/*
* BT_RROOT -- Fix up the recno root page after it has been split.
*
* Parameters:
* t: tree
* h: root page
* l: left page
* r: right page
*
* Returns:
* RET_ERROR, RET_SUCCESS
*/
static int
bt_rroot(BTREE *t, PAGE *h, PAGE *l, PAGE *r)
{
char *dest;
/* Insert the left and right keys, set the header information. */
h->linp[0] = h->upper = t->bt_psize - NRINTERNAL;
dest = (char *)h + h->upper;
WR_RINTERNAL(dest,
l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
__PAST_END(h->linp, 1) = h->upper -= NRINTERNAL;
dest = (char *)h + h->upper;
WR_RINTERNAL(dest,
r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
h->lower = BTDATAOFF + 2 * sizeof(indx_t);
/* Unpin the root page, set to recno internal page. */
h->flags &= ~P_TYPE;
h->flags |= P_RINTERNAL;
mpool_put(t->bt_mp, h, MPOOL_DIRTY);
return (RET_SUCCESS);
}
void
unlinkpg(DB *dbp)
{
BTREE *t = dbp->internal;
PAGE *h = NULL;
pgno_t pg;
for (pg = P_ROOT; pg < t->bt_mp->npages;
mpool_put(t->bt_mp, h, 0), pg++) {
if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
break;
/* Look for a nonempty leaf page that has both left
* and right siblings. */
if (h->prevpg == P_INVALID || h->nextpg == P_INVALID)
continue;
if (NEXTINDEX(h) == 0)
continue;
if ((h->flags & (P_BLEAF | P_RLEAF)))
break;
}
if (h == NULL || pg == t->bt_mp->npages) {
errx(1, "%s: no appropriate page found", __func__);
return;
}
if (__bt_relink(t, h) != 0) {
perror("unlinkpg");
goto cleanup;
}
h->prevpg = P_INVALID;
h->nextpg = P_INVALID;
cleanup:
mpool_put(t->bt_mp, h, MPOOL_DIRTY);
}
/*
* REC_TOTAL -- Return the number of recno entries below a page.
*
* Parameters:
* h: page
*
* Returns:
* The number of recno entries below a page.
*
* XXX
* These values could be set by the bt_psplit routine. The problem is that the
* entry has to be popped off of the stack etc. or the values have to be passed
* all the way back to bt_split/bt_rroot and it's not very clean.
*/
static recno_t
rec_total(PAGE *h)
{
recno_t recs;
indx_t nxt, top;
for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
recs += GETRINTERNAL(h, nxt)->nrecs;
return (recs);
}
/*
* __bt_dleaf --
* Delete a single record from a leaf page.
*
* Parameters:
* t: tree
* key: referenced key
* h: page
* idx: index on page to delete
*
* Returns:
* RET_SUCCESS, RET_ERROR.
*/
int
__bt_dleaf(BTREE *t, const DBT *key, PAGE *h, u_int idx)
{
BLEAF *bl;
indx_t cnt, *ip, offset;
u_int32_t nbytes;
void *to;
char *from;
/* If this record is referenced by the cursor, delete the cursor. */
if (F_ISSET(&t->bt_cursor, CURS_INIT) &&
!F_ISSET(&t->bt_cursor, CURS_ACQUIRE) &&
t->bt_cursor.pg.pgno == h->pgno && t->bt_cursor.pg.index == idx &&
__bt_curdel(t, key, h, idx))
return (RET_ERROR);
/* If the entry uses overflow pages, make them available for reuse. */
to = bl = GETBLEAF(h, idx);
if (bl->flags & P_BIGKEY && __ovfl_delete(t, bl->bytes) == RET_ERROR)
return (RET_ERROR);
if (bl->flags & P_BIGDATA &&
__ovfl_delete(t, bl->bytes + bl->ksize) == RET_ERROR)
return (RET_ERROR);
/* Pack the remaining key/data items at the end of the page. */
nbytes = NBLEAF(bl);
from = (char *)h + h->upper;
memmove(from + nbytes, from, (char *)to - from);
h->upper += nbytes;
/* Adjust the indices' offsets, shift the indices down. */
offset = h->linp[idx];
for (cnt = idx, ip = &h->linp[0]; cnt--; ++ip)
if (ip[0] < offset)
ip[0] += nbytes;
for (cnt = NEXTINDEX(h) - idx; --cnt; ++ip)
ip[0] = ip[1] < offset ? ip[1] + nbytes : ip[1];
h->lower -= sizeof(indx_t);
/* If the cursor is on this page, adjust it as necessary. */
if (F_ISSET(&t->bt_cursor, CURS_INIT) &&
!F_ISSET(&t->bt_cursor, CURS_ACQUIRE) &&
t->bt_cursor.pg.pgno == h->pgno && t->bt_cursor.pg.index > idx)
--t->bt_cursor.pg.index;
return (RET_SUCCESS);
}
/*
* __REC_DLEAF -- Delete a single record from a recno leaf page.
