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 bool 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 false; // 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 true;
 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 (!try_invoke_on_locked_down_task(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 #ifdef CONFIG_RCU_FAST_NO_HZ
 351
 352 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
 353 {
 354         struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 355
 356         sprintf(cp, "last_accelerate: %04lx/%04lx dyntick_enabled: %d",
 357                 rdp->last_accelerate & 0xffff, jiffies & 0xffff,
 358                 !!rdp->tick_nohz_enabled_snap);
 359 }
 360
 361 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
 362
 363 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
 364 {
 365         *cp = '\0';
 366 }
 367
 368 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
 369
 370 static const char * const gp_state_names[] = {
 371         [RCU_GP_IDLE] = "RCU_GP_IDLE",
 372         [RCU_GP_WAIT_GPS] = "RCU_GP_WAIT_GPS",
 373         [RCU_GP_DONE_GPS] = "RCU_GP_DONE_GPS",
 374         [RCU_GP_ONOFF] = "RCU_GP_ONOFF",
 375         [RCU_GP_INIT] = "RCU_GP_INIT",
 376         [RCU_GP_WAIT_FQS] = "RCU_GP_WAIT_FQS",
 377         [RCU_GP_DOING_FQS] = "RCU_GP_DOING_FQS",
 378         [RCU_GP_CLEANUP] = "RCU_GP_CLEANUP",
 379         [RCU_GP_CLEANED] = "RCU_GP_CLEANED",
 380 };
 381
 382 /*
 383  * Convert a ->gp_state value to a character string.
 384  */
 385 static const char *gp_state_getname(short gs)
 386 {
 387         if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
 388                 return "???";
 389         return gp_state_names[gs];
 390 }
 391
 392 /* Is the RCU grace-period kthread being starved of CPU time? */
 393 static bool rcu_is_gp_kthread_starving(unsigned long *jp)
 394 {
 395         unsigned long j = jiffies - READ_ONCE(rcu_state.gp_activity);
 396
 397         if (jp)
 398                 *jp = j;
 399         return j > 2 * HZ;
 400 }
 401
 402 /*
 403  * Print out diagnostic information for the specified stalled CPU.
 404  *
 405  * If the specified CPU is aware of the current RCU grace period, then
 406  * print the number of scheduling clock interrupts the CPU has taken
 407  * during the time that it has been aware.  Otherwise, print the number
 408  * of RCU grace periods that this CPU is ignorant of, for example, "1"
 409  * if the CPU was aware of the previous grace period.
 410  *
 411  * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
 412  */
 413 static void print_cpu_stall_info(int cpu)
 414 {
 415         unsigned long delta;
 416         bool falsepositive;
 417         char fast_no_hz[72];
 418         struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 419         char *ticks_title;
 420         unsigned long ticks_value;
 421
 422         /*
 423          * We could be printing a lot while holding a spinlock.  Avoid
 424          * triggering hard lockup.
 425          */
 426         touch_nmi_watchdog();
 427
 428         ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
 429         if (ticks_value) {
 430                 ticks_title = "GPs behind";
 431         } else {
 432                 ticks_title = "ticks this GP";
 433                 ticks_value = rdp->ticks_this_gp;
 434         }
 435         print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
 436         delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
 437         falsepositive = rcu_is_gp_kthread_starving(NULL) &&
 438                         rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp));
 439         pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%03x/%ld/%#lx softirq=%u/%u fqs=%ld %s%s\n",
 440                cpu,
 441                "O."[!!cpu_online(cpu)],
 442                "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
 443                "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
 444                !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
 445                         rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
 446                                 "!."[!delta],
 447                ticks_value, ticks_title,
 448                rcu_dynticks_snap(rdp) & 0xfff,
 449                rdp->dynticks_nesting, rdp->dynticks_nmi_nesting,
 450                rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
 451                data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
 452                fast_no_hz,
 453                falsepositive ? " (false positive?)" : "");
 454 }
 455
 456 /* Complain about starvation of grace-period kthread.  */
