GNU Linux-libre 6.1.90-gnu
[releases.git] / arch / ia64 / include / asm / spinlock.h
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _ASM_IA64_SPINLOCK_H
3 #define _ASM_IA64_SPINLOCK_H
4
5 /*
6  * Copyright (C) 1998-2003 Hewlett-Packard Co
7  *      David Mosberger-Tang <davidm@hpl.hp.com>
8  * Copyright (C) 1999 Walt Drummond <drummond@valinux.com>
9  *
10  * This file is used for SMP configurations only.
11  */
12
13 #include <linux/compiler.h>
14 #include <linux/kernel.h>
15 #include <linux/bitops.h>
16
17 #include <linux/atomic.h>
18 #include <asm/intrinsics.h>
19 #include <asm/barrier.h>
20 #include <asm/processor.h>
21
22 #define arch_spin_lock_init(x)                  ((x)->lock = 0)
23
24 /*
25  * Ticket locks are conceptually two parts, one indicating the current head of
26  * the queue, and the other indicating the current tail. The lock is acquired
27  * by atomically noting the tail and incrementing it by one (thus adding
28  * ourself to the queue and noting our position), then waiting until the head
29  * becomes equal to the initial value of the tail.
30  * The pad bits in the middle are used to prevent the next_ticket number
31  * overflowing into the now_serving number.
32  *
33  *   31             17  16    15  14                    0
34  *  +----------------------------------------------------+
35  *  |  now_serving     | padding |   next_ticket         |
36  *  +----------------------------------------------------+
37  */
38
39 #define TICKET_SHIFT    17
40 #define TICKET_BITS     15
41 #define TICKET_MASK     ((1 << TICKET_BITS) - 1)
42
43 static __always_inline void __ticket_spin_lock(arch_spinlock_t *lock)
44 {
45         int     *p = (int *)&lock->lock, ticket, serve;
46
47         ticket = ia64_fetchadd(1, p, acq);
48
49         if (!(((ticket >> TICKET_SHIFT) ^ ticket) & TICKET_MASK))
50                 return;
51
52         ia64_invala();
53
54         for (;;) {
55                 asm volatile ("ld4.c.nc %0=[%1]" : "=r"(serve) : "r"(p) : "memory");
56
57                 if (!(((serve >> TICKET_SHIFT) ^ ticket) & TICKET_MASK))
58                         return;
59                 cpu_relax();
60         }
61 }
62
63 static __always_inline int __ticket_spin_trylock(arch_spinlock_t *lock)
64 {
65         int tmp = READ_ONCE(lock->lock);
66
67         if (!(((tmp >> TICKET_SHIFT) ^ tmp) & TICKET_MASK))
68                 return ia64_cmpxchg(acq, &lock->lock, tmp, tmp + 1, sizeof (tmp)) == tmp;
69         return 0;
70 }
71
72 static __always_inline void __ticket_spin_unlock(arch_spinlock_t *lock)
73 {
74         unsigned short  *p = (unsigned short *)&lock->lock + 1, tmp;
75
76         /* This could be optimised with ARCH_HAS_MMIOWB */
77         mmiowb();
78         asm volatile ("ld2.bias %0=[%1]" : "=r"(tmp) : "r"(p));
79         WRITE_ONCE(*p, (tmp + 2) & ~1);
80 }
81
82 static inline int __ticket_spin_is_locked(arch_spinlock_t *lock)
83 {
84         long tmp = READ_ONCE(lock->lock);
85
86         return !!(((tmp >> TICKET_SHIFT) ^ tmp) & TICKET_MASK);
87 }
88
89 static inline int __ticket_spin_is_contended(arch_spinlock_t *lock)
90 {
91         long tmp = READ_ONCE(lock->lock);
92
93         return ((tmp - (tmp >> TICKET_SHIFT)) & TICKET_MASK) > 1;
94 }
95
96 static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock)
97 {
98         return !(((lock.lock >> TICKET_SHIFT) ^ lock.lock) & TICKET_MASK);
99 }
100
101 static inline int arch_spin_is_locked(arch_spinlock_t *lock)
102 {
103         return __ticket_spin_is_locked(lock);
104 }
105
106 static inline int arch_spin_is_contended(arch_spinlock_t *lock)
107 {
108         return __ticket_spin_is_contended(lock);
109 }
110 #define arch_spin_is_contended  arch_spin_is_contended
111
112 static __always_inline void arch_spin_lock(arch_spinlock_t *lock)
113 {
114         __ticket_spin_lock(lock);
115 }
116
117 static __always_inline int arch_spin_trylock(arch_spinlock_t *lock)
118 {
119         return __ticket_spin_trylock(lock);
120 }
121
122 static __always_inline void arch_spin_unlock(arch_spinlock_t *lock)
123 {
124         __ticket_spin_unlock(lock);
125 }
126
127 #ifdef ASM_SUPPORTED
128
129 static __always_inline void
130 arch_read_lock(arch_rwlock_t *lock)
131 {
132         unsigned long flags = 0;
133
134         __asm__ __volatile__ (
135                 "tbit.nz p6, p0 = %1,%2\n"
136                 "br.few 3f\n"
137                 "1:\n"
138                 "fetchadd4.rel r2 = [%0], -1;;\n"
139                 "(p6) ssm psr.i\n"
140                 "2:\n"
141                 "hint @pause\n"
142                 "ld4 r2 = [%0];;\n"
143                 "cmp4.lt p7,p0 = r2, r0\n"
144                 "(p7) br.cond.spnt.few 2b\n"
145                 "(p6) rsm psr.i\n"
146                 ";;\n"
147                 "3:\n"
148                 "fetchadd4.acq r2 = [%0], 1;;\n"
149                 "cmp4.lt p7,p0 = r2, r0\n"
150                 "(p7) br.cond.spnt.few 1b\n"
151                 : : "r"(lock), "r"(flags), "i"(IA64_PSR_I_BIT)
152                 : "p6", "p7", "r2", "memory");
153 }
154
155 #else /* !ASM_SUPPORTED */
156
157 #define arch_read_lock(rw)                                                              \
