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{ 33if(!context) { 34 context =BN_CTX_new(); 35} 36return context; 37} 38 39static struct fraction *new_zero() 40{ 41struct fraction *result =xmalloc(sizeof(*result)); 42BN_init(&result->numerator); 43BN_init(&result->denominator); 44BN_zero(&result->numerator); 45BN_one(&result->denominator); 46return result; 47} 48 49static voidclear_fraction(struct fraction *fraction) 50{ 51BN_clear(&fraction->numerator); 52BN_clear(&fraction->denominator); 53} 54 55static struct fraction *divide(struct fraction *result,struct fraction *fraction,int divisor) 56{ 57 BIGNUM bn_divisor; 58 59BN_init(&bn_divisor); 60BN_set_word(&bn_divisor, divisor); 61 62BN_copy(&result->numerator, &fraction->numerator); 63BN_mul(&result->denominator, &fraction->denominator, &bn_divisor,get_BN_CTX()); 64 65BN_clear(&bn_divisor); 66return result; 67} 68 69static struct fraction *init_fraction(struct fraction *fraction) 70{ 71BN_init(&fraction->numerator); 72BN_init(&fraction->denominator); 73BN_zero(&fraction->numerator); 74BN_one(&fraction->denominator); 75return fraction; 76} 77 78static struct fraction *get_one() 79{ 80if(!one) { 81 one =new_zero(); 82BN_one(&one->numerator); 83} 84return one; 85} 86 87static struct fraction *get_zero() 88{ 89if(!zero) { 90 zero =new_zero(); 91} 92return zero; 93} 94 95static struct fraction *copy(struct fraction *to,struct fraction *from) 96{ 97BN_copy(&to->numerator, &from->numerator); 98BN_copy(&to->denominator, &from->denominator); 99return to; 100} 101 102static struct fraction *add(struct fraction *result,struct fraction *left,struct fraction *right) 103{ 104 BIGNUM a, b, gcd; 105 106BN_init(&a); 107BN_init(&b); 108BN_init(&gcd); 109 110BN_mul(&a, &left->numerator, &right->denominator,get_BN_CTX()); 111BN_mul(&b, &left->denominator, &right->numerator,get_BN_CTX()); 112BN_mul(&result->denominator, &left->denominator, &right->denominator,get_BN_CTX()); 113BN_add(&result->numerator, &a, &b); 114 115BN_gcd(&gcd, &result->denominator, &result->numerator,get_BN_CTX()); 116BN_div(&result->denominator, NULL, &result->denominator, &gcd,get_BN_CTX()); 117BN_div(&result->numerator, NULL, &result->numerator, &gcd,get_BN_CTX()); 118 119BN_clear(&a); 120BN_clear(&b); 121BN_clear(&gcd); 122 123return result; 124} 125 126static intcompare(struct fraction *left,struct fraction *right) 127{ 128 BIGNUM a, b; 129int result; 130 131BN_init(&a); 132BN_init(&b); 133 134BN_mul(&a, &left->numerator, &right->denominator,get_BN_CTX()); 135BN_mul(&b, &left->denominator, &right->numerator,get_BN_CTX()); 136 137 result =BN_cmp(&a, &b); 138 139BN_clear(&a); 140BN_clear(&b); 141 142return result; 143} 144 145struct mass_counter { 146struct fraction seen; 147struct fraction pending; 148}; 149 150static struct mass_counter *new_mass_counter(struct commit *commit,struct fraction *pending) 151{ 152struct mass_counter *mass_counter =xmalloc(sizeof(*mass_counter)); 153memset(mass_counter,0,sizeof(*mass_counter)); 154 155init_fraction(&mass_counter->seen); 156init_fraction(&mass_counter->pending); 157 158copy(&mass_counter->pending, pending); 159copy(&mass_counter->seen,get_zero()); 160 161if(commit->object.util) { 162die("multiple attempts to initialize mass counter for%s", 163sha1_to_hex(commit->object.sha1)); 164} 165 166 commit->object.util = mass_counter; 167 168return mass_counter; 169} 170 171static voidfree_mass_counter(struct mass_counter *counter) 172{ 173clear_fraction(&counter->seen); 174clear_fraction(&counter->pending); 175free(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 intfind_base_for_list(struct commit_list *list,struct commit **boundary) 216{ 217int ret =0; 218struct commit_list *cleaner = NULL; 219struct commit_list *pending = NULL; 220struct fraction injected; 221init_fraction(&injected); 222*boundary = NULL; 223 224for(; list; list = list->next) { 225struct commit *item = list->item; 226 227if(!item->object.util) { 228new_mass_counter(list->item,get_one()); 229add(&injected, &injected,get_one()); 230 231commit_list_insert(list->item, &cleaner); 232commit_list_insert(list->item, &pending); 233} 234} 235 236while(!