1#include "cache.h" 2#include "split-index.h" 3#include "ewah/ewok.h" 4 5struct split_index *init_split_index(struct index_state *istate) 6{ 7 if (!istate->split_index) { 8 istate->split_index = xcalloc(1, sizeof(*istate->split_index)); 9 istate->split_index->refcount = 1; 10 } 11 return istate->split_index; 12} 13 14int read_link_extension(struct index_state *istate, 15 const void *data_, unsigned long sz) 16{ 17 const unsigned char *data = data_; 18 struct split_index *si; 19 int ret; 20 21 if (sz < the_hash_algo->rawsz) 22 return error("corrupt link extension (too short)"); 23 si = init_split_index(istate); 24 hashcpy(si->base_oid.hash, data); 25 data += the_hash_algo->rawsz; 26 sz -= the_hash_algo->rawsz; 27 if (!sz) 28 return 0; 29 si->delete_bitmap = ewah_new(); 30 ret = ewah_read_mmap(si->delete_bitmap, data, sz); 31 if (ret < 0) 32 return error("corrupt delete bitmap in link extension"); 33 data += ret; 34 sz -= ret; 35 si->replace_bitmap = ewah_new(); 36 ret = ewah_read_mmap(si->replace_bitmap, data, sz); 37 if (ret < 0) 38 return error("corrupt replace bitmap in link extension"); 39 if (ret != sz) 40 return error("garbage at the end of link extension"); 41 return 0; 42} 43 44int write_link_extension(struct strbuf *sb, 45 struct index_state *istate) 46{ 47 struct split_index *si = istate->split_index; 48 strbuf_add(sb, si->base_oid.hash, the_hash_algo->rawsz); 49 if (!si->delete_bitmap && !si->replace_bitmap) 50 return 0; 51 ewah_serialize_strbuf(si->delete_bitmap, sb); 52 ewah_serialize_strbuf(si->replace_bitmap, sb); 53 return 0; 54} 55 56static void mark_base_index_entries(struct index_state *base) 57{ 58 int i; 59 /* 60 * To keep track of the shared entries between 61 * istate->base->cache[] and istate->cache[], base entry 62 * position is stored in each base entry. All positions start 63 * from 1 instead of 0, which is reserved to say "this is a new 64 * entry". 65 */ 66 for (i = 0; i < base->cache_nr; i++) 67 base->cache[i]->index = i + 1; 68} 69 70void move_cache_to_base_index(struct index_state *istate) 71{ 72 struct split_index *si = istate->split_index; 73 int i; 74 75 /* 76 * If there was a previous base index, then transfer ownership of allocated 77 * entries to the parent index. 78 */ 79 if (si->base && 80 si->base->ce_mem_pool) { 81 82 if (!istate->ce_mem_pool) 83 mem_pool_init(&istate->ce_mem_pool, 0); 84 85 mem_pool_combine(istate->ce_mem_pool, istate->split_index->base->ce_mem_pool); 86 } 87 88 si->base = xcalloc(1, sizeof(*si->base)); 89 si->base->version = istate->version; 90 /* zero timestamp disables racy test in ce_write_index() */ 91 si->base->timestamp = istate->timestamp; 92 ALLOC_GROW(si->base->cache, istate->cache_nr, si->base->cache_alloc); 93 si->base->cache_nr = istate->cache_nr; 94 95 /* 96 * The mem_pool needs to move with the allocated entries. 97 */ 98 si->base->ce_mem_pool = istate->ce_mem_pool; 99 istate->ce_mem_pool = NULL; 100 101 COPY_ARRAY(si->base->cache, istate->cache, istate->cache_nr); 102 mark_base_index_entries(si->base); 103 for (i = 0; i < si->base->cache_nr; i++) 104 si->base->cache[i]->ce_flags &= ~CE_UPDATE_IN_BASE; 105} 106 107static void mark_entry_for_delete(size_t pos, void *data) 108{ 109 struct index_state *istate = data; 110 if (pos >= istate->cache_nr) 111 die("position for delete %d exceeds base index size %d", 112 (int)pos, istate->cache_nr); 113 istate->cache[pos]->ce_flags |= CE_REMOVE; 114 istate->split_index->nr_deletions++; 115} 116 117static void replace_entry(size_t pos, void *data) 118{ 119 struct index_state *istate = data; 120 struct split_index *si = istate->split_index; 