Loading...
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#include <linux/irq_work.h>
45#include <linux/rcupdate_trace.h>
46
47#define CREATE_TRACE_POINTS
48
49#include "rcu.h"
50
51#ifdef MODULE_PARAM_PREFIX
52#undef MODULE_PARAM_PREFIX
53#endif
54#define MODULE_PARAM_PREFIX "rcupdate."
55
56#ifndef CONFIG_TINY_RCU
57module_param(rcu_expedited, int, 0444);
58module_param(rcu_normal, int, 0444);
59static int rcu_normal_after_boot = IS_ENABLED(CONFIG_PREEMPT_RT);
60#if !defined(CONFIG_PREEMPT_RT) || defined(CONFIG_NO_HZ_FULL)
61module_param(rcu_normal_after_boot, int, 0444);
62#endif
63#endif /* #ifndef CONFIG_TINY_RCU */
64
65#ifdef CONFIG_DEBUG_LOCK_ALLOC
66/**
67 * rcu_read_lock_held_common() - might we be in RCU-sched read-side critical section?
68 * @ret: Best guess answer if lockdep cannot be relied on
69 *
70 * Returns true if lockdep must be ignored, in which case ``*ret`` contains
71 * the best guess described below. Otherwise returns false, in which
72 * case ``*ret`` tells the caller nothing and the caller should instead
73 * consult lockdep.
74 *
75 * If CONFIG_DEBUG_LOCK_ALLOC is selected, set ``*ret`` to nonzero iff in an
76 * RCU-sched read-side critical section. In absence of
77 * CONFIG_DEBUG_LOCK_ALLOC, this assumes we are in an RCU-sched read-side
78 * critical section unless it can prove otherwise. Note that disabling
79 * of preemption (including disabling irqs) counts as an RCU-sched
80 * read-side critical section. This is useful for debug checks in functions
81 * that required that they be called within an RCU-sched read-side
82 * critical section.
83 *
84 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot
85 * and while lockdep is disabled.
86 *
87 * Note that if the CPU is in the idle loop from an RCU point of view (ie:
88 * that we are in the section between ct_idle_enter() and ct_idle_exit())
89 * then rcu_read_lock_held() sets ``*ret`` to false even if the CPU did an
90 * rcu_read_lock(). The reason for this is that RCU ignores CPUs that are
91 * in such a section, considering these as in extended quiescent state,
92 * so such a CPU is effectively never in an RCU read-side critical section
93 * regardless of what RCU primitives it invokes. This state of affairs is
94 * required --- we need to keep an RCU-free window in idle where the CPU may
95 * possibly enter into low power mode. This way we can notice an extended
96 * quiescent state to other CPUs that started a grace period. Otherwise
97 * we would delay any grace period as long as we run in the idle task.
98 *
99 * Similarly, we avoid claiming an RCU read lock held if the current
100 * CPU is offline.
101 */
102static bool rcu_read_lock_held_common(bool *ret)
103{
104 if (!debug_lockdep_rcu_enabled()) {
105 *ret = true;
106 return true;
107 }
108 if (!rcu_is_watching()) {
109 *ret = false;
110 return true;
111 }
112 if (!rcu_lockdep_current_cpu_online()) {
113 *ret = false;
114 return true;
115 }
116 return false;
117}
118
119int rcu_read_lock_sched_held(void)
120{
121 bool ret;
122
123 if (rcu_read_lock_held_common(&ret))
124 return ret;
125 return lock_is_held(&rcu_sched_lock_map) || !preemptible();
126}
127EXPORT_SYMBOL(rcu_read_lock_sched_held);
128#endif
129
130#ifndef CONFIG_TINY_RCU
131
132/*
133 * Should expedited grace-period primitives always fall back to their
134 * non-expedited counterparts? Intended for use within RCU. Note
135 * that if the user specifies both rcu_expedited and rcu_normal, then
136 * rcu_normal wins. (Except during the time period during boot from
137 * when the first task is spawned until the rcu_set_runtime_mode()
138 * core_initcall() is invoked, at which point everything is expedited.)
139 */
140bool rcu_gp_is_normal(void)
141{
142 return READ_ONCE(rcu_normal) &&
143 rcu_scheduler_active != RCU_SCHEDULER_INIT;
144}
145EXPORT_SYMBOL_GPL(rcu_gp_is_normal);
146
147static atomic_t rcu_expedited_nesting = ATOMIC_INIT(1);
148
149/*
150 * Should normal grace-period primitives be expedited? Intended for
151 * use within RCU. Note that this function takes the rcu_expedited
152 * sysfs/boot variable and rcu_scheduler_active into account as well
153 * as the rcu_expedite_gp() nesting. So looping on rcu_unexpedite_gp()
154 * until rcu_gp_is_expedited() returns false is a -really- bad idea.
155 */
156bool rcu_gp_is_expedited(void)
157{
158 return rcu_expedited || atomic_read(&rcu_expedited_nesting);
159}
160EXPORT_SYMBOL_GPL(rcu_gp_is_expedited);
161
162/**
163 * rcu_expedite_gp - Expedite future RCU grace periods
164 *
165 * After a call to this function, future calls to synchronize_rcu() and
166 * friends act as the corresponding synchronize_rcu_expedited() function
167 * had instead been called.
168 */
169void rcu_expedite_gp(void)
170{
171 atomic_inc(&rcu_expedited_nesting);
172}
173EXPORT_SYMBOL_GPL(rcu_expedite_gp);
174
175/**
176 * rcu_unexpedite_gp - Cancel prior rcu_expedite_gp() invocation
177 *
178 * Undo a prior call to rcu_expedite_gp(). If all prior calls to
179 * rcu_expedite_gp() are undone by a subsequent call to rcu_unexpedite_gp(),
180 * and if the rcu_expedited sysfs/boot parameter is not set, then all
181 * subsequent calls to synchronize_rcu() and friends will return to
182 * their normal non-expedited behavior.
183 */
184void rcu_unexpedite_gp(void)
185{
186 atomic_dec(&rcu_expedited_nesting);
187}
188EXPORT_SYMBOL_GPL(rcu_unexpedite_gp);
189
190static bool rcu_boot_ended __read_mostly;
191
192/*
193 * Inform RCU of the end of the in-kernel boot sequence.
194 */
195void rcu_end_inkernel_boot(void)
196{
197 rcu_unexpedite_gp();
198 if (rcu_normal_after_boot)
199 WRITE_ONCE(rcu_normal, 1);
200 rcu_boot_ended = true;
201}
202
203/*
204 * Let rcutorture know when it is OK to turn it up to eleven.
205 */
206bool rcu_inkernel_boot_has_ended(void)
207{
208 return rcu_boot_ended;
209}
210EXPORT_SYMBOL_GPL(rcu_inkernel_boot_has_ended);
211
212#endif /* #ifndef CONFIG_TINY_RCU */
213
214/*
215 * Test each non-SRCU synchronous grace-period wait API. This is
216 * useful just after a change in mode for these primitives, and
217 * during early boot.
218 */
219void rcu_test_sync_prims(void)
220{
221 if (!IS_ENABLED(CONFIG_PROVE_RCU))
222 return;
223 synchronize_rcu();
224 synchronize_rcu_expedited();
225}
226
227#if !defined(CONFIG_TINY_RCU)
228
229/*
230 * Switch to run-time mode once RCU has fully initialized.
