xref: /freebsd/lib/libc/stdlib/random.c (revision 780fb4a2)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1983, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 #if defined(LIBC_SCCS) && !defined(lint)
33 static char sccsid[] = "@(#)random.c	8.2 (Berkeley) 5/19/95";
34 #endif /* LIBC_SCCS and not lint */
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include "namespace.h"
39 #include <sys/param.h>
40 #include <sys/sysctl.h>
41 #include <stdint.h>
42 #include <stdlib.h>
43 #include "un-namespace.h"
44 
45 /*
46  * random.c:
47  *
48  * An improved random number generation package.  In addition to the standard
49  * rand()/srand() like interface, this package also has a special state info
50  * interface.  The initstate() routine is called with a seed, an array of
51  * bytes, and a count of how many bytes are being passed in; this array is
52  * then initialized to contain information for random number generation with
53  * that much state information.  Good sizes for the amount of state
54  * information are 32, 64, 128, and 256 bytes.  The state can be switched by
55  * calling the setstate() routine with the same array as was initiallized
56  * with initstate().  By default, the package runs with 128 bytes of state
57  * information and generates far better random numbers than a linear
58  * congruential generator.  If the amount of state information is less than
59  * 32 bytes, a simple linear congruential R.N.G. is used.
60  *
61  * Internally, the state information is treated as an array of uint32_t's; the
62  * zeroeth element of the array is the type of R.N.G. being used (small
63  * integer); the remainder of the array is the state information for the
64  * R.N.G.  Thus, 32 bytes of state information will give 7 ints worth of
65  * state information, which will allow a degree seven polynomial.  (Note:
66  * the zeroeth word of state information also has some other information
67  * stored in it -- see setstate() for details).
68  *
69  * The random number generation technique is a linear feedback shift register
70  * approach, employing trinomials (since there are fewer terms to sum up that
71  * way).  In this approach, the least significant bit of all the numbers in
72  * the state table will act as a linear feedback shift register, and will
73  * have period 2^deg - 1 (where deg is the degree of the polynomial being
74  * used, assuming that the polynomial is irreducible and primitive).  The
75  * higher order bits will have longer periods, since their values are also
76  * influenced by pseudo-random carries out of the lower bits.  The total
77  * period of the generator is approximately deg*(2**deg - 1); thus doubling
78  * the amount of state information has a vast influence on the period of the
79  * generator.  Note: the deg*(2**deg - 1) is an approximation only good for
80  * large deg, when the period of the shift is the dominant factor.
81  * With deg equal to seven, the period is actually much longer than the
82  * 7*(2**7 - 1) predicted by this formula.
83  *
84  * Modified 28 December 1994 by Jacob S. Rosenberg.
85  * The following changes have been made:
86  * All references to the type u_int have been changed to unsigned long.
87  * All references to type int have been changed to type long.  Other
88  * cleanups have been made as well.  A warning for both initstate and
89  * setstate has been inserted to the effect that on Sparc platforms
90  * the 'arg_state' variable must be forced to begin on word boundaries.
91  * This can be easily done by casting a long integer array to char *.
92  * The overall logic has been left STRICTLY alone.  This software was
93  * tested on both a VAX and Sun SpacsStation with exactly the same
94  * results.  The new version and the original give IDENTICAL results.
95  * The new version is somewhat faster than the original.  As the
96  * documentation says:  "By default, the package runs with 128 bytes of
97  * state information and generates far better random numbers than a linear
98  * congruential generator.  If the amount of state information is less than
99  * 32 bytes, a simple linear congruential R.N.G. is used."  For a buffer of
100  * 128 bytes, this new version runs about 19 percent faster and for a 16
101  * byte buffer it is about 5 percent faster.
102  */
103 
104 /*
105  * For each of the currently supported random number generators, we have a
106  * break value on the amount of state information (you need at least this
107  * many bytes of state info to support this random number generator), a degree
108  * for the polynomial (actually a trinomial) that the R.N.G. is based on, and
109  * the separation between the two lower order coefficients of the trinomial.
