1 // SPDX-License-Identifier: GPL-2.0+ 2 /* 3 * Read-Copy Update mechanism for mutual exclusion 4 * 5 * Copyright IBM Corporation, 2001 6 * 7 * Authors: Dipankar Sarma <dipankar@in.ibm.com> 8 * Manfred Spraul <manfred@colorfullife.com> 9 * 10 * Based on the original work by Paul McKenney <paulmck@linux.ibm.com> 11 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen. 12 * Papers: 13 * http://www.rdrop.com/users/paulmck/paper/rclockpdcsproof.pdf 14 * http://lse.sourceforge.net/locking/rclock_OLS.2001.05.01c.sc.pdf (OLS2001) 15 * 16 * For detailed explanation of Read-Copy Update mechanism see - 17 * http://lse.sourceforge.net/locking/rcupdate.html 18 * 19 */ 20 #include <linux/types.h> 21 #include <linux/kernel.h> 22 #include <linux/init.h> 23 #include <linux/spinlock.h> 24 #include <linux/smp.h> 25 #include <linux/interrupt.h> 26 #include <linux/sched/signal.h> 27 #include <linux/sched/debug.h> 28 #include <linux/atomic.h> 29 #include <linux/bitops.h> 30 #include <linux/percpu.h> 31 #include <linux/notifier.h> 32 #include <linux/cpu.h> 33 #include <linux/mutex.h> 34 #include <linux/export.h> 35 #include <linux/hardirq.h> 36 #include <linux/delay.h> 37 #include <linux/moduleparam.h> 38 #include <linux/kthread.h> 39 #include <linux/tick.h> 40 #include <linux/rcupdate_wait.h> 41 #include <linux/sched/isolation.h> 42 #include <linux/kprobes.h> 43 #include <linux/slab.h> 44 45 #define CREATE_TRACE_POINTS 46 47 #include "rcu.h" 48 49 #ifdef MODULE_PARAM_PREFIX 50 #undef MODULE_PARAM_PREFIX 51 #endif 52 #define MODULE_PARAM_PREFIX "rcupdate." 53 54 #ifndef CONFIG_TINY_RCU 55 module_param(rcu_expedited, int, 0); 56 module_param(rcu_normal, int, 0); 57 static int rcu_normal_after_boot; 58 module_param(rcu_normal_after_boot, int, 0); 59 #endif /* #ifndef CONFIG_TINY_RCU */ 60 61 #ifdef CONFIG_DEBUG_LOCK_ALLOC 62 /** 63 * rcu_read_lock_held_common() - might we be in RCU-sched read-side critical section? 64 * @ret: Best guess answer if lockdep cannot be relied on 65 * 66 * Returns true if lockdep must be ignored, in which case *ret contains 67 * the best guess described below. Otherwise returns false, in which 68 * case *ret tells the caller nothing and the caller should instead 69 * consult lockdep. 70 * 71 * If CONFIG_DEBUG_LOCK_ALLOC is selected, set *ret to nonzero iff in an 72 * RCU-sched read-side critical section. In absence of 73 * CONFIG_DEBUG_LOCK_ALLOC, this assumes we are in an RCU-sched read-side 74 * critical section unless it can prove otherwise. Note that disabling 75 * of preemption (including disabling irqs) counts as an RCU-sched 76 * read-side critical section. This is useful for debug checks in functions 77 * that required that they be called within an RCU-sched read-side 78 * critical section. 79 * 80 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot 81 * and while lockdep is disabled. 82 * 83 * Note that if the CPU is in the idle loop from an RCU point of view (ie: 84 * that we are in the section between rcu_idle_enter() and rcu_idle_exit()) 85 * then rcu_read_lock_held() sets *ret to false even if the CPU did an 86 * rcu_read_lock(). The reason for this is that RCU ignores CPUs that are 87 * in such a section, considering these as in extended quiescent state, 88 * so such a CPU is effectively never in an RCU read-side critical section 89 * regardless of what RCU primitives it invokes. This state of affairs is 90 * required --- we need to keep an RCU-free window in idle where the CPU may 91 * possibly enter into low power mode. This way we can notice an extended 92 * quiescent state to other CPUs that started a grace period. Otherwise 93 * we would delay any grace period as long as we run in the idle task. 94 * 95 * Similarly, we avoid claiming an RCU read lock held if the current 96 * CPU is offline. 