xref: /openbmc/linux/kernel/locking/osq_lock.c (revision 2f61c664)
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 	/*
138 	 * Wait to acquire the lock or cancellation. Note that need_resched()
139 	 * will come with an IPI, which will wake smp_cond_load_relaxed() if it
140 	 * is implemented with a monitor-wait. vcpu_is_preempted() relies on
141 	 * polling, be careful.
142 	 */
143 	if (smp_cond_load_relaxed(&node->locked, VAL || need_resched() ||
144 				  vcpu_is_preempted(node_cpu(node->prev))))
145 		return true;
146 
147 	/* unqueue */
148 	/*
149 	 * Step - A  -- stabilize @prev
150 	 *
151 	 * Undo our @prev->next assignment; this will make @prev's
152 	 * unlock()/unqueue() wait for a next pointer since @lock points to us
153 	 * (or later).
154 	 */
155 
156 	for (;;) {
157 		/*
158 		 * cpu_relax() below implies a compiler barrier which would
159 		 * prevent this comparison being optimized away.
160 		 */
161 		if (data_race(prev->next) == node &&
162 		    cmpxchg(&prev->next, node, NULL) == node)
163 			break;
164 
165 		/*
166 		 * We can only fail the cmpxchg() racing against an unlock(),
167 		 * in which case we should observe @node->locked becoming
168 		 * true.
169 		 */
170 		if (smp_load_acquire(&node->locked))
171 			return true;
172 
173 		cpu_relax();
174 
175 		/*
176 		 * Or we race against a concurrent unqueue()'s step-B, in which
177 		 * case its step-C will write us a new @node->prev pointer.
178 		 */
179 		prev = READ_ONCE(node->prev);
180 	}
181 
182 	/*
183 	 * Step - B -- stabilize @next
184 	 *
185 	 * Similar to unlock(), wait for @node->next or move @lock from @node
186 	 * back to @prev.
187 	 */
188 
189 	next = osq_wait_next(lock, node, prev);
190 	if (!next)
191 		return false;
192 
193 	/*
194 	 * Step - C -- unlink
195 	 *
196 	 * @prev is stable because its still waiting for a new @prev->next
197 	 * pointer, @next is stable because our @node->next pointer is NULL and
198 	 * it will wait in Step-A.
199 	 */
200 
201 	WRITE_ONCE(next->prev, prev);
202 	WRITE_ONCE(prev->next, next);
203 
204 	return false;
205 }
206 
207 void osq_unlock(struct optimistic_spin_queue *lock)
208 {
209 	struct optimistic_spin_node *node, *next;
210 	int curr = encode_cpu(smp_processor_id());
211 
212 	/*
213 	 * Fast path for the uncontended case.
214 	 */
215 	if (likely(atomic_cmpxchg_release(&lock->tail, curr,
216 					  OSQ_UNLOCKED_VAL) == curr))
217 		return;
218 
219 	/*
220 	 * Second most likely case.
221 	 */
222 	node = this_cpu_ptr(&osq_node);
223 	next = xchg(&node->next, NULL);
224 	if (next) {
225 		WRITE_ONCE(next->locked, 1);
226 		return;
227 	}
228 
229 	next = osq_wait_next(lock, node, NULL);
230 	if (next)
231 		WRITE_ONCE(next->locked, 1);
232 }
233