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/random.h> 132 #include <sys/systimer.h> 133 #include <sys/time.h> 134 #include <sys/proc.h> 135 #include <sys/lock.h> 136 #include <sys/sysctl.h> 137 #include <sys/spinlock.h> 138 #include <machine/clock.h> 139 140 #include <sys/thread2.h> 141 #include <sys/spinlock2.h> 142 143 /* 144 * Portability note: The u_char/unsigned char type is used where 145 * uint8_t from <stdint.h> or u_int8_t from <sys/types.h> should really 146 * be being used. On FreeBSD, it is safe to make the assumption that these 147 * different types are equivalent (on all architectures). 148 * The FreeBSD <sys/crypto/rc4> module also makes this assumption. 149 */ 150 151 /*------------------------------ IBAA ----------------------------------*/ 152 153 /*-------------------------- IBAA CSPRNG -------------------------------*/ 154 155 /* 156 * NOTE: The original source code from which this source code (IBAA) 157 * was taken has no copyright/license. The algorithm has no patent 158 * and is freely/publicly available from: 159 * 160 * http://www.burtleburtle.net/bob/rand/isaac.html 161 */ 162 163 /* 164 * ^ means XOR, & means bitwise AND, a<<b means shift a by b. 165 * barrel(a) shifts a 19 bits to the left, and bits wrap around 166 * ind(x) is (x AND 255), or (x mod 256) 167 */ 168 typedef u_int32_t u4; /* unsigned four bytes, 32 bits */ 169 170 #define ALPHA (8) 171 #define SIZE (1 << ALPHA) 172 #define MASK (SIZE - 1) 173 #define ind(x) ((x) & (SIZE - 1)) 174 #define barrel(a) (((a) << 19) ^ ((a) >> 13)) /* beta=32,shift=19 */ 175 176 static void IBAA 177 ( 178 u4 *m, /* Memory: array of SIZE ALPHA-bit terms */ 179 u4 *r, /* Results: the sequence, same size as m */ 180 u4 *aa, /* Accumulator: a single value */ 181 u4 *bb /* the previous result */ 182 ) 183 { 184 u4 a, b, x, y, i; 185 186 a = *aa; b = *bb; 187 for (i = 0; i < SIZE; ++i) { 188 x = m[i]; 189 a = barrel(a) + m[ind(i + (SIZE / 2))]; /* set a */ 190 m[i] = y = m[ind(x)] + a + b; /* set m */ 191 r[i] = b = m[ind(y >> ALPHA)] + x; /* set r */ 192 } 193 *bb = b; *aa = a; 194 } 195 196 /*-------------------------- IBAA CSPRNG -------------------------------*/ 197 198 199 static u4 IBAA_memory[SIZE]; 200 static u4 IBAA_results[SIZE]; 201 static u4 IBAA_aa; 202 static u4 IBAA_bb; 203 204 static volatile int IBAA_byte_index; 205 206 207 static void IBAA_Init(void); 208 static void IBAA_Call(void); 209 static void IBAA_Seed(const u_int32_t val); 210 static u_char IBAA_Byte(void); 211 212 /* 213 * Initialize IBAA. 214 */ 215 static void 216 IBAA_Init(void) 217 { 218 size_t i; 219 220 for (i = 0; i < SIZE; ++i) { 221 IBAA_memory[i] = i; 222 } 223 IBAA_aa = IBAA_bb = 0; 224 IBAA_byte_index = sizeof(IBAA_results); /* force IBAA_Call() */ 225 } 226 227 /* 228 * PRIVATE: Call IBAA to produce 256 32-bit u4 results. 229 */ 230 static void 231 IBAA_Call (void) 232 { 233 IBAA(IBAA_memory, IBAA_results, &IBAA_aa, &IBAA_bb); 234 IBAA_byte_index = 0; 235 } 236 237 /* 238 * Add a 32-bit u4 seed value into IBAAs memory. Mix the low 4 bits 239 * with 4 bits of PNG data to reduce the possibility of a seeding-based 240 * attack. 