97 */ 98 static bool rcu_read_lock_held_common(bool *ret) 99 { 100 if (!debug_lockdep_rcu_enabled()) { 101 *ret = 1; 102 return true; 103 } 104 if (!rcu_is_watching()) { 105 *ret = 0; 106 return true; 107 } 108 if (!rcu_lockdep_current_cpu_online()) { 109 *ret = 0; 110 return true; 111 } 112 return false; 113 } 114 115 int rcu_read_lock_sched_held(void) 116 { 117 bool ret; 118 119 if (rcu_read_lock_held_common(&ret)) 120 return ret; 121 return lock_is_held(&rcu_sched_lock_map) || !preemptible(); 122 } 123 EXPORT_SYMBOL(rcu_read_lock_sched_held); 124 #endif 125 126 #ifndef CONFIG_TINY_RCU 127 128 /* 129 * Should expedited grace-period primitives always fall back to their 130 * non-expedited counterparts? Intended for use within RCU. Note 131 * that if the user specifies both rcu_expedited and rcu_normal, then 132 * rcu_normal wins. (Except during the time period during boot from 133 * when the first task is spawned until the rcu_set_runtime_mode() 134 * core_initcall() is invoked, at which point everything is expedited.) 135 */ 136 bool rcu_gp_is_normal(void) 137 { 138 return READ_ONCE(rcu_normal) && 139 rcu_scheduler_active != RCU_SCHEDULER_INIT; 140 } 141 EXPORT_SYMBOL_GPL(rcu_gp_is_normal); 142 143 static atomic_t rcu_expedited_nesting = ATOMIC_INIT(1); 144 145 /* 146 * Should normal grace-period primitives be expedited? Intended for 147 * use within RCU. Note that this function takes the rcu_expedited 148 * sysfs/boot variable and rcu_scheduler_active into account as well 149 * as the rcu_expedite_gp() nesting. So looping on rcu_unexpedite_gp() 150 * until rcu_gp_is_expedited() returns false is a -really- bad idea. 151 */ 152 bool rcu_gp_is_expedited(void) 153 { 154 return rcu_expedited || atomic_read(&rcu_expedited_nesting); 155 } 156 EXPORT_SYMBOL_GPL(rcu_gp_is_expedited); 157 158 /** 159 * rcu_expedite_gp - Expedite future RCU grace periods 160 * 161 * After a call to this function, future calls to synchronize_rcu() and 162 * friends act as the corresponding synchronize_rcu_expedited() function 163 * had instead been called. 164 */ 165 void rcu_expedite_gp(void) 166 { 167 atomic_inc(&rcu_expedited_nesting); 168 } 169 EXPORT_SYMBOL_GPL(rcu_expedite_gp); 170 171 /** 172 * rcu_unexpedite_gp - Cancel prior rcu_expedite_gp() invocation 173 * 174 * Undo a prior call to rcu_expedite_gp(). If all prior calls to 175 * rcu_expedite_gp() are undone by a subsequent call to rcu_unexpedite_gp(), 176 * and if the rcu_expedited sysfs/boot parameter is not set, then all 177 * subsequent calls to synchronize_rcu() and friends will return to 178 * their normal non-expedited behavior. 179 */ 180 void rcu_unexpedite_gp(void) 181 { 182 atomic_dec(&rcu_expedited_nesting); 183 } 184 EXPORT_SYMBOL_GPL(rcu_unexpedite_gp); 185 186 static bool rcu_boot_ended __read_mostly; 187 188 /* 189 * Inform RCU of the end of the in-kernel boot sequence. 190 */ 191 void rcu_end_inkernel_boot(void) 192 { 193 rcu_unexpedite_gp(); 194 if (rcu_normal_after_boot) 195 WRITE_ONCE(rcu_normal, 1); 196 rcu_boot_ended = 1; 197 } 198 199 /* 200 * Let rcutorture know when it is OK to turn it up to eleven. 201 */ 202 bool rcu_inkernel_boot_has_ended(void) 203 { 204 return rcu_boot_ended; 205 } 206 EXPORT_SYMBOL_GPL(rcu_inkernel_boot_has_ended); 207 208 #endif /* #ifndef CONFIG_TINY_RCU */ 209 210 /* 211 * Test each non-SRCU synchronous grace-period wait API. This is 212 * useful just after a change in mode for these primitives, and 213 * during early boot. 214 */ 215 void rcu_test_sync_prims(void) 216 { 217 if (!IS_ENABLED(CONFIG_PROVE_RCU)) 218 return; 219 synchronize_rcu(); 220 synchronize_rcu_expedited(); 221 } 222 223 #if !defined(CONFIG_TINY_RCU) || defined(CONFIG_SRCU) 224 225 /* 226 * Switch to run-time mode once RCU has fully initialized. 227 */ 228 static int __init rcu_set_runtime_mode(void) 229 { 230 rcu_test_sync_prims(); 231 rcu_scheduler_active = RCU_SCHEDULER_RUNNING; 232 kfree_rcu_scheduler_running(); 233 rcu_test_sync_prims(); 234 return 0; 235 } 236 core_initcall(rcu_set_runtime_mode); 237 238 #endif /* #if !defined(CONFIG_TINY_RCU) || defined(CONFIG_SRCU) */ 239 240 #ifdef CONFIG_DEBUG_LOCK_ALLOC 241 static struct lock_class_key rcu_lock_key; 242 struct lockdep_map rcu_lock_map = 243 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock", &rcu_lock_key); 244 EXPORT_SYMBOL_GPL(rcu_lock_map); 245 246 static struct lock_class_key rcu_bh_lock_key; 247 struct lockdep_map