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 int rcu_exp_jiffies_till_stall_check(void) 29 { 30 int cpu_stall_timeout = READ_ONCE(rcu_exp_cpu_stall_timeout); 31 int exp_stall_delay_delta = 0; 32 int till_stall_check; 33 34 // Zero says to use rcu_cpu_stall_timeout, but in milliseconds. 35 if (!cpu_stall_timeout) 36 cpu_stall_timeout = jiffies_to_msecs(rcu_jiffies_till_stall_check()); 37 38 // Limit check must be consistent with the Kconfig limits for 39 // CONFIG_RCU_EXP_CPU_STALL_TIMEOUT, so check the allowed range. 40 // The minimum clamped value is "2UL", because at least one full 41 // tick has to be guaranteed. 42 till_stall_check = clamp(msecs_to_jiffies(cpu_stall_timeout), 2UL, 300UL * HZ); 43 44 if (cpu_stall_timeout && jiffies_to_msecs(till_stall_check) != cpu_stall_timeout) 45 WRITE_ONCE(rcu_exp_cpu_stall_timeout, jiffies_to_msecs(till_stall_check)); 46 47 #ifdef CONFIG_PROVE_RCU 48 /* Add extra ~25% out of till_stall_check. */ 49 exp_stall_delay_delta = ((till_stall_check * 25) / 100) + 1; 50 #endif 51 52 return till_stall_check + exp_stall_delay_delta; 53 } 54 EXPORT_SYMBOL_GPL(rcu_exp_jiffies_till_stall_check); 55 56 /* Limit-check stall timeouts specified at boottime and runtime. */ 57 int rcu_jiffies_till_stall_check(void) 58 { 59 int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout); 60 61 /* 62 * Limit check must be consistent with the Kconfig limits 63 * for CONFIG_RCU_CPU_STALL_TIMEOUT. 64 */ 65 if (till_stall_check < 3) { 66 WRITE_ONCE(rcu_cpu_stall_timeout, 3); 67 till_stall_check = 3; 68 } else if (till_stall_check > 300) { 69 WRITE_ONCE(rcu_cpu_stall_timeout, 300); 70 till_stall_check = 300; 71 } 72 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA; 73 } 74 EXPORT_SYMBOL_GPL(rcu_jiffies_till_stall_check); 75 76 /** 77 * rcu_gp_might_be_stalled - Is it likely that the grace period is stalled? 78 * 79 * Returns @true if the current grace period is sufficiently old that 80 * it is reasonable to assume that it might be stalled. This can be 81 * useful when deciding whether to allocate memory to enable RCU-mediated 82 * freeing on the one hand or just invoking synchronize_rcu() on the other. 83 * The latter is preferable when the grace period is stalled. 84 * 85 * Note that sampling of the .gp_start and .gp_seq fields must be done 86 * carefully to avoid false positives at the beginnings and ends of 87 * grace periods. 88 */ 89 bool rcu_gp_might_be_stalled(void) 90 { 91 unsigned long d = rcu_jiffies_till_stall_check() / RCU_STALL_MIGHT_DIV; 92 unsigned long j = jiffies; 93 94 if (d < RCU_STALL_MIGHT_MIN) 95 d = RCU_STALL_MIGHT_MIN; 96 smp_mb(); // jiffies before .gp_seq to avoid false positives. 97 if (!rcu_gp_in_progress()) 98 return false; 99 // Long delays at this point avoids false positive, but a delay 100 // of ULONG_MAX/4 jiffies voids your no-false-positive warranty. 101 smp_mb(); // .gp_seq before second .gp_start 102 // And ditto here. 103 return !time_before(j, READ_ONCE(rcu_state.gp_start) + d); 104 } 105 106 /* Don't do RCU CPU stall warnings during long sysrq printouts. */ 107 void rcu_sysrq_start(void) 108 { 109 if (!rcu_cpu_stall_suppress) 110 rcu_cpu_stall_suppress = 2; 111 } 112 113 void rcu_sysrq_end(void) 114 { 115 if (rcu_cpu_stall_suppress == 2) 116 rcu_cpu_stall_suppress = 0; 117 } 118 119 /* Don't print RCU CPU stall warnings during a kernel panic. */ 120 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr) 121 { 122 rcu_cpu_stall_suppress = 1; 123 return NOTIFY_DONE; 124 } 125 126 static struct notifier_block rcu_panic_block = { 127 .notifier_call = rcu_panic, 128 }; 129 130 static int __init check_cpu_stall_init(void) 131 { 132 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block); 133 return 0; 134 } 135 early_initcall(check_cpu_stall_init); 136 137 /* If so specified via sysctl, panic, yielding cleaner stall-warning output. */ 138 static void panic_on_rcu_stall(void) 139 { 140 static int cpu_stall; 141 142 if (++cpu_stall < sysctl_max_rcu_stall_to_panic) 143 return; 144 145 if (sysctl_panic_on_rcu_stall) 146 panic("RCU Stall\n"); 147 } 148 149 /** 150 * rcu_cpu_stall_reset - restart stall-warning timeout for current grace period 151 * 152 * To perform the reset request from the caller, disable stall detection until 153 * 3 fqs loops have passed. This is required to ensure a fresh jiffies is 154 * loaded. It should be safe to do from the fqs loop as enough timer 155 * interrupts and context switches should have passed. 