xref: /openbmc/linux/arch/x86/include/asm/spinlock.h (revision 9d749629)
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 support up to 2^16 CPUs.
16  *
17  * (the type definitions are in asm/spinlock_types.h)
18  */
19 
20 #ifdef CONFIG_X86_32
21 # define LOCK_PTR_REG "a"
22 #else
23 # define LOCK_PTR_REG "D"
24 #endif
25 
26 #if defined(CONFIG_X86_32) && \
27 	(defined(CONFIG_X86_OOSTORE) || defined(CONFIG_X86_PPRO_FENCE))
28 /*
29  * On PPro SMP or if we are using OOSTORE, we use a locked operation to unlock
30  * (PPro errata 66, 92)
31  */
32 # define UNLOCK_LOCK_PREFIX LOCK_PREFIX
33 #else
34 # define UNLOCK_LOCK_PREFIX
35 #endif
36 
37 /*
38  * Ticket locks are conceptually two parts, one indicating the current head of
39  * the queue, and the other indicating the current tail. The lock is acquired
40  * by atomically noting the tail and incrementing it by one (thus adding
41  * ourself to the queue and noting our position), then waiting until the head
42  * becomes equal to the the initial value of the tail.
43  *
44  * We use an xadd covering *both* parts of the lock, to increment the tail and
45  * also load the position of the head, which takes care of memory ordering
46  * issues and should be optimal for the uncontended case. Note the tail must be
47  * in the high part, because a wide xadd increment of the low part would carry
48  * up and contaminate the high part.
49  */
50 static __always_inline void __ticket_spin_lock(arch_spinlock_t *lock)
51 {
52 	register struct __raw_tickets inc = { .tail = 1 };
53 
54 	inc = xadd(&lock->tickets, inc);
55 
56 	for (;;) {
57 		if (inc.head == inc.tail)
58 			break;
59 		cpu_relax();
60 		inc.head = ACCESS_ONCE(lock->tickets.head);
61 	}
62 	barrier();		/* make sure nothing creeps before the lock is taken */
63 }
64 
65 static __always_inline int __ticket_spin_trylock(arch_spinlock_t *lock)
66 {
67 	arch_spinlock_t old, new;
68 
69 	old.tickets = ACCESS_ONCE(lock->tickets);
70 	if (old.tickets.head != old.tickets.tail)
71 		return 0;
72 
73 	new.head_tail = old.head_tail + (1 << TICKET_SHIFT);
74 
75 	/* cmpxchg is a full barrier, so nothing can move before it */
76 	return cmpxchg(&lock->head_tail, old.head_tail, new.head_tail) == old.head_tail;
77 }
78 
79 static __always_inline void __ticket_spin_unlock(arch_spinlock_t *lock)
80 {
81 	__add(&lock->tickets.head, 1, UNLOCK_LOCK_PREFIX);
82 }
83 
84 static inline int __ticket_spin_is_locked(arch_spinlock_t *lock)
85 {
86 	struct __raw_tickets tmp = ACCESS_ONCE(lock->tickets);
87 
88 	return tmp.tail != tmp.head;
89 }
90 
91 static inline int __ticket_spin_is_contended(arch_spinlock_t *lock)
92 {
93 	struct __raw_tickets tmp = ACCESS_ONCE(lock->tickets);
94 
95 	return (__ticket_t)(tmp.tail - tmp.head) > 1;
96 }
97 
98 #ifndef CONFIG_PARAVIRT_SPINLOCKS
99 
100 static inline int arch_spin_is_locked(arch_spinlock_t *lock)
101 {
102 	return __ticket_spin_is_locked(lock);
103 }
104 
105 static inline int arch_spin_is_contended(arch_spinlock_t *lock)
106 {
107 	return __ticket_spin_is_contended(lock);
108 }
109 #define arch_spin_is_contended	arch_spin_is_contended
110 
111 static __always_inline void arch_spin_lock(arch_spinlock_t *lock)
112 {
113 	__ticket_spin_lock(lock);
114 }
115 
116 static __always_inline int arch_spin_trylock(arch_spinlock_t *lock)
117 {
118 	return __ticket_spin_trylock(lock);
119 }
120 
121 static __always_inline void arch_spin_unlock(arch_spinlock_t *lock)
122 {
123 	__ticket_spin_unlock(lock);
124 }
125 
126 static __always_inline void arch_spin_lock_flags(arch_spinlock_t *lock,
127 						  unsigned long flags)
128 {
129 	arch_spin_lock(lock);
130 }
131 
132 #endif	/* CONFIG_PARAVIRT_SPINLOCKS */
133 
134 static inline void arch_spin_unlock_wait(arch_spinlock_t *lock)
135 {
136 	while (arch_spin_is_locked(lock))
137 		cpu_relax();
138 }
139 
140 /*
141  * Read-write spinlocks, allowing multiple readers
142  * but only one writer.
