xref: /freebsd/lib/libc/stdlib/random.c (revision e14ddd1f16e7e5788392c50de21ea7c927e0690c)
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  * 4. Neither the name of the University nor the names of its contributors
14  *    may be used to endorse or promote products derived from this software
15  *    without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #if defined(LIBC_SCCS) && !defined(lint)
31 static char sccsid[] = "@(#)random.c	8.2 (Berkeley) 5/19/95";
32 #endif /* LIBC_SCCS and not lint */
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 #include "namespace.h"
37 #include <sys/param.h>
38 #include <sys/sysctl.h>
39 #include <stdint.h>
40 #include <stdio.h>
41 #include <stdlib.h>
42 #include "un-namespace.h"
43 
44 /*
45  * random.c:
46  *
47  * An improved random number generation package.  In addition to the standard
48  * rand()/srand() like interface, this package also has a special state info
49  * interface.  The initstate() routine is called with a seed, an array of
50  * bytes, and a count of how many bytes are being passed in; this array is
51  * then initialized to contain information for random number generation with
52  * that much state information.  Good sizes for the amount of state
53  * information are 32, 64, 128, and 256 bytes.  The state can be switched by
54  * calling the setstate() routine with the same array as was initiallized
55  * with initstate().  By default, the package runs with 128 bytes of state
56  * information and generates far better random numbers than a linear
57  * congruential generator.  If the amount of state information is less than
58  * 32 bytes, a simple linear congruential R.N.G. is used.
59  *
60  * Internally, the state information is treated as an array of uint32_t's; the
61  * zeroeth element of the array is the type of R.N.G. being used (small
62  * integer); the remainder of the array is the state information for the
63  * R.N.G.  Thus, 32 bytes of state information will give 7 ints worth of
64  * state information, which will allow a degree seven polynomial.  (Note:
65  * the zeroeth word of state information also has some other information
66  * stored in it -- see setstate() for details).
67  *
68  * The random number generation technique is a linear feedback shift register
69  * approach, employing trinomials (since there are fewer terms to sum up that
70  * way).  In this approach, the least significant bit of all the numbers in
71  * the state table will act as a linear feedback shift register, and will
72  * have period 2^deg - 1 (where deg is the degree of the polynomial being
73  * used, assuming that the polynomial is irreducible and primitive).  The
74  * higher order bits will have longer periods, since their values are also
75  * influenced by pseudo-random carries out of the lower bits.  The total
76  * period of the generator is approximately deg*(2**deg - 1); thus doubling
77  * the amount of state information has a vast influence on the period of the
78  * generator.  Note: the deg*(2**deg - 1) is an approximation only good for
79  * large deg, when the period of the shift is the dominant factor.
80  * With deg equal to seven, the period is actually much longer than the
81  * 7*(2**7 - 1) predicted by this formula.
82  *
83  * Modified 28 December 1994 by Jacob S. Rosenberg.
84  * The following changes have been made:
85  * All references to the type u_int have been changed to unsigned long.
86  * All references to type int have been changed to type long.  Other
87  * cleanups have been made as well.  A warning for both initstate and
88  * setstate has been inserted to the effect that on Sparc platforms
89  * the 'arg_state' variable must be forced to begin on word boundaries.
90  * This can be easily done by casting a long integer array to char *.
91  * The overall logic has been left STRICTLY alone.  This software was
92  * tested on both a VAX and Sun SpacsStation with exactly the same
93  * results.  The new version and the original give IDENTICAL results.
94  * The new version is somewhat faster than the original.  As the
95  * documentation says:  "By default, the package runs with 128 bytes of
96  * state information and generates far better random numbers than a linear
97  * congruential generator.  If the amount of state information is less than
98  * 32 bytes, a simple linear congruential R.N.G. is used."  For a buffer of
99  * 128 bytes, this new version runs about 19 percent faster and for a 16
100  * byte buffer it is about 5 percent faster.
101  */
102 
103 /*
104  * For each of the currently supported random number generators, we have a
105  * break value on the amount of state information (you need at least this
106  * many bytes of state info to support this random number generator), a degree
107  * for the polynomial (actually a trinomial) that the R.N.G. is based on, and
108  * the separation between the two lower order coefficients of the trinomial.
