1/* 2 * Copyright (c) 2005, Jon Seymour 3 * 4 * For more information about epoch theory on which this module is based, 5 * refer to http://blackcubes.dyndns.org/epoch/. That web page defines 6 * terms such as "epoch" and "minimal, non-linear epoch" and provides rationales 7 * for some of the algorithms used here. 8 * 9 */ 10#include <stdlib.h> 11 12/* Provides arbitrary precision integers required to accurately represent 13 * fractional mass: */ 14#include <openssl/bn.h> 15 16#include "cache.h" 17#include "commit.h" 18#include "epoch.h" 19 20struct fraction { 21 BIGNUM numerator; 22 BIGNUM denominator; 23}; 24 25#define HAS_EXACTLY_ONE_PARENT(n) ((n)->parents && !(n)->parents->next) 26 27static BN_CTX *context = NULL; 28static struct fraction *one = NULL; 29static struct fraction *zero = NULL; 30 31static BN_CTX *get_BN_CTX() 32{ 33 if (!context) { 34 context = BN_CTX_new(); 35 } 36 return context; 37} 38 39static struct fraction *new_zero() 40{ 41 struct fraction *result = xmalloc(sizeof(*result)); 42 BN_init(&result->numerator); 43 BN_init(&result->denominator); 44 BN_zero(&result->numerator); 45 BN_one(&result->denominator); 46 return result; 47} 48 49static void clear_fraction(struct fraction *fraction) 50{ 51 BN_clear(&fraction->numerator); 52 BN_clear(&fraction->denominator); 53} 54 55static struct fraction *divide(struct fraction *result, struct fraction *fraction, int divisor) 56{ 57 BIGNUM bn_divisor; 58 59 BN_init(&bn_divisor); 60 BN_set_word(&bn_divisor, divisor); 61 62 BN_copy(&result->numerator, &fraction->numerator); 63 BN_mul(&result->denominator, &fraction->denominator, &bn_divisor, get_BN_CTX()); 64 65 BN_clear(&bn_divisor); 66 return result; 67} 68 69static struct fraction *init_fraction(struct fraction *fraction) 70{ 71 BN_init(&fraction->numerator); 72 BN_init(&fraction->denominator); 73 BN_zero(&fraction->numerator); 74 BN_one(&fraction->denominator); 75 return fraction; 76} 77 78static struct fraction *get_one() 79{ 80 if (!one) { 81 one = new_zero(); 82 BN_one(&one->numerator); 83 } 84 return one; 85} 86 87static struct fraction *get_zero() 88{ 89 if (!zero) { 90 zero = new_zero(); 91 } 92 return zero; 93} 94 95static struct fraction *copy(struct fraction *to, struct fraction *from) 96{ 97 BN_copy(&to->numerator, &from->numerator); 98 BN_copy(&to->denominator, &from->denominator); 99 return to; 100} 101 102static struct fraction *add(struct fraction *result, struct fraction *left, struct fraction *right) 103{ 104 BIGNUM a, b, gcd; 105 106 BN_init(&a); 107 BN_init(&b); 108 BN_init(&gcd); 109 110 BN_mul(&a, &left->numerator, &right->denominator, get_BN_CTX()); 111 BN_mul(&b, &left->denominator, &right->numerator, get_BN_CTX()); 112 BN_mul(&result->denominator, &left->denominator, &right->denominator, get_BN_CTX()); 113 BN_add(&result->numerator, &a, &b); 114 115 BN_gcd(&gcd, &result->denominator, &result->numerator, get_BN_CTX()); 116 BN_div(&result->denominator, NULL, &result->denominator, &gcd, get_BN_CTX()); 117 BN_div(&result->numerator, NULL, &result->numerator, &gcd, get_BN_CTX()); 118 119 BN_clear(&a); 120 BN_clear(&b); 121 BN_clear(&gcd); 122 123 return result; 124} 125 126static int compare(struct fraction *left, struct fraction *right) 127{ 128 BIGNUM a, b; 129 int result; 130 131 BN_init(&a); 132 BN_init(&b); 133 134 BN_mul(&a, &left->numerator, &right->denominator, get_BN_CTX()); 135 BN_mul(&b, &left->denominator, &right->numerator, get_BN_CTX()); 136 137 result = BN_cmp(&a, &b); 138 139 BN_clear(&a); 140 BN_clear(&b); 141 142 return result; 143} 144 145struct mass_counter { 146 struct fraction seen; 147 struct fraction pending; 148}; 149 150static struct mass_counter *new_mass_counter(struct commit *commit, struct fraction *pending) 151{ 152 struct mass_counter *mass_counter = xmalloc(sizeof(*mass_counter)); 153 memset(mass_counter, 0, sizeof(*mass_counter)); 