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