241 */ 242 static void 243 IBAA_Seed (const u_int32_t val) 244 { 245 static int memIndex; 246 u4 *iptr; 247 248 iptr = &IBAA_memory[memIndex & MASK]; 249 *iptr = ((*iptr << 3) | (*iptr >> 29)) + (val ^ (IBAA_Byte() & 15)); 250 ++memIndex; 251 } 252 253 /* 254 * Extract a byte from IBAAs 256 32-bit u4 results array. 255 * 256 * NOTE: This code is designed to prevent MP races from taking 257 * IBAA_byte_index out of bounds. 258 */ 259 static u_char 260 IBAA_Byte(void) 261 { 262 u_char result; 263 int index; 264 265 index = IBAA_byte_index; 266 if (index == sizeof(IBAA_results)) { 267 IBAA_Call(); 268 index = 0; 269 } 270 result = ((u_char *)IBAA_results)[index]; 271 IBAA_byte_index = index + 1; 272 return result; 273 } 274 275 /*------------------------------ IBAA ----------------------------------*/ 276 277 278 /*------------------------------- L15 ----------------------------------*/ 279 280 /* 281 * IMPORTANT NOTE: LByteType must be exactly 8-bits in size or this software 282 * will not function correctly. 283 */ 284 typedef unsigned char LByteType; 285 286 #define L15_STATE_SIZE 256 287 288 static LByteType L15_x, L15_y; 289 static LByteType L15_start_x; 290 static LByteType L15_state[L15_STATE_SIZE]; 291 292 /* 293 * PRIVATE FUNCS: 294 */ 295 296 static void L15_Swap(const LByteType pos1, const LByteType pos2); 297 static void L15_InitState(void); 298 static void L15_KSA(const LByteType * const key, 299 const size_t keyLen); 300 static void L15_Discard(const LByteType numCalls); 301 302 /* 303 * PUBLIC INTERFACE: 304 */ 305 static void L15(const LByteType * const key, const size_t keyLen); 306 static LByteType L15_Byte(void); 307 static void L15_Vector(const LByteType * const key, 308 const size_t keyLen); 309 310 static __inline void 311 L15_Swap(const LByteType pos1, const LByteType pos2) 312 { 313 const LByteType save1 = L15_state[pos1]; 314 315 L15_state[pos1] = L15_state[pos2]; 316 L15_state[pos2] = save1; 317 } 318 319 static void 320 L15_InitState (void) 321 { 322 size_t i; 323 for (i = 0; i < L15_STATE_SIZE; ++i) 324 L15_state[i] = i; 325 } 326 327 #define L_SCHEDULE(xx) \ 328 \ 329 for (i = 0; i < L15_STATE_SIZE; ++i) { \ 330 L15_Swap(i, (stateIndex += (L15_state[i] + (xx)))); \ 331 } 332 333 static void 334 L15_KSA (const LByteType * const key, const size_t keyLen) 335 { 336 size_t i, keyIndex; 337 LByteType stateIndex = 0; 338 339 L_SCHEDULE(keyLen); 340 for (keyIndex = 0; keyIndex < keyLen; ++keyIndex) { 341 L_SCHEDULE(key[keyIndex]); 342 } 343 } 344 345 static void 346 L15_Discard(const LByteType numCalls) 347 { 348 LByteType i; 349 for (i = 0; i < numCalls; ++i) { 350 (void)L15_Byte(); 351 } 352 } 353 354 355 /* 356 * PUBLIC INTERFACE: 357 */ 358 static void 359 L15(const LByteType * const key, const size_t keyLen) 360 { 361 L15_x = L15_start_x = 0; 362 L15_y = L15_STATE_SIZE - 1; 363 L15_InitState(); 364 L15_KSA(key, keyLen); 365 L15_Discard(L15_Byte()); 366 } 367 368 static LByteType 369 L15_Byte(void) 370 { 371 LByteType z; 372 373 L15_Swap(L15_state[L15_x], L15_y); 374 z = (L15_state [L15_x++] + L15_state[L15_y--]); 375 if (L15_x == L15_start_x) { 376 --L15_y; 377 } 378 return (L15_state[z]); 379 } 380 381 static void 382 L15_Vector (const LByteType * const key, const size_t keyLen) 383 { 384 L15_KSA(key, keyLen); 385 } 386 387 /*------------------------------- L15 ----------------------------------*/ 388 389 /************************************************************************ 390 * KERNEL INTERFACE * 391 ************************************************************************ 392 * 393 * By Robin J Carey and Matthew Dillon. 