 457 static void rcu_check_gp_kthread_starvation(void)
 458 {
 459         int cpu;
 460         struct task_struct *gpk = rcu_state.gp_kthread;
 461         unsigned long j;
 462
 463         if (rcu_is_gp_kthread_starving(&j)) {
 464                 cpu = gpk ? task_cpu(gpk) : -1;
 465                 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x ->cpu=%d\n",
 466                        rcu_state.name, j,
 467                        (long)rcu_seq_current(&rcu_state.gp_seq),
 468                        data_race(READ_ONCE(rcu_state.gp_flags)),
 469                        gp_state_getname(rcu_state.gp_state),
 470                        data_race(READ_ONCE(rcu_state.gp_state)),
 471                        gpk ? data_race(READ_ONCE(gpk->__state)) : ~0, cpu);
 472                 if (gpk) {
 473                         pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name);
 474                         pr_err("RCU grace-period kthread stack dump:\n");
 475                         sched_show_task(gpk);
 476                         if (cpu >= 0) {
 477                                 if (cpu_is_offline(cpu)) {
 478                                         pr_err("RCU GP kthread last ran on offline CPU %d.\n", cpu);
 479                                 } else  {
 480                                         pr_err("Stack dump where RCU GP kthread last ran:\n");
 481                                         if (!trigger_single_cpu_backtrace(cpu))
 482                                                 dump_cpu_task(cpu);
 483                                 }
 484                         }
 485                         wake_up_process(gpk);
 486                 }
 487         }
 488 }
 489
 490 /* Complain about missing wakeups from expired fqs wait timer */
 491 static void rcu_check_gp_kthread_expired_fqs_timer(void)
 492 {
 493         struct task_struct *gpk = rcu_state.gp_kthread;
 494         short gp_state;
 495         unsigned long jiffies_fqs;
 496         int cpu;
 497
 498         /*
 499          * Order reads of .gp_state and .jiffies_force_qs.
 500          * Matching smp_wmb() is present in rcu_gp_fqs_loop().
 501          */
 502         gp_state = smp_load_acquire(&rcu_state.gp_state);
 503         jiffies_fqs = READ_ONCE(rcu_state.jiffies_force_qs);
 504
 505         if (gp_state == RCU_GP_WAIT_FQS &&
 506             time_after(jiffies, jiffies_fqs + RCU_STALL_MIGHT_MIN) &&
 507             gpk && !READ_ONCE(gpk->on_rq)) {
 508                 cpu = task_cpu(gpk);
 509                 pr_err("%s kthread timer wakeup didn't happen for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x\n",
 510                        rcu_state.name, (jiffies - jiffies_fqs),
 511                        (long)rcu_seq_current(&rcu_state.gp_seq),
 512                        data_race(rcu_state.gp_flags),
 513                        gp_state_getname(RCU_GP_WAIT_FQS), RCU_GP_WAIT_FQS,
 514                        data_race(READ_ONCE(gpk->__state)));
 515                 pr_err("\tPossible timer handling issue on cpu=%d timer-softirq=%u\n",
 516                        cpu, kstat_softirqs_cpu(TIMER_SOFTIRQ, cpu));
 517         }
 518 }
 519
 520 static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
 521 {
 522         int cpu;
 523         unsigned long flags;
 524         unsigned long gpa;
 525         unsigned long j;
 526         int ndetected = 0;
 527         struct rcu_node *rnp;
 528         long totqlen = 0;
 529
 530         lockdep_assert_irqs_disabled();
 531
 532         /* Kick and suppress, if so configured. */
 533         rcu_stall_kick_kthreads();
 534         if (rcu_stall_is_suppressed())
 535                 return;
 536
 537         /*
 538          * OK, time to rat on our buddy...
 539          * See Documentation/RCU/stallwarn.rst for info on how to debug
 540          * RCU CPU stall warnings.
 541          */
 542         trace_rcu_stall_warning(rcu_state.name, TPS("StallDetected"));
 543         pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
 544         rcu_for_each_leaf_node(rnp) {
 545                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 546                 if (rnp->qsmask != 0) {
 547                         for_each_leaf_node_possible_cpu(rnp, cpu)
 548                                 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
 549                                         print_cpu_stall_info(cpu);
 550                                         ndetected++;
 551                                 }
 552                 }
 553                 ndetected += rcu_print_task_stall(rnp, flags); // Releases rnp->lock.
