xref: /openbsd-src/sys/dev/rnd.c (revision ae3cb403620ab940fbaabb3055fac045a63d56b7)
1 /*	$OpenBSD: rnd.c,v 1.195 2017/11/26 17:06:46 mikeb Exp $	*/
2 
3 /*
4  * Copyright (c) 2011 Theo de Raadt.
5  * Copyright (c) 2008 Damien Miller.
6  * Copyright (c) 1996, 1997, 2000-2002 Michael Shalayeff.
7  * Copyright (c) 2013 Markus Friedl.
8  * Copyright Theodore Ts'o, 1994, 1995, 1996, 1997, 1998, 1999.
9  * All rights reserved.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, and the entire permission notice in its entirety,
16  *    including the disclaimer of warranties.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. The name of the author may not be used to endorse or promote
21  *    products derived from this software without specific prior
22  *    written permission.
23  *
24  * ALTERNATIVELY, this product may be distributed under the terms of
25  * the GNU Public License, in which case the provisions of the GPL are
26  * required INSTEAD OF the above restrictions.  (This clause is
27  * necessary due to a potential bad interaction between the GPL and
28  * the restrictions contained in a BSD-style copyright.)
29  *
30  * THIS SOFTWARE IS PROVIDED ``AS IS'' AND ANY EXPRESS OR IMPLIED
31  * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
32  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
33  * DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT,
34  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
35  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
36  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
38  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
39  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
40  * OF THE POSSIBILITY OF SUCH DAMAGE.
41  */
42 
43 /*
44  * Computers are very predictable devices.  Hence it is extremely hard
45  * to produce truly random numbers on a computer --- as opposed to
46  * pseudo-random numbers, which can be easily generated by using an
47  * algorithm.  Unfortunately, it is very easy for attackers to guess
48  * the sequence of pseudo-random number generators, and for some
49  * applications this is not acceptable.  Instead, we must try to
50  * gather "environmental noise" from the computer's environment, which
51  * must be hard for outside attackers to observe and use to
52  * generate random numbers.  In a Unix environment, this is best done
53  * from inside the kernel.
54  *
55  * Sources of randomness from the environment include inter-keyboard
56  * timings, inter-interrupt timings from some interrupts, and other
57  * events which are both (a) non-deterministic and (b) hard for an
58  * outside observer to measure.  Randomness from these sources is
59  * added to the "rnd states" queue; this is used as much of the
60  * source material which is mixed on occasion using a CRC-like function
61  * into the "entropy pool".  This is not cryptographically strong, but
62  * it is adequate assuming the randomness is not chosen maliciously,
63  * and it is very fast because the interrupt-time event is only to add
64  * a small random token to the "rnd states" queue.
65  *
66  * When random bytes are desired, they are obtained by pulling from
67  * the entropy pool and running a SHA512 hash. The SHA512 hash avoids
68  * exposing the internal state of the entropy pool.  Even if it is
69  * possible to analyze SHA512 in some clever way, as long as the amount
70  * of data returned from the generator is less than the inherent
71  * entropy in the pool, the output data is totally unpredictable.  For
72  * this reason, the routine decreases its internal estimate of how many
73  * bits of "true randomness" are contained in the entropy pool as it
74  * outputs random numbers.
75  *
76  * If this estimate goes to zero, the SHA512 hash will continue to generate
77  * output since there is no true risk because the SHA512 output is not
78  * exported outside this subsystem.  It is next used as input to seed a
79  * ChaCha20 stream cipher, which is re-seeded from time to time.  This
80  * design provides very high amounts of output data from a potentially
81  * small entropy base, at high enough speeds to encourage use of random
82  * numbers in nearly any situation.  Before OpenBSD 5.5, the RC4 stream
83  * cipher (also known as ARC4) was used instead of ChaCha20.
84  *
85  * The output of this single ChaCha20 engine is then shared amongst many
86  * consumers in the kernel and userland via a few interfaces:
87  * arc4random_buf(), arc4random(), arc4random_uniform(), randomread()
88  * for the set of /dev/random nodes and the system call getentropy(),
89  * which provides seeds for process-context pseudorandom generators.
