xref: /freebsd/lib/libc/stdlib/random.c (revision 74bf4e164ba5851606a27d4feff27717452583e5)
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 
34 #if defined(LIBC_SCCS) && !defined(lint)
35 static char sccsid[] = "@(#)random.c	8.2 (Berkeley) 5/19/95";
36 #endif /* LIBC_SCCS and not lint */
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
39 
40 #include "namespace.h"
41 #include <sys/time.h>          /* for srandomdev() */
42 #include <fcntl.h>             /* for srandomdev() */
43 #include <stdint.h>
44 #include <stdio.h>
45 #include <stdlib.h>
46 #include <unistd.h>            /* for srandomdev() */
47 #include "un-namespace.h"
48 
49 /*
50  * random.c:
51  *
52  * An improved random number generation package.  In addition to the standard
53  * rand()/srand() like interface, this package also has a special state info
54  * interface.  The initstate() routine is called with a seed, an array of
55  * bytes, and a count of how many bytes are being passed in; this array is
56  * then initialized to contain information for random number generation with
57  * that much state information.  Good sizes for the amount of state
58  * information are 32, 64, 128, and 256 bytes.  The state can be switched by
59  * calling the setstate() routine with the same array as was initiallized
60  * with initstate().  By default, the package runs with 128 bytes of state
61  * information and generates far better random numbers than a linear
62  * congruential generator.  If the amount of state information is less than
63  * 32 bytes, a simple linear congruential R.N.G. is used.
64  *
65  * Internally, the state information is treated as an array of uint32_t's; the
66  * zeroeth element of the array is the type of R.N.G. being used (small
67  * integer); the remainder of the array is the state information for the
68  * R.N.G.  Thus, 32 bytes of state information will give 7 ints worth of
69  * state information, which will allow a degree seven polynomial.  (Note:
70  * the zeroeth word of state information also has some other information
71  * stored in it -- see setstate() for details).
72  *
73  * The random number generation technique is a linear feedback shift register
74  * approach, employing trinomials (since there are fewer terms to sum up that
75  * way).  In this approach, the least significant bit of all the numbers in
76  * the state table will act as a linear feedback shift register, and will
77  * have period 2^deg - 1 (where deg is the degree of the polynomial being
78  * used, assuming that the polynomial is irreducible and primitive).  The
79  * higher order bits will have longer periods, since their values are also
80  * influenced by pseudo-random carries out of the lower bits.  The total
81  * period of the generator is approximately deg*(2**deg - 1); thus doubling
82  * the amount of state information has a vast influence on the period of the
83  * generator.  Note: the deg*(2**deg - 1) is an approximation only good for
84  * large deg, when the period of the shift is the dominant factor.
85  * With deg equal to seven, the period is actually much longer than the
86  * 7*(2**7 - 1) predicted by this formula.
87  *
88  * Modified 28 December 1994 by Jacob S. Rosenberg.
89  * The following changes have been made:
90  * All references to the type u_int have been changed to unsigned long.
91  * All references to type int have been changed to type long.  Other
92  * cleanups have been made as well.  A warning for both initstate and
93  * setstate has been inserted to the effect that on Sparc platforms
94  * the 'arg_state' variable must be forced to begin on word boundaries.
95  * This can be easily done by casting a long integer array to char *.
96  * The overall logic has been left STRICTLY alone.  This software was
97  * tested on both a VAX and Sun SpacsStation with exactly the same
98  * results.  The new version and the original give IDENTICAL results.
99  * The new version is somewhat faster than the original.  As the
100  * documentation says:  "By default, the package runs with 128 bytes of
101  * state information and generates far better random numbers than a linear
102  * congruential generator.  If the amount of state information is less than
103  * 32 bytes, a simple linear congruential R.N.G. is used."  For a buffer of
104  * 128 bytes, this new version runs about 19 percent faster and for a 16
105  * byte buffer it is about 5 percent faster.