 554                 lockdep_assert_irqs_disabled();
 555         }
 556
 557         for_each_possible_cpu(cpu)
 558                 totqlen += rcu_get_n_cbs_cpu(cpu);
 559         pr_cont("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu)\n",
 560                smp_processor_id(), (long)(jiffies - gps),
 561                (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
 562         if (ndetected) {
 563                 rcu_dump_cpu_stacks();
 564
 565                 /* Complain about tasks blocking the grace period. */
 566                 rcu_for_each_leaf_node(rnp)
 567                         rcu_print_detail_task_stall_rnp(rnp);
 568         } else {
 569                 if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
 570                         pr_err("INFO: Stall ended before state dump start\n");
 571                 } else {
 572                         j = jiffies;
 573                         gpa = data_race(READ_ONCE(rcu_state.gp_activity));
 574                         pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
 575                                rcu_state.name, j - gpa, j, gpa,
 576                                data_race(READ_ONCE(jiffies_till_next_fqs)),
 577                                data_race(READ_ONCE(rcu_get_root()->qsmask)));
 578                 }
 579         }
 580         /* Rewrite if needed in case of slow consoles. */
 581         if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
 582                 WRITE_ONCE(rcu_state.jiffies_stall,
 583                            jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
 584
 585         rcu_check_gp_kthread_expired_fqs_timer();
 586         rcu_check_gp_kthread_starvation();
 587
 588         panic_on_rcu_stall();
 589
 590         rcu_force_quiescent_state();  /* Kick them all. */
 591 }
 592
 593 static void print_cpu_stall(unsigned long gps)
 594 {
 595         int cpu;
 596         unsigned long flags;
 597         struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 598         struct rcu_node *rnp = rcu_get_root();
 599         long totqlen = 0;
 600
 601         lockdep_assert_irqs_disabled();
 602
 603         /* Kick and suppress, if so configured. */
 604         rcu_stall_kick_kthreads();
 605         if (rcu_stall_is_suppressed())
 606                 return;
 607
 608         /*
 609          * OK, time to rat on ourselves...
 610          * See Documentation/RCU/stallwarn.rst for info on how to debug
 611          * RCU CPU stall warnings.
 612          */
 613         trace_rcu_stall_warning(rcu_state.name, TPS("SelfDetected"));
 614         pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
 615         raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
 616         print_cpu_stall_info(smp_processor_id());
 617         raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
 618         for_each_possible_cpu(cpu)
 619                 totqlen += rcu_get_n_cbs_cpu(cpu);
 620         pr_cont("\t(t=%lu jiffies g=%ld q=%lu)\n",
 621                 jiffies - gps,
 622                 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
 623
 624         rcu_check_gp_kthread_expired_fqs_timer();
 625         rcu_check_gp_kthread_starvation();
 626
 627         rcu_dump_cpu_stacks();
 628
 629         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 630         /* Rewrite if needed in case of slow consoles. */
 631         if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
 632                 WRITE_ONCE(rcu_state.jiffies_stall,
 633                            jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
 634         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 635
 636         panic_on_rcu_stall();
 637
 638         /*
 639          * Attempt to revive the RCU machinery by forcing a context switch.
 640          *
 641          * A context switch would normally allow the RCU state machine to make
 642          * progress and it could be we're stuck in kernel space without context
 643          * switches for an entirely unreasonable amount of time.
 644          */
 645         set_tsk_need_resched(current);
 646         set_preempt_need_resched();
 647 }
 648
 649 static void check_cpu_stall(struct rcu_data *rdp)
 650 {
 651         bool didstall = false;
 652         unsigned long gs1;
 653         unsigned long gs2;
 654         unsigned long gps;
 655         unsigned long j;
 656         unsigned long jn;
 657         unsigned long js;
 658         struct rcu_node *rnp;
 659
 660         lockdep_assert_irqs_disabled();
 661         if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
 662             !rcu_gp_in_progress())
 663                 return;
 664         rcu_stall_kick_kthreads();
 665         j = jiffies;
 666
 667         /*
 668          * Lots of memory barriers to reject false positives.
 669          *
 670          * The idea is to pick up rcu_state.gp_seq, then
 671          * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
 672          * another copy of rcu_state.gp_seq.  These values are updated in
 673          * the opposite order with memory barriers (or equivalent) during
 674          * grace-period initialization and cleanup.  Now, a false positive
 675          * can occur if we get an new value of rcu_state.gp_start and a old
 676          * value of rcu_state.jiffies_stall.  But given the memory barriers,
 677          * the only way that this can happen is if one grace period ends
 678          * and another starts between these two fetches.  This is detected
 679          * by comparing the second fetch of rcu_state.gp_seq with the
 680          * previous fetch from rcu_state.gp_seq.
