xref: /netbsd-src/lib/libc/gen/arc4random.c (revision 63372caa2f74032c7c1cb34e7cd32f28ad65b703)
1 /*	$NetBSD: arc4random.c,v 1.38 2024/08/29 13:39:42 riastradh Exp $	*/
2 
3 /*-
4  * Copyright (c) 2014 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Taylor R. Campbell.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*
33  * Legacy arc4random(3) API from OpenBSD reimplemented using the
34  * ChaCha20 PRF, with per-thread state.
35  *
36  * Security model:
37  * - An attacker who sees some outputs cannot predict past or future
38  *   outputs.
39  * - An attacker who sees the PRNG state cannot predict past outputs.
40  * - An attacker who sees a child's PRNG state cannot predict past or
41  *   future outputs in the parent, or in other children.
42  *
43  * The arc4random(3) API may abort the process if:
44  *
45  * (a) the crypto self-test fails,
46  * (b) pthread_atfork or thr_keycreate fail, or
47  * (c) sysctl(KERN_ARND) fails when reseeding the PRNG.
48  *
49  * The crypto self-test, pthread_atfork, and thr_keycreate occur only
50  * once, on the first use of any of the arc4random(3) API.  KERN_ARND
51  * is unlikely to fail later unless the kernel is seriously broken.
52  */
53 
54 #include <sys/cdefs.h>
55 __RCSID("$NetBSD: arc4random.c,v 1.38 2024/08/29 13:39:42 riastradh Exp $");
56 
57 #include "namespace.h"
58 #include "reentrant.h"
59 
60 #include <sys/bitops.h>
61 #include <sys/endian.h>
62 #include <sys/errno.h>
63 #include <sys/mman.h>
64 #include <sys/sysctl.h>
65 
66 #include <assert.h>
67 #include <sha2.h>
68 #include <stdatomic.h>
69 #include <stdbool.h>
70 #include <stdint.h>
71 #include <stdlib.h>
72 #include <string.h>
73 #include <unistd.h>
74 
75 #include "arc4random.h"
76 #include "reentrant.h"
77 
78 #ifdef __weak_alias
79 __weak_alias(arc4random,_arc4random)
80 __weak_alias(arc4random_addrandom,_arc4random_addrandom)
81 __weak_alias(arc4random_buf,_arc4random_buf)
82 __weak_alias(arc4random_stir,_arc4random_stir)
83 __weak_alias(arc4random_uniform,_arc4random_uniform)
84 #endif
85 
86 /*
87  * For standard ChaCha, use le32dec/le32enc.  We don't need that for
88  * the purposes of a nondeterministic random number generator -- we
89  * don't need to be bit-for-bit compatible over any wire.
90  */
91 
92 static inline uint32_t
93 crypto_le32dec(const void *p)
94 {
95 	uint32_t v;
96 
97 	(void)memcpy(&v, p, sizeof v);
98 
99 	return v;
100 }
101 
102 static inline void
103 crypto_le32enc(void *p, uint32_t v)
104 {
105 
106 	(void)memcpy(p, &v, sizeof v);
107 }
108 
109 /* ChaCha core */
110 
111 #define	crypto_core_OUTPUTBYTES	64
112 #define	crypto_core_INPUTBYTES	16
113 #define	crypto_core_KEYBYTES	32
114 #define	crypto_core_CONSTBYTES	16
115 
116 #define	crypto_core_ROUNDS	20
117 
118 static uint32_t