*boundary && pending && !ret) { 237struct commit *latest =pop_commit(&pending); 238struct mass_counter *latest_node = (struct mass_counter *) latest->object.util; 239int num_parents; 240 241if((ret =parse_commit(latest))) 242continue; 243add(&latest_node->seen, &latest_node->seen, &latest_node->pending); 244 245 num_parents =count_parents(latest); 246if(num_parents) { 247struct fraction distribution; 248struct commit_list *parents; 249 250divide(init_fraction(&distribution), &latest_node->pending, num_parents); 251 252for(parents = latest->parents; parents; parents = parents->next) { 253struct commit *parent = parents->item; 254struct mass_counter *parent_node = (struct mass_counter *) parent->object.util; 255 256if(!parent_node) { 257 parent_node =new_mass_counter(parent, &distribution); 258insert_by_date(&pending, parent); 259commit_list_insert(parent, &cleaner); 260}else{ 261if(!compare(&parent_node->pending,get_zero())) 262insert_by_date(&pending, parent); 263add(&parent_node->pending, &parent_node->pending, &distribution); 264} 265} 266 267clear_fraction(&distribution); 268} 269 270if(!compare(&latest_node->seen, &injected)) 271*boundary = latest; 272copy(&latest_node->pending,get_zero()); 273} 274 275while(cleaner) { 276struct commit *next =pop_commit(&cleaner); 277free_mass_counter((struct mass_counter *) next->object.util); 278 next->object.util = NULL; 279} 280 281if(pending) 282free_commit_list(pending); 283 284clear_fraction(&injected); 285return ret; 286} 287 288 289/* 290 * Finds the base of an minimal, non-linear epoch, headed at head, by 291 * applying the find_base_for_list to a list consisting of the parents 292 */ 293static intfind_base(struct commit *head,struct commit **boundary) 294{ 295int ret =0; 296struct commit_list *pending = NULL; 297struct commit_list *next; 298 299for(next = head->parents; next; next = next->next) { 300commit_list_insert(next->item, &pending); 301} 302 ret =find_base_for_list(pending, boundary); 303free_commit_list(pending); 304 305return ret; 306} 307 308/* 309 * This procedure traverses to the boundary of the first epoch in the epoch 310 * sequence of the epoch headed at head_of_epoch. This is either the end of 311 * the maximal linear epoch or the base of a minimal non-linear epoch. 312 * 313 * The queue of pending nodes is sorted in reverse date order and each node 314 * is currently in the queue at most once. 315 */ 316static intfind_next_epoch_boundary(struct commit *head_of_epoch,struct commit **boundary) 317{ 318int ret; 319struct commit *item = head_of_epoch; 320 321 ret =parse_commit(item); 322if(ret) 323return ret; 324 325if(HAS_EXACTLY_ONE_PARENT(item)) { 326/* 327 * We are at the start of a maximimal linear epoch. 328 * Traverse to the end. 329 */ 330while(HAS_EXACTLY_ONE_PARENT(item) && !ret) { 331 item = item->parents->item; 332 ret =parse_commit(item); 333} 334*boundary = item; 335 336}else{ 337/* 338 * Otherwise, we are at the start of a minimal, non-linear 339 * epoch - find the common base of all parents. 340 */ 341 ret =find_base(item, boundary); 342} 343 344return ret; 345} 346 347/* 348 * Returns non-zero if parent is known to be a parent of child. 349 */ 350static intis_parent_of(struct commit *parent,struct commit *child) 351{ 352struct commit_list *parents; 353for(parents = child->parents; parents; parents = parents->next) { 354if(!memcmp(parent->object.sha1, parents->item->object.sha1, 355sizeof(parents->item->object.sha1))) 356return1; 357} 358return0; 359} 360 361/* 362 * Pushes an item onto the merge order stack. If the top of the stack is 363 * marked as being a possible "break", we check to see whether it actually 364 * is a break. 