109  */
110 #define	TYPE_0		0		/* linear congruential */
111 #define	BREAK_0		8
112 #define	DEG_0		0
113 #define	SEP_0		0
114 
115 #define	TYPE_1		1		/* x**7 + x**3 + 1 */
116 #define	BREAK_1		32
117 #define	DEG_1		7
118 #define	SEP_1		3
119 
120 #define	TYPE_2		2		/* x**15 + x + 1 */
121 #define	BREAK_2		64
122 #define	DEG_2		15
123 #define	SEP_2		1
124 
125 #define	TYPE_3		3		/* x**31 + x**3 + 1 */
126 #define	BREAK_3		128
127 #define	DEG_3		31
128 #define	SEP_3		3
129 
130 #define	TYPE_4		4		/* x**63 + x + 1 */
131 #define	BREAK_4		256
132 #define	DEG_4		63
133 #define	SEP_4		1
134 
135 /*
136  * Array versions of the above information to make code run faster --
137  * relies on fact that TYPE_i == i.
138  */
139 #define	MAX_TYPES	5		/* max number of types above */
140 
141 #ifdef  USE_WEAK_SEEDING
142 #define NSHUFF 0
143 #else   /* !USE_WEAK_SEEDING */
144 #define NSHUFF 50       /* to drop some "seed -> 1st value" linearity */
145 #endif  /* !USE_WEAK_SEEDING */
146 
147 static const int degrees[MAX_TYPES] =	{ DEG_0, DEG_1, DEG_2, DEG_3, DEG_4 };
148 static const int seps [MAX_TYPES] =	{ SEP_0, SEP_1, SEP_2, SEP_3, SEP_4 };
149 
150 /*
151  * Initially, everything is set up as if from:
152  *
153  *	initstate(1, randtbl, 128);
154  *
155  * Note that this initialization takes advantage of the fact that srandom()
156  * advances the front and rear pointers 10*rand_deg times, and hence the
157  * rear pointer which starts at 0 will also end up at zero; thus the zeroeth
158  * element of the state information, which contains info about the current
159  * position of the rear pointer is just
160  *
161  *	MAX_TYPES * (rptr - state) + TYPE_3 == TYPE_3.
162  */
163 
164 static uint32_t randtbl[DEG_3 + 1] = {
165 	TYPE_3,
166 #ifdef  USE_WEAK_SEEDING
167 /* Historic implementation compatibility */
168 /* The random sequences do not vary much with the seed */
169 	0x9a319039, 0x32d9c024, 0x9b663182, 0x5da1f342, 0xde3b81e0, 0xdf0a6fb5,
170 	0xf103bc02, 0x48f340fb, 0x7449e56b, 0xbeb1dbb0, 0xab5c5918, 0x946554fd,
171 	0x8c2e680f, 0xeb3d799f, 0xb11ee0b7, 0x2d436b86, 0xda672e2a, 0x1588ca88,
172 	0xe369735d, 0x904f35f7, 0xd7158fd6, 0x6fa6f051, 0x616e6b96, 0xac94efdc,
173 	0x36413f93, 0xc622c298, 0xf5a42ab8, 0x8a88d77b, 0xf5ad9d0e, 0x8999220b,
174 	0x27fb47b9,
175 #else   /* !USE_WEAK_SEEDING */
176 	0x991539b1, 0x16a5bce3, 0x6774a4cd, 0x3e01511e, 0x4e508aaa, 0x61048c05,
177 	0xf5500617, 0x846b7115, 0x6a19892c, 0x896a97af, 0xdb48f936, 0x14898454,
178 	0x37ffd106, 0xb58bff9c, 0x59e17104, 0xcf918a49, 0x09378c83, 0x52c7a471,
179 	0x8d293ea9, 0x1f4fc301, 0xc3db71be, 0x39b44e1c, 0xf8a44ef9, 0x4c8b80b1,
180 	0x19edc328, 0x87bf4bdd, 0xc9b240e5, 0xe9ee4b1b, 0x4382aee7, 0x535b6b41,
181 	0xf3bec5da
182 #endif  /* !USE_WEAK_SEEDING */
183 };
184 
185 /*
186  * fptr and rptr are two pointers into the state info, a front and a rear
187  * pointer.  These two pointers are always rand_sep places aparts, as they
188  * cycle cyclically through the state information.  (Yes, this does mean we
189  * could get away with just one pointer, but the code for random() is more
190  * efficient this way).  The pointers are left positioned as they would be
191  * from the call
192  *
193  *	initstate(1, randtbl, 128);
194  *
195  * (The position of the rear pointer, rptr, is really 0 (as explained above
196  * in the initialization of randtbl) because the state table pointer is set
197  * to point to randtbl[1] (as explained below).