119 rotate(uint32_t u, unsigned c)
120 {
121 
122 	return (u << c) | (u >> (32 - c));
123 }
124 
125 #define	QUARTERROUND(a, b, c, d) do {					      \
126 	(a) += (b); (d) ^= (a); (d) = rotate((d), 16);			      \
127 	(c) += (d); (b) ^= (c); (b) = rotate((b), 12);			      \
128 	(a) += (b); (d) ^= (a); (d) = rotate((d),  8);			      \
129 	(c) += (d); (b) ^= (c); (b) = rotate((b),  7);			      \
130 } while (0)
131 
132 static const uint8_t crypto_core_constant32[16] = "expand 32-byte k";
133 
134 static void
135 crypto_core(uint8_t *out, const uint8_t *in, const uint8_t *k,
136     const uint8_t *c)
137 {
138 	uint32_t x0,x1,x2,x3,x4,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15;
139 	uint32_t j0,j1,j2,j3,j4,j5,j6,j7,j8,j9,j10,j11,j12,j13,j14,j15;
140 	int i;
141 
142 	j0 = x0 = crypto_le32dec(c + 0);
143 	j1 = x1 = crypto_le32dec(c + 4);
144 	j2 = x2 = crypto_le32dec(c + 8);
145 	j3 = x3 = crypto_le32dec(c + 12);
146 	j4 = x4 = crypto_le32dec(k + 0);
147 	j5 = x5 = crypto_le32dec(k + 4);
148 	j6 = x6 = crypto_le32dec(k + 8);
149 	j7 = x7 = crypto_le32dec(k + 12);
150 	j8 = x8 = crypto_le32dec(k + 16);
151 	j9 = x9 = crypto_le32dec(k + 20);
152 	j10 = x10 = crypto_le32dec(k + 24);
153 	j11 = x11 = crypto_le32dec(k + 28);
154 	j12 = x12 = crypto_le32dec(in + 0);
155 	j13 = x13 = crypto_le32dec(in + 4);
156 	j14 = x14 = crypto_le32dec(in + 8);
157 	j15 = x15 = crypto_le32dec(in + 12);
158 
159 	for (i = crypto_core_ROUNDS; i > 0; i -= 2) {
160 		QUARTERROUND( x0, x4, x8,x12);
161 		QUARTERROUND( x1, x5, x9,x13);
162 		QUARTERROUND( x2, x6,x10,x14);
163 		QUARTERROUND( x3, x7,x11,x15);
164 		QUARTERROUND( x0, x5,x10,x15);
165 		QUARTERROUND( x1, x6,x11,x12);
166 		QUARTERROUND( x2, x7, x8,x13);
167 		QUARTERROUND( x3, x4, x9,x14);
168 	}
169 
170 	crypto_le32enc(out + 0, x0 + j0);
171 	crypto_le32enc(out + 4, x1 + j1);
172 	crypto_le32enc(out + 8, x2 + j2);
173 	crypto_le32enc(out + 12, x3 + j3);
174 	crypto_le32enc(out + 16, x4 + j4);
175 	crypto_le32enc(out + 20, x5 + j5);
176 	crypto_le32enc(out + 24, x6 + j6);
177 	crypto_le32enc(out + 28, x7 + j7);
178 	crypto_le32enc(out + 32, x8 + j8);
179 	crypto_le32enc(out + 36, x9 + j9);
180 	crypto_le32enc(out + 40, x10 + j10);
181 	crypto_le32enc(out + 44, x11 + j11);
182 	crypto_le32enc(out + 48, x12 + j12);
183 	crypto_le32enc(out + 52, x13 + j13);
184 	crypto_le32enc(out + 56, x14 + j14);
185 	crypto_le32enc(out + 60, x15 + j15);
186 }
187 
188 /* ChaCha self-test */
189 
190 #ifdef _DIAGNOSTIC
191 
192 /*
193  * Test vector for ChaCha20 from
194  * <http://tools.ietf.org/html/draft-strombergson-chacha-test-vectors-00>,
195  * test vectors for ChaCha12 and ChaCha8 and for big-endian machines
196  * generated by the same crypto_core code with crypto_core_ROUNDS and
197  * crypto_le32enc/dec varied.
