xref: /freebsd/sys/sys/_atomic_subword.h (revision c697fb7f)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2019 Kyle Evans <kevans@FreeBSD.org>
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  *
27  * $FreeBSD$
28  */
29 #ifndef _SYS__ATOMIC_SUBWORD_H_
30 #define	_SYS__ATOMIC_SUBWORD_H_
31 
32 /*
33  * This header is specifically for platforms that either do not have ways to or
34  * simply do not do sub-word atomic operations.  These are not ideal as they
35  * require a little more effort to make sure our atomic operations are failing
36  * because of the bits of the word we're trying to write rather than the rest
37  * of the word.
38  */
39 #ifndef _MACHINE_ATOMIC_H_
40 #error do not include this header, use machine/atomic.h
41 #endif
42 
43 #include <machine/endian.h>
44 
45 #ifndef NBBY
46 #define	NBBY	8
47 #endif
48 
49 #define	_ATOMIC_WORD_ALIGNED(p)		\
50     (uint32_t *)((__uintptr_t)(p) - ((__uintptr_t)(p) % 4))
51 
52 #if _BYTE_ORDER == _BIG_ENDIAN
53 #define	_ATOMIC_BYTE_SHIFT(p)		\
54     ((3 - ((__uintptr_t)(p) % 4)) * NBBY)
55 
56 #define	_ATOMIC_HWORD_SHIFT(p)		\
57     ((2 - ((__uintptr_t)(p) % 4)) * NBBY)
58 #else
59 #define	_ATOMIC_BYTE_SHIFT(p)		\
60     ((((__uintptr_t)(p) % 4)) * NBBY)
61 
62 #define	_ATOMIC_HWORD_SHIFT(p)		\
63     ((((__uintptr_t)(p) % 4)) * NBBY)
64 #endif
65 
66 #ifndef	_atomic_cmpset_masked_word
67 /*
68  * Pass these bad boys a couple words and a mask of the bits you care about,
69  * they'll loop until we either succeed or fail because of those bits rather
70  * than the ones we're not masking.  old and val should already be preshifted to
71  * the proper position.
72  */
73 static __inline int
74 _atomic_cmpset_masked_word(uint32_t *addr, uint32_t old, uint32_t val,
75     uint32_t mask)
76 {
77 	int ret;
78 	uint32_t wcomp;
79 
80 	wcomp = old;
81 
82 	/*
83 	 * We'll attempt the cmpset on the entire word.  Loop here in case the
84 	 * operation fails due to the other half-word resident in that word,
85 	 * rather than the half-word we're trying to operate on.  Ideally we
86 	 * only take one trip through here.  We'll have to recalculate the old
87 	 * value since it's the other part of the word changing.
88 	 */
89 	do {
90 		old = (*addr & ~mask) | wcomp;
91 		ret = atomic_fcmpset_32(addr, &old, (old & ~mask) | val);
92 	} while (ret == 0 && (old & mask) == wcomp);
93 
94 	return (ret);
95 }
96 #endif
97 
98 #ifndef	_atomic_fcmpset_masked_word
99 static __inline int
100 _atomic_fcmpset_masked_word(uint32_t *addr, uint32_t *old, uint32_t val,
101     uint32_t mask)
102 {
103 
104 	/*
105 	 * fcmpset_* is documented in atomic(9) to allow spurious failures where
106 	 * *old == val on ll/sc architectures because the sc may fail due to
107 	 * parallel writes or other reasons.  We take advantage of that here
108 	 * and only attempt once, because the caller should be compensating for
109 	 * that possibility.
110 	 */
111 	*old = (*addr & ~mask) | *old;
112 	return (atomic_fcmpset_32(addr, old, (*old & ~mask) | val));
113 }
114 #endif
115 
116 static __inline int
117 atomic_cmpset_8(__volatile uint8_t *addr, uint8_t old, uint8_t val)
118 {
119 	int shift;
120 
121 	shift = _ATOMIC_BYTE_SHIFT(addr);
122 
123 	return (_atomic_cmpset_masked_word(_ATOMIC_WORD_ALIGNED(addr),
124 	    old << shift, val << shift, 0xff << shift));
125 }
126 
127 static __inline int
128 atomic_fcmpset_8(__volatile uint8_t *addr, uint8_t *old, uint8_t val)
129 {
130 	int ret, shift;
131 	uint32_t wold;
132 
133 	shift = _ATOMIC_BYTE_SHIFT(addr);
134 	wold = *old << shift;
135 	ret = _atomic_fcmpset_masked_word(_ATOMIC_WORD_ALIGNED(addr),
136 	    &wold, val << shift, 0xff << shift);
137 	if (ret == 0)
138 		*old = (wold >> shift) & 0xff;
139 	return (ret);
140 }
141 
142 static __inline int
143 atomic_cmpset_16(__volatile uint16_t *addr, uint16_t old, uint16_t val)
144 {
145 	int shift;
146 
147 	shift = _ATOMIC_HWORD_SHIFT(addr);
148 
149 	return (_atomic_cmpset_masked_word(_ATOMIC_WORD_ALIGNED(addr),
150 	    old << shift, val << shift, 0xffff << shift));
151 }
152 
153 static __inline int
154 atomic_fcmpset_16(__volatile uint16_t *addr, uint16_t *old, uint16_t val)
155 {
156 	int ret, shift;
157 	uint32_t wold;
158 
159 	shift = _ATOMIC_HWORD_SHIFT(addr);
160 	wold = *old << shift;
161 	ret = _atomic_fcmpset_masked_word(_ATOMIC_WORD_ALIGNED(addr),
162 	    &wold, val << shift, 0xffff << shift);
163 	if (ret == 0)
164 		*old = (wold >> shift) & 0xffff;
165 	return (ret);
166 }
167 
168 #undef _ATOMIC_WORD_ALIGNED
169 #undef _ATOMIC_BYTE_SHIFT
170 #undef _ATOMIC_HWORD_SHIFT
171 
172 #endif	/* _SYS__ATOMIC_SUBWORD_H_ */
173