394 */ 395 396 static int rand_thread_signal = 1; 397 static void NANOUP_EVENT(void); 398 static thread_t rand_td; 399 static struct spinlock rand_spin; 400 401 static int nrandevents; 402 SYSCTL_INT(_kern, OID_AUTO, nrandevents, CTLFLAG_RD, &nrandevents, 0, ""); 403 static int seedenable; 404 SYSCTL_INT(_kern, OID_AUTO, seedenable, CTLFLAG_RW, &seedenable, 0, ""); 405 406 /* 407 * Called from early boot 408 */ 409 void 410 rand_initialize(void) 411 { 412 struct timespec now; 413 int i; 414 415 spin_init(&rand_spin); 416 417 /* Initialize IBAA. */ 418 IBAA_Init(); 419 420 /* Initialize L15. */ 421 nanouptime(&now); 422 L15((const LByteType *)&now.tv_nsec, sizeof(now.tv_nsec)); 423 for (i = 0; i < (SIZE / 2); ++i) { 424 nanotime(&now); 425 IBAA_Seed(now.tv_nsec); 426 L15_Vector((const LByteType *)&now.tv_nsec, 427 sizeof(now.tv_nsec)); 428 nanouptime(&now); 429 IBAA_Seed(now.tv_nsec); 430 L15_Vector((const LByteType *)&now.tv_nsec, 431 sizeof(now.tv_nsec)); 432 } 433 434 /* 435 * Warm up the generator to get rid of weak initial states. 436 */ 437 for (i = 0; i < 10; ++i) 438 IBAA_Call(); 439 } 440 441 /* 442 * Keyboard events 443 */ 444 void 445 add_keyboard_randomness(u_char scancode) 446 { 447 spin_lock_wr(&rand_spin); 448 L15_Vector((const LByteType *) &scancode, sizeof (scancode)); 449 spin_unlock_wr(&rand_spin); 450 add_interrupt_randomness(0); 451 } 452 453 /* 454 * Interrupt events. This is SMP safe and allowed to race. 455 */ 456 void 457 add_interrupt_randomness(int intr) 458 { 459 if (rand_thread_signal == 0) { 460 rand_thread_signal = 1; 461 lwkt_schedule(rand_td); 462 } 463 } 464 465 /* 466 * True random number source 467 */ 468 void 469 add_true_randomness(int val) 470 { 471 spin_lock_wr(&rand_spin); 472 IBAA_Seed(val); 473 L15_Vector((const LByteType *) &val, sizeof (val)); 474 ++nrandevents; 475 spin_unlock_wr(&rand_spin); 476 } 477 478 int 479 add_buffer_randomness(const char *buf, int bytes) 480 { 481 int error; 482 int i; 483 484 if (seedenable && securelevel <= 0) { 485 while (bytes >= sizeof(int)) { 486 add_true_randomness(*(const int *)buf); 487 buf += sizeof(int); 488 bytes -= sizeof(int); 489 } 490 error = 0; 491 492 /* 493 * Warm up the generator to get rid of weak initial states. 494 */ 495 for (i = 0; i < 10; ++i) 496 IBAA_Call(); 497 } else { 498 error = EPERM; 499 } 500 return (error); 501 } 502 503 /* 504 * Poll (always succeeds) 505 */ 506 int 507 random_poll(cdev_t dev, int events) 508 { 509 int revents = 0; 510 511 if (events & (POLLIN | POLLRDNORM)) 512 revents |= events & (POLLIN | POLLRDNORM); 513 if (events & (POLLOUT | POLLWRNORM)) 514 revents |= events & (POLLOUT | POLLWRNORM); 515 516 return (revents); 517 } 518 519 /* 520 * Heavy weight random number generator. May return less then the 521 * requested number of bytes. 