154 155 init_fraction(&mass_counter->seen); 156 init_fraction(&mass_counter->pending); 157 158 copy(&mass_counter->pending, pending); 159 copy(&mass_counter->seen, get_zero()); 160 161 if (commit->object.util) { 162 die("multiple attempts to initialize mass counter for %s", 163 sha1_to_hex(commit->object.sha1)); 164 } 165 166 commit->object.util = mass_counter; 167 168 return mass_counter; 169} 170 171static void free_mass_counter(struct mass_counter *counter) 172{ 173 clear_fraction(&counter->seen); 174 clear_fraction(&counter->pending); 175 free(counter); 176} 177 178/* 179 * Finds the base commit of a list of commits. 180 * 181 * One property of the commit being searched for is that every commit reachable 182 * from the base commit is reachable from the commits in the starting list only 183 * via paths that include the base commit. 184 * 185 * This algorithm uses a conservation of mass approach to find the base commit. 186 * 187 * We start by injecting one unit of mass into the graph at each 188 * of the commits in the starting list. Injecting mass into a commit 189 * is achieved by adding to its pending mass counter and, if it is not already 190 * enqueued, enqueuing the commit in a list of pending commits, in latest 191 * commit date first order. 192 * 193 * The algorithm then preceeds to visit each commit in the pending queue. 194 * Upon each visit, the pending mass is added to the mass already seen for that 195 * commit and then divided into N equal portions, where N is the number of 196 * parents of the commit being visited. The divided portions are then injected 197 * into each of the parents. 198 * 199 * The algorithm continues until we discover a commit which has seen all the 200 * mass originally injected or until we run out of things to do. 201 * 202 * If we find a commit that has seen all the original mass, we have found 203 * the common base of all the commits in the starting list. 204 * 205 * The algorithm does _not_ depend on accurate timestamps for correct operation. 206 * However, reasonably sane (e.g. non-random) timestamps are required in order 207 * to prevent an exponential performance characteristic. The occasional 208 * timestamp inaccuracy will not dramatically affect performance but may 209 * result in more nodes being processed than strictly necessary. 210 * 211 * This procedure sets *boundary to the address of the base commit. It returns 212 * non-zero if, and only if, there was a problem parsing one of the 213 * commits discovered during the traversal. 214 */ 215static int find_base_for_list(struct commit_list *list, struct commit **boundary) 216{ 217 int ret = 0; 218 struct commit_list *cleaner = NULL; 219 struct commit_list *pending = NULL; 220 struct fraction injected; 221 init_fraction(&injected); 222 *boundary = NULL; 223 224 for (; list; list = list->next) { 225 struct commit *item = list->item; 226 227 if (item->object.util) { 228 die("%s:%d:%s: logic error: this should not have happened - commit %s", 229 __FILE__, __LINE__, __FUNCTION__, 230 sha1_to_hex(item->object.sha1)); 231 } 232 233 new_mass_counter(list->item, get_one()); 234 add(&injected, &injected, get_one()); 235 236 commit_list_insert(list->item, &cleaner); 237 commit_list_insert(list->item, &pending); 238 } 239 240 while (!