1 // SPDX-License-Identifier: GPL-2.0+ <<
2 /* 1 /*
3 * RCU-based infrastructure for lightweight re 2 * RCU-based infrastructure for lightweight reader-writer locking
4 * 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 *
5 * Copyright (c) 2015, Red Hat, Inc. 18 * Copyright (c) 2015, Red Hat, Inc.
6 * 19 *
7 * Author: Oleg Nesterov <oleg@redhat.com> 20 * Author: Oleg Nesterov <oleg@redhat.com>
8 */ 21 */
9 22
10 #include <linux/rcu_sync.h> 23 #include <linux/rcu_sync.h>
11 #include <linux/sched.h> 24 #include <linux/sched.h>
12 25
13 enum { GP_IDLE = 0, GP_ENTER, GP_PASSED, GP_EX !! 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 };
14 62
15 #define rss_lock gp_wait.lock 63 #define rss_lock gp_wait.lock
16 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
17 /** 75 /**
18 * rcu_sync_init() - Initialize an rcu_sync st 76 * rcu_sync_init() - Initialize an rcu_sync structure
19 * @rsp: Pointer to rcu_sync structure to be i 77 * @rsp: Pointer to rcu_sync structure to be initialized
>> 78 * @type: Flavor of RCU with which to synchronize rcu_sync structure
20 */ 79 */
21 void rcu_sync_init(struct rcu_sync *rsp) !! 80 void rcu_sync_init(struct rcu_sync *rsp, enum rcu_sync_type type)
22 { 81 {
23 memset(rsp, 0, sizeof(*rsp)); 82 memset(rsp, 0, sizeof(*rsp));
24 init_waitqueue_head(&rsp->gp_wait); 83 init_waitqueue_head(&rsp->gp_wait);
>> 84 rsp->gp_type = type;
25 } 85 }
26 86
27 static void rcu_sync_func(struct rcu_head *rhp !! 87 /**
>> 88 * rcu_sync_enter_start - Force readers onto slow path for multiple updates
>> 89 * @rsp: Pointer to rcu_sync structure to use for synchronization
>> 90 *
>> 91 * Must be called after rcu_sync_init() and before first use.
>> 92 *
>> 93 * Ensures rcu_sync_is_idle() returns false and rcu_sync_{enter,exit}()
>> 94 * pairs turn into NO-OPs.
>> 95 */
>> 96 void rcu_sync_enter_start(struct rcu_sync *rsp)
>> 97 {
>> 98 rsp->gp_count++;
>> 99 rsp->gp_state = GP_PASSED;
>> 100 }
28 101
29 static void rcu_sync_call(struct rcu_sync *rsp !! 102 /**
>> 103 * rcu_sync_enter() - Force readers onto slowpath
>> 104 * @rsp: Pointer to rcu_sync structure to use for synchronization
>> 105 *
>> 106 * This function is used by updaters who need readers to make use of
>> 107 * a slowpath during the update. After this function returns, all
>> 108 * subsequent calls to rcu_sync_is_idle() will return false, which
>> 109 * tells readers to stay off their fastpaths. A later call to
>> 110 * rcu_sync_exit() re-enables reader slowpaths.
>> 111 *
>> 112 * When called in isolation, rcu_sync_enter() must wait for a grace
>> 113 * period, however, closely spaced calls to rcu_sync_enter() can
>> 114 * optimize away the grace-period wait via a state machine implemented
>> 115 * by rcu_sync_enter(), rcu_sync_exit(), and rcu_sync_func().
>> 116 */
>> 117 void rcu_sync_enter(struct rcu_sync *rsp)
30 { 118 {
31 call_rcu_hurry(&rsp->cb_head, rcu_sync !! 119 bool need_wait, need_sync;
>> 120
>> 121 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 spin_unlock_irq(&rsp->rss_lock);
>> 127
>> 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 {
>> 137 /*
>> 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'.
>> 141 */
>> 142 BUG_ON(rsp->gp_state != GP_PASSED);
>> 143 }
32 } 144 }
33 145
34 /** 146 /**
35 * rcu_sync_func() - Callback function managin 147 * rcu_sync_func() - Callback function managing reader access to fastpath
36 * @rhp: Pointer to rcu_head in rcu_sync struc 148 * @rhp: Pointer to rcu_head in rcu_sync structure to use for synchronization
37 * 149 *
38 * This function is passed to call_rcu() funct !! 150 * This function is passed to one of the call_rcu() functions by
39 * rcu_sync_exit(), so that it is invoked afte 151 * rcu_sync_exit(), so that it is invoked after a grace period following the
40 * that invocation of enter/exit. !! 152 * that invocation of rcu_sync_exit(). It takes action based on events that
41 * <<
42 * If it is called by rcu_sync_enter() it sign <<
43 * switched onto slow path. <<
44 * <<
45 * If it is called by rcu_sync_exit() it takes <<
46 * have taken place in the meantime, so that c 153 * have taken place in the meantime, so that closely spaced rcu_sync_enter()
47 * and rcu_sync_exit() pairs need not wait for 154 * and rcu_sync_exit() pairs need not wait for a grace period.