231 */
232static int __init rcu_set_runtime_mode(void)
233{
234 rcu_test_sync_prims();
235 rcu_scheduler_active = RCU_SCHEDULER_RUNNING;
236 kfree_rcu_scheduler_running();
237 rcu_test_sync_prims();
238 return 0;
239}
240core_initcall(rcu_set_runtime_mode);
241
242#endif /* #if !defined(CONFIG_TINY_RCU) */
243
244#ifdef CONFIG_DEBUG_LOCK_ALLOC
245static struct lock_class_key rcu_lock_key;
246struct lockdep_map rcu_lock_map = {
247 .name = "rcu_read_lock",
248 .key = &rcu_lock_key,
249 .wait_type_outer = LD_WAIT_FREE,
250 .wait_type_inner = LD_WAIT_CONFIG, /* PREEMPT_RT implies PREEMPT_RCU */
251};
252EXPORT_SYMBOL_GPL(rcu_lock_map);
253
254static struct lock_class_key rcu_bh_lock_key;
255struct lockdep_map rcu_bh_lock_map = {
256 .name = "rcu_read_lock_bh",
257 .key = &rcu_bh_lock_key,
258 .wait_type_outer = LD_WAIT_FREE,
259 .wait_type_inner = LD_WAIT_CONFIG, /* PREEMPT_RT makes BH preemptible. */
260};
261EXPORT_SYMBOL_GPL(rcu_bh_lock_map);
262
263static struct lock_class_key rcu_sched_lock_key;
264struct lockdep_map rcu_sched_lock_map = {
265 .name = "rcu_read_lock_sched",
266 .key = &rcu_sched_lock_key,
267 .wait_type_outer = LD_WAIT_FREE,
268 .wait_type_inner = LD_WAIT_SPIN,
269};
270EXPORT_SYMBOL_GPL(rcu_sched_lock_map);
271
272// Tell lockdep when RCU callbacks are being invoked.
273static struct lock_class_key rcu_callback_key;
274struct lockdep_map rcu_callback_map =
275 STATIC_LOCKDEP_MAP_INIT("rcu_callback", &rcu_callback_key);
276EXPORT_SYMBOL_GPL(rcu_callback_map);
277
278noinstr int notrace debug_lockdep_rcu_enabled(void)
279{
280 return rcu_scheduler_active != RCU_SCHEDULER_INACTIVE && READ_ONCE(debug_locks) &&
281 current->lockdep_recursion == 0;
282}
283EXPORT_SYMBOL_GPL(debug_lockdep_rcu_enabled);
284
285/**
286 * rcu_read_lock_held() - might we be in RCU read-side critical section?
287 *
288 * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an RCU
289 * read-side critical section. In absence of CONFIG_DEBUG_LOCK_ALLOC,
290 * this assumes we are in an RCU read-side critical section unless it can
291 * prove otherwise. This is useful for debug checks in functions that
292 * require that they be called within an RCU read-side critical section.
293 *
294 * Checks debug_lockdep_rcu_enabled() to prevent false positives during boot
295 * and while lockdep is disabled.
296 *
297 * Note that rcu_read_lock() and the matching rcu_read_unlock() must
298 * occur in the same context, for example, it is illegal to invoke
299 * rcu_read_unlock() in process context if the matching rcu_read_lock()
300 * was invoked from within an irq handler.
301 *
302 * Note that rcu_read_lock() is disallowed if the CPU is either idle or
303 * offline from an RCU perspective, so check for those as well.
304 */
305int rcu_read_lock_held(void)
306{
307 bool ret;
308
309 if (rcu_read_lock_held_common(&ret))
310 return ret;
311 return lock_is_held(&rcu_lock_map);
312}
313EXPORT_SYMBOL_GPL(rcu_read_lock_held);
314
315/**
316 * rcu_read_lock_bh_held() - might we be in RCU-bh read-side critical section?
317 *
318 * Check for bottom half being disabled, which covers both the
319 * CONFIG_PROVE_RCU and not cases. Note that if someone uses
320 * rcu_read_lock_bh(), but then later enables BH, lockdep (if enabled)
321 * will show the situation. This is useful for debug checks in functions
322 * that require that they be called within an RCU read-side critical
323 * section.
324 *
325 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot.
326 *
327 * Note that rcu_read_lock_bh() is disallowed if the CPU is either idle or
328 * offline from an RCU perspective, so check for those as well.
329 */
330int rcu_read_lock_bh_held(void)
331{
332 bool ret;
333
334 if (rcu_read_lock_held_common(&ret))
335 return ret;
336 return in_softirq() || irqs_disabled();
337}
338EXPORT_SYMBOL_GPL(rcu_read_lock_bh_held);
339
340int rcu_read_lock_any_held(void)
341{
342 bool ret;
343
344 if (rcu_read_lock_held_common(&ret))
345 return ret;
346 if (lock_is_held(&rcu_lock_map) ||
347 lock_is_held(&rcu_bh_lock_map) ||
348 lock_is_held(&rcu_sched_lock_map))
349 return 1;
350 return !preemptible();
351}
352EXPORT_SYMBOL_GPL(rcu_read_lock_any_held);
353
354#endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
355
356/**
357 * wakeme_after_rcu() - Callback function to awaken a task after grace period
358 * @head: Pointer to rcu_head member within rcu_synchronize structure
359 *
360 * Awaken the corresponding task now that a grace period has elapsed.
