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 int sysctl_max_rcu_stall_to_panic __read_mostly;
  19
  20 #ifdef CONFIG_PROVE_RCU
  21 #define RCU_STALL_DELAY_DELTA           (5 * HZ)
  22 #else
  23 #define RCU_STALL_DELAY_DELTA           0
  24 #endif
  25 #define RCU_STALL_MIGHT_DIV             8
  26 #define RCU_STALL_MIGHT_MIN             (2 * HZ)
  27
  28 /* Limit-check stall timeouts specified at boottime and runtime. */
  29 int rcu_jiffies_till_stall_check(void)
  30 {
  31         int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
  32
  33         /*
  34          * Limit check must be consistent with the Kconfig limits
  35          * for CONFIG_RCU_CPU_STALL_TIMEOUT.
  36          */
  37         if (till_stall_check < 3) {
  38                 WRITE_ONCE(rcu_cpu_stall_timeout, 3);
  39                 till_stall_check = 3;
  40         } else if (till_stall_check > 300) {
  41                 WRITE_ONCE(rcu_cpu_stall_timeout, 300);
  42                 till_stall_check = 300;
  43         }
  44         return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
  45 }
  46 EXPORT_SYMBOL_GPL(rcu_jiffies_till_stall_check);
  47
  48 /**
  49  * rcu_gp_might_be_stalled - Is it likely that the grace period is stalled?
  50  *
  51  * Returns @true if the current grace period is sufficiently old that
  52  * it is reasonable to assume that it might be stalled.  This can be
  53  * useful when deciding whether to allocate memory to enable RCU-mediated
  54  * freeing on the one hand or just invoking synchronize_rcu() on the other.
  55  * The latter is preferable when the grace period is stalled.
  56  *
  57  * Note that sampling of the .gp_start and .gp_seq fields must be done
  58  * carefully to avoid false positives at the beginnings and ends of
  59  * grace periods.
  60  */
  61 bool rcu_gp_might_be_stalled(void)
  62 {
  63         unsigned long d = rcu_jiffies_till_stall_check() / RCU_STALL_MIGHT_DIV;
  64         unsigned long j = jiffies;
  65
  66         if (d < RCU_STALL_MIGHT_MIN)
  67                 d = RCU_STALL_MIGHT_MIN;
  68         smp_mb(); // jiffies before .gp_seq to avoid false positives.
  69         if (!rcu_gp_in_progress())
  70                 return false;
  71         // Long delays at this point avoids false positive, but a delay
  72         // of ULONG_MAX/4 jiffies voids your no-false-positive warranty.
  73         smp_mb(); // .gp_seq before second .gp_start
  74         // And ditto here.
  75         return !time_before(j, READ_ONCE(rcu_state.gp_start) + d);
  76 }
  77
  78 /* Don't do RCU CPU stall warnings during long sysrq printouts. */
  79 void rcu_sysrq_start(void)
  80 {
  81         if (!rcu_cpu_stall_suppress)
  82                 rcu_cpu_stall_suppress = 2;
  83 }
  84
  85 void rcu_sysrq_end(void)
  86 {
  87         if (rcu_cpu_stall_suppress == 2)
  88                 rcu_cpu_stall_suppress = 0;
  89 }
  90
  91 /* Don't print RCU CPU stall warnings during a kernel panic. */
  92 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
  93 {
  94         rcu_cpu_stall_suppress = 1;
  95         return NOTIFY_DONE;
  96 }
  97
  98 static struct notifier_block rcu_panic_block = {
  99         .notifier_call = rcu_panic,
 100 };
 101
 102 static int __init check_cpu_stall_init(void)
 103 {
 104         atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
 105         return 0;
 106 }
 107 early_initcall(check_cpu_stall_init);
 108
 109 /* If so specified via sysctl, panic, yielding cleaner stall-warning output. */
 110 static void panic_on_rcu_stall(void)
 111 {
 112         static int cpu_stall;
 113
 114         if (++cpu_stall < sysctl_max_rcu_stall_to_panic)
 115                 return;
 116
 117         if (sysctl_panic_on_rcu_stall)
 118                 panic("RCU Stall\n");
 119 }
 120
 121 /**
 122  * rcu_cpu_stall_reset - restart stall-warning timeout for current grace period
 123  *
 124  * The caller must disable hard irqs.
 125  */
 126 void rcu_cpu_stall_reset(void)
 127 {
 128         WRITE_ONCE(rcu_state.jiffies_stall,
 129                    jiffies + rcu_jiffies_till_stall_check());
 130 }
 131
 132 //////////////////////////////////////////////////////////////////////////////
 133 //
 134 // Interaction with RCU grace periods
 135
 136 /* Start of new grace period, so record stall time (and forcing times). */
 137 static void record_gp_stall_check_time(void)
 138 {
 139         unsigned long j = jiffies;
 140         unsigned long j1;
 141
 142         WRITE_ONCE(rcu_state.gp_start, j);
 143         j1 = rcu_jiffies_till_stall_check();
 144         smp_mb(); // ->gp_start before ->jiffies_stall and caller's ->gp_seq.