106  */
107 
108 /*
109  * For each of the currently supported random number generators, we have a
110  * break value on the amount of state information (you need at least this
111  * many bytes of state info to support this random number generator), a degree
112  * for the polynomial (actually a trinomial) that the R.N.G. is based on, and
113  * the separation between the two lower order coefficients of the trinomial.
114  */
115 #define	TYPE_0		0		/* linear congruential */
116 #define	BREAK_0		8
117 #define	DEG_0		0
118 #define	SEP_0		0
119 
120 #define	TYPE_1		1		/* x**7 + x**3 + 1 */
121 #define	BREAK_1		32
122 #define	DEG_1		7
123 #define	SEP_1		3
124 
125 #define	TYPE_2		2		/* x**15 + x + 1 */
126 #define	BREAK_2		64
127 #define	DEG_2		15
128 #define	SEP_2		1
129 
130 #define	TYPE_3		3		/* x**31 + x**3 + 1 */
131 #define	BREAK_3		128
132 #define	DEG_3		31
133 #define	SEP_3		3
134 
135 #define	TYPE_4		4		/* x**63 + x + 1 */
136 #define	BREAK_4		256
137 #define	DEG_4		63
138 #define	SEP_4		1
139 
140 /*
141  * Array versions of the above information to make code run faster --
142  * relies on fact that TYPE_i == i.
143  */
144 #define	MAX_TYPES	5		/* max number of types above */
145 
146 #ifdef  USE_WEAK_SEEDING
147 #define NSHUFF 0
148 #else   /* !USE_WEAK_SEEDING */
149 #define NSHUFF 50       /* to drop some "seed -> 1st value" linearity */
150 #endif  /* !USE_WEAK_SEEDING */
151 
152 static const int degrees[MAX_TYPES] =	{ DEG_0, DEG_1, DEG_2, DEG_3, DEG_4 };
153 static const int seps [MAX_TYPES] =	{ SEP_0, SEP_1, SEP_2, SEP_3, SEP_4 };
154 
155 /*
156  * Initially, everything is set up as if from:
157  *
158  *	initstate(1, randtbl, 128);
159  *
160  * Note that this initialization takes advantage of the fact that srandom()
161  * advances the front and rear pointers 10*rand_deg times, and hence the
162  * rear pointer which starts at 0 will also end up at zero; thus the zeroeth
163  * element of the state information, which contains info about the current
164  * position of the rear pointer is just
165  *
166  *	MAX_TYPES * (rptr - state) + TYPE_3 == TYPE_3.
167  */
168 
169 static uint32_t randtbl[DEG_3 + 1] = {
170 	TYPE_3,
171 #ifdef  USE_WEAK_SEEDING
172 /* Historic implementation compatibility */
173 /* The random sequences do not vary much with the seed */
174 	0x9a319039, 0x32d9c024, 0x9b663182, 0x5da1f342, 0xde3b81e0, 0xdf0a6fb5,
175 	0xf103bc02, 0x48f340fb, 0x7449e56b, 0xbeb1dbb0, 0xab5c5918, 0x946554fd,
176 	0x8c2e680f, 0xeb3d799f, 0xb11ee0b7, 0x2d436b86, 0xda672e2a, 0x1588ca88,
177 	0xe369735d, 0x904f35f7, 0xd7158fd6, 0x6fa6f051, 0x616e6b96, 0xac94efdc,
178 	0x36413f93, 0xc622c298, 0xf5a42ab8, 0x8a88d77b, 0xf5ad9d0e, 0x8999220b,
179 	0x27fb47b9,
180 #else   /* !USE_WEAK_SEEDING */
181 	0x991539b1, 0x16a5bce3, 0x6774a4cd, 0x3e01511e, 0x4e508aaa, 0x61048c05,
182 	0xf5500617, 0x846b7115, 0x6a19892c, 0x896a97af, 0xdb48f936, 0x14898454,
183 	0x37ffd106, 0xb58bff9c, 0x59e17104, 0xcf918a49, 0x09378c83, 0x52c7a471,
184 	0x8d293ea9, 0x1f4fc301, 0xc3db71be, 0x39b44e1c, 0xf8a44ef9, 0x4c8b80b1,
185 	0x19edc328, 0x87bf4bdd, 0xc9b240e5, 0xe9ee4b1b, 0x4382aee7, 0x535b6b41,
186 	0xf3bec5da
187 #endif  /* !USE_WEAK_SEEDING */
188 };
189 
190 /*
191  * fptr and rptr are two pointers into the state info, a front and a rear
192  * pointer.  These two pointers are always rand_sep places aparts, as they
193  * cycle cyclically through the state information.  (Yes, this does mean we
194  * could get away with just one pointer, but the code for random() is more
195  * efficient this way).  The pointers are left positioned as they would be
196  * from the call
197  *
198  *	initstate(1, randtbl, 128);
199  *
200  * (The position of the rear pointer, rptr, is really 0 (as explained above
201  * in the initialization of randtbl) because the state table pointer is set
202  * to point to randtbl[1] (as explained below).
