1 /* 2 * Copyright (c) 2004, 2005, 2006 Robin J Carey. All rights reserved. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions, and the following disclaimer, 9 * without modification, immediately at the beginning of the file. 10 * 2. The name of the author may not be used to endorse or promote products 11 * derived from this software without specific prior written permission. 12 * 13 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR 17 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23 * SUCH DAMAGE. 24 * 25 * $DragonFly: src/sys/kern/kern_nrandom.c,v 1.7 2008/08/01 04:42:30 dillon Exp $ 26 */ 27 /* --- NOTES --- 28 * 29 * Note: The word "entropy" is often incorrectly used to describe 30 * random data. The word "entropy" originates from the science of 31 * Physics. The correct descriptive definition would be something 32 * along the lines of "seed", "unpredictable numbers" or 33 * "unpredictable data". 34 * 35 * Note: Some /dev/[u]random implementations save "seed" between 36 * boots which represents a security hazard since an adversary 37 * could acquire this data (since it is stored in a file). If 38 * the unpredictable data used in the above routines is only 39 * generated during Kernel operation, then an adversary can only 40 * acquire that data through a Kernel security compromise and/or 41 * a cryptographic algorithm failure/cryptanalysis. 42 * 43 * Note: On FreeBSD-4.11, interrupts have to be manually enabled 44 * using the rndcontrol(8) command. 45 * 46 * --- DESIGN (FreeBSD-4.11 based) --- 47 * 48 * The rnddev module automatically initializes itself the first time 49 * it is used (client calls any public rnddev_*() interface routine). 50 * Both CSPRNGs are initially seeded from the precise nano[up]time() routines. 51 * Tests show this method produces good enough results, suitable for intended 52 * use. It is necessary for both CSPRNGs to be completely seeded, initially. 53 * 54 * After initialization and during Kernel operation the only suitable 55 * unpredictable data available is: 56 * 57 * (1) Keyboard scan-codes. 58 * (2) Nanouptime acquired by a Keyboard/Read-Event. 59 * (3) Suitable interrupt source; hard-disk/ATA-device. 60 * 61 * (X) Mouse-event (xyz-data unsuitable); NOT IMPLEMENTED. 62 * 63 * This data is added to both CSPRNGs in real-time as it happens/ 64 * becomes-available. Additionally, unpredictable (?) data may be 65 * acquired from a true-random number generator if such a device is 66 * available to the system (not advisable !). 67 * Nanouptime() acquired by a Read-Event is a very important aspect of 68 * this design, since it ensures that unpredictable data is added to 69 * the CSPRNGs even if there are no other sources. 70 * The nanouptime() Kernel routine is used since time relative to 71 * boot is less adversary-known than time itself. 72 * 73 * This design has been thoroughly tested with debug logging 74 * and the output from both /dev/random and /dev/urandom has 75 * been tested with the DIEHARD test-suite; both pass. 76 * 77 * MODIFICATIONS MADE TO ORIGINAL "kern_random.c": 78 * 79 * 6th July 2005: 80 * 81 * o Changed ReadSeed() function to schedule future read-seed-events 82 * by at least one second. Previous implementation used a randomised 83 * scheduling { 0, 1, 2, 3 seconds }. 84 * o Changed SEED_NANOUP() function to use a "previous" accumulator 85 * algorithm similar to ReadSeed(). This ensures that there is no 86 * way that an adversary can tell what number is being added to the 87 * CSPRNGs, since the number added to the CSPRNGs at Event-Time is 88 * the sum of nanouptime()@Event and an unknown/secret number. 