121 struct cache_entry *dst, *src; 122 123 if (pos >= istate->cache_nr) 124 die("position for replacement %d exceeds base index size %d", 125 (int)pos, istate->cache_nr); 126 if (si->nr_replacements >= si->saved_cache_nr) 127 die("too many replacements (%d vs %d)", 128 si->nr_replacements, si->saved_cache_nr); 129 dst = istate->cache[pos]; 130 if (dst->ce_flags & CE_REMOVE) 131 die("entry %d is marked as both replaced and deleted", 132 (int)pos); 133 src = si->saved_cache[si->nr_replacements]; 134 if (ce_namelen(src)) 135 die("corrupt link extension, entry %d should have " 136 "zero length name", (int)pos); 137 src->index = pos + 1; 138 src->ce_flags |= CE_UPDATE_IN_BASE; 139 src->ce_namelen = dst->ce_namelen; 140 copy_cache_entry(dst, src); 141 discard_cache_entry(src); 142 si->nr_replacements++; 143} 144 145void merge_base_index(struct index_state *istate) 146{ 147 struct split_index *si = istate->split_index; 148 unsigned int i; 149 150 mark_base_index_entries(si->base); 151 152 si->saved_cache = istate->cache; 153 si->saved_cache_nr = istate->cache_nr; 154 istate->cache_nr = si->base->cache_nr; 155 istate->cache = NULL; 156 istate->cache_alloc = 0; 157 ALLOC_GROW(istate->cache, istate->cache_nr, istate->cache_alloc); 158 COPY_ARRAY(istate->cache, si->base->cache, istate->cache_nr); 159 160 si->nr_deletions = 0; 161 si->nr_replacements = 0; 162 ewah_each_bit(si->replace_bitmap, replace_entry, istate); 163 ewah_each_bit(si->delete_bitmap, mark_entry_for_delete, istate); 164 if (si->nr_deletions) 165 remove_marked_cache_entries(istate); 166 167 for (i = si->nr_replacements; i < si->saved_cache_nr; i++) { 168 if (!ce_namelen(si->saved_cache[i])) 169 die("corrupt link extension, entry %d should " 170 "have non-zero length name", i); 171 add_index_entry(istate, si->saved_cache[i], 172 ADD_CACHE_OK_TO_ADD | 173 ADD_CACHE_KEEP_CACHE_TREE | 174 /* 175 * we may have to replay what 176 * merge-recursive.c:update_stages() 177 * does, which has this flag on 178 */ 179 ADD_CACHE_SKIP_DFCHECK); 180 si->saved_cache[i] = NULL; 181 } 182 183 ewah_free(si->delete_bitmap); 184 ewah_free(si->replace_bitmap); 185 FREE_AND_NULL(si->saved_cache); 186 si->delete_bitmap = NULL; 187 si->replace_bitmap = NULL; 188 si->saved_cache_nr = 0; 189} 190 191/* 192 * Compare most of the fields in two cache entries, i.e. all except the 193 * hashmap_entry and the name. 194 */ 195static int compare_ce_content(struct cache_entry *a, struct cache_entry *b) 196{ 197 const unsigned int ondisk_flags = CE_STAGEMASK | CE_VALID | 198 CE_EXTENDED_FLAGS; 199 unsigned int ce_flags = a->ce_flags; 200 unsigned int base_flags = b->ce_flags; 201 int ret; 202 203 /* only on-disk flags matter */ 204 a->ce_flags &= ondisk_flags; 205 b->ce_flags &= ondisk_flags; 206 ret = memcmp(&a->ce_stat_data, &b->ce_stat_data, 207 offsetof(struct cache_entry, name) - 208 offsetof(struct cache_entry, ce_stat_data)); 209 a->ce_flags = ce_flags; 210 b->ce_flags = base_flags; 211 212 return ret; 213} 214 215void prepare_to_write_split_index(struct index_state *istate) 216{ 217 struct split_index *si = init_split_index(istate); 218 struct cache_entry **entries = NULL, *ce; 219 int i, nr_entries = 0, nr_alloc = 0; 220 221 si->delete_bitmap = ewah_new(); 222 si->replace_bitmap = ewah_new(); 223 224 if (si->base) { 225 /* Go through istate->cache[] and mark CE_MATCHED to 226 * entry with positive index. We'll go through 227 * base->cache[] later to delete all entries in base 228 * that are not marked with either CE_MATCHED or 229 * CE_UPDATE_IN_BASE. If istate->cache[i] is a 230 * duplicate, deduplicate it. 231 */ 232 for (i = 0; i < istate->cache_nr; i++) { 233 struct cache_entry *base; 234 ce = istate->cache[i]; 235 if (!ce->index) { 236 /* 237 * During simple update index operations this 238 * is a cache entry that is not present in 239 * the shared index. It will be added to the 240 * split index. 