xref: /freebsd/lib/libc/stdlib/random.c (revision 952d112864d8008aa87278a30a539d888a8493cd)
1 /*
2  * Copyright (c) 1983, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 3. All advertising materials mentioning features or use of this software
14  *    must display the following acknowledgement:
15  *	This product includes software developed by the University of
16  *	California, Berkeley and its contributors.
17  * 4. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  * $Id$
34  *
35  */
36 
37 #if defined(LIBC_SCCS) && !defined(lint)
38 static char sccsid[] = "@(#)random.c	8.2 (Berkeley) 5/19/95";
39 #endif /* LIBC_SCCS and not lint */
40 
41 #include <fcntl.h>             /* for srandomdev() */
42 #include <stdio.h>
43 #include <stdlib.h>
44 #include <unistd.h>            /* for srandomdev() */
45 
46 /*
47  * random.c:
48  *
49  * An improved random number generation package.  In addition to the standard
50  * rand()/srand() like interface, this package also has a special state info
51  * interface.  The initstate() routine is called with a seed, an array of
52  * bytes, and a count of how many bytes are being passed in; this array is
53  * then initialized to contain information for random number generation with
54  * that much state information.  Good sizes for the amount of state
55  * information are 32, 64, 128, and 256 bytes.  The state can be switched by
56  * calling the setstate() routine with the same array as was initiallized
57  * with initstate().  By default, the package runs with 128 bytes of state
58  * information and generates far better random numbers than a linear
59  * congruential generator.  If the amount of state information is less than
60  * 32 bytes, a simple linear congruential R.N.G. is used.
61  *
62  * Internally, the state information is treated as an array of longs; the
63  * zeroeth element of the array is the type of R.N.G. being used (small
64  * integer); the remainder of the array is the state information for the
65  * R.N.G.  Thus, 32 bytes of state information will give 7 longs worth of
66  * state information, which will allow a degree seven polynomial.  (Note:
67  * the zeroeth word of state information also has some other information
68  * stored in it -- see setstate() for details).
69  *
70  * The random number generation technique is a linear feedback shift register
71  * approach, employing trinomials (since there are fewer terms to sum up that
72  * way).  In this approach, the least significant bit of all the numbers in
73  * the state table will act as a linear feedback shift register, and will
74  * have period 2^deg - 1 (where deg is the degree of the polynomial being
75  * used, assuming that the polynomial is irreducible and primitive).  The
76  * higher order bits will have longer periods, since their values are also
77  * influenced by pseudo-random carries out of the lower bits.  The total
78  * period of the generator is approximately deg*(2**deg - 1); thus doubling
79  * the amount of state information has a vast influence on the period of the
80  * generator.  Note: the deg*(2**deg - 1) is an approximation only good for
81  * large deg, when the period of the shift register is the dominant factor.
82  * With deg equal to seven, the period is actually much longer than the
83  * 7*(2**7 - 1) predicted by this formula.
84  *
85  * Modified 28 December 1994 by Jacob S. Rosenberg.
86  * The following changes have been made:
87  * All references to the type u_int have been changed to unsigned long.
88  * All references to type int have been changed to type long.  Other
89  * cleanups have been made as well.  A warning for both initstate and
90  * setstate has been inserted to the effect that on Sparc platforms
91  * the 'arg_state' variable must be forced to begin on word boundaries.
92  * This can be easily done by casting a long integer array to char *.
93  * The overall logic has been left STRICTLY alone.  This software was
94  * tested on both a VAX and Sun SpacsStation with exactly the same
95  * results.  The new version and the original give IDENTICAL results.
96  * The new version is somewhat faster than the original.  As the
97  * documentation says:  "By default, the package runs with 128 bytes of
98  * state information and generates far better random numbers than a linear
99  * congruential generator.  If the amount of state information is less than
100  * 32 bytes, a simple linear congruential R.N.G. is used."  For a buffer of
101  * 128 bytes, this new version runs about 19 percent faster and for a 16
102  * byte buffer it is about 5 percent faster.
103  */
104 
105 /*
106  * For each of the currently supported random number generators, we have a
107  * break value on the amount of state information (you need at least this
108  * many bytes of state info to support this random number generator), a degree
109  * for the polynomial (actually a trinomial) that the R.N.G. is based on, and
110  * the separation between the two lower order coefficients of the trinomial.