90  *
91  * Acknowledgements:
92  * =================
93  *
94  * Ideas for constructing this random number generator were derived
95  * from Pretty Good Privacy's random number generator, and from private
96  * discussions with Phil Karn.  Colin Plumb provided a faster random
97  * number generator, which speeds up the mixing function of the entropy
98  * pool, taken from PGPfone.  Dale Worley has also contributed many
99  * useful ideas and suggestions to improve this driver.
100  *
101  * Any flaws in the design are solely my responsibility, and should
102  * not be attributed to the Phil, Colin, or any of the authors of PGP.
103  *
104  * Further background information on this topic may be obtained from
105  * RFC 1750, "Randomness Recommendations for Security", by Donald
106  * Eastlake, Steve Crocker, and Jeff Schiller.
107  *
108  * Using a RC4 stream cipher as 2nd stage after the MD5 (now SHA512) output
109  * is the result of work by David Mazieres.
110  */
111 
112 #include <sys/param.h>
113 #include <sys/systm.h>
114 #include <sys/disk.h>
115 #include <sys/event.h>
116 #include <sys/limits.h>
117 #include <sys/time.h>
118 #include <sys/ioctl.h>
119 #include <sys/malloc.h>
120 #include <sys/fcntl.h>
121 #include <sys/timeout.h>
122 #include <sys/mutex.h>
123 #include <sys/task.h>
124 #include <sys/msgbuf.h>
125 #include <sys/mount.h>
126 #include <sys/syscallargs.h>
127 
128 #include <crypto/sha2.h>
129 
130 #define KEYSTREAM_ONLY
131 #include <crypto/chacha_private.h>
132 
133 #include <dev/rndvar.h>
134 
135 #include <uvm/uvm_param.h>
136 #include <uvm/uvm_extern.h>
137 
138 /*
139  * For the purposes of better mixing, we use the CRC-32 polynomial as
140  * well to make a twisted Generalized Feedback Shift Register
141  *
142  * (See M. Matsumoto & Y. Kurita, 1992.  Twisted GFSR generators.  ACM
143  * Transactions on Modeling and Computer Simulation 2(3):179-194.
144  * Also see M. Matsumoto & Y. Kurita, 1994.  Twisted GFSR generators
145  * II.  ACM Transactions on Modeling and Computer Simulation 4:254-266)
146  *
147  * Thanks to Colin Plumb for suggesting this.
148  *
149  * We have not analyzed the resultant polynomial to prove it primitive;
150  * in fact it almost certainly isn't.  Nonetheless, the irreducible factors
151  * of a random large-degree polynomial over GF(2) are more than large enough
152  * that periodicity is not a concern.
153  *
154  * The input hash is much less sensitive than the output hash.  All
155  * we want from it is to be a good non-cryptographic hash -
156  * i.e. to not produce collisions when fed "random" data of the sort
157  * we expect to see.  As long as the pool state differs for different
158  * inputs, we have preserved the input entropy and done a good job.
159  * The fact that an intelligent attacker can construct inputs that
160  * will produce controlled alterations to the pool's state is not
161  * important because we don't consider such inputs to contribute any
162  * randomness.  The only property we need with respect to them is that
163  * the attacker can't increase his/her knowledge of the pool's state.
164  * Since all additions are reversible (knowing the final state and the
165  * input, you can reconstruct the initial state), if an attacker has
166  * any uncertainty about the initial state, he/she can only shuffle
167  * that uncertainty about, but never cause any collisions (which would
168  * decrease the uncertainty).
169  *
170  * The chosen system lets the state of the pool be (essentially) the input
171  * modulo the generator polynomial.  Now, for random primitive polynomials,
172  * this is a universal class of hash functions, meaning that the chance
173  * of a collision is limited by the attacker's knowledge of the generator
174  * polynomial, so if it is chosen at random, an attacker can never force
175  * a collision.  Here, we use a fixed polynomial, but we *can* assume that
176  * ###--> it is unknown to the processes generating the input entropy. <-###
177  * Because of this important property, this is a good, collision-resistant
178  * hash; hash collisions will occur no more often than chance.
179  */
180 
181 /*
182  * Stirring polynomials over GF(2) for various pool sizes. Used in
183  * add_entropy_words() below.
184  *
185  * The polynomial terms are chosen to be evenly spaced (minimum RMS
186  * distance from evenly spaced; except for the last tap, which is 1 to
187  * get the twisting happening as fast as possible.
