/* * Called when we encounter an overflow or big key/data page during split * handling. This is special cased since we have to begin checking whether * the key/data pairs fit on their respective pages and because we may need * overflow pages for both the old and new pages. * * The first page might be a page with regular key/data pairs in which case * we have a regular overflow condition and just need to go on to the next * page or it might be a big key/data pair in which case we need to fix the * big key/data pair. * * Returns: * 0 ==> success * -1 ==> failure */ static int ugly_split( HTAB *hashp, uint32_t obucket, /* Same as __split_page. */ BUFHEAD *old_bufp, BUFHEAD *new_bufp, int copyto, /* First byte on page which contains key/data values. */ int moved /* Number of pairs moved to new page. */ ) { BUFHEAD *bufp; /* Buffer header for ino */ uint16_t *ino; /* Page keys come off of */ uint16_t *np; /* New page */ uint16_t *op; /* Page keys go on to if they aren't moving */ size_t temp; BUFHEAD *last_bfp; /* Last buf header OVFL needing to be freed */ DBT key, val; SPLIT_RETURN ret; uint16_t n, off, ov_addr, scopyto; char *cino; /* Character value of ino */ bufp = old_bufp; ino = (uint16_t *)(void *)old_bufp->page; np = (uint16_t *)(void *)new_bufp->page; op = (uint16_t *)(void *)old_bufp->page; last_bfp = NULL; scopyto = (uint16_t)copyto; /* ANSI */ n = ino[0] - 1; while (n < ino[0]) { if (ino[2] < REAL_KEY && ino[2] != OVFLPAGE) { if (__big_split(hashp, old_bufp, new_bufp, bufp, (int)bufp->addr, obucket, &ret)) return (-1); old_bufp = ret.oldp; if (!old_bufp) return (-1); op = (uint16_t *)(void *)old_bufp->page; new_bufp = ret.newp; if (!new_bufp) return (-1); np = (uint16_t *)(void *)new_bufp->page; bufp = ret.nextp; if (!bufp) return (0); cino = (char *)bufp->page; ino = (uint16_t *)(void *)cino; last_bfp = ret.nextp; } else if (ino[n + 1] == OVFLPAGE) { ov_addr = ino[n]; /* * Fix up the old page -- the extra 2 are the fields * which contained the overflow information. */ ino[0] -= (moved + 2); temp = sizeof(uint16_t) * (ino[0] + 3); _DIAGASSERT(scopyto >= temp); FREESPACE(ino) = (uint16_t)(scopyto - temp); OFFSET(ino) = scopyto; bufp = __get_buf(hashp, (uint32_t)ov_addr, bufp, 0); if (!bufp) return (-1); ino = (uint16_t *)(void *)bufp->page; n = 1; scopyto = hashp->BSIZE; moved = 0; if (last_bfp) __free_ovflpage(hashp, last_bfp); last_bfp = bufp; } /* Move regular sized pairs of there are any */ off = hashp->BSIZE; for (n = 1; (n < ino[0]) && (ino[n + 1] >= REAL_KEY); n += 2) { cino = (char *)(void *)ino; key.data = (uint8_t *)cino + ino[n]; key.size = off - ino[n]; val.data = (uint8_t *)cino + ino[n + 1]; val.size = ino[n] - ino[n + 1]; off = ino[n + 1]; if (__call_hash(hashp, key.data, (int)key.size) == obucket) { /* Keep on old page */ if (PAIRFITS(op, (&key), (&val))) putpair((char *)(void *)op, &key, &val); else { old_bufp = __add_ovflpage(hashp, old_bufp); if (!old_bufp) return (-1); op = (uint16_t *)(void *)old_bufp->page; putpair((char *)(void *)op, &key, &val); } old_bufp->flags |= BUF_MOD; } else { /* Move to new page */ if (PAIRFITS(np, (&key), (&val))) putpair((char *)(void *)np, &key, &val); else { new_bufp = __add_ovflpage(hashp, new_bufp); if (!new_bufp) return (-1); np = (uint16_t *)(void *)new_bufp->page; putpair((char *)(void *)np, &key, &val); } new_bufp->flags |= BUF_MOD; } } } if (last_bfp) __free_ovflpage(hashp, last_bfp); return (0); }