203  */
204 static uint32_t *fptr = &randtbl[SEP_3 + 1];
205 static uint32_t *rptr = &randtbl[1];
206 
207 /*
208  * The following things are the pointer to the state information table, the
209  * type of the current generator, the degree of the current polynomial being
210  * used, and the separation between the two pointers.  Note that for efficiency
211  * of random(), we remember the first location of the state information, not
212  * the zeroeth.  Hence it is valid to access state[-1], which is used to
213  * store the type of the R.N.G.  Also, we remember the last location, since
214  * this is more efficient than indexing every time to find the address of
215  * the last element to see if the front and rear pointers have wrapped.
216  */
217 static uint32_t *state = &randtbl[1];
218 static int rand_type = TYPE_3;
219 static int rand_deg = DEG_3;
220 static int rand_sep = SEP_3;
221 static uint32_t *end_ptr = &randtbl[DEG_3 + 1];
222 
223 static inline uint32_t good_rand(int32_t);
224 
225 static inline uint32_t good_rand (x)
226 	int32_t x;
227 {
228 #ifdef  USE_WEAK_SEEDING
229 /*
230  * Historic implementation compatibility.
231  * The random sequences do not vary much with the seed,
232  * even with overflowing.
233  */
234 	return (1103515245 * x + 12345);
235 #else   /* !USE_WEAK_SEEDING */
236 /*
237  * Compute x = (7^5 * x) mod (2^31 - 1)
238  * wihout overflowing 31 bits:
239  *      (2^31 - 1) = 127773 * (7^5) + 2836
240  * From "Random number generators: good ones are hard to find",
241  * Park and Miller, Communications of the ACM, vol. 31, no. 10,
242  * October 1988, p. 1195.
243  */
244 	int32_t hi, lo;
245 
246 	/* Can't be initialized with 0, so use another value. */
247 	if (x == 0)
248 		x = 123459876;
249 	hi = x / 127773;
250 	lo = x % 127773;
251 	x = 16807 * lo - 2836 * hi;
252 	if (x < 0)
253 		x += 0x7fffffff;
254 	return (x);
255 #endif  /* !USE_WEAK_SEEDING */
256 }
257 
258 /*
259  * srandom:
260  *
261  * Initialize the random number generator based on the given seed.  If the
262  * type is the trivial no-state-information type, just remember the seed.
263  * Otherwise, initializes state[] based on the given "seed" via a linear
264  * congruential generator.  Then, the pointers are set to known locations
265  * that are exactly rand_sep places apart.  Lastly, it cycles the state
266  * information a given number of times to get rid of any initial dependencies
267  * introduced by the L.C.R.N.G.  Note that the initialization of randtbl[]
268  * for default usage relies on values produced by this routine.
269  */
270 void
271 srandom(x)
272 	unsigned long x;
273 {
274 	int i, lim;
275 
276 	state[0] = (uint32_t)x;
277 	if (rand_type == TYPE_0)
278 		lim = NSHUFF;
279 	else {
280 		for (i = 1; i < rand_deg; i++)
281 			state[i] = good_rand(state[i - 1]);
282 		fptr = &state[rand_sep];
283 		rptr = &state[0];
284 		lim = 10 * rand_deg;
285 	}
286 	for (i = 0; i < lim; i++)
287 		(void)random();
288 }
289 
290 /*
291  * srandomdev:
292  *
293  * Many programs choose the seed value in a totally predictable manner.