188  *
189  * The resultant polynomial is:
190  *   2^POOLWORDS + 2^POOL_TAP1 + 2^POOL_TAP2 + 2^POOL_TAP3 + 2^POOL_TAP4 + 1
191  */
192 #define POOLWORDS	2048
193 #define POOLBYTES	(POOLWORDS*4)
194 #define POOLMASK	(POOLWORDS - 1)
195 #define	POOL_TAP1	1638
196 #define	POOL_TAP2	1231
197 #define	POOL_TAP3	819
198 #define	POOL_TAP4	411
199 
200 /*
201  * Raw entropy collection from device drivers; at interrupt context or not.
202  * add_*_randomness() provide data which is put into the entropy queue.
203  */
204 
205 #define QEVLEN	128		 /* must be a power of 2 */
206 #define QEVSLOW (QEVLEN * 3 / 4) /* yet another 0.75 for 60-minutes hour /-; */
207 
208 #define KEYSZ	32
209 #define IVSZ	8
210 #define BLOCKSZ	64
211 #define RSBUFSZ	(16*BLOCKSZ)
212 #define EBUFSIZE KEYSZ + IVSZ
213 
214 struct rand_event {
215 	u_int re_time;
216 	u_int re_val;
217 } rnd_event_space[QEVLEN];
218 
219 u_int rnd_event_cons;
220 u_int rnd_event_prod;
221 
222 struct mutex rnd_enqlck = MUTEX_INITIALIZER(IPL_HIGH);
223 struct mutex rnd_deqlck = MUTEX_INITIALIZER(IPL_HIGH);
224 
225 struct timeout rnd_timeout;
226 
227 static u_int32_t entropy_pool[POOLWORDS];
228 u_int32_t entropy_pool0[POOLWORDS] __attribute__((section(".openbsd.randomdata")));
229 u_int	entropy_add_ptr;
230 u_char	entropy_input_rotate;
231 
232 void	dequeue_randomness(void *);
233 void	add_entropy_words(const u_int32_t *, u_int);
234 void	extract_entropy(u_int8_t *)
235     __attribute__((__bounded__(__minbytes__,1,EBUFSIZE)));
236 
237 int	filt_randomread(struct knote *, long);
238 void	filt_randomdetach(struct knote *);
239 int	filt_randomwrite(struct knote *, long);
240 
241 static void _rs_seed(u_char *, size_t);
242 static void _rs_clearseed(const void *p, size_t s);
243 
244 struct filterops randomread_filtops =
245 	{ 1, NULL, filt_randomdetach, filt_randomread };
246 struct filterops randomwrite_filtops =
247 	{ 1, NULL, filt_randomdetach, filt_randomwrite };
248 
249 static inline struct rand_event *
250 rnd_get(void)
251 {
252 	u_int idx;
253 
254 	/* nothing to do if queue is empty */
255 	if (rnd_event_prod == rnd_event_cons)
256 		return NULL;
257 
258 	if (rnd_event_prod - rnd_event_cons > QEVLEN)
259 		rnd_event_cons = rnd_event_prod - QEVLEN;
260 	idx = rnd_event_cons++;
261 	return &rnd_event_space[idx & (QEVLEN - 1)];
262 }
263 
264 static inline struct rand_event *
265 rnd_put(void)
266 {
267 	u_int idx = rnd_event_prod++;
268 
269 	/* allow wrapping. caller will mix it in. */
270 	return &rnd_event_space[idx & (QEVLEN - 1)];
271 }
272 
273 static inline u_int
274 rnd_qlen(void)
275 {
276 	return rnd_event_prod - rnd_event_cons;
277 }
278 
279 /*
280  * This function adds entropy to the entropy pool by using timing delays.
281  *
282  * The number "val" is also added to the pool - it should somehow describe
283  * the type of event which just happened.  Currently the values of 0-255
284  * are for keyboard scan codes, 256 and upwards - for interrupts.
