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