1/*
2 * Copyright 2010 Tilera Corporation. All Rights Reserved.
3 *
4 *   This program is free software; you can redistribute it and/or
5 *   modify it under the terms of the GNU General Public License
6 *   as published by the Free Software Foundation, version 2.
7 *
8 *   This program is distributed in the hope that it will be useful, but
9 *   WITHOUT ANY WARRANTY; without even the implied warranty of
10 *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
11 *   NON INFRINGEMENT.  See the GNU General Public License for
12 *   more details.
13 *
14 * These routines make two important assumptions:
15 *
16 * 1. atomic_t is really an int and can be freely cast back and forth
17 *    (validated in __init_atomic_per_cpu).
18 *
19 * 2. userspace uses sys_cmpxchg() for all atomic operations, thus using
20 *    the same locking convention that all the kernel atomic routines use.
21 */
22
23#ifndef _ASM_TILE_FUTEX_H
24#define _ASM_TILE_FUTEX_H
25
26#ifndef __ASSEMBLY__
27
28#include <linux/futex.h>
29#include <linux/uaccess.h>
30#include <linux/errno.h>
31#include <asm/atomic.h>
32
33/*
34 * Support macros for futex operations.  Do not use these macros directly.
35 * They assume "ret", "val", "oparg", and "uaddr" in the lexical context.
36 * __futex_cmpxchg() additionally assumes "oldval".
37 */
38
39#ifdef __tilegx__
40
41#define __futex_asm(OP) \
42	asm("1: {" #OP " %1, %3, %4; movei %0, 0 }\n"		\
43	    ".pushsection .fixup,\"ax\"\n"			\
44	    "0: { movei %0, %5; j 9f }\n"			\
45	    ".section __ex_table,\"a\"\n"			\
46	    ".align 8\n"					\
47	    ".quad 1b, 0b\n"					\
48	    ".popsection\n"					\
49	    "9:"						\
50	    : "=r" (ret), "=r" (val), "+m" (*(uaddr))		\
51	    : "r" (uaddr), "r" (oparg), "i" (-EFAULT))
52
53#define __futex_set() __futex_asm(exch4)
54#define __futex_add() __futex_asm(fetchadd4)
55#define __futex_or() __futex_asm(fetchor4)
56#define __futex_andn() ({ oparg = ~oparg; __futex_asm(fetchand4); })
57#define __futex_cmpxchg() \
58	({ __insn_mtspr(SPR_CMPEXCH_VALUE, oldval); __futex_asm(cmpexch4); })
59
60#define __futex_xor()						\
61	({							\
62		u32 oldval, n = oparg;				\
63		if ((ret = __get_user(oldval, uaddr)) == 0) {	\
64			do {					\
65				oparg = oldval ^ n;		\
66				__futex_cmpxchg();		\
67			} while (ret == 0 && oldval != val);	\
68		}						\
69	})
70
71/* No need to prefetch, since the atomic ops go to the home cache anyway. */
72#define __futex_prolog()
73
74#else
75
76#define __futex_call(FN)						\
77	{								\
78		struct __get_user gu = FN((u32 __force *)uaddr, lock, oparg); \
79		val = gu.val;						\
80		ret = gu.err;						\
81	}
82
83#define __futex_set() __futex_call(__atomic_xchg)
84#define __futex_add() __futex_call(__atomic_xchg_add)
85#define __futex_or() __futex_call(__atomic_or)
86#define __futex_andn() __futex_call(__atomic_andn)
87#define __futex_xor() __futex_call(__atomic_xor)
88
89#define __futex_cmpxchg()						\
90	{								\
91		struct __get_user gu = __atomic_cmpxchg((u32 __force *)uaddr, \
92							lock, oldval, oparg); \
93		val = gu.val;						\
94		ret = gu.err;						\
95	}
96
97/*
98 * Find the lock pointer for the atomic calls to use, and issue a
99 * prefetch to the user address to bring it into cache.  Similar to
100 * __atomic_setup(), but we can't do a read into the L1 since it might
101 * fault; instead we do a prefetch into the L2.
102 */
103#define __futex_prolog()					\
104	int *lock;						\
105	__insn_prefetch(uaddr);					\
106	lock = __atomic_hashed_lock((int __force *)uaddr)
107#endif
108
109static inline int futex_atomic_op_inuser(int encoded_op, u32 __user *uaddr)
110{
111	int op = (encoded_op >> 28) & 7;
112	int cmp = (encoded_op >> 24) & 15;
113	int oparg = (encoded_op << 8) >> 20;
114	int cmparg = (encoded_op << 20) >> 20;
115	int uninitialized_var(val), ret;
116
117	__futex_prolog();
118
119	/* The 32-bit futex code makes this assumption, so validate it here. */
120	BUILD_BUG_ON(sizeof(atomic_t) != sizeof(int));
121
122	if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28))
123		oparg = 1 << oparg;
124
125	if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
126		return -EFAULT;
127
128	pagefault_disable();
129	switch (op) {
130	case FUTEX_OP_SET:
131		__futex_set();
132		break;
133	case FUTEX_OP_ADD:
134		__futex_add();
135		break;
136	case FUTEX_OP_OR:
137		__futex_or();
138		break;
139	case FUTEX_OP_ANDN:
140		__futex_andn();
141		break;
142	case FUTEX_OP_XOR:
143		__futex_xor();
144		break;
145	default:
146		ret = -ENOSYS;
147		break;
148	}
149	pagefault_enable();
150
151	if (!ret) {
152		switch (cmp) {
153		case FUTEX_OP_CMP_EQ:
154			ret = (val == cmparg);
155			break;
156		case FUTEX_OP_CMP_NE:
157			ret = (val != cmparg);
158			break;
159		case FUTEX_OP_CMP_LT:
160			ret = (val < cmparg);
161			break;
162		case FUTEX_OP_CMP_GE:
163			ret = (val >= cmparg);
164			break;
165		case FUTEX_OP_CMP_LE:
166			ret = (val <= cmparg);
167			break;
168		case FUTEX_OP_CMP_GT:
169			ret = (val > cmparg);
170			break;
171		default:
172			ret = -ENOSYS;
173		}
174	}
175	return ret;
176}
177
178static inline int futex_atomic_cmpxchg_inatomic(u32 *uval, u32 __user *uaddr,
179						u32 oldval, u32 oparg)
180{
181	int ret, val;
182
183	__futex_prolog();
184
185	if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
186		return -EFAULT;
187
188	__futex_cmpxchg();
189
190	*uval = val;
191	return ret;
192}
193
194#endif /* !__ASSEMBLY__ */
195
196#endif /* _ASM_TILE_FUTEX_H */
197