data_race(rdp->n_cbs_invoked);
 711                 if (rcu_segcblist_is_offloaded(&rdp->cblist))
 712                         show_rcu_nocb_state(rdp);
 713         }
 714         pr_info("RCU callbacks invoked since boot: %lu\n", cbs);
 715         show_rcu_tasks_gp_kthreads();
 716 }
 717 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
 718
 719 /*
 720  * This function checks for grace-period requests that fail to motivate
 721  * RCU to come out of its idle mode.
 722  */
 723 static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
 724                                      const unsigned long gpssdelay)
 725 {
 726         unsigned long flags;
 727         unsigned long j;
 728         struct rcu_node *rnp_root = rcu_get_root();
 729         static atomic_t warned = ATOMIC_INIT(0);
 730
 731         if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
 732             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 733                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 734             !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread.
 735                 return;
 736         j = jiffies; /* Expensive access, and in common case don't get here. */
 737         if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 738             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 739             atomic_read(&warned))
 740                 return;
 741
 742         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 743         j = jiffies;
 744         if (rcu_gp_in_progress() ||
 745             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 746                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 747             time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 748             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 749             atomic_read(&warned)) {
 750                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 751                 return;
 752         }
 753         /* Hold onto the leaf lock to make others see warned==1. */
 754
 755         if (rnp_root != rnp)
 756                 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
 757         j = jiffies;
 758         if (rcu_gp_in_progress() ||
 759             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 760                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 761             time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 762             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 763             atomic_xchg(&warned, 1)) {
 764                 if (rnp_root != rnp)
 765                         /* irqs remain disabled. */
 766                         raw_spin_unlock_rcu_node(rnp_root);
 767                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 768                 return;
 769         }
 770         WARN_ON(1);
 771         if (rnp_root != rnp)
 772                 raw_spin_unlock_rcu_node(rnp_root);
 773         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 774         show_rcu_gp_kthreads();
 775 }
 776
 777 /*
 778  * Do a forward-progress check for rcutorture.  This is normally invoked
 779  * due to an OOM event.  The argument "j" gives the time period during
 780  * which rcutorture would like progress to have been made.
 781  */
 782 void rcu_fwd_progress_check(unsigned long j)
 783 {
 784         unsigned long cbs;
 785         int cpu;
 786         unsigned long max_cbs = 0;
 787         int max_cpu = -1;
 788         struct rcu_data *rdp;
 789
 790         if (rcu_gp_in_progress()) {
 791                 pr_info("%s: GP age %lu jiffies\n",
 792                         __func__, jiffies - rcu_state.gp_start);
 793                 show_rcu_gp_kthreads();
 794         } else {
 795                 pr_info("%s: Last GP end %lu jiffies ago\n",
 796                         __func__, jiffies - rcu_state.gp_end);
 797                 preempt_disable();
 798                 rdp = this_cpu_ptr(&rcu_data);
 799                 rcu_check_gp_start_stall(rdp->mynode, rdp, j);
 800                 preempt_enable();
 801         }
 802         for_each_possible_cpu(cpu) {
 803                 cbs = rcu_get_n_cbs_cpu(cpu);
 804                 if (!cbs)
 805                         continue;
 806                 if (max_cpu < 0)
 807                         pr_info("%s: callbacks", __func__);
 808                 pr_cont(" %d: %lu", cpu, cbs);
 809                 if (cbs <= max_cbs)
 810                         continue;
 811                 max_cbs = cbs;
 812                 max_cpu = cpu;
 813         }
 814         if (max_cpu >= 0)
 815                 pr_cont("\n");
 816 }
 817 EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
 818
 819 /* Commandeer a sysrq key to dump RCU's tree. */
 820 static bool sysrq_rcu;
 821 module_param(sysrq_rcu, bool, 0444);
 822
 823 /* Dump grace-period-request information due to commandeered sysrq. */
 824 static void sysrq_show_rcu(int key)
 825 {
 826         show_rcu_gp_kthreads();
 827 }
 828
 829 static const struct sysrq_key_op sysrq_rcudump_op = {
 830         .handler = sysrq_show_rcu,
 831         .help_msg = "show-rcu(y)",
 832         .action_msg = "Show RCU tree",
 833         .enable_mask = SYSRQ_ENABLE_DUMP,
 834 };
 835
 836 static int __init rcu_sysrq_init(void)
 837 {
 838         if (sysrq_rcu)
 839                 return register_sysrq_key('y', &sysrq_rcudump_op);
 840         return 0;
 841 }
 842 early_initcall(rcu_sysrq_init);