89 * o Changed rnddev_add_interrupt() function to schedule future 90 * interrupt-events by at least one second. Previous implementation 91 * had no scheduling algorithm which allowed an "interrupt storm" 92 * to occur resulting in skewed data entering into the CSPRNGs. 93 * 94 * 95 * 9th July 2005: 96 * 97 * o Some small cleanups and change all internal functions to be 98 * static/private. 99 * o Removed ReadSeed() since its functionality is already performed 100 * by another function { rnddev_add_interrupt_OR_read() } and remove 101 * the silly rndByte accumulator/feedback-thing (since multipying by 102 * rndByte could yield a value of 0). 103 * o Made IBAA/L14 public interface become static/private; 104 * Local to this file (not changed to that in the original C modules). 105 * 106 * 16th July 2005: 107 * 108 * o SEED_NANOUP() -> NANOUP_EVENT() function rename. 109 * o Make NANOUP_EVENT() handle the time-buffering directly so that all 110 * time-stamp-events use this single time-buffer (including keyboard). 111 * This removes dependancy on "time_second" Kernel variable. 112 * o Removed second-time-buffer code in rnddev_add_interrupt_OR_read (void). 113 * o Rewrote the time-buffering algorithm in NANOUP_EVENT() to use a 114 * randomised time-delay range. 115 * 116 * 12th Dec 2005: 117 * 118 * o Updated to (hopefully final) L15 algorithm. 119 * 120 * 12th June 2006: 121 * 122 * o Added missing (u_char *) cast in RnddevRead() function. 123 * o Changed copyright to 3-clause BSD license and cleaned up the layout 124 * of this file. 125 */ 126 127 #include <sys/types.h> 128 #include <sys/kernel.h> 129 #include <sys/systm.h> 130 #include <sys/poll.h> 131 #include <sys/event.h> 132 #include <sys/random.h> 133 #include <sys/systimer.h> 134 #include <sys/time.h> 135 #include <sys/proc.h> 136 #include <sys/lock.h> 137 #include <sys/sysctl.h> 138 #include <sys/spinlock.h> 139 #include <machine/clock.h> 140 141 #include <sys/thread2.h> 142 #include <sys/spinlock2.h> 143 144 /* 145 * Portability note: The u_char/unsigned char type is used where 146 * uint8_t from <stdint.h> or u_int8_t from <sys/types.h> should really 147 * be being used. On FreeBSD, it is safe to make the assumption that these 148 * different types are equivalent (on all architectures). 149 * The FreeBSD <sys/crypto/rc4> module also makes this assumption. 150 */ 151 152 /*------------------------------ IBAA ----------------------------------*/ 153 154 /*-------------------------- IBAA CSPRNG -------------------------------*/ 155 156 /* 157 * NOTE: The original source code from which this source code (IBAA) 158 * was taken has no copyright/license. The algorithm has no patent 159 * and is freely/publicly available from: 160 * 161 * http://www.burtleburtle.net/bob/rand/isaac.html 162 */ 163 164 /* 165 * ^ means XOR, & means bitwise AND, a<<b means shift a by b. 166 * barrel(a) shifts a 19 bits to the left, and bits wrap around 167 * ind(x) is (x AND 255), or (x mod 256) 168 */ 169 typedef u_int32_t u4; /* unsigned four bytes, 32 bits */ 170 171 #define ALPHA (8) 172 #define SIZE (1 << ALPHA) 173 #define MASK (SIZE - 1) 174 #define ind(x) ((x) & (SIZE - 1)) 175 #define barrel(a) (((a) << 20) ^ ((a) >> 12)) /* beta=32,shift=20 */ 176 177 static void IBAA 178 ( 179 u4 *m, /* Memory: array of SIZE ALPHA-bit terms */ 180 u4 *r, /* Results: the sequence, same size as m */ 181 u4 *aa, /* Accumulator: a single value */ 182 u4 *bb, /* the previous result */ 183 u4 *counter /* counter */ 184 ) 185 { 186 u4 a, b, x, y, i; 187 188 a = *aa; 189 b = *bb + *counter; 190 ++*counter; 191 for (i = 0; i < SIZE; ++i) { 192 x = m[i]; 193 a = barrel(a) + m[ind(i + (SIZE / 2))]; /* set a */ 194 m[i] = y = m[ind(x)] + a + b; /* set