156 * 157 * The caller must disable hard irqs. 158 */ 159 void rcu_cpu_stall_reset(void) 160 { 161 WRITE_ONCE(rcu_state.nr_fqs_jiffies_stall, 3); 162 WRITE_ONCE(rcu_state.jiffies_stall, ULONG_MAX); 163 } 164 165 ////////////////////////////////////////////////////////////////////////////// 166 // 167 // Interaction with RCU grace periods 168 169 /* Start of new grace period, so record stall time (and forcing times). */ 170 static void record_gp_stall_check_time(void) 171 { 172 unsigned long j = jiffies; 173 unsigned long j1; 174 175 WRITE_ONCE(rcu_state.gp_start, j); 176 j1 = rcu_jiffies_till_stall_check(); 177 smp_mb(); // ->gp_start before ->jiffies_stall and caller's ->gp_seq. 178 WRITE_ONCE(rcu_state.nr_fqs_jiffies_stall, 0); 179 WRITE_ONCE(rcu_state.jiffies_stall, j + j1); 180 rcu_state.jiffies_resched = j + j1 / 2; 181 rcu_state.n_force_qs_gpstart = READ_ONCE(rcu_state.n_force_qs); 182 } 183 184 /* Zero ->ticks_this_gp and snapshot the number of RCU softirq handlers. */ 185 static void zero_cpu_stall_ticks(struct rcu_data *rdp) 186 { 187 rdp->ticks_this_gp = 0; 188 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id()); 189 WRITE_ONCE(rdp->last_fqs_resched, jiffies); 190 } 191 192 /* 193 * If too much time has passed in the current grace period, and if 194 * so configured, go kick the relevant kthreads. 195 */ 196 static void rcu_stall_kick_kthreads(void) 197 { 198 unsigned long j; 199 200 if (!READ_ONCE(rcu_kick_kthreads)) 201 return; 202 j = READ_ONCE(rcu_state.jiffies_kick_kthreads); 203 if (time_after(jiffies, j) && rcu_state.gp_kthread && 204 (rcu_gp_in_progress() || READ_ONCE(rcu_state.gp_flags))) { 205 WARN_ONCE(1, "Kicking %s grace-period kthread\n", 206 rcu_state.name); 207 rcu_ftrace_dump(DUMP_ALL); 208 wake_up_process(rcu_state.gp_kthread); 209 WRITE_ONCE(rcu_state.jiffies_kick_kthreads, j + HZ); 210 } 211 } 212 213 /* 214 * Handler for the irq_work request posted about halfway into the RCU CPU 215 * stall timeout, and used to detect excessive irq disabling. Set state 216 * appropriately, but just complain if there is unexpected state on entry. 217 */ 218 static void rcu_iw_handler(struct irq_work *iwp) 219 { 220 struct rcu_data *rdp; 221 struct rcu_node *rnp; 222 223 rdp = container_of(iwp, struct rcu_data, rcu_iw); 224 rnp = rdp->mynode; 225 raw_spin_lock_rcu_node(rnp); 226 if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) { 227 rdp->rcu_iw_gp_seq = rnp->gp_seq; 228 rdp->rcu_iw_pending = false; 229 } 230 raw_spin_unlock_rcu_node(rnp); 231 } 232 233 ////////////////////////////////////////////////////////////////////////////// 234 // 235 // Printing RCU CPU stall warnings 236 237 #ifdef CONFIG_PREEMPT_RCU 238 239 /* 240 * Dump detailed information for all tasks blocking the current RCU 241 * grace period on the specified rcu_node structure. 242 */ 243 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp) 244 { 245 unsigned long flags; 246 struct task_struct *t; 247 248 raw_spin_lock_irqsave_rcu_node(rnp, flags); 249 if (!rcu_preempt_blocked_readers_cgp(rnp)) { 250 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 251 return; 252 } 253 t = list_entry(rnp->gp_tasks->prev, 254 struct task_struct, rcu_node_entry); 255 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) { 256 /* 257 * We could be printing a lot while holding a spinlock. 258 * Avoid triggering hard lockup. 259 */ 260 touch_nmi_watchdog(); 261 sched_show_task(t); 262 } 263 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 264 } 265 266 // Communicate task state back to the RCU CPU stall warning request. 267 struct rcu_stall_chk_rdr { 268 int nesting; 269 union rcu_special rs; 270 bool on_blkd_list; 271 }; 272 273 /* 274 * Report out the state of a not-running task that is stalling the 275 * current RCU grace period. 