1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3 * Read-Copy Update mechanism for mutual exclusion, the Bloatwatch edition.
4 *
5 * Copyright IBM Corporation, 2008
6 *
7 * Author: Paul E. McKenney <paulmck@linux.ibm.com>
8 *
9 * For detailed explanation of Read-Copy Update mechanism see -
10 * Documentation/RCU
11 */
12 #include <linux/completion.h>
13 #include <linux/interrupt.h>
14 #include <linux/notifier.h>
15 #include <linux/rcupdate_wait.h>
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/mutex.h>
19 #include <linux/sched.h>
20 #include <linux/types.h>
21 #include <linux/init.h>
22 #include <linux/time.h>
23 #include <linux/cpu.h>
24 #include <linux/prefetch.h>
25 #include <linux/slab.h>
26 #include <linux/mm.h>
27
28 #include "rcu.h"
29
30 /* Global control variables for rcupdate callback mechanism. */
31 struct rcu_ctrlblk {
32 struct rcu_head *rcucblist; /* List of pending callbacks (CBs). */
33 struct rcu_head **donetail; /* ->next pointer of last "done" CB. */
34 struct rcu_head **curtail; /* ->next pointer of last CB. */
35 unsigned long gp_seq; /* Grace-period counter. */
36 };
37
38 /* Definition for rcupdate control block. */
39 static struct rcu_ctrlblk rcu_ctrlblk = {
40 .donetail = &rcu_ctrlblk.rcucblist,
41 .curtail = &rcu_ctrlblk.rcucblist,
42 .gp_seq = 0 - 300UL,
43 };
44
rcu_barrier(void)45 void rcu_barrier(void)
46 {
47 wait_rcu_gp(call_rcu_hurry);
48 }
49 EXPORT_SYMBOL(rcu_barrier);
50
51 /* Record an rcu quiescent state. */
rcu_qs(void)52 void rcu_qs(void)
53 {
54 unsigned long flags;
55
56 local_irq_save(flags);
57 if (rcu_ctrlblk.donetail != rcu_ctrlblk.curtail) {
58 rcu_ctrlblk.donetail = rcu_ctrlblk.curtail;
59 raise_softirq_irqoff(RCU_SOFTIRQ);
60 }
61 WRITE_ONCE(rcu_ctrlblk.gp_seq, rcu_ctrlblk.gp_seq + 2);
62 local_irq_restore(flags);
63 }
64
65 /*
66 * Check to see if the scheduling-clock interrupt came from an extended
67 * quiescent state, and, if so, tell RCU about it. This function must
68 * be called from hardirq context. It is normally called from the
69 * scheduling-clock interrupt.
70 */
rcu_sched_clock_irq(int user)71 void rcu_sched_clock_irq(int user)
72 {
73 if (user)
74 rcu_qs();
75 else if (rcu_ctrlblk.donetail != rcu_ctrlblk.curtail)
76 set_need_resched_current();
77 }
78
79 /*
80 * Reclaim the specified callback, either by invoking it for non-kfree cases or
81 * freeing it directly (for kfree). Return true if kfreeing, false otherwise.
82 */
rcu_reclaim_tiny(struct rcu_head * head)83 static inline bool rcu_reclaim_tiny(struct rcu_head *head)
84 {
85 rcu_callback_t f;
86
87 rcu_lock_acquire(&rcu_callback_map);
88
89 trace_rcu_invoke_callback("", head);
90 f = head->func;
91 debug_rcu_head_callback(head);
92 WRITE_ONCE(head->func, (rcu_callback_t)0L);
93 f(head);
94 rcu_lock_release(&rcu_callback_map);
95 return false;
96 }
97
98 /* Invoke the RCU callbacks whose grace period has elapsed. */
rcu_process_callbacks(void)99 static __latent_entropy void rcu_process_callbacks(void)
100 {
101 struct rcu_head *next, *list;
102 unsigned long flags;
103
104 /* Move the ready-to-invoke callbacks to a local list. */
105 local_irq_save(flags);
106 if (rcu_ctrlblk.donetail == &rcu_ctrlblk.rcucblist) {
107 /* No callbacks ready, so just leave. */
108 local_irq_restore(flags);
109 return;
110 }
111 list = rcu_ctrlblk.rcucblist;
112 rcu_ctrlblk.rcucblist = *rcu_ctrlblk.donetail;
113 *rcu_ctrlblk.donetail = NULL;
114 if (rcu_ctrlblk.curtail == rcu_ctrlblk.donetail)
115 rcu_ctrlblk.curtail = &rcu_ctrlblk.rcucblist;
116 rcu_ctrlblk.donetail = &rcu_ctrlblk.rcucblist;
117 local_irq_restore(flags);
118
119 /* Invoke the callbacks on the local list. */
120 while (list) {
121 next = list->next;
122 prefetch(next);
123 debug_rcu_head_unqueue(list);
124 rcu_reclaim_tiny(list);
125 list = next;
126 }
127 }
128
129 /*
130 * Wait for a grace period to elapse. But it is illegal to invoke
131 * synchronize_rcu() from within an RCU read-side critical section.