 145         WRITE_ONCE(rcu_state.jiffies_stall, j + j1);
 146         rcu_state.jiffies_resched = j + j1 / 2;
 147         rcu_state.n_force_qs_gpstart = READ_ONCE(rcu_state.n_force_qs);
 148 }
 149
 150 /* Zero ->ticks_this_gp and snapshot the number of RCU softirq handlers. */
 151 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
 152 {
 153         rdp->ticks_this_gp = 0;
 154         rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
 155         WRITE_ONCE(rdp->last_fqs_resched, jiffies);
 156 }
 157
 158 /*
 159  * If too much time has passed in the current grace period, and if
 160  * so configured, go kick the relevant kthreads.
 161  */
 162 static void rcu_stall_kick_kthreads(void)
 163 {
 164         unsigned long j;
 165
 166         if (!READ_ONCE(rcu_kick_kthreads))
 167                 return;
 168         j = READ_ONCE(rcu_state.jiffies_kick_kthreads);
 169         if (time_after(jiffies, j) && rcu_state.gp_kthread &&
 170             (rcu_gp_in_progress() || READ_ONCE(rcu_state.gp_flags))) {
 171                 WARN_ONCE(1, "Kicking %s grace-period kthread\n",
 172                           rcu_state.name);
 173                 rcu_ftrace_dump(DUMP_ALL);
 174                 wake_up_process(rcu_state.gp_kthread);
 175                 WRITE_ONCE(rcu_state.jiffies_kick_kthreads, j + HZ);
 176         }
 177 }
 178
 179 /*
 180  * Handler for the irq_work request posted about halfway into the RCU CPU
 181  * stall timeout, and used to detect excessive irq disabling.  Set state
 182  * appropriately, but just complain if there is unexpected state on entry.
 183  */
 184 static void rcu_iw_handler(struct irq_work *iwp)
 185 {
 186         struct rcu_data *rdp;
 187         struct rcu_node *rnp;
 188
 189         rdp = container_of(iwp, struct rcu_data, rcu_iw);
 190         rnp = rdp->mynode;
 191         raw_spin_lock_rcu_node(rnp);
 192         if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
 193                 rdp->rcu_iw_gp_seq = rnp->gp_seq;
 194                 rdp->rcu_iw_pending = false;
 195         }
 196         raw_spin_unlock_rcu_node(rnp);
 197 }
 198
 199 //////////////////////////////////////////////////////////////////////////////
 200 //
 201 // Printing RCU CPU stall warnings
 202
 203 #ifdef CONFIG_PREEMPT_RCU
 204
 205 /*
 206  * Dump detailed information for all tasks blocking the current RCU
 207  * grace period on the specified rcu_node structure.
 208  */
 209 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
 210 {
 211         unsigned long flags;
 212         struct task_struct *t;
 213
 214         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 215         if (!rcu_preempt_blocked_readers_cgp(rnp)) {
 216                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 217                 return;
 218         }
 219         t = list_entry(rnp->gp_tasks->prev,
 220                        struct task_struct, rcu_node_entry);
 221         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
 222                 /*
 223                  * We could be printing a lot while holding a spinlock.
 224                  * Avoid triggering hard lockup.
 225                  */
 226                 touch_nmi_watchdog();
 227                 sched_show_task(t);
 228         }
 229         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 230 }
 231
 232 // Communicate task state back to the RCU CPU stall warning request.
 233 struct rcu_stall_chk_rdr {
 234         int nesting;
 235         union rcu_special rs;
 236         bool on_blkd_list;
 237 };
 238
 239 /*
 240  * Report out the state of a not-running task that is stalling the
 241  * current RCU grace period.
 242  */
 243 static int check_slow_task(struct task_struct *t, void *arg)
 244 {
 245         struct rcu_stall_chk_rdr *rscrp = arg;
 246
 247         if (task_curr(t))
 248                 return -EBUSY; // It is running, so decline to inspect it.
 249         rscrp->nesting = t->rcu_read_lock_nesting;
 250         rscrp->rs = t->rcu_read_unlock_special;
 251         rscrp->on_blkd_list = !list_empty(&t->rcu_node_entry);
 252         return 0;
 253 }
 254
 255 /*
 256  * Scan the current list of tasks blocked within RCU read-side critical
 257  * sections, printing out the tid of each of the first few of them.