111  */
112 #define	TYPE_0		0		/* linear congruential */
113 #define	BREAK_0		8
114 #define	DEG_0		0
115 #define	SEP_0		0
116 
117 #define	TYPE_1		1		/* x**7 + x**3 + 1 */
118 #define	BREAK_1		32
119 #define	DEG_1		7
120 #define	SEP_1		3
121 
122 #define	TYPE_2		2		/* x**15 + x + 1 */
123 #define	BREAK_2		64
124 #define	DEG_2		15
125 #define	SEP_2		1
126 
127 #define	TYPE_3		3		/* x**31 + x**3 + 1 */
128 #define	BREAK_3		128
129 #define	DEG_3		31
130 #define	SEP_3		3
131 
132 #define	TYPE_4		4		/* x**63 + x + 1 */
133 #define	BREAK_4		256
134 #define	DEG_4		63
135 #define	SEP_4		1
136 
137 /*
138  * Array versions of the above information to make code run faster --
139  * relies on fact that TYPE_i == i.
140  */
141 #define	MAX_TYPES	5		/* max number of types above */
142 
143 static long degrees[MAX_TYPES] =	{ DEG_0, DEG_1, DEG_2, DEG_3, DEG_4 };
144 static long seps [MAX_TYPES] =	{ SEP_0, SEP_1, SEP_2, SEP_3, SEP_4 };
145 
146 /*
147  * Initially, everything is set up as if from:
148  *
149  *	initstate(1, randtbl, 128);
150  *
151  * Note that this initialization takes advantage of the fact that srandom()
152  * advances the front and rear pointers 10*rand_deg times, and hence the
153  * rear pointer which starts at 0 will also end up at zero; thus the zeroeth
154  * element of the state information, which contains info about the current
155  * position of the rear pointer is just
156  *
157  *	MAX_TYPES * (rptr - state) + TYPE_3 == TYPE_3.
158  */
159 
160 static long randtbl[DEG_3 + 1] = {
161 	TYPE_3,
162 #ifdef  USE_WEAK_SEEDING
163 /* Historic implementation compatibility */
164 /* The random sequences do not vary much with the seed */
165 	0x9a319039, 0x32d9c024, 0x9b663182, 0x5da1f342, 0xde3b81e0, 0xdf0a6fb5,
166 	0xf103bc02, 0x48f340fb, 0x7449e56b, 0xbeb1dbb0, 0xab5c5918, 0x946554fd,
167 	0x8c2e680f, 0xeb3d799f, 0xb11ee0b7, 0x2d436b86, 0xda672e2a, 0x1588ca88,
168 	0xe369735d, 0x904f35f7, 0xd7158fd6, 0x6fa6f051, 0x616e6b96, 0xac94efdc,
169 	0x36413f93, 0xc622c298, 0xf5a42ab8, 0x8a88d77b, 0xf5ad9d0e, 0x8999220b,
170 	0x27fb47b9,
171 #else   /* !USE_WEAK_SEEDING */
172 	0x991539b1, 0x16a5bce3, 0x6774a4cd, 0x3e01511e, 0x4e508aaa, 0x61048c05,
173 	0xf5500617, 0x846b7115, 0x6a19892c, 0x896a97af, 0xdb48f936, 0x14898454,
174 	0x37ffd106, 0xb58bff9c, 0x59e17104, 0xcf918a49, 0x09378c83, 0x52c7a471,
175 	0x8d293ea9, 0x1f4fc301, 0xc3db71be, 0x39b44e1c, 0xf8a44ef9, 0x4c8b80b1,
176 	0x19edc328, 0x87bf4bdd, 0xc9b240e5, 0xe9ee4b1b, 0x4382aee7, 0x535b6b41,
177 	0xf3bec5da
178 #endif  /* !USE_WEAK_SEEDING */
179 };
180 
181 /*
182  * fptr and rptr are two pointers into the state info, a front and a rear
183  * pointer.  These two pointers are always rand_sep places aparts, as they
184  * cycle cyclically through the state information.  (Yes, this does mean we
185  * could get away with just one pointer, but the code for random() is more
186  * efficient this way).  The pointers are left positioned as they would be
187  * from the call
188  *
189  *	initstate(1, randtbl, 128);
190  *
191  * (The position of the rear pointer, rptr, is really 0 (as explained above
192  * in the initialization of randtbl) because the state table pointer is set
193  * to point to randtbl[1] (as explained below).
194  */
195 static long *fptr = &randtbl[SEP_3 + 1];
196 static long *rptr = &randtbl[1];
197 
198 /*
199  * The following things are the pointer to the state information table, the
200  * type of the current generator, the degree of the current polynomial being
201  * used, and the separation between the two pointers.  Note that for efficiency
202  * of random(), we remember the first location of the state information, not
203  * the zeroeth.  Hence it is valid to access state[-1], which is used to
204  * store the type of the R.N.G.  Also, we remember the last location, since
205  * this is more efficient than indexing every time to find the address of
206  * the last element to see if the front and rear pointers have wrapped.
207  */
208 static long *state = &randtbl[1];
209 static long rand_type = TYPE_3;
210 static long rand_deg = DEG_3;
211 static long rand_sep = SEP_3;
212 static long *end_ptr = &randtbl[DEG_3 + 1];
213 
214 static inline long good_rand __P((long));
215 
216 static inline long good_rand (x)
217 	register long x;
218 {
219 #ifdef  USE_WEAK_SEEDING
220 /*
221  * Historic implementation compatibility.