m */ 195 r[i] = b = m[ind(y >> ALPHA)] + x; /* set r */ 196 } 197 *bb = b; *aa = a; 198 } 199 200 /*-------------------------- IBAA CSPRNG -------------------------------*/ 201 202 203 static u4 IBAA_memory[SIZE]; 204 static u4 IBAA_results[SIZE]; 205 static u4 IBAA_aa; 206 static u4 IBAA_bb; 207 static u4 IBAA_counter; 208 209 static volatile int IBAA_byte_index; 210 211 212 static void IBAA_Init(void); 213 static void IBAA_Call(void); 214 static void IBAA_Seed(const u_int32_t val); 215 static u_char IBAA_Byte(void); 216 217 /* 218 * Initialize IBAA. 219 */ 220 static void 221 IBAA_Init(void) 222 { 223 size_t i; 224 225 for (i = 0; i < SIZE; ++i) { 226 IBAA_memory[i] = i; 227 } 228 IBAA_aa = IBAA_bb = 0; 229 IBAA_counter = 0; 230 IBAA_byte_index = sizeof(IBAA_results); /* force IBAA_Call() */ 231 } 232 233 /* 234 * PRIVATE: Call IBAA to produce 256 32-bit u4 results. 235 */ 236 static void 237 IBAA_Call (void) 238 { 239 IBAA(IBAA_memory, IBAA_results, &IBAA_aa, &IBAA_bb, &IBAA_counter); 240 IBAA_byte_index = 0; 241 } 242 243 /* 244 * Add a 32-bit u4 seed value into IBAAs memory. Mix the low 4 bits 245 * with 4 bits of PNG data to reduce the possibility of a seeding-based 246 * attack. 247 */ 248 static void 249 IBAA_Seed (const u_int32_t val) 250 { 251 static int memIndex; 252 u4 *iptr; 253 254 iptr = &IBAA_memory[memIndex & MASK]; 255 *iptr = ((*iptr << 3) | (*iptr >> 29)) + (val ^ (IBAA_Byte() & 15)); 256 ++memIndex; 257 } 258 259 /* 260 * Extract a byte from IBAAs 256 32-bit u4 results array. 261 * 262 * NOTE: This code is designed to prevent MP races from taking 263 * IBAA_byte_index out of bounds. 264 */ 265 static u_char 266 IBAA_Byte(void) 267 { 268 u_char result; 269 int index; 270 271 index = IBAA_byte_index; 272 if (index == sizeof(IBAA_results)) { 273 IBAA_Call(); 274 index = 0; 275 } 276 result = ((u_char *)IBAA_results)[index]; 277 IBAA_byte_index = index + 1; 278 return result; 279 } 280 281 /*------------------------------ IBAA ----------------------------------*/ 282 283 284 /*------------------------------- L15 ----------------------------------*/ 285 286 /* 287 * IMPORTANT NOTE: LByteType must be exactly 8-bits in size or this software 288 * will not function correctly. 289 */ 290 typedef unsigned char LByteType; 291 292 #define L15_STATE_SIZE 256 293 294 static LByteType L15_x, L15_y; 295 static LByteType L15_start_x; 296 static LByteType L15_state[L15_STATE_SIZE]; 297 298 /* 299 * PRIVATE FUNCS: 300 */ 301 302 static void L15_Swap(const LByteType pos1, const LByteType pos2); 303 static void L15_InitState(void); 304 static void L15_KSA(const LByteType * const key, 305 const size_t keyLen); 306 static void L15_Discard(const LByteType numCalls); 307 308 /* 309 * PUBLIC INTERFACE: 310 */ 311 static void L15(const LByteType * const key, const size_t keyLen); 312 static LByteType L15_Byte(void); 313 static void L15_Vector(const LByteType * const key, 314 const size_t keyLen); 315 316 static __inline void 317 L15_Swap(const LByteType pos1, const LByteType pos2) 318 { 319 const LByteType save1 = L15_state[pos1]; 320 321 L15_state[pos1] = L15_state[pos2]; 322 L15_state[pos2] = save1; 323 } 324 325 static void 326 L15_InitState (void) 327 { 328 size_t i; 329 for (i = 0; i < L15_STATE_SIZE; ++i) 330 L15_state[i] = i; 331 } 332 333 #define L_SCHEDULE(xx) \ 334 \ 335 for (i = 0; i < L15_STATE_SIZE; ++i) { \ 336 L15_Swap(i, (stateIndex += (L15_state[i] + (xx)))); \ 337 } 338 339 static void 340 L15_KSA (const LByteType * const key, const size_t keyLen) 341 { 342 size_t i, keyIndex; 343 LByteType stateIndex = 0; 344 345 L_SCHEDULE(keyLen); 346 for (keyIndex = 0; keyIndex < keyLen; ++keyIndex) { 347 