*boundary && pending && !ret) { 241 struct commit *latest = pop_commit(&pending); 242 struct mass_counter *latest_node = (struct mass_counter *) latest->object.util; 243 int num_parents; 244 245 if ((ret = parse_commit(latest))) 246 continue; 247 add(&latest_node->seen, &latest_node->seen, &latest_node->pending); 248 249 num_parents = count_parents(latest); 250 if (num_parents) { 251 struct fraction distribution; 252 struct commit_list *parents; 253 254 divide(init_fraction(&distribution), &latest_node->pending, num_parents); 255 256 for (parents = latest->parents; parents; parents = parents->next) { 257 struct commit *parent = parents->item; 258 struct mass_counter *parent_node = (struct mass_counter *) parent->object.util; 259 260 if (!parent_node) { 261 parent_node = new_mass_counter(parent, &distribution); 262 insert_by_date(&pending, parent); 263 commit_list_insert(parent, &cleaner); 264 } else { 265 if (!compare(&parent_node->pending, get_zero())) 266 insert_by_date(&pending, parent); 267 add(&parent_node->pending, &parent_node->pending, &distribution); 268 } 269 } 270 271 clear_fraction(&distribution); 272 } 273 274 if (!compare(&latest_node->seen, &injected)) 275 *boundary = latest; 276 copy(&latest_node->pending, get_zero()); 277 } 278 279 while (cleaner) { 280 struct commit *next = pop_commit(&cleaner); 281 free_mass_counter((struct mass_counter *) next->object.util); 282 next->object.util = NULL; 283 } 284 285 if (pending) 286 free_commit_list(pending); 287 288 clear_fraction(&injected); 289 return ret; 290} 291 292 293/* 294 * Finds the base of an minimal, non-linear epoch, headed at head, by 295 * applying the find_base_for_list to a list consisting of the parents 296 */ 297static int find_base(struct commit *head, struct commit **boundary) 298{ 299 int ret = 0; 300 struct commit_list *pending = NULL; 301 struct commit_list *next; 302 303 for (next = head->parents; next; next = next->next) { 304 commit_list_insert(next->item, &pending); 305 } 306 ret = find_base_for_list(pending, boundary); 307 free_commit_list(pending); 308 309 return ret; 310} 311 312/* 313 * This procedure traverses to the boundary of the first epoch in the epoch 314 * sequence of the epoch headed at head_of_epoch. This is either the end of 315 * the maximal linear epoch or the base of a minimal non-linear epoch. 316 * 317 * The queue of pending nodes is sorted in reverse date order and each node 318 * is currently in the queue at most once. 319 */ 320static int find_next_epoch_boundary(struct commit *head_of_epoch, struct commit **boundary) 321{ 322 int ret; 323 struct commit *item = head_of_epoch; 324 325 ret = parse_commit(item); 326 if (ret) 327 return ret; 328 329 if (HAS_EXACTLY_ONE_PARENT(item)) { 330 /* 331 * We are at the start of a maximimal linear epoch. 332 * Traverse to the end. 333 */ 334 while (HAS_EXACTLY_ONE_PARENT(item) && !ret) { 335 item = item->parents->item; 336 ret = parse_commit(item); 337 } 338 *boundary = item; 339 340 } else { 341 /* 342 * Otherwise, we are at the start of a minimal, non-linear 343 * epoch - find the common base of all parents. 344 */ 345 ret = find_base(item, boundary); 346 } 347 348 return ret; 349} 350 351/* 352 * Returns non-zero if parent is known to be a parent of child. 353 */ 354static int is_parent_of(struct commit *parent, struct commit *child) 355{ 356 struct commit_list *parents; 357 for (parents = child->parents; parents; parents = parents->next) { 358 if (!memcmp(parent->object.sha1, parents->item->object.sha1, 359 sizeof(parents->item->object.sha1))) 360 return 1; 361 } 362 return 0; 363} 364 365/* 366 * Pushes an item onto the merge order stack. If the top of the stack is 367 * marked as being a possible "break", we check to see whether it actually 368 * is a break. 369 */ 370static void push_onto_merge_order_stack(struct commit_list **stack, struct commit *item) 371{ 372 struct commit_list *top = *stack; 373 if (top && (top->item->object.flags & DISCONTINUITY)) { 374 if (is_parent_of(top->item, item)) { 375 top->item->object.flags &= ~DISCONTINUITY; 376 } 377 } 378 commit_list_insert(item, stack); 379} 380 381/* 382 * Marks all interesting, visited commits reachable from this commit 383 * as uninteresting. We stop recursing when we reach the epoch boundary, 384 * an unvisited node or a node that has already been marking uninteresting. 