110  */
111 #define	TYPE_0		0		/* linear congruential */
112 #define	BREAK_0		8
113 #define	DEG_0		0
114 #define	SEP_0		0
115 
116 #define	TYPE_1		1		/* x**7 + x**3 + 1 */
117 #define	BREAK_1		32
118 #define	DEG_1		7
119 #define	SEP_1		3
120 
121 #define	TYPE_2		2		/* x**15 + x + 1 */
122 #define	BREAK_2		64
123 #define	DEG_2		15
124 #define	SEP_2		1
125 
126 #define	TYPE_3		3		/* x**31 + x**3 + 1 */
127 #define	BREAK_3		128
128 #define	DEG_3		31
129 #define	SEP_3		3
130 
131 #define	TYPE_4		4		/* x**63 + x + 1 */
132 #define	BREAK_4		256
133 #define	DEG_4		63
134 #define	SEP_4		1
135 
136 /*
137  * Array versions of the above information to make code run faster --
138  * relies on fact that TYPE_i == i.
139  */
140 #define	MAX_TYPES	5		/* max number of types above */
141 
142 #define NSHUFF 50       /* to drop some "seed -> 1st value" linearity */
143 
144 static const int degrees[MAX_TYPES] =	{ DEG_0, DEG_1, DEG_2, DEG_3, DEG_4 };
145 static const int seps [MAX_TYPES] =	{ SEP_0, SEP_1, SEP_2, SEP_3, SEP_4 };
146 
147 /*
148  * Initially, everything is set up as if from:
149  *
150  *	initstate(1, randtbl, 128);
151  *
152  * Note that this initialization takes advantage of the fact that srandom()
153  * advances the front and rear pointers 10*rand_deg times, and hence the
154  * rear pointer which starts at 0 will also end up at zero; thus the zeroeth
155  * element of the state information, which contains info about the current
156  * position of the rear pointer is just
157  *
158  *	MAX_TYPES * (rptr - state) + TYPE_3 == TYPE_3.
159  */
160 
161 static uint32_t randtbl[DEG_3 + 1] = {
162 	TYPE_3,
163 	0x2cf41758, 0x27bb3711, 0x4916d4d1, 0x7b02f59f, 0x9b8e28eb, 0xc0e80269,
164 	0x696f5c16, 0x878f1ff5, 0x52d9c07f, 0x916a06cd, 0xb50b3a20, 0x2776970a,
165 	0xee4eb2a6, 0xe94640ec, 0xb1d65612, 0x9d1ed968, 0x1043f6b7, 0xa3432a76,
166 	0x17eacbb9, 0x3c09e2eb, 0x4f8c2b3,  0x708a1f57, 0xee341814, 0x95d0e4d2,
167 	0xb06f216c, 0x8bd2e72e, 0x8f7c38d7, 0xcfc6a8fc, 0x2a59495,  0xa20d2a69,
168 	0xe29d12d1
169 };
170 
171 /*
172  * fptr and rptr are two pointers into the state info, a front and a rear
173  * pointer.  These two pointers are always rand_sep places aparts, as they
174  * cycle cyclically through the state information.  (Yes, this does mean we
175  * could get away with just one pointer, but the code for random() is more
176  * efficient this way).  The pointers are left positioned as they would be
177  * from the call
178  *
179  *	initstate(1, randtbl, 128);
180  *
181  * (The position of the rear pointer, rptr, is really 0 (as explained above
182  * in the initialization of randtbl) because the state table pointer is set
183  * to point to randtbl[1] (as explained below).
184  */
185 static uint32_t *fptr = &randtbl[SEP_3 + 1];
186 static uint32_t *rptr = &randtbl[1];
187 
188 /*
189  * The following things are the pointer to the state information table, the
190  * type of the current generator, the degree of the current polynomial being
191  * used, and the separation between the two pointers.  Note that for efficiency
192  * of random(), we remember the first location of the state information, not
193  * the zeroeth.  Hence it is valid to access state[-1], which is used to
194  * store the type of the R.N.G.  Also, we remember the last location, since
195  * this is more efficient than indexing every time to find the address of
196  * the last element to see if the front and rear pointers have wrapped.
197  */
198 static uint32_t *state = &randtbl[1];
199 static int rand_type = TYPE_3;
200 static int rand_deg = DEG_3;
201 static int rand_sep = SEP_3;
202 static uint32_t *end_ptr = &randtbl[DEG_3 + 1];
203 
204 static inline uint32_t
205 good_rand(uint32_t ctx)
206 {
207 /*
208  * Compute x = (7^5 * x) mod (2^31 - 1)
209  * wihout overflowing 31 bits:
210  *      (2^31 - 1) = 127773 * (7^5) + 2836
211  * From "Random number generators: good ones are hard to find",
212  * Park and Miller, Communications of the ACM, vol. 31, no. 10,
213  * October 1988, p. 1195.