143  *
144  * NOTE! it is quite common to have readers in interrupts
145  * but no interrupt writers. For those circumstances we
146  * can "mix" irq-safe locks - any writer needs to get a
147  * irq-safe write-lock, but readers can get non-irqsafe
148  * read-locks.
149  *
150  * On x86, we implement read-write locks as a 32-bit counter
151  * with the high bit (sign) being the "contended" bit.
152  */
153 
154 /**
155  * read_can_lock - would read_trylock() succeed?
156  * @lock: the rwlock in question.
157  */
158 static inline int arch_read_can_lock(arch_rwlock_t *lock)
159 {
160 	return lock->lock > 0;
161 }
162 
163 /**
164  * write_can_lock - would write_trylock() succeed?
165  * @lock: the rwlock in question.
166  */
167 static inline int arch_write_can_lock(arch_rwlock_t *lock)
168 {
169 	return lock->write == WRITE_LOCK_CMP;
170 }
171 
172 static inline void arch_read_lock(arch_rwlock_t *rw)
173 {
174 	asm volatile(LOCK_PREFIX READ_LOCK_SIZE(dec) " (%0)\n\t"
175 		     "jns 1f\n"
176 		     "call __read_lock_failed\n\t"
177 		     "1:\n"
178 		     ::LOCK_PTR_REG (rw) : "memory");
179 }
180 
181 static inline void arch_write_lock(arch_rwlock_t *rw)
182 {
183 	asm volatile(LOCK_PREFIX WRITE_LOCK_SUB(%1) "(%0)\n\t"
184 		     "jz 1f\n"
185 		     "call __write_lock_failed\n\t"
186 		     "1:\n"
187 		     ::LOCK_PTR_REG (&rw->write), "i" (RW_LOCK_BIAS)
188 		     : "memory");
189 }
190 
191 static inline int arch_read_trylock(arch_rwlock_t *lock)
192 {
193 	READ_LOCK_ATOMIC(t) *count = (READ_LOCK_ATOMIC(t) *)lock;
194 
195 	if (READ_LOCK_ATOMIC(dec_return)(count) >= 0)
196 		return 1;
197 	READ_LOCK_ATOMIC(inc)(count);
198 	return 0;
199 }
200 
201 static inline int arch_write_trylock(arch_rwlock_t *lock)
202 {
203 	atomic_t *count = (atomic_t *)&lock->write;
204 
205 	if (atomic_sub_and_test(WRITE_LOCK_CMP, count))
206 		return 1;
207 	atomic_add(WRITE_LOCK_CMP, count);
208 	return 0;
209 }
210 
211 static inline void arch_read_unlock(arch_rwlock_t *rw)
212 {
213 	asm volatile(LOCK_PREFIX READ_LOCK_SIZE(inc) " %0"
214 		     :"+m" (rw->lock) : : "memory");
215 }
216 
217 static inline void arch_write_unlock(arch_rwlock_t *rw)
218 {
219 	asm volatile(LOCK_PREFIX WRITE_LOCK_ADD(%1) "%0"
220 		     : "+m" (rw->write) : "i" (RW_LOCK_BIAS) : "memory");
221 }
222 
223 #define arch_read_lock_flags(lock, flags) arch_read_lock(lock)
224 #define arch_write_lock_flags(lock, flags) arch_write_lock(lock)
225 
226 #undef READ_LOCK_SIZE
227 #undef READ_LOCK_ATOMIC
228 #undef WRITE_LOCK_ADD
229 #undef WRITE_LOCK_SUB
230 #undef WRITE_LOCK_CMP
231 
232 #define arch_spin_relax(lock)	cpu_relax()
233 #define arch_read_relax(lock)	cpu_relax()
234 #define arch_write_relax(lock)	cpu_relax()
235 
236 /* The {read|write|spin}_lock() on x86 are full memory barriers. */
237 static inline void smp_mb__after_lock(void) { }
238 #define ARCH_HAS_SMP_MB_AFTER_LOCK
239 
240 #endif /* _ASM_X86_SPINLOCK_H */
241