276 */ 277 static int check_slow_task(struct task_struct *t, void *arg) 278 { 279 struct rcu_stall_chk_rdr *rscrp = arg; 280 281 if (task_curr(t)) 282 return -EBUSY; // It is running, so decline to inspect it. 283 rscrp->nesting = t->rcu_read_lock_nesting; 284 rscrp->rs = t->rcu_read_unlock_special; 285 rscrp->on_blkd_list = !list_empty(&t->rcu_node_entry); 286 return 0; 287 } 288 289 /* 290 * Scan the current list of tasks blocked within RCU read-side critical 291 * sections, printing out the tid of each of the first few of them. 292 */ 293 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags) 294 __releases(rnp->lock) 295 { 296 int i = 0; 297 int ndetected = 0; 298 struct rcu_stall_chk_rdr rscr; 299 struct task_struct *t; 300 struct task_struct *ts[8]; 301 302 lockdep_assert_irqs_disabled(); 303 if (!rcu_preempt_blocked_readers_cgp(rnp)) { 304 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 305 return 0; 306 } 307 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):", 308 rnp->level, rnp->grplo, rnp->grphi); 309 t = list_entry(rnp->gp_tasks->prev, 310 struct task_struct, rcu_node_entry); 311 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) { 312 get_task_struct(t); 313 ts[i++] = t; 314 if (i >= ARRAY_SIZE(ts)) 315 break; 316 } 317 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 318 while (i) { 319 t = ts[--i]; 320 if (task_call_func(t, check_slow_task, &rscr)) 321 pr_cont(" P%d", t->pid); 322 else 323 pr_cont(" P%d/%d:%c%c%c%c", 324 t->pid, rscr.nesting, 325 ".b"[rscr.rs.b.blocked], 326 ".q"[rscr.rs.b.need_qs], 327 ".e"[rscr.rs.b.exp_hint], 328 ".l"[rscr.on_blkd_list]); 329 lockdep_assert_irqs_disabled(); 330 put_task_struct(t); 331 ndetected++; 332 } 333 pr_cont("\n"); 334 return ndetected; 335 } 336 337 #else /* #ifdef CONFIG_PREEMPT_RCU */ 338 339 /* 340 * Because preemptible RCU does not exist, we never have to check for 341 * tasks blocked within RCU read-side critical sections. 342 */ 343 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp) 344 { 345 } 346 347 /* 348 * Because preemptible RCU does not exist, we never have to check for 349 * tasks blocked within RCU read-side critical sections. 350 */ 351 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags) 352 __releases(rnp->lock) 353 { 354 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 355 return 0; 356 } 357 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */ 358 359 /* 360 * Dump stacks of all tasks running on stalled CPUs. First try using 361 * NMIs, but fall back to manual remote stack tracing on architectures 362 * that don't support NMI-based stack dumps. The NMI-triggered stack 363 * traces are more accurate because they are printed by the target CPU. 364 */ 365 static void rcu_dump_cpu_stacks(void) 366 { 367 int cpu; 368 unsigned long flags; 369 struct rcu_node *rnp; 370 371 rcu_for_each_leaf_node(rnp) { 372 raw_spin_lock_irqsave_rcu_node(rnp, flags); 373 for_each_leaf_node_possible_cpu(rnp, cpu) 374 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) { 375 if (cpu_is_offline(cpu)) 376 pr_err("Offline CPU %d blocking current GP.\n", cpu); 377 else 378 dump_cpu_task(cpu); 379 } 380 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 381 } 382 } 383 384 static const char * const gp_state_names[] = { 385 [RCU_GP_IDLE] = "RCU_GP_IDLE", 386 [RCU_GP_WAIT_GPS] = "RCU_GP_WAIT_GPS", 387 [RCU_GP_DONE_GPS] = "RCU_GP_DONE_GPS", 388 [RCU_GP_ONOFF] = "RCU_GP_ONOFF", 389 [RCU_GP_INIT] = "RCU_GP_INIT", 390 [RCU_GP_WAIT_FQS] = "RCU_GP_WAIT_FQS", 391 [RCU_GP_DOING_FQS] = "RCU_GP_DOING_FQS", 392 [RCU_GP_CLEANUP] = "RCU_GP_CLEANUP", 393 [RCU_GP_CLEANED] = "RCU_GP_CLEANED", 394 }; 395 396 /* 397 * Convert a ->gp_state value to a character string. 398 */ 399 static const char *gp_state_getname(short gs) 400 { 401 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names)) 402 return "???"; 403 return gp_state_names[gs]; 404 } 405 406 /* Is the RCU grace-period kthread being starved of CPU time? */ 407 static bool rcu_is_gp_kthread_starving(unsigned long *jp) 408 { 409 unsigned long j = jiffies - READ_ONCE(rcu_state.gp_activity); 410 411 if (jp) 412 *jp = j; 413 return j > 2 * HZ; 414 } 415 416 static bool rcu_is_rcuc_kthread_starving(struct rcu_data *rdp, unsigned long *jp) 417 { 418 int cpu; 419 struct task_struct *rcuc; 420 unsigned long j; 421 422 rcuc = rdp->rcu_cpu_kthread_task; 423 if (!rcuc) 424 return false; 425 426 cpu = task_cpu(rcuc); 427 if (cpu_is_offline(cpu) || idle_cpu(cpu)) 428 return false; 429 430 j = jiffies - READ_ONCE(rdp->rcuc_activity); 431 432 if (jp) 433 *jp = j; 434 return j > 2 * HZ; 435 } 436 437 static void print_cpu_stat_info(int cpu) 438 { 439 struct rcu_snap_record rsr, *rsrp; 440 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); 441 struct kernel_cpustat *kcsp = &kcpustat_cpu(cpu); 442 443 if (!rcu_cpu_stall_cputime) 444 return; 445 446 rsrp = &rdp->snap_record; 447 if (rsrp->gp_seq != rdp->gp_seq) 448 return; 449 450 rsr.cputime_irq = kcpustat_field(kcsp, CPUTIME_IRQ, cpu); 451 rsr.cputime_softirq = kcpustat_field(kcsp, CPUTIME_SOFTIRQ, cpu); 452 rsr.cputime_system = kcpustat_field(kcsp, CPUTIME_SYSTEM, cpu); 453 454 pr_err("\t hardirqs softirqs csw/system\n"); 455 pr_err("\t number: %8ld %10d %12lld\n", 456 kstat_cpu_irqs_sum(cpu) - rsrp->nr_hardirqs, 457 kstat_cpu_softirqs_sum(cpu) - rsrp->nr_softirqs, 458 nr_context_switches_cpu(cpu) - rsrp->nr_csw); 459 pr_err("\tcputime: %8lld %10lld %12lld ==> %d(ms)\n", 460 div_u64(rsr.cputime_irq - rsrp->cputime_irq, NSEC_PER_MSEC), 461 div_u64(rsr.cputime_softirq - rsrp->cputime_softirq, NSEC_PER_MSEC), 462 div_u64(rsr.cputime_system - rsrp->cputime_system, NSEC_PER_MSEC), 463 jiffies_to_msecs(jiffies - rsrp->jiffies)); 464 } 465 466 /* 467 * Print out diagnostic information for the specified stalled CPU. 468 * 469 * If the specified CPU is aware of the current RCU grace period, then 470 * print the number of scheduling clock interrupts the CPU has taken 471 * during the time that it has been aware. Otherwise, print the number 472 * of RCU grace periods that this CPU is ignorant of, for example, "1" 473 * if the CPU was aware of the previous grace period. 474 * 475 * Also print out idle info. 476 */ 477 static void print_cpu_stall_info(int cpu) 478 { 479 unsigned long delta; 480 bool falsepositive; 481 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); 482 char *ticks_title; 483 unsigned long ticks_value; 484 bool rcuc_starved; 485 unsigned long j; 486 char buf[32]; 487 488 /* 489 * We could be printing a lot while holding a spinlock. Avoid 490 * triggering hard lockup. 491 */ 492 touch_nmi_watchdog(); 493 494 ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq); 495 if (ticks_value) { 496 ticks_title = "GPs behind"; 497 } else { 498 ticks_title = "ticks this GP"; 499 ticks_value = rdp->ticks_this_gp; 500 } 501 delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq); 502 falsepositive = rcu_is_gp_kthread_starving(NULL) && 503 rcu_dynticks_in_eqs(rcu_dynticks_snap(cpu)); 504 rcuc_starved = rcu_is_rcuc_kthread_starving(rdp, &j); 505 if (rcuc_starved) 506 sprintf(buf, " rcuc=%ld jiffies(starved)", j); 507 pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%04x/%ld/%#lx softirq=%u/%u fqs=%ld%s%s\n", 508 cpu, 509 "O."[!!cpu_online(cpu)], 510 "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)], 511 "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)], 512 !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' : 513 rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' : 514 "!."[!delta], 515 ticks_value, ticks_title, 516 rcu_dynticks_snap(cpu) & 0xffff, 517 ct_dynticks_nesting_cpu(cpu), ct_dynticks_nmi_nesting_cpu(cpu), 518 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu), 519 data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart, 520 rcuc_starved ? buf : "", 521 falsepositive ? " (false positive?)" : ""); 522 523 print_cpu_stat_info(cpu); 524 } 525 526 /* Complain about starvation of grace-period kthread. */ 527 static void rcu_check_gp_kthread_starvation(void) 528 { 529 int cpu; 530 struct task_struct *gpk = rcu_state.gp_kthread; 531 unsigned long j; 532 533 if (rcu_is_gp_kthread_starving(&j)) { 534 cpu = gpk ? task_cpu(gpk) : -1; 535 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x ->cpu=%d\n", 536 rcu_state.name, j, 537 (long)rcu_seq_current(&rcu_state.gp_seq), 538 data_race(READ_ONCE(rcu_state.gp_flags)), 539 gp_state_getname(rcu_state.gp_state), 540 data_race(READ_ONCE(rcu_state.gp_state)), 541 gpk ? data_race(READ_ONCE(gpk->__state)) : ~0, cpu); 542 if (gpk) { 543 pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name); 544 pr_err("RCU grace-period kthread stack dump:\n"); 545 sched_show_task(gpk); 546 if (cpu >= 0) { 547 if (cpu_is_offline(cpu)) { 548 pr_err("RCU GP kthread last ran on offline CPU %d.\n", cpu); 549 } else { 550 pr_err("Stack dump where RCU GP kthread last ran:\n"); 551 dump_cpu_task(cpu); 552 } 553 } 554 wake_up_process(gpk); 555 } 556 } 557 } 558 559 /* Complain about missing wakeups from expired fqs wait timer */ 560 static void rcu_check_gp_kthread_expired_fqs_timer(void) 561 { 562 struct task_struct *gpk = rcu_state.gp_kthread; 563 short gp_state; 564 unsigned long jiffies_fqs; 565 int cpu; 566 567 /* 568 * Order reads of .gp_state and .jiffies_force_qs. 569 * Matching smp_wmb() is present in rcu_gp_fqs_loop(). 