385 * 386 * This doesn't actually mark all ancestors between the start node and the 387 * epoch boundary uninteresting, but does ensure that they will eventually 388 * be marked uninteresting when the main sort_first_epoch() traversal 389 * eventually reaches them. 390 */ 391static void mark_ancestors_uninteresting(struct commit *commit) 392{ 393 unsigned int flags = commit->object.flags; 394 int visited = flags & VISITED; 395 int boundary = flags & BOUNDARY; 396 int uninteresting = flags & UNINTERESTING; 397 struct commit_list *next; 398 399 commit->object.flags |= UNINTERESTING; 400 401 /* 402 * We only need to recurse if 403 * we are not on the boundary and 404 * we have not already been marked uninteresting and 405 * we have already been visited. 406 * 407 * The main sort_first_epoch traverse will mark unreachable 408 * all uninteresting, unvisited parents as they are visited 409 * so there is no need to duplicate that traversal here. 410 * 411 * Similarly, if we are already marked uninteresting 412 * then either all ancestors have already been marked 413 * uninteresting or will be once the sort_first_epoch 414 * traverse reaches them. 415 */ 416 417 if (uninteresting || boundary || !visited) 418 return; 419 420 for (next = commit->parents; next; next = next->next) 421 mark_ancestors_uninteresting(next->item); 422} 423 424/* 425 * Sorts the nodes of the first epoch of the epoch sequence of the epoch headed at head 426 * into merge order. 427 */ 428static void sort_first_epoch(struct commit *head, struct commit_list **stack) 429{ 430 struct commit_list *parents; 431 struct commit_list *reversed_parents = NULL; 432 433 head->object.flags |= VISITED; 434 435 /* 436 * parse_commit() builds the parent list in reverse order with respect 437 * to the order of the git-commit-tree arguments. So we need to reverse 438 * this list to output the oldest (or most "local") commits last. 439 */ 440 for (parents = head->parents; parents; parents = parents->next) 441 commit_list_insert(parents->item, &reversed_parents); 442 443 /* 444 * TODO: By sorting the parents in a different order, we can alter the 445 * merge order to show contemporaneous changes in parallel branches 446 * occurring after "local" changes. This is useful for a developer 447 * when a developer wants to see all changes that were incorporated 448 * into the same merge as her own changes occur after her own 449 * changes. 450 */ 451 452 while (reversed_parents) { 453 struct commit *parent = pop_commit(&reversed_parents); 454 455 if (head->object.flags & UNINTERESTING) { 456 /* 457 * Propagates the uninteresting bit to all parents. 458 * if we have already visited this parent, then 459 * the uninteresting bit will be propagated to each 460 * reachable commit that is still not marked 461 * uninteresting and won't otherwise be reached. 462 */ 463 mark_ancestors_uninteresting(parent); 464 } 465 466 if (!(parent->object.flags & VISITED)) { 467 if (parent->object.flags & BOUNDARY) { 468 if (*stack) { 469 die("something else is on the stack - %s", 470 sha1_to_hex((*stack)->item->object.sha1)); 471 } 472 push_onto_merge_order_stack(stack, parent); 473 parent->object.flags |= VISITED; 474 475 } else { 476 sort_first_epoch(parent, stack); 477 if (reversed_parents) { 478 /* 479 * This indicates a possible 480 * discontinuity it may not be be 481 * actual discontinuity if the head 482 * of parent N happens to be the tail 483 * of parent N+1. 484 * 485 * The next push onto the stack will 486 * resolve the question. 487 */ 488 (*stack)->item->object.flags |= DISCONTINUITY; 489 } 490 } 491 } 492 } 493 494 push_onto_merge_order_stack(stack, head); 495} 496 497/* 498 * Emit the contents of the stack. 499 * 500 * The stack is freed and replaced by NULL. 501 * 502 * Sets the return value to STOP if no further output should be generated. 