*
* Parameters:
* t: tree
* idx: index on current page to delete
*
* Returns:
* RET_SUCCESS, RET_ERROR.
*/
int
__rec_dleaf(BTREE *t, PAGE *h, u_int32_t idx)
{
RLEAF *rl;
indx_t *ip, cnt, offset;
u_int32_t nbytes;
char *from;
void *to;
/*
* Delete a record from a recno leaf page. Internal records are never
* deleted from internal pages, regardless of the records that caused
* them to be added being deleted. Pages made empty by deletion are
* not reclaimed. They are, however, made available for reuse.
*
* Pack the remaining entries at the end of the page, shift the indices
* down, overwriting the deleted record and its index. If the record
* uses overflow pages, make them available for reuse.
*/
to = rl = GETRLEAF(h, idx);
if (rl->flags & P_BIGDATA && __ovfl_delete(t, rl->bytes) == RET_ERROR)
return (RET_ERROR);
nbytes = NRLEAF(rl);
/*
* Compress the key/data pairs. Compress and adjust the [BR]LEAF
* offsets. Reset the headers.
*/
from = (char *)h + h->upper;
memmove(from + nbytes, from, (char *)to - from);
h->upper += nbytes;
offset = h->linp[idx];
for (cnt = &h->linp[idx] - (ip = &h->linp[0]); cnt--; ++ip)
if (ip[0] < offset)
ip[0] += nbytes;
for (cnt = &h->linp[NEXTINDEX(h)] - ip; --cnt; ++ip)
ip[0] = ip[1] < offset ? ip[1] + nbytes : ip[1];
h->lower -= sizeof(indx_t);
--t->bt_nrecs;
return (RET_SUCCESS);
}
/*
* __bt_sprev --
* Check for an exact match before the key.
*
* Parameters:
* t: tree
* h: current page
* key: key
* exactp: pointer to exact match flag
*
* Returns:
* If an exact match found.
*/
static int
__bt_sprev(BTREE *t, PAGE *h, const DBT *key, int *exactp)
{
EPG e;
/*
* Get the previous page. The key is either an exact
* match, or not as good as the one we already have.
*/
if ((e.page = mpool_get(t->bt_mp, h->prevpg, 0)) == NULL)
return (0);
e.index = NEXTINDEX(e.page) - 1;
if (__bt_cmp(t, key, &e) == 0) {
mpool_put(t->bt_mp, h, 0);
t->bt_cur = e;
*exactp = 1;
return (1);
}
mpool_put(t->bt_mp, e.page, 0);
return (0);
}
/*
* __bt_seqset --
* Set the sequential scan to a specific key.
*
* Parameters:
* t: tree
* ep: storage for returned key
* key: key for initial scan position
* flags: R_CURSOR, R_FIRST, R_LAST, R_NEXT, R_PREV
*
* Side effects:
* Pins the page the cursor references.
*
* Returns:
* RET_ERROR, RET_SUCCESS or RET_SPECIAL if there's no next key.
*/
static int
__bt_seqset(BTREE *t, EPG *ep, DBT *key, int flags)
{
PAGE *h;
pgno_t pg;
int exact;
/*
* Find the first, last or specific key in the tree and point the
* cursor at it. The cursor may not be moved until a new key has
* been found.
*/
switch (flags) {
case R_CURSOR: /* Keyed scan. */
/*
* Find the first instance of the key or the smallest key
* which is greater than or equal to the specified key.