365 */ 366static voidpush_onto_merge_order_stack(struct commit_list **stack,struct commit *item) 367{ 368struct commit_list *top = *stack; 369if(top && (top->item->object.flags & DISCONTINUITY)) { 370if(is_parent_of(top->item, item)) { 371 top->item->object.flags &= ~DISCONTINUITY; 372} 373} 374commit_list_insert(item, stack); 375} 376 377/* 378 * Marks all interesting, visited commits reachable from this commit 379 * as uninteresting. We stop recursing when we reach the epoch boundary, 380 * an unvisited node or a node that has already been marking uninteresting. 381 * 382 * This doesn't actually mark all ancestors between the start node and the 383 * epoch boundary uninteresting, but does ensure that they will eventually 384 * be marked uninteresting when the main sort_first_epoch() traversal 385 * eventually reaches them. 386 */ 387static voidmark_ancestors_uninteresting(struct commit *commit) 388{ 389unsigned int flags = commit->object.flags; 390int visited = flags & VISITED; 391int boundary = flags & BOUNDARY; 392int uninteresting = flags & UNINTERESTING; 393struct commit_list *next; 394 395 commit->object.flags |= UNINTERESTING; 396 397/* 398 * We only need to recurse if 399 * we are not on the boundary and 400 * we have not already been marked uninteresting and 401 * we have already been visited. 402 * 403 * The main sort_first_epoch traverse will mark unreachable 404 * all uninteresting, unvisited parents as they are visited 405 * so there is no need to duplicate that traversal here. 406 * 407 * Similarly, if we are already marked uninteresting 408 * then either all ancestors have already been marked 409 * uninteresting or will be once the sort_first_epoch 410 * traverse reaches them. 411 */ 412 413if(uninteresting || boundary || !visited) 414return; 415 416for(next = commit->parents; next; next = next->next) 417mark_ancestors_uninteresting(next->item); 418} 419 420/* 421 * Sorts the nodes of the first epoch of the epoch sequence of the epoch headed at head 422 * into merge order. 423 */ 424static voidsort_first_epoch(struct commit *head,struct commit_list **stack) 425{ 426struct commit_list *parents; 427 428 head->object.flags |= VISITED; 429 430/* 431 * TODO: By sorting the parents in a different order, we can alter the 432 * merge order to show contemporaneous changes in parallel branches 433 * occurring after "local" changes. This is useful for a developer 434 * when a developer wants to see all changes that were incorporated 435 * into the same merge as her own changes occur after her own 436 * changes. 437 */ 438 439for(parents = head->parents; parents; parents = parents->next) { 440struct commit *parent = parents->item; 441 442if(head->object.flags & UNINTERESTING) { 443/* 444 * Propagates the uninteresting bit to all parents. 445 * if we have already visited this parent, then 446 * the uninteresting bit will be propagated to each 447 * reachable commit that is still not marked 448 * uninteresting and won't otherwise be reached. 449 */ 450mark_ancestors_uninteresting(parent); 451} 452 453if(!(parent->object.flags & VISITED)) { 454if(parent->object.flags & BOUNDARY) { 455if(*stack) { 456die("something else is on the stack -%s", 457sha1_to_hex((*stack)->item->object.sha1)); 458} 459push_onto_merge_order_stack(stack, parent); 460 parent->object.flags |= VISITED; 461 462}else{ 463sort_first_epoch(parent, stack); 464if(parents) { 465/* 466 * This indicates a possible 467 * discontinuity it may not be be 468 * actual discontinuity if the head 469 * of parent N happens to be the tail 470 * of parent N+1. 471 * 472 * The next push onto the stack will 473 * resolve the question. 474 */ 475(*stack)->item->object.flags |= DISCONTINUITY; 476} 477} 478} 479} 480 481push_onto_merge_order_stack(stack, head); 482} 483 484/* 485 * Emit the contents of the stack. 486 * 487 * The stack is freed and replaced by NULL. 488 * 489 * Sets the return value to STOP if no further output should be generated. 