 681          *
 682          * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
 683          * and rcu_state.gp_start suffice to forestall false positives.
 684          */
 685         gs1 = READ_ONCE(rcu_state.gp_seq);
 686         smp_rmb(); /* Pick up ->gp_seq first... */
 687         js = READ_ONCE(rcu_state.jiffies_stall);
 688         smp_rmb(); /* ...then ->jiffies_stall before the rest... */
 689         gps = READ_ONCE(rcu_state.gp_start);
 690         smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
 691         gs2 = READ_ONCE(rcu_state.gp_seq);
 692         if (gs1 != gs2 ||
 693             ULONG_CMP_LT(j, js) ||
 694             ULONG_CMP_GE(gps, js))
 695                 return; /* No stall or GP completed since entering function. */
 696         rnp = rdp->mynode;
 697         jn = jiffies + ULONG_MAX / 2;
 698         if (rcu_gp_in_progress() &&
 699             (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
 700             cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
 701
 702                 /*
 703                  * If a virtual machine is stopped by the host it can look to
 704                  * the watchdog like an RCU stall. Check to see if the host
 705                  * stopped the vm.
 706                  */
 707                 if (kvm_check_and_clear_guest_paused())
 708                         return;
 709
 710                 /* We haven't checked in, so go dump stack. */
 711                 print_cpu_stall(gps);
 712                 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
 713                         rcu_ftrace_dump(DUMP_ALL);
 714                 didstall = true;
 715
 716         } else if (rcu_gp_in_progress() &&
 717                    ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
 718                    cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
 719
 720                 /*
 721                  * If a virtual machine is stopped by the host it can look to
 722                  * the watchdog like an RCU stall. Check to see if the host
 723                  * stopped the vm.
 724                  */
 725                 if (kvm_check_and_clear_guest_paused())
 726                         return;
 727
 728                 /* They had a few time units to dump stack, so complain. */
 729                 print_other_cpu_stall(gs2, gps);
 730                 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
 731                         rcu_ftrace_dump(DUMP_ALL);
 732                 didstall = true;
 733         }
 734         if (didstall && READ_ONCE(rcu_state.jiffies_stall) == jn) {
 735                 jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
 736                 WRITE_ONCE(rcu_state.jiffies_stall, jn);
 737         }
 738 }
 739
 740 //////////////////////////////////////////////////////////////////////////////
 741 //
 742 // RCU forward-progress mechanisms, including of callback invocation.
 743
 744
 745 /*
 746  * Check to see if a failure to end RCU priority inversion was due to
 747  * a CPU not passing through a quiescent state.  When this happens, there
 748  * is nothing that RCU priority boosting can do to help, so we shouldn't
 749  * count this as an RCU priority boosting failure.  A return of true says
 750  * RCU priority boosting is to blame, and false says otherwise.  If false
 751  * is returned, the first of the CPUs to blame is stored through cpup.
 752  * If there was no CPU blocking the current grace period, but also nothing
 753  * in need of being boosted, *cpup is set to -1.  This can happen in case
 754  * of vCPU preemption while the last CPU is reporting its quiscent state,
 755  * for example.
 756  *
 757  * If cpup is NULL, then a lockless quick check is carried out, suitable
 758  * for high-rate usage.  On the other hand, if cpup is non-NULL, each
 759  * rcu_node structure's ->lock is acquired, ruling out high-rate usage.
 760  */
 761 bool rcu_check_boost_fail(unsigned long gp_state, int *cpup)
 762 {
 763         bool atb = false;
 764         int cpu;
 765         unsigned long flags;
 766         struct rcu_node *rnp;
 767
 768         rcu_for_each_leaf_node(rnp) {
 769                 if (!cpup) {
 770                         if (data_race(READ_ONCE(rnp->qsmask))) {
 771                                 return false;
 772                         } else {
 773                                 if (READ_ONCE(rnp->gp_tasks))
 774                                         atb = true;
 775                                 continue;
 776                         }
 777                 }
 778                 *cpup = -1;
 779                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 780                 if (rnp->gp_tasks)
 781                         atb = true;
 782                 if (!rnp->qsmask) {
 783                         // No CPUs without quiescent states for this rnp.
 784                         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 785                         continue;
 786                 }
 787                 // Find the first holdout CPU.