503 */ 504static int emit_stack(struct commit_list **stack, emitter_func emitter) 505{ 506 unsigned int seen = 0; 507 int action = CONTINUE; 508 509 while (*stack && (action != STOP)) { 510 struct commit *next = pop_commit(stack); 511 seen |= next->object.flags; 512 if (*stack) 513 action = (*emitter) (next); 514 } 515 516 if (*stack) { 517 free_commit_list(*stack); 518 *stack = NULL; 519 } 520 521 return (action == STOP || (seen & UNINTERESTING)) ? STOP : CONTINUE; 522} 523 524/* 525 * Sorts an arbitrary epoch into merge order by sorting each epoch 526 * of its epoch sequence into order. 527 * 528 * Note: this algorithm currently leaves traces of its execution in the 529 * object flags of nodes it discovers. This should probably be fixed. 530 */ 531static int sort_in_merge_order(struct commit *head_of_epoch, emitter_func emitter) 532{ 533 struct commit *next = head_of_epoch; 534 int ret = 0; 535 int action = CONTINUE; 536 537 ret = parse_commit(head_of_epoch); 538 539 next->object.flags |= BOUNDARY; 540 541 while (next && next->parents && !ret && (action != STOP)) { 542 struct commit *base = NULL; 543 544 ret = find_next_epoch_boundary(next, &base); 545 if (ret) 546 return ret; 547 next->object.flags |= BOUNDARY; 548 if (base) 549 base->object.flags |= BOUNDARY; 550 551 if (HAS_EXACTLY_ONE_PARENT(next)) { 552 while (HAS_EXACTLY_ONE_PARENT(next) 553 && (action != STOP) 554 && !ret) { 555 if (next->object.flags & UNINTERESTING) { 556 action = STOP; 557 } else { 558 action = (*emitter) (next); 559 } 560 if (action != STOP) { 561 next = next->parents->item; 562 ret = parse_commit(next); 563 } 564 } 565 566 } else { 567 struct commit_list *stack = NULL; 568 sort_first_epoch(next, &stack); 569 action = emit_stack(&stack, emitter); 570 next = base; 571 } 572 } 573 574 if (next && (action != STOP) && !ret) { 575 (*emitter) (next); 576 } 577 578 return ret; 579} 580 581/* 582 * Sorts the nodes reachable from a starting list in merge order, we 583 * first find the base for the starting list and then sort all nodes 584 * in this subgraph using the sort_first_epoch algorithm. Once we have 585 * reached the base we can continue sorting using sort_in_merge_order. 586 */ 587int sort_list_in_merge_order(struct commit_list *list, emitter_func emitter) 588{ 589 struct commit_list *stack = NULL; 590 struct commit *base; 591 int ret = 0; 592 int action = CONTINUE; 593 struct commit_list *reversed = NULL; 594 595 for (; list; list = list->next) { 596 struct commit *next = list->item; 597 598 if (!(next->object.flags & UNINTERESTING)) { 599 if (next->object.flags & DUPCHECK) { 600 fprintf(stderr, "%s: duplicate commit %s ignored\n", 601 __FUNCTION__, sha1_to_hex(next->object.sha1)); 602 } else { 603 next->object.flags |= DUPCHECK; 604 commit_list_insert(list->item, &reversed); 605 } 606 } 607 } 608 609 if (!reversed) 610 return ret; 611 else if (!reversed->next) { 612 /* 613 * If there is only one element in the list, we can sort it 614 * using sort_in_merge_order. 615 */ 616 base = reversed->item; 617 } else { 618 /* 619 * Otherwise, we search for the base of the list. 620 */ 621 ret = find_base_for_list(reversed, &base); 622 if (ret) 623 return ret; 624 if (base) 625 base->object.flags |= BOUNDARY; 626 627 while (reversed) { 628 sort_first_epoch(pop_commit(&reversed), &stack); 629 if (reversed) { 630 /* 631 * If we have more commits to push, then the 632 * first push for the next parent may (or may 633 * not) represent a discontinuity with respect 634 * to the parent currently on the top of 635 * the stack. 636 * 637 * Mark it for checking here, and check it 638 * with the next push. See sort_first_epoch() 639 * for more details. 640 */ 641 stack->item->object.flags |= DISCONTINUITY; 642 } 643 } 644 645 action = emit_stack(&stack, emitter); 646 } 647 648 if (base && (action != STOP)) { 649 ret = sort_in_merge_order(base, emitter); 650 } 651 652 return ret; 653}