522 */ 523 u_int 524 read_random(void *buf, u_int nbytes) 525 { 526 u_int i; 527 528 spin_lock_wr(&rand_spin); 529 for (i = 0; i < nbytes; ++i) 530 ((u_char *)buf)[i] = IBAA_Byte(); 531 spin_unlock_wr(&rand_spin); 532 add_interrupt_randomness(0); 533 return(i); 534 } 535 536 /* 537 * Lightweight random number generator. Must return requested number of 538 * bytes. 539 */ 540 u_int 541 read_random_unlimited(void *buf, u_int nbytes) 542 { 543 u_int i; 544 545 spin_lock_wr(&rand_spin); 546 for (i = 0; i < nbytes; ++i) 547 ((u_char *)buf)[i] = L15_Byte(); 548 spin_unlock_wr(&rand_spin); 549 add_interrupt_randomness(0); 550 return (i); 551 } 552 553 /* 554 * Random number generator helper thread. This limits code overhead from 555 * high frequency events by delaying the clearing of rand_thread_signal. 556 */ 557 static 558 void 559 rand_thread_loop(void *dummy) 560 { 561 int count; 562 563 for (;;) { 564 NANOUP_EVENT (); 565 spin_lock_wr(&rand_spin); 566 count = (int)(L15_Byte() * hz / (256 * 10) + hz / 10); 567 spin_unlock_wr(&rand_spin); 568 tsleep(rand_td, 0, "rwait", count); 569 crit_enter(); 570 lwkt_deschedule_self(rand_td); 571 cpu_sfence(); 572 rand_thread_signal = 0; 573 crit_exit(); 574 lwkt_switch(); 575 } 576 } 577 578 static 579 void 580 rand_thread_init(void) 581 { 582 lwkt_create(rand_thread_loop, NULL, &rand_td, NULL, 0, 0, "random"); 583 } 584 585 SYSINIT(rand, SI_SUB_HELPER_THREADS, SI_ORDER_ANY, rand_thread_init, 0); 586 587 /* 588 * Time-buffered event time-stamping. This is necessary to cutoff higher 589 * event frequencies, e.g. an interrupt occuring at 25Hz. In such cases 590 * the CPU is being chewed and the timestamps are skewed (minimal variation). 591 * Use a nano-second time-delay to limit how many times an Event can occur 592 * in one second; <= 5Hz. Note that this doesn't prevent time-stamp skewing. 593 * This implementation randmoises the time-delay between events, which adds 594 * a layer of security/unpredictability with regard to read-events (a user 595 * controlled input). 596 * 597 * Note: now.tv_nsec should range [ 0 - 1000,000,000 ]. 598 * Note: "ACCUM" is a security measure (result = capped-unknown + unknown), 599 * and also produces an uncapped (>=32-bit) value. 600 */ 601 static void 602 NANOUP_EVENT(void) 603 { 604 static struct timespec ACCUM = { 0, 0 }; 605 static struct timespec NEXT = { 0, 0 }; 606 struct timespec now; 607 608 nanouptime(&now); 609 spin_lock_wr(&rand_spin); 610 if ((now.tv_nsec > NEXT.tv_nsec) || (now.tv_sec != NEXT.tv_sec)) { 611 /* 612 * Randomised time-delay: 200e6 - 350e6 ns; 5 - 2.86 Hz. 613 */ 614 unsigned long one_mil; 615 unsigned long timeDelay; 616 617 one_mil = 1000000UL; /* 0.001 s */ 618 timeDelay = (one_mil * 200) + 619 (((unsigned long)ACCUM.tv_nsec % 151) * one_mil); 620 NEXT.tv_nsec = now.tv_nsec + timeDelay; 621 NEXT.tv_sec = now.tv_sec; 622 ACCUM.tv_nsec += now.tv_nsec; 623 624 /* 625 * The TSC, if present, generally has an even higher 626 * resolution. Integrate a portion of it into our seed. 627 */ 628 if (tsc_present) 629 ACCUM.tv_nsec ^= rdtsc() & 255; 630 631 IBAA_Seed(ACCUM.tv_nsec); 632 L15_Vector((const LByteType *)&ACCUM.tv_nsec, 633 sizeof(ACCUM.tv_nsec)); 634 ++nrandevents; 635 } 636 spin_unlock_wr(&rand_spin); 637 } 638 639