rcu_bh_lock_map = 248 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_bh", &rcu_bh_lock_key); 249 EXPORT_SYMBOL_GPL(rcu_bh_lock_map); 250 251 static struct lock_class_key rcu_sched_lock_key; 252 struct lockdep_map rcu_sched_lock_map = 253 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_sched", &rcu_sched_lock_key); 254 EXPORT_SYMBOL_GPL(rcu_sched_lock_map); 255 256 static struct lock_class_key rcu_callback_key; 257 struct lockdep_map rcu_callback_map = 258 STATIC_LOCKDEP_MAP_INIT("rcu_callback", &rcu_callback_key); 259 EXPORT_SYMBOL_GPL(rcu_callback_map); 260 261 int notrace debug_lockdep_rcu_enabled(void) 262 { 263 return rcu_scheduler_active != RCU_SCHEDULER_INACTIVE && debug_locks && 264 current->lockdep_recursion == 0; 265 } 266 EXPORT_SYMBOL_GPL(debug_lockdep_rcu_enabled); 267 NOKPROBE_SYMBOL(debug_lockdep_rcu_enabled); 268 269 /** 270 * rcu_read_lock_held() - might we be in RCU read-side critical section? 271 * 272 * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an RCU 273 * read-side critical section. In absence of CONFIG_DEBUG_LOCK_ALLOC, 274 * this assumes we are in an RCU read-side critical section unless it can 275 * prove otherwise. This is useful for debug checks in functions that 276 * require that they be called within an RCU read-side critical section. 277 * 278 * Checks debug_lockdep_rcu_enabled() to prevent false positives during boot 279 * and while lockdep is disabled. 280 * 281 * Note that rcu_read_lock() and the matching rcu_read_unlock() must 282 * occur in the same context, for example, it is illegal to invoke 283 * rcu_read_unlock() in process context if the matching rcu_read_lock() 284 * was invoked from within an irq handler. 285 * 286 * Note that rcu_read_lock() is disallowed if the CPU is either idle or 287 * offline from an RCU perspective, so check for those as well. 288 */ 289 int rcu_read_lock_held(void) 290 { 291 bool ret; 292 293 if (rcu_read_lock_held_common(&ret)) 294 return ret; 295 return lock_is_held(&rcu_lock_map); 296 } 297 EXPORT_SYMBOL_GPL(rcu_read_lock_held); 298 299 /** 300 * rcu_read_lock_bh_held() - might we be in RCU-bh read-side critical section? 301 * 302 * Check for bottom half being disabled, which covers both the 303 * CONFIG_PROVE_RCU and not cases. Note that if someone uses 304 * rcu_read_lock_bh(), but then later enables BH, lockdep (if enabled) 305 * will show the situation. This is useful for debug checks in functions 306 * that require that they be called within an RCU read-side critical 307 * section. 308 * 309 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot. 310 * 311 * Note that rcu_read_lock_bh() is disallowed if the CPU is either idle or 312 * offline from an RCU perspective, so check for those as well. 313 */ 314 int rcu_read_lock_bh_held(void) 315 { 316 bool ret; 317 318 if (rcu_read_lock_held_common(&ret)) 319 return ret; 320 return in_softirq() || irqs_disabled(); 321 } 322 EXPORT_SYMBOL_GPL(rcu_read_lock_bh_held); 323 324 int rcu_read_lock_any_held(void) 325 { 326 bool ret; 327 328 if (rcu_read_lock_held_common(&ret)) 329 return ret; 330 if (lock_is_held(&rcu_lock_map) || 331 lock_is_held(&rcu_bh_lock_map) || 332 lock_is_held(&rcu_sched_lock_map)) 333 return 1; 334 return !preemptible(); 335 } 336 EXPORT_SYMBOL_GPL(rcu_read_lock_any_held); 337 338 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */ 339 340 /** 341 * wakeme_after_rcu() - Callback function to awaken a task after grace period 342 * @head: Pointer to rcu_head member within rcu_synchronize structure 343 * 344 * Awaken the corresponding task now that a grace period has elapsed. 