.. | .. |
---|
| 1 | +// SPDX-License-Identifier: GPL-2.0+ |
---|
1 | 2 | /* |
---|
2 | 3 | * RCU-based infrastructure for lightweight reader-writer locking |
---|
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 | 4 | * |
---|
18 | 5 | * Copyright (c) 2015, Red Hat, Inc. |
---|
19 | 6 | * |
---|
.. | .. |
---|
23 | 10 | #include <linux/rcu_sync.h> |
---|
24 | 11 | #include <linux/sched.h> |
---|
25 | 12 | |
---|
26 | | -#ifdef CONFIG_PROVE_RCU |
---|
27 | | -#define __INIT_HELD(func) .held = func, |
---|
28 | | -#else |
---|
29 | | -#define __INIT_HELD(func) |
---|
30 | | -#endif |
---|
31 | | - |
---|
32 | | -static const struct { |
---|
33 | | - void (*sync)(void); |
---|
34 | | - void (*call)(struct rcu_head *, void (*)(struct rcu_head *)); |
---|
35 | | - void (*wait)(void); |
---|
36 | | -#ifdef CONFIG_PROVE_RCU |
---|
37 | | - int (*held)(void); |
---|
38 | | -#endif |
---|
39 | | -} gp_ops[] = { |
---|
40 | | - [RCU_SYNC] = { |
---|
41 | | - .sync = synchronize_rcu, |
---|
42 | | - .call = call_rcu, |
---|
43 | | - .wait = rcu_barrier, |
---|
44 | | - __INIT_HELD(rcu_read_lock_held) |
---|
45 | | - }, |
---|
46 | | - [RCU_SCHED_SYNC] = { |
---|
47 | | - .sync = synchronize_sched, |
---|
48 | | - .call = call_rcu_sched, |
---|
49 | | - .wait = rcu_barrier_sched, |
---|
50 | | - __INIT_HELD(rcu_read_lock_sched_held) |
---|
51 | | - }, |
---|
52 | | - [RCU_BH_SYNC] = { |
---|
53 | | - .sync = synchronize_rcu_bh, |
---|
54 | | - .call = call_rcu_bh, |
---|
55 | | - .wait = rcu_barrier_bh, |
---|
56 | | - __INIT_HELD(rcu_read_lock_bh_held) |
---|
57 | | - }, |
---|
58 | | -}; |
---|
59 | | - |
---|
60 | | -enum { GP_IDLE = 0, GP_PENDING, GP_PASSED }; |
---|
61 | | -enum { CB_IDLE = 0, CB_PENDING, CB_REPLAY }; |
---|
| 13 | +enum { GP_IDLE = 0, GP_ENTER, GP_PASSED, GP_EXIT, GP_REPLAY }; |
---|
62 | 14 | |
---|
63 | 15 | #define rss_lock gp_wait.lock |
---|
64 | | - |
---|
65 | | -#ifdef CONFIG_PROVE_RCU |
---|
66 | | -void rcu_sync_lockdep_assert(struct rcu_sync *rsp) |
---|
67 | | -{ |
---|
68 | | - RCU_LOCKDEP_WARN(!gp_ops[rsp->gp_type].held(), |
---|
69 | | - "suspicious rcu_sync_is_idle() usage"); |
---|
70 | | -} |
---|
71 | | - |
---|
72 | | -EXPORT_SYMBOL_GPL(rcu_sync_lockdep_assert); |
---|
73 | | -#endif |
---|
74 | 16 | |
---|
75 | 17 | /** |
---|
76 | 18 | * rcu_sync_init() - Initialize an rcu_sync structure |
---|
77 | 19 | * @rsp: Pointer to rcu_sync structure to be initialized |
---|
78 | | - * @type: Flavor of RCU with which to synchronize rcu_sync structure |
---|
79 | 20 | */ |
---|
80 | | -void rcu_sync_init(struct rcu_sync *rsp, enum rcu_sync_type type) |
---|
| 21 | +void rcu_sync_init(struct rcu_sync *rsp) |
---|
81 | 22 | { |
---|
82 | 23 | memset(rsp, 0, sizeof(*rsp)); |
---|
83 | 24 | init_waitqueue_head(&rsp->gp_wait); |
---|
84 | | - rsp->gp_type = type; |
---|
85 | 25 | } |
---|
86 | 26 | |
---|
87 | 27 | /** |
---|
.. | .. |
---|
97 | 37 | { |
---|
98 | 38 | rsp->gp_count++; |
---|
99 | 39 | rsp->gp_state = GP_PASSED; |
---|
| 40 | +} |
---|
| 41 | + |
---|
| 42 | + |
---|
| 43 | +static void rcu_sync_func(struct rcu_head *rhp); |
---|
| 44 | + |
---|
| 45 | +static void rcu_sync_call(struct rcu_sync *rsp) |
---|
| 46 | +{ |
---|
| 47 | + call_rcu(&rsp->cb_head, rcu_sync_func); |
---|
| 48 | +} |
---|
| 49 | + |
---|
| 50 | +/** |
---|
| 51 | + * rcu_sync_func() - Callback function managing reader access to fastpath |
---|
