xref: /openbmc/linux/kernel/locking/osq_lock.c (revision cd4d09ec)
1 #include <linux/percpu.h>
2 #include <linux/sched.h>
3 #include <linux/osq_lock.h>
4 
5 /*
6  * An MCS like lock especially tailored for optimistic spinning for sleeping
7  * lock implementations (mutex, rwsem, etc).
8  *
9  * Using a single mcs node per CPU is safe because sleeping locks should not be
10  * called from interrupt context and we have preemption disabled while
11  * spinning.
12  */
13 static DEFINE_PER_CPU_SHARED_ALIGNED(struct optimistic_spin_node, osq_node);
14 
15 /*
16  * We use the value 0 to represent "no CPU", thus the encoded value
17  * will be the CPU number incremented by 1.
18  */
19 static inline int encode_cpu(int cpu_nr)
20 {
21 	return cpu_nr + 1;
22 }
23 
24 static inline struct optimistic_spin_node *decode_cpu(int encoded_cpu_val)
25 {
26 	int cpu_nr = encoded_cpu_val - 1;
27 
28 	return per_cpu_ptr(&osq_node, cpu_nr);
29 }
30 
31 /*
32  * Get a stable @node->next pointer, either for unlock() or unqueue() purposes.
33  * Can return NULL in case we were the last queued and we updated @lock instead.
34  */
35 static inline struct optimistic_spin_node *
36 osq_wait_next(struct optimistic_spin_queue *lock,
37 	      struct optimistic_spin_node *node,
38 	      struct optimistic_spin_node *prev)
39 {
40 	struct optimistic_spin_node *next = NULL;
41 	int curr = encode_cpu(smp_processor_id());
42 	int old;
43 
44 	/*
45 	 * If there is a prev node in queue, then the 'old' value will be
46 	 * the prev node's CPU #, else it's set to OSQ_UNLOCKED_VAL since if
47 	 * we're currently last in queue, then the queue will then become empty.
48 	 */
49 	old = prev ? prev->cpu : OSQ_UNLOCKED_VAL;
50 
51 	for (;;) {
52 		if (atomic_read(&lock->tail) == curr &&
53 		    atomic_cmpxchg_acquire(&lock->tail, curr, old) == curr) {
54 			/*
55 			 * We were the last queued, we moved @lock back. @prev
56 			 * will now observe @lock and will complete its
57 			 * unlock()/unqueue().
58 			 */
59 			break;
60 		}
61 
62 		/*
63 		 * We must xchg() the @node->next value, because if we were to
64 		 * leave it in, a concurrent unlock()/unqueue() from
65 		 * @node->next might complete Step-A and think its @prev is
66 		 * still valid.
67 		 *
68 		 * If the concurrent unlock()/unqueue() wins the race, we'll
69 		 * wait for either @lock to point to us, through its Step-B, or
70 		 * wait for a new @node->next from its Step-C.
71 		 */
72 		if (node->next) {
73 			next = xchg(&node->next, NULL);
74 			if (next)
75 				break;
76 		}
77 
78 		cpu_relax_lowlatency();
79 	}
80 
81 	return next;
82 }
83 
84 bool osq_lock(struct optimistic_spin_queue *lock)
85 {
86 	struct optimistic_spin_node *node = this_cpu_ptr(&osq_node);
87 	struct optimistic_spin_node *prev, *next;
88 	int curr = encode_cpu(smp_processor_id());
89 	int old;
90 
91 	node->locked = 0;
92 	node->next = NULL;
93 	node->cpu = curr;
94 
95 	/*
96 	 * We need both ACQUIRE (pairs with corresponding RELEASE in
97 	 * unlock() uncontended, or fastpath) and RELEASE (to publish
98 	 * the node fields we just initialised) semantics when updating
99 	 * the lock tail.
100 	 */
101 	old = atomic_xchg(&lock->tail, curr);
102 	if (old == OSQ_UNLOCKED_VAL)
103 		return true;
104 
105 	prev = decode_cpu(old);
106 	node->prev = prev;
107 	WRITE_ONCE(prev->next, node);
108 
109 	/*
110 	 * Normally @prev is untouchable after the above store; because at that
111 	 * moment unlock can proceed and wipe the node element from stack.
112 	 *
113 	 * However, since our nodes are static per-cpu storage, we're
114 	 * guaranteed their existence -- this allows us to apply
115 	 * cmpxchg in an attempt to undo our queueing.
116 	 */
117 
118 	while (!READ_ONCE(node->locked)) {
119 		/*
120 		 * If we need to reschedule bail... so we can block.
121 		 */
122 		if (need_resched())
123 			goto unqueue;
124 
125 		cpu_relax_lowlatency();
126 	}
127 	return true;
128 
129 unqueue:
130 	/*
131 	 * Step - A  -- stabilize @prev
132 	 *
133 	 * Undo our @prev->next assignment; this will make @prev's
134 	 * unlock()/unqueue() wait for a next pointer since @lock points to us
135 	 * (or later).
136 	 */
137 
138 	for (;;) {
139 		if (prev->next == node &&
140 		    cmpxchg(&prev->next, node, NULL) == node)
141 			break;
142 
143 		/*
144 		 * We can only fail the cmpxchg() racing against an unlock(),
145 		 * in which case we should observe @node->locked becomming
146 		 * true.
147 		 */
148 		if (smp_load_acquire(&node->locked))
149 			return true;
150 
151 		cpu_relax_lowlatency();
152 
153 		/*
154 		 * Or we race against a concurrent unqueue()'s step-B, in which
155 		 * case its step-C will write us a new @node->prev pointer.
156 		 */
157 		prev = READ_ONCE(node->prev);
158 	}
159 
160 	/*
161 	 * Step - B -- stabilize @next
162 	 *
163 	 * Similar to unlock(), wait for @node->next or move @lock from @node
164 	 * back to @prev.
165 	 */
166 
167 	next = osq_wait_next(lock, node, prev);
168 	if (!next)
169 		return false;
170 
171 	/*
172 	 * Step - C -- unlink
173 	 *
174 	 * @prev is stable because its still waiting for a new @prev->next
175 	 * pointer, @next is stable because our @node->next pointer is NULL and
176 	 * it will wait in Step-A.
177 	 */
178 
179 	WRITE_ONCE(next->prev, prev);
180 	WRITE_ONCE(prev->next, next);
181 
182 	return false;
183 }
184 
185 void osq_unlock(struct optimistic_spin_queue *lock)
186 {
187 	struct optimistic_spin_node *node, *next;
188 	int curr = encode_cpu(smp_processor_id());
189 
190 	/*
191 	 * Fast path for the uncontended case.
192 	 */
193 	if (likely(atomic_cmpxchg_release(&lock->tail, curr,
194 					  OSQ_UNLOCKED_VAL) == curr))
195 		return;
196 
197 	/*
198 	 * Second most likely case.
199 	 */
200 	node = this_cpu_ptr(&osq_node);
201 	next = xchg(&node->next, NULL);
202 	if (next) {
203 		WRITE_ONCE(next->locked, 1);
204 		return;
205 	}
206 
207 	next = osq_wait_next(lock, node, NULL);
208 	if (next)
209 		WRITE_ONCE(next->locked, 1);
210 }
211