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 #include <sys/mplock2.h> 144 145 /* 146 * Portability note: The u_char/unsigned char type is used where 147 * uint8_t from <stdint.h> or u_int8_t from <sys/types.h> should really 148 * be being used. On FreeBSD, it is safe to make the assumption that these 149 * different types are equivalent (on all architectures). 150 * The FreeBSD <sys/crypto/rc4> module also makes this assumption. 151 */ 152 153 /*------------------------------ IBAA ----------------------------------*/ 154 155 /*-------------------------- IBAA CSPRNG -------------------------------*/ 156 157 /* 158 * NOTE: The original source code from which this source code (IBAA) 159 * was taken has no copyright/license. The algorithm has no patent 160 * and is freely/publicly available from: 161 * 162 * http://www.burtleburtle.net/bob/rand/isaac.html 163 */ 164 165 /* 166 * ^ means XOR, & means bitwise AND, a<<b means shift a by b. 167 * barrel(a) shifts a 19 bits to the left, and bits wrap around 168 * ind(x) is (x AND 255), or (x mod 256) 169 */ 170 typedef u_int32_t u4; /* unsigned four bytes, 32 bits */ 171 172 #define ALPHA (8) 173 #define SIZE (1 << ALPHA) 174 #define MASK (SIZE - 1) 175 #define ind(x) ((x) & (SIZE - 1)) 176 #define barrel(a) (((a) << 20) ^ ((a) >> 12)) /* beta=32,shift=20 */ 177 178 static void IBAA 179 ( 180 u4 *m, /* Memory: array of SIZE ALPHA-bit terms */ 181 u4 *r, /* Results: the sequence, same size as m */ 182 u4 *aa, /* Accumulator: a single value */ 183 u4 *bb, /* the previous result */ 184 u4 *counter /* counter */ 185 ) 186 { 187 u4 a, b, x, y, i; 188 189 a = *aa; 190 b = *bb + *counter; 191 ++*counter; 192 for (i = 0; i < SIZE; ++i) { 193 x = m[i]; 194 a = barrel(a) + m[ind(i + (SIZE / 2))]; /* set a */ 195 m[i] = y = m[ind(x)] + a + b; /* set m */ 196 r[i] = b = m[ind(y >> ALPHA)] + x; /* set r */ 197 } 198 *bb = b; *aa = a; 199 } 200 201 /*-------------------------- IBAA CSPRNG -------------------------------*/ 202 203 204 static u4 IBAA_memory[SIZE]; 205 static u4 IBAA_results[SIZE]; 206 static u4 IBAA_aa; 207 static u4 IBAA_bb; 208 static u4 IBAA_counter; 209 210 static volatile int IBAA_byte_index; 211 212 213 static void IBAA_Init(void); 214 static void IBAA_Call(void); 215 static void IBAA_Seed(const u_int32_t val); 216 static u_char IBAA_Byte(void); 217 218 /* 219 * Initialize IBAA. 220 */ 221 static void 222 IBAA_Init(void) 223 { 224 size_t i; 225 226 for (i = 0; i < SIZE; ++i) { 227 IBAA_memory[i] = i; 228 } 229 IBAA_aa = IBAA_bb = 0; 230 IBAA_counter = 0; 231 IBAA_byte_index = sizeof(IBAA_results); /* force IBAA_Call() */ 232 } 233 234 /* 235 * PRIVATE: Call IBAA to produce 256 32-bit u4 results. 236 */ 237 static void 238 IBAA_Call (void) 239 { 240 IBAA(IBAA_memory, IBAA_results, &IBAA_aa, &IBAA_bb, &IBAA_counter); 241 IBAA_byte_index = 0; 242 } 243 244 /* 245 * Add a 32-bit u4 seed value into IBAAs memory. Mix the low 4 bits 246 * with 4 bits of PNG data to reduce the possibility of a seeding-based 247 * attack. 248 */ 249 static void 250 IBAA_Seed (const u_int32_t val) 251 { 252 static int memIndex; 253 u4 *iptr; 254 255 iptr = &IBAA_memory[memIndex & MASK]; 256 *iptr = ((*iptr << 3) | (*iptr >> 29)) + (val ^ (IBAA_Byte() & 15)); 257 ++memIndex; 258 } 259 260 /* 261 * Extract a byte from IBAAs 256 32-bit u4 results array. 