1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/percpu.h>
3 #include <linux/sched.h>
4 #include <linux/osq_lock.h>
5
6 /*
7 * An MCS like lock especially tailored for optimistic spinning for sleeping
8 * lock implementations (mutex, rwsem, etc).
9 *
10 * Using a single mcs node per CPU is safe because sleeping locks should not be
11 * called from interrupt context and we have preemption disabled while
12 * spinning.
13 */
14 static DEFINE_PER_CPU_SHARED_ALIGNED(struct optimistic_spin_node, osq_node);
15
16 /*
17 * We use the value 0 to represent "no CPU", thus the encoded value
18 * will be the CPU number incremented by 1.
19 */
20 static inline int encode_cpu(int cpu_nr)
21 {
22 return cpu_nr + 1;
23 }
24
25 static inline int node_cpu(struct optimistic_spin_node *node)
26 {
27 return node->cpu - 1;
28 }
29
30 static inline struct optimistic_spin_node *decode_cpu(int encoded_cpu_val)
31 {
32 int cpu_nr = encoded_cpu_val - 1;
33
34 return per_cpu_ptr(&osq_node, cpu_nr);
35 }
36
37 /*
38 * Get a stable @node->next pointer, either for unlock() or unqueue() purposes.
39 * Can return NULL in case we were the last queued and we updated @lock instead.
40 */
41 static inline struct optimistic_spin_node *
42 osq_wait_next(struct optimistic_spin_queue *lock,
43 struct optimistic_spin_node *node,
44 struct optimistic_spin_node *prev)
45 {
46 struct optimistic_spin_node *next = NULL;
47 int curr = encode_cpu(smp_processor_id());
48 int old;
49
50 /*
51 * If there is a prev node in queue, then the 'old' value will be
52 * the prev node's CPU #, else it's set to OSQ_UNLOCKED_VAL since if
53 * we're currently last in queue, then the queue will then become empty.
54 */
55 old = prev ? prev->cpu : OSQ_UNLOCKED_VAL;
56
57 for (;;) {
58 if (atomic_read(&lock->tail) == curr &&
59 atomic_cmpxchg_acquire(&lock->tail, curr, old) == curr) {
60 /*
61 * We were the last queued, we moved @lock back. @prev
62 * will now observe @lock and will complete its
63 * unlock()/unqueue().
64 */
65 break;
66 }
67
68 /*
69 * We must xchg() the @node->next value, because if we were to
70 * leave it in, a concurrent unlock()/unqueue() from
71 * @node->next might complete Step-A and think its @prev is
72 * still valid.
73 *
74 * If the concurrent unlock()/unqueue() wins the race, we'll
75 * wait for either @lock to point to us, through its Step-B, or
76 * wait for a new @node->next from its Step-C.
77 */
78 if (node->next) {
79 next = xchg(&node->next, NULL);
80 if (next)
81 break;
82 }
83
84 cpu_relax();
85 }
86
87 return next;
88 }
89
90 bool osq_lock(struct optimistic_spin_queue *lock)
91 {
92 struct optimistic_spin_node *node = this_cpu_ptr(&osq_node);
93 struct optimistic_spin_node *prev, *next;
94 int curr = encode_cpu(smp_processor_id());
95 int old;
96
97 node->locked = 0;
98 node->next = NULL;
99 node->cpu = curr;
100
101 /*
102 * We need both ACQUIRE (pairs with corresponding RELEASE in
103 * unlock() uncontended, or fastpath) and RELEASE (to publish
104 * the node fields we just initialised) semantics when updating
105 * the lock tail.
106 */
107 old = atomic_xchg(&lock->tail, curr);
108 if (old == OSQ_UNLOCKED_VAL)
109 return true;
110
111 prev = decode_cpu(old);
112 node->prev = prev;
113
114 /*
115 * osq_lock() unqueue
116 *
117 * node->prev = prev osq_wait_next()
118 * WMB MB
119 * prev->next = node next->prev = prev // unqueue-C
120 *
121 * Here 'node->prev' and 'next->prev' are the same variable and we need
122 * to ensure these stores happen in-order to avoid corrupting the list.
123 */
124 smp_wmb();
125
126 WRITE_ONCE(prev->next, node);
127
128 /*
129 * Normally @prev is untouchable after the above store; because at that
130 * moment unlock can proceed and wipe the node element from stack.
131 *
132 * However, since our nodes are static per-cpu storage, we're
133 * guaranteed their existence -- this allows us to apply
134 * cmpxchg in an attempt to undo our queueing.
135 */
136
137 while (!READ_ONCE(node->locked)) {
138 /*
139 * If we need to reschedule bail... so we can block.
140 * Use vcpu_is_preempted() to avoid waiting for a preempted
141 * lock holder:
142 */
143 if (need_resched() || vcpu_is_preempted(node_cpu(node->prev)))
144 goto unqueue;
145
146 cpu_relax();
147 }
148 return true;
149
150 unqueue:
151 /*
152 * Step - A -- stabilize @prev
153 *
154 * Undo our @prev->next assignment; this will make @prev's
155 * unlock()/unqueue() wait for a next pointer since @lock points to us
156 * (or later).
157 */
158
159 for (;;) {
160 if (prev->next == node &&
161 cmpxchg(&prev->next, node, NULL) == node)
162 break;
163
164 /*
165 * We can only fail the cmpxchg() racing against an unlock(),
166 * in which case we should observe @node->locked becomming
167 * true.
168 */
169 if (smp_load_acquire(&node->locked))
170 return true;
171
172 cpu_relax();
173
174 /*
175 * Or we race against a concurrent unqueue()'s step-B, in which
176 * case its step-C will write us a new @node->prev pointer.
177 */
178 prev = READ_ONCE(node->prev);
179 }
180
181 /*
182 * Step - B -- stabilize @next
183 *
184 * Similar to unlock(), wait for @node->next or move @lock from @node
185 * back to @prev.
186 */
187
188 next = osq_wait_next(lock, node, prev);
189 if (!next)
190 return false;
191
192 /*
193 * Step - C -- unlink
194 *
195 * @prev is stable because its still waiting for a new @prev->next
196 * pointer, @next is stable because our @node->next pointer is NULL and
197 * it will wait in Step-A.
198 */
199
200 WRITE_ONCE(next->prev, prev);
201 WRITE_ONCE(prev->next, next);
202
203 return false;
204 }
205
206 void osq_unlock(struct optimistic_spin_queue *lock)
207 {
208 struct optimistic_spin_node *node, *next;
209 int curr = encode_cpu(smp_processor_id());
210
211 /*
212 * Fast path for the uncontended case.
213 */
214 if (likely(atomic_cmpxchg_release(&lock->tail, curr,
215 OSQ_UNLOCKED_VAL) == curr))
216 return;
217
218 /*
219 * Second most likely case.
220 */
221 node = this_cpu_ptr(&osq_node);
222 next = xchg(&node->next, NULL);
223 if (next) {
224 WRITE_ONCE(next->locked, 1);
225 return;
226 }
227
228 next = osq_wait_next(lock, node, NULL);
229 if (next)
230 WRITE_ONCE(next->locked, 1);
231 }