222  * The random sequences do not vary much with the seed,
223  * even with overflowing.
224  */
225 	return (1103515245 * x + 12345);
226 #else   /* !USE_WEAK_SEEDING */
227 /*
228  * Compute x = (7^5 * x) mod (2^31 - 1)
229  * wihout overflowing 31 bits:
230  *      (2^31 - 1) = 127773 * (7^5) + 2836
231  * From "Random number generators: good ones are hard to find",
232  * Park and Miller, Communications of the ACM, vol. 31, no. 10,
233  * October 1988, p. 1195.
234  */
235 	register long hi, lo;
236 
237 	hi = x / 127773;
238 	lo = x % 127773;
239 	x = 16807 * lo - 2836 * hi;
240 	if (x <= 0)
241 		x += 0x7fffffff;
242 	return (x);
243 #endif  /* !USE_WEAK_SEEDING */
244 }
245 
246 /*
247  * srandom:
248  *
249  * Initialize the random number generator based on the given seed.  If the
250  * type is the trivial no-state-information type, just remember the seed.
251  * Otherwise, initializes state[] based on the given "seed" via a linear
252  * congruential generator.  Then, the pointers are set to known locations
253  * that are exactly rand_sep places apart.  Lastly, it cycles the state
254  * information a given number of times to get rid of any initial dependencies
255  * introduced by the L.C.R.N.G.  Note that the initialization of randtbl[]
256  * for default usage relies on values produced by this routine.
257  */
258 void
259 srandom(x)
260 	unsigned long x;
261 {
262 	register long i;
263 
264 	if (rand_type == TYPE_0)
265 		state[0] = x;
266 	else {
267 		state[0] = x;
268 		for (i = 1; i < rand_deg; i++)
269 			state[i] = good_rand(state[i - 1]);
270 		fptr = &state[rand_sep];
271 		rptr = &state[0];
272 		for (i = 0; i < 10 * rand_deg; i++)
273 			(void)random();
274 	}
275 }
276 
277 /*
278  * srandomdev:
279  *
280  * Many programs choose the seed value in a totally predictable manner.
281  * This often causes problems.  We seed the generator using the much more
282  * secure urandom(4) interface.  Note that this particular seeding
283  * procedure can generate states which are impossible to reproduce by
284  * calling srandom() with any value, since the succeeding terms in the
285  * state buffer are no longer derived from the LC algorithm applied to
286  * a fixed seed.
287  */
288 int
289 srandomdev()
290 {
291 	int fd;
292 	size_t len;
293 
294 	if (rand_type == TYPE_0)
295 		len = sizeof state[0];
296 	else
297 		len = rand_deg * sizeof state[0];
298 
299 	fd = open("/dev/urandom", O_RDONLY, 0);
300 	if (fd < 0)
301 		return -1;
302 	if (read(fd, (void *) state, len) < (ssize_t) len) {
303 		close(fd);
304 		return -1;
305 	}
306 	close(fd);
307 
308 	if (rand_type != TYPE_0) {
309 		fptr = &state[rand_sep];
310 		rptr = &state[0];
311 	}
312 	return 0;
313 }
314 
315 /*
316  * initstate:
317  *
318  * Initialize the state information in the given array of n bytes for future
319  * random number generation.  Based on the number of bytes we are given, and
320  * the break values for the different R.N.G.'s, we choose the best (largest)
321  * one we can and set things up for it.  srandom() is then called to
322  * initialize the state information.
323  *
324  * Note that on return from srandom(), we set state[-1] to be the type
325  * multiplexed with the current value of the rear pointer; this is so
326  * successive calls to initstate() won't lose this information and will be
327  * able to restart with setstate().
328  *
329  * Note: the first thing we do is save the current state, if any, just like
330  * setstate() so that it doesn't matter when initstate is called.
331  *
332  * Returns a pointer to the old state.
333  *
334  * Note: The Sparc platform requires that arg_state begin on a long
335  * word boundary; otherwise a bus error will occur. Even so, lint will
336  * complain about mis-alignment, but you should disregard these messages.