361 */
362void wakeme_after_rcu(struct rcu_head *head)
363{
364 struct rcu_synchronize *rcu;
365
366 rcu = container_of(head, struct rcu_synchronize, head);
367 complete(&rcu->completion);
368}
369EXPORT_SYMBOL_GPL(wakeme_after_rcu);
370
371void __wait_rcu_gp(bool checktiny, int n, call_rcu_func_t *crcu_array,
372 struct rcu_synchronize *rs_array)
373{
374 int i;
375 int j;
376
377 /* Initialize and register callbacks for each crcu_array element. */
378 for (i = 0; i < n; i++) {
379 if (checktiny &&
380 (crcu_array[i] == call_rcu)) {
381 might_sleep();
382 continue;
383 }
384 for (j = 0; j < i; j++)
385 if (crcu_array[j] == crcu_array[i])
386 break;
387 if (j == i) {
388 init_rcu_head_on_stack(&rs_array[i].head);
389 init_completion(&rs_array[i].completion);
390 (crcu_array[i])(&rs_array[i].head, wakeme_after_rcu);
391 }
392 }
393
394 /* Wait for all callbacks to be invoked. */
395 for (i = 0; i < n; i++) {
396 if (checktiny &&
397 (crcu_array[i] == call_rcu))
398 continue;
399 for (j = 0; j < i; j++)
400 if (crcu_array[j] == crcu_array[i])
401 break;
402 if (j == i) {
403 wait_for_completion(&rs_array[i].completion);
404 destroy_rcu_head_on_stack(&rs_array[i].head);
405 }
406 }
407}
408EXPORT_SYMBOL_GPL(__wait_rcu_gp);
409
410void finish_rcuwait(struct rcuwait *w)
411{
412 rcu_assign_pointer(w->task, NULL);
413 __set_current_state(TASK_RUNNING);
414}
415EXPORT_SYMBOL_GPL(finish_rcuwait);
416
417#ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
418void init_rcu_head(struct rcu_head *head)
419{
420 debug_object_init(head, &rcuhead_debug_descr);
421}
422EXPORT_SYMBOL_GPL(init_rcu_head);
423
424void destroy_rcu_head(struct rcu_head *head)
425{
426 debug_object_free(head, &rcuhead_debug_descr);
427}
428EXPORT_SYMBOL_GPL(destroy_rcu_head);
429
430static bool rcuhead_is_static_object(void *addr)
431{
432 return true;
433}
434
435/**
436 * init_rcu_head_on_stack() - initialize on-stack rcu_head for debugobjects
437 * @head: pointer to rcu_head structure to be initialized
438 *
439 * This function informs debugobjects of a new rcu_head structure that
440 * has been allocated as an auto variable on the stack. This function
441 * is not required for rcu_head structures that are statically defined or
442 * that are dynamically allocated on the heap. This function has no
443 * effect for !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
444 */
445void init_rcu_head_on_stack(struct rcu_head *head)
446{
447 debug_object_init_on_stack(head, &rcuhead_debug_descr);
448}
449EXPORT_SYMBOL_GPL(init_rcu_head_on_stack);
450
451/**
452 * destroy_rcu_head_on_stack() - destroy on-stack rcu_head for debugobjects
453 * @head: pointer to rcu_head structure to be initialized
454 *
455 * This function informs debugobjects that an on-stack rcu_head structure
456 * is about to go out of scope. As with init_rcu_head_on_stack(), this
457 * function is not required for rcu_head structures that are statically
458 * defined or that are dynamically allocated on the heap. Also as with
459 * init_rcu_head_on_stack(), this function has no effect for
460 * !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
461 */
462void destroy_rcu_head_on_stack(struct rcu_head *head)
463{
464 debug_object_free(head, &rcuhead_debug_descr);
465}
466EXPORT_SYMBOL_GPL(destroy_rcu_head_on_stack);
467
468const struct debug_obj_descr rcuhead_debug_descr = {
469 .name = "rcu_head",
470 .is_static_object = rcuhead_is_static_object,
471};
472EXPORT_SYMBOL_GPL(rcuhead_debug_descr);
473#endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
474
475#if defined(CONFIG_TREE_RCU) || defined(CONFIG_RCU_TRACE)
476void do_trace_rcu_torture_read(const char *rcutorturename, struct rcu_head *rhp,
477 unsigned long secs,
478 unsigned long c_old, unsigned long c)
479{
480 trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c);
481}
482EXPORT_SYMBOL_GPL(do_trace_rcu_torture_read);
483#else
484#define do_trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c) \
485 do { } while (0)
486#endif
487
488#if IS_ENABLED(CONFIG_RCU_TORTURE_TEST) || IS_MODULE(CONFIG_RCU_TORTURE_TEST)
489/* Get rcutorture access to sched_setaffinity(). */
490long rcutorture_sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
491{
492 int ret;
493
494 ret = sched_setaffinity(pid, in_mask);
495 WARN_ONCE(ret, "%s: sched_setaffinity() returned %d\n", __func__, ret);
496 return ret;
497}
498EXPORT_SYMBOL_GPL(rcutorture_sched_setaffinity);
499#endif
500
501#ifdef CONFIG_RCU_STALL_COMMON
502int rcu_cpu_stall_ftrace_dump __read_mostly;
503module_param(rcu_cpu_stall_ftrace_dump, int, 0644);
504int rcu_cpu_stall_suppress __read_mostly; // !0 = suppress stall warnings.
505EXPORT_SYMBOL_GPL(rcu_cpu_stall_suppress);
506module_param(rcu_cpu_stall_suppress, int, 0644);
507int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
508module_param(rcu_cpu_stall_timeout, int, 0644);
509int rcu_exp_cpu_stall_timeout __read_mostly = CONFIG_RCU_EXP_CPU_STALL_TIMEOUT;
510module_param(rcu_exp_cpu_stall_timeout, int, 0644);
511#endif /* #ifdef CONFIG_RCU_STALL_COMMON */
512
513// Suppress boot-time RCU CPU stall warnings and rcutorture writer stall
514// warnings. Also used by rcutorture even if stall warnings are excluded.
515int rcu_cpu_stall_suppress_at_boot __read_mostly; // !0 = suppress boot stalls.
516EXPORT_SYMBOL_GPL(rcu_cpu_stall_suppress_at_boot);
517module_param(rcu_cpu_stall_suppress_at_boot, int, 0444);
518
519/**
520 * get_completed_synchronize_rcu - Return a pre-completed polled state cookie
521 *
522 * Returns a value that will always be treated by functions like
523 * poll_state_synchronize_rcu() as a cookie whose grace period has already
524 * completed.
525 */
526unsigned long get_completed_synchronize_rcu(void)
527{
528 return RCU_GET_STATE_COMPLETED;
529}
530EXPORT_SYMBOL_GPL(get_completed_synchronize_rcu);
531
532#ifdef CONFIG_PROVE_RCU
533
534/*
535 * Early boot self test parameters.
536 */
537static bool rcu_self_test;
538module_param(rcu_self_test, bool, 0444);
539
540static int rcu_self_test_counter;
541
542static void test_callback(struct rcu_head *r)
543{
544 rcu_self_test_counter++;
545 pr_info("RCU test callback executed %d\n", rcu_self_test_counter);
546}
547
548DEFINE_STATIC_SRCU(early_srcu);
549static unsigned long early_srcu_cookie;
550
551struct early_boot_kfree_rcu {
552 struct rcu_head rh;
553};
554
555static void early_boot_test_call_rcu(void)
556{
557 static struct rcu_head head;
558 static struct rcu_head shead;
559 struct early_boot_kfree_rcu *rhp;
560
561 call_rcu(&head, test_callback);
562 early_srcu_cookie = start_poll_synchronize_srcu(&early_srcu);
563 call_srcu(&early_srcu, &shead, test_callback);
564 rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
565 if (!WARN_ON_ONCE(!rhp))
566 kfree_rcu(rhp, rh);
567}
568
569void rcu_early_boot_tests(void)
570{
571 pr_info("Running RCU self tests\n");
572
573 if (rcu_self_test)
574 early_boot_test_call_rcu();
575 rcu_test_sync_prims();
576}
577
578static int rcu_verify_early_boot_tests(void)
579{
580 int ret = 0;
581 int early_boot_test_counter = 0;
582
583 if (rcu_self_test) {
584 early_boot_test_counter++;
585 rcu_barrier();
586 early_boot_test_counter++;
587 srcu_barrier(&early_srcu);
588 WARN_ON_ONCE(!poll_state_synchronize_srcu(&early_srcu, early_srcu_cookie));
589 }
590 if (rcu_self_test_counter != early_boot_test_counter) {
591 WARN_ON(1);
592 ret = -1;
593 }
594
595 return ret;
596}
597late_initcall(rcu_verify_early_boot_tests);
598#else
599void rcu_early_boot_tests(void) {}
600#endif /* CONFIG_PROVE_RCU */
601
602#include "tasks.h"
603
604#ifndef CONFIG_TINY_RCU
605
606/*
607 * Print any significant non-default boot-time settings.