294  * This often causes problems.  We seed the generator using the much more
295  * secure random(4) interface.  Note that this particular seeding
296  * procedure can generate states which are impossible to reproduce by
297  * calling srandom() with any value, since the succeeding terms in the
298  * state buffer are no longer derived from the LC algorithm applied to
299  * a fixed seed.
300  */
301 void
302 srandomdev()
303 {
304 	int fd, done;
305 	size_t len;
306 
307 	if (rand_type == TYPE_0)
308 		len = sizeof state[0];
309 	else
310 		len = rand_deg * sizeof state[0];
311 
312 	done = 0;
313 	fd = _open("/dev/random", O_RDONLY, 0);
314 	if (fd >= 0) {
315 		if (_read(fd, (void *) state, len) == (ssize_t) len)
316 			done = 1;
317 		_close(fd);
318 	}
319 
320 	if (!done) {
321 		struct timeval tv;
322 		unsigned long junk;
323 
324 		gettimeofday(&tv, NULL);
325 		srandom((getpid() << 16) ^ tv.tv_sec ^ tv.tv_usec ^ junk);
326 		return;
327 	}
328 
329 	if (rand_type != TYPE_0) {
330 		fptr = &state[rand_sep];
331 		rptr = &state[0];
332 	}
333 }
334 
335 /*
336  * initstate:
337  *
338  * Initialize the state information in the given array of n bytes for future
339  * random number generation.  Based on the number of bytes we are given, and
340  * the break values for the different R.N.G.'s, we choose the best (largest)
341  * one we can and set things up for it.  srandom() is then called to
342  * initialize the state information.
343  *
344  * Note that on return from srandom(), we set state[-1] to be the type
345  * multiplexed with the current value of the rear pointer; this is so
346  * successive calls to initstate() won't lose this information and will be
347  * able to restart with setstate().
348  *
349  * Note: the first thing we do is save the current state, if any, just like
350  * setstate() so that it doesn't matter when initstate is called.
351  *
352  * Returns a pointer to the old state.
353  *
354  * Note: The Sparc platform requires that arg_state begin on an int
355  * word boundary; otherwise a bus error will occur. Even so, lint will
356  * complain about mis-alignment, but you should disregard these messages.
357  */
358 char *
359 initstate(seed, arg_state, n)
360 	unsigned long seed;		/* seed for R.N.G. */
361 	char *arg_state;		/* pointer to state array */
362 	long n;				/* # bytes of state info */
363 {
364 	char *ostate = (char *)(&state[-1]);
365 	uint32_t *int_arg_state = (uint32_t *)arg_state;
366 
367 	if (rand_type == TYPE_0)
368 		state[-1] = rand_type;
369 	else
370 		state[-1] = MAX_TYPES * (rptr - state) + rand_type;
371 	if (n < BREAK_0) {
372 		(void)fprintf(stderr,
373 		    "random: not enough state (%ld bytes); ignored.\n", n);
374 		return(0);
375 	}
376 	if (n < BREAK_1) {
377 		rand_type = TYPE_0;
378 		rand_deg = DEG_0;
379 		rand_sep = SEP_0;
380 	} else if (n < BREAK_2) {
381 		rand_type = TYPE_1;
382 		rand_deg = DEG_1;
383 		rand_sep = SEP_1;
384 	} else if (n < BREAK_3) {
385 		rand_type = TYPE_2;
386 		rand_deg = DEG_2;
387 		rand_sep = SEP_2;
388 	} else if (n < BREAK_4) {
389 		rand_type = TYPE_3;
390 		rand_deg = DEG_3;
391 		rand_sep = SEP_3;
392 	} else {
393 		rand_type = TYPE_4;
394 		rand_deg = DEG_4;
395 		rand_sep = SEP_4;
396 	}
397 	state = int_arg_state + 1; /* first location */
398 	end_ptr = &state[rand_deg];	/* must set end_ptr before srandom */
399 	srandom(seed);
400 	if (rand_type == TYPE_0)
401 		int_arg_state[0] = rand_type;
402 	else
403 		int_arg_state[0] = MAX_TYPES * (rptr - state) + rand_type;
404 	return(ostate);
405 }
406 
407 /*
408  * setstate:
409  *
410  * Restore the state from the given state array.
