xref: /netbsd-src/lib/libc/gen/arc4random.c (revision ba65fde2d7fefa7d39838fa5fa855e62bd606b5e)
1 /*	$NetBSD: arc4random.c,v 1.20 2012/08/20 21:38:09 dsl Exp $	*/
2 /*	$OpenBSD: arc4random.c,v 1.6 2001/06/05 05:05:38 pvalchev Exp $	*/
3 
4 /*
5  * Arc4 random number generator for OpenBSD.
6  * Copyright 1996 David Mazieres <dm@lcs.mit.edu>.
7  *
8  * Modification and redistribution in source and binary forms is
9  * permitted provided that due credit is given to the author and the
10  * OpenBSD project by leaving this copyright notice intact.
11  */
12 
13 /*
14  * This code is derived from section 17.1 of Applied Cryptography,
15  * second edition, which describes a stream cipher allegedly
16  * compatible with RSA Labs "RC4" cipher (the actual description of
17  * which is a trade secret).  The same algorithm is used as a stream
18  * cipher called "arcfour" in Tatu Ylonen's ssh package.
19  *
20  * Here the stream cipher has been modified always to include the time
21  * when initializing the state.  That makes it impossible to
22  * regenerate the same random sequence twice, so this can't be used
23  * for encryption, but will generate good random numbers.
24  *
25  * RC4 is a registered trademark of RSA Laboratories.
26  */
27 
28 #include <sys/cdefs.h>
29 #if defined(LIBC_SCCS) && !defined(lint)
30 __RCSID("$NetBSD: arc4random.c,v 1.20 2012/08/20 21:38:09 dsl Exp $");
31 #endif /* LIBC_SCCS and not lint */
32 
33 #include "namespace.h"
34 #include "reentrant.h"
35 #include <fcntl.h>
36 #include <stdlib.h>
37 #include <unistd.h>
38 #include <sys/types.h>
39 #include <sys/param.h>
40 #include <sys/time.h>
41 #include <sys/sysctl.h>
42 
43 #ifdef __weak_alias
44 __weak_alias(arc4random,_arc4random)
45 __weak_alias(arc4random_addrandom,_arc4random_addrandom)
46 __weak_alias(arc4random_buf,_arc4random_buf)
47 __weak_alias(arc4random_stir,_arc4random_stir)
48 __weak_alias(arc4random_uniform,_arc4random_uniform)
49 #endif
50 
51 struct arc4_stream {
52 	uint8_t stirred;
53 	uint8_t pad;
54 	uint8_t i;
55 	uint8_t j;
56 	uint8_t s[(uint8_t)~0u + 1u];	/* 256 to you and me */
57 	mutex_t mtx;
58 };
59 
60 #ifdef _REENTRANT
61 #define LOCK(rs) { \
62 		int isthreaded = __isthreaded; \
63 		if (isthreaded)        \
64 			mutex_lock(&(rs)->mtx);
65 #define UNLOCK(rs) \
66 		if (isthreaded)        \
67 			mutex_unlock(&(rs)->mtx);      \
68 	}
69 #else
70 #define LOCK(rs)
71 #define UNLOCK(rs)
72 #endif
73 
74 #define S(n) (n)
75 #define S4(n) S(n), S(n + 1), S(n + 2), S(n + 3)
76 #define S16(n) S4(n), S4(n + 4), S4(n + 8), S4(n + 12)
77 #define S64(n) S16(n), S16(n + 16), S16(n + 32), S16(n + 48)
78 #define S256 S64(0), S64(64), S64(128), S64(192)
79 
80 static struct arc4_stream rs = { .i = 0xff, .j = 0, .s = { S256 },
81 		.stirred = 0, .mtx = MUTEX_INITIALIZER };
82 
83 #undef S
84 #undef S4
85 #undef S16
86 #undef S64
87 #undef S256
88 
89 static inline void arc4_addrandom(struct arc4_stream *, u_char *, int);
90 static __noinline void arc4_stir(struct arc4_stream *);
91 static inline uint8_t arc4_getbyte(struct arc4_stream *);
92 static inline uint32_t arc4_getword(struct arc4_stream *);
93 
94 static inline int
95 arc4_check_init(struct arc4_stream *as)
96 {
97 	if (__predict_true(rs.stirred))
98 		return 0;
99 
100 	arc4_stir(as);
101 	return 1;
102 }
103 
104 static inline void
105 arc4_addrandom(struct arc4_stream *as, u_char *dat, int datlen)
106 {
107 	uint8_t si;
108 	size_t n;
109 
110 	for (n = 0; n < __arraycount(as->s); n++) {
111 		as->i = (as->i + 1);
112 		si = as->s[as->i];
113 		as->j = (as->j + si + dat[n % datlen]);
114 		as->s[as->i] = as->s[as->j];
115 		as->s[as->j] = si;
116 	}
117 }
118 
119 static __noinline void
120 arc4_stir(struct arc4_stream *as)
121 {
122 	int rdat[32];
123 	int mib[] = { CTL_KERN, KERN_URND };
124 	size_t len;
125 	size_t i, j;
126 
127 	/*
128 	 * This code once opened and read /dev/urandom on each
129 	 * call.  That causes repeated rekeying of the kernel stream
130 	 * generator, which is very wasteful.  Because of application
131 	 * behavior, caching the fd doesn't really help.  So we just
132 	 * fill up the tank from sysctl, which is a tiny bit slower
133 	 * for us but much friendlier to other entropy consumers.