608 */
609void __init rcupdate_announce_bootup_oddness(void)
610{
611 if (rcu_normal)
612 pr_info("\tNo expedited grace period (rcu_normal).\n");
613 else if (rcu_normal_after_boot)
614 pr_info("\tNo expedited grace period (rcu_normal_after_boot).\n");
615 else if (rcu_expedited)
616 pr_info("\tAll grace periods are expedited (rcu_expedited).\n");
617 if (rcu_cpu_stall_suppress)
618 pr_info("\tRCU CPU stall warnings suppressed (rcu_cpu_stall_suppress).\n");
619 if (rcu_cpu_stall_timeout != CONFIG_RCU_CPU_STALL_TIMEOUT)
620 pr_info("\tRCU CPU stall warnings timeout set to %d (rcu_cpu_stall_timeout).\n", rcu_cpu_stall_timeout);
621 rcu_tasks_bootup_oddness();
622}
623
624#endif /* #ifndef CONFIG_TINY_RCU */
1/*
2 * Read-Copy Update mechanism for mutual exclusion
3 *
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
17 *
18 * Copyright IBM Corporation, 2001
19 *
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 *
23 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
24 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
25 * Papers:
26 * http://www.rdrop.com/users/paulmck/paper/rclockpdcsproof.pdf
27 * http://lse.sourceforge.net/locking/rclock_OLS.2001.05.01c.sc.pdf (OLS2001)
28 *
29 * For detailed explanation of Read-Copy Update mechanism see -
30 * http://lse.sourceforge.net/locking/rcupdate.html
31 *
32 */
33#include <linux/types.h>
34#include <linux/kernel.h>
35#include <linux/init.h>
36#include <linux/spinlock.h>
37#include <linux/smp.h>
38#include <linux/interrupt.h>
39#include <linux/sched/signal.h>
40#include <linux/sched/debug.h>
41#include <linux/atomic.h>
42#include <linux/bitops.h>
43#include <linux/percpu.h>
44#include <linux/notifier.h>
45#include <linux/cpu.h>
46#include <linux/mutex.h>
47#include <linux/export.h>
48#include <linux/hardirq.h>
49#include <linux/delay.h>
50#include <linux/moduleparam.h>
51#include <linux/kthread.h>
52#include <linux/tick.h>
53#include <linux/rcupdate_wait.h>
54#include <linux/sched/isolation.h>
55
56#define CREATE_TRACE_POINTS
57
58#include "rcu.h"
59
60#ifdef MODULE_PARAM_PREFIX
61#undef MODULE_PARAM_PREFIX
62#endif
63#define MODULE_PARAM_PREFIX "rcupdate."
64
65#ifndef CONFIG_TINY_RCU
66extern int rcu_expedited; /* from sysctl */
67module_param(rcu_expedited, int, 0);
68extern int rcu_normal; /* from sysctl */
69module_param(rcu_normal, int, 0);
70static int rcu_normal_after_boot;
71module_param(rcu_normal_after_boot, int, 0);
72#endif /* #ifndef CONFIG_TINY_RCU */
73
74#ifdef CONFIG_DEBUG_LOCK_ALLOC
75/**
76 * rcu_read_lock_sched_held() - might we be in RCU-sched read-side critical section?
77 *
78 * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an
79 * RCU-sched read-side critical section. In absence of
80 * CONFIG_DEBUG_LOCK_ALLOC, this assumes we are in an RCU-sched read-side
81 * critical section unless it can prove otherwise. Note that disabling
82 * of preemption (including disabling irqs) counts as an RCU-sched
83 * read-side critical section. This is useful for debug checks in functions
84 * that required that they be called within an RCU-sched read-side
85 * critical section.
86 *
87 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot
88 * and while lockdep is disabled.
89 *
90 * Note that if the CPU is in the idle loop from an RCU point of
91 * view (ie: that we are in the section between rcu_idle_enter() and
92 * rcu_idle_exit()) then rcu_read_lock_held() returns false even if the CPU
93 * did an rcu_read_lock(). The reason for this is that RCU ignores CPUs
94 * that are in such a section, considering these as in extended quiescent
95 * state, so such a CPU is effectively never in an RCU read-side critical
96 * section regardless of what RCU primitives it invokes. This state of
97 * affairs is required --- we need to keep an RCU-free window in idle
98 * where the CPU may possibly enter into low power mode. This way we can
99 * notice an extended quiescent state to other CPUs that started a grace
100 * period. Otherwise we would delay any grace period as long as we run in
101 * the idle task.
102 *
103 * Similarly, we avoid claiming an SRCU read lock held if the current
104 * CPU is offline.
105 */
106int rcu_read_lock_sched_held(void)
107{
108 int lockdep_opinion = 0;
109
110 if (!debug_lockdep_rcu_enabled())
111 return 1;
112 if (!rcu_is_watching())
113 return 0;
114 if (!rcu_lockdep_current_cpu_online())
115 return 0;
116 if (debug_locks)
117 lockdep_opinion = lock_is_held(&rcu_sched_lock_map);
118 return lockdep_opinion || !preemptible();
119}
120EXPORT_SYMBOL(rcu_read_lock_sched_held);
121#endif
122
123#ifndef CONFIG_TINY_RCU
124
125/*
126 * Should expedited grace-period primitives always fall back to their
127 * non-expedited counterparts? Intended for use within RCU. Note
128 * that if the user specifies both rcu_expedited and rcu_normal, then
129 * rcu_normal wins. (Except during the time period during boot from
130 * when the first task is spawned until the rcu_set_runtime_mode()
131 * core_initcall() is invoked, at which point everything is expedited.)
132 */
133bool rcu_gp_is_normal(void)
134{
135 return READ_ONCE(rcu_normal) &&
136 rcu_scheduler_active != RCU_SCHEDULER_INIT;
137}
138EXPORT_SYMBOL_GPL(rcu_gp_is_normal);
139
140static atomic_t rcu_expedited_nesting = ATOMIC_INIT(1);
141
142/*
143 * Should normal grace-period primitives be expedited? Intended for
144 * use within RCU. Note that this function takes the rcu_expedited
145 * sysfs/boot variable and rcu_scheduler_active into account as well
146 * as the rcu_expedite_gp() nesting. So looping on rcu_unexpedite_gp()
147 * until rcu_gp_is_expedited() returns false is a -really- bad idea.
148 */
149bool rcu_gp_is_expedited(void)
150{
151 return rcu_expedited || atomic_read(&rcu_expedited_nesting) ||
152 rcu_scheduler_active == RCU_SCHEDULER_INIT;
153}
154EXPORT_SYMBOL_GPL(rcu_gp_is_expedited);
155
156/**
157 * rcu_expedite_gp - Expedite future RCU grace periods
158 *
159 * After a call to this function, future calls to synchronize_rcu() and
160 * friends act as the corresponding synchronize_rcu_expedited() function
161 * had instead been called.
162 */
163void rcu_expedite_gp(void)
164{
165 atomic_inc(&rcu_expedited_nesting);
166}
167EXPORT_SYMBOL_GPL(rcu_expedite_gp);
168
169/**
170 * rcu_unexpedite_gp - Cancel prior rcu_expedite_gp() invocation
171 *
172 * Undo a prior call to rcu_expedite_gp(). If all prior calls to
173 * rcu_expedite_gp() are undone by a subsequent call to rcu_unexpedite_gp(),
174 * and if the rcu_expedited sysfs/boot parameter is not set, then all
175 * subsequent calls to synchronize_rcu() and friends will return to
176 * their normal non-expedited behavior.