570 */ 571 gp_state = smp_load_acquire(&rcu_state.gp_state); 572 jiffies_fqs = READ_ONCE(rcu_state.jiffies_force_qs); 573 574 if (gp_state == RCU_GP_WAIT_FQS && 575 time_after(jiffies, jiffies_fqs + RCU_STALL_MIGHT_MIN) && 576 gpk && !READ_ONCE(gpk->on_rq)) { 577 cpu = task_cpu(gpk); 578 pr_err("%s kthread timer wakeup didn't happen for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x\n", 579 rcu_state.name, (jiffies - jiffies_fqs), 580 (long)rcu_seq_current(&rcu_state.gp_seq), 581 data_race(rcu_state.gp_flags), 582 gp_state_getname(RCU_GP_WAIT_FQS), RCU_GP_WAIT_FQS, 583 data_race(READ_ONCE(gpk->__state))); 584 pr_err("\tPossible timer handling issue on cpu=%d timer-softirq=%u\n", 585 cpu, kstat_softirqs_cpu(TIMER_SOFTIRQ, cpu)); 586 } 587 } 588 589 static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps) 590 { 591 int cpu; 592 unsigned long flags; 593 unsigned long gpa; 594 unsigned long j; 595 int ndetected = 0; 596 struct rcu_node *rnp; 597 long totqlen = 0; 598 599 lockdep_assert_irqs_disabled(); 600 601 /* Kick and suppress, if so configured. */ 602 rcu_stall_kick_kthreads(); 603 if (rcu_stall_is_suppressed()) 604 return; 605 606 /* 607 * OK, time to rat on our buddy... 608 * See Documentation/RCU/stallwarn.rst for info on how to debug 609 * RCU CPU stall warnings. 610 */ 611 trace_rcu_stall_warning(rcu_state.name, TPS("StallDetected")); 612 pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name); 613 rcu_for_each_leaf_node(rnp) { 614 raw_spin_lock_irqsave_rcu_node(rnp, flags); 615 if (rnp->qsmask != 0) { 616 for_each_leaf_node_possible_cpu(rnp, cpu) 617 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) { 618 print_cpu_stall_info(cpu); 619 ndetected++; 620 } 621 } 622 ndetected += rcu_print_task_stall(rnp, flags); // Releases rnp->lock. 623 lockdep_assert_irqs_disabled(); 624 } 625 626 for_each_possible_cpu(cpu) 627 totqlen += rcu_get_n_cbs_cpu(cpu); 628 pr_err("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu ncpus=%d)\n", 629 smp_processor_id(), (long)(jiffies - gps), 630 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen, rcu_state.n_online_cpus); 631 if (ndetected) { 632 rcu_dump_cpu_stacks(); 633 634 /* Complain about tasks blocking the grace period. */ 635 rcu_for_each_leaf_node(rnp) 636 rcu_print_detail_task_stall_rnp(rnp); 637 } else { 638 if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) { 639 pr_err("INFO: Stall ended before state dump start\n"); 640 } else { 641 j = jiffies; 642 gpa = data_race(READ_ONCE(rcu_state.gp_activity)); 643 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n", 644 rcu_state.name, j - gpa, j, gpa, 645 data_race(READ_ONCE(jiffies_till_next_fqs)), 646 data_race(READ_ONCE(rcu_get_root()->qsmask))); 647 } 648 } 649 /* Rewrite if needed in case of slow consoles. */ 650 if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall))) 651 WRITE_ONCE(rcu_state.jiffies_stall, 652 jiffies + 3 * rcu_jiffies_till_stall_check() + 3); 653 654 rcu_check_gp_kthread_expired_fqs_timer(); 655 rcu_check_gp_kthread_starvation(); 656 657 panic_on_rcu_stall(); 658 659 rcu_force_quiescent_state(); /* Kick them all. */ 660 } 661 662 static void print_cpu_stall(unsigned long gps) 663 { 664 int cpu; 665 unsigned long flags; 666 struct rcu_data *rdp = this_cpu_ptr(&rcu_data); 667 struct rcu_node *rnp = rcu_get_root(); 668 long totqlen = 0; 669 670 lockdep_assert_irqs_disabled(); 671 672 /* Kick and suppress, if so configured. */ 673 rcu_stall_kick_kthreads(); 674 if (rcu_stall_is_suppressed()) 675 return; 676 677 /* 678 * OK, time to rat on ourselves... 679 * See Documentation/RCU/stallwarn.rst for info on how to debug 680 * RCU CPU stall warnings. 681 */ 682 trace_rcu_stall_warning(rcu_state.name, TPS("SelfDetected")); 683 pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name); 684 raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags); 685 print_cpu_stall_info(smp_processor_id()); 686 raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags); 687 for_each_possible_cpu(cpu) 688 totqlen += rcu_get_n_cbs_cpu(cpu); 689 pr_err("\t(t=%lu jiffies g=%ld q=%lu ncpus=%d)\n", 690 jiffies - gps, 691 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen, rcu_state.n_online_cpus); 692 693 rcu_check_gp_kthread_expired_fqs_timer(); 694 rcu_check_gp_kthread_starvation(); 695 696 rcu_dump_cpu_stacks(); 697 698 raw_spin_lock_irqsave_rcu_node(rnp, flags); 699 /* Rewrite if needed in case of slow consoles. */ 700 if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall))) 701 WRITE_ONCE(rcu_state.jiffies_stall, 702 jiffies + 3 * rcu_jiffies_till_stall_check() + 3); 703 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 704 705 panic_on_rcu_stall(); 706 707 /* 708 * Attempt to revive the RCU machinery by forcing a context switch. 