198  */
199 
200 static const uint8_t crypto_core_selftest_vector[64] = {
201 #if _BYTE_ORDER == _LITTLE_ENDIAN
202 #  if crypto_core_ROUNDS == 8
203 	0x3e,0x00,0xef,0x2f,0x89,0x5f,0x40,0xd6,
204 	0x7f,0x5b,0xb8,0xe8,0x1f,0x09,0xa5,0xa1,
205 	0x2c,0x84,0x0e,0xc3,0xce,0x9a,0x7f,0x3b,
206 	0x18,0x1b,0xe1,0x88,0xef,0x71,0x1a,0x1e,
207 	0x98,0x4c,0xe1,0x72,0xb9,0x21,0x6f,0x41,
208 	0x9f,0x44,0x53,0x67,0x45,0x6d,0x56,0x19,
209 	0x31,0x4a,0x42,0xa3,0xda,0x86,0xb0,0x01,
210 	0x38,0x7b,0xfd,0xb8,0x0e,0x0c,0xfe,0x42,
211 #  elif crypto_core_ROUNDS == 12
212 	0x9b,0xf4,0x9a,0x6a,0x07,0x55,0xf9,0x53,
213 	0x81,0x1f,0xce,0x12,0x5f,0x26,0x83,0xd5,
214 	0x04,0x29,0xc3,0xbb,0x49,0xe0,0x74,0x14,
215 	0x7e,0x00,0x89,0xa5,0x2e,0xae,0x15,0x5f,
216 	0x05,0x64,0xf8,0x79,0xd2,0x7a,0xe3,0xc0,
217 	0x2c,0xe8,0x28,0x34,0xac,0xfa,0x8c,0x79,
218 	0x3a,0x62,0x9f,0x2c,0xa0,0xde,0x69,0x19,
219 	0x61,0x0b,0xe8,0x2f,0x41,0x13,0x26,0xbe,
220 #  elif crypto_core_ROUNDS == 20
221 	0x76,0xb8,0xe0,0xad,0xa0,0xf1,0x3d,0x90,
222 	0x40,0x5d,0x6a,0xe5,0x53,0x86,0xbd,0x28,
223 	0xbd,0xd2,0x19,0xb8,0xa0,0x8d,0xed,0x1a,
224 	0xa8,0x36,0xef,0xcc,0x8b,0x77,0x0d,0xc7,
225 	0xda,0x41,0x59,0x7c,0x51,0x57,0x48,0x8d,
226 	0x77,0x24,0xe0,0x3f,0xb8,0xd8,0x4a,0x37,
227 	0x6a,0x43,0xb8,0xf4,0x15,0x18,0xa1,0x1c,
228 	0xc3,0x87,0xb6,0x69,0xb2,0xee,0x65,0x86,
229 #  else
230 #    error crypto_core_ROUNDS must be 8, 12, or 20.
231 #  endif
232 #elif _BYTE_ORDER == _BIG_ENDIAN
233 #  if crypto_core_ROUNDS == 8
234 	0x9a,0x13,0x07,0xe3,0x38,0x18,0x9e,0x99,
235 	0x15,0x37,0x16,0x4d,0x04,0xe6,0x48,0x9a,
236 	0x07,0xd6,0xe8,0x7a,0x02,0xf9,0xf5,0xc7,
237 	0x3f,0xa9,0xc2,0x0a,0xe1,0xc6,0x62,0xea,
238 	0x80,0xaf,0xb6,0x51,0xca,0x52,0x43,0x87,
239 	0xe3,0xa6,0xa6,0x61,0x11,0xf5,0xe6,0xcf,
240 	0x09,0x0f,0xdc,0x9d,0xc3,0xc3,0xbb,0x43,
241 	0xd7,0xfa,0x70,0x42,0xbf,0xa5,0xee,0xa2,
242 #  elif crypto_core_ROUNDS == 12
243 	0xcf,0x6c,0x16,0x48,0xbf,0xf4,0xba,0x85,
244 	0x32,0x69,0xd3,0x98,0xc8,0x7d,0xcd,0x3f,
245 	0xdc,0x76,0x6b,0xa2,0x7b,0xcb,0x17,0x4d,
246 	0x05,0xda,0xdd,0xd8,0x62,0x54,0xbf,0xe0,
247 	0x65,0xed,0x0e,0xf4,0x01,0x7e,0x3c,0x05,
248 	0x35,0xb2,0x7a,0x60,0xf3,0x8f,0x12,0x33,
249 	0x24,0x60,0xcd,0x85,0xfe,0x4c,0xf3,0x39,
250 	0xb1,0x0e,0x3e,0xe0,0xba,0xa6,0x2f,0xa9,
251 #  elif crypto_core_ROUNDS == 20
252 	0x83,0x8b,0xf8,0x75,0xf7,0xde,0x9d,0x8c,
253 	0x33,0x14,0x72,0x28,0xd1,0xbe,0x88,0xe5,
254 	0x94,0xb5,0xed,0xb8,0x56,0xb5,0x9e,0x0c,
255 	0x64,0x6a,0xaf,0xd9,0xa7,0x49,0x10,0x59,
256 	0xba,0x3a,0x82,0xf8,0x4a,0x70,0x9c,0x00,
257 	0x82,0x2c,0xae,0xc6,0xd7,0x1c,0x2e,0xda,
258 	0x2a,0xfb,0x61,0x70,0x2b,0xd1,0xbf,0x8b,
259 	0x95,0xbc,0x23,0xb6,0x4b,0x60,0x02,0xec,
260 #  else
261 #    error crypto_core_ROUNDS must be 8, 12, or 20.