 258  */
 259 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
 260         __releases(rnp->lock)
 261 {
 262         int i = 0;
 263         int ndetected = 0;
 264         struct rcu_stall_chk_rdr rscr;
 265         struct task_struct *t;
 266         struct task_struct *ts[8];
 267
 268         lockdep_assert_irqs_disabled();
 269         if (!rcu_preempt_blocked_readers_cgp(rnp)) {
 270                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 271                 return 0;
 272         }
 273         pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
 274                rnp->level, rnp->grplo, rnp->grphi);
 275         t = list_entry(rnp->gp_tasks->prev,
 276                        struct task_struct, rcu_node_entry);
 277         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
 278                 get_task_struct(t);
 279                 ts[i++] = t;
 280                 if (i >= ARRAY_SIZE(ts))
 281                         break;
 282         }
 283         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 284         while (i) {
 285                 t = ts[--i];
 286                 if (task_call_func(t, check_slow_task, &rscr))
 287                         pr_cont(" P%d", t->pid);
 288                 else
 289                         pr_cont(" P%d/%d:%c%c%c%c",
 290                                 t->pid, rscr.nesting,
 291                                 ".b"[rscr.rs.b.blocked],
 292                                 ".q"[rscr.rs.b.need_qs],
 293                                 ".e"[rscr.rs.b.exp_hint],
 294                                 ".l"[rscr.on_blkd_list]);
 295                 lockdep_assert_irqs_disabled();
 296                 put_task_struct(t);
 297                 ndetected++;
 298         }
 299         pr_cont("\n");
 300         return ndetected;
 301 }
 302
 303 #else /* #ifdef CONFIG_PREEMPT_RCU */
 304
 305 /*
 306  * Because preemptible RCU does not exist, we never have to check for
 307  * tasks blocked within RCU read-side critical sections.
 308  */
 309 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
 310 {
 311 }
 312
 313 /*
 314  * Because preemptible RCU does not exist, we never have to check for
 315  * tasks blocked within RCU read-side critical sections.
 316  */
 317 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
 318         __releases(rnp->lock)
 319 {
 320         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 321         return 0;
 322 }
 323 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
 324
 325 /*
 326  * Dump stacks of all tasks running on stalled CPUs.  First try using
 327  * NMIs, but fall back to manual remote stack tracing on architectures
 328  * that don't support NMI-based stack dumps.  The NMI-triggered stack
 329  * traces are more accurate because they are printed by the target CPU.
 330  */
 331 static void rcu_dump_cpu_stacks(void)
 332 {
 333         int cpu;
 334         unsigned long flags;
 335         struct rcu_node *rnp;
 336
 337         rcu_for_each_leaf_node(rnp) {
 338                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 339                 for_each_leaf_node_possible_cpu(rnp, cpu)
 340                         if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
 341                                 if (cpu_is_offline(cpu))
 342                                         pr_err("Offline CPU %d blocking current GP.\n", cpu);
 343                                 else if (!trigger_single_cpu_backtrace(cpu))
 344                                         dump_cpu_task(cpu);
 345                         }
 346                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 347         }
 348 }
 349
 350 static const char * const gp_state_names[] = {
 351         [RCU_GP_IDLE] = "RCU_GP_IDLE",
 352         [RCU_GP_WAIT_GPS] = "RCU_GP_WAIT_GPS",
 353         [RCU_GP_DONE_GPS] = "RCU_GP_DONE_GPS",
 354         [RCU_GP_ONOFF] = "RCU_GP_ONOFF",
 355         [RCU_GP_INIT] = "RCU_GP_INIT",
 356         [RCU_GP_WAIT_FQS] = "RCU_GP_WAIT_FQS",
 357         [RCU_GP_DOING_FQS] = "RCU_GP_DOING_FQS",
 358         [RCU_GP_CLEANUP] = "RCU_GP_CLEANUP",
 359         [RCU_GP_CLEANED] = "RCU_GP_CLEANED",
 360 };
 361
 362 /*
 363  * Convert a ->gp_state value to a character string.
 364  */
 365 static const char *gp_state_getname(short gs)
 366 {
 367         if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
 368                 return "???";
 369         return gp_state_names[gs];
 370 }
 371
 372 /* Is the RCU grace-period kthread being starved of CPU time? */
 373 static bool rcu_is_gp_kthread_starving(unsigned long *jp)
 374 {
 375         unsigned long j = jiffies - READ_ONCE(rcu_state.gp_activity);
 376
 377         if (jp)
 378                 *jp = j;
 379         return j > 2 * HZ;
 380 }
 381
 382 static bool rcu_is_rcuc_kthread_starving(struct rcu_data *rdp, unsigned long *jp)
 383 {
 384         unsigned long j = jiffies - READ_ONCE(rdp->rcuc_activity);
 385
 386         if (jp)
 387                 *jp = j;
 388         return j > 2 * HZ;
 389 }
 390
 391 /*
 392  * Print out diagnostic information for the specified stalled CPU.
 393  *
 394  * If the specified CPU is aware of the current RCU grace period, then
 395  * print the number of scheduling clock interrupts the CPU has taken
 396  * during the time that it has been aware.  Otherwise, print the number
 397  * of RCU grace periods that this CPU is ignorant of, for example, "1"
 398  * if the CPU was aware of the previous grace period.
 399  *
 400  * Also print out idle info.