709 * 710 * A context switch would normally allow the RCU state machine to make 711 * progress and it could be we're stuck in kernel space without context 712 * switches for an entirely unreasonable amount of time. 713 */ 714 set_tsk_need_resched(current); 715 set_preempt_need_resched(); 716 } 717 718 static void check_cpu_stall(struct rcu_data *rdp) 719 { 720 bool didstall = false; 721 unsigned long gs1; 722 unsigned long gs2; 723 unsigned long gps; 724 unsigned long j; 725 unsigned long jn; 726 unsigned long js; 727 struct rcu_node *rnp; 728 729 lockdep_assert_irqs_disabled(); 730 if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) || 731 !rcu_gp_in_progress()) 732 return; 733 rcu_stall_kick_kthreads(); 734 735 /* 736 * Check if it was requested (via rcu_cpu_stall_reset()) that the FQS 737 * loop has to set jiffies to ensure a non-stale jiffies value. This 738 * is required to have good jiffies value after coming out of long 739 * breaks of jiffies updates. Not doing so can cause false positives. 740 */ 741 if (READ_ONCE(rcu_state.nr_fqs_jiffies_stall) > 0) 742 return; 743 744 j = jiffies; 745 746 /* 747 * Lots of memory barriers to reject false positives. 748 * 749 * The idea is to pick up rcu_state.gp_seq, then 750 * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally 751 * another copy of rcu_state.gp_seq. These values are updated in 752 * the opposite order with memory barriers (or equivalent) during 753 * grace-period initialization and cleanup. Now, a false positive 754 * can occur if we get an new value of rcu_state.gp_start and a old 755 * value of rcu_state.jiffies_stall. But given the memory barriers, 756 * the only way that this can happen is if one grace period ends 757 * and another starts between these two fetches. This is detected 758 * by comparing the second fetch of rcu_state.gp_seq with the 759 * previous fetch from rcu_state.gp_seq. 760 * 761 * Given this check, comparisons of jiffies, rcu_state.jiffies_stall, 762 * and rcu_state.gp_start suffice to forestall false positives. 763 */ 764 gs1 = READ_ONCE(rcu_state.gp_seq); 765 smp_rmb(); /* Pick up ->gp_seq first... */ 766 js = READ_ONCE(rcu_state.jiffies_stall); 767 smp_rmb(); /* ...then ->jiffies_stall before the rest... */ 768 gps = READ_ONCE(rcu_state.gp_start); 769 smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */ 770 gs2 = READ_ONCE(rcu_state.gp_seq); 771 if (gs1 != gs2 || 772 ULONG_CMP_LT(j, js) || 773 ULONG_CMP_GE(gps, js)) 774 return; /* No stall or GP completed since entering function. */ 775 rnp = rdp->mynode; 776 jn = jiffies + ULONG_MAX / 2; 777 if (rcu_gp_in_progress() && 778 (READ_ONCE(rnp->qsmask) & rdp->grpmask) && 779 cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) { 780 781 /* 782 * If a virtual machine is stopped by the host it can look to 783 * the watchdog like an RCU stall. Check to see if the host 784 * stopped the vm. 785 */ 786 if (kvm_check_and_clear_guest_paused()) 787 return; 788 789 /* We haven't checked in, so go dump stack. */ 790 print_cpu_stall(gps); 791 if (READ_ONCE(rcu_cpu_stall_ftrace_dump)) 792 rcu_ftrace_dump(DUMP_ALL); 793 didstall = true; 794 795 } else if (rcu_gp_in_progress() && 796 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) && 797 cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) { 798 799 /* 800 * If a virtual machine is stopped by the host it can look to 801 * the watchdog like an RCU stall. Check to see if the host 802 * stopped the vm. 803 */ 804 if (kvm_check_and_clear_guest_paused()) 805 return; 806 807 /* They had a few time units to dump stack, so complain. */ 808 print_other_cpu_stall(gs2, gps); 809 if (READ_ONCE(rcu_cpu_stall_ftrace_dump)) 810 rcu_ftrace_dump(DUMP_ALL); 811 didstall = true; 812 } 813 if (didstall && READ_ONCE(rcu_state.jiffies_stall) == jn) { 814 jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3; 815 WRITE_ONCE(rcu_state.jiffies_stall, jn); 816 } 817 } 818 819 ////////////////////////////////////////////////////////////////////////////// 820 // 821 // RCU forward-progress mechanisms, including of callback invocation. 