285  */
286 void
287 enqueue_randomness(u_int state, u_int val)
288 {
289 	struct rand_event *rep;
290 	struct timespec	ts;
291 	u_int qlen;
292 
293 #ifdef DIAGNOSTIC
294 	if (state >= RND_SRC_NUM)
295 		return;
296 #endif
297 
298 	if (timeout_initialized(&rnd_timeout))
299 		nanotime(&ts);
300 
301 	val += state << 13;
302 
303 	mtx_enter(&rnd_enqlck);
304 	rep = rnd_put();
305 	rep->re_time += ts.tv_nsec ^ (ts.tv_sec << 20);
306 	rep->re_val += val;
307 	qlen = rnd_qlen();
308 	mtx_leave(&rnd_enqlck);
309 
310 	if (qlen > QEVSLOW/2 && timeout_initialized(&rnd_timeout) &&
311 	    !timeout_pending(&rnd_timeout))
312 		timeout_add(&rnd_timeout, 1);
313 }
314 
315 /*
316  * This function adds a byte into the entropy pool.  It does not
317  * update the entropy estimate.  The caller must do this if appropriate.
318  *
319  * The pool is stirred with a polynomial of degree POOLWORDS over GF(2);
320  * see POOL_TAP[1-4] above
321  *
322  * Rotate the input word by a changing number of bits, to help assure
323  * that all bits in the entropy get toggled.  Otherwise, if the pool
324  * is consistently fed small numbers (such as keyboard scan codes)
325  * then the upper bits of the entropy pool will frequently remain
326  * untouched.
327  */
328 void
329 add_entropy_words(const u_int32_t *buf, u_int n)
330 {
331 	/* derived from IEEE 802.3 CRC-32 */
332 	static const u_int32_t twist_table[8] = {
333 		0x00000000, 0x3b6e20c8, 0x76dc4190, 0x4db26158,
334 		0xedb88320, 0xd6d6a3e8, 0x9b64c2b0, 0xa00ae278
335 	};
336 
337 	for (; n--; buf++) {
338 		u_int32_t w = (*buf << entropy_input_rotate) |
339 		    (*buf >> ((32 - entropy_input_rotate) & 31));
340 		u_int i = entropy_add_ptr =
341 		    (entropy_add_ptr - 1) & POOLMASK;
342 		/*
343 		 * Normally, we add 7 bits of rotation to the pool.
344 		 * At the beginning of the pool, add an extra 7 bits
345 		 * rotation, so that successive passes spread the
346 		 * input bits across the pool evenly.
347 		 */
348 		entropy_input_rotate =
349 		    (entropy_input_rotate + (i ? 7 : 14)) & 31;
350 
351 		/* XOR pool contents corresponding to polynomial terms */
352 		w ^= entropy_pool[(i + POOL_TAP1) & POOLMASK] ^
353 		     entropy_pool[(i + POOL_TAP2) & POOLMASK] ^
354 		     entropy_pool[(i + POOL_TAP3) & POOLMASK] ^
355 		     entropy_pool[(i + POOL_TAP4) & POOLMASK] ^
356 		     entropy_pool[(i + 1) & POOLMASK] ^
357 		     entropy_pool[i]; /* + 2^POOLWORDS */
358 
359 		entropy_pool[i] = (w >> 3) ^ twist_table[w & 7];
360 	}
361 }
362 
363 /*
364  * Pulls entropy out of the queue and merges it into the pool
365  * with the CRC.
366  */
367 /* ARGSUSED */
368 void
369 dequeue_randomness(void *v)
370 {
371 	struct rand_event *rep;
372 	u_int32_t buf[2];
373 
374 	if (timeout_initialized(&rnd_timeout))
375 		timeout_del(&rnd_timeout);
376 
377 	mtx_enter(&rnd_deqlck);
378 	while ((rep = rnd_get())) {
379 		buf[0] = rep->re_time;
380 		buf[1] = rep->re_val;
381 		mtx_leave(&rnd_deqlck);
382 		add_entropy_words(buf, 2);
383 		mtx_enter(&rnd_deqlck);
384 	}
385 	mtx_leave(&rnd_deqlck);
386 }
387 
388 /*
389  * Grabs a chunk from the entropy_pool[] and slams it through SHA512 when
390  * requested.
391  */
392 void
393 extract_entropy(u_int8_t *buf)
394 {
395 	static u_int32_t extract_pool[POOLWORDS];
396 	u_char digest[SHA512_DIGEST_LENGTH];
397 	SHA2_CTX shactx;
398 
399 #if SHA512_DIGEST_LENGTH < EBUFSIZE
400 #error "need more bigger hash output"
401 #endif
402 
403 	/*
404 	 * INTENTIONALLY not protected by any lock.  Races during
405 	 * memcpy() result in acceptable input data; races during
406 	 * SHA512Update() would create nasty data dependencies.  We
407 	 * do not rely on this as a benefit, but if it happens, cool.