345 */ 346 void wakeme_after_rcu(struct rcu_head *head) 347 { 348 struct rcu_synchronize *rcu; 349 350 rcu = container_of(head, struct rcu_synchronize, head); 351 complete(&rcu->completion); 352 } 353 EXPORT_SYMBOL_GPL(wakeme_after_rcu); 354 355 void __wait_rcu_gp(bool checktiny, int n, call_rcu_func_t *crcu_array, 356 struct rcu_synchronize *rs_array) 357 { 358 int i; 359 int j; 360 361 /* Initialize and register callbacks for each crcu_array element. */ 362 for (i = 0; i < n; i++) { 363 if (checktiny && 364 (crcu_array[i] == call_rcu)) { 365 might_sleep(); 366 continue; 367 } 368 init_rcu_head_on_stack(&rs_array[i].head); 369 init_completion(&rs_array[i].completion); 370 for (j = 0; j < i; j++) 371 if (crcu_array[j] == crcu_array[i]) 372 break; 373 if (j == i) 374 (crcu_array[i])(&rs_array[i].head, wakeme_after_rcu); 375 } 376 377 /* Wait for all callbacks to be invoked. */ 378 for (i = 0; i < n; i++) { 379 if (checktiny && 380 (crcu_array[i] == call_rcu)) 381 continue; 382 for (j = 0; j < i; j++) 383 if (crcu_array[j] == crcu_array[i]) 384 break; 385 if (j == i) 386 wait_for_completion(&rs_array[i].completion); 387 destroy_rcu_head_on_stack(&rs_array[i].head); 388 } 389 } 390 EXPORT_SYMBOL_GPL(__wait_rcu_gp); 391 392 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD 393 void init_rcu_head(struct rcu_head *head) 394 { 395 debug_object_init(head, &rcuhead_debug_descr); 396 } 397 EXPORT_SYMBOL_GPL(init_rcu_head); 398 399 void destroy_rcu_head(struct rcu_head *head) 400 { 401 debug_object_free(head, &rcuhead_debug_descr); 402 } 403 EXPORT_SYMBOL_GPL(destroy_rcu_head); 404 405 static bool rcuhead_is_static_object(void *addr) 406 { 407 return true; 408 } 409 410 /** 411 * init_rcu_head_on_stack() - initialize on-stack rcu_head for debugobjects 412 * @head: pointer to rcu_head structure to be initialized 413 * 414 * This function informs debugobjects of a new rcu_head structure that 415 * has been allocated as an auto variable on the stack. This function 416 * is not required for rcu_head structures that are statically defined or 417 * that are dynamically allocated on the heap. This function has no 418 * effect for !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds. 419 */ 420 void init_rcu_head_on_stack(struct rcu_head *head) 421 { 422 debug_object_init_on_stack(head, &rcuhead_debug_descr); 423 } 424 EXPORT_SYMBOL_GPL(init_rcu_head_on_stack); 425 426 /** 427 * destroy_rcu_head_on_stack() - destroy on-stack rcu_head for debugobjects 428 * @head: pointer to rcu_head structure to be initialized 429 * 430 * This function informs debugobjects that an on-stack rcu_head structure 431 * is about to go out of scope. As with init_rcu_head_on_stack(), this 432 * function is not required for rcu_head structures that are statically 433 * defined or that are dynamically allocated on the heap. Also as with 434 * init_rcu_head_on_stack(), this function has no effect for 435 * !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds. 436 */ 437 void destroy_rcu_head_on_stack(struct rcu_head *head) 438 { 439 debug_object_free(head, &rcuhead_debug_descr); 440 } 441 EXPORT_SYMBOL_GPL(destroy_rcu_head_on_stack); 442 443 struct debug_obj_descr rcuhead_debug_descr = { 444 .name = "rcu_head", 445 .is_static_object = rcuhead_is_static_object, 446 }; 447 EXPORT_SYMBOL_GPL(rcuhead_debug_descr); 448 #endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */ 449 450 #if defined(CONFIG_TREE_RCU) || defined(CONFIG_RCU_TRACE) 451 void do_trace_rcu_torture_read(const char *rcutorturename, struct rcu_head *rhp, 452 unsigned long secs, 453 unsigned long c_old, unsigned long c) 454 { 455 trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c); 456 } 457 EXPORT_SYMBOL_GPL(do_trace_rcu_torture_read); 458 #else 459 #define do_trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c) \ 460 do { } while (0) 461 #endif 462 463 #if IS_ENABLED(CONFIG_RCU_TORTURE_TEST) || IS_MODULE(CONFIG_RCU_TORTURE_TEST) 464 /* Get rcutorture access to sched_setaffinity(). */ 465 long rcutorture_sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 466 { 467 int ret; 468 469 ret = sched_setaffinity(pid, in_mask); 470 WARN_ONCE(ret, "%s: sched_setaffinity() returned %d\n", __func__, ret); 471 return ret; 472 } 473 EXPORT_SYMBOL_GPL(rcutorture_sched_setaffinity); 474 #endif 475 476 #ifdef CONFIG_RCU_STALL_COMMON 477 int rcu_cpu_stall_ftrace_dump __read_mostly; 478 module_param(rcu_cpu_stall_ftrace_dump, int, 0644); 479 int rcu_cpu_stall_suppress __read_mostly; // !0 = suppress stall warnings. 