198  */
199 static uint32_t *fptr = &randtbl[SEP_3 + 1];
200 static uint32_t *rptr = &randtbl[1];
201 
202 /*
203  * The following things are the pointer to the state information table, the
204  * type of the current generator, the degree of the current polynomial being
205  * used, and the separation between the two pointers.  Note that for efficiency
206  * of random(), we remember the first location of the state information, not
207  * the zeroeth.  Hence it is valid to access state[-1], which is used to
208  * store the type of the R.N.G.  Also, we remember the last location, since
209  * this is more efficient than indexing every time to find the address of
210  * the last element to see if the front and rear pointers have wrapped.
211  */
212 static uint32_t *state = &randtbl[1];
213 static int rand_type = TYPE_3;
214 static int rand_deg = DEG_3;
215 static int rand_sep = SEP_3;
216 static uint32_t *end_ptr = &randtbl[DEG_3 + 1];
217 
218 static inline uint32_t
219 good_rand(int32_t x)
220 {
221 #ifdef  USE_WEAK_SEEDING
222 /*
223  * Historic implementation compatibility.
224  * The random sequences do not vary much with the seed,
225  * even with overflowing.
226  */
227 	return (1103515245 * x + 12345);
228 #else   /* !USE_WEAK_SEEDING */
229 /*
230  * Compute x = (7^5 * x) mod (2^31 - 1)
231  * wihout overflowing 31 bits:
232  *      (2^31 - 1) = 127773 * (7^5) + 2836
233  * From "Random number generators: good ones are hard to find",
234  * Park and Miller, Communications of the ACM, vol. 31, no. 10,
235  * October 1988, p. 1195.
236  */
237 	int32_t hi, lo;
238 
239 	/* Can't be initialized with 0, so use another value. */
240 	if (x == 0)
241 		x = 123459876;
242 	hi = x / 127773;
243 	lo = x % 127773;
244 	x = 16807 * lo - 2836 * hi;
245 	if (x < 0)
246 		x += 0x7fffffff;
247 	return (x);
248 #endif  /* !USE_WEAK_SEEDING */
249 }
250 
251 /*
252  * srandom:
253  *
254  * Initialize the random number generator based on the given seed.  If the
255  * type is the trivial no-state-information type, just remember the seed.
256  * Otherwise, initializes state[] based on the given "seed" via a linear
257  * congruential generator.  Then, the pointers are set to known locations
258  * that are exactly rand_sep places apart.  Lastly, it cycles the state
259  * information a given number of times to get rid of any initial dependencies
260  * introduced by the L.C.R.N.G.  Note that the initialization of randtbl[]
261  * for default usage relies on values produced by this routine.
262  */
263 void
264 srandom(unsigned long x)
265 {
266 	int i, lim;
267 
268 	state[0] = (uint32_t)x;
269 	if (rand_type == TYPE_0)
270 		lim = NSHUFF;
271 	else {
272 		for (i = 1; i < rand_deg; i++)
273 			state[i] = good_rand(state[i - 1]);
274 		fptr = &state[rand_sep];
275 		rptr = &state[0];
276 		lim = 10 * rand_deg;
277 	}
278 	for (i = 0; i < lim; i++)
279 		(void)random();
280 }
281 
282 /*
283  * srandomdev:
284  *
285  * Many programs choose the seed value in a totally predictable manner.
286  * This often causes problems.  We seed the generator using pseudo-random
287  * data from the kernel.
288  *
289  * Note that this particular seeding procedure can generate states
290  * which are impossible to reproduce by calling srandom() with any
291  * value, since the succeeding terms in the state buffer are no longer
292  * derived from the LC algorithm applied to a fixed seed.
