xref: /openbmc/linux/arch/x86/include/asm/spinlock.h (revision 63dc02bd)
1 #ifndef _ASM_X86_SPINLOCK_H
2 #define _ASM_X86_SPINLOCK_H
3 
4 #include <linux/atomic.h>
5 #include <asm/page.h>
6 #include <asm/processor.h>
7 #include <linux/compiler.h>
8 #include <asm/paravirt.h>
9 /*
10  * Your basic SMP spinlocks, allowing only a single CPU anywhere
11  *
12  * Simple spin lock operations.  There are two variants, one clears IRQ's
13  * on the local processor, one does not.
14  *
15  * These are fair FIFO ticket locks, which are currently limited to 256
16  * CPUs.
17  *
18  * (the type definitions are in asm/spinlock_types.h)
19  */
20 
21 #ifdef CONFIG_X86_32
22 # define LOCK_PTR_REG "a"
23 # define REG_PTR_MODE "k"
24 #else
25 # define LOCK_PTR_REG "D"
26 # define REG_PTR_MODE "q"
27 #endif
28 
29 #if defined(CONFIG_X86_32) && \
30 	(defined(CONFIG_X86_OOSTORE) || defined(CONFIG_X86_PPRO_FENCE))
31 /*
32  * On PPro SMP or if we are using OOSTORE, we use a locked operation to unlock
33  * (PPro errata 66, 92)
34  */
35 # define UNLOCK_LOCK_PREFIX LOCK_PREFIX
36 #else
37 # define UNLOCK_LOCK_PREFIX
38 #endif
39 
40 /*
41  * Ticket locks are conceptually two parts, one indicating the current head of
42  * the queue, and the other indicating the current tail. The lock is acquired
43  * by atomically noting the tail and incrementing it by one (thus adding
44  * ourself to the queue and noting our position), then waiting until the head
45  * becomes equal to the the initial value of the tail.
46  *
47  * We use an xadd covering *both* parts of the lock, to increment the tail and
48  * also load the position of the head, which takes care of memory ordering
49  * issues and should be optimal for the uncontended case. Note the tail must be
50  * in the high part, because a wide xadd increment of the low part would carry
51  * up and contaminate the high part.
52  */
53 static __always_inline void __ticket_spin_lock(arch_spinlock_t *lock)
54 {
55 	register struct __raw_tickets inc = { .tail = 1 };
56 
57 	inc = xadd(&lock->tickets, inc);
58 
59 	for (;;) {
60 		if (inc.head == inc.tail)
61 			break;
62 		cpu_relax();
63 		inc.head = ACCESS_ONCE(lock->tickets.head);
64 	}
65 	barrier();		/* make sure nothing creeps before the lock is taken */
66 }
67 
68 static __always_inline int __ticket_spin_trylock(arch_spinlock_t *lock)
69 {
70 	arch_spinlock_t old, new;
71 
72 	old.tickets = ACCESS_ONCE(lock->tickets);
73 	if (old.tickets.head != old.tickets.tail)
74 		return 0;
75 
76 	new.head_tail = old.head_tail + (1 << TICKET_SHIFT);
77 
78 	/* cmpxchg is a full barrier, so nothing can move before it */
79 	return cmpxchg(&lock->head_tail, old.head_tail, new.head_tail) == old.head_tail;
80 }
81 
82 static __always_inline void __ticket_spin_unlock(arch_spinlock_t *lock)
83 {
84 	__add(&lock->tickets.head, 1, UNLOCK_LOCK_PREFIX);
85 }
86 
87 static inline int __ticket_spin_is_locked(arch_spinlock_t *lock)
88 {
89 	struct __raw_tickets tmp = ACCESS_ONCE(lock->tickets);
90 
91 	return tmp.tail != tmp.head;
92 }
93 
94 static inline int __ticket_spin_is_contended(arch_spinlock_t *lock)
95 {
96 	struct __raw_tickets tmp = ACCESS_ONCE(lock->tickets);
97 
98 	return (__ticket_t)(tmp.tail - tmp.head) > 1;
99 }
100 
101 #ifndef CONFIG_PARAVIRT_SPINLOCKS
102 
103 static inline int arch_spin_is_locked(arch_spinlock_t *lock)
104 {
105 	return __ticket_spin_is_locked(lock);
106 }
107 
108 static inline int arch_spin_is_contended(arch_spinlock_t *lock)
109 {
110 	return __ticket_spin_is_contended(lock);
111 }
112 #define arch_spin_is_contended	arch_spin_is_contended
113 
114 static __always_inline void arch_spin_lock(arch_spinlock_t *lock)
115 {
116 	__ticket_spin_lock(lock);
117 }
118 
119 static __always_inline int arch_spin_trylock(arch_spinlock_t *lock)
120 {
121 	return __ticket_spin_trylock(lock);
122 }
123 
124 static __always_inline void arch_spin_unlock(arch_spinlock_t *lock)
125 {
126 	__ticket_spin_unlock(lock);
127 }
128 
129 static __always_inline void arch_spin_lock_flags(arch_spinlock_t *lock,
130 						  unsigned long flags)
131 {
132 	arch_spin_lock(lock);
133 }
134 
135 #endif	/* CONFIG_PARAVIRT_SPINLOCKS */
136 
137 static inline void arch_spin_unlock_wait(arch_spinlock_t *lock)
138 {
139 	while (arch_spin_is_locked(lock))
140 		cpu_relax();
141 }
142 
143 /*
144  * Read-write spinlocks, allowing multiple readers
145  * but only one writer.