480 EXPORT_SYMBOL_GPL(rcu_cpu_stall_suppress); 481 module_param(rcu_cpu_stall_suppress, int, 0644); 482 int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT; 483 module_param(rcu_cpu_stall_timeout, int, 0644); 484 #endif /* #ifdef CONFIG_RCU_STALL_COMMON */ 485 486 // Suppress boot-time RCU CPU stall warnings and rcutorture writer stall 487 // warnings. Also used by rcutorture even if stall warnings are excluded. 488 int rcu_cpu_stall_suppress_at_boot __read_mostly; // !0 = suppress boot stalls. 489 EXPORT_SYMBOL_GPL(rcu_cpu_stall_suppress_at_boot); 490 module_param(rcu_cpu_stall_suppress_at_boot, int, 0444); 491 492 #ifdef CONFIG_TASKS_RCU 493 494 /* 495 * Simple variant of RCU whose quiescent states are voluntary context 496 * switch, cond_resched_rcu_qs(), user-space execution, and idle. 497 * As such, grace periods can take one good long time. There are no 498 * read-side primitives similar to rcu_read_lock() and rcu_read_unlock() 499 * because this implementation is intended to get the system into a safe 500 * state for some of the manipulations involved in tracing and the like. 501 * Finally, this implementation does not support high call_rcu_tasks() 502 * rates from multiple CPUs. If this is required, per-CPU callback lists 503 * will be needed. 504 */ 505 506 /* Global list of callbacks and associated lock. */ 507 static struct rcu_head *rcu_tasks_cbs_head; 508 static struct rcu_head **rcu_tasks_cbs_tail = &rcu_tasks_cbs_head; 509 static DECLARE_WAIT_QUEUE_HEAD(rcu_tasks_cbs_wq); 510 static DEFINE_RAW_SPINLOCK(rcu_tasks_cbs_lock); 511 512 /* Track exiting tasks in order to allow them to be waited for. */ 513 DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu); 514 515 /* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */ 516 #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10) 517 static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT; 518 module_param(rcu_task_stall_timeout, int, 0644); 519 520 static struct task_struct *rcu_tasks_kthread_ptr; 521 522 /** 523 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period 524 * @rhp: structure to be used for queueing the RCU updates. 525 * @func: actual callback function to be invoked after the grace period 526 * 527 * The callback function will be invoked some time after a full grace 528 * period elapses, in other words after all currently executing RCU 529 * read-side critical sections have completed. call_rcu_tasks() assumes 530 * that the read-side critical sections end at a voluntary context 531 * switch (not a preemption!), cond_resched_rcu_qs(), entry into idle, 532 * or transition to usermode execution. As such, there are no read-side 533 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because 534 * this primitive is intended to determine that all tasks have passed 535 * through a safe state, not so much for data-strcuture synchronization. 536 * 537 * See the description of call_rcu() for more detailed information on 538 * memory ordering guarantees. 539 */ 540 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func) 541 { 542 unsigned long flags; 543 bool needwake; 544 545 rhp->next = NULL; 546 rhp->func = func; 547 raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags); 548 needwake = !rcu_tasks_cbs_head; 549 WRITE_ONCE(*rcu_tasks_cbs_tail, rhp); 550 rcu_tasks_cbs_tail = &rhp->next; 551 raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags); 552 /* We can't create the thread unless interrupts are enabled. */ 553 if (needwake && READ_ONCE(rcu_tasks_kthread_ptr)) 554 wake_up(&rcu_tasks_cbs_wq); 555 } 556 EXPORT_SYMBOL_GPL(call_rcu_tasks); 557 558 /** 559 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed. 560 * 561 * Control will return to the caller some time after a full rcu-tasks 562 * grace period has elapsed, in other words after all currently 563 * executing rcu-tasks read-side critical sections have elapsed. These 564 * read-side critical sections are delimited by calls to schedule(), 565 * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls 566 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched(). 