*/
if (key->data == NULL || key->size == 0) {
errno = EINVAL;
return (RET_ERROR);
}
return (__bt_first(t, key, ep, &exact));
case R_FIRST: /* First record. */
case R_NEXT:
/* Walk down the left-hand side of the tree. */
for (pg = P_ROOT;;) {
if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
return (RET_ERROR);
/* Check for an empty tree. */
if (NEXTINDEX(h) == 0) {
mpool_put(t->bt_mp, h, 0);
return (RET_SPECIAL);
}
if (h->flags & (P_BLEAF | P_RLEAF))
break;
pg = GETBINTERNAL(h, 0)->pgno;
mpool_put(t->bt_mp, h, 0);
}
ep->page = h;
ep->index = 0;
break;
case R_LAST: /* Last record. */
case R_PREV:
/* Walk down the right-hand side of the tree. */
for (pg = P_ROOT;;) {
if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
return (RET_ERROR);
/* Check for an empty tree. */
if (NEXTINDEX(h) == 0) {
mpool_put(t->bt_mp, h, 0);
return (RET_SPECIAL);
}
if (h->flags & (P_BLEAF | P_RLEAF))
break;
pg = GETBINTERNAL(h, NEXTINDEX(h) - 1)->pgno;
mpool_put(t->bt_mp, h, 0);
}
ep->page = h;
ep->index = NEXTINDEX(h) - 1;
break;
}
return (RET_SUCCESS);
}
/*
* __bt_delete
* Delete the item(s) referenced by a key.
*
* Return RET_SPECIAL if the key is not found.
*/
int
__bt_delete(const DB *dbp, const DBT *key, u_int flags)
{
BTREE *t;
CURSOR *c;
PAGE *h;
int status;
t = dbp->internal;
/* Toss any page pinned across calls. */
if (t->bt_pinned != NULL) {
mpool_put(t->bt_mp, t->bt_pinned, 0);
t->bt_pinned = NULL;
}
/* Check for change to a read-only tree. */
if (F_ISSET(t, B_RDONLY)) {
errno = EPERM;
return (RET_ERROR);
}
switch (flags) {
case 0:
status = __bt_bdelete(t, key);
break;
case R_CURSOR:
/*
* If flags is R_CURSOR, delete the cursor. Must already
* have started a scan and not have already deleted it.
*/
c = &t->bt_cursor;
if (F_ISSET(c, CURS_INIT)) {
if (F_ISSET(c, CURS_ACQUIRE | CURS_AFTER | CURS_BEFORE))
return (RET_SPECIAL);
if ((h = mpool_get(t->bt_mp, c->pg.pgno, 0)) == NULL)
return (RET_ERROR);
/*
* If the page is about to be emptied, we'll need to
* delete it, which means we have to acquire a stack.
*/
if (NEXTINDEX(h) == 1)
if (__bt_stkacq(t, &h, &t->bt_cursor))
return (RET_ERROR);
status = __bt_dleaf(t, NULL, h, c->pg.index);
if (NEXTINDEX(h) == 0 && status == RET_SUCCESS) {
if (__bt_pdelete(t, h))
return (RET_ERROR);
} else
mpool_put(t->bt_mp,
h, status == RET_SUCCESS ? MPOOL_DIRTY : 0);
break;
}
/* FALLTHROUGH */
default:
errno = EINVAL;
return (RET_ERROR);
}
if (status == RET_SUCCESS)
F_SET(t, B_MODIFIED);
return (status);
}
/*
* __REC_SEARCH -- Search a btree for a key.
*
* Parameters:
* t: tree to search
* recno: key to find
* op: search operation
*
* Returns:
* EPG for matching record, if any, or the EPG for the location of the
* key, if it were inserted into the tree.
*
* Returns:
* The EPG for matching record, if any, or the EPG for the location
* of the key, if it were inserted into the tree, is entered into
* the bt_cur field of the tree. A pointer to the field is returned.