408 	 */
409 	memcpy(extract_pool, entropy_pool, sizeof(extract_pool));
410 
411 	/* Hash the pool to get the output */
412 	SHA512Init(&shactx);
413 	SHA512Update(&shactx, (u_int8_t *)extract_pool, sizeof(extract_pool));
414 	SHA512Final(digest, &shactx);
415 
416 	/* Copy data to destination buffer */
417 	memcpy(buf, digest, EBUFSIZE);
418 
419 	/* Modify pool so next hash will produce different results */
420 	add_timer_randomness(EBUFSIZE);
421 	dequeue_randomness(NULL);
422 
423 	/* Wipe data from memory */
424 	explicit_bzero(extract_pool, sizeof(extract_pool));
425 	explicit_bzero(digest, sizeof(digest));
426 }
427 
428 /* random keystream by ChaCha */
429 
430 void arc4_reinit(void *v);		/* timeout to start reinit */
431 void arc4_init(void *);			/* actually do the reinit */
432 
433 struct mutex rndlock = MUTEX_INITIALIZER(IPL_HIGH);
434 struct timeout arc4_timeout;
435 struct task arc4_task = TASK_INITIALIZER(arc4_init, NULL);
436 
437 static chacha_ctx rs;		/* chacha context for random keystream */
438 /* keystream blocks (also chacha seed from boot) */
439 static u_char rs_buf[RSBUFSZ];
440 u_char rs_buf0[RSBUFSZ] __attribute__((section(".openbsd.randomdata")));
441 static size_t rs_have;		/* valid bytes at end of rs_buf */
442 static size_t rs_count;		/* bytes till reseed */
443 
444 void
445 suspend_randomness(void)
446 {
447 	struct timespec ts;
448 
449 	getnanotime(&ts);
450 	add_true_randomness(ts.tv_sec);
451 	add_true_randomness(ts.tv_nsec);
452 
453 	dequeue_randomness(NULL);
454 	rs_count = 0;
455 	arc4random_buf(entropy_pool, sizeof(entropy_pool));
456 }
457 
458 void
459 resume_randomness(char *buf, size_t buflen)
460 {
461 	struct timespec ts;
462 
463 	if (buf && buflen)
464 		_rs_seed(buf, buflen);
465 	getnanotime(&ts);
466 	add_true_randomness(ts.tv_sec);
467 	add_true_randomness(ts.tv_nsec);
468 
469 	dequeue_randomness(NULL);
470 	rs_count = 0;
471 }
472 
473 static inline void _rs_rekey(u_char *dat, size_t datlen);
474 
475 static inline void
476 _rs_init(u_char *buf, size_t n)
477 {
478 	KASSERT(n >= KEYSZ + IVSZ);
479 	chacha_keysetup(&rs, buf, KEYSZ * 8);
480 	chacha_ivsetup(&rs, buf + KEYSZ, NULL);
481 }
482 
483 static void
484 _rs_seed(u_char *buf, size_t n)
485 {
486 	_rs_rekey(buf, n);
487 
488 	/* invalidate rs_buf */
489 	rs_have = 0;
490 	memset(rs_buf, 0, RSBUFSZ);
491 
492 	rs_count = 1600000;
493 }
494 
495 static void
496 _rs_stir(int do_lock)
497 {
498 	struct timespec ts;
499 	u_int8_t buf[EBUFSIZE], *p;
500 	int i;
501 
502 	/*
503 	 * Use SHA512 PRNG data and a system timespec; early in the boot
504 	 * process this is the best we can do -- some architectures do
505 	 * not collect entropy very well during this time, but may have
506 	 * clock information which is better than nothing.