| 52 | + * @rhp: Pointer to rcu_head in rcu_sync structure to use for synchronization |
---|
| 53 | + * |
---|
| 54 | + * This function is passed to call_rcu() function by rcu_sync_enter() and |
---|
| 55 | + * rcu_sync_exit(), so that it is invoked after a grace period following the |
---|
| 56 | + * that invocation of enter/exit. |
---|
| 57 | + * |
---|
| 58 | + * If it is called by rcu_sync_enter() it signals that all the readers were |
---|
| 59 | + * switched onto slow path. |
---|
| 60 | + * |
---|
| 61 | + * If it is called by rcu_sync_exit() it takes action based on events that |
---|
| 62 | + * have taken place in the meantime, so that closely spaced rcu_sync_enter() |
---|
| 63 | + * and rcu_sync_exit() pairs need not wait for a grace period. |
---|
| 64 | + * |
---|
| 65 | + * If another rcu_sync_enter() is invoked before the grace period |
---|
| 66 | + * ended, reset state to allow the next rcu_sync_exit() to let the |
---|
| 67 | + * readers back onto their fastpaths (after a grace period). If both |
---|
| 68 | + * another rcu_sync_enter() and its matching rcu_sync_exit() are invoked |
---|
| 69 | + * before the grace period ended, re-invoke call_rcu() on behalf of that |
---|
| 70 | + * rcu_sync_exit(). Otherwise, set all state back to idle so that readers |
---|
| 71 | + * can again use their fastpaths. |
---|
| 72 | + */ |
---|
| 73 | +static void rcu_sync_func(struct rcu_head *rhp) |
---|
| 74 | +{ |
---|
| 75 | + struct rcu_sync *rsp = container_of(rhp, struct rcu_sync, cb_head); |
---|
| 76 | + unsigned long flags; |
---|
| 77 | + |
---|
| 78 | + WARN_ON_ONCE(READ_ONCE(rsp->gp_state) == GP_IDLE); |
---|
| 79 | + WARN_ON_ONCE(READ_ONCE(rsp->gp_state) == GP_PASSED); |
---|
| 80 | + |
---|
| 81 | + spin_lock_irqsave(&rsp->rss_lock, flags); |
---|
| 82 | + if (rsp->gp_count) { |
---|
| 83 | + /* |
---|
| 84 | + * We're at least a GP after the GP_IDLE->GP_ENTER transition. |
---|
| 85 | + */ |
---|
| 86 | + WRITE_ONCE(rsp->gp_state, GP_PASSED); |
---|
| 87 | + wake_up_locked(&rsp->gp_wait); |
---|
| 88 | + } else if (rsp->gp_state == GP_REPLAY) { |
---|
| 89 | + /* |
---|
| 90 | + * A new rcu_sync_exit() has happened; requeue the callback to |
---|
| 91 | + * catch a later GP. |
---|
| 92 | + */ |
---|
| 93 | + WRITE_ONCE(rsp->gp_state, GP_EXIT); |
---|
| 94 | + rcu_sync_call(rsp); |
---|
| 95 | + } else { |
---|
| 96 | + /* |
---|
| 97 | + * We're at least a GP after the last rcu_sync_exit(); eveybody |
---|
| 98 | + * will now have observed the write side critical section. |
---|
| 99 | + * Let 'em rip!. |
---|
| 100 | + */ |
---|
| 101 | + WRITE_ONCE(rsp->gp_state, GP_IDLE); |
---|
| 102 | + } |
---|
| 103 | + spin_unlock_irqrestore(&rsp->rss_lock, flags); |
---|
100 | 104 | } |
---|
101 | 105 | |
---|
102 | 106 | /** |
---|
.. | .. |
---|
116 | 120 | */ |
---|
117 | 121 | void rcu_sync_enter(struct rcu_sync *rsp) |
---|
118 | 122 | { |
---|
119 | | - bool need_wait, need_sync; |
---|
| 123 | + int gp_state; |
---|
120 | 124 | |
---|
121 | 125 | spin_lock_irq(&rsp->rss_lock); |
---|
122 | | - need_wait = rsp->gp_count++; |
---|
123 | | - need_sync = rsp->gp_state == GP_IDLE; |
---|
124 | | - if (need_sync) |
---|
125 | | - rsp->gp_state = GP_PENDING; |
---|
| 126 | + gp_state = rsp->gp_state; |
---|
| 127 | + if (gp_state == GP_IDLE) { |
---|
| 128 | + WRITE_ONCE(rsp->gp_state, GP_ENTER); |
---|
| 129 | + WARN_ON_ONCE(rsp->gp_count); |
---|
| 130 | + /* |
---|
| 131 | + * Note that we could simply do rcu_sync_call(rsp) here and |