146  *
147  * NOTE! it is quite common to have readers in interrupts
148  * but no interrupt writers. For those circumstances we
149  * can "mix" irq-safe locks - any writer needs to get a
150  * irq-safe write-lock, but readers can get non-irqsafe
151  * read-locks.
152  *
153  * On x86, we implement read-write locks as a 32-bit counter
154  * with the high bit (sign) being the "contended" bit.
155  */
156 
157 /**
158  * read_can_lock - would read_trylock() succeed?
159  * @lock: the rwlock in question.
160  */
161 static inline int arch_read_can_lock(arch_rwlock_t *lock)
162 {
163 	return lock->lock > 0;
164 }
165 
166 /**
167  * write_can_lock - would write_trylock() succeed?
168  * @lock: the rwlock in question.
169  */
170 static inline int arch_write_can_lock(arch_rwlock_t *lock)
171 {
172 	return lock->write == WRITE_LOCK_CMP;
173 }
174 
175 static inline void arch_read_lock(arch_rwlock_t *rw)
176 {
177 	asm volatile(LOCK_PREFIX READ_LOCK_SIZE(dec) " (%0)\n\t"
178 		     "jns 1f\n"
179 		     "call __read_lock_failed\n\t"
180 		     "1:\n"
181 		     ::LOCK_PTR_REG (rw) : "memory");
182 }
183 
184 static inline void arch_write_lock(arch_rwlock_t *rw)
185 {
186 	asm volatile(LOCK_PREFIX WRITE_LOCK_SUB(%1) "(%0)\n\t"
187 		     "jz 1f\n"
188 		     "call __write_lock_failed\n\t"
189 		     "1:\n"
190 		     ::LOCK_PTR_REG (&rw->write), "i" (RW_LOCK_BIAS)
191 		     : "memory");
192 }
193 
194 static inline int arch_read_trylock(arch_rwlock_t *lock)
195 {
196 	READ_LOCK_ATOMIC(t) *count = (READ_LOCK_ATOMIC(t) *)lock;
197 
198 	if (READ_LOCK_ATOMIC(dec_return)(count) >= 0)
199 		return 1;
200 	READ_LOCK_ATOMIC(inc)(count);
201 	return 0;
202 }
203 
204 static inline int arch_write_trylock(arch_rwlock_t *lock)
205 {
206 	atomic_t *count = (atomic_t *)&lock->write;
207 
208 	if (atomic_sub_and_test(WRITE_LOCK_CMP, count))
209 		return 1;
210 	atomic_add(WRITE_LOCK_CMP, count);
211 	return 0;
212 }
213 
214 static inline void arch_read_unlock(arch_rwlock_t *rw)
215 {
216 	asm volatile(LOCK_PREFIX READ_LOCK_SIZE(inc) " %0"
217 		     :"+m" (rw->lock) : : "memory");
218 }
219 
220 static inline void arch_write_unlock(arch_rwlock_t *rw)
221 {
222 	asm volatile(LOCK_PREFIX WRITE_LOCK_ADD(%1) "%0"
223 		     : "+m" (rw->write) : "i" (RW_LOCK_BIAS) : "memory");
224 }
225 
226 #define arch_read_lock_flags(lock, flags) arch_read_lock(lock)
227 #define arch_write_lock_flags(lock, flags) arch_write_lock(lock)
228 
229 #undef READ_LOCK_SIZE
230 #undef READ_LOCK_ATOMIC
231 #undef WRITE_LOCK_ADD
232 #undef WRITE_LOCK_SUB
233 #undef WRITE_LOCK_CMP
234 
235 #define arch_spin_relax(lock)	cpu_relax()
236 #define arch_read_relax(lock)	cpu_relax()
237 #define arch_write_relax(lock)	cpu_relax()
238 
239 /* The {read|write|spin}_lock() on x86 are full memory barriers. */
240 static inline void smp_mb__after_lock(void) { }
241 #define ARCH_HAS_SMP_MB_AFTER_LOCK
242 
243 #endif /* _ASM_X86_SPINLOCK_H */
244