507 	 */
508 	extract_entropy(buf);
509 
510 	nanotime(&ts);
511 	for (p = (u_int8_t *)&ts, i = 0; i < sizeof(ts); i++)
512 		buf[i] ^= p[i];
513 
514 	if (do_lock)
515 		mtx_enter(&rndlock);
516 	_rs_seed(buf, sizeof(buf));
517 	if (do_lock)
518 		mtx_leave(&rndlock);
519 
520 	explicit_bzero(buf, sizeof(buf));
521 }
522 
523 static inline void
524 _rs_stir_if_needed(size_t len)
525 {
526 	static int rs_initialized;
527 
528 	if (!rs_initialized) {
529 		memcpy(entropy_pool, entropy_pool0, sizeof entropy_pool);
530 		memcpy(rs_buf, rs_buf0, sizeof rs_buf);
531 		/* seeds cannot be cleaned yet, random_start() will do so */
532 		_rs_init(rs_buf, KEYSZ + IVSZ);
533 		rs_count = 1024 * 1024 * 1024;	/* until main() runs */
534 		rs_initialized = 1;
535 	} else if (rs_count <= len)
536 		_rs_stir(0);
537 	else
538 		rs_count -= len;
539 }
540 
541 static void
542 _rs_clearseed(const void *p, size_t s)
543 {
544 	struct kmem_dyn_mode kd_avoidalias;
545 	vaddr_t va = trunc_page((vaddr_t)p);
546 	vsize_t off = (vaddr_t)p - va;
547 	vsize_t len;
548 	vaddr_t rwva;
549 	paddr_t pa;
550 
551 	while (s > 0) {
552 		pmap_extract(pmap_kernel(), va, &pa);
553 
554 		memset(&kd_avoidalias, 0, sizeof kd_avoidalias);
555 		kd_avoidalias.kd_prefer = pa;
556 		kd_avoidalias.kd_waitok = 1;
557 		rwva = (vaddr_t)km_alloc(PAGE_SIZE, &kv_any, &kp_none,
558 		    &kd_avoidalias);
559 		if (!rwva)
560 			panic("_rs_clearseed");
561 
562 		pmap_kenter_pa(rwva, pa, PROT_READ | PROT_WRITE);
563 		pmap_update(pmap_kernel());
564 
565 		len = MIN(s, PAGE_SIZE - off);
566 		explicit_bzero((void *)(rwva + off), len);
567 
568 		pmap_kremove(rwva, PAGE_SIZE);
569 		km_free((void *)rwva, PAGE_SIZE, &kv_any, &kp_none);
570 
571 		va += PAGE_SIZE;
572 		s -= len;
573 		off = 0;
574 	}
575 }
576 
577 static inline void
578 _rs_rekey(u_char *dat, size_t datlen)
579 {
580 #ifndef KEYSTREAM_ONLY
581 	memset(rs_buf, 0, RSBUFSZ);
582 #endif
583 	/* fill rs_buf with the keystream */
584 	chacha_encrypt_bytes(&rs, rs_buf, rs_buf, RSBUFSZ);
585 	/* mix in optional user provided data */
586 	if (dat) {
587 		size_t i, m;
588 
589 		m = MIN(datlen, KEYSZ + IVSZ);
590 		for (i = 0; i < m; i++)
591 			rs_buf[i] ^= dat[i];
592 	}
593 	/* immediately reinit for backtracking resistance */
594 	_rs_init(rs_buf, KEYSZ + IVSZ);
595 	memset(rs_buf, 0, KEYSZ + IVSZ);
596 	rs_have = RSBUFSZ - KEYSZ - IVSZ;
597 }
598 
599 static inline void
600 _rs_random_buf(void *_buf, size_t n)
601 {
602 	u_char *buf = (u_char *)_buf;
603 	size_t m;
604 
605 	_rs_stir_if_needed(n);
606 	while (n > 0) {
607 		if (rs_have > 0) {
608 			m = MIN(n, rs_have);
609 			memcpy(buf, rs_buf + RSBUFSZ - rs_have, m);
610 			memset(rs_buf + RSBUFSZ - rs_have, 0, m);
611 			buf += m;
612 			n -= m;
613 			rs_have -= m;
614 		}
615 		if (rs_have == 0)
616 			_rs_rekey(NULL, 0);
617 	}
618 }
619 
620 static inline void
621 _rs_random_u32(u_int32_t *val)
622 {
623 	_rs_stir_if_needed(sizeof(*val));
624 	if (rs_have < sizeof(*val))
625 		_rs_rekey(NULL, 0);
626 	memcpy(val, rs_buf + RSBUFSZ - rs_have, sizeof(*val));
627 	memset(rs_buf + RSBUFSZ - rs_have, 0, sizeof(*val));
628 	rs_have -= sizeof(*val);
629 }
630 
631 /* Return one word of randomness from a ChaCha20 generator */
632 u_int32_t
633 arc4random(void)
634 {
635 	u_int32_t ret;
636 
637 	mtx_enter(&rndlock);
638 	_rs_random_u32(&ret);
639 	mtx_leave(&rndlock);
640 	return ret;
641 }
642 
643 /*
644  * Fill a buffer of arbitrary length with ChaCha20-derived randomness.
