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