*/
EPG *
__rec_search(BTREE *t, recno_t recno, enum SRCHOP op)
{
indx_t idx;
PAGE *h;
EPGNO *parent;
RINTERNAL *r;
pgno_t pg;
indx_t top;
recno_t total;
int sverrno;
BT_CLR(t);
for (pg = P_ROOT, total = 0;;) {
if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
goto err;
if (h->flags & P_RLEAF) {
t->bt_cur.page = h;
t->bt_cur.index = recno - total;
return (&t->bt_cur);
}
for (idx = 0, top = NEXTINDEX(h);;) {
r = GETRINTERNAL(h, idx);
if (++idx == top || total + r->nrecs > recno)
break;
total += r->nrecs;
}
BT_PUSH(t, pg, idx - 1);
pg = r->pgno;
switch (op) {
case SDELETE:
--GETRINTERNAL(h, (idx - 1))->nrecs;
mpool_put(t->bt_mp, h, MPOOL_DIRTY);
break;
case SINSERT:
++GETRINTERNAL(h, (idx - 1))->nrecs;
mpool_put(t->bt_mp, h, MPOOL_DIRTY);
break;
case SEARCH:
mpool_put(t->bt_mp, h, 0);
break;
}
}
/* Try and recover the tree. */
err: sverrno = errno;
if (op != SEARCH)
while ((parent = BT_POP(t)) != NULL) {
if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
break;
if (op == SINSERT)
--GETRINTERNAL(h, parent->index)->nrecs;
else
++GETRINTERNAL(h, parent->index)->nrecs;
mpool_put(t->bt_mp, h, MPOOL_DIRTY);
}
errno = sverrno;
return (NULL);
}
/*
* BT_DPAGE -- Dump the page
*
* Parameters:
* h: pointer to the PAGE
*/
void
__bt_dpage(PAGE *h)
{
BINTERNAL *bi;
BLEAF *bl;
RINTERNAL *ri;
RLEAF *rl;
indx_t cur, top;
char *sep;
(void)fprintf(stderr, " page %u: (", h->pgno);
#undef X
#define X(flag, name) \
if (h->flags & flag) { \
(void)fprintf(stderr, "%s%s", sep, name); \
sep = ", "; \
}
sep = "";
X(P_BINTERNAL, "BINTERNAL") /* types */
X(P_BLEAF, "BLEAF")
X(P_RINTERNAL, "RINTERNAL") /* types */
X(P_RLEAF, "RLEAF")
X(P_OVERFLOW, "OVERFLOW")
X(P_PRESERVE, "PRESERVE");
(void)fprintf(stderr, ")\n");
#undef X
(void)fprintf(stderr, "\tprev %2u next %2u", h->prevpg, h->nextpg);
if (h->flags & P_OVERFLOW)
return;
top = NEXTINDEX(h);
(void)fprintf(stderr, " lower %3d upper %3d nextind %d\n",
h->lower, h->upper, top);
for (cur = 0; cur < top; cur++) {
(void)fprintf(stderr, "\t[%03d] %4d ", cur, h->linp[cur]);
switch (h->flags & P_TYPE) {
case P_BINTERNAL:
bi = GETBINTERNAL(h, cur);
(void)fprintf(stderr,
"size %03d pgno %03d", bi->ksize, bi->pgno);
if (bi->flags & P_BIGKEY)
(void)fprintf(stderr, " (indirect)");
else if (bi->ksize)
(void)fprintf(stderr,
" {%.*s}", (int)bi->ksize, bi->bytes);
break;
case P_RINTERNAL:
ri = GETRINTERNAL(h, cur);
(void)fprintf(stderr, "entries %03d pgno %03d",
ri->nrecs, ri->pgno);
break;
case P_BLEAF:
bl = GETBLEAF(h, cur);
if (bl->flags & P_BIGKEY)
(void)fprintf(stderr,
"big key page %u size %u/",
*(pgno_t *)bl->bytes,
*(u_int32_t *)(bl->bytes + sizeof(pgno_t)));
else if (bl->ksize)
(void)fprintf(stderr, "%s/", bl->bytes);
if (bl->flags & P_BIGDATA)
(void)fprintf(stderr,
"big data page %u size %u",
*(pgno_t *)(bl->bytes + bl->ksize),
*(u_int32_t *)(bl->bytes + bl->ksize +
sizeof(pgno_t)));
else if (bl->dsize)
(void)fprintf(stderr, "%.*s",
(int)bl->dsize, bl->bytes + bl->ksize);
break;
case P_RLEAF:
rl = GETRLEAF(h, cur);
if (rl->flags & P_BIGDATA)
(void)fprintf(stderr,
"big data page %u size %u",
*(pgno_t *)rl->bytes,
*(u_int32_t *)(rl->bytes + sizeof(pgno_t)));
else if (rl->dsize)
(void)fprintf(stderr,
"%.*s", (int)rl->dsize, rl->bytes);
break;
}
(void)fprintf(stderr, "\n");
}
}
/*
* __bt_curdel --
* Delete the cursor.