262 #  endif
263 #else
264 #  error Byte order must be little-endian or big-endian.
265 #endif
266 };
267 
268 static int
269 crypto_core_selftest(void)
270 {
271 	const uint8_t nonce[crypto_core_INPUTBYTES] = {0};
272 	const uint8_t key[crypto_core_KEYBYTES] = {0};
273 	uint8_t block[64];
274 	unsigned i;
275 
276 	crypto_core(block, nonce, key, crypto_core_constant32);
277 	for (i = 0; i < 64; i++) {
278 		if (block[i] != crypto_core_selftest_vector[i])
279 			return EIO;
280 	}
281 
282 	return 0;
283 }
284 
285 #else  /* !_DIAGNOSTIC */
286 
287 static int
288 crypto_core_selftest(void)
289 {
290 
291 	return 0;
292 }
293 
294 #endif
295 
296 /* PRNG */
297 
298 /*
299  * For a state s, rather than use ChaCha20 as a stream cipher to
300  * generate the concatenation ChaCha20_s(0) || ChaCha20_s(1) || ..., we
301  * split ChaCha20_s(0) into s' || x and yield x for the first request,
302  * split ChaCha20_s'(0) into s'' || y and yield y for the second
303  * request, &c.  This provides backtracking resistance: an attacker who
304  * finds s'' can't recover s' or x.
305  */
306 
307 #define	crypto_prng_SEEDBYTES		crypto_core_KEYBYTES
308 #define	crypto_prng_MAXOUTPUTBYTES	\
309 	(crypto_core_OUTPUTBYTES - crypto_prng_SEEDBYTES)
310 
311 __CTASSERT(sizeof(struct crypto_prng) == crypto_prng_SEEDBYTES);
312 
313 static void
314 crypto_prng_seed(struct crypto_prng *prng, const void *seed)
315 {
316 
317 	(void)memcpy(prng->state, seed, crypto_prng_SEEDBYTES);
318 }
319 
320 static void
321 crypto_prng_buf(struct crypto_prng *prng, void *buf, size_t n)
322 {
323 	const uint8_t nonce[crypto_core_INPUTBYTES] = {0};
324 	uint8_t output[crypto_core_OUTPUTBYTES];
325 
326 	_DIAGASSERT(n <= crypto_prng_MAXOUTPUTBYTES);
327 	__CTASSERT(sizeof prng->state + crypto_prng_MAXOUTPUTBYTES
328 	    <= sizeof output);
329 
330 	crypto_core(output, nonce, prng->state, crypto_core_constant32);
331 	(void)memcpy(prng->state, output, sizeof prng->state);
332 	(void)memcpy(buf, output + sizeof prng->state, n);
333 	(void)explicit_memset(output, 0, sizeof output);
334 }
335 
336 /* One-time stream: expand short single-use secret into long secret */
337 
338 #define	crypto_onetimestream_SEEDBYTES	crypto_core_KEYBYTES
339 
340 static void
341 crypto_onetimestream(const void *seed, void *buf, size_t n)
342 {
343 	uint32_t nonce[crypto_core_INPUTBYTES / sizeof(uint32_t)] = {0};
344 	uint8_t block[crypto_core_OUTPUTBYTES];
345 	uint8_t *p8, *p32;
346 	const uint8_t *nonce8 = (const uint8_t *)(void *)nonce;
347 	size_t ni, nb, nf;
348 
349 	/*
350 	 * Guarantee we can generate up to n bytes.  We have
351 	 * 2^(8*INPUTBYTES) possible inputs yielding output of
352 	 * OUTPUTBYTES*2^(8*INPUTBYTES) bytes.  It suffices to require
353 	 * that sizeof n > (1/CHAR_BIT) log_2 n be less than
354 	 * (1/CHAR_BIT) log_2 of the total output stream length.  We
355 	 * have
356 	 *
357 	 *	log_2 (o 2^(8 i)) = log_2 o + log_2 2^(8 i)
358 	 *	  = log_2 o + 8 i.