 401  */
 402 static void print_cpu_stall_info(int cpu)
 403 {
 404         unsigned long delta;
 405         bool falsepositive;
 406         struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 407         char *ticks_title;
 408         unsigned long ticks_value;
 409
 410         /*
 411          * We could be printing a lot while holding a spinlock.  Avoid
 412          * triggering hard lockup.
 413          */
 414         touch_nmi_watchdog();
 415
 416         ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
 417         if (ticks_value) {
 418                 ticks_title = "GPs behind";
 419         } else {
 420                 ticks_title = "ticks this GP";
 421                 ticks_value = rdp->ticks_this_gp;
 422         }
 423         delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
 424         falsepositive = rcu_is_gp_kthread_starving(NULL) &&
 425                         rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp));
 426         pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%03x/%ld/%#lx softirq=%u/%u fqs=%ld %s\n",
 427                cpu,
 428                "O."[!!cpu_online(cpu)],
 429                "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
 430                "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
 431                !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
 432                         rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
 433                                 "!."[!delta],
 434                ticks_value, ticks_title,
 435                rcu_dynticks_snap(rdp) & 0xfff,
 436                rdp->dynticks_nesting, rdp->dynticks_nmi_nesting,
 437                rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
 438                data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
 439                falsepositive ? " (false positive?)" : "");
 440 }
 441
 442 static void rcuc_kthread_dump(struct rcu_data *rdp)
 443 {
 444         int cpu;
 445         unsigned long j;
 446         struct task_struct *rcuc;
 447
 448         rcuc = rdp->rcu_cpu_kthread_task;
 449         if (!rcuc)
 450                 return;
 451
 452         cpu = task_cpu(rcuc);
 453         if (cpu_is_offline(cpu) || idle_cpu(cpu))
 454                 return;
 455
 456         if (!rcu_is_rcuc_kthread_starving(rdp, &j))
 457                 return;
 458
 459         pr_err("%s kthread starved for %ld jiffies\n", rcuc->comm, j);
 460         sched_show_task(rcuc);
 461         if (!trigger_single_cpu_backtrace(cpu))
 462                 dump_cpu_task(cpu);
 463 }
 464
 465 /* Complain about starvation of grace-period kthread.  */
 466 static void rcu_check_gp_kthread_starvation(void)
 467 {
 468         int cpu;
 469         struct task_struct *gpk = rcu_state.gp_kthread;
 470         unsigned long j;
 471
 472         if (rcu_is_gp_kthread_starving(&j)) {
 473                 cpu = gpk ? task_cpu(gpk) : -1;
 474                 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x ->cpu=%d\n",
 475                        rcu_state.name, j,
 476                        (long)rcu_seq_current(&rcu_state.gp_seq),
 477                        data_race(READ_ONCE(rcu_state.gp_flags)),
 478                        gp_state_getname(rcu_state.gp_state),
 479                        data_race(READ_ONCE(rcu_state.gp_state)),
 480                        gpk ? data_race(READ_ONCE(gpk->__state)) : ~0, cpu);
 481                 if (gpk) {
 482                         pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name);
 483                         pr_err("RCU grace-period kthread stack dump:\n");
 484                         sched_show_task(gpk);
 485                         if (cpu >= 0) {
 486                                 if (cpu_is_offline(cpu)) {
 487                                         pr_err("RCU GP kthread last ran on offline CPU %d.\n", cpu);
 488                                 } else  {
 489                                         pr_err("Stack dump where RCU GP kthread last ran:\n");
 490                                         if (!trigger_single_cpu_backtrace(cpu))
 491                                                 dump_cpu_task(cpu);
 492                                 }
 493                         }
 494                         wake_up_process(gpk);
 495                 }
 496         }
 497 }
 498
 499 /* Complain about missing wakeups from expired fqs wait timer */
 500 static void rcu_check_gp_kthread_expired_fqs_timer(void)
 501 {
 502         struct task_struct *gpk = rcu_state.gp_kthread;
 503         short gp_state;
 504         unsigned long jiffies_fqs;
 505         int cpu;
 506
 507         /*
 508          * Order reads of .gp_state and .jiffies_force_qs.
 509          * Matching smp_wmb() is present in rcu_gp_fqs_loop().