645  */
646 void
647 arc4random_buf(void *buf, size_t n)
648 {
649 	mtx_enter(&rndlock);
650 	_rs_random_buf(buf, n);
651 	mtx_leave(&rndlock);
652 }
653 
654 /*
655  * Calculate a uniformly distributed random number less than upper_bound
656  * avoiding "modulo bias".
657  *
658  * Uniformity is achieved by generating new random numbers until the one
659  * returned is outside the range [0, 2**32 % upper_bound).  This
660  * guarantees the selected random number will be inside
661  * [2**32 % upper_bound, 2**32) which maps back to [0, upper_bound)
662  * after reduction modulo upper_bound.
663  */
664 u_int32_t
665 arc4random_uniform(u_int32_t upper_bound)
666 {
667 	u_int32_t r, min;
668 
669 	if (upper_bound < 2)
670 		return 0;
671 
672 	/* 2**32 % x == (2**32 - x) % x */
673 	min = -upper_bound % upper_bound;
674 
675 	/*
676 	 * This could theoretically loop forever but each retry has
677 	 * p > 0.5 (worst case, usually far better) of selecting a
678 	 * number inside the range we need, so it should rarely need
679 	 * to re-roll.
680 	 */
681 	for (;;) {
682 		r = arc4random();
683 		if (r >= min)
684 			break;
685 	}
686 
687 	return r % upper_bound;
688 }
689 
690 /* ARGSUSED */
691 void
692 arc4_init(void *null)
693 {
694 	_rs_stir(1);
695 }
696 
697 /*
698  * Called by timeout to mark arc4 for stirring,
699  */
700 void
701 arc4_reinit(void *v)
702 {
703 	task_add(systq, &arc4_task);
704 	/* 10 minutes, per dm@'s suggestion */
705 	timeout_add_sec(&arc4_timeout, 10 * 60);
706 }
707 
708 /*
709  * Start periodic services inside the random subsystem, which pull
710  * entropy forward, hash it, and re-seed the random stream as needed.
711  */
712 void
713 random_start(void)
714 {
715 #if !defined(NO_PROPOLICE)
716 	extern long __guard_local;
717 
718 	if (__guard_local == 0)
719 		printf("warning: no entropy supplied by boot loader\n");
720 #endif
721 
722 	_rs_clearseed(entropy_pool0, sizeof entropy_pool0);
723 	_rs_clearseed(rs_buf0, sizeof rs_buf0);
724 
725 	/* Message buffer may contain data from previous boot */
726 	if (msgbufp->msg_magic == MSG_MAGIC)
727 		add_entropy_words((u_int32_t *)msgbufp->msg_bufc,
728 		    msgbufp->msg_bufs / sizeof(u_int32_t));
729 
730 	dequeue_randomness(NULL);
731 	arc4_init(NULL);
732 	timeout_set(&arc4_timeout, arc4_reinit, NULL);
733 	arc4_reinit(NULL);
734 	timeout_set(&rnd_timeout, dequeue_randomness, NULL);
735 }
736 
737 int
738 randomopen(dev_t dev, int flag, int mode, struct proc *p)
739 {
740 	return 0;
741 }
742 
743 int
744 randomclose(dev_t dev, int flag, int mode, struct proc *p)
745 {
746 	return 0;
747 }
748 
749 /*
750  * Maximum number of bytes to serve directly from the main ChaCha
751  * pool. Larger requests are served from a discrete ChaCha instance keyed
752  * from the main pool.