359 	 */
360 #ifndef __lint__
361 	__CTASSERT(CHAR_BIT * sizeof n <= (ilog2(crypto_core_OUTPUTBYTES) +
362 		8 * crypto_core_INPUTBYTES));
363 #endif
364 
365 	p8 = buf;
366 	p32 = (uint8_t *)roundup2((uintptr_t)p8, 4);
367 	ni = p32 - p8;
368 	if (n < ni)
369 		ni = n;
370 	nb = (n - ni) / sizeof block;
371 	nf = (n - ni) % sizeof block;
372 
373 	_DIAGASSERT(((uintptr_t)p32 & 3) == 0);
374 	_DIAGASSERT(ni <= n);
375 	_DIAGASSERT(nb <= (n / sizeof block));
376 	_DIAGASSERT(nf <= n);
377 	_DIAGASSERT(n == (ni + (nb * sizeof block) + nf));
378 	_DIAGASSERT(ni < 4);
379 	_DIAGASSERT(nf < sizeof block);
380 
381 	if (ni) {
382 		crypto_core(block, nonce8, seed, crypto_core_constant32);
383 		nonce[0]++;
384 		(void)memcpy(p8, block, ni);
385 	}
386 	while (nb--) {
387 		crypto_core(p32, nonce8, seed, crypto_core_constant32);
388 		if (++nonce[0] == 0)
389 			nonce[1]++;
390 		p32 += crypto_core_OUTPUTBYTES;
391 	}
392 	if (nf) {
393 		crypto_core(block, nonce8, seed, crypto_core_constant32);
394 		if (++nonce[0] == 0)
395 			nonce[1]++;
396 		(void)memcpy(p32, block, nf);
397 	}
398 
399 	if (ni | nf)
400 		(void)explicit_memset(block, 0, sizeof block);
401 }
402 
403 /*
404  * entropy_epoch()
405  *
406  *	Return the current entropy epoch, from the sysctl node
407  *	kern.entropy.epoch.
408  *
409  *	The entropy epoch is never zero.  Initially, or on error, it is
410  *	(unsigned)-1.  It may wrap around but it skips (unsigned)-1 and
411  *	0 when it does.  Changes happen less than once per second, so
412  *	wraparound will only affect systems after 136 years of uptime.
413  *
414  *	XXX This should get it from a page shared read-only by kernel
415  *	with userland, but until we implement such a mechanism, this
416  *	sysctl -- incurring the cost of a syscall -- will have to
417  *	serve.
418  */
419 static unsigned
420 entropy_epoch(void)
421 {
422 	static atomic_int mib0[3];
423 	static atomic_bool initialized = false;
424 	int mib[3];
425 	unsigned epoch = (unsigned)-1;
426 	size_t epochlen = sizeof(epoch);
427 
428 	/*
429 	 * Resolve kern.entropy.epoch if we haven't already.  Cache it
430 	 * for the next caller.  Initialization is idempotent, so it's
431 	 * OK if two threads do it at once.
432 	 */
433 	if (atomic_load_explicit(&initialized, memory_order_acquire)) {
434 		mib[0] = atomic_load_explicit(&mib0[0], memory_order_relaxed);
435 		mib[1] = atomic_load_explicit(&mib0[1], memory_order_relaxed);
436 		mib[2] = atomic_load_explicit(&mib0[2], memory_order_relaxed);
437 	} else {
438 		size_t nmib = __arraycount(mib);
439 
440 		if (sysctlnametomib("kern.entropy.epoch", mib, &nmib) == -1)
441 			return (unsigned)-1;
442 		if (nmib != __arraycount(mib))
443 			return (unsigned)-1;
444 		atomic_store_explicit(&mib0[0], mib[0], memory_order_relaxed);
445 		atomic_store_explicit(&mib0[1], mib[1], memory_order_relaxed);
446 		atomic_store_explicit(&mib0[2], mib[2], memory_order_relaxed);
447 		atomic_store_explicit(&initialized, true,
448 		    memory_order_release);
449 	}
450 
451 	if (sysctl(mib, __arraycount(mib), &epoch, &epochlen, NULL, 0) == -1)
452 		return (unsigned)-1;
453 	if (epochlen != sizeof(epoch))
454 		return (unsigned)-1;
455 
456 	return epoch;
457 }
458 
459 /* arc4random state: per-thread, per-process (zeroed in child on fork) */
460 
461 static void