 510          */
 511         gp_state = smp_load_acquire(&rcu_state.gp_state);
 512         jiffies_fqs = READ_ONCE(rcu_state.jiffies_force_qs);
 513
 514         if (gp_state == RCU_GP_WAIT_FQS &&
 515             time_after(jiffies, jiffies_fqs + RCU_STALL_MIGHT_MIN) &&
 516             gpk && !READ_ONCE(gpk->on_rq)) {
 517                 cpu = task_cpu(gpk);
 518                 pr_err("%s kthread timer wakeup didn't happen for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x\n",
 519                        rcu_state.name, (jiffies - jiffies_fqs),
 520                        (long)rcu_seq_current(&rcu_state.gp_seq),
 521                        data_race(rcu_state.gp_flags),
 522                        gp_state_getname(RCU_GP_WAIT_FQS), RCU_GP_WAIT_FQS,
 523                        data_race(READ_ONCE(gpk->__state)));
 524                 pr_err("\tPossible timer handling issue on cpu=%d timer-softirq=%u\n",
 525                        cpu, kstat_softirqs_cpu(TIMER_SOFTIRQ, cpu));
 526         }
 527 }
 528
 529 static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
 530 {
 531         int cpu;
 532         unsigned long flags;
 533         unsigned long gpa;
 534         unsigned long j;
 535         int ndetected = 0;
 536         struct rcu_node *rnp;
 537         long totqlen = 0;
 538
 539         lockdep_assert_irqs_disabled();
 540
 541         /* Kick and suppress, if so configured. */
 542         rcu_stall_kick_kthreads();
 543         if (rcu_stall_is_suppressed())
 544                 return;
 545
 546         /*
 547          * OK, time to rat on our buddy...
 548          * See Documentation/RCU/stallwarn.rst for info on how to debug
 549          * RCU CPU stall warnings.
 550          */
 551         trace_rcu_stall_warning(rcu_state.name, TPS("StallDetected"));
 552         pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
 553         rcu_for_each_leaf_node(rnp) {
 554                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 555                 if (rnp->qsmask != 0) {
 556                         for_each_leaf_node_possible_cpu(rnp, cpu)
 557                                 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
 558                                         print_cpu_stall_info(cpu);
 559                                         ndetected++;
 560                                 }
 561                 }
 562                 ndetected += rcu_print_task_stall(rnp, flags); // Releases rnp->lock.
 563                 lockdep_assert_irqs_disabled();
 564         }
 565
 566         for_each_possible_cpu(cpu)
 567                 totqlen += rcu_get_n_cbs_cpu(cpu);
 568         pr_cont("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu)\n",
 569                smp_processor_id(), (long)(jiffies - gps),
 570                (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
 571         if (ndetected) {
 572                 rcu_dump_cpu_stacks();
 573
 574                 /* Complain about tasks blocking the grace period. */
 575                 rcu_for_each_leaf_node(rnp)
 576                         rcu_print_detail_task_stall_rnp(rnp);
 577         } else {
 578                 if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
 579                         pr_err("INFO: Stall ended before state dump start\n");
 580                 } else {
 581                         j = jiffies;
 582                         gpa = data_race(READ_ONCE(rcu_state.gp_activity));
 583                         pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
 584                                rcu_state.name, j - gpa, j, gpa,
 585                                data_race(READ_ONCE(jiffies_till_next_fqs)),
 586                                data_race(READ_ONCE(rcu_get_root()->qsmask)));
 587                 }
 588         }
 589         /* Rewrite if needed in case of slow consoles. */
 590         if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
 591                 WRITE_ONCE(rcu_state.jiffies_stall,
 592                            jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
 593
 594         rcu_check_gp_kthread_expired_fqs_timer();
 595         rcu_check_gp_kthread_starvation();
 596
 597         panic_on_rcu_stall();
 598
 599         rcu_force_quiescent_state();  /* Kick them all. */
 600 }
 601
 602 static void print_cpu_stall(unsigned long gps)
 603 {
 604         int cpu;
 605         unsigned long flags;
 606         struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 607         struct rcu_node *rnp = rcu_get_root();
 608         long totqlen = 0;
 609
 610         lockdep_assert_irqs_disabled();
 611
 612         /* Kick and suppress, if so configured. */
 613         rcu_stall_kick_kthreads();
 614         if (rcu_stall_is_suppressed())
 615                 return;
 616
 617         /*
 618          * OK, time to rat on ourselves...
 619          * See Documentation/RCU/stallwarn.rst for info on how to debug
 620          * RCU CPU stall warnings.
 621          */
 622         trace_rcu_stall_warning(rcu_state.name, TPS("SelfDetected"));
 623         pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
 624         raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
 625         print_cpu_stall_info(smp_processor_id());
 626         raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
 627         for_each_possible_cpu(cpu)
 628                 totqlen += rcu_get_n_cbs_cpu(cpu);
 629         pr_cont("\t(t=%lu jiffies g=%ld q=%lu)\n",
 630                 jiffies - gps,
 631                 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
 632
 633         rcu_check_gp_kthread_expired_fqs_timer();
 634         rcu_check_gp_kthread_starvation();
 635
 636         if (!use_softirq)
 637                 rcuc_kthread_dump(rdp);
 638
 639         rcu_dump_cpu_stacks();
 640
 641         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 642         /* Rewrite if needed in case of slow consoles. */
 643         if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
 644                 WRITE_ONCE(rcu_state.jiffies_stall,
 645                            jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
 646         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 647
 648         panic_on_rcu_stall();
 649
 650         /*
 651          * Attempt to revive the RCU machinery by forcing a context switch.