---|
| 132 | + * avoid the "if (gp_state == GP_IDLE)" block below. |
---|
| 133 | + * |
---|
| 134 | + * However, synchronize_rcu() can be faster if rcu_expedited |
---|
| 135 | + * or rcu_blocking_is_gp() is true. |
---|
| 136 | + * |
---|
| 137 | + * Another reason is that we can't wait for rcu callback if |
---|
| 138 | + * we are called at early boot time but this shouldn't happen. |
---|
| 139 | + */ |
---|
| 140 | + } |
---|
| 141 | + rsp->gp_count++; |
---|
126 | 142 | spin_unlock_irq(&rsp->rss_lock); |
---|
127 | 143 | |
---|
128 | | - BUG_ON(need_wait && need_sync); |
---|
129 | | - |
---|
130 | | - if (need_sync) { |
---|
131 | | - gp_ops[rsp->gp_type].sync(); |
---|
132 | | - rsp->gp_state = GP_PASSED; |
---|
133 | | - wake_up_all(&rsp->gp_wait); |
---|
134 | | - } else if (need_wait) { |
---|
135 | | - wait_event(rsp->gp_wait, rsp->gp_state == GP_PASSED); |
---|
136 | | - } else { |
---|
| 144 | + if (gp_state == GP_IDLE) { |
---|
137 | 145 | /* |
---|
138 | | - * Possible when there's a pending CB from a rcu_sync_exit(). |
---|
139 | | - * Nobody has yet been allowed the 'fast' path and thus we can |
---|
140 | | - * avoid doing any sync(). The callback will get 'dropped'. |
---|
| 146 | + * See the comment above, this simply does the "synchronous" |
---|
| 147 | + * call_rcu(rcu_sync_func) which does GP_ENTER -> GP_PASSED. |
---|
141 | 148 | */ |
---|
142 | | - BUG_ON(rsp->gp_state != GP_PASSED); |
---|
| 149 | + synchronize_rcu(); |
---|
| 150 | + rcu_sync_func(&rsp->cb_head); |
---|
| 151 | + /* Not really needed, wait_event() would see GP_PASSED. */ |
---|
| 152 | + return; |
---|
143 | 153 | } |
---|
| 154 | + |
---|
| 155 | + wait_event(rsp->gp_wait, READ_ONCE(rsp->gp_state) >= GP_PASSED); |
---|
144 | 156 | } |
---|
145 | 157 | |
---|
146 | 158 | /** |
---|
147 | | - * rcu_sync_func() - Callback function managing reader access to fastpath |
---|
148 | | - * @rhp: Pointer to rcu_head in rcu_sync structure to use for synchronization |
---|
149 | | - * |
---|
150 | | - * This function is passed to one of the call_rcu() functions by |
---|
151 | | - * rcu_sync_exit(), so that it is invoked after a grace period following the |
---|
152 | | - * that invocation of rcu_sync_exit(). It takes action based on events that |
---|
153 | | - * have taken place in the meantime, so that closely spaced rcu_sync_enter() |
---|
154 | | - * and rcu_sync_exit() pairs need not wait for a grace period. |
---|
155 | | - * |
---|
156 | | - * If another rcu_sync_enter() is invoked before the grace period |
---|
157 | | - * ended, reset state to allow the next rcu_sync_exit() to let the |
---|
158 | | - * readers back onto their fastpaths (after a grace period). If both |
---|
159 | | - * another rcu_sync_enter() and its matching rcu_sync_exit() are invoked |
---|
160 | | - * before the grace period ended, re-invoke call_rcu() on behalf of that |
---|
161 | | - * rcu_sync_exit(). Otherwise, set all state back to idle so that readers |
---|
162 | | - * can again use their fastpaths. |
---|
163 | | - */ |
---|
164 | | -static void rcu_sync_func(struct rcu_head *rhp) |
---|
165 | | -{ |
---|
166 | | - struct rcu_sync *rsp = container_of(rhp, struct rcu_sync, cb_head); |
---|
167 | | - unsigned long flags; |
---|
168 | | - |
---|
169 | | - BUG_ON(rsp->gp_state != GP_PASSED); |
---|
170 | | - BUG_ON(rsp->cb_state == CB_IDLE); |
---|
171 | | - |
---|
172 | | - spin_lock_irqsave(&rsp->rss_lock, flags); |