462 arc4random_prng_addrandom(struct arc4random_prng *prng, const void *data,
463     size_t datalen)
464 {
465 	const int mib[] = { CTL_KERN, KERN_ARND };
466 	SHA256_CTX ctx;
467 	uint8_t buf[crypto_prng_SEEDBYTES];
468 	size_t buflen = sizeof buf;
469 	unsigned epoch = entropy_epoch();
470 
471 	__CTASSERT(sizeof buf == SHA256_DIGEST_LENGTH);
472 
473 	SHA256_Init(&ctx);
474 
475 	crypto_prng_buf(&prng->arc4_prng, buf, sizeof buf);
476 	SHA256_Update(&ctx, buf, sizeof buf);
477 
478 	if (sysctl(mib, (u_int)__arraycount(mib), buf, &buflen, NULL, 0) == -1)
479 		abort();
480 	if (buflen != sizeof buf)
481 		abort();
482 	SHA256_Update(&ctx, buf, sizeof buf);
483 
484 	if (data != NULL)
485 		SHA256_Update(&ctx, data, datalen);
486 
487 	SHA256_Final(buf, &ctx);
488 	(void)explicit_memset(&ctx, 0, sizeof ctx);
489 
490 	/* reseed(SHA256(prng() || sysctl(KERN_ARND) || data)) */
491 	crypto_prng_seed(&prng->arc4_prng, buf);
492 	(void)explicit_memset(buf, 0, sizeof buf);
493 	prng->arc4_epoch = epoch;
494 }
495 
496 #ifdef _REENTRANT
497 static struct arc4random_prng *
498 arc4random_prng_create(void)
499 {
500 	struct arc4random_prng *prng;
501 	const size_t size = roundup(sizeof(*prng), sysconf(_SC_PAGESIZE));
502 
503 	prng = mmap(NULL, size, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANON, -1,
504 	    0);
505 	if (prng == MAP_FAILED)
506 		goto fail0;
507 	if (minherit(prng, size, MAP_INHERIT_ZERO) == -1)
508 		goto fail1;
509 
510 	return prng;
511 
512 fail1:	(void)munmap(prng, size);
513 fail0:	return NULL;
514 }
515 #endif
516 
517 #ifdef _REENTRANT
518 static void
519 arc4random_prng_destroy(struct arc4random_prng *prng)
520 {
521 	const size_t size = roundup(sizeof(*prng), sysconf(_SC_PAGESIZE));
522 
523 	(void)explicit_memset(prng, 0, sizeof(*prng));
524 	(void)munmap(prng, size);
525 }
526 #endif
527 
528 /* Library state */
529 
530 struct arc4random_global_state arc4random_global = {
531 #ifdef _REENTRANT
532 	.lock		= MUTEX_INITIALIZER,
533 #endif
534 	.initialized	= false,
535 };
536 
537 static void
538 arc4random_atfork_prepare(void)
539 {
540 
541 	mutex_lock(&arc4random_global.lock);
542 	(void)explicit_memset(&arc4random_global.prng, 0,
543 	    sizeof arc4random_global.prng);
544 }
545 
546 static void
547 arc4random_atfork_parent(void)
548 {
549 
550 	mutex_unlock(&arc4random_global.lock);
551 }
552 
553 static void
554 arc4random_atfork_child(void)
555 {
556 
557 	mutex_unlock(&arc4random_global.lock);
558 }
559 
560 #ifdef _REENTRANT
561 static void
562 arc4random_tsd_destructor(void *p)
563 {
564 	struct arc4random_prng *const prng = p;
565 
566 	arc4random_prng_destroy(prng);
567 }
568 #endif
569 
570 static void
571 arc4random_initialize(void)
572 {
573 
574 	mutex_lock(&arc4random_global.lock);
575 	if (!arc4random_global.initialized) {
576 		if (crypto_core_selftest() != 0)
577 			abort();
578 		if (pthread_atfork(&arc4random_atfork_prepare,
579 			&arc4random_atfork_parent, &arc4random_atfork_child)
580 		    != 0)
581 			abort();
582 #ifdef _REENTRANT
583 		if (thr_keycreate(&arc4random_global.thread_key,
584 			&arc4random_tsd_destructor) != 0)
585 			abort();
586 #endif
587 		arc4random_global.initialized = true;
588 	}
589 	mutex_unlock(&arc4random_global.lock);
590 }
591 
592 static struct arc4random_prng *
593 arc4random_prng_get(void)
594 {
