Merge branch 'for-linus' into for-next
[linux-2.6-microblaze.git] / kernel / locking / osq_lock.c
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 }