48 * 155 *
49 * If another rcu_sync_enter() is invoked befo 156 * If another rcu_sync_enter() is invoked before the grace period
50 * ended, reset state to allow the next rcu_sy 157 * ended, reset state to allow the next rcu_sync_exit() to let the
51 * readers back onto their fastpaths (after a 158 * readers back onto their fastpaths (after a grace period). If both
52 * another rcu_sync_enter() and its matching r 159 * another rcu_sync_enter() and its matching rcu_sync_exit() are invoked
53 * before the grace period ended, re-invoke ca 160 * before the grace period ended, re-invoke call_rcu() on behalf of that
54 * rcu_sync_exit(). Otherwise, set all state 161 * rcu_sync_exit(). Otherwise, set all state back to idle so that readers
55 * can again use their fastpaths. 162 * can again use their fastpaths.
56 */ 163 */
57 static void rcu_sync_func(struct rcu_head *rhp 164 static void rcu_sync_func(struct rcu_head *rhp)
58 { 165 {
59 struct rcu_sync *rsp = container_of(rh 166 struct rcu_sync *rsp = container_of(rhp, struct rcu_sync, cb_head);
60 unsigned long flags; 167 unsigned long flags;
61 168
62 WARN_ON_ONCE(READ_ONCE(rsp->gp_state) !! 169 BUG_ON(rsp->gp_state != GP_PASSED);
63 WARN_ON_ONCE(READ_ONCE(rsp->gp_state) !! 170 BUG_ON(rsp->cb_state == CB_IDLE);
64 171
65 spin_lock_irqsave(&rsp->rss_lock, flag 172 spin_lock_irqsave(&rsp->rss_lock, flags);
66 if (rsp->gp_count) { 173 if (rsp->gp_count) {
67 /* 174 /*
68 * We're at least a GP after t !! 175 * A new rcu_sync_begin() has happened; drop the callback.
69 */ 176 */
70 WRITE_ONCE(rsp->gp_state, GP_P !! 177 rsp->cb_state = CB_IDLE;
71 wake_up_locked(&rsp->gp_wait); !! 178 } else if (rsp->cb_state == CB_REPLAY) {
72 } else if (rsp->gp_state == GP_REPLAY) <<
73 /* 179 /*
74 * A new rcu_sync_exit() has h !! 180 * A new rcu_sync_exit() has happened; requeue the callback
75 * catch a later GP. !! 181 * to catch a later GP.
76 */ 182 */
77 WRITE_ONCE(rsp->gp_state, GP_E !! 183 rsp->cb_state = CB_PENDING;
78 rcu_sync_call(rsp); !! 184 gp_ops[rsp->gp_type].call(&rsp->cb_head, rcu_sync_func);
79 } else { 185 } else {
80 /* 186 /*
81 * We're at least a GP after t !! 187 * We're at least a GP after rcu_sync_exit(); eveybody will now
82 * will now have observed the !! 188 * have observed the write side critical section. Let 'em rip!.
83 * Let 'em rip! <<
84 */ 189 */
85 WRITE_ONCE(rsp->gp_state, GP_I !! 190 rsp->cb_state = CB_IDLE;
>> 191 rsp->gp_state = GP_IDLE;
86 } 192 }
87 spin_unlock_irqrestore(&rsp->rss_lock, 193 spin_unlock_irqrestore(&rsp->rss_lock, flags);
88 } 194 }
89 195
90 /** 196 /**
91 * rcu_sync_enter() - Force readers onto slowp !! 197 * rcu_sync_exit() - Allow readers back onto fast patch after grace period
92 * @rsp: Pointer to rcu_sync structure to use <<
93 * <<
94 * This function is used by updaters who need <<
95 * a slowpath during the update. After this f <<
96 * subsequent calls to rcu_sync_is_idle() will <<
97 * tells readers to stay off their fastpaths. <<
98 * rcu_sync_exit() re-enables reader fastpaths <<
99 * <<
100 * When called in isolation, rcu_sync_enter() <<
101 * period, however, closely spaced calls to rc <<
102 * optimize away the grace-period wait via a s <<
103 * by rcu_sync_enter(), rcu_sync_exit(), and r <<
104 */ <<
105 void rcu_sync_enter(struct rcu_sync *rsp) <<
106 { <<
107 int gp_state; <<
108 <<
109 spin_lock_irq(&rsp->rss_lock); <<
110 gp_state = rsp->gp_state; <<
111 if (gp_state == GP_IDLE) { <<
112 WRITE_ONCE(rsp->gp_state, GP_E <<