262 * 263 * NOTE: This code is designed to prevent MP races from taking 264 * IBAA_byte_index out of bounds. 265 */ 266 static u_char 267 IBAA_Byte(void) 268 { 269 u_char result; 270 int index; 271 272 index = IBAA_byte_index; 273 if (index == sizeof(IBAA_results)) { 274 IBAA_Call(); 275 index = 0; 276 } 277 result = ((u_char *)IBAA_results)[index]; 278 IBAA_byte_index = index + 1; 279 return result; 280 } 281 282 /*------------------------------ IBAA ----------------------------------*/ 283 284 285 /*------------------------------- L15 ----------------------------------*/ 286 287 /* 288 * IMPORTANT NOTE: LByteType must be exactly 8-bits in size or this software 289 * will not function correctly. 290 */ 291 typedef unsigned char LByteType; 292 293 #define L15_STATE_SIZE 256 294 295 static LByteType L15_x, L15_y; 296 static LByteType L15_start_x; 297 static LByteType L15_state[L15_STATE_SIZE]; 298 299 /* 300 * PRIVATE FUNCS: 301 */ 302 303 static void L15_Swap(const LByteType pos1, const LByteType pos2); 304 static void L15_InitState(void); 305 static void L15_KSA(const LByteType * const key, 306 const size_t keyLen); 307 static void L15_Discard(const LByteType numCalls); 308 309 /* 310 * PUBLIC INTERFACE: 311 */ 312 static void L15(const LByteType * const key, const size_t keyLen); 313 static LByteType L15_Byte(void); 314 static void L15_Vector(const LByteType * const key, 315 const size_t keyLen); 316 317 static __inline void 318 L15_Swap(const LByteType pos1, const LByteType pos2) 319 { 320 const LByteType save1 = L15_state[pos1]; 321 322 L15_state[pos1] = L15_state[pos2]; 323 L15_state[pos2] = save1; 324 } 325 326 static void 327 L15_InitState (void) 328 { 329 size_t i; 330 for (i = 0; i < L15_STATE_SIZE; ++i) 331 L15_state[i] = i; 332 } 333 334 #define L_SCHEDULE(xx) \ 335 \ 336 for (i = 0; i < L15_STATE_SIZE; ++i) { \ 337 L15_Swap(i, (stateIndex += (L15_state[i] + (xx)))); \ 338 } 339 340 static void 341 L15_KSA (const LByteType * const key, const size_t keyLen) 342 { 343 size_t i, keyIndex; 344 LByteType stateIndex = 0; 345 346 L_SCHEDULE(keyLen); 347 for (keyIndex = 0; keyIndex < keyLen; ++keyIndex) { 348 L_SCHEDULE(key[keyIndex]); 349 } 350 } 351 352 static void 353 L15_Discard(const LByteType numCalls) 354 { 355 LByteType i; 356 for (i = 0; i < numCalls; ++i) { 357 (void)L15_Byte(); 358 } 359 } 360 361 362 /* 363 * PUBLIC INTERFACE: 364 */ 365 static void 366 L15(const LByteType * const key, const size_t keyLen) 367 { 368 L15_x = L15_start_x = 0; 369 L15_y = L15_STATE_SIZE - 1; 370 L15_InitState(); 371 L15_KSA(key, keyLen); 372 L15_Discard(L15_Byte()); 373 } 374 375 static LByteType 376 L15_Byte(void) 377 { 378 LByteType z; 379 380 L15_Swap(L15_state[L15_x], L15_y); 381 z = (L15_state [L15_x++] + L15_state[L15_y--]); 382 if (L15_x == L15_start_x) { 383 --L15_y; 384 } 385 return (L15_state[z]); 386 } 387 388 static void 389 L15_Vector (const LByteType * const key, const size_t keyLen) 390 { 391 L15_KSA(key, keyLen); 392 } 393 394 /*------------------------------- L15 ----------------------------------*/ 395 396 /************************************************************************ 397 * KERNEL INTERFACE * 398 ************************************************************************ 399 * 400 * By Robin J Carey and Matthew Dillon. 