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_tsk_need_resched(current);
77 set_preempt_need_resched();
78 }
79 }
80
81 /*
82 * Reclaim the specified callback, either by invoking it for non-kfree cases or
83 * freeing it directly (for kfree). Return true if kfreeing, false otherwise.
84 */
rcu_reclaim_tiny(struct rcu_head * head)85 static inline bool rcu_reclaim_tiny(struct rcu_head *head)
86 {
87 rcu_callback_t f;
88
89 rcu_lock_acquire(&rcu_callback_map);
90
91 trace_rcu_invoke_callback("", head);
92 f = head->func;
93 debug_rcu_head_callback(head);
94 WRITE_ONCE(head->func, (rcu_callback_t)0L);
95 f(head);
96 rcu_lock_release(&rcu_callback_map);
97 return false;
98 }
99
100 /* Invoke the RCU callbacks whose grace period has elapsed. */
rcu_process_callbacks(void)101 static __latent_entropy void rcu_process_callbacks(void)
102 {
103 struct rcu_head *next, *list;
104 unsigned long flags;
105
106 /* Move the ready-to-invoke callbacks to a local list. */
107 local_irq_save(flags);
108 if (rcu_ctrlblk.donetail == &rcu_ctrlblk.rcucblist) {
109 /* No callbacks ready, so just leave. */
110 local_irq_restore(flags);
111 return;
112 }
113 list = rcu_ctrlblk.rcucblist;
114 rcu_ctrlblk.rcucblist = *rcu_ctrlblk.donetail;
115 *rcu_ctrlblk.donetail = NULL;
116 if (rcu_ctrlblk.curtail == rcu_ctrlblk.donetail)
117 rcu_ctrlblk.curtail = &rcu_ctrlblk.rcucblist;
118 rcu_ctrlblk.donetail = &rcu_ctrlblk.rcucblist;
119 local_irq_restore(flags);
120
121 /* Invoke the callbacks on the local list. */
122 while (list) {
123 next = list->next;
124 prefetch(next);
125 debug_rcu_head_unqueue(list);
126 rcu_reclaim_tiny(list);
127 list = next;
128 }
129 }
130
131 /*
132 * Wait for a grace period to elapse. But it is illegal to invoke
133 * synchronize_rcu() from within an RCU read-side critical section.
134 * Therefore, any legal call to synchronize_rcu() is a quiescent state,
135 * and so on a UP system, synchronize_rcu() need do nothing, other than
136 * let the polled APIs know that another grace period elapsed.
137 *
138 * (But Lai Jiangshan points out the benefits of doing might_sleep()
139 * to reduce latency.)
140 *
141 * Cool, huh? (Due to Josh Triplett.)
142 */
synchronize_rcu(void)143 void synchronize_rcu(void)
144 {
145 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
146 lock_is_held(&rcu_lock_map) ||
147 lock_is_held(&rcu_sched_lock_map),
148 "Illegal synchronize_rcu() in RCU read-side critical section");
149 preempt_disable();
150 WRITE_ONCE(rcu_ctrlblk.gp_seq, rcu_ctrlblk.gp_seq + 2);
151 preempt_enable();
152 }
153 EXPORT_SYMBOL_GPL(synchronize_rcu);
154
155 /*
156 * Post an RCU callback to be invoked after the end of an RCU grace
157 * period. But since we have but one CPU, that would be after any
158 * quiescent state.
159 */
call_rcu(struct rcu_head * head,rcu_callback_t func)160 void call_rcu(struct rcu_head *head, rcu_callback_t func)
161 {
162 static atomic_t doublefrees;
163 unsigned long flags;
164
165 if (debug_rcu_head_queue(head)) {
166 if (atomic_inc_return(&doublefrees) < 4) {
167 pr_err("%s(): Double-freed CB %p->%pS()!!! ", __func__, head, head->func);
168 mem_dump_obj(head);
169 }
170 return;
171 }
172
173 head->func = func;
174 head->next = NULL;
175
176 local_irq_save(flags);
177 *rcu_ctrlblk.curtail = head;
178 rcu_ctrlblk.curtail = &head->next;
179 local_irq_restore(flags);
180
181 if (unlikely(is_idle_task(current))) {
182 /* force scheduling for rcu_qs() */
183 resched_cpu(0);
184 }
185 }
186 EXPORT_SYMBOL_GPL(call_rcu);
187
188 /*
189 * Store a grace-period-counter "cookie". For more information,
190 * see the Tree RCU header comment.
191 */
get_completed_synchronize_rcu_full(struct rcu_gp_oldstate * rgosp)192 void get_completed_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
193 {
194 rgosp->rgos_norm = RCU_GET_STATE_COMPLETED;
195 }
196 EXPORT_SYMBOL_GPL(get_completed_synchronize_rcu_full);
197
198 /*
199 * Return a grace-period-counter "cookie". For more information,
200 * see the Tree RCU header comment.
201 */
get_state_synchronize_rcu(void)202 unsigned long get_state_synchronize_rcu(void)
203 {
204 return READ_ONCE(rcu_ctrlblk.gp_seq);
205 }
206 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
207
208 /*
209 * Return a grace-period-counter "cookie" and ensure that a future grace
210 * period completes. For more information, see the Tree RCU header comment.
211 */
start_poll_synchronize_rcu(void)212 unsigned long start_poll_synchronize_rcu(void)
213 {
214 unsigned long gp_seq = get_state_synchronize_rcu();
215
216 if (unlikely(is_idle_task(current))) {
217 /* force scheduling for rcu_qs() */
218 resched_cpu(0);
219 }
220 return gp_seq;
221 }
222 EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu);
223
224 /*
225 * Return true if the grace period corresponding to oldstate has completed
226 * and false otherwise. For more information, see the Tree RCU header
227 * comment.
228 */
poll_state_synchronize_rcu(unsigned long oldstate)229 bool poll_state_synchronize_rcu(unsigned long oldstate)
230 {
231 return oldstate == RCU_GET_STATE_COMPLETED || READ_ONCE(rcu_ctrlblk.gp_seq) != oldstate;
232 }
233 EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu);
234
235 #if IS_ENABLED(CONFIG_RCU_TORTURE_TEST)
rcutorture_gather_gp_seqs(void)236 unsigned long long rcutorture_gather_gp_seqs(void)
237 {
238 return READ_ONCE(rcu_ctrlblk.gp_seq) & 0xffffULL;
239 }
240 EXPORT_SYMBOL_GPL(rcutorture_gather_gp_seqs);
241
rcutorture_format_gp_seqs(unsigned long long seqs,char * cp,size_t len)242 void rcutorture_format_gp_seqs(unsigned long long seqs, char *cp, size_t len)
243 {
244 snprintf(cp, len, "g%04llx", seqs & 0xffffULL);
245 }
246 EXPORT_SYMBOL_GPL(rcutorture_format_gp_seqs);
247 #endif
248
rcu_init(void)249 void __init rcu_init(void)
250 {
251 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
252 rcu_early_boot_tests();
253 tasks_cblist_init_generic();
254 }
255