158 do {                                                                                    \
159         arch_rwlock_t *__read_lock_ptr = (rw);                                          \
160                                                                                         \
161         while (unlikely(ia64_fetchadd(1, (int *) __read_lock_ptr, acq) < 0)) {          \
162                 ia64_fetchadd(-1, (int *) __read_lock_ptr, rel);                        \
163                 while (*(volatile int *)__read_lock_ptr < 0)                            \
164                         cpu_relax();                                                    \
165         }                                                                               \
166 } while (0)
167
168 #endif /* !ASM_SUPPORTED */
169
170 #define arch_read_unlock(rw)                                    \
171 do {                                                            \
172         arch_rwlock_t *__read_lock_ptr = (rw);                  \
173         ia64_fetchadd(-1, (int *) __read_lock_ptr, rel);        \
174 } while (0)
175
176 #ifdef ASM_SUPPORTED
177
178 static __always_inline void
179 arch_write_lock(arch_rwlock_t *lock)
180 {
181         unsigned long flags = 0;
182
183         __asm__ __volatile__ (
184                 "tbit.nz p6, p0 = %1, %2\n"
185                 "mov ar.ccv = r0\n"
186                 "dep r29 = -1, r0, 31, 1\n"
187                 "br.few 3f;;\n"
188                 "1:\n"
189                 "(p6) ssm psr.i\n"
190                 "2:\n"
191                 "hint @pause\n"
192                 "ld4 r2 = [%0];;\n"
193                 "cmp4.eq p0,p7 = r0, r2\n"
194                 "(p7) br.cond.spnt.few 2b\n"
195                 "(p6) rsm psr.i\n"
196                 ";;\n"
197                 "3:\n"
198                 "cmpxchg4.acq r2 = [%0], r29, ar.ccv;;\n"
199                 "cmp4.eq p0,p7 = r0, r2\n"
200                 "(p7) br.cond.spnt.few 1b;;\n"
201                 : : "r"(lock), "r"(flags), "i"(IA64_PSR_I_BIT)
202                 : "ar.ccv", "p6", "p7", "r2", "r29", "memory");
203 }
204
205 #define arch_write_trylock(rw)                                                  \
206 ({                                                                              \
207         register long result;                                                   \
208                                                                                 \
209         __asm__ __volatile__ (                                                  \
210                 "mov ar.ccv = r0\n"                                             \
211                 "dep r29 = -1, r0, 31, 1;;\n"                                   \
212                 "cmpxchg4.acq %0 = [%1], r29, ar.ccv\n"                         \
213                 : "=r"(result) : "r"(rw) : "ar.ccv", "r29", "memory");          \
214         (result == 0);                                                          \
215 })
216
217 static inline void arch_write_unlock(arch_rwlock_t *x)
218 {
219         u8 *y = (u8 *)x;
220         barrier();
221         asm volatile ("st1.rel.nta [%0] = r0\n\t" :: "r"(y+3) : "memory" );
222 }
223
224 #else /* !ASM_SUPPORTED */
225
226 #define arch_write_lock(l)                                                              \
227 ({                                                                                      \
228         __u64 ia64_val, ia64_set_val = ia64_dep_mi(-1, 0, 31, 1);                       \
229         __u32 *ia64_write_lock_ptr = (__u32 *) (l);                                     \
230         do {                                                                            \
231                 while (*ia64_write_lock_ptr)                                            \
232                         ia64_barrier();                                                 \
233                 ia64_val = ia64_cmpxchg4_acq(ia64_write_lock_ptr, ia64_set_val, 0);     \
234         } while (ia64_val);                                                             \
235 })
236
237 #define arch_write_trylock(rw)                                          \
238 ({                                                                      \
239         __u64 ia64_val;                                                 \
240         __u64 ia64_set_val = ia64_dep_mi(-1, 0, 31,1);                  \
241         ia64_val = ia64_cmpxchg4_acq((__u32 *)(rw), ia64_set_val, 0);   \
242         (ia64_val == 0);                                                \
243 })
244
245 static inline void arch_write_unlock(arch_rwlock_t *x)
246 {
247         barrier();
248         x->write_lock = 0;
249 }
250
251 #endif /* !ASM_SUPPORTED */
252
253 static inline int arch_read_trylock(arch_rwlock_t *x)
254 {
255         union {
256                 arch_rwlock_t lock;
257                 __u32 word;
258         } old, new;
259         old.lock = new.lock = *x;
260         old.lock.write_lock = new.lock.write_lock = 0;
261         ++new.lock.read_counter;
262         return (u32)ia64_cmpxchg4_acq((__u32 *)(x), new.word, old.word) == old.word;
263 }
264
265 #endif /*  _ASM_IA64_SPINLOCK_H */