/* * Assume that hashp has been set in wrapper routine. */ static int hash_access(HTAB *hashp, ACTION action, DBT *key, DBT *val) { BUFHEAD *rbufp; BUFHEAD *bufp, *save_bufp; u_int16_t *bp; int n, ndx, off, size; char *kp; u_int16_t pageno; #ifdef HASH_STATISTICS hash_accesses++; #endif off = hashp->BSIZE; size = key->size; kp = (char *)key->data; rbufp = __get_buf(hashp, __call_hash(hashp, kp, size), NULL, 0); if (!rbufp) return (ERROR); save_bufp = rbufp; /* Pin the bucket chain */ rbufp->flags |= BUF_PIN; for (bp = (u_int16_t *)rbufp->page, n = *bp++, ndx = 1; ndx < n;) if (bp[1] >= REAL_KEY) { /* Real key/data pair */ if (size == off - *bp && memcmp(kp, rbufp->page + *bp, size) == 0) goto found; off = bp[1]; #ifdef HASH_STATISTICS hash_collisions++; #endif bp += 2; ndx += 2; } else if (bp[1] == OVFLPAGE) { rbufp = __get_buf(hashp, *bp, rbufp, 0); if (!rbufp) { save_bufp->flags &= ~BUF_PIN; return (ERROR); } /* FOR LOOP INIT */ bp = (u_int16_t *)rbufp->page; n = *bp++; ndx = 1; off = hashp->BSIZE; } else if (bp[1] < REAL_KEY) { if ((ndx = __find_bigpair(hashp, rbufp, ndx, kp, size)) > 0) goto found; if (ndx == -2) { bufp = rbufp; if (!(pageno = __find_last_page(hashp, &bufp))) { ndx = 0; rbufp = bufp; break; /* FOR */ } rbufp = __get_buf(hashp, pageno, bufp, 0); if (!rbufp) { save_bufp->flags &= ~BUF_PIN; return (ERROR); } /* FOR LOOP INIT */ bp = (u_int16_t *)rbufp->page; n = *bp++; ndx = 1; off = hashp->BSIZE; } else { save_bufp->flags &= ~BUF_PIN; return (ERROR); } } /* Not found */ switch (action) { case HASH_PUT: case HASH_PUTNEW: if (__addel(hashp, rbufp, key, val)) { save_bufp->flags &= ~BUF_PIN; return (ERROR); } else { save_bufp->flags &= ~BUF_PIN; return (SUCCESS); } case HASH_GET: case HASH_DELETE: default: save_bufp->flags &= ~BUF_PIN; return (ABNORMAL); } found: switch (action) { case HASH_PUTNEW: save_bufp->flags &= ~BUF_PIN; return (ABNORMAL); case HASH_GET: bp = (u_int16_t *)rbufp->page; if (bp[ndx + 1] < REAL_KEY) { if (__big_return(hashp, rbufp, ndx, val, 0)) return (ERROR); } else { val->data = (u_char *)rbufp->page + (int)bp[ndx + 1]; val->size = bp[ndx] - bp[ndx + 1]; } break; case HASH_PUT: if ((__delpair(hashp, rbufp, ndx)) || (__addel(hashp, rbufp, key, val))) { save_bufp->flags &= ~BUF_PIN; return (ERROR); } break; case HASH_DELETE: if (__delpair(hashp, rbufp, ndx)) return (ERROR); break; default: abort(); } save_bufp->flags &= ~BUF_PIN; return (SUCCESS); }