177 */
178void rcu_unexpedite_gp(void)
179{
180 atomic_dec(&rcu_expedited_nesting);
181}
182EXPORT_SYMBOL_GPL(rcu_unexpedite_gp);
183
184/*
185 * Inform RCU of the end of the in-kernel boot sequence.
186 */
187void rcu_end_inkernel_boot(void)
188{
189 rcu_unexpedite_gp();
190 if (rcu_normal_after_boot)
191 WRITE_ONCE(rcu_normal, 1);
192}
193
194#endif /* #ifndef CONFIG_TINY_RCU */
195
196/*
197 * Test each non-SRCU synchronous grace-period wait API. This is
198 * useful just after a change in mode for these primitives, and
199 * during early boot.
200 */
201void rcu_test_sync_prims(void)
202{
203 if (!IS_ENABLED(CONFIG_PROVE_RCU))
204 return;
205 synchronize_rcu();
206 synchronize_rcu_bh();
207 synchronize_sched();
208 synchronize_rcu_expedited();
209 synchronize_rcu_bh_expedited();
210 synchronize_sched_expedited();
211}
212
213#if !defined(CONFIG_TINY_RCU) || defined(CONFIG_SRCU)
214
215/*
216 * Switch to run-time mode once RCU has fully initialized.
217 */
218static int __init rcu_set_runtime_mode(void)
219{
220 rcu_test_sync_prims();
221 rcu_scheduler_active = RCU_SCHEDULER_RUNNING;
222 rcu_test_sync_prims();
223 return 0;
224}
225core_initcall(rcu_set_runtime_mode);
226
227#endif /* #if !defined(CONFIG_TINY_RCU) || defined(CONFIG_SRCU) */
228
229#ifdef CONFIG_PREEMPT_RCU
230
231/*
232 * Preemptible RCU implementation for rcu_read_lock().
233 * Just increment ->rcu_read_lock_nesting, shared state will be updated
234 * if we block.
235 */
236void __rcu_read_lock(void)
237{
238 current->rcu_read_lock_nesting++;
239 barrier(); /* critical section after entry code. */
240}
241EXPORT_SYMBOL_GPL(__rcu_read_lock);
242
243/*
244 * Preemptible RCU implementation for rcu_read_unlock().
245 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
246 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
247 * invoke rcu_read_unlock_special() to clean up after a context switch
248 * in an RCU read-side critical section and other special cases.
249 */
250void __rcu_read_unlock(void)
251{
252 struct task_struct *t = current;
253
254 if (t->rcu_read_lock_nesting != 1) {
255 --t->rcu_read_lock_nesting;
256 } else {
257 barrier(); /* critical section before exit code. */
258 t->rcu_read_lock_nesting = INT_MIN;
259 barrier(); /* assign before ->rcu_read_unlock_special load */
260 if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
261 rcu_read_unlock_special(t);
262 barrier(); /* ->rcu_read_unlock_special load before assign */
263 t->rcu_read_lock_nesting = 0;
264 }
265#ifdef CONFIG_PROVE_LOCKING
266 {
267 int rrln = READ_ONCE(t->rcu_read_lock_nesting);
268
269 WARN_ON_ONCE(rrln < 0 && rrln > INT_MIN / 2);
270 }
271#endif /* #ifdef CONFIG_PROVE_LOCKING */
272}
273EXPORT_SYMBOL_GPL(__rcu_read_unlock);
274
275#endif /* #ifdef CONFIG_PREEMPT_RCU */
276
277#ifdef CONFIG_DEBUG_LOCK_ALLOC
278static struct lock_class_key rcu_lock_key;
279struct lockdep_map rcu_lock_map =
280 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock", &rcu_lock_key);
281EXPORT_SYMBOL_GPL(rcu_lock_map);
282
283static struct lock_class_key rcu_bh_lock_key;
284struct lockdep_map rcu_bh_lock_map =
285 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_bh", &rcu_bh_lock_key);
286EXPORT_SYMBOL_GPL(rcu_bh_lock_map);
287
288static struct lock_class_key rcu_sched_lock_key;
289struct lockdep_map rcu_sched_lock_map =
290 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_sched", &rcu_sched_lock_key);
291EXPORT_SYMBOL_GPL(rcu_sched_lock_map);
292
293static struct lock_class_key rcu_callback_key;
294struct lockdep_map rcu_callback_map =
295 STATIC_LOCKDEP_MAP_INIT("rcu_callback", &rcu_callback_key);
296EXPORT_SYMBOL_GPL(rcu_callback_map);
297
298int notrace debug_lockdep_rcu_enabled(void)
299{
300 return rcu_scheduler_active != RCU_SCHEDULER_INACTIVE && debug_locks &&
301 current->lockdep_recursion == 0;
302}
303EXPORT_SYMBOL_GPL(debug_lockdep_rcu_enabled);
304
305/**
306 * rcu_read_lock_held() - might we be in RCU read-side critical section?
307 *
308 * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an RCU
309 * read-side critical section. In absence of CONFIG_DEBUG_LOCK_ALLOC,
310 * this assumes we are in an RCU read-side critical section unless it can
311 * prove otherwise. This is useful for debug checks in functions that
312 * require that they be called within an RCU read-side critical section.
313 *
314 * Checks debug_lockdep_rcu_enabled() to prevent false positives during boot
315 * and while lockdep is disabled.
316 *
317 * Note that rcu_read_lock() and the matching rcu_read_unlock() must
318 * occur in the same context, for example, it is illegal to invoke
319 * rcu_read_unlock() in process context if the matching rcu_read_lock()
320 * was invoked from within an irq handler.
321 *
322 * Note that rcu_read_lock() is disallowed if the CPU is either idle or
323 * offline from an RCU perspective, so check for those as well.
324 */
325int rcu_read_lock_held(void)
326{
327 if (!debug_lockdep_rcu_enabled())
328 return 1;
329 if (!rcu_is_watching())
330 return 0;
331 if (!rcu_lockdep_current_cpu_online())
332 return 0;
333 return lock_is_held(&rcu_lock_map);
334}
335EXPORT_SYMBOL_GPL(rcu_read_lock_held);
336
337/**
338 * rcu_read_lock_bh_held() - might we be in RCU-bh read-side critical section?
339 *
340 * Check for bottom half being disabled, which covers both the
341 * CONFIG_PROVE_RCU and not cases. Note that if someone uses
342 * rcu_read_lock_bh(), but then later enables BH, lockdep (if enabled)
343 * will show the situation. This is useful for debug checks in functions
344 * that require that they be called within an RCU read-side critical
345 * section.
346 *
347 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot.
348 *
349 * Note that rcu_read_lock() is disallowed if the CPU is either idle or
350 * offline from an RCU perspective, so check for those as well.
351 */
352int rcu_read_lock_bh_held(void)
353{
354 if (!debug_lockdep_rcu_enabled())
355 return 1;
356 if (!rcu_is_watching())
357 return 0;
358 if (!rcu_lockdep_current_cpu_online())
359 return 0;
360 return in_softirq() || irqs_disabled();
361}
362EXPORT_SYMBOL_GPL(rcu_read_lock_bh_held);
363
364#endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
365
366/**
367 * wakeme_after_rcu() - Callback function to awaken a task after grace period
368 * @head: Pointer to rcu_head member within rcu_synchronize structure
369 *
370 * Awaken the corresponding task now that a grace period has elapsed.