411  *
412  * Note: it is important that we also remember the locations of the pointers
413  * in the current state information, and restore the locations of the pointers
414  * from the old state information.  This is done by multiplexing the pointer
415  * location into the zeroeth word of the state information.
416  *
417  * Note that due to the order in which things are done, it is OK to call
418  * setstate() with the same state as the current state.
419  *
420  * Returns a pointer to the old state information.
421  *
422  * Note: The Sparc platform requires that arg_state begin on an int
423  * word boundary; otherwise a bus error will occur. Even so, lint will
424  * complain about mis-alignment, but you should disregard these messages.
425  */
426 char *
427 setstate(arg_state)
428 	char *arg_state;		/* pointer to state array */
429 {
430 	uint32_t *new_state = (uint32_t *)arg_state;
431 	uint32_t type = new_state[0] % MAX_TYPES;
432 	uint32_t rear = new_state[0] / MAX_TYPES;
433 	char *ostate = (char *)(&state[-1]);
434 
435 	if (rand_type == TYPE_0)
436 		state[-1] = rand_type;
437 	else
438 		state[-1] = MAX_TYPES * (rptr - state) + rand_type;
439 	switch(type) {
440 	case TYPE_0:
441 	case TYPE_1:
442 	case TYPE_2:
443 	case TYPE_3:
444 	case TYPE_4:
445 		rand_type = type;
446 		rand_deg = degrees[type];
447 		rand_sep = seps[type];
448 		break;
449 	default:
450 		(void)fprintf(stderr,
451 		    "random: state info corrupted; not changed.\n");
452 	}
453 	state = new_state + 1;
454 	if (rand_type != TYPE_0) {
455 		rptr = &state[rear];
456 		fptr = &state[(rear + rand_sep) % rand_deg];
457 	}
458 	end_ptr = &state[rand_deg];		/* set end_ptr too */
459 	return(ostate);
460 }
461 
462 /*
463  * random:
464  *
465  * If we are using the trivial TYPE_0 R.N.G., just do the old linear
466  * congruential bit.  Otherwise, we do our fancy trinomial stuff, which is
467  * the same in all the other cases due to all the global variables that have
468  * been set up.  The basic operation is to add the number at the rear pointer
469  * into the one at the front pointer.  Then both pointers are advanced to
470  * the next location cyclically in the table.  The value returned is the sum
471  * generated, reduced to 31 bits by throwing away the "least random" low bit.
472  *
473  * Note: the code takes advantage of the fact that both the front and
474  * rear pointers can't wrap on the same call by not testing the rear
475  * pointer if the front one has wrapped.
476  *
477  * Returns a 31-bit random number.
478  */
479 long
480 random()
481 {
482 	uint32_t i;
483 	uint32_t *f, *r;
484 
485 	if (rand_type == TYPE_0) {
486 		i = state[0];
487 		state[0] = i = (good_rand(i)) & 0x7fffffff;
488 	} else {
489 		/*
490 		 * Use local variables rather than static variables for speed.
491 		 */
492 		f = fptr; r = rptr;
493 		*f += *r;
494 		i = (*f >> 1) & 0x7fffffff;	/* chucking least random bit */
495 		if (++f >= end_ptr) {
496 			f = state;
497 			++r;
498 		}
499 		else if (++r >= end_ptr) {
500 			r = state;
501 		}
502 
503 		fptr = f; rptr = r;
504 	}
505 	return((long)i);
506 }
507