490 */ 491static intemit_stack(struct commit_list **stack, emitter_func emitter,int include_last) 492{ 493unsigned int seen =0; 494int action = CONTINUE; 495 496while(*stack && (action != STOP)) { 497struct commit *next =pop_commit(stack); 498 seen |= next->object.flags; 499if(*stack || include_last) { 500if(!*stack) 501 next->object.flags |= BOUNDARY; 502 action = (*emitter) (next); 503} 504} 505 506if(*stack) { 507free_commit_list(*stack); 508*stack = NULL; 509} 510 511return(action == STOP || (seen & UNINTERESTING)) ? STOP : CONTINUE; 512} 513 514/* 515 * Sorts an arbitrary epoch into merge order by sorting each epoch 516 * of its epoch sequence into order. 517 * 518 * Note: this algorithm currently leaves traces of its execution in the 519 * object flags of nodes it discovers. This should probably be fixed. 520 */ 521static intsort_in_merge_order(struct commit *head_of_epoch, emitter_func emitter) 522{ 523struct commit *next = head_of_epoch; 524int ret =0; 525int action = CONTINUE; 526 527 ret =parse_commit(head_of_epoch); 528 529 next->object.flags |= BOUNDARY; 530 531while(next && next->parents && !ret && (action != STOP)) { 532struct commit *base = NULL; 533 534 ret =find_next_epoch_boundary(next, &base); 535if(ret) 536return ret; 537 next->object.flags |= BOUNDARY; 538if(base) 539 base->object.flags |= BOUNDARY; 540 541if(HAS_EXACTLY_ONE_PARENT(next)) { 542while(HAS_EXACTLY_ONE_PARENT(next) 543&& (action != STOP) 544&& !ret) { 545if(next->object.flags & UNINTERESTING) { 546 action = STOP; 547}else{ 548 action = (*emitter) (next); 549} 550if(action != STOP) { 551 next = next->parents->item; 552 ret =parse_commit(next); 553} 554} 555 556}else{ 557struct commit_list *stack = NULL; 558sort_first_epoch(next, &stack); 559 action =emit_stack(&stack, emitter, (base == NULL)); 560 next = base; 561} 562} 563 564if(next && (action != STOP) && !ret) { 565(*emitter) (next); 566} 567 568return ret; 569} 570 571/* 572 * Sorts the nodes reachable from a starting list in merge order, we 573 * first find the base for the starting list and then sort all nodes 574 * in this subgraph using the sort_first_epoch algorithm. Once we have 575 * reached the base we can continue sorting using sort_in_merge_order. 576 */ 577intsort_list_in_merge_order(struct commit_list *list, emitter_func emitter) 578{ 579struct commit_list *stack = NULL; 580struct commit *base; 581int ret =0; 582int action = CONTINUE; 583struct commit_list *reversed = NULL; 584 585for(; list; list = list->next) { 586struct commit *next = list->item; 587 588if(!(next->object.flags & DUPCHECK)) { 589 next->object.flags |= DUPCHECK; 590commit_list_insert(list->item, &reversed); 591} 592} 593 594if(!reversed) 595return ret; 596else if(!reversed->next) { 597/* 598 * If there is only one element in the list, we can sort it 599 * using sort_in_merge_order. 600 */ 601 base = reversed->item; 602}else{ 603/* 604 * Otherwise, we search for the base of the list. 605 */ 606 ret =find_base_for_list(reversed, &base); 607if(ret) 608return ret; 609if(base) 610 base->object.flags |= BOUNDARY; 611 612while(reversed) { 613struct commit * next =pop_commit(&reversed); 614 615if(!(next->object.flags & VISITED)) { 616sort_first_epoch(next, &stack); 617if(reversed) { 618/* 619 * If we have more commits 620 * to push, then the first 621 * push for the next parent may 622 * (or may * not) represent a 623 * discontinuity with respect 624 * to the parent currently on 625 * the top of the stack. 626 * 627 * Mark it for checking here, 628 * and check it with the next 629 * push. See sort_first_epoch() 630 * for more details. 631 */ 632 stack->item->object.flags |= DISCONTINUITY; 633} 634} 635} 636 637 action =emit_stack(&stack, emitter, (base==NULL)); 638} 639 640if(base && (action != STOP)) { 641 ret =sort_in_merge_order(base, emitter); 642} 643 644return ret; 645}