293  */
294 void
295 srandomdev(void)
296 {
297 	int mib[2];
298 	size_t len;
299 
300 	if (rand_type == TYPE_0)
301 		len = sizeof(state[0]);
302 	else
303 		len = rand_deg * sizeof(state[0]);
304 
305 	mib[0] = CTL_KERN;
306 	mib[1] = KERN_ARND;
307 	sysctl(mib, 2, state, &len, NULL, 0);
308 
309 	if (rand_type != TYPE_0) {
310 		fptr = &state[rand_sep];
311 		rptr = &state[0];
312 	}
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 an int
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(unsigned long seed, char *arg_state, long n)
340 {
341 	char *ostate = (char *)(&state[-1]);
342 	uint32_t *int_arg_state = (uint32_t *)arg_state;
343 
344 	if (rand_type == TYPE_0)
345 		state[-1] = rand_type;
346 	else
347 		state[-1] = MAX_TYPES * (rptr - state) + rand_type;
348 	if (n < BREAK_0) {
349 		(void)fprintf(stderr,
350 		    "random: not enough state (%ld bytes); ignored.\n", n);
351 		return (0);
352 	}
353 	if (n < BREAK_1) {
354 		rand_type = TYPE_0;
355 		rand_deg = DEG_0;
356 		rand_sep = SEP_0;
357 	} else if (n < BREAK_2) {
358 		rand_type = TYPE_1;
359 		rand_deg = DEG_1;
360 		rand_sep = SEP_1;
361 	} else if (n < BREAK_3) {
362 		rand_type = TYPE_2;
363 		rand_deg = DEG_2;
364 		rand_sep = SEP_2;
365 	} else if (n < BREAK_4) {
366 		rand_type = TYPE_3;
367 		rand_deg = DEG_3;
368 		rand_sep = SEP_3;
369 	} else {
370 		rand_type = TYPE_4;
371 		rand_deg = DEG_4;
372 		rand_sep = SEP_4;
373 	}
374 	state = int_arg_state + 1; /* first location */
375 	end_ptr = &state[rand_deg];	/* must set end_ptr before srandom */
376 	srandom(seed);
377 	if (rand_type == TYPE_0)
378 		int_arg_state[0] = rand_type;
379 	else
380 		int_arg_state[0] = MAX_TYPES * (rptr - state) + rand_type;
381 	return (ostate);
382 }
383 
384 /*
385  * setstate:
386  *
387  * Restore the state from the given state array.
388  *
389  * Note: it is important that we also remember the locations of the pointers
390  * in the current state information, and restore the locations of the pointers
391  * from the old state information.  This is done by multiplexing the pointer
392  * location into the zeroeth word of the state information.
393  *
394  * Note that due to the order in which things are done, it is OK to call
395  * setstate() with the same state as the current state.
396  *
397  * Returns a pointer to the old state information.
398  *
399  * Note: The Sparc platform requires that arg_state begin on an int
400  * word boundary; otherwise a bus error will occur. Even so, lint will
401  * complain about mis-alignment, but you should disregard these messages.
402  */
403 char *
404 setstate(char *arg_state)
405 {
406 	uint32_t *new_state = (uint32_t *)arg_state;
407 	uint32_t type = new_state[0] % MAX_TYPES;
408 	uint32_t rear = new_state[0] / MAX_TYPES;
409 	char *ostate = (char *)(&state[-1]);
410 
411 	if (rand_type == TYPE_0)
412 		state[-1] = rand_type;
413 	else
414 		state[-1] = MAX_TYPES * (rptr - state) + rand_type;
415 	switch(type) {
416 	case TYPE_0:
417 	case TYPE_1:
418 	case TYPE_2:
419 	case TYPE_3:
420 	case TYPE_4:
421 		rand_type = type;
422 		rand_deg = degrees[type];
423 		rand_sep = seps[type];
424 		break;
425 	default:
426 		(void)fprintf(stderr,
427 		    "random: state info corrupted; not changed.\n");
428 	}
429 	state = new_state + 1;
430 	if (rand_type != TYPE_0) {
431 		rptr = &state[rear];
432 		fptr = &state[(rear + rand_sep) % rand_deg];
433 	}
434 	end_ptr = &state[rand_deg];		/* set end_ptr too */
435 	return (ostate);
436 }
437 
438 /*
439  * random:
440  *
441  * If we are using the trivial TYPE_0 R.N.G., just do the old linear
442  * congruential bit.  Otherwise, we do our fancy trinomial stuff, which is
443  * the same in all the other cases due to all the global variables that have
444  * been set up.  The basic operation is to add the number at the rear pointer
445  * into the one at the front pointer.  Then both pointers are advanced to
446  * the next location cyclically in the table.  The value returned is the sum
447  * generated, reduced to 31 bits by throwing away the "least random" low bit.
448  *
449  * Note: the code takes advantage of the fact that both the front and
450  * rear pointers can't wrap on the same call by not testing the rear
451  * pointer if the front one has wrapped.
452  *
453  * Returns a 31-bit random number.
454  */
455 long
456 random(void)
457 {
458 	uint32_t i;
459 	uint32_t *f, *r;
460 
461 	if (rand_type == TYPE_0) {
462 		i = state[0];
463 		state[0] = i = (good_rand(i)) & 0x7fffffff;
464 	} else {
465 		/*
466 		 * Use local variables rather than static variables for speed.
467 		 */
468 		f = fptr; r = rptr;
469 		*f += *r;
470 		i = (*f >> 1) & 0x7fffffff;	/* chucking least random bit */
471 		if (++f >= end_ptr) {
472 			f = state;
473 			++r;
474 		}
475 		else if (++r >= end_ptr) {
476 			r = state;
477 		}
478 
479 		fptr = f; rptr = r;
480 	}
481 	return ((long)i);
482 }
483