567 * 568 * This is a very specialized primitive, intended only for a few uses in 569 * tracing and other situations requiring manipulation of function 570 * preambles and profiling hooks. The synchronize_rcu_tasks() function 571 * is not (yet) intended for heavy use from multiple CPUs. 572 * 573 * Note that this guarantee implies further memory-ordering guarantees. 574 * On systems with more than one CPU, when synchronize_rcu_tasks() returns, 575 * each CPU is guaranteed to have executed a full memory barrier since the 576 * end of its last RCU-tasks read-side critical section whose beginning 577 * preceded the call to synchronize_rcu_tasks(). In addition, each CPU 578 * having an RCU-tasks read-side critical section that extends beyond 579 * the return from synchronize_rcu_tasks() is guaranteed to have executed 580 * a full memory barrier after the beginning of synchronize_rcu_tasks() 581 * and before the beginning of that RCU-tasks read-side critical section. 582 * Note that these guarantees include CPUs that are offline, idle, or 583 * executing in user mode, as well as CPUs that are executing in the kernel. 584 * 585 * Furthermore, if CPU A invoked synchronize_rcu_tasks(), which returned 586 * to its caller on CPU B, then both CPU A and CPU B are guaranteed 587 * to have executed a full memory barrier during the execution of 588 * synchronize_rcu_tasks() -- even if CPU A and CPU B are the same CPU 589 * (but again only if the system has more than one CPU). 590 */ 591 void synchronize_rcu_tasks(void) 592 { 593 /* Complain if the scheduler has not started. */ 594 RCU_LOCKDEP_WARN(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE, 595 "synchronize_rcu_tasks called too soon"); 596 597 /* Wait for the grace period. */ 598 wait_rcu_gp(call_rcu_tasks); 599 } 600 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks); 601 602 /** 603 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks. 604 * 605 * Although the current implementation is guaranteed to wait, it is not 606 * obligated to, for example, if there are no pending callbacks. 607 */ 608 void rcu_barrier_tasks(void) 609 { 610 /* There is only one callback queue, so this is easy. ;-) */ 611 synchronize_rcu_tasks(); 612 } 613 EXPORT_SYMBOL_GPL(rcu_barrier_tasks); 614 615 /* See if tasks are still holding out, complain if so. */ 616 static void check_holdout_task(struct task_struct *t, 617 bool needreport, bool *firstreport) 618 { 619 int cpu; 620 621 if (!READ_ONCE(t->rcu_tasks_holdout) || 622 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) || 623 !READ_ONCE(t->on_rq) || 624 (IS_ENABLED(CONFIG_NO_HZ_FULL) && 625 !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) { 626 WRITE_ONCE(t->rcu_tasks_holdout, false); 627 list_del_init(&t->rcu_tasks_holdout_list); 628 put_task_struct(t); 629 return; 630 } 631 rcu_request_urgent_qs_task(t); 632 if (!needreport) 633 return; 634 if (*firstreport) { 635 pr_err("INFO: rcu_tasks detected stalls on tasks:\n"); 636 *firstreport = false; 637 } 638 cpu = task_cpu(t); 639 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n", 640 t, ".I"[is_idle_task(t)], 641 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)], 642 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout, 643 t->rcu_tasks_idle_cpu, cpu); 644 sched_show_task(t); 645 } 646 647 /* RCU-tasks kthread that detects grace periods and invokes callbacks. */ 648 static int __noreturn rcu_tasks_kthread(void *arg) 649 { 650 unsigned long flags; 651 struct task_struct *g, *t; 652 unsigned long lastreport; 653 struct rcu_head *list; 654 struct rcu_head *next; 655 LIST_HEAD(rcu_tasks_holdouts); 656 int fract; 657 658 /* Run on housekeeping CPUs by default. Sysadm can move if desired. */ 659 housekeeping_affine(current, HK_FLAG_RCU); 660 661 /* 662 * Each pass through the following loop makes one check for 663 * newly arrived callbacks, and, if there are some, waits for 664 * one RCU-tasks grace period and then invokes the callbacks. 