214  */
215 	int32_t hi, lo, x;
216 
217 	/* Transform to [1, 0x7ffffffe] range. */
218 	x = (ctx % 0x7ffffffe) + 1;
219 	hi = x / 127773;
220 	lo = x % 127773;
221 	x = 16807 * lo - 2836 * hi;
222 	if (x < 0)
223 		x += 0x7fffffff;
224 	/* Transform to [0, 0x7ffffffd] range. */
225 	return (x - 1);
226 }
227 
228 /*
229  * srandom:
230  *
231  * Initialize the random number generator based on the given seed.  If the
232  * type is the trivial no-state-information type, just remember the seed.
233  * Otherwise, initializes state[] based on the given "seed" via a linear
234  * congruential generator.  Then, the pointers are set to known locations
235  * that are exactly rand_sep places apart.  Lastly, it cycles the state
236  * information a given number of times to get rid of any initial dependencies
237  * introduced by the L.C.R.N.G.  Note that the initialization of randtbl[]
238  * for default usage relies on values produced by this routine.
239  */
240 void
241 srandom(unsigned int x)
242 {
243 	int i, lim;
244 
245 	state[0] = (uint32_t)x;
246 	if (rand_type == TYPE_0)
247 		lim = NSHUFF;
248 	else {
249 		for (i = 1; i < rand_deg; i++)
250 			state[i] = good_rand(state[i - 1]);
251 		fptr = &state[rand_sep];
252 		rptr = &state[0];
253 		lim = 10 * rand_deg;
254 	}
255 	for (i = 0; i < lim; i++)
256 		(void)random();
257 }
258 
259 /*
260  * srandomdev:
261  *
262  * Many programs choose the seed value in a totally predictable manner.
263  * This often causes problems.  We seed the generator using pseudo-random
264  * data from the kernel.
265  *
266  * Note that this particular seeding procedure can generate states
267  * which are impossible to reproduce by calling srandom() with any
268  * value, since the succeeding terms in the state buffer are no longer
269  * derived from the LC algorithm applied to a fixed seed.
270  */
271 void
272 srandomdev(void)
273 {
274 	int mib[2];
275 	size_t expected, len;
276 
277 	if (rand_type == TYPE_0)
278 		expected = len = sizeof(state[0]);
279 	else
280 		expected = len = rand_deg * sizeof(state[0]);
281 
282 	mib[0] = CTL_KERN;
283 	mib[1] = KERN_ARND;
284 	if (sysctl(mib, 2, state, &len, NULL, 0) == -1 || len != expected) {
285 		/*
286 		 * The sysctl cannot fail. If it does fail on some FreeBSD
287 		 * derivative or after some future change, just abort so that
288 		 * the problem will be found and fixed. abort is not normally
289 		 * suitable for a library but makes sense here.
290 		 */
291 		abort();
292 	}
293 
294 	if (rand_type != TYPE_0) {
295 		fptr = &state[rand_sep];
296 		rptr = &state[0];
297 	}
298 }
299 
300 /*
301  * initstate:
302  *
303  * Initialize the state information in the given array of n bytes for future
304  * random number generation.  Based on the number of bytes we are given, and
305  * the break values for the different R.N.G.'s, we choose the best (largest)
306  * one we can and set things up for it.  srandom() is then called to
307  * initialize the state information.
308  *
309  * Note that on return from srandom(), we set state[-1] to be the type
310  * multiplexed with the current value of the rear pointer; this is so
311  * successive calls to initstate() won't lose this information and will be
312  * able to restart with setstate().
313  *
314  * Note: the first thing we do is save the current state, if any, just like
315  * setstate() so that it doesn't matter when initstate is called.
316  *
317  * Returns a pointer to the old state.
318  *
319  * Note: The Sparc platform requires that arg_state begin on an int
320  * word boundary; otherwise a bus error will occur. Even so, lint will
321  * complain about mis-alignment, but you should disregard these messages.
