1 #include <linux/percpu.h>
2 #include <linux/sched.h>
3 #include <linux/osq_lock.h>
6 * An MCS like lock especially tailored for optimistic spinning for sleeping
7 * lock implementations (mutex, rwsem, etc).
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
13 static DEFINE_PER_CPU_SHARED_ALIGNED(struct optimistic_spin_node, osq_node);
16 * We use the value 0 to represent "no CPU", thus the encoded value
17 * will be the CPU number incremented by 1.
19 static inline int encode_cpu(int cpu_nr)
24 static inline int node_cpu(struct optimistic_spin_node *node)
29 static inline struct optimistic_spin_node *decode_cpu(int encoded_cpu_val)
31 int cpu_nr = encoded_cpu_val - 1;
33 return per_cpu_ptr(&osq_node, cpu_nr);
37 * Get a stable @node->next pointer, either for unlock() or unqueue() purposes.
38 * Can return NULL in case we were the last queued and we updated @lock instead.
40 static inline struct optimistic_spin_node *
41 osq_wait_next(struct optimistic_spin_queue *lock,
42 struct optimistic_spin_node *node,
43 struct optimistic_spin_node *prev)
45 struct optimistic_spin_node *next = NULL;
46 int curr = encode_cpu(smp_processor_id());
50 * If there is a prev node in queue, then the 'old' value will be
51 * the prev node's CPU #, else it's set to OSQ_UNLOCKED_VAL since if
52 * we're currently last in queue, then the queue will then become empty.
54 old = prev ? prev->cpu : OSQ_UNLOCKED_VAL;
57 if (atomic_read(&lock->tail) == curr &&
58 atomic_cmpxchg_acquire(&lock->tail, curr, old) == curr) {
60 * We were the last queued, we moved @lock back. @prev
61 * will now observe @lock and will complete its
68 * We must xchg() the @node->next value, because if we were to
69 * leave it in, a concurrent unlock()/unqueue() from
70 * @node->next might complete Step-A and think its @prev is
73 * If the concurrent unlock()/unqueue() wins the race, we'll
74 * wait for either @lock to point to us, through its Step-B, or
75 * wait for a new @node->next from its Step-C.
78 next = xchg(&node->next, NULL);
89 bool osq_lock(struct optimistic_spin_queue *lock)
91 struct optimistic_spin_node *node = this_cpu_ptr(&osq_node);
92 struct optimistic_spin_node *prev, *next;
93 int curr = encode_cpu(smp_processor_id());
101 * We need both ACQUIRE (pairs with corresponding RELEASE in
102 * unlock() uncontended, or fastpath) and RELEASE (to publish
103 * the node fields we just initialised) semantics when updating
106 old = atomic_xchg(&lock->tail, curr);
107 if (old == OSQ_UNLOCKED_VAL)
110 prev = decode_cpu(old);
112 WRITE_ONCE(prev->next, node);
115 * Normally @prev is untouchable after the above store; because at that
116 * moment unlock can proceed and wipe the node element from stack.
118 * However, since our nodes are static per-cpu storage, we're
119 * guaranteed their existence -- this allows us to apply
120 * cmpxchg in an attempt to undo our queueing.
123 while (!READ_ONCE(node->locked)) {
125 * If we need to reschedule bail... so we can block.
126 * Use vcpu_is_preempted() to avoid waiting for a preempted
129 if (need_resched() || vcpu_is_preempted(node_cpu(node->prev)))
138 * Step - A -- stabilize @prev
140 * Undo our @prev->next assignment; this will make @prev's
141 * unlock()/unqueue() wait for a next pointer since @lock points to us
146 if (prev->next == node &&
147 cmpxchg(&prev->next, node, NULL) == node)
151 * We can only fail the cmpxchg() racing against an unlock(),
152 * in which case we should observe @node->locked becomming
155 if (smp_load_acquire(&node->locked))
161 * Or we race against a concurrent unqueue()'s step-B, in which
162 * case its step-C will write us a new @node->prev pointer.
164 prev = READ_ONCE(node->prev);
168 * Step - B -- stabilize @next
170 * Similar to unlock(), wait for @node->next or move @lock from @node
174 next = osq_wait_next(lock, node, prev);
181 * @prev is stable because its still waiting for a new @prev->next
182 * pointer, @next is stable because our @node->next pointer is NULL and
183 * it will wait in Step-A.
186 WRITE_ONCE(next->prev, prev);
187 WRITE_ONCE(prev->next, next);
192 void osq_unlock(struct optimistic_spin_queue *lock)
194 struct optimistic_spin_node *node, *next;
195 int curr = encode_cpu(smp_processor_id());
198 * Fast path for the uncontended case.
200 if (likely(atomic_cmpxchg_release(&lock->tail, curr,
201 OSQ_UNLOCKED_VAL) == curr))
205 * Second most likely case.
207 node = this_cpu_ptr(&osq_node);
208 next = xchg(&node->next, NULL);
210 WRITE_ONCE(next->locked, 1);
214 next = osq_wait_next(lock, node, NULL);
216 WRITE_ONCE(next->locked, 1);