xref: /openbsd-src/lib/libc/stdlib/random.c (revision ac9b4aacc1da35008afea06a5d23c2f2dea9b93e)
1 /*	$OpenBSD: random.c,v 1.17 2012/06/01 01:01:57 guenther Exp $ */
2 /*
3  * Copyright (c) 1983 Regents of the University of California.
4  * All rights reserved.
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  * 3. Neither the name of the University nor the names of its contributors
15  *    may be used to endorse or promote products derived from this software
16  *    without specific prior written permission.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #include <sys/param.h>
32 #include <sys/sysctl.h>
33 #include <sys/time.h>
34 #include <fcntl.h>
35 #include <stdio.h>
36 #include <stdlib.h>
37 #include <unistd.h>
38 
39 /*
40  * random.c:
41  *
42  * An improved random number generation package.  In addition to the standard
43  * rand()/srand() like interface, this package also has a special state info
44  * interface.  The initstate() routine is called with a seed, an array of
45  * bytes, and a count of how many bytes are being passed in; this array is
46  * then initialized to contain information for random number generation with
47  * that much state information.  Good sizes for the amount of state
48  * information are 32, 64, 128, and 256 bytes.  The state can be switched by
49  * calling the setstate() routine with the same array as was initiallized
50  * with initstate().  By default, the package runs with 128 bytes of state
51  * information and generates far better random numbers than a linear
52  * congruential generator.  If the amount of state information is less than
53  * 32 bytes, a simple linear congruential R.N.G. is used.
54  *
55  * Internally, the state information is treated as an array of int32_t; the
56  * zeroeth element of the array is the type of R.N.G. being used (small
57  * integer); the remainder of the array is the state information for the
58  * R.N.G.  Thus, 32 bytes of state information will give 7 int32_ts worth of
59  * state information, which will allow a degree seven polynomial.  (Note:
60  * the zeroeth word of state information also has some other information
61  * stored in it -- see setstate() for details).
62  *
63  * The random number generation technique is a linear feedback shift register
64  * approach, employing trinomials (since there are fewer terms to sum up that
65  * way).  In this approach, the least significant bit of all the numbers in
66  * the state table will act as a linear feedback shift register, and will
67  * have period 2^deg - 1 (where deg is the degree of the polynomial being
68  * used, assuming that the polynomial is irreducible and primitive).  The
69  * higher order bits will have longer periods, since their values are also
70  * influenced by pseudo-random carries out of the lower bits.  The total
71  * period of the generator is approximately deg*(2**deg - 1); thus doubling
72  * the amount of state information has a vast influence on the period of the
73  * generator.  Note: the deg*(2**deg - 1) is an approximation only good for
74  * large deg, when the period of the shift register is the dominant factor.
75  * With deg equal to seven, the period is actually much longer than the
76  * 7*(2**7 - 1) predicted by this formula.
77  */
78 
79 /*
80  * For each of the currently supported random number generators, we have a
81  * break value on the amount of state information (you need at least this
82  * many bytes of state info to support this random number generator), a degree
83  * for the polynomial (actually a trinomial) that the R.N.G. is based on, and
84  * the separation between the two lower order coefficients of the trinomial.
85  */
86 #define	TYPE_0		0		/* linear congruential */
87 #define	BREAK_0		8
88 #define	DEG_0		0
89 #define	SEP_0		0
90 
91 #define	TYPE_1		1		/* x**7 + x**3 + 1 */
92 #define	BREAK_1		32
93 #define	DEG_1		7
94 #define	SEP_1		3
95 
96 #define	TYPE_2		2		/* x**15 + x + 1 */
97 #define	BREAK_2		64
98 #define	DEG_2		15
99 #define	SEP_2		1
100 
101 #define	TYPE_3		3		/* x**31 + x**3 + 1 */
102 #define	BREAK_3		128
103 #define	DEG_3		31
104 #define	SEP_3		3
105 
106 #define	TYPE_4		4		/* x**63 + x + 1 */
107 #define	BREAK_4		256
108 #define	DEG_4		63
109 #define	SEP_4		1
110 
111 /*
112  * Array versions of the above information to make code run faster --
113  * relies on fact that TYPE_i == i.
114  */
115 #define	MAX_TYPES	5		/* max number of types above */
116 
117 static int degrees[MAX_TYPES] =	{ DEG_0, DEG_1, DEG_2, DEG_3, DEG_4 };
118 static int seps [MAX_TYPES] =	{ SEP_0, SEP_1, SEP_2, SEP_3, SEP_4 };
119 
120 /*
121  * Initially, everything is set up as if from:
122  *
123  *	initstate(1, &randtbl, 128);
124  *
125  * Note that this initialization takes advantage of the fact that srandom()
126  * advances the front and rear pointers 10*rand_deg times, and hence the
127  * rear pointer which starts at 0 will also end up at zero; thus the zeroeth
128  * element of the state information, which contains info about the current
129  * position of the rear pointer is just
130  *
131  *	MAX_TYPES * (rptr - state) + TYPE_3 == TYPE_3.
