xref: /linux/kernel/locking/osq_lock.c (revision 021bc4b9)
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 
15 struct optimistic_spin_node {
16 	struct optimistic_spin_node *next, *prev;
17 	int locked; /* 1 if lock acquired */
18 	int cpu; /* encoded CPU # + 1 value */
19 };
20 
21 static DEFINE_PER_CPU_SHARED_ALIGNED(struct optimistic_spin_node, osq_node);
22 
23 /*
24  * We use the value 0 to represent "no CPU", thus the encoded value
25  * will be the CPU number incremented by 1.
26  */
27 static inline int encode_cpu(int cpu_nr)
28 {
29 	return cpu_nr + 1;
30 }
31 
32 static inline int node_cpu(struct optimistic_spin_node *node)
33 {
34 	return node->cpu - 1;
35 }
36 
37 static inline struct optimistic_spin_node *decode_cpu(int encoded_cpu_val)
38 {
39 	int cpu_nr = encoded_cpu_val - 1;
40 
41 	return per_cpu_ptr(&osq_node, cpu_nr);
42 }
43 
44 /*
45  * Get a stable @node->next pointer, either for unlock() or unqueue() purposes.
46  * Can return NULL in case we were the last queued and we updated @lock instead.
47  *
48  * If osq_lock() is being cancelled there must be a previous node
49  * and 'old_cpu' is its CPU #.
50  * For osq_unlock() there is never a previous node and old_cpu is
51  * set to OSQ_UNLOCKED_VAL.
52  */
53 static inline struct optimistic_spin_node *
54 osq_wait_next(struct optimistic_spin_queue *lock,
55 	      struct optimistic_spin_node *node,
56 	      int old_cpu)
57 {
58 	int curr = encode_cpu(smp_processor_id());
59 
60 	for (;;) {
61 		if (atomic_read(&lock->tail) == curr &&
62 		    atomic_cmpxchg_acquire(&lock->tail, curr, old_cpu) == curr) {
63 			/*
64 			 * We were the last queued, we moved @lock back. @prev
65 			 * will now observe @lock and will complete its
66 			 * unlock()/unqueue().
67 			 */
68 			return NULL;
69 		}
70 
71 		/*
72 		 * We must xchg() the @node->next value, because if we were to
73 		 * leave it in, a concurrent unlock()/unqueue() from
74 		 * @node->next might complete Step-A and think its @prev is
75 		 * still valid.
76 		 *
77 		 * If the concurrent unlock()/unqueue() wins the race, we'll
78 		 * wait for either @lock to point to us, through its Step-B, or
79 		 * wait for a new @node->next from its Step-C.
80 		 */
81 		if (node->next) {
82 			struct optimistic_spin_node *next;
83 
84 			next = xchg(&node->next, NULL);
85 			if (next)
86 				return next;
87 		}
88 
89 		cpu_relax();
90 	}
91 }
92 
93 bool osq_lock(struct optimistic_spin_queue *lock)
94 {
95 	struct optimistic_spin_node *node = this_cpu_ptr(&osq_node);
96 	struct optimistic_spin_node *prev, *next;
97 	int curr = encode_cpu(smp_processor_id());
98 	int old;
99 
100 	node->locked = 0;
101 	node->next = NULL;
102 	node->cpu = curr;
103 
104 	/*
105 	 * We need both ACQUIRE (pairs with corresponding RELEASE in
106 	 * unlock() uncontended, or fastpath) and RELEASE (to publish
107 	 * the node fields we just initialised) semantics when updating
108 	 * the lock tail.
109 	 */
110 	old = atomic_xchg(&lock->tail, curr);
111 	if (old == OSQ_UNLOCKED_VAL)
112 		return true;
113 
114 	prev = decode_cpu(old);
115 	node->prev = prev;
116 
117 	/*
118 	 * osq_lock()			unqueue
119 	 *
120 	 * node->prev = prev		osq_wait_next()
121 	 * WMB				MB
122 	 * prev->next = node		next->prev = prev // unqueue-C
123 	 *
124 	 * Here 'node->prev' and 'next->prev' are the same variable and we need
125 	 * to ensure these stores happen in-order to avoid corrupting the list.
126 	 */
127 	smp_wmb();
128 
129 	WRITE_ONCE(prev->next, node);
130 
131 	/*
132 	 * Normally @prev is untouchable after the above store; because at that
133 	 * moment unlock can proceed and wipe the node element from stack.
134 	 *
135 	 * However, since our nodes are static per-cpu storage, we're
136 	 * guaranteed their existence -- this allows us to apply
137 	 * cmpxchg in an attempt to undo our queueing.
138 	 */
139 
140 	/*
141 	 * Wait to acquire the lock or cancellation. Note that need_resched()
142 	 * will come with an IPI, which will wake smp_cond_load_relaxed() if it
143 	 * is implemented with a monitor-wait. vcpu_is_preempted() relies on
144 	 * polling, be careful.
145 	 */
146 	if (smp_cond_load_relaxed(&node->locked, VAL || need_resched() ||
147 				  vcpu_is_preempted(node_cpu(node->prev))))
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 		/*
161 		 * cpu_relax() below implies a compiler barrier which would
162 		 * prevent this comparison being optimized away.
163 		 */
164 		if (data_race(prev->next) == node &&
165 		    cmpxchg(&prev->next, node, NULL) == node)
166 			break;
167 
168 		/*
169 		 * We can only fail the cmpxchg() racing against an unlock(),
170 		 * in which case we should observe @node->locked becoming
171 		 * true.
172 		 */
173 		if (smp_load_acquire(&node->locked))
174 			return true;
175 
176 		cpu_relax();
177 
178 		/*
179 		 * Or we race against a concurrent unqueue()'s step-B, in which
180 		 * case its step-C will write us a new @node->prev pointer.
181 		 */
182 		prev = READ_ONCE(node->prev);
183 	}
184 
185 	/*
186 	 * Step - B -- stabilize @next
187 	 *
188 	 * Similar to unlock(), wait for @node->next or move @lock from @node
189 	 * back to @prev.
190 	 */
191 
192 	next = osq_wait_next(lock, node, prev->cpu);
193 	if (!next)
194 		return false;
195 
196 	/*
197 	 * Step - C -- unlink
198 	 *
199 	 * @prev is stable because its still waiting for a new @prev->next
200 	 * pointer, @next is stable because our @node->next pointer is NULL and
201 	 * it will wait in Step-A.
202 	 */
203 
204 	WRITE_ONCE(next->prev, prev);
205 	WRITE_ONCE(prev->next, next);
206 
207 	return false;
208 }
209 
210 void osq_unlock(struct optimistic_spin_queue *lock)
211 {
212 	struct optimistic_spin_node *node, *next;
213 	int curr = encode_cpu(smp_processor_id());
214 
215 	/*
216 	 * Fast path for the uncontended case.
217 	 */
218 	if (likely(atomic_cmpxchg_release(&lock->tail, curr,
219 					  OSQ_UNLOCKED_VAL) == curr))
220 		return;
221 
222 	/*
223 	 * Second most likely case.
224 	 */
225 	node = this_cpu_ptr(&osq_node);
226 	next = xchg(&node->next, NULL);
227 	if (next) {
228 		WRITE_ONCE(next->locked, 1);
229 		return;
230 	}
231 
232 	next = osq_wait_next(lock, node, OSQ_UNLOCKED_VAL);
233 	if (next)
234 		WRITE_ONCE(next->locked, 1);
235 }
236