 788                 for_each_leaf_node_possible_cpu(rnp, cpu) {
 789                         if (rnp->qsmask & (1UL << (cpu - rnp->grplo))) {
 790                                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 791                                 *cpup = cpu;
 792                                 return false;
 793                         }
 794                 }
 795                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 796         }
 797         // Can't blame CPUs, so must blame RCU priority boosting.
 798         return atb;
 799 }
 800 EXPORT_SYMBOL_GPL(rcu_check_boost_fail);
 801
 802 /*
 803  * Show the state of the grace-period kthreads.
 804  */
 805 void show_rcu_gp_kthreads(void)
 806 {
 807         unsigned long cbs = 0;
 808         int cpu;
 809         unsigned long j;
 810         unsigned long ja;
 811         unsigned long jr;
 812         unsigned long js;
 813         unsigned long jw;
 814         struct rcu_data *rdp;
 815         struct rcu_node *rnp;
 816         struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
 817
 818         j = jiffies;
 819         ja = j - data_race(READ_ONCE(rcu_state.gp_activity));
 820         jr = j - data_race(READ_ONCE(rcu_state.gp_req_activity));
 821         js = j - data_race(READ_ONCE(rcu_state.gp_start));
 822         jw = j - data_race(READ_ONCE(rcu_state.gp_wake_time));
 823         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",
 824                 rcu_state.name, gp_state_getname(rcu_state.gp_state),
 825                 data_race(READ_ONCE(rcu_state.gp_state)),
 826                 t ? data_race(READ_ONCE(t->__state)) : 0x1ffff, t ? t->rt_priority : 0xffU,
 827                 js, ja, jr, jw, (long)data_race(READ_ONCE(rcu_state.gp_wake_seq)),
 828                 (long)data_race(READ_ONCE(rcu_state.gp_seq)),
 829                 (long)data_race(READ_ONCE(rcu_get_root()->gp_seq_needed)),
 830                 data_race(READ_ONCE(rcu_state.gp_max)),
 831                 data_race(READ_ONCE(rcu_state.gp_flags)));
 832         rcu_for_each_node_breadth_first(rnp) {
 833                 if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq), READ_ONCE(rnp->gp_seq_needed)) &&
 834                     !data_race(READ_ONCE(rnp->qsmask)) && !data_race(READ_ONCE(rnp->boost_tasks)) &&
 835                     !data_race(READ_ONCE(rnp->exp_tasks)) && !data_race(READ_ONCE(rnp->gp_tasks)))
 836                         continue;
 837                 pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld ->qsmask %#lx %c%c%c%c ->n_boosts %ld\n",
 838                         rnp->grplo, rnp->grphi,
 839                         (long)data_race(READ_ONCE(rnp->gp_seq)),
 840                         (long)data_race(READ_ONCE(rnp->gp_seq_needed)),
 841                         data_race(READ_ONCE(rnp->qsmask)),
 842                         ".b"[!!data_race(READ_ONCE(rnp->boost_kthread_task))],
 843                         ".B"[!!data_race(READ_ONCE(rnp->boost_tasks))],
 844                         ".E"[!!data_race(READ_ONCE(rnp->exp_tasks))],
 845                         ".G"[!!data_race(READ_ONCE(rnp->gp_tasks))],
 846                         data_race(READ_ONCE(rnp->n_boosts)));
 847                 if (!rcu_is_leaf_node(rnp))
 848                         continue;
 849                 for_each_leaf_node_possible_cpu(rnp, cpu) {
 850                         rdp = per_cpu_ptr(&rcu_data, cpu);
 851                         if (READ_ONCE(rdp->gpwrap) ||
 852                             ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
 853                                          READ_ONCE(rdp->gp_seq_needed)))
 854                                 continue;
 855                         pr_info("\tcpu %d ->gp_seq_needed %ld\n",
 856                                 cpu, (long)data_race(READ_ONCE(rdp->gp_seq_needed)));
 857                 }
 858         }
 859         for_each_possible_cpu(cpu) {
 860                 rdp = per_cpu_ptr(&rcu_data, cpu);
 861                 cbs += data_race(READ_ONCE(rdp->n_cbs_invoked));
 862                 if (rcu_segcblist_is_offloaded(&rdp->cblist))
 863                         show_rcu_nocb_state(rdp);
 864         }
 865         pr_info("RCU callbacks invoked since boot: %lu\n", cbs);
 866         show_rcu_tasks_gp_kthreads();
 867 }
 868 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
 869
 870 /*
 871  * This function checks for grace-period requests that fail to motivate
 872  * RCU to come out of its idle mode.