822 823 824 /* 825 * Check to see if a failure to end RCU priority inversion was due to 826 * a CPU not passing through a quiescent state. When this happens, there 827 * is nothing that RCU priority boosting can do to help, so we shouldn't 828 * count this as an RCU priority boosting failure. A return of true says 829 * RCU priority boosting is to blame, and false says otherwise. If false 830 * is returned, the first of the CPUs to blame is stored through cpup. 831 * If there was no CPU blocking the current grace period, but also nothing 832 * in need of being boosted, *cpup is set to -1. This can happen in case 833 * of vCPU preemption while the last CPU is reporting its quiscent state, 834 * for example. 835 * 836 * If cpup is NULL, then a lockless quick check is carried out, suitable 837 * for high-rate usage. On the other hand, if cpup is non-NULL, each 838 * rcu_node structure's ->lock is acquired, ruling out high-rate usage. 839 */ 840 bool rcu_check_boost_fail(unsigned long gp_state, int *cpup) 841 { 842 bool atb = false; 843 int cpu; 844 unsigned long flags; 845 struct rcu_node *rnp; 846 847 rcu_for_each_leaf_node(rnp) { 848 if (!cpup) { 849 if (data_race(READ_ONCE(rnp->qsmask))) { 850 return false; 851 } else { 852 if (READ_ONCE(rnp->gp_tasks)) 853 atb = true; 854 continue; 855 } 856 } 857 *cpup = -1; 858 raw_spin_lock_irqsave_rcu_node(rnp, flags); 859 if (rnp->gp_tasks) 860 atb = true; 861 if (!rnp->qsmask) { 862 // No CPUs without quiescent states for this rnp. 863 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 864 continue; 865 } 866 // Find the first holdout CPU. 867 for_each_leaf_node_possible_cpu(rnp, cpu) { 868 if (rnp->qsmask & (1UL << (cpu - rnp->grplo))) { 869 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 870 *cpup = cpu; 871 return false; 872 } 873 } 874 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 875 } 876 // Can't blame CPUs, so must blame RCU priority boosting. 877 return atb; 878 } 879 EXPORT_SYMBOL_GPL(rcu_check_boost_fail); 880 881 /* 882 * Show the state of the grace-period kthreads. 883 */ 884 void show_rcu_gp_kthreads(void) 885 { 886 unsigned long cbs = 0; 887 int cpu; 888 unsigned long j; 889 unsigned long ja; 890 unsigned long jr; 891 unsigned long js; 892 unsigned long jw; 893 struct rcu_data *rdp; 894 struct rcu_node *rnp; 895 struct task_struct *t = READ_ONCE(rcu_state.gp_kthread); 896 897 j = jiffies; 898 ja = j - data_race(READ_ONCE(rcu_state.gp_activity)); 899 jr = j - data_race(READ_ONCE(rcu_state.gp_req_activity)); 900 js = j - data_race(READ_ONCE(rcu_state.gp_start)); 901 jw = j - data_race(READ_ONCE(rcu_state.gp_wake_time)); 902 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", 903 rcu_state.name, gp_state_getname(rcu_state.gp_state), 904 data_race(READ_ONCE(rcu_state.gp_state)), 905 t ? data_race(READ_ONCE(t->__state)) : 0x1ffff, t ? t->rt_priority : 0xffU, 906 js, ja, jr, jw, (long)data_race(READ_ONCE(rcu_state.gp_wake_seq)), 907 (long)data_race(READ_ONCE(rcu_state.gp_seq)), 908 (long)data_race(READ_ONCE(rcu_get_root()->gp_seq_needed)), 909 data_race(READ_ONCE(rcu_state.gp_max)), 910 data_race(READ_ONCE(rcu_state.gp_flags))); 911 rcu_for_each_node_breadth_first(rnp) { 912 if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq), READ_ONCE(rnp->gp_seq_needed)) && 913 !data_race(READ_ONCE(rnp->qsmask)) && !data_race(READ_ONCE(rnp->boost_tasks)) && 914 !data_race(READ_ONCE(rnp->exp_tasks)) && !data_race(READ_ONCE(rnp->gp_tasks))) 915 continue; 916 pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld ->qsmask %#lx %c%c%c%c ->n_boosts %ld\n", 917 rnp->grplo, rnp->grphi, 918 (long)data_race(READ_ONCE(rnp->gp_seq)), 919 (long)data_race(READ_ONCE(rnp->gp_seq_needed)), 920 data_race(READ_ONCE(rnp->qsmask)), 921 ".b"[!!data_race(READ_ONCE(rnp->boost_kthread_task))], 922 ".B"[!!data_race(READ_ONCE(rnp->boost_tasks))], 923 ".E"[!!data_race(READ_ONCE(rnp->exp_tasks))], 924 ".G"[!!data_race(READ_ONCE(rnp->gp_tasks))], 925 data_race(READ_ONCE(rnp->n_boosts))); 926 if (!rcu_is_leaf_node(rnp)) 927 continue; 928 for_each_leaf_node_possible_cpu(rnp, cpu) { 929 rdp = per_cpu_ptr(&rcu_data, cpu); 930 if (READ_ONCE(rdp->gpwrap) || 931 ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq), 932 READ_ONCE(rdp->gp_seq_needed))) 933 continue; 934 pr_info("\tcpu %d ->gp_seq_needed %ld\n", 935 cpu, (long)data_race(READ_ONCE(rdp->gp_seq_needed))); 936 } 937 } 938 for_each_possible_cpu(cpu) { 