371 */
372void wakeme_after_rcu(struct rcu_head *head)
373{
374 struct rcu_synchronize *rcu;
375
376 rcu = container_of(head, struct rcu_synchronize, head);
377 complete(&rcu->completion);
378}
379EXPORT_SYMBOL_GPL(wakeme_after_rcu);
380
381void __wait_rcu_gp(bool checktiny, int n, call_rcu_func_t *crcu_array,
382 struct rcu_synchronize *rs_array)
383{
384 int i;
385 int j;
386
387 /* Initialize and register callbacks for each flavor specified. */
388 for (i = 0; i < n; i++) {
389 if (checktiny &&
390 (crcu_array[i] == call_rcu ||
391 crcu_array[i] == call_rcu_bh)) {
392 might_sleep();
393 continue;
394 }
395 init_rcu_head_on_stack(&rs_array[i].head);
396 init_completion(&rs_array[i].completion);
397 for (j = 0; j < i; j++)
398 if (crcu_array[j] == crcu_array[i])
399 break;
400 if (j == i)
401 (crcu_array[i])(&rs_array[i].head, wakeme_after_rcu);
402 }
403
404 /* Wait for all callbacks to be invoked. */
405 for (i = 0; i < n; i++) {
406 if (checktiny &&
407 (crcu_array[i] == call_rcu ||
408 crcu_array[i] == call_rcu_bh))
409 continue;
410 for (j = 0; j < i; j++)
411 if (crcu_array[j] == crcu_array[i])
412 break;
413 if (j == i)
414 wait_for_completion(&rs_array[i].completion);
415 destroy_rcu_head_on_stack(&rs_array[i].head);
416 }
417}
418EXPORT_SYMBOL_GPL(__wait_rcu_gp);
419
420#ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
421void init_rcu_head(struct rcu_head *head)
422{
423 debug_object_init(head, &rcuhead_debug_descr);
424}
425EXPORT_SYMBOL_GPL(init_rcu_head);
426
427void destroy_rcu_head(struct rcu_head *head)
428{
429 debug_object_free(head, &rcuhead_debug_descr);
430}
431EXPORT_SYMBOL_GPL(destroy_rcu_head);
432
433static bool rcuhead_is_static_object(void *addr)
434{
435 return true;
436}
437
438/**
439 * init_rcu_head_on_stack() - initialize on-stack rcu_head for debugobjects
440 * @head: pointer to rcu_head structure to be initialized
441 *
442 * This function informs debugobjects of a new rcu_head structure that
443 * has been allocated as an auto variable on the stack. This function
444 * is not required for rcu_head structures that are statically defined or
445 * that are dynamically allocated on the heap. This function has no
446 * effect for !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
447 */
448void init_rcu_head_on_stack(struct rcu_head *head)
449{
450 debug_object_init_on_stack(head, &rcuhead_debug_descr);
451}
452EXPORT_SYMBOL_GPL(init_rcu_head_on_stack);
453
454/**
455 * destroy_rcu_head_on_stack() - destroy on-stack rcu_head for debugobjects
456 * @head: pointer to rcu_head structure to be initialized
457 *
458 * This function informs debugobjects that an on-stack rcu_head structure
459 * is about to go out of scope. As with init_rcu_head_on_stack(), this
460 * function is not required for rcu_head structures that are statically
461 * defined or that are dynamically allocated on the heap. Also as with
462 * init_rcu_head_on_stack(), this function has no effect for
463 * !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
464 */
465void destroy_rcu_head_on_stack(struct rcu_head *head)
466{
467 debug_object_free(head, &rcuhead_debug_descr);
468}
469EXPORT_SYMBOL_GPL(destroy_rcu_head_on_stack);
470
471struct debug_obj_descr rcuhead_debug_descr = {
472 .name = "rcu_head",
473 .is_static_object = rcuhead_is_static_object,
474};
475EXPORT_SYMBOL_GPL(rcuhead_debug_descr);
476#endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
477
478#if defined(CONFIG_TREE_RCU) || defined(CONFIG_PREEMPT_RCU) || defined(CONFIG_RCU_TRACE)
479void do_trace_rcu_torture_read(const char *rcutorturename, struct rcu_head *rhp,
480 unsigned long secs,
481 unsigned long c_old, unsigned long c)
482{
483 trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c);
484}
485EXPORT_SYMBOL_GPL(do_trace_rcu_torture_read);
486#else
487#define do_trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c) \
488 do { } while (0)
489#endif
490
491#ifdef CONFIG_RCU_STALL_COMMON
492
493#ifdef CONFIG_PROVE_RCU
494#define RCU_STALL_DELAY_DELTA (5 * HZ)
495#else
496#define RCU_STALL_DELAY_DELTA 0
497#endif
498
499int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
500EXPORT_SYMBOL_GPL(rcu_cpu_stall_suppress);
501static int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
502
503module_param(rcu_cpu_stall_suppress, int, 0644);
504module_param(rcu_cpu_stall_timeout, int, 0644);
505
506int rcu_jiffies_till_stall_check(void)
507{
508 int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
509
510 /*
511 * Limit check must be consistent with the Kconfig limits
512 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
513 */
514 if (till_stall_check < 3) {
515 WRITE_ONCE(rcu_cpu_stall_timeout, 3);
516 till_stall_check = 3;
517 } else if (till_stall_check > 300) {
518 WRITE_ONCE(rcu_cpu_stall_timeout, 300);
519 till_stall_check = 300;
520 }
521 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
522}
523
524void rcu_sysrq_start(void)
525{
526 if (!rcu_cpu_stall_suppress)
527 rcu_cpu_stall_suppress = 2;
528}
529
530void rcu_sysrq_end(void)
531{
532 if (rcu_cpu_stall_suppress == 2)
533 rcu_cpu_stall_suppress = 0;
534}
535
536static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
537{
538 rcu_cpu_stall_suppress = 1;
539 return NOTIFY_DONE;
540}
541
542static struct notifier_block rcu_panic_block = {
543 .notifier_call = rcu_panic,
544};
545
546static int __init check_cpu_stall_init(void)
547{
548 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
549 return 0;
550}
551early_initcall(check_cpu_stall_init);
552
553#endif /* #ifdef CONFIG_RCU_STALL_COMMON */
554
555#ifdef CONFIG_TASKS_RCU
556
557/*
558 * Simple variant of RCU whose quiescent states are voluntary context switch,
559 * user-space execution, and idle. As such, grace periods can take one good
560 * long time. There are no read-side primitives similar to rcu_read_lock()
561 * and rcu_read_unlock() because this implementation is intended to get
562 * the system into a safe state for some of the manipulations involved in
563 * tracing and the like. Finally, this implementation does not support
564 * high call_rcu_tasks() rates from multiple CPUs. If this is required,
565 * per-CPU callback lists will be needed.
566 */
567
568/* Global list of callbacks and associated lock. */
569static struct rcu_head *rcu_tasks_cbs_head;
570static struct rcu_head **rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
571static DECLARE_WAIT_QUEUE_HEAD(rcu_tasks_cbs_wq);
572static DEFINE_RAW_SPINLOCK(rcu_tasks_cbs_lock);
573
574/* Track exiting tasks in order to allow them to be waited for. */
575DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu);
576
577/* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */
578#define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
579static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
580module_param(rcu_task_stall_timeout, int, 0644);
581
582static struct task_struct *rcu_tasks_kthread_ptr;
583
584/**
585 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
586 * @rhp: structure to be used for queueing the RCU updates.