753  */
754 #define ARC4_MAIN_MAX_BYTES	2048
755 
756 int
757 randomread(dev_t dev, struct uio *uio, int ioflag)
758 {
759 	u_char		lbuf[KEYSZ+IVSZ];
760 	chacha_ctx	lctx;
761 	size_t		total = uio->uio_resid;
762 	u_char		*buf;
763 	int		myctx = 0, ret = 0;
764 
765 	if (uio->uio_resid == 0)
766 		return 0;
767 
768 	buf = malloc(POOLBYTES, M_TEMP, M_WAITOK);
769 	if (total > ARC4_MAIN_MAX_BYTES) {
770 		arc4random_buf(lbuf, sizeof(lbuf));
771 		chacha_keysetup(&lctx, lbuf, KEYSZ * 8);
772 		chacha_ivsetup(&lctx, lbuf + KEYSZ, NULL);
773 		explicit_bzero(lbuf, sizeof(lbuf));
774 		myctx = 1;
775 	}
776 
777 	while (ret == 0 && uio->uio_resid > 0) {
778 		size_t	n = ulmin(POOLBYTES, uio->uio_resid);
779 
780 		if (myctx) {
781 #ifndef KEYSTREAM_ONLY
782 			memset(buf, 0, n);
783 #endif
784 			chacha_encrypt_bytes(&lctx, buf, buf, n);
785 		} else
786 			arc4random_buf(buf, n);
787 		ret = uiomove(buf, n, uio);
788 		if (ret == 0 && uio->uio_resid > 0)
789 			yield();
790 	}
791 	if (myctx)
792 		explicit_bzero(&lctx, sizeof(lctx));
793 	explicit_bzero(buf, POOLBYTES);
794 	free(buf, M_TEMP, POOLBYTES);
795 	return ret;
796 }
797 
798 int
799 randomwrite(dev_t dev, struct uio *uio, int flags)
800 {
801 	int		ret = 0, newdata = 0;
802 	u_int32_t	*buf;
803 
804 	if (uio->uio_resid == 0)
805 		return 0;
806 
807 	buf = malloc(POOLBYTES, M_TEMP, M_WAITOK);
808 
809 	while (ret == 0 && uio->uio_resid > 0) {
810 		size_t	n = ulmin(POOLBYTES, uio->uio_resid);
811 
812 		ret = uiomove(buf, n, uio);
813 		if (ret != 0)
814 			break;
815 		while (n % sizeof(u_int32_t))
816 			((u_int8_t *)buf)[n++] = 0;
817 		add_entropy_words(buf, n / 4);
818 		if (uio->uio_resid > 0)
819 			yield();
820 		newdata = 1;
821 	}
822 
823 	if (newdata)
824 		arc4_init(NULL);
825 
826 	explicit_bzero(buf, POOLBYTES);
827 	free(buf, M_TEMP, POOLBYTES);
828 	return ret;
829 }
830 
831 int
832 randomkqfilter(dev_t dev, struct knote *kn)
833 {
834 	switch (kn->kn_filter) {
835 	case EVFILT_READ:
836 		kn->kn_fop = &randomread_filtops;
837 		break;
838 	case EVFILT_WRITE:
839 		kn->kn_fop = &randomwrite_filtops;
840 		break;
841 	default:
842 		return (EINVAL);
843 	}
844 
845 	return (0);
846 }
847 
848 void
849 filt_randomdetach(struct knote *kn)
850 {
851 }
852 
853 int
854 filt_randomread(struct knote *kn, long hint)
855 {
856 	kn->kn_data = ARC4_MAIN_MAX_BYTES;
857 	return (1);
858 }
859 
860 int
861 filt_randomwrite(struct knote *kn, long hint)
862 {
863 	kn->kn_data = POOLBYTES;
864 	return (1);
865 }
866 
867 int
868 randomioctl(dev_t dev, u_long cmd, caddr_t data, int flag, struct proc *p)
869 {
870 	switch (cmd) {
871 	case FIOASYNC:
872 		/* No async flag in softc so this is a no-op. */
873 		break;
874 	case FIONBIO:
875 		/* Handled in the upper FS layer. */
876 		break;
877 	default:
878 		return ENOTTY;
879 	}
880 	return 0;
881 }
882 
883 int
884 sys_getentropy(struct proc *p, void *v, register_t *retval)
885 {
886 	struct sys_getentropy_args /* {
887 		syscallarg(void *) buf;
888 		syscallarg(size_t) nbyte;
889 	} */ *uap = v;
890 	char buf[256];
891 	int error;
892 
893 	if (SCARG(uap, nbyte) > sizeof(buf))
894 		return (EIO);
895 	arc4random_buf(buf, SCARG(uap, nbyte));
896 	if ((error = copyout(buf, SCARG(uap, buf), SCARG(uap, nbyte))) != 0)
897 		return (error);
898 	explicit_bzero(buf, sizeof(buf));
899 	retval[0] = 0;
900 	return (0);
901 }
902