 652          *
 653          * A context switch would normally allow the RCU state machine to make
 654          * progress and it could be we're stuck in kernel space without context
 655          * switches for an entirely unreasonable amount of time.
 656          */
 657         set_tsk_need_resched(current);
 658         set_preempt_need_resched();
 659 }
 660
 661 static void check_cpu_stall(struct rcu_data *rdp)
 662 {
 663         bool didstall = false;
 664         unsigned long gs1;
 665         unsigned long gs2;
 666         unsigned long gps;
 667         unsigned long j;
 668         unsigned long jn;
 669         unsigned long js;
 670         struct rcu_node *rnp;
 671
 672         lockdep_assert_irqs_disabled();
 673         if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
 674             !rcu_gp_in_progress())
 675                 return;
 676         rcu_stall_kick_kthreads();
 677         j = jiffies;
 678
 679         /*
 680          * Lots of memory barriers to reject false positives.
 681          *
 682          * The idea is to pick up rcu_state.gp_seq, then
 683          * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
 684          * another copy of rcu_state.gp_seq.  These values are updated in
 685          * the opposite order with memory barriers (or equivalent) during
 686          * grace-period initialization and cleanup.  Now, a false positive
 687          * can occur if we get an new value of rcu_state.gp_start and a old
 688          * value of rcu_state.jiffies_stall.  But given the memory barriers,
 689          * the only way that this can happen is if one grace period ends
 690          * and another starts between these two fetches.  This is detected
 691          * by comparing the second fetch of rcu_state.gp_seq with the
 692          * previous fetch from rcu_state.gp_seq.
 693          *
 694          * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
 695          * and rcu_state.gp_start suffice to forestall false positives.
 696          */
 697         gs1 = READ_ONCE(rcu_state.gp_seq);
 698         smp_rmb(); /* Pick up ->gp_seq first... */
 699         js = READ_ONCE(rcu_state.jiffies_stall);
 700         smp_rmb(); /* ...then ->jiffies_stall before the rest... */
 701         gps = READ_ONCE(rcu_state.gp_start);
 702         smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
 703         gs2 = READ_ONCE(rcu_state.gp_seq);
 704         if (gs1 != gs2 ||
 705             ULONG_CMP_LT(j, js) ||
 706             ULONG_CMP_GE(gps, js))
 707                 return; /* No stall or GP completed since entering function. */
 708         rnp = rdp->mynode;
 709         jn = jiffies + ULONG_MAX / 2;
 710         if (rcu_gp_in_progress() &&
 711             (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
 712             cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
 713
 714                 /*
 715                  * If a virtual machine is stopped by the host it can look to
 716                  * the watchdog like an RCU stall. Check to see if the host
 717                  * stopped the vm.
 718                  */
 719                 if (kvm_check_and_clear_guest_paused())
 720                         return;
 721
 722                 /* We haven't checked in, so go dump stack. */
 723                 print_cpu_stall(gps);
 724                 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
 725                         rcu_ftrace_dump(DUMP_ALL);
 726                 didstall = true;
 727
 728         } else if (rcu_gp_in_progress() &&
 729                    ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
 730                    cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
 731
 732                 /*
 733                  * If a virtual machine is stopped by the host it can look to
 734                  * the watchdog like an RCU stall. Check to see if the host
 735                  * stopped the vm.
 736                  */
 737                 if (kvm_check_and_clear_guest_paused())
 738                         return;
 739
 740                 /* They had a few time units to dump stack, so complain. */
 741                 print_other_cpu_stall(gs2, gps);
 742                 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
 743                         rcu_ftrace_dump(DUMP_ALL);
 744                 didstall = true;
 745         }
 746         if (didstall && READ_ONCE(rcu_state.jiffies_stall) == jn) {
 747                 jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
 748                 WRITE_ONCE(rcu_state.jiffies_stall, jn);
 749         }
 750 }
 751
 752 //////////////////////////////////////////////////////////////////////////////
 753 //
 754 // RCU forward-progress mechanisms, including of callback invocation.
 755
 756
 757 /*
 758  * Check to see if a failure to end RCU priority inversion was due to
 759  * a CPU not passing through a quiescent state.  When this happens, there
 760  * is nothing that RCU priority boosting can do to help, so we shouldn't
 761  * count this as an RCU priority boosting failure.  A return of true says
 762  * RCU priority boosting is to blame, and false says otherwise.  If false
 763  * is returned, the first of the CPUs to blame is stored through cpup.
 764  * If there was no CPU blocking the current grace period, but also nothing
 765  * in need of being boosted, *cpup is set to -1.  This can happen in case
 766  * of vCPU preemption while the last CPU is reporting its quiscent state,
 767  * for example.
 768  *
 769  * If cpup is NULL, then a lockless quick check is carried out, suitable
 770  * for high-rate usage.  On the other hand, if cpup is non-NULL, each
 771  * rcu_node structure's ->lock is acquired, ruling out high-rate usage.