/* * Returns: * 0 ==> OK * -1 ==> Error */ int __split_page(HTAB *hashp, uint32_t obucket, uint32_t nbucket) { BUFHEAD *new_bufp, *old_bufp; uint16_t *ino; char *np; DBT key, val; int n, ndx, retval; uint16_t copyto, diff, off, moved; char *op; size_t temp; copyto = (uint16_t)hashp->BSIZE; off = (uint16_t)hashp->BSIZE; old_bufp = __get_buf(hashp, obucket, NULL, 0); if (old_bufp == NULL) return (-1); new_bufp = __get_buf(hashp, nbucket, NULL, 0); if (new_bufp == NULL) return (-1); old_bufp->flags |= (BUF_MOD | BUF_PIN); new_bufp->flags |= (BUF_MOD | BUF_PIN); ino = (uint16_t *)(void *)(op = old_bufp->page); np = new_bufp->page; moved = 0; for (n = 1, ndx = 1; n < ino[0]; n += 2) { if (ino[n + 1] < REAL_KEY) { retval = ugly_split(hashp, obucket, old_bufp, new_bufp, (int)copyto, (int)moved); old_bufp->flags &= ~BUF_PIN; new_bufp->flags &= ~BUF_PIN; return (retval); } key.data = (uint8_t *)op + ino[n]; key.size = off - ino[n]; if (__call_hash(hashp, key.data, (int)key.size) == obucket) { /* Don't switch page */ diff = copyto - off; if (diff) { copyto = ino[n + 1] + diff; memmove(op + copyto, op + ino[n + 1], (size_t)(off - ino[n + 1])); ino[ndx] = copyto + ino[n] - ino[n + 1]; ino[ndx + 1] = copyto; } else copyto = ino[n + 1]; ndx += 2; } else { /* Switch page */ val.data = (uint8_t *)op + ino[n + 1]; val.size = ino[n] - ino[n + 1]; putpair(np, &key, &val); moved += 2; } off = ino[n + 1]; } /* Now clean up the page */ ino[0] -= moved; temp = sizeof(uint16_t) * (ino[0] + 3); _DIAGASSERT(copyto >= temp); FREESPACE(ino) = (uint16_t)(copyto - temp); OFFSET(ino) = copyto; #ifdef DEBUG3 (void)fprintf(stderr, "split %d/%d\n", ((uint16_t *)np)[0] / 2, ((uint16_t *)op)[0] / 2); #endif /* unpin both pages */ old_bufp->flags &= ~BUF_PIN; new_bufp->flags &= ~BUF_PIN; return (0); }
/* * Returns: * 0 => OK * -1 => error */ int __big_split( HTAB *hashp, BUFHEAD *op, /* Pointer to where to put keys that go in old bucket */ BUFHEAD *np, /* Pointer to new bucket page */ /* Pointer to first page containing the big key/data */ BUFHEAD *big_keyp, int addr, /* Address of big_keyp */ uint32_t obucket,/* Old Bucket */ SPLIT_RETURN *ret ) { BUFHEAD *tmpp; uint16_t *tp; BUFHEAD *bp; DBT key, val; uint32_t change; uint16_t free_space, n, off; size_t temp; bp = big_keyp; /* Now figure out where the big key/data goes */ if (__big_keydata(hashp, big_keyp, &key, &val, 0)) return (-1); change = (__call_hash(hashp, key.data, (int)key.size) != obucket); if ((ret->next_addr = __find_last_page(hashp, &big_keyp)) != 0) { if (!(ret->nextp = __get_buf(hashp, (uint32_t)ret->next_addr, big_keyp, 0))) return (-1); } else ret->nextp = NULL; /* Now make one of np/op point to the big key/data pair */ _DIAGASSERT(np->ovfl == NULL); if (change) tmpp = np; else tmpp = op; tmpp->flags |= BUF_MOD; #ifdef DEBUG1 (void)fprintf(stderr, "BIG_SPLIT: %d->ovfl was %d is now %d\n", tmpp->addr, (tmpp->ovfl ? tmpp->ovfl->addr : 0), (bp ? bp->addr : 0)); #endif tmpp->ovfl = bp; /* one of op/np point to big_keyp */ tp = (uint16_t *)(void *)tmpp->page; _DIAGASSERT(FREESPACE(tp) >= OVFLSIZE); n = tp[0]; off = OFFSET(tp); free_space = FREESPACE(tp); tp[++n] = (uint16_t)addr; tp[++n] = OVFLPAGE; tp[0] = n; OFFSET(tp) = off; temp = free_space - OVFLSIZE; _DBFIT(temp, uint16_t); FREESPACE(tp) = (uint16_t)temp; /* * Finally, set the new and old return values. BIG_KEYP contains a * pointer to the last page of the big key_data pair. Make sure that * big_keyp has no following page (2 elements) or create an empty * following page. */ ret->newp = np; ret->oldp = op; tp = (uint16_t *)(void *)big_keyp->page; big_keyp->flags |= BUF_MOD; if (tp[0] > 2) { /* * There may be either one or two offsets on this page. If * there is one, then the overflow page is linked on normally * and tp[4] is OVFLPAGE. If there are two, tp[4] contains * the second offset and needs to get stuffed in after the * next overflow page is added. */ n = tp[4]; free_space = FREESPACE(tp); off = OFFSET(tp); tp[0] -= 2; temp = free_space + OVFLSIZE; _DBFIT(temp, uint16_t); FREESPACE(tp) = (uint16_t)temp; OFFSET(tp) = off; tmpp = __add_ovflpage(hashp, big_keyp); if (!tmpp) return (-1); tp[4] = n; } else tmpp = big_keyp; if (change) ret->newp = tmpp; else ret->oldp = tmpp; return (0); }
/* * Assume that hashp has been set in wrapper routine. */ static int hash_access( HTAB *hashp, ACTION action, DBT *key, DBT *val) { register BUFHEAD *rbufp; BUFHEAD *bufp, *save_bufp; register uint16 *bp; register long n, ndx, off; register size_t size; register char *kp; uint16 pageno; uint32 ovfl_loop_count=0; int32 last_overflow_page_no = -1; #ifdef HASH_STATISTICS hash_accesses++; #endif off = hashp->BSIZE; size = key->size; kp = (char *)key->data; rbufp = __get_buf(hashp, __call_hash(hashp, kp, size), NULL, 0); if (!rbufp) return (DATABASE_CORRUPTED_ERROR); save_bufp = rbufp; /* Pin the bucket chain */ rbufp->flags |= BUF_PIN; for (bp = (uint16 *)rbufp->page, n = *bp++, ndx = 1; ndx < n;) { if (bp[1] >= REAL_KEY) { /* Real key/data pair */ if (size == (unsigned long)(off - *bp) && memcmp(kp, rbufp->page + *bp, size) == 0) goto found; off = bp[1]; #ifdef HASH_STATISTICS hash_collisions++; #endif bp += 2; ndx += 2; } else if (bp[1] == OVFLPAGE) { /* database corruption: overflow loop detection */ if(last_overflow_page_no == (int32)*bp) return (DATABASE_CORRUPTED_ERROR); last_overflow_page_no = *bp; rbufp = __get_buf(hashp, *bp, rbufp, 0); if (!rbufp) { save_bufp->flags &= ~BUF_PIN; return (DBM_ERROR); } ovfl_loop_count++; if(ovfl_loop_count > MAX_OVERFLOW_HASH_ACCESS_LOOPS) return (DATABASE_CORRUPTED_ERROR); /* FOR LOOP INIT */ bp = (uint16 *)rbufp->page; n = *bp++; ndx = 1; off = hashp->BSIZE; } else if (bp[1] < REAL_KEY) { if ((ndx = __find_bigpair(hashp, rbufp, ndx, kp, (int)size)) > 0) goto found; if (ndx == -2) { bufp = rbufp; if (!