113 WARN_ON_ONCE(rsp->gp_count); <<
114 /* <<
115 * Note that we could simply d <<
116 * avoid the "if (gp_state == <<
117 * <<
118 * However, synchronize_rcu() <<
119 * or rcu_blocking_is_gp() is <<
120 * <<
121 * Another reason is that we c <<
122 * we are called at early boot <<
123 */ <<
124 } <<
125 rsp->gp_count++; <<
126 spin_unlock_irq(&rsp->rss_lock); <<
127 <<
128 if (gp_state == GP_IDLE) { <<
129 /* <<
130 * See the comment above, this <<
131 * call_rcu(rcu_sync_func) whi <<
132 */ <<
133 synchronize_rcu(); <<
134 rcu_sync_func(&rsp->cb_head); <<
135 /* Not really needed, wait_eve <<
136 return; <<
137 } <<
138 <<
139 wait_event(rsp->gp_wait, READ_ONCE(rsp <<
140 } <<
141 <<
142 /** <<
143 * rcu_sync_exit() - Allow readers back onto f <<
144 * @rsp: Pointer to rcu_sync structure to use 198 * @rsp: Pointer to rcu_sync structure to use for synchronization
145 * 199 *
146 * This function is used by updaters who have 200 * This function is used by updaters who have completed, and can therefore
147 * now allow readers to make use of their fast 201 * now allow readers to make use of their fastpaths after a grace period
148 * has elapsed. After this grace period has c 202 * has elapsed. After this grace period has completed, all subsequent
149 * calls to rcu_sync_is_idle() will return tru 203 * calls to rcu_sync_is_idle() will return true, which tells readers that
150 * they can once again use their fastpaths. 204 * they can once again use their fastpaths.
151 */ 205 */
152 void rcu_sync_exit(struct rcu_sync *rsp) 206 void rcu_sync_exit(struct rcu_sync *rsp)
153 { 207 {
154 WARN_ON_ONCE(READ_ONCE(rsp->gp_state) <<
155 <<
156 spin_lock_irq(&rsp->rss_lock); 208 spin_lock_irq(&rsp->rss_lock);
157 WARN_ON_ONCE(rsp->gp_count == 0); <<
158 if (!--rsp->gp_count) { 209 if (!--rsp->gp_count) {
159 if (rsp->gp_state == GP_PASSED !! 210 if (rsp->cb_state == CB_IDLE) {
160 WRITE_ONCE(rsp->gp_sta !! 211 rsp->cb_state = CB_PENDING;
161 rcu_sync_call(rsp); !! 212 gp_ops[rsp->gp_type].call(&rsp->cb_head, rcu_sync_func);
162 } else if (rsp->gp_state == GP !! 213 } else if (rsp->cb_state == CB_PENDING) {
163 WRITE_ONCE(rsp->gp_sta !! 214 rsp->cb_state = CB_REPLAY;
164 } 215 }
165 } 216 }
166 spin_unlock_irq(&rsp->rss_lock); 217 spin_unlock_irq(&rsp->rss_lock);
167 } 218 }
168 219
169 /** 220 /**
170 * rcu_sync_dtor() - Clean up an rcu_sync stru 221 * rcu_sync_dtor() - Clean up an rcu_sync structure
171 * @rsp: Pointer to rcu_sync structure to be c 222 * @rsp: Pointer to rcu_sync structure to be cleaned up
172 */ 223 */
173 void rcu_sync_dtor(struct rcu_sync *rsp) 224 void rcu_sync_dtor(struct rcu_sync *rsp)
174 { 225 {
175 int gp_state; !! 226 int cb_state;
176 227
177 WARN_ON_ONCE(READ_ONCE(rsp->gp_state) !! 228 BUG_ON(rsp->gp_count);
178 229
179 spin_lock_irq(&rsp->rss_lock); 230 spin_lock_irq(&rsp->rss_lock);
180 WARN_ON_ONCE(rsp->gp_count); !! 231 if (rsp->cb_state == CB_REPLAY)
181 if (rsp->gp_state == GP_REPLAY) !! 232 rsp->cb_state = CB_PENDING;
182 WRITE_ONCE(rsp->gp_state, GP_E !! 233 cb_state = rsp->cb_state;
183 gp_state = rsp->gp_state; <<
184 spin_unlock_irq(&rsp->rss_lock); 234 spin_unlock_irq(&rsp->rss_lock);
185 235
186 if (gp_state != GP_IDLE) { !! 236 if (cb_state != CB_IDLE) {
187 rcu_barrier(); !! 237 gp_ops[rsp->gp_type].wait();
188 WARN_ON_ONCE(rsp->gp_state != !! 238 BUG_ON(rsp->cb_state != CB_IDLE);
189 } 239 }
190 } 240 }
191 241
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.