132  */
133 
134 static int32_t randtbl[DEG_3 + 1] = {
135 	TYPE_3,
136 	0x991539b1, 0x16a5bce3, 0x6774a4cd, 0x3e01511e, 0x4e508aaa, 0x61048c05,
137 	0xf5500617, 0x846b7115, 0x6a19892c, 0x896a97af, 0xdb48f936, 0x14898454,
138 	0x37ffd106, 0xb58bff9c, 0x59e17104, 0xcf918a49, 0x09378c83, 0x52c7a471,
139 	0x8d293ea9, 0x1f4fc301, 0xc3db71be, 0x39b44e1c, 0xf8a44ef9, 0x4c8b80b1,
140 	0x19edc328, 0x87bf4bdd, 0xc9b240e5, 0xe9ee4b1b, 0x4382aee7, 0x535b6b41,
141 	0xf3bec5da,
142 };
143 
144 /*
145  * fptr and rptr are two pointers into the state info, a front and a rear
146  * pointer.  These two pointers are always rand_sep places aparts, as they
147  * cycle cyclically through the state information.  (Yes, this does mean we
148  * could get away with just one pointer, but the code for random() is more
149  * efficient this way).  The pointers are left positioned as they would be
150  * from the call
151  *
152  *	initstate(1, randtbl, 128);
153  *
154  * (The position of the rear pointer, rptr, is really 0 (as explained above
155  * in the initialization of randtbl) because the state table pointer is set
156  * to point to randtbl[1] (as explained below).
157  */
158 static int32_t *fptr = &randtbl[SEP_3 + 1];
159 static int32_t *rptr = &randtbl[1];
160 
161 /*
162  * The following things are the pointer to the state information table, the
163  * type of the current generator, the degree of the current polynomial being
164  * used, and the separation between the two pointers.  Note that for efficiency
165  * of random(), we remember the first location of the state information, not
166  * the zeroeth.  Hence it is valid to access state[-1], which is used to
167  * store the type of the R.N.G.  Also, we remember the last location, since
168  * this is more efficient than indexing every time to find the address of
169  * the last element to see if the front and rear pointers have wrapped.
170  */
171 static int32_t *state = &randtbl[1];
172 static int32_t *end_ptr = &randtbl[DEG_3 + 1];
173 static int rand_type = TYPE_3;
174 static int rand_deg = DEG_3;
175 static int rand_sep = SEP_3;
176 
177 /*
178  * srandom:
179  *
180  * Initialize the random number generator based on the given seed.  If the
181  * type is the trivial no-state-information type, just remember the seed.
182  * Otherwise, initializes state[] based on the given "seed" via a linear
183  * congruential generator.  Then, the pointers are set to known locations
184  * that are exactly rand_sep places apart.  Lastly, it cycles the state
185  * information a given number of times to get rid of any initial dependencies
186  * introduced by the L.C.R.N.G.  Note that the initialization of randtbl[]
187  * for default usage relies on values produced by this routine.
188  */
189 void
190 srandom(unsigned int x)
191 {
192 	int i;
193 	int32_t test;
194 	div_t val;
195 
196 	if (rand_type == TYPE_0)
197 		state[0] = x;
198 	else {
199 		/* A seed of 0 would result in state[] always being zero. */
200 		state[0] = x ? x : 1;
201 		for (i = 1; i < rand_deg; i++) {
202 			/*
203 			 * Implement the following, without overflowing 31 bits:
204 			 *
205 			 *	state[i] = (16807 * state[i - 1]) % 2147483647;
206 			 *
207 			 *	2^31-1 (prime) = 2147483647 = 127773*16807+2836
208 			 */
209 			val = div(state[i-1], 127773);
210 			test = 16807 * val.rem - 2836 * val.quot;
211 			state[i] = test + (test < 0 ? 2147483647 : 0);
212 		}
213 		fptr = &state[rand_sep];
214 		rptr = &state[0];
215 		for (i = 0; i < 10 * rand_deg; i++)
216 			(void)random();
217 	}
218 }
219 
220 /*
221  * srandomdev:
222  *
223  * Many programs choose the seed value in a totally predictable manner.
224  * This often causes problems.  We seed the generator using random
225  * data from the kernel.
226  * Note that this particular seeding procedure can generate states
227  * which are impossible to reproduce by calling srandom() with any
228  * value, since the succeeding terms in the state buffer are no longer
229  * derived from the LC algorithm applied to a fixed seed.
230  */
231 void
232 srandomdev(void)
233 {
234 	int mib[2];
235 	size_t len;
236 
237 	if (rand_type == TYPE_0)
238 		len = sizeof(state[0]);
239 	else
240 		len = rand_deg * sizeof(state[0]);
241 
242 	mib[0] = CTL_KERN;
243 	mib[1] = KERN_ARND;
244 	sysctl(mib, 2, state, &len, NULL, 0);
245 
246 	if (rand_type != TYPE_0) {
247 		fptr = &state[rand_sep];
248 		rptr = &state[0];
249 	}
250 }
251 
252 /*
253  * initstate:
254  *
255  * Initialize the state information in the given array of n bytes for future
256  * random number generation.  Based on the number of bytes we are given, and
257  * the break values for the different R.N.G.'s, we choose the best (largest)
258  * one we can and set things up for it.  srandom() is then called to
259  * initialize the state information.