(pageno = __find_last_page(hashp, &bufp))) { ndx = 0; rbufp = bufp; break; /* FOR */ } rbufp = __get_buf(hashp, pageno, bufp, 0); if (!rbufp) { save_bufp->flags &= ~BUF_PIN; return (DBM_ERROR); } /* FOR LOOP INIT */ bp = (uint16 *)rbufp->page; n = *bp++; ndx = 1; off = hashp->BSIZE; } else { save_bufp->flags &= ~BUF_PIN; return (DBM_ERROR); } } } /* Not found */ switch (action) { case HASH_PUT: case HASH_PUTNEW: if (__addel(hashp, rbufp, key, val)) { save_bufp->flags &= ~BUF_PIN; return (DBM_ERROR); } else { save_bufp->flags &= ~BUF_PIN; return (SUCCESS); } case HASH_GET: case HASH_DELETE: default: save_bufp->flags &= ~BUF_PIN; return (ABNORMAL); } found: switch (action) { case HASH_PUTNEW: save_bufp->flags &= ~BUF_PIN; return (ABNORMAL); case HASH_GET: bp = (uint16 *)rbufp->page; if (bp[ndx + 1] < REAL_KEY) { if (__big_return(hashp, rbufp, ndx, val, 0)) return (DBM_ERROR); } else { val->data = (uint8 *)rbufp->page + (int)bp[ndx + 1]; val->size = bp[ndx] - bp[ndx + 1]; } break; case HASH_PUT: if ((__delpair(hashp, rbufp, ndx)) || (__addel(hashp, rbufp, key, val))) { save_bufp->flags &= ~BUF_PIN; return (DBM_ERROR); } break; case HASH_DELETE: if (__delpair(hashp, rbufp, ndx)) return (DBM_ERROR); break; default: abort(); } save_bufp->flags &= ~BUF_PIN; return (SUCCESS); }
static int ugly_split(HTAB *hashp, uint32 obucket, BUFHEAD *old_bufp, BUFHEAD *new_bufp, /* Same as __split_page. */ int copyto, int moved) /* int copyto; First byte on page which contains key/data values. */ /* int moved; Number of pairs moved to new page. */ { register BUFHEAD *bufp; /* Buffer header for ino */ register uint16 *ino; /* Page keys come off of */ register uint16 *np; /* New page */ register uint16 *op; /* Page keys go on to if they aren't moving */ uint32 loop_detection = 0; BUFHEAD *last_bfp; /* Last buf header OVFL needing to be freed */ DBT key, val; SPLIT_RETURN ret; uint16 n, off, ov_addr, scopyto; char *cino; /* Character value of ino */ int status; bufp = old_bufp; ino = (uint16 *)old_bufp->page; np = (uint16 *)new_bufp->page; op = (uint16 *)old_bufp->page; last_bfp = NULL; scopyto = (uint16)copyto; /* ANSI */ n = ino[0] - 1; while (n < ino[0]) { /* this function goes nuts sometimes and never returns. * I havent found the problem yet but I need a solution * so if we loop too often we assume a database curruption error * :LJM */ loop_detection++; if (loop_detection > MAX_UGLY_SPLIT_LOOPS) return DATABASE_CORRUPTED_ERROR; if (ino[2] < REAL_KEY && ino[2] != OVFLPAGE) { if ((status = __big_split(hashp, old_bufp, new_bufp, bufp, bufp->addr, obucket, &ret))) return (status); old_bufp = ret.oldp; if (!old_bufp) return (-1); op = (uint16 *)old_bufp->page; new_bufp = ret.newp; if (!new_bufp) return (-1); np = (uint16 *)new_bufp->page; bufp = ret.nextp; if (!bufp) return (0); cino = (char *)bufp->page; ino = (uint16 *)cino; last_bfp = ret.nextp; } else if (ino[n + 1] == OVFLPAGE) { ov_addr = ino[n]; /* * Fix up the old page -- the extra 2 are the fields * which contained the overflow information. */ ino[0] -= (moved + 2); FREESPACE(ino) = scopyto - sizeof(uint16) * (ino[0] + 3); OFFSET(ino) = scopyto; bufp = __get_buf(hashp, ov_addr, bufp, 0); if (!bufp) return (-1); ino = (uint16 *)bufp->page; n = 1; scopyto = hashp->BSIZE; moved = 0; if (last_bfp) __free_ovflpage(hashp, last_bfp); last_bfp = bufp; } /* Move regular sized pairs of there are any */ off = hashp->BSIZE; for (n = 1; (n < ino[0]) && (ino[n + 1] >= REAL_KEY); n += 2) { cino = (char *)ino; key.data = (uint8 *)cino + ino[n]; key.size = off - ino[n]; val.data = (uint8 *)cino + ino[n + 1]; val.size = ino[n] - ino[n + 1]; off = ino[n + 1]; if (__call_hash(hashp, (char *)key.data, key.size) == obucket) { /* Keep on old page */ if (PAIRFITS(op, (&key), (&val))) putpair((char *)op, &key, &val); else { old_bufp = __add_ovflpage(hashp, old_bufp); if (!old_bufp) return (-1); op = (uint16 *)old_bufp->page; putpair((char *)op, &key, &val); } old_bufp->flags |= BUF_MOD; } else { /* Move to new page */ if (PAIRFITS(np, (&key), (&val))) putpair((char *)np, &key, &val); else { new_bufp = __add_ovflpage(hashp, new_bufp); if (!new_bufp) return (-1); np = (uint16 *)new_bufp->page; putpair((char *)np, &key, &val); } new_bufp->flags |= BUF_MOD; } } } if (last_bfp) __free_ovflpage(hashp, last_bfp); return (0); }
/* * Returns: * 0 ==> OK * -1 ==> Error */ extern int __split_page(HTAB *hashp, uint32 obucket, uint32 nbucket) { register BUFHEAD *new_bufp, *old_bufp; register uint16 *ino; register uint16 *tmp_uint16_array; register char *np; DBT key, val; uint16 n, ndx; int retval; uint16 copyto, diff, moved; size_t off; char *op; copyto = (uint16)hashp->BSIZE; off = (uint16)hashp->BSIZE; old_bufp = __get_buf(hashp, obucket, NULL, 0); if (old_bufp == NULL) return (-1); new_bufp = __get_buf(hashp, nbucket, NULL, 0); if (new_bufp == NULL) return (-1); old_bufp->flags |= (BUF_MOD | BUF_PIN); new_bufp->flags |= (BUF_MOD | BUF_PIN); ino = (uint16 *)(op = old_bufp->page); np = new_bufp->page; moved = 0; for (n = 1, ndx = 1; n < ino[0]; n += 2) { if (ino[n + 1] < REAL_KEY) { retval = ugly_split(hashp, obucket, old_bufp, new_bufp, (int)copyto, (int)moved); old_bufp->flags &= ~BUF_PIN; new_bufp->flags &= ~BUF_PIN; return (retval); } key.data = (uint8 *)op + ino[n]; /* check here for ino[n] being greater than * off. If it is then the database has * been corrupted. */ if (ino[n] > off) return (DATABASE_CORRUPTED_ERROR); key.size = off - ino[n]; #ifdef DEBUG /* make sure the size is positive */ assert(((int)key.size) > -1); #endif if (__call_hash(hashp, (char *)key.data, key.size) == obucket) { /* Don't switch page */ diff = copyto - off; if (diff) { copyto = ino[n + 1] + diff; memmove(op + copyto, op + ino[n + 1], off - ino[n + 1]); ino[ndx] = copyto + ino[n] - ino[n + 1]; ino[ndx + 1] = copyto; } else copyto = ino[n + 1]; ndx += 2; } else { /* Switch page */ val.data = (uint8 *)op + ino[n + 1]; val.size = ino[n] - ino[n + 1]; /* if the pair doesn't fit something is horribly * wrong. LJM */ tmp_uint16_array = (uint16 *)np; if (!PAIRFITS(tmp_uint16_array, &key, &val)) return (DATABASE_CORRUPTED_ERROR); putpair(np, &key, &val); moved += 2; } off = ino[n + 1]; } /* Now clean up the page */ ino[0] -= moved; FREESPACE(ino) = copyto - sizeof(uint16) * (ino[0] + 3); OFFSET(ino) = copyto; #ifdef DEBUG3 (void)fprintf(stderr, "split %d/%d\n", ((uint16 *)np)[0] / 2, ((uint16 *)op)[0] / 2); #endif /* unpin both pages */ old_bufp->flags &= ~BUF_PIN; new_bufp->flags &= ~BUF_PIN; return (0); }