337  */
338 char *
339 initstate(seed, arg_state, n)
340 	unsigned long seed;		/* seed for R.N.G. */
341 	char *arg_state;		/* pointer to state array */
342 	long n;				/* # bytes of state info */
343 {
344 	register char *ostate = (char *)(&state[-1]);
345 	register long *long_arg_state = (long *) arg_state;
346 
347 	if (rand_type == TYPE_0)
348 		state[-1] = rand_type;
349 	else
350 		state[-1] = MAX_TYPES * (rptr - state) + rand_type;
351 	if (n < BREAK_0) {
352 		(void)fprintf(stderr,
353 		    "random: not enough state (%ld bytes); ignored.\n", n);
354 		return(0);
355 	}
356 	if (n < BREAK_1) {
357 		rand_type = TYPE_0;
358 		rand_deg = DEG_0;
359 		rand_sep = SEP_0;
360 	} else if (n < BREAK_2) {
361 		rand_type = TYPE_1;
362 		rand_deg = DEG_1;
363 		rand_sep = SEP_1;
364 	} else if (n < BREAK_3) {
365 		rand_type = TYPE_2;
366 		rand_deg = DEG_2;
367 		rand_sep = SEP_2;
368 	} else if (n < BREAK_4) {
369 		rand_type = TYPE_3;
370 		rand_deg = DEG_3;
371 		rand_sep = SEP_3;
372 	} else {
373 		rand_type = TYPE_4;
374 		rand_deg = DEG_4;
375 		rand_sep = SEP_4;
376 	}
377 	state = (long *) (long_arg_state + 1); /* first location */
378 	end_ptr = &state[rand_deg];	/* must set end_ptr before srandom */
379 	srandom(seed);
380 	if (rand_type == TYPE_0)
381 		long_arg_state[0] = rand_type;
382 	else
383 		long_arg_state[0] = MAX_TYPES * (rptr - state) + rand_type;
384 	return(ostate);
385 }
386 
387 /*
388  * setstate:
389  *
390  * Restore the state from the given state array.
391  *
392  * Note: it is important that we also remember the locations of the pointers
393  * in the current state information, and restore the locations of the pointers
394  * from the old state information.  This is done by multiplexing the pointer
395  * location into the zeroeth word of the state information.
396  *
397  * Note that due to the order in which things are done, it is OK to call
398  * setstate() with the same state as the current state.
399  *
400  * Returns a pointer to the old state information.
401  *
402  * Note: The Sparc platform requires that arg_state begin on a long
403  * word boundary; otherwise a bus error will occur. Even so, lint will
404  * complain about mis-alignment, but you should disregard these messages.
405  */
406 char *
407 setstate(arg_state)
408 	char *arg_state;		/* pointer to state array */
409 {
410 	register long *new_state = (long *) arg_state;
411 	register long type = new_state[0] % MAX_TYPES;
412 	register long rear = new_state[0] / MAX_TYPES;
413 	char *ostate = (char *)(&state[-1]);
414 
415 	if (rand_type == TYPE_0)
416 		state[-1] = rand_type;
417 	else
418 		state[-1] = MAX_TYPES * (rptr - state) + rand_type;
419 	switch(type) {
420 	case TYPE_0:
421 	case TYPE_1:
422 	case TYPE_2:
423 	case TYPE_3:
424 	case TYPE_4:
425 		rand_type = type;
426 		rand_deg = degrees[type];
427 		rand_sep = seps[type];
428 		break;
429 	default:
430 		(void)fprintf(stderr,
431 		    "random: state info corrupted; not changed.\n");
432 	}
433 	state = (long *) (new_state + 1);
434 	if (rand_type != TYPE_0) {
435 		rptr = &state[rear];
436 		fptr = &state[(rear + rand_sep) % rand_deg];
437 	}
438 	end_ptr = &state[rand_deg];		/* set end_ptr too */
439 	return(ostate);
440 }
441 
442 /*
443  * random:
444  *
445  * If we are using the trivial TYPE_0 R.N.G., just do the old linear
446  * congruential bit.  Otherwise, we do our fancy trinomial stuff, which is
447  * the same in all the other cases due to all the global variables that have
448  * been set up.  The basic operation is to add the number at the rear pointer
449  * into the one at the front pointer.  Then both pointers are advanced to
450  * the next location cyclically in the table.  The value returned is the sum
451  * generated, reduced to 31 bits by throwing away the "least random" low bit.
452  *
453  * Note: the code takes advantage of the fact that both the front and
454  * rear pointers can't wrap on the same call by not testing the rear
455  * pointer if the front one has wrapped.
456  *
457  * Returns a 31-bit random number.
458  */
459 long
460 random()
461 {
462 	register long i;
463 	register long *f, *r;
464 
465 	if (rand_type == TYPE_0) {
466 		i = state[0];
467 		state[0] = i = (good_rand(i)) & 0x7fffffff;
468 	} else {
469 		/*
470 		 * Use local variables rather than static variables for speed.
471 		 */
472 		f = fptr; r = rptr;
473 		*f += *r;
474 		i = (*f >> 1) & 0x7fffffff;	/* chucking least random bit */
475 		if (++f >= end_ptr) {
476 			f = state;
477 			++r;
478 		}
479 		else if (++r >= end_ptr) {
480 			r = state;
481 		}
482 
483 		fptr = f; rptr = r;
484 	}
485 	return(i);
486 }
487