241 * 242 * However, it might also represent a file 243 * that already has a cache entry in the 244 * shared index, but a new index has just 245 * been constructed by unpack_trees(), and 246 * this entry now refers to different content 247 * than what was recorded in the original 248 * index, e.g. during 'read-tree -m HEAD^' or 249 * 'checkout HEAD^'. In this case the 250 * original entry in the shared index will be 251 * marked as deleted, and this entry will be 252 * added to the split index. 253 */ 254 continue; 255 } 256 if (ce->index > si->base->cache_nr) { 257 BUG("ce refers to a shared ce at %d, which is beyond the shared index size %d", 258 ce->index, si->base->cache_nr); 259 } 260 ce->ce_flags |= CE_MATCHED; /* or "shared" */ 261 base = si->base->cache[ce->index - 1]; 262 if (ce == base) { 263 /* The entry is present in the shared index. */ 264 if (ce->ce_flags & CE_UPDATE_IN_BASE) { 265 /* 266 * Already marked for inclusion in 267 * the split index, either because 268 * the corresponding file was 269 * modified and the cached stat data 270 * was refreshed, or because there 271 * is already a replacement entry in 272 * the split index. 273 * Nothing more to do here. 274 */ 275 } else if (!ce_uptodate(ce) && 276 is_racy_timestamp(istate, ce)) { 277 /* 278 * A racily clean cache entry stored 279 * only in the shared index: it must 280 * be added to the split index, so 281 * the subsequent do_write_index() 282 * can smudge its stat data. 283 */ 284 ce->ce_flags |= CE_UPDATE_IN_BASE; 285 } else { 286 /* 287 * The entry is only present in the 288 * shared index and it was not 289 * refreshed. 290 * Just leave it there. 291 */ 292 } 293 continue; 294 } 295 if (ce->ce_namelen != base->ce_namelen || 296 strcmp(ce->name, base->name)) { 297 ce->index = 0; 298 continue; 299 } 300 /* 301 * This is the copy of a cache entry that is present 302 * in the shared index, created by unpack_trees() 303 * while it constructed a new index. 304 */ 305 if (ce->ce_flags & CE_UPDATE_IN_BASE) { 306 /* 307 * Already marked for inclusion in the split 308 * index, either because the corresponding 309 * file was modified and the cached stat data 310 * was refreshed, or because the original 311 * entry already had a replacement entry in 312 * the split index. 313 * Nothing to do. 314 */ 315 } else if (!ce_uptodate(ce) && 316 is_racy_timestamp(istate, ce)) { 317 /* 318 * A copy of a racily clean cache entry from 319 * the shared index. It must be added to 320 * the split index, so the subsequent 321 * do_write_index() can smudge its stat data. 322 */ 323 ce->ce_flags |= CE_UPDATE_IN_BASE; 324 } else { 325 /* 326 * Thoroughly compare the cached data to see 327 * whether it should be marked for inclusion 328 * in the split index. 329 * 330 * This comparison might be unnecessary, as 331 * code paths modifying the cached data do 332 * set CE_UPDATE_IN_BASE as well. 333 */ 334 if (compare_ce_content(ce, base)) 335 ce->ce_flags |= CE_UPDATE_IN_BASE; 336 } 337 discard_cache_entry(base); 338 si->base->cache[ce->index - 1] = ce; 339 } 340 for (i = 0; i < si->base->cache_nr; i++) { 341 ce = si->base->cache[i]; 342 if ((ce->ce_flags & CE_REMOVE) || 343 !