401 */ 402 403 static int rand_thread_signal = 1; 404 static void NANOUP_EVENT(void); 405 static thread_t rand_td; 406 static struct spinlock rand_spin; 407 408 static int nrandevents; 409 SYSCTL_INT(_kern, OID_AUTO, nrandevents, CTLFLAG_RD, &nrandevents, 0, ""); 410 static int seedenable; 411 SYSCTL_INT(_kern, OID_AUTO, seedenable, CTLFLAG_RW, &seedenable, 0, ""); 412 413 /* 414 * Called from early boot 415 */ 416 void 417 rand_initialize(void) 418 { 419 struct timespec now; 420 int i; 421 422 spin_init(&rand_spin); 423 424 /* Initialize IBAA. */ 425 IBAA_Init(); 426 427 /* Initialize L15. */ 428 nanouptime(&now); 429 L15((const LByteType *)&now.tv_nsec, sizeof(now.tv_nsec)); 430 for (i = 0; i < (SIZE / 2); ++i) { 431 nanotime(&now); 432 IBAA_Seed(now.tv_nsec); 433 L15_Vector((const LByteType *)&now.tv_nsec, 434 sizeof(now.tv_nsec)); 435 nanouptime(&now); 436 IBAA_Seed(now.tv_nsec); 437 L15_Vector((const LByteType *)&now.tv_nsec, 438 sizeof(now.tv_nsec)); 439 } 440 441 /* 442 * Warm up the generator to get rid of weak initial states. 443 */ 444 for (i = 0; i < 10; ++i) 445 IBAA_Call(); 446 } 447 448 /* 449 * Keyboard events 450 */ 451 void 452 add_keyboard_randomness(u_char scancode) 453 { 454 spin_lock(&rand_spin); 455 L15_Vector((const LByteType *) &scancode, sizeof (scancode)); 456 spin_unlock(&rand_spin); 457 add_interrupt_randomness(0); 458 } 459 460 /* 461 * Interrupt events. This is SMP safe and allowed to race. 462 */ 463 void 464 add_interrupt_randomness(int intr) 465 { 466 if (rand_thread_signal == 0) { 467 rand_thread_signal = 1; 468 lwkt_schedule(rand_td); 469 } 470 } 471 472 /* 473 * True random number source 474 */ 475 void 476 add_true_randomness(int val) 477 { 478 spin_lock(&rand_spin); 479 IBAA_Seed(val); 480 L15_Vector((const LByteType *) &val, sizeof (val)); 481 ++nrandevents; 482 spin_unlock(&rand_spin); 483 } 484 485 int 486 add_buffer_randomness(const char *buf, int bytes) 487 { 488 int error; 489 int i; 490 491 if (seedenable && securelevel <= 0) { 492 while (bytes >= sizeof(int)) { 493 add_true_randomness(*(const int *)buf); 494 buf += sizeof(int); 495 bytes -= sizeof(int); 496 } 497 error = 0; 498 499 /* 500 * Warm up the generator to get rid of weak initial states. 501 */ 502 for (i = 0; i < 10; ++i) 503 IBAA_Call(); 504 } else { 505 error = EPERM; 506 } 507 return (error); 508 } 509 510 /* 511 * Poll (always succeeds) 512 */ 513 int 514 random_poll(cdev_t dev, int events) 515 { 516 int revents = 0; 517 518 if (events & (POLLIN | POLLRDNORM)) 519 revents |= events & (POLLIN | POLLRDNORM); 520 if (events & (POLLOUT | POLLWRNORM)) 521 revents |= events & (POLLOUT | POLLWRNORM); 522 523 return (revents); 524 } 525 526 /* 527 * Kqueue filter (always succeeds) 528 */ 529 int 530 random_filter_read(struct knote *kn, long hint) 531 { 532 return (1); 533 } 534 535 /* 536 * Heavy weight random number generator. May return less then the 537 * requested number of bytes. 