 772  */
 773 bool rcu_check_boost_fail(unsigned long gp_state, int *cpup)
 774 {
 775         bool atb = false;
 776         int cpu;
 777         unsigned long flags;
 778         struct rcu_node *rnp;
 779
 780         rcu_for_each_leaf_node(rnp) {
 781                 if (!cpup) {
 782                         if (data_race(READ_ONCE(rnp->qsmask))) {
 783                                 return false;
 784                         } else {
 785                                 if (READ_ONCE(rnp->gp_tasks))
 786                                         atb = true;
 787                                 continue;
 788                         }
 789                 }
 790                 *cpup = -1;
 791                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 792                 if (rnp->gp_tasks)
 793                         atb = true;
 794                 if (!rnp->qsmask) {
 795                         // No CPUs without quiescent states for this rnp.
 796                         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 797                         continue;
 798                 }
 799                 // Find the first holdout CPU.
 800                 for_each_leaf_node_possible_cpu(rnp, cpu) {
 801                         if (rnp->qsmask & (1UL << (cpu - rnp->grplo))) {
 802                                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 803                                 *cpup = cpu;
 804                                 return false;
 805                         }
 806                 }
 807                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 808         }
 809         // Can't blame CPUs, so must blame RCU priority boosting.
 810         return atb;
 811 }
 812 EXPORT_SYMBOL_GPL(rcu_check_boost_fail);
 813
 814 /*
 815  * Show the state of the grace-period kthreads.
 816  */
 817 void show_rcu_gp_kthreads(void)
 818 {
 819         unsigned long cbs = 0;
 820         int cpu;
 821         unsigned long j;
 822         unsigned long ja;
 823         unsigned long jr;
 824         unsigned long js;
 825         unsigned long jw;
 826         struct rcu_data *rdp;
 827         struct rcu_node *rnp;
 828         struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
 829
 830         j = jiffies;
 831         ja = j - data_race(READ_ONCE(rcu_state.gp_activity));
 832         jr = j - data_race(READ_ONCE(rcu_state.gp_req_activity));
 833         js = j - data_race(READ_ONCE(rcu_state.gp_start));
 834         jw = j - data_race(READ_ONCE(rcu_state.gp_wake_time));
 835         pr_info("%s: wait state: %s(%d) ->state: %#x ->rt_priority %u delta ->gp_start %lu ->gp_activity %lu ->gp_req_activity %lu ->gp_wake_time %lu ->gp_wake_seq %ld ->gp_seq %ld ->gp_seq_needed %ld ->gp_max %lu ->gp_flags %#x\n",
 836                 rcu_state.name, gp_state_getname(rcu_state.gp_state),
 837                 data_race(READ_ONCE(rcu_state.gp_state)),
 838                 t ? data_race(READ_ONCE(t->__state)) : 0x1ffff, t ? t->rt_priority : 0xffU,
 839                 js, ja, jr, jw, (long)data_race(READ_ONCE(rcu_state.gp_wake_seq)),
 840                 (long)data_race(READ_ONCE(rcu_state.gp_seq)),
 841                 (long)data_race(READ_ONCE(rcu_get_root()->gp_seq_needed)),
 842                 data_race(READ_ONCE(rcu_state.gp_max)),
 843                 data_race(READ_ONCE(rcu_state.gp_flags)));
 844         rcu_for_each_node_breadth_first(rnp) {
 845                 if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq), READ_ONCE(rnp->gp_seq_needed)) &&
 846                     !data_race(READ_ONCE(rnp->qsmask)) && !data_race(READ_ONCE(rnp->boost_tasks)) &&
 847                     !data_race(READ_ONCE(rnp->exp_tasks)) && !data_race(READ_ONCE(rnp->gp_tasks)))
 848                         continue;
 849                 pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld ->qsmask %#lx %c%c%c%c ->n_boosts %ld\n",
 850                         rnp->grplo, rnp->grphi,
 851                         (long)data_race(READ_ONCE(rnp->gp_seq)),
 852                         (long)data_race(READ_ONCE(rnp->gp_seq_needed)),
 853                         data_race(READ_ONCE(rnp->qsmask)),
 854                         ".b"[!!data_race(READ_ONCE(rnp->boost_kthread_task))],
 855                         ".B"[!!data_race(READ_ONCE(rnp->boost_tasks))],
 856                         ".E"[!!data_race(READ_ONCE(rnp->exp_tasks))],
 857                         ".G"[!!data_race(READ_ONCE(rnp->gp_tasks))],
 858                         data_race(READ_ONCE(rnp->n_boosts)));
 859                 if (!rcu_is_leaf_node(rnp))
 860                         continue;
 861                 for_each_leaf_node_possible_cpu(rnp, cpu) {
 862                         rdp = per_cpu_ptr(&rcu_data, cpu);
 863                         if (READ_ONCE(rdp->gpwrap) ||
 864                             ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
 865                                          READ_ONCE(rdp->gp_seq_needed)))
 866                                 continue;
 867                         pr_info("\tcpu %d ->gp_seq_needed %ld\n",
 868                                 cpu, (long)data_race(READ_ONCE(rdp->gp_seq_needed)));
 869                 }
 870         }
 871         for_each_possible_cpu(cpu) {
 872                 rdp = per_cpu_ptr(&rcu_data, cpu);
 873                 cbs += data_race(READ_ONCE(rdp->n_cbs_invoked));
 874                 if (rcu_segcblist_is_offloaded(&rdp->cblist))
 875                         show_rcu_nocb_state(rdp);
 876         }
 877         pr_info("RCU callbacks invoked since boot: %lu\n", cbs);
 878         show_rcu_tasks_gp_kthreads();
 879 }
 880 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
 881
 882 /*
 883  * This function checks for grace-period requests that fail to motivate
 884  * RCU to come out of its idle mode.