260  *
261  * Note that on return from srandom(), we set state[-1] to be the type
262  * multiplexed with the current value of the rear pointer; this is so
263  * successive calls to initstate() won't lose this information and will be
264  * able to restart with setstate().
265  *
266  * Note: the first thing we do is save the current state, if any, just like
267  * setstate() so that it doesn't matter when initstate is called.
268  *
269  * Returns a pointer to the old state.
270  */
271 char *
272 initstate(u_int seed, char *arg_state, size_t n)
273 {
274 	char *ostate = (char *)(&state[-1]);
275 
276 	if (rand_type == TYPE_0)
277 		state[-1] = rand_type;
278 	else
279 		state[-1] = MAX_TYPES * (rptr - state) + rand_type;
280 	if (n < BREAK_0)
281 		return(NULL);
282 	if (n < BREAK_1) {
283 		rand_type = TYPE_0;
284 		rand_deg = DEG_0;
285 		rand_sep = SEP_0;
286 	} else if (n < BREAK_2) {
287 		rand_type = TYPE_1;
288 		rand_deg = DEG_1;
289 		rand_sep = SEP_1;
290 	} else if (n < BREAK_3) {
291 		rand_type = TYPE_2;
292 		rand_deg = DEG_2;
293 		rand_sep = SEP_2;
294 	} else if (n < BREAK_4) {
295 		rand_type = TYPE_3;
296 		rand_deg = DEG_3;
297 		rand_sep = SEP_3;
298 	} else {
299 		rand_type = TYPE_4;
300 		rand_deg = DEG_4;
301 		rand_sep = SEP_4;
302 	}
303 	state = &(((int32_t *)arg_state)[1]);	/* first location */
304 	end_ptr = &state[rand_deg];	/* must set end_ptr before srandom */
305 	srandom(seed);
306 	if (rand_type == TYPE_0)
307 		state[-1] = rand_type;
308 	else
309 		state[-1] = MAX_TYPES*(rptr - state) + rand_type;
310 	return(ostate);
311 }
312 
313 /*
314  * setstate:
315  *
316  * Restore the state from the given state array.
317  *
318  * Note: it is important that we also remember the locations of the pointers
319  * in the current state information, and restore the locations of the pointers
320  * from the old state information.  This is done by multiplexing the pointer
321  * location into the zeroeth word of the state information.
322  *
323  * Note that due to the order in which things are done, it is OK to call
324  * setstate() with the same state as the current state.
325  *
326  * Returns a pointer to the old state information.
327  */
328 char *
329 setstate(char *arg_state)
330 {
331 	int32_t *new_state = (int32_t *)arg_state;
332 	int32_t type = new_state[0] % MAX_TYPES;
333 	int32_t rear = new_state[0] / MAX_TYPES;
334 	char *ostate = (char *)(&state[-1]);
335 
336 	if (rand_type == TYPE_0)
337 		state[-1] = rand_type;
338 	else
339 		state[-1] = MAX_TYPES * (rptr - state) + rand_type;
340 	switch(type) {
341 	case TYPE_0:
342 	case TYPE_1:
343 	case TYPE_2:
344 	case TYPE_3:
345 	case TYPE_4:
346 		rand_type = type;
347 		rand_deg = degrees[type];
348 		rand_sep = seps[type];
349 		break;
350 	default:
351 		return(NULL);
352 	}
353 	state = &new_state[1];
354 	if (rand_type != TYPE_0) {
355 		rptr = &state[rear];
356 		fptr = &state[(rear + rand_sep) % rand_deg];
357 	}
358 	end_ptr = &state[rand_deg];		/* set end_ptr too */
359 	return(ostate);
360 }
361 
362 /*
363  * random:
364  *
365  * If we are using the trivial TYPE_0 R.N.G., just do the old linear
366  * congruential bit.  Otherwise, we do our fancy trinomial stuff, which is
367  * the same in all the other cases due to all the global variables that have
368  * been set up.  The basic operation is to add the number at the rear pointer
369  * into the one at the front pointer.  Then both pointers are advanced to
370  * the next location cyclically in the table.  The value returned is the sum
371  * generated, reduced to 31 bits by throwing away the "least random" low bit.
372  *
373  * Note: the code takes advantage of the fact that both the front and
374  * rear pointers can't wrap on the same call by not testing the rear
375  * pointer if the front one has wrapped.
376  *
377  * Returns a 31-bit random number.
378  */
379 long
380 random(void)
381 {
382 	int32_t i;
383 
384 	if (rand_type == TYPE_0)
385 		i = state[0] = (state[0] * 1103515245 + 12345) & 0x7fffffff;
386 	else {
387 		*fptr += *rptr;
388 		i = (*fptr >> 1) & 0x7fffffff;	/* chucking least random bit */
389 		if (++fptr >= end_ptr) {
390 			fptr = state;
391 			++rptr;
392 		} else if (++rptr >= end_ptr)
393 			rptr = state;
394 	}
395 	return((long)i);
396 }
397