595 	struct arc4random_prng *prng = NULL;
596 
597 	/* Make sure the library is initialized.  */
598 	if (__predict_false(!arc4random_global.initialized))
599 		arc4random_initialize();
600 
601 #ifdef _REENTRANT
602 	/* Get or create the per-thread PRNG state.  */
603 	prng = thr_getspecific(arc4random_global.thread_key);
604 	if (__predict_false(prng == NULL)) {
605 		prng = arc4random_prng_create();
606 		thr_setspecific(arc4random_global.thread_key, prng);
607 	}
608 #endif
609 
610 	/* If we can't create it, fall back to the global PRNG.  */
611 	if (__predict_false(prng == NULL)) {
612 		mutex_lock(&arc4random_global.lock);
613 		prng = &arc4random_global.prng;
614 	}
615 
616 	/* Guarantee the PRNG is seeded.  */
617 	if (__predict_false(prng->arc4_epoch != entropy_epoch()))
618 		arc4random_prng_addrandom(prng, NULL, 0);
619 
620 	return prng;
621 }
622 
623 static void
624 arc4random_prng_put(struct arc4random_prng *prng)
625 {
626 
627 	/* If we had fallen back to the global PRNG, unlock it.  */
628 	if (__predict_false(prng == &arc4random_global.prng))
629 		mutex_unlock(&arc4random_global.lock);
630 }
631 
632 /* Public API */
633 
634 uint32_t
635 arc4random(void)
636 {
637 	struct arc4random_prng *prng;
638 	uint32_t v;
639 
640 	prng = arc4random_prng_get();
641 	crypto_prng_buf(&prng->arc4_prng, &v, sizeof v);
642 	arc4random_prng_put(prng);
643 
644 	return v;
645 }
646 
647 void
648 arc4random_buf(void *buf, size_t len)
649 {
650 	struct arc4random_prng *prng;
651 
652 	if (len <= crypto_prng_MAXOUTPUTBYTES) {
653 		prng = arc4random_prng_get();
654 		crypto_prng_buf(&prng->arc4_prng, buf, len);
655 		arc4random_prng_put(prng);
656 	} else {
657 		uint8_t seed[crypto_onetimestream_SEEDBYTES];
658 
659 		prng = arc4random_prng_get();
660 		crypto_prng_buf(&prng->arc4_prng, seed, sizeof seed);
661 		arc4random_prng_put(prng);
662 
663 		crypto_onetimestream(seed, buf, len);
664 		(void)explicit_memset(seed, 0, sizeof seed);
665 	}
666 }
667 
668 uint32_t
669 arc4random_uniform(uint32_t bound)
670 {
671 	struct arc4random_prng *prng;
672 	uint32_t minimum, r;
673 
674 	/*
675 	 * We want a uniform random choice in [0, n), and arc4random()
676 	 * makes a uniform random choice in [0, 2^32).  If we reduce
677 	 * that modulo n, values in [0, 2^32 mod n) will be represented
678 	 * slightly more than values in [2^32 mod n, n).  Instead we
679 	 * choose only from [2^32 mod n, 2^32) by rejecting samples in
680 	 * [0, 2^32 mod n), to avoid counting the extra representative
681 	 * of [0, 2^32 mod n).  To compute 2^32 mod n, note that
682 	 *
683 	 *	2^32 mod n = 2^32 mod n - 0
684 	 *	  = 2^32 mod n - n mod n
685 	 *	  = (2^32 - n) mod n,
686 	 *
687 	 * the last of which is what we compute in 32-bit arithmetic.
688 	 */
689 	minimum = (-bound % bound);
690 
691 	prng = arc4random_prng_get();
692 	do crypto_prng_buf(&prng->arc4_prng, &r, sizeof r);
693 	while (__predict_false(r < minimum));
694 	arc4random_prng_put(prng);
695 
696 	return (r % bound);
697 }
698 
699 void
700 arc4random_stir(void)
701 {
702 	struct arc4random_prng *prng;
703 
704 	prng = arc4random_prng_get();
705 	arc4random_prng_addrandom(prng, NULL, 0);
706 	arc4random_prng_put(prng);
707 }
708 
709 /*
710  * Silly signature here is for hysterical raisins.  Should instead be
711  * const void *data and size_t datalen.