L_SCHEDULE(key[keyIndex]); 348 } 349 } 350 351 static void 352 L15_Discard(const LByteType numCalls) 353 { 354 LByteType i; 355 for (i = 0; i < numCalls; ++i) { 356 (void)L15_Byte(); 357 } 358 } 359 360 361 /* 362 * PUBLIC INTERFACE: 363 */ 364 static void 365 L15(const LByteType * const key, const size_t keyLen) 366 { 367 L15_x = L15_start_x = 0; 368 L15_y = L15_STATE_SIZE - 1; 369 L15_InitState(); 370 L15_KSA(key, keyLen); 371 L15_Discard(L15_Byte()); 372 } 373 374 static LByteType 375 L15_Byte(void) 376 { 377 LByteType z; 378 379 L15_Swap(L15_state[L15_x], L15_y); 380 z = (L15_state [L15_x++] + L15_state[L15_y--]); 381 if (L15_x == L15_start_x) { 382 --L15_y; 383 } 384 return (L15_state[z]); 385 } 386 387 static void 388 L15_Vector (const LByteType * const key, const size_t keyLen) 389 { 390 L15_KSA(key, keyLen); 391 } 392 393 /*------------------------------- L15 ----------------------------------*/ 394 395 /************************************************************************ 396 * KERNEL INTERFACE * 397 ************************************************************************ 398 * 399 * By Robin J Carey and Matthew Dillon. 400 */ 401 402 static int rand_thread_signal = 1; 403 static void NANOUP_EVENT(void); 404 static thread_t rand_td; 405 static struct spinlock rand_spin; 406 407 static int nrandevents; 408 SYSCTL_INT(_kern, OID_AUTO, nrandevents, CTLFLAG_RD, &nrandevents, 0, ""); 409 static int seedenable; 410 SYSCTL_INT(_kern, OID_AUTO, seedenable, CTLFLAG_RW, &seedenable, 0, ""); 411 412 /* 413 * Called from early boot 414 */ 415 void 416 rand_initialize(void) 417 { 418 struct timespec now; 419 int i; 420 421 spin_init(&rand_spin); 422 423 /* Initialize IBAA. */ 424 IBAA_Init(); 425 426 /* Initialize L15. */ 427 nanouptime(&now); 428 L15((const LByteType *)&now.tv_nsec, sizeof(now.tv_nsec)); 429 for (i = 0; i < (SIZE / 2); ++i) { 430 nanotime(&now); 431 IBAA_Seed(now.tv_nsec); 432 L15_Vector((const LByteType *)&now.tv_nsec, 433 sizeof(now.tv_nsec)); 434 nanouptime(&now); 435 IBAA_Seed(now.tv_nsec); 436 L15_Vector((const LByteType *)&now.tv_nsec, 437 sizeof(now.tv_nsec)); 438 } 439 440 /* 441 * Warm up the generator to get rid of weak initial states. 442 */ 443 for (i = 0; i < 10; ++i) 444 IBAA_Call(); 445 } 446 447 /* 448 * Keyboard events 449 */ 450 void 451 add_keyboard_randomness(u_char scancode) 452 { 453 spin_lock_wr(&rand_spin); 454 L15_Vector((const LByteType *) &scancode, sizeof (scancode)); 455 spin_unlock_wr(&rand_spin); 456 add_interrupt_randomness(0); 457 } 458 459 /* 460 * Interrupt events. This is SMP safe and allowed to race. 461 */ 462 void 463 add_interrupt_randomness(int intr) 464 { 465 if (rand_thread_signal == 0) { 466 rand_thread_signal = 1; 467 lwkt_schedule(rand_td); 468 } 469 } 470 471 /* 472 * True random number source 473 */ 474 void 475 add_true_randomness(int val) 476 { 477 spin_lock_wr(&rand_spin); 478 IBAA_Seed(val); 479 L15_Vector((const LByteType *) &val, sizeof (val)); 480 ++nrandevents; 481 spin_unlock_wr(&rand_spin); 482 } 483 484 int 485 add_buffer_randomness(const char *buf, int bytes) 486 { 487 int error; 488 int i; 489 490 if (seedenable && securelevel <= 0) { 491 while (bytes >= sizeof(int)) { 492 add_true_randomness(*(const int *)buf); 493 buf += sizeof(int); 494 bytes -= sizeof(int); 495 } 496 error = 0; 497 498 /* 499 * Warm up the generator to get rid of weak initial states. 