(ce->ce_flags & CE_MATCHED)) 344 ewah_set(si->delete_bitmap, i); 345 else if (ce->ce_flags & CE_UPDATE_IN_BASE) { 346 ewah_set(si->replace_bitmap, i); 347 ce->ce_flags |= CE_STRIP_NAME; 348 ALLOC_GROW(entries, nr_entries+1, nr_alloc); 349 entries[nr_entries++] = ce; 350 } 351 if (is_null_oid(&ce->oid)) 352 istate->drop_cache_tree = 1; 353 } 354 } 355 356 for (i = 0; i < istate->cache_nr; i++) { 357 ce = istate->cache[i]; 358 if ((!si->base || !ce->index) && !(ce->ce_flags & CE_REMOVE)) { 359 assert(!(ce->ce_flags & CE_STRIP_NAME)); 360 ALLOC_GROW(entries, nr_entries+1, nr_alloc); 361 entries[nr_entries++] = ce; 362 } 363 ce->ce_flags &= ~CE_MATCHED; 364 } 365 366 /* 367 * take cache[] out temporarily, put entries[] in its place 368 * for writing 369 */ 370 si->saved_cache = istate->cache; 371 si->saved_cache_nr = istate->cache_nr; 372 istate->cache = entries; 373 istate->cache_nr = nr_entries; 374} 375 376void finish_writing_split_index(struct index_state *istate) 377{ 378 struct split_index *si = init_split_index(istate); 379 380 ewah_free(si->delete_bitmap); 381 ewah_free(si->replace_bitmap); 382 si->delete_bitmap = NULL; 383 si->replace_bitmap = NULL; 384 free(istate->cache); 385 istate->cache = si->saved_cache; 386 istate->cache_nr = si->saved_cache_nr; 387} 388 389void discard_split_index(struct index_state *istate) 390{ 391 struct split_index *si = istate->split_index; 392 if (!si) 393 return; 394 istate->split_index = NULL; 395 si->refcount--; 396 if (si->refcount) 397 return; 398 if (si->base) { 399 discard_index(si->base); 400 free(si->base); 401 } 402 free(si); 403} 404 405void save_or_free_index_entry(struct index_state *istate, struct cache_entry *ce) 406{ 407 if (ce->index && 408 istate->split_index && 409 istate->split_index->base && 410 ce->index <= istate->split_index->base->cache_nr && 411 ce == istate->split_index->base->cache[ce->index - 1]) 412 ce->ce_flags |= CE_REMOVE; 413 else 414 discard_cache_entry(ce); 415} 416 417void replace_index_entry_in_base(struct index_state *istate, 418 struct cache_entry *old_entry, 419 struct cache_entry *new_entry) 420{ 421 if (old_entry->index && 422 istate->split_index && 423 istate->split_index->base && 424 old_entry->index <= istate->split_index->base->cache_nr) { 425 new_entry->index = old_entry->index; 426 if (old_entry != istate->split_index->base->cache[new_entry->index - 1]) 427 discard_cache_entry(istate->split_index->base->cache[new_entry->index - 1]); 428 istate->split_index->base->cache[new_entry->index - 1] = new_entry; 429 } 430} 431 432void add_split_index(struct index_state *istate) 433{ 434 if (!istate->split_index) { 435 init_split_index(istate); 436 istate->cache_changed |= SPLIT_INDEX_ORDERED; 437 } 438} 439 440void remove_split_index(struct index_state *istate) 441{ 442 if (istate->split_index) { 443 /* 444 * When removing the split index, we need to move 445 * ownership of the mem_pool associated with the 446 * base index to the main index. There may be cache entries 447 * allocated from the base's memory pool that are shared with 448 * the_index.cache[]. 449 */ 450 mem_pool_combine(istate->ce_mem_pool, istate->split_index->base->ce_mem_pool); 451 452 /* 453 * The split index no longer owns the mem_pool backing 454 * its cache array. As we are discarding this index, 455 * mark the index as having no cache entries, so it 456 * will not attempt to clean up the cache entries or 457 * validate them. 458 */ 459 if (istate->split_index->base) 460 istate->split_index->base->cache_nr = 0; 461 462 /* 463 * We can discard the split index because its 464 * memory pool has been incorporated into the 465 * memory pool associated with the the_index. 466 */ 467 discard_split_index(istate); 468 469 istate->cache_changed |= SOMETHING_CHANGED; 470 } 471}