538 */ 539 u_int 540 read_random(void *buf, u_int nbytes) 541 { 542 u_int i; 543 544 spin_lock(&rand_spin); 545 for (i = 0; i < nbytes; ++i) 546 ((u_char *)buf)[i] = IBAA_Byte(); 547 spin_unlock(&rand_spin); 548 add_interrupt_randomness(0); 549 return(i); 550 } 551 552 /* 553 * Lightweight random number generator. Must return requested number of 554 * bytes. 555 */ 556 u_int 557 read_random_unlimited(void *buf, u_int nbytes) 558 { 559 u_int i; 560 561 spin_lock(&rand_spin); 562 for (i = 0; i < nbytes; ++i) 563 ((u_char *)buf)[i] = L15_Byte(); 564 spin_unlock(&rand_spin); 565 add_interrupt_randomness(0); 566 return (i); 567 } 568 569 /* 570 * Random number generator helper thread. This limits code overhead from 571 * high frequency events by delaying the clearing of rand_thread_signal. 572 * 573 * MPSAFE thread 574 */ 575 static 576 void 577 rand_thread_loop(void *dummy) 578 { 579 int count; 580 581 for (;;) { 582 NANOUP_EVENT (); 583 spin_lock(&rand_spin); 584 count = (int)(L15_Byte() * hz / (256 * 10) + hz / 10 + 1); 585 spin_unlock(&rand_spin); 586 tsleep(rand_td, 0, "rwait", count); 587 crit_enter(); 588 lwkt_deschedule_self(rand_td); 589 cpu_sfence(); 590 rand_thread_signal = 0; 591 crit_exit(); 592 lwkt_switch(); 593 } 594 } 595 596 static 597 void 598 rand_thread_init(void) 599 { 600 lwkt_create(rand_thread_loop, NULL, &rand_td, NULL, 0, 0, "random"); 601 } 602 603 SYSINIT(rand, SI_SUB_HELPER_THREADS, SI_ORDER_ANY, rand_thread_init, 0); 604 605 /* 606 * Time-buffered event time-stamping. This is necessary to cutoff higher 607 * event frequencies, e.g. an interrupt occuring at 25Hz. In such cases 608 * the CPU is being chewed and the timestamps are skewed (minimal variation). 609 * Use a nano-second time-delay to limit how many times an Event can occur 610 * in one second; <= 5Hz. Note that this doesn't prevent time-stamp skewing. 611 * This implementation randmoises the time-delay between events, which adds 612 * a layer of security/unpredictability with regard to read-events (a user 613 * controlled input). 614 * 615 * Note: now.tv_nsec should range [ 0 - 1000,000,000 ]. 616 * Note: "ACCUM" is a security measure (result = capped-unknown + unknown), 617 * and also produces an uncapped (>=32-bit) value. 618 */ 619 static void 620 NANOUP_EVENT(void) 621 { 622 static struct timespec ACCUM = { 0, 0 }; 623 static struct timespec NEXT = { 0, 0 }; 624 struct timespec now; 625 626 nanouptime(&now); 627 spin_lock(&rand_spin); 628 if ((now.tv_nsec > NEXT.tv_nsec) || (now.tv_sec != NEXT.tv_sec)) { 629 /* 630 * Randomised time-delay: 200e6 - 350e6 ns; 5 - 2.86 Hz. 631 */ 632 unsigned long one_mil; 633 unsigned long timeDelay; 634 635 one_mil = 1000000UL; /* 0.001 s */ 636 timeDelay = (one_mil * 200) + 637 (((unsigned long)ACCUM.tv_nsec % 151) * one_mil); 638 NEXT.tv_nsec = now.tv_nsec + timeDelay; 639 NEXT.tv_sec = now.tv_sec; 640 ACCUM.tv_nsec += now.tv_nsec; 641 642 /* 643 * The TSC, if present, generally has an even higher 644 * resolution. Integrate a portion of it into our seed. 645 */ 646 if (tsc_present) 647 ACCUM.tv_nsec ^= rdtsc() & 255; 648 649 IBAA_Seed(ACCUM.tv_nsec); 650 L15_Vector((const LByteType *)&ACCUM.tv_nsec, 651 sizeof(ACCUM.tv_nsec)); 652 ++nrandevents; 653 } 654 spin_unlock(&rand_spin); 655 } 656 657