500 */ 501 for (i = 0; i < 10; ++i) 502 IBAA_Call(); 503 } else { 504 error = EPERM; 505 } 506 return (error); 507 } 508 509 /* 510 * Poll (always succeeds) 511 */ 512 int 513 random_poll(cdev_t dev, int events) 514 { 515 int revents = 0; 516 517 if (events & (POLLIN | POLLRDNORM)) 518 revents |= events & (POLLIN | POLLRDNORM); 519 if (events & (POLLOUT | POLLWRNORM)) 520 revents |= events & (POLLOUT | POLLWRNORM); 521 522 return (revents); 523 } 524 525 /* 526 * Kqueue filter (always succeeds) 527 */ 528 int 529 random_filter_read(struct knote *kn, long hint) 530 { 531 return (1); 532 } 533 534 /* 535 * Heavy weight random number generator. May return less then the 536 * requested number of bytes. 537 */ 538 u_int 539 read_random(void *buf, u_int nbytes) 540 { 541 u_int i; 542 543 spin_lock_wr(&rand_spin); 544 for (i = 0; i < nbytes; ++i) 545 ((u_char *)buf)[i] = IBAA_Byte(); 546 spin_unlock_wr(&rand_spin); 547 add_interrupt_randomness(0); 548 return(i); 549 } 550 551 /* 552 * Lightweight random number generator. Must return requested number of 553 * bytes. 554 */ 555 u_int 556 read_random_unlimited(void *buf, u_int nbytes) 557 { 558 u_int i; 559 560 spin_lock_wr(&rand_spin); 561 for (i = 0; i < nbytes; ++i) 562 ((u_char *)buf)[i] = L15_Byte(); 563 spin_unlock_wr(&rand_spin); 564 add_interrupt_randomness(0); 565 return (i); 566 } 567 568 /* 569 * Random number generator helper thread. This limits code overhead from 570 * high frequency events by delaying the clearing of rand_thread_signal. 571 */ 572 static 573 void 574 rand_thread_loop(void *dummy) 575 { 576 int count; 577 578 for (;;) { 579 NANOUP_EVENT (); 580 spin_lock_wr(&rand_spin); 581 count = (int)(L15_Byte() * hz / (256 * 10) + hz / 10); 582 spin_unlock_wr(&rand_spin); 583 tsleep(rand_td, 0, "rwait", count); 584 crit_enter(); 585 lwkt_deschedule_self(rand_td); 586 cpu_sfence(); 587 rand_thread_signal = 0; 588 crit_exit(); 589 lwkt_switch(); 590 } 591 } 592 593 static 594 void 595 rand_thread_init(void) 596 { 597 lwkt_create(rand_thread_loop, NULL, &rand_td, NULL, 0, 0, "random"); 598 } 599 600 SYSINIT(rand, SI_SUB_HELPER_THREADS, SI_ORDER_ANY, rand_thread_init, 0); 601 602 /* 603 * Time-buffered event time-stamping. This is necessary to cutoff higher 604 * event frequencies, e.g. an interrupt occuring at 25Hz. In such cases 605 * the CPU is being chewed and the timestamps are skewed (minimal variation). 606 * Use a nano-second time-delay to limit how many times an Event can occur 607 * in one second; <= 5Hz. Note that this doesn't prevent time-stamp skewing. 608 * This implementation randmoises the time-delay between events, which adds 609 * a layer of security/unpredictability with regard to read-events (a user 610 * controlled input). 611 * 612 * Note: now.tv_nsec should range [ 0 - 1000,000,000 ]. 613 * Note: "ACCUM" is a security measure (result = capped-unknown + unknown), 614 * and also produces an uncapped (>=32-bit) value. 615 */ 616 static void 617 NANOUP_EVENT(void) 618 { 619 static struct timespec ACCUM = { 0, 0 }; 620 static struct timespec NEXT = { 0, 0 }; 621 struct timespec now; 622 623 nanouptime(&now); 624 spin_lock_wr(&rand_spin); 625 if ((now.tv_nsec > NEXT.tv_nsec) || (now.tv_sec != NEXT.tv_sec)) { 626 /* 627 * Randomised time-delay: 200e6 - 350e6 ns; 5 - 2.86 Hz. 628 */ 629 unsigned long one_mil; 630 unsigned long timeDelay; 631 632 one_mil = 1000000UL; /* 0.001 s */ 633 timeDelay = (one_mil * 200) + 634 (((unsigned long)ACCUM.tv_nsec % 151) * one_mil); 635 NEXT.tv_nsec = now.tv_nsec + timeDelay; 636 NEXT.tv_sec = now.tv_sec; 637 ACCUM.tv_nsec += now.tv_nsec; 638 639 /* 640 * The TSC, if present, generally has an even higher 641 * resolution. Integrate a portion of it into our seed. 642 */ 643 if (tsc_present) 644 ACCUM.tv_nsec ^= rdtsc() & 255; 645 646 IBAA_Seed(ACCUM.tv_nsec); 647 L15_Vector((const LByteType *)&ACCUM.tv_nsec, 648 sizeof(ACCUM.tv_nsec)); 649 ++nrandevents; 650 } 651 spin_unlock_wr(&rand_spin); 652 } 653 654