 885  */
 886 static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
 887                                      const unsigned long gpssdelay)
 888 {
 889         unsigned long flags;
 890         unsigned long j;
 891         struct rcu_node *rnp_root = rcu_get_root();
 892         static atomic_t warned = ATOMIC_INIT(0);
 893
 894         if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
 895             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 896                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 897             !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread.
 898                 return;
 899         j = jiffies; /* Expensive access, and in common case don't get here. */
 900         if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 901             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 902             atomic_read(&warned))
 903                 return;
 904
 905         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 906         j = jiffies;
 907         if (rcu_gp_in_progress() ||
 908             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 909                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 910             time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 911             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 912             atomic_read(&warned)) {
 913                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 914                 return;
 915         }
 916         /* Hold onto the leaf lock to make others see warned==1. */
 917
 918         if (rnp_root != rnp)
 919                 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
 920         j = jiffies;
 921         if (rcu_gp_in_progress() ||
 922             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 923                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 924             time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 925             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 926             atomic_xchg(&warned, 1)) {
 927                 if (rnp_root != rnp)
 928                         /* irqs remain disabled. */
 929                         raw_spin_unlock_rcu_node(rnp_root);
 930                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 931                 return;
 932         }
 933         WARN_ON(1);
 934         if (rnp_root != rnp)
 935                 raw_spin_unlock_rcu_node(rnp_root);
 936         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 937         show_rcu_gp_kthreads();
 938 }
 939
 940 /*
 941  * Do a forward-progress check for rcutorture.  This is normally invoked
 942  * due to an OOM event.  The argument "j" gives the time period during
 943  * which rcutorture would like progress to have been made.
 944  */
 945 void rcu_fwd_progress_check(unsigned long j)
 946 {
 947         unsigned long cbs;
 948         int cpu;
 949         unsigned long max_cbs = 0;
 950         int max_cpu = -1;
 951         struct rcu_data *rdp;
 952
 953         if (rcu_gp_in_progress()) {
 954                 pr_info("%s: GP age %lu jiffies\n",
 955                         __func__, jiffies - data_race(READ_ONCE(rcu_state.gp_start)));
 956                 show_rcu_gp_kthreads();
 957         } else {
 958                 pr_info("%s: Last GP end %lu jiffies ago\n",
 959                         __func__, jiffies - data_race(READ_ONCE(rcu_state.gp_end)));
 960                 preempt_disable();
 961                 rdp = this_cpu_ptr(&rcu_data);
 962                 rcu_check_gp_start_stall(rdp->mynode, rdp, j);
 963                 preempt_enable();
 964         }
 965         for_each_possible_cpu(cpu) {
 966                 cbs = rcu_get_n_cbs_cpu(cpu);
 967                 if (!cbs)
 968                         continue;
 969                 if (max_cpu < 0)
 970                         pr_info("%s: callbacks", __func__);
 971                 pr_cont(" %d: %lu", cpu, cbs);
 972                 if (cbs <= max_cbs)
 973                         continue;
 974                 max_cbs = cbs;
 975                 max_cpu = cpu;
 976         }
 977         if (max_cpu >= 0)
 978                 pr_cont("\n");
 979 }
 980 EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
 981
 982 /* Commandeer a sysrq key to dump RCU's tree. */
 983 static bool sysrq_rcu;
 984 module_param(sysrq_rcu, bool, 0444);
 985
 986 /* Dump grace-period-request information due to commandeered sysrq. */
 987 static void sysrq_show_rcu(int key)
 988 {
 989         show_rcu_gp_kthreads();
 990 }
 991
 992 static const struct sysrq_key_op sysrq_rcudump_op = {
 993         .handler = sysrq_show_rcu,
 994         .help_msg = "show-rcu(y)",
 995         .action_msg = "Show RCU tree",
 996         .enable_mask = SYSRQ_ENABLE_DUMP,
 997 };
 998
 999 static int __init rcu_sysrq_init(void)
1000 {
1001         if (sysrq_rcu)
1002                 return register_sysrq_key('y', &sysrq_rcudump_op);
1003         return 0;
1004 }
1005 early_initcall(rcu_sysrq_init);