---|
173 | | - if (rsp->gp_count) { |
---|
174 | | - /* |
---|
175 | | - * A new rcu_sync_begin() has happened; drop the callback. |
---|
176 | | - */ |
---|
177 | | - rsp->cb_state = CB_IDLE; |
---|
178 | | - } else if (rsp->cb_state == CB_REPLAY) { |
---|
179 | | - /* |
---|
180 | | - * A new rcu_sync_exit() has happened; requeue the callback |
---|
181 | | - * to catch a later GP. |
---|
182 | | - */ |
---|
183 | | - rsp->cb_state = CB_PENDING; |
---|
184 | | - gp_ops[rsp->gp_type].call(&rsp->cb_head, rcu_sync_func); |
---|
185 | | - } else { |
---|
186 | | - /* |
---|
187 | | - * We're at least a GP after rcu_sync_exit(); eveybody will now |
---|
188 | | - * have observed the write side critical section. Let 'em rip!. |
---|
189 | | - */ |
---|
190 | | - rsp->cb_state = CB_IDLE; |
---|
191 | | - rsp->gp_state = GP_IDLE; |
---|
192 | | - } |
---|
193 | | - spin_unlock_irqrestore(&rsp->rss_lock, flags); |
---|
194 | | -} |
---|
195 | | - |
---|
196 | | -/** |
---|
197 | | - * rcu_sync_exit() - Allow readers back onto fast patch after grace period |
---|
| 159 | + * rcu_sync_exit() - Allow readers back onto fast path after grace period |
---|
198 | 160 | * @rsp: Pointer to rcu_sync structure to use for synchronization |
---|
199 | 161 | * |
---|
200 | 162 | * This function is used by updaters who have completed, and can therefore |
---|
.. | .. |
---|
205 | 167 | */ |
---|
206 | 168 | void rcu_sync_exit(struct rcu_sync *rsp) |
---|
207 | 169 | { |
---|
| 170 | + WARN_ON_ONCE(READ_ONCE(rsp->gp_state) == GP_IDLE); |
---|
| 171 | + WARN_ON_ONCE(READ_ONCE(rsp->gp_count) == 0); |
---|
| 172 | + |
---|
208 | 173 | spin_lock_irq(&rsp->rss_lock); |
---|
209 | 174 | if (!--rsp->gp_count) { |
---|
210 | | - if (rsp->cb_state == CB_IDLE) { |
---|
211 | | - rsp->cb_state = CB_PENDING; |
---|
212 | | - gp_ops[rsp->gp_type].call(&rsp->cb_head, rcu_sync_func); |
---|
213 | | - } else if (rsp->cb_state == CB_PENDING) { |
---|
214 | | - rsp->cb_state = CB_REPLAY; |
---|
| 175 | + if (rsp->gp_state == GP_PASSED) { |
---|
| 176 | + WRITE_ONCE(rsp->gp_state, GP_EXIT); |
---|
| 177 | + rcu_sync_call(rsp); |
---|
| 178 | + } else if (rsp->gp_state == GP_EXIT) { |
---|
| 179 | + WRITE_ONCE(rsp->gp_state, GP_REPLAY); |
---|
215 | 180 | } |
---|
216 | 181 | } |
---|
217 | 182 | spin_unlock_irq(&rsp->rss_lock); |
---|
.. | .. |
---|
223 | 188 | */ |
---|
224 | 189 | void rcu_sync_dtor(struct rcu_sync *rsp) |
---|
225 | 190 | { |
---|
226 | | - int cb_state; |
---|
| 191 | + int gp_state; |
---|
227 | 192 | |
---|
228 | | - BUG_ON(rsp->gp_count); |
---|
| 193 | + WARN_ON_ONCE(READ_ONCE(rsp->gp_count)); |
---|
| 194 | + WARN_ON_ONCE(READ_ONCE(rsp->gp_state) == GP_PASSED); |
---|
229 | 195 | |
---|
230 | 196 | spin_lock_irq(&rsp->rss_lock); |
---|
231 | | - if (rsp->cb_state == CB_REPLAY) |
---|
232 | | - rsp->cb_state = CB_PENDING; |
---|
233 | | - cb_state = rsp->cb_state; |
---|
| 197 | + if (rsp->gp_state == GP_REPLAY) |
---|
| 198 | + WRITE_ONCE(rsp->gp_state, GP_EXIT); |
---|
| 199 | + gp_state = rsp->gp_state; |
---|
234 | 200 | spin_unlock_irq(&rsp->rss_lock); |
---|
235 | 201 | |
---|
236 | | - if (cb_state != CB_IDLE) { |
---|
237 | | - gp_ops[rsp->gp_type].wait(); |
---|
238 | | - BUG_ON(rsp->cb_state != CB_IDLE); |
---|
| 202 | + if (gp_state != GP_IDLE) { |
---|
| 203 | + rcu_barrier(); |
---|
| 204 | + WARN_ON_ONCE(rsp->gp_state != GP_IDLE); |
---|
239 | 205 | } |
---|
240 | 206 | } |
---|