134 	 */
135 
136 	for (i = 0; i < __arraycount(rdat); i++) {
137 		len = sizeof(rdat[i]);
138 		if (sysctl(mib, 2, &rdat[i], &len, NULL, 0) == -1)
139 			abort();
140 	}
141 
142 	arc4_addrandom(as, (void *) &rdat, (int)sizeof(rdat));
143 
144 	/*
145 	 * Throw away the first N words of output, as suggested in the
146 	 * paper "Weaknesses in the Key Scheduling Algorithm of RC4"
147 	 * by Fluher, Mantin, and Shamir.  (N = 256 in our case.)
148 	 */
149 	for (j = 0; j < __arraycount(as->s) * 4; j++)
150 		arc4_getbyte(as);
151 
152 	as->stirred = 1;
153 }
154 
155 static __always_inline uint8_t
156 arc4_getbyte_ij(struct arc4_stream *as, uint8_t *i, uint8_t *j)
157 {
158 	uint8_t si, sj;
159 
160 	*i = *i + 1;
161 	si = as->s[*i];
162 	*j = *j + si;
163 	sj = as->s[*j];
164 	as->s[*i] = sj;
165 	as->s[*j] = si;
166 	return (as->s[(si + sj) & 0xff]);
167 }
168 
169 static inline uint8_t
170 arc4_getbyte(struct arc4_stream *as)
171 {
172 	return arc4_getbyte_ij(as, &as->i, &as->j);
173 }
174 
175 static inline uint32_t
176 arc4_getword(struct arc4_stream *as)
177 {
178 	uint32_t val;
179 	val = arc4_getbyte(as) << 24;
180 	val |= arc4_getbyte(as) << 16;
181 	val |= arc4_getbyte(as) << 8;
182 	val |= arc4_getbyte(as);
183 	return val;
184 }
185 
186 void
187 arc4random_stir(void)
188 {
189 	LOCK(&rs);
190 	arc4_stir(&rs);
191 	UNLOCK(&rs);
192 }
193 
194 void
195 arc4random_addrandom(u_char *dat, int datlen)
196 {
197 	LOCK(&rs);
198 	arc4_check_init(&rs);
199 	arc4_addrandom(&rs, dat, datlen);
200 	UNLOCK(&rs);
201 }
202 
203 uint32_t
204 arc4random(void)
205 {
206 	uint32_t v;
207 
208 	LOCK(&rs);
209 	arc4_check_init(&rs);
210 	v = arc4_getword(&rs);
211 	UNLOCK(&rs);
212 	return v;
213 }
214 
215 void
216 arc4random_buf(void *buf, size_t len)
217 {
218 	uint8_t *bp = buf;
219 	uint8_t *ep = bp + len;
220 	uint8_t i, j;
221 
222 	LOCK(&rs);
223 	arc4_check_init(&rs);
224 
225 	/* cache i and j - compiler can't know 'buf' doesn't alias them */
226 	i = rs.i;
227 	j = rs.j;
228 
229 	while (bp < ep)
230 		*bp++ = arc4_getbyte_ij(&rs, &i, &j);
231 	rs.i = i;
232 	rs.j = j;
233 
234 	UNLOCK(&rs);
235 }
236 
237 /*-
238  * Written by Damien Miller.
239  * With simplifications by Jinmei Tatuya.
240  */
241 
242 /*
243  * Calculate a uniformly distributed random number less than
244  * upper_bound avoiding "modulo bias".
245  *
246  * Uniformity is achieved by generating new random numbers
247  * until the one returned is outside the range
248  * [0, 2^32 % upper_bound[. This guarantees the selected
249  * random number will be inside the range
250  * [2^32 % upper_bound, 2^32[ which maps back to
251  * [0, upper_bound[ after reduction modulo upper_bound.
252  */
253 uint32_t
254 arc4random_uniform(uint32_t upper_bound)
255 {
256 	uint32_t r, min;
257 
258 	if (upper_bound < 2)
259 		return 0;
260 
261 	/* calculate (2^32 % upper_bound) avoiding 64-bit math */
262 	/* ((2^32 - x) % x) == (2^32 % x) when x <= 2^31 */
263 	min = (0xFFFFFFFFU - upper_bound + 1) % upper_bound;
264 
265 	LOCK(&rs);
266 	arc4_check_init(&rs);
267 
268 	/*
269 	 * This could theoretically loop forever but each retry has
270 	 * p > 0.5 (worst case, usually far better) of selecting a
271 	 * number inside the range we need, so it should rarely need
272 	 * to re-roll (at all).
273 	 */
274 	do
275 		r = arc4_getword(&rs);
276 	while (r < min);
277 	UNLOCK(&rs);
278 
279 	return r % upper_bound;
280 }
281