587 * @func: actual callback function to be invoked after the grace period
588 *
589 * The callback function will be invoked some time after a full grace
590 * period elapses, in other words after all currently executing RCU
591 * read-side critical sections have completed. call_rcu_tasks() assumes
592 * that the read-side critical sections end at a voluntary context
593 * switch (not a preemption!), entry into idle, or transition to usermode
594 * execution. As such, there are no read-side primitives analogous to
595 * rcu_read_lock() and rcu_read_unlock() because this primitive is intended
596 * to determine that all tasks have passed through a safe state, not so
597 * much for data-strcuture synchronization.
598 *
599 * See the description of call_rcu() for more detailed information on
600 * memory ordering guarantees.
601 */
602void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
603{
604 unsigned long flags;
605 bool needwake;
606
607 rhp->next = NULL;
608 rhp->func = func;
609 raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
610 needwake = !rcu_tasks_cbs_head;
611 *rcu_tasks_cbs_tail = rhp;
612 rcu_tasks_cbs_tail = &rhp->next;
613 raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
614 /* We can't create the thread unless interrupts are enabled. */
615 if (needwake && READ_ONCE(rcu_tasks_kthread_ptr))
616 wake_up(&rcu_tasks_cbs_wq);
617}
618EXPORT_SYMBOL_GPL(call_rcu_tasks);
619
620/**
621 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
622 *
623 * Control will return to the caller some time after a full rcu-tasks
624 * grace period has elapsed, in other words after all currently
625 * executing rcu-tasks read-side critical sections have elapsed. These
626 * read-side critical sections are delimited by calls to schedule(),
627 * cond_resched_rcu_qs(), idle execution, userspace execution, calls
628 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
629 *
630 * This is a very specialized primitive, intended only for a few uses in
631 * tracing and other situations requiring manipulation of function
632 * preambles and profiling hooks. The synchronize_rcu_tasks() function
633 * is not (yet) intended for heavy use from multiple CPUs.
634 *
635 * Note that this guarantee implies further memory-ordering guarantees.
636 * On systems with more than one CPU, when synchronize_rcu_tasks() returns,
637 * each CPU is guaranteed to have executed a full memory barrier since the
638 * end of its last RCU-tasks read-side critical section whose beginning
639 * preceded the call to synchronize_rcu_tasks(). In addition, each CPU
640 * having an RCU-tasks read-side critical section that extends beyond
641 * the return from synchronize_rcu_tasks() is guaranteed to have executed
642 * a full memory barrier after the beginning of synchronize_rcu_tasks()
643 * and before the beginning of that RCU-tasks read-side critical section.
644 * Note that these guarantees include CPUs that are offline, idle, or
645 * executing in user mode, as well as CPUs that are executing in the kernel.
646 *
647 * Furthermore, if CPU A invoked synchronize_rcu_tasks(), which returned
648 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
649 * to have executed a full memory barrier during the execution of
650 * synchronize_rcu_tasks() -- even if CPU A and CPU B are the same CPU
651 * (but again only if the system has more than one CPU).
652 */
653void synchronize_rcu_tasks(void)
654{
655 /* Complain if the scheduler has not started. */
656 RCU_LOCKDEP_WARN(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
657 "synchronize_rcu_tasks called too soon");
658
659 /* Wait for the grace period. */
660 wait_rcu_gp(call_rcu_tasks);
661}
662EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
663
664/**
665 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
666 *
667 * Although the current implementation is guaranteed to wait, it is not
668 * obligated to, for example, if there are no pending callbacks.
669 */
670void rcu_barrier_tasks(void)
671{
672 /* There is only one callback queue, so this is easy. ;-) */
673 synchronize_rcu_tasks();
674}
675EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
676
677/* See if tasks are still holding out, complain if so. */
678static void check_holdout_task(struct task_struct *t,
679 bool needreport, bool *firstreport)
680{
681 int cpu;
682
683 if (!READ_ONCE(t->rcu_tasks_holdout) ||
684 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
685 !READ_ONCE(t->on_rq) ||
686 (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
687 !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
688 WRITE_ONCE(t->rcu_tasks_holdout, false);
689 list_del_init(&t->rcu_tasks_holdout_list);
690 put_task_struct(t);
691 return;
692 }
693 rcu_request_urgent_qs_task(t);
694 if (!needreport)
695 return;
696 if (*firstreport) {
697 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
698 *firstreport = false;
699 }
700 cpu = task_cpu(t);
701 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
702 t, ".I"[is_idle_task(t)],
703 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
704 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
705 t->rcu_tasks_idle_cpu, cpu);
706 sched_show_task(t);
707}
708
709/* RCU-tasks kthread that detects grace periods and invokes callbacks. */
710static int __noreturn rcu_tasks_kthread(void *arg)
711{
712 unsigned long flags;
713 struct task_struct *g, *t;
714 unsigned long lastreport;
715 struct rcu_head *list;
716 struct rcu_head *next;
717 LIST_HEAD(rcu_tasks_holdouts);
718
719 /* Run on housekeeping CPUs by default. Sysadm can move if desired. */
720 housekeeping_affine(current, HK_FLAG_RCU);
721
722 /*
723 * Each pass through the following loop makes one check for
724 * newly arrived callbacks, and, if there are some, waits for
725 * one RCU-tasks grace period and then invokes the callbacks.
726 * This loop is terminated by the system going down. ;-)
727 */
728 for (;;) {
729
730 /* Pick up any new callbacks. */
731 raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
732 list = rcu_tasks_cbs_head;
733 rcu_tasks_cbs_head = NULL;
734 rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
735 raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
736
737 /* If there were none, wait a bit and start over. */
738 if (!list) {
739 wait_event_interruptible(rcu_tasks_cbs_wq,
740 rcu_tasks_cbs_head);
741 if (!rcu_tasks_cbs_head) {
742 WARN_ON(signal_pending(current));
743 schedule_timeout_interruptible(HZ/10);
744 }
745 continue;
746 }
747
748 /*
749 * Wait for all pre-existing t->on_rq and t->nvcsw
750 * transitions to complete. Invoking synchronize_sched()
751 * suffices because all these transitions occur with
752 * interrupts disabled. Without this synchronize_sched(),
753 * a read-side critical section that started before the
754 * grace period might be incorrectly seen as having started
755 * after the grace period.
756 *
757 * This synchronize_sched() also dispenses with the
758 * need for a memory barrier on the first store to
759 * ->rcu_tasks_holdout, as it forces the store to happen
760 * after the beginning of the grace period.
761 */
762 synchronize_sched();
763
764 /*
765 * There were callbacks, so we need to wait for an
766 * RCU-tasks grace period. Start off by scanning
767 * the task list for tasks that are not already
768 * voluntarily blocked. Mark these tasks and make
769 * a list of them in rcu_tasks_holdouts.
770 */
771 rcu_read_lock();
772 for_each_process_thread(g, t) {
773 if (t != current && READ_ONCE(t->on_rq) &&
774 !is_idle_task(t)) {
775 get_task_struct(t);
776 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
777 WRITE_ONCE(t->rcu_tasks_holdout, true);
778 list_add(&t->rcu_tasks_holdout_list,
779 &rcu_tasks_holdouts);
780 }
781 }
782 rcu_read_unlock();
783
784 /*
785 * Wait for tasks that are in the process of exiting.