712  */
713 void
714 arc4random_addrandom(u_char *data, int datalen)
715 {
716 	struct arc4random_prng *prng;
717 
718 	_DIAGASSERT(0 <= datalen);
719 
720 	prng = arc4random_prng_get();
721 	arc4random_prng_addrandom(prng, data, datalen);
722 	arc4random_prng_put(prng);
723 }
724 
725 #ifdef _ARC4RANDOM_TEST
726 
727 #include <sys/wait.h>
728 
729 #include <err.h>
730 #include <stdio.h>
731 
732 int
733 main(int argc __unused, char **argv __unused)
734 {
735 	unsigned char gubbish[] = "random gubbish";
736 	const uint8_t zero64[64] = {0};
737 	uint8_t buf[2048];
738 	unsigned i, a, n;
739 
740 	/* Test arc4random: should not be deterministic.  */
741 	if (printf("arc4random: %08"PRIx32"\n", arc4random()) < 0)
742 		err(1, "printf");
743 
744 	/* Test stirring: should definitely not be deterministic.  */
745 	arc4random_stir();
746 
747 	/* Test small buffer.  */
748 	arc4random_buf(buf, 8);
749 	if (printf("arc4randombuf small:") < 0)
750 		err(1, "printf");
751 	for (i = 0; i < 8; i++)
752 		if (printf(" %02x", buf[i]) < 0)
753 			err(1, "printf");
754 	if (printf("\n") < 0)
755 		err(1, "printf");
756 
757 	/* Test addrandom: should not make the rest deterministic.  */
758 	arc4random_addrandom(gubbish, sizeof gubbish);
759 
760 	/* Test large buffer.  */
761 	arc4random_buf(buf, sizeof buf);
762 	if (printf("arc4randombuf_large:") < 0)
763 		err(1, "printf");
764 	for (i = 0; i < sizeof buf; i++)
765 		if (printf(" %02x", buf[i]) < 0)
766 			err(1, "printf");
767 	if (printf("\n") < 0)
768 		err(1, "printf");
769 
770 	/* Test misaligned small and large.  */
771 	for (a = 0; a < 64; a++) {
772 		for (n = a; n < sizeof buf; n++) {
773 			(void)memset(buf, 0, sizeof buf);
774 			arc4random_buf(buf, n - a);
775 			if (memcmp(buf + n - a, zero64, a) != 0)
776 				errx(1, "arc4random buffer overflow 0");
777 
778 			(void)memset(buf, 0, sizeof buf);
779 			arc4random_buf(buf + a, n - a);
780 			if (memcmp(buf, zero64, a) != 0)
781 				errx(1, "arc4random buffer overflow 1");
782 
783 			if ((2*a) <= n) {
784 				(void)memset(buf, 0, sizeof buf);
785 				arc4random_buf(buf + a, n - a - a);
786 				if (memcmp(buf + n - a, zero64, a) != 0)
787 					errx(1,
788 					    "arc4random buffer overflow 2");
789 			}
790 		}
791 	}
792 
793 	/* Test fork-safety.  */
794     {
795 	pid_t pid, rpid;
796 	int status;
797 
798 	pid = fork();
799 	switch (pid) {
800 	case -1:
801 		err(1, "fork");
802 	case 0: {
803 		/*
804 		 * Verify the epoch has been set to zero by fork.
805 		 */
806 		struct arc4random_prng *prng = NULL;
807 #ifdef _REENTRANT
808 		prng = thr_getspecific(arc4random_global.thread_key);
809 #endif
810 		if (prng == NULL)
811 			prng = &arc4random_global.prng;
812 		_exit(prng->arc4_epoch != 0);
813 	}
814 	default:
815 		rpid = waitpid(pid, &status, 0);
816 		if (rpid == -1)
817 			err(1, "waitpid");
818 		if (rpid != pid)
819 			errx(1, "waitpid returned wrong pid"
820 			    ": %"PRIdMAX" != %"PRIdMAX,
821 			    (intmax_t)rpid,
822 			    (intmax_t)pid);
823 		if (WIFEXITED(status)) {
824 			if (WEXITSTATUS(status) != 0)
825 				errx(1, "child exited with %d",
826 				    WEXITSTATUS(status));
827 		} else if (WIFSIGNALED(status)) {
828 			errx(1, "child terminated on signal %d",
829 			    WTERMSIG(status));
830 		} else {
831 			errx(1, "child died mysteriously: %d", status);
832 		}
833 	}
834     }
835 
836 	/* XXX Test multithreaded fork safety...?  */
837 
838 	return 0;
839 }
840 #endif
841