665 * This loop is terminated by the system going down. ;-) 666 */ 667 for (;;) { 668 669 /* Pick up any new callbacks. */ 670 raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags); 671 list = rcu_tasks_cbs_head; 672 rcu_tasks_cbs_head = NULL; 673 rcu_tasks_cbs_tail = &rcu_tasks_cbs_head; 674 raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags); 675 676 /* If there were none, wait a bit and start over. */ 677 if (!list) { 678 wait_event_interruptible(rcu_tasks_cbs_wq, 679 READ_ONCE(rcu_tasks_cbs_head)); 680 if (!rcu_tasks_cbs_head) { 681 WARN_ON(signal_pending(current)); 682 schedule_timeout_interruptible(HZ/10); 683 } 684 continue; 685 } 686 687 /* 688 * Wait for all pre-existing t->on_rq and t->nvcsw 689 * transitions to complete. Invoking synchronize_rcu() 690 * suffices because all these transitions occur with 691 * interrupts disabled. Without this synchronize_rcu(), 692 * a read-side critical section that started before the 693 * grace period might be incorrectly seen as having started 694 * after the grace period. 695 * 696 * This synchronize_rcu() also dispenses with the 697 * need for a memory barrier on the first store to 698 * ->rcu_tasks_holdout, as it forces the store to happen 699 * after the beginning of the grace period. 700 */ 701 synchronize_rcu(); 702 703 /* 704 * There were callbacks, so we need to wait for an 705 * RCU-tasks grace period. Start off by scanning 706 * the task list for tasks that are not already 707 * voluntarily blocked. Mark these tasks and make 708 * a list of them in rcu_tasks_holdouts. 709 */ 710 rcu_read_lock(); 711 for_each_process_thread(g, t) { 712 if (t != current && READ_ONCE(t->on_rq) && 713 !is_idle_task(t)) { 714 get_task_struct(t); 715 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw); 716 WRITE_ONCE(t->rcu_tasks_holdout, true); 717 list_add(&t->rcu_tasks_holdout_list, 718 &rcu_tasks_holdouts); 719 } 720 } 721 rcu_read_unlock(); 722 723 /* 724 * Wait for tasks that are in the process of exiting. 725 * This does only part of the job, ensuring that all 726 * tasks that were previously exiting reach the point 727 * where they have disabled preemption, allowing the 728 * later synchronize_rcu() to finish the job. 729 */ 730 synchronize_srcu(&tasks_rcu_exit_srcu); 731 732 /* 733 * Each pass through the following loop scans the list 734 * of holdout tasks, removing any that are no longer 735 * holdouts. When the list is empty, we are done. 736 */ 737 lastreport = jiffies; 738 739 /* Start off with HZ/10 wait and slowly back off to 1 HZ wait*/ 740 fract = 10; 741 742 for (;;) { 743 bool firstreport; 744 bool needreport; 745 int rtst; 746 struct task_struct *t1; 747 748 if (list_empty(&rcu_tasks_holdouts)) 749 break; 750 751 /* Slowly back off waiting for holdouts */ 752 schedule_timeout_interruptible(HZ/fract); 753 754 if (fract > 1) 755 fract--; 756 757 rtst = READ_ONCE(rcu_task_stall_timeout); 758 needreport = rtst > 0 && 759 time_after(jiffies, lastreport + rtst); 760 if (needreport) 761 lastreport = jiffies; 762 firstreport = true; 763 WARN_ON(signal_pending(current)); 764 list_for_each_entry_safe(t, t1, &rcu_tasks_holdouts, 765 rcu_tasks_holdout_list) { 766 check_holdout_task(t, needreport, &firstreport); 767 cond_resched(); 768 } 769 } 770 771 /* 772 * Because ->on_rq and ->nvcsw are not guaranteed 773 * to have a full memory barriers prior to them in the 774 * schedule() path, memory reordering on other CPUs could 775 * cause their RCU-tasks read-side critical sections to 776 * extend past the end of the grace period. However, 777 * because these ->nvcsw updates are carried out with 778 * interrupts disabled, we can use synchronize_rcu() 779 * to force the needed ordering on all such CPUs. 780 * 781 * This synchronize_rcu() also confines all 782 * ->rcu_tasks_holdout accesses to be within the grace 783 * period, avoiding the need for memory barriers for 784 * ->rcu_tasks_holdout accesses. 785 * 786 * In addition, this synchronize_rcu() waits for exiting 787 * tasks to complete their final preempt_disable() region 788 * of execution, cleaning up after the synchronize_srcu() 789 * above. 