786 * This does only part of the job, ensuring that all
787 * tasks that were previously exiting reach the point
788 * where they have disabled preemption, allowing the
789 * later synchronize_sched() to finish the job.
790 */
791 synchronize_srcu(&tasks_rcu_exit_srcu);
792
793 /*
794 * Each pass through the following loop scans the list
795 * of holdout tasks, removing any that are no longer
796 * holdouts. When the list is empty, we are done.
797 */
798 lastreport = jiffies;
799 while (!list_empty(&rcu_tasks_holdouts)) {
800 bool firstreport;
801 bool needreport;
802 int rtst;
803 struct task_struct *t1;
804
805 schedule_timeout_interruptible(HZ);
806 rtst = READ_ONCE(rcu_task_stall_timeout);
807 needreport = rtst > 0 &&
808 time_after(jiffies, lastreport + rtst);
809 if (needreport)
810 lastreport = jiffies;
811 firstreport = true;
812 WARN_ON(signal_pending(current));
813 list_for_each_entry_safe(t, t1, &rcu_tasks_holdouts,
814 rcu_tasks_holdout_list) {
815 check_holdout_task(t, needreport, &firstreport);
816 cond_resched();
817 }
818 }
819
820 /*
821 * Because ->on_rq and ->nvcsw are not guaranteed
822 * to have a full memory barriers prior to them in the
823 * schedule() path, memory reordering on other CPUs could
824 * cause their RCU-tasks read-side critical sections to
825 * extend past the end of the grace period. However,
826 * because these ->nvcsw updates are carried out with
827 * interrupts disabled, we can use synchronize_sched()
828 * to force the needed ordering on all such CPUs.
829 *
830 * This synchronize_sched() also confines all
831 * ->rcu_tasks_holdout accesses to be within the grace
832 * period, avoiding the need for memory barriers for
833 * ->rcu_tasks_holdout accesses.
834 *
835 * In addition, this synchronize_sched() waits for exiting
836 * tasks to complete their final preempt_disable() region
837 * of execution, cleaning up after the synchronize_srcu()
838 * above.
839 */
840 synchronize_sched();
841
842 /* Invoke the callbacks. */
843 while (list) {
844 next = list->next;
845 local_bh_disable();
846 list->func(list);
847 local_bh_enable();
848 list = next;
849 cond_resched();
850 }
851 schedule_timeout_uninterruptible(HZ/10);
852 }
853}
854
855/* Spawn rcu_tasks_kthread() at core_initcall() time. */
856static int __init rcu_spawn_tasks_kthread(void)
857{
858 struct task_struct *t;
859
860 t = kthread_run(rcu_tasks_kthread, NULL, "rcu_tasks_kthread");
861 BUG_ON(IS_ERR(t));
862 smp_mb(); /* Ensure others see full kthread. */
863 WRITE_ONCE(rcu_tasks_kthread_ptr, t);
864 return 0;
865}
866core_initcall(rcu_spawn_tasks_kthread);
867
868/* Do the srcu_read_lock() for the above synchronize_srcu(). */
869void exit_tasks_rcu_start(void)
870{
871 preempt_disable();
872 current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
873 preempt_enable();
874}
875
876/* Do the srcu_read_unlock() for the above synchronize_srcu(). */
877void exit_tasks_rcu_finish(void)
878{
879 preempt_disable();
880 __srcu_read_unlock(&tasks_rcu_exit_srcu, current->rcu_tasks_idx);
881 preempt_enable();
882}
883
884#endif /* #ifdef CONFIG_TASKS_RCU */
885
886#ifndef CONFIG_TINY_RCU
887
888/*
889 * Print any non-default Tasks RCU settings.
890 */
891static void __init rcu_tasks_bootup_oddness(void)
892{
893#ifdef CONFIG_TASKS_RCU
894 if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
895 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
896 else
897 pr_info("\tTasks RCU enabled.\n");
898#endif /* #ifdef CONFIG_TASKS_RCU */
899}
900
901#endif /* #ifndef CONFIG_TINY_RCU */
902
903#ifdef CONFIG_PROVE_RCU
904
905/*
906 * Early boot self test parameters, one for each flavor
907 */
908static bool rcu_self_test;
909static bool rcu_self_test_bh;
910static bool rcu_self_test_sched;
911
912module_param(rcu_self_test, bool, 0444);
913module_param(rcu_self_test_bh, bool, 0444);
914module_param(rcu_self_test_sched, bool, 0444);
915
916static int rcu_self_test_counter;
917
918static void test_callback(struct rcu_head *r)
919{
920 rcu_self_test_counter++;
921 pr_info("RCU test callback executed %d\n", rcu_self_test_counter);
922}
923
924static void early_boot_test_call_rcu(void)
925{
926 static struct rcu_head head;
927
928 call_rcu(&head, test_callback);
929}
930
931static void early_boot_test_call_rcu_bh(void)
932{
933 static struct rcu_head head;
934
935 call_rcu_bh(&head, test_callback);
936}
937
938static void early_boot_test_call_rcu_sched(void)
939{
940 static struct rcu_head head;
941
942 call_rcu_sched(&head, test_callback);
943}
944
945void rcu_early_boot_tests(void)
946{
947 pr_info("Running RCU self tests\n");
948
949 if (rcu_self_test)
950 early_boot_test_call_rcu();
951 if (rcu_self_test_bh)
952 early_boot_test_call_rcu_bh();
953 if (rcu_self_test_sched)
954 early_boot_test_call_rcu_sched();
955 rcu_test_sync_prims();
956}
957
958static int rcu_verify_early_boot_tests(void)
959{
960 int ret = 0;
961 int early_boot_test_counter = 0;
962
963 if (rcu_self_test) {
964 early_boot_test_counter++;
965 rcu_barrier();
966 }
967 if (rcu_self_test_bh) {
968 early_boot_test_counter++;
969 rcu_barrier_bh();
970 }
971 if (rcu_self_test_sched) {
972 early_boot_test_counter++;
973 rcu_barrier_sched();
974 }
975
976 if (rcu_self_test_counter != early_boot_test_counter) {
977 WARN_ON(1);
978 ret = -1;
979 }
980
981 return ret;
982}
983late_initcall(rcu_verify_early_boot_tests);
984#else
985void rcu_early_boot_tests(void) {}
986#endif /* CONFIG_PROVE_RCU */
987
988#ifndef CONFIG_TINY_RCU
989
990/*
991 * Print any significant non-default boot-time settings.
992 */
993void __init rcupdate_announce_bootup_oddness(void)
994{
995 if (rcu_normal)
996 pr_info("\tNo expedited grace period (rcu_normal).\n");
997 else if (rcu_normal_after_boot)
998 pr_info("\tNo expedited grace period (rcu_normal_after_boot).\n");
999 else if (rcu_expedited)
1000 pr_info("\tAll grace periods are expedited (rcu_expedited).\n");
1001 if (rcu_cpu_stall_suppress)
1002 pr_info("\tRCU CPU stall warnings suppressed (rcu_cpu_stall_suppress).\n");
1003 if (rcu_cpu_stall_timeout != CONFIG_RCU_CPU_STALL_TIMEOUT)
1004 pr_info("\tRCU CPU stall warnings timeout set to %d (rcu_cpu_stall_timeout).\n", rcu_cpu_stall_timeout);
1005 rcu_tasks_bootup_oddness();
1006}
1007
1008#endif /* #ifndef CONFIG_TINY_RCU */