322  */
323 char *
324 initstate(unsigned int seed, char *arg_state, size_t n)
325 {
326 	char *ostate = (char *)(&state[-1]);
327 	uint32_t *int_arg_state = (uint32_t *)arg_state;
328 
329 	if (n < BREAK_0)
330 		return (NULL);
331 	if (rand_type == TYPE_0)
332 		state[-1] = rand_type;
333 	else
334 		state[-1] = MAX_TYPES * (rptr - state) + rand_type;
335 	if (n < BREAK_1) {
336 		rand_type = TYPE_0;
337 		rand_deg = DEG_0;
338 		rand_sep = SEP_0;
339 	} else if (n < BREAK_2) {
340 		rand_type = TYPE_1;
341 		rand_deg = DEG_1;
342 		rand_sep = SEP_1;
343 	} else if (n < BREAK_3) {
344 		rand_type = TYPE_2;
345 		rand_deg = DEG_2;
346 		rand_sep = SEP_2;
347 	} else if (n < BREAK_4) {
348 		rand_type = TYPE_3;
349 		rand_deg = DEG_3;
350 		rand_sep = SEP_3;
351 	} else {
352 		rand_type = TYPE_4;
353 		rand_deg = DEG_4;
354 		rand_sep = SEP_4;
355 	}
356 	state = int_arg_state + 1; /* first location */
357 	end_ptr = &state[rand_deg];	/* must set end_ptr before srandom */
358 	srandom(seed);
359 	if (rand_type == TYPE_0)
360 		int_arg_state[0] = rand_type;
361 	else
362 		int_arg_state[0] = MAX_TYPES * (rptr - state) + rand_type;
363 	return (ostate);
364 }
365 
366 /*
367  * setstate:
368  *
369  * Restore the state from the given state array.
370  *
371  * Note: it is important that we also remember the locations of the pointers
372  * in the current state information, and restore the locations of the pointers
373  * from the old state information.  This is done by multiplexing the pointer
374  * location into the zeroeth word of the state information.
375  *
376  * Note that due to the order in which things are done, it is OK to call
377  * setstate() with the same state as the current state.
378  *
379  * Returns a pointer to the old state information.
380  *
381  * Note: The Sparc platform requires that arg_state begin on an int
382  * word boundary; otherwise a bus error will occur. Even so, lint will
383  * complain about mis-alignment, but you should disregard these messages.
384  */
385 char *
386 setstate(char *arg_state)
387 {
388 	uint32_t *new_state = (uint32_t *)arg_state;
389 	uint32_t type = new_state[0] % MAX_TYPES;
390 	uint32_t rear = new_state[0] / MAX_TYPES;
391 	char *ostate = (char *)(&state[-1]);
392 
393 	if (type != TYPE_0 && rear >= degrees[type])
394 		return (NULL);
395 	if (rand_type == TYPE_0)
396 		state[-1] = rand_type;
397 	else
398 		state[-1] = MAX_TYPES * (rptr - state) + rand_type;
399 	rand_type = type;
400 	rand_deg = degrees[type];
401 	rand_sep = seps[type];
402 	state = new_state + 1;
403 	if (rand_type != TYPE_0) {
404 		rptr = &state[rear];
405 		fptr = &state[(rear + rand_sep) % rand_deg];
406 	}
407 	end_ptr = &state[rand_deg];		/* set end_ptr too */
408 	return (ostate);
409 }
410 
411 /*
412  * random:
413  *
414  * If we are using the trivial TYPE_0 R.N.G., just do the old linear
415  * congruential bit.  Otherwise, we do our fancy trinomial stuff, which is
416  * the same in all the other cases due to all the global variables that have
417  * been set up.  The basic operation is to add the number at the rear pointer
418  * into the one at the front pointer.  Then both pointers are advanced to
419  * the next location cyclically in the table.  The value returned is the sum
420  * generated, reduced to 31 bits by throwing away the "least random" low bit.
421  *
422  * Note: the code takes advantage of the fact that both the front and
423  * rear pointers can't wrap on the same call by not testing the rear
424  * pointer if the front one has wrapped.
425  *
426  * Returns a 31-bit random number.
427  */
428 long
429 random(void)
430 {
431 	uint32_t i;
432 	uint32_t *f, *r;
433 
434 	if (rand_type == TYPE_0) {
435 		i = state[0];
436 		state[0] = i = good_rand(i);
437 	} else {
438 		/*
439 		 * Use local variables rather than static variables for speed.
440 		 */
441 		f = fptr; r = rptr;
442 		*f += *r;
443 		i = *f >> 1;	/* chucking least random bit */
444 		if (++f >= end_ptr) {
445 			f = state;
446 			++r;
447 		}
448 		else if (++r >= end_ptr) {
449 			r = state;
450 		}
451 
452 		fptr = f; rptr = r;
453 	}
454 	return ((long)i);
455 }
456