 873  */
 874 static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
 875                                      const unsigned long gpssdelay)
 876 {
 877         unsigned long flags;
 878         unsigned long j;
 879         struct rcu_node *rnp_root = rcu_get_root();
 880         static atomic_t warned = ATOMIC_INIT(0);
 881
 882         if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
 883             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 884                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 885             !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread.
 886                 return;
 887         j = jiffies; /* Expensive access, and in common case don't get here. */
 888         if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 889             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 890             atomic_read(&warned))
 891                 return;
 892
 893         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 894         j = jiffies;
 895         if (rcu_gp_in_progress() ||
 896             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 897                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 898             time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 899             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 900             atomic_read(&warned)) {
 901                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 902                 return;
 903         }
 904         /* Hold onto the leaf lock to make others see warned==1. */
 905
 906         if (rnp_root != rnp)
 907                 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
 908         j = jiffies;
 909         if (rcu_gp_in_progress() ||
 910             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 911                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 912             time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 913             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 914             atomic_xchg(&warned, 1)) {
 915                 if (rnp_root != rnp)
 916                         /* irqs remain disabled. */
 917                         raw_spin_unlock_rcu_node(rnp_root);
 918                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 919                 return;
 920         }
 921         WARN_ON(1);
 922         if (rnp_root != rnp)
 923                 raw_spin_unlock_rcu_node(rnp_root);
 924         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 925         show_rcu_gp_kthreads();
 926 }
 927
 928 /*
 929  * Do a forward-progress check for rcutorture.  This is normally invoked
 930  * due to an OOM event.  The argument "j" gives the time period during
 931  * which rcutorture would like progress to have been made.
 932  */
 933 void rcu_fwd_progress_check(unsigned long j)
 934 {
 935         unsigned long cbs;
 936         int cpu;
 937         unsigned long max_cbs = 0;
 938         int max_cpu = -1;
 939         struct rcu_data *rdp;
 940
 941         if (rcu_gp_in_progress()) {
 942                 pr_info("%s: GP age %lu jiffies\n",
 943                         __func__, jiffies - data_race(READ_ONCE(rcu_state.gp_start)));
 944                 show_rcu_gp_kthreads();
 945         } else {
 946                 pr_info("%s: Last GP end %lu jiffies ago\n",
 947                         __func__, jiffies - data_race(READ_ONCE(rcu_state.gp_end)));
 948                 preempt_disable();
 949                 rdp = this_cpu_ptr(&rcu_data);
 950                 rcu_check_gp_start_stall(rdp->mynode, rdp, j);
 951                 preempt_enable();
 952         }
 953         for_each_possible_cpu(cpu) {
 954                 cbs = rcu_get_n_cbs_cpu(cpu);
 955                 if (!cbs)
 956                         continue;
 957                 if (max_cpu < 0)
 958                         pr_info("%s: callbacks", __func__);
 959                 pr_cont(" %d: %lu", cpu, cbs);
 960                 if (cbs <= max_cbs)
 961                         continue;
 962                 max_cbs = cbs;
 963                 max_cpu = cpu;
 964         }
 965         if (max_cpu >= 0)
 966                 pr_cont("\n");
 967 }
 968 EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
 969
 970 /* Commandeer a sysrq key to dump RCU's tree. */
 971 static bool sysrq_rcu;
 972 module_param(sysrq_rcu, bool, 0444);
 973
 974 /* Dump grace-period-request information due to commandeered sysrq. */
 975 static void sysrq_show_rcu(int key)
 976 {
 977         show_rcu_gp_kthreads();
 978 }
 979
 980 static const struct sysrq_key_op sysrq_rcudump_op = {
 981         .handler = sysrq_show_rcu,
 982         .help_msg = "show-rcu(y)",
 983         .action_msg = "Show RCU tree",
 984         .enable_mask = SYSRQ_ENABLE_DUMP,
 985 };
 986
 987 static int __init rcu_sysrq_init(void)
 988 {
 989         if (sysrq_rcu)
 990                 return register_sysrq_key('y', &sysrq_rcudump_op);
 991         return 0;
 992 }
 993 early_initcall(rcu_sysrq_init);