790 */ 791 synchronize_rcu(); 792 793 /* Invoke the callbacks. */ 794 while (list) { 795 next = list->next; 796 local_bh_disable(); 797 list->func(list); 798 local_bh_enable(); 799 list = next; 800 cond_resched(); 801 } 802 /* Paranoid sleep to keep this from entering a tight loop */ 803 schedule_timeout_uninterruptible(HZ/10); 804 } 805 } 806 807 /* Spawn rcu_tasks_kthread() at core_initcall() time. */ 808 static int __init rcu_spawn_tasks_kthread(void) 809 { 810 struct task_struct *t; 811 812 t = kthread_run(rcu_tasks_kthread, NULL, "rcu_tasks_kthread"); 813 if (WARN_ONCE(IS_ERR(t), "%s: Could not start Tasks-RCU grace-period kthread, OOM is now expected behavior\n", __func__)) 814 return 0; 815 smp_mb(); /* Ensure others see full kthread. */ 816 WRITE_ONCE(rcu_tasks_kthread_ptr, t); 817 return 0; 818 } 819 core_initcall(rcu_spawn_tasks_kthread); 820 821 /* Do the srcu_read_lock() for the above synchronize_srcu(). */ 822 void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu) 823 { 824 preempt_disable(); 825 current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu); 826 preempt_enable(); 827 } 828 829 /* Do the srcu_read_unlock() for the above synchronize_srcu(). */ 830 void exit_tasks_rcu_finish(void) __releases(&tasks_rcu_exit_srcu) 831 { 832 preempt_disable(); 833 __srcu_read_unlock(&tasks_rcu_exit_srcu, current->rcu_tasks_idx); 834 preempt_enable(); 835 } 836 837 #endif /* #ifdef CONFIG_TASKS_RCU */ 838 839 #ifndef CONFIG_TINY_RCU 840 841 /* 842 * Print any non-default Tasks RCU settings. 843 */ 844 static void __init rcu_tasks_bootup_oddness(void) 845 { 846 #ifdef CONFIG_TASKS_RCU 847 if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT) 848 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout); 849 else 850 pr_info("\tTasks RCU enabled.\n"); 851 #endif /* #ifdef CONFIG_TASKS_RCU */ 852 } 853 854 #endif /* #ifndef CONFIG_TINY_RCU */ 855 856 #ifdef CONFIG_PROVE_RCU 857 858 /* 859 * Early boot self test parameters. 860 */ 861 static bool rcu_self_test; 862 module_param(rcu_self_test, bool, 0444); 863 864 static int rcu_self_test_counter; 865 866 static void test_callback(struct rcu_head *r) 867 { 868 rcu_self_test_counter++; 869 pr_info("RCU test callback executed %d\n", rcu_self_test_counter); 870 } 871 872 DEFINE_STATIC_SRCU(early_srcu); 873 874 struct early_boot_kfree_rcu { 875 struct rcu_head rh; 876 }; 877 878 static void early_boot_test_call_rcu(void) 879 { 880 static struct rcu_head head; 881 static struct rcu_head shead; 882 struct early_boot_kfree_rcu *rhp; 883 884 call_rcu(&head, test_callback); 885 if (IS_ENABLED(CONFIG_SRCU)) 886 call_srcu(&early_srcu, &shead, test_callback); 887 rhp = kmalloc(sizeof(*rhp), GFP_KERNEL); 888 if (!WARN_ON_ONCE(!rhp)) 889 kfree_rcu(rhp, rh); 890 } 891 892 void rcu_early_boot_tests(void) 893 { 894 pr_info("Running RCU self tests\n"); 895 896 if (rcu_self_test) 897 early_boot_test_call_rcu(); 898 rcu_test_sync_prims(); 899 } 900 901 static int rcu_verify_early_boot_tests(void) 902 { 903 int ret = 0; 904 int early_boot_test_counter = 0; 905 906 if (rcu_self_test) { 907 early_boot_test_counter++; 908 rcu_barrier(); 909 if (IS_ENABLED(CONFIG_SRCU)) { 910 early_boot_test_counter++; 911 srcu_barrier(&early_srcu); 912 } 913 } 914 if (rcu_self_test_counter != early_boot_test_counter) { 915 WARN_ON(1); 916 ret = -1; 917 } 918 919 return ret; 920 } 921 late_initcall(rcu_verify_early_boot_tests); 922 #else 923 void rcu_early_boot_tests(void) {} 924 #endif /* CONFIG_PROVE_RCU */ 925 926 #ifndef CONFIG_TINY_RCU 927 928 /* 929 * Print any significant non-default boot-time settings. 930 */ 931 void __init rcupdate_announce_bootup_oddness(void) 932 { 933 if (rcu_normal) 934 pr_info("\tNo expedited grace period (rcu_normal).\n"); 935 else if (rcu_normal_after_boot) 936 pr_info("\tNo expedited grace period (rcu_normal_after_boot).\n"); 937 else if (rcu_expedited) 938 pr_info("\tAll grace periods are expedited (rcu_expedited).\n"); 939 if (rcu_cpu_stall_suppress) 940 pr_info("\tRCU CPU stall warnings suppressed (rcu_cpu_stall_suppress).\n"); 941 if (rcu_cpu_stall_timeout != CONFIG_RCU_CPU_STALL_TIMEOUT) 942 pr_info("\tRCU CPU stall warnings timeout set to %d (rcu_cpu_stall_timeout).\n", rcu_cpu_stall_timeout); 943 rcu_tasks_bootup_oddness(); 944 } 945 946 #endif /* #ifndef CONFIG_TINY_RCU */ 947