939 rdp = per_cpu_ptr(&rcu_data, cpu); 940 cbs += data_race(READ_ONCE(rdp->n_cbs_invoked)); 941 if (rcu_segcblist_is_offloaded(&rdp->cblist)) 942 show_rcu_nocb_state(rdp); 943 } 944 pr_info("RCU callbacks invoked since boot: %lu\n", cbs); 945 show_rcu_tasks_gp_kthreads(); 946 } 947 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads); 948 949 /* 950 * This function checks for grace-period requests that fail to motivate 951 * RCU to come out of its idle mode. 952 */ 953 static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp, 954 const unsigned long gpssdelay) 955 { 956 unsigned long flags; 957 unsigned long j; 958 struct rcu_node *rnp_root = rcu_get_root(); 959 static atomic_t warned = ATOMIC_INIT(0); 960 961 if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() || 962 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq), 963 READ_ONCE(rnp_root->gp_seq_needed)) || 964 !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread. 965 return; 966 j = jiffies; /* Expensive access, and in common case don't get here. */ 967 if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) || 968 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) || 969 atomic_read(&warned)) 970 return; 971 972 raw_spin_lock_irqsave_rcu_node(rnp, flags); 973 j = jiffies; 974 if (rcu_gp_in_progress() || 975 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq), 976 READ_ONCE(rnp_root->gp_seq_needed)) || 977 time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) || 978 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) || 979 atomic_read(&warned)) { 980 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 981 return; 982 } 983 /* Hold onto the leaf lock to make others see warned==1. */ 984 985 if (rnp_root != rnp) 986 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */ 987 j = jiffies; 988 if (rcu_gp_in_progress() || 989 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq), 990 READ_ONCE(rnp_root->gp_seq_needed)) || 991 time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) || 992 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) || 993 atomic_xchg(&warned, 1)) { 994 if (rnp_root != rnp) 995 /* irqs remain disabled. */ 996 raw_spin_unlock_rcu_node(rnp_root); 997 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 998 return; 999 } 1000 WARN_ON(1); 1001 if (rnp_root != rnp) 1002 raw_spin_unlock_rcu_node(rnp_root); 1003 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 1004 show_rcu_gp_kthreads(); 1005 } 1006 1007 /* 1008 * Do a forward-progress check for rcutorture. This is normally invoked 1009 * due to an OOM event. The argument "j" gives the time period during 1010 * which rcutorture would like progress to have been made. 1011 */ 1012 void rcu_fwd_progress_check(unsigned long j) 1013 { 1014 unsigned long cbs; 1015 int cpu; 1016 unsigned long max_cbs = 0; 1017 int max_cpu = -1; 1018 struct rcu_data *rdp; 1019 1020 if (rcu_gp_in_progress()) { 1021 pr_info("%s: GP age %lu jiffies\n", 1022 __func__, jiffies - data_race(READ_ONCE(rcu_state.gp_start))); 1023 show_rcu_gp_kthreads(); 1024 } else { 1025 pr_info("%s: Last GP end %lu jiffies ago\n", 1026 __func__, jiffies - data_race(READ_ONCE(rcu_state.gp_end))); 1027 preempt_disable(); 1028 rdp = this_cpu_ptr(&rcu_data); 1029 rcu_check_gp_start_stall(rdp->mynode, rdp, j); 1030 preempt_enable(); 1031 } 1032 for_each_possible_cpu(cpu) { 1033 cbs = rcu_get_n_cbs_cpu(cpu); 1034 if (!cbs) 1035 continue; 1036 if (max_cpu < 0) 1037 pr_info("%s: callbacks", __func__); 1038 pr_cont(" %d: %lu", cpu, cbs); 1039 if (cbs <= max_cbs) 1040 continue; 1041 max_cbs = cbs; 1042 max_cpu = cpu; 1043 } 1044 if (max_cpu >= 0) 1045 pr_cont("\n"); 1046 } 1047 EXPORT_SYMBOL_GPL(rcu_fwd_progress_check); 1048 1049 /* Commandeer a sysrq key to dump RCU's tree. */ 1050 static bool sysrq_rcu; 1051 module_param(sysrq_rcu, bool, 0444); 1052 1053 /* Dump grace-period-request information due to commandeered sysrq. */ 1054 static void sysrq_show_rcu(u8 key) 1055 { 1056 show_rcu_gp_kthreads(); 1057 } 1058 1059 static const struct sysrq_key_op sysrq_rcudump_op = { 1060 .handler = sysrq_show_rcu, 1061 .help_msg = "show-rcu(y)", 1062 .action_msg = "Show RCU tree", 1063 .enable_mask = SYSRQ_ENABLE_DUMP, 1064 }; 1065 1066 static int __init rcu_sysrq_init(void) 1067 { 1068 if (sysrq_rcu) 1069 return register_sysrq_key('y', &sysrq_rcudump_op); 1070 return 0; 1071 } 1072 early_initcall(rcu_sysrq_init); 1073