132 * Therefore, any legal call to synchronize_rcu() is a quiescent state,
133 * and so on a UP system, synchronize_rcu() need do nothing, other than
134 * let the polled APIs know that another grace period elapsed.
135 *
136 * (But Lai Jiangshan points out the benefits of doing might_sleep()
137 * to reduce latency.)
138 *
139 * Cool, huh? (Due to Josh Triplett.)
140 */
synchronize_rcu(void)141 void synchronize_rcu(void)
142 {
143 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
144 lock_is_held(&rcu_lock_map) ||
145 lock_is_held(&rcu_sched_lock_map),
146 "Illegal synchronize_rcu() in RCU read-side critical section");
147 preempt_disable();
148 WRITE_ONCE(rcu_ctrlblk.gp_seq, rcu_ctrlblk.gp_seq + 2);
149 preempt_enable();
150 }
151 EXPORT_SYMBOL_GPL(synchronize_rcu);
152
153 /*
154 * Post an RCU callback to be invoked after the end of an RCU grace
155 * period. But since we have but one CPU, that would be after any
156 * quiescent state.
157 */
call_rcu(struct rcu_head * head,rcu_callback_t func)158 void call_rcu(struct rcu_head *head, rcu_callback_t func)
159 {
160 static atomic_t doublefrees;
161 unsigned long flags;
162
163 if (debug_rcu_head_queue(head)) {
164 if (atomic_inc_return(&doublefrees) < 4) {
165 pr_err("%s(): Double-freed CB %p->%pS()!!! ", __func__, head, head->func);
166 mem_dump_obj(head);
167 }
168 return;
169 }
170
171 head->func = func;
172 head->next = NULL;
173
174 local_irq_save(flags);
175 *rcu_ctrlblk.curtail = head;
176 rcu_ctrlblk.curtail = &head->next;
177 local_irq_restore(flags);
178
179 if (unlikely(is_idle_task(current))) {
180 /* force scheduling for rcu_qs() */
181 resched_cpu(0);
182 }
183 }
184 EXPORT_SYMBOL_GPL(call_rcu);
185
186 /*
187 * Store a grace-period-counter "cookie". For more information,
188 * see the Tree RCU header comment.
189 */
get_completed_synchronize_rcu_full(struct rcu_gp_oldstate * rgosp)190 void get_completed_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
191 {
192 rgosp->rgos_norm = RCU_GET_STATE_COMPLETED;
193 }
194 EXPORT_SYMBOL_GPL(get_completed_synchronize_rcu_full);
195
196 /*
197 * Return a grace-period-counter "cookie". For more information,
198 * see the Tree RCU header comment.
199 */
get_state_synchronize_rcu(void)200 unsigned long get_state_synchronize_rcu(void)
201 {
202 return READ_ONCE(rcu_ctrlblk.gp_seq);
203 }
204 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
205
206 /*
207 * Return a grace-period-counter "cookie" and ensure that a future grace
208 * period completes. For more information, see the Tree RCU header comment.
209 */
start_poll_synchronize_rcu(void)210 unsigned long start_poll_synchronize_rcu(void)
211 {
212 unsigned long gp_seq = get_state_synchronize_rcu();
213
214 if (unlikely(is_idle_task(current))) {
215 /* force scheduling for rcu_qs() */
216 resched_cpu(0);
217 }
218 return gp_seq;
219 }
220 EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu);
221
222 /*
223 * Return true if the grace period corresponding to oldstate has completed
224 * and false otherwise. For more information, see the Tree RCU header
225 * comment.
226 */
poll_state_synchronize_rcu(unsigned long oldstate)227 bool poll_state_synchronize_rcu(unsigned long oldstate)
228 {
229 return oldstate == RCU_GET_STATE_COMPLETED || READ_ONCE(rcu_ctrlblk.gp_seq) != oldstate;
230 }
231 EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu);
232
233 #if IS_ENABLED(CONFIG_RCU_TORTURE_TEST)
rcutorture_gather_gp_seqs(void)234 unsigned long long rcutorture_gather_gp_seqs(void)
235 {
236 return READ_ONCE(rcu_ctrlblk.gp_seq) & 0xffffULL;
237 }
238 EXPORT_SYMBOL_GPL(rcutorture_gather_gp_seqs);
239
rcutorture_format_gp_seqs(unsigned long long seqs,char * cp,size_t len)240 void rcutorture_format_gp_seqs(unsigned long long seqs, char *cp, size_t len)
241 {
242 snprintf(cp, len, "g%04llx", seqs & 0xffffULL);
243 }
244 EXPORT_SYMBOL_GPL(rcutorture_format_gp_seqs);
245 #endif
246
rcu_init(void)247 void __init rcu_init(void)
248 {
249 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
250 rcu_early_boot_tests();
251 tasks_cblist_init_generic();
252 }
253