*
* Parameters:
* t: tree
* key: referenced key (or NULL)
* h: page
* idx: index on page to delete
*
* Returns:
* RET_SUCCESS, RET_ERROR.
*/
static int
__bt_curdel(BTREE *t, const DBT *key, PAGE *h, u_int idx)
{
CURSOR *c;
EPG e;
PAGE *pg;
int curcopy, status;
/*
* If there are duplicates, move forward or backward to one.
* Otherwise, copy the key into the cursor area.
*/
c = &t->bt_cursor;
F_CLR(c, CURS_AFTER | CURS_BEFORE | CURS_ACQUIRE);
curcopy = 0;
if (!F_ISSET(t, B_NODUPS)) {
/*
* We're going to have to do comparisons. If we weren't
* provided a copy of the key, i.e. the user is deleting
* the current cursor position, get one.
*/
if (key == NULL) {
e.page = h;
e.index = idx;
if ((status = __bt_ret(t, &e,
&c->key, &c->key, NULL, NULL, 1)) != RET_SUCCESS)
return (status);
curcopy = 1;
key = &c->key;
}
/* Check previous key, if not at the beginning of the page. */
if (idx > 0) {
e.page = h;
e.index = idx - 1;
if (__bt_cmp(t, key, &e) == 0) {
F_SET(c, CURS_BEFORE);
goto dup2;
}
}
/* Check next key, if not at the end of the page. */
if (idx < NEXTINDEX(h) - 1) {
e.page = h;
e.index = idx + 1;
if (__bt_cmp(t, key, &e) == 0) {
F_SET(c, CURS_AFTER);
goto dup2;
}
}
/* Check previous key if at the beginning of the page. */
if (idx == 0 && h->prevpg != P_INVALID) {
if ((pg = mpool_get(t->bt_mp, h->prevpg, 0)) == NULL)
return (RET_ERROR);
e.page = pg;
e.index = NEXTINDEX(pg) - 1;
if (__bt_cmp(t, key, &e) == 0) {
F_SET(c, CURS_BEFORE);
goto dup1;
}
mpool_put(t->bt_mp, pg, 0);
}
/* Check next key if at the end of the page. */
if (idx == NEXTINDEX(h) - 1 && h->nextpg != P_INVALID) {
if ((pg = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL)
return (RET_ERROR);
e.page = pg;
e.index = 0;
if (__bt_cmp(t, key, &e) == 0) {
F_SET(c, CURS_AFTER);
dup1: mpool_put(t->bt_mp, pg, 0);
dup2: c->pg.pgno = e.page->pgno;
c->pg.index = e.index;
return (RET_SUCCESS);
}
mpool_put(t->bt_mp, pg, 0);
}
}
e.page = h;
e.index = idx;
if (curcopy || (status =
__bt_ret(t, &e, &c->key, &c->key, NULL, NULL, 1)) == RET_SUCCESS) {
F_SET(c, CURS_ACQUIRE);
return (RET_SUCCESS);
}
return (status);
}
/*
* __bt_bdelete --
* Delete all key/data pairs matching the specified key.
*
* Parameters:
* t: tree
* key: key to delete
*
* Returns:
* RET_ERROR, RET_SUCCESS and RET_SPECIAL if the key not found.
*/
static int
__bt_bdelete(BTREE *t, const DBT *key)
{
EPG *e;
PAGE *h;
int deleted, exact, redo;
deleted = 0;
/* Find any matching record; __bt_search pins the page. */
loop: if ((e = __bt_search(t, key, &exact)) == NULL)
return (deleted ? RET_SUCCESS : RET_ERROR);
if (!exact) {
mpool_put(t->bt_mp, e->page, 0);
return (deleted ? RET_SUCCESS : RET_SPECIAL);
}
/*
* Delete forward, then delete backward, from the found key. If
* there are duplicates and we reach either side of the page, do
* the key search again, so that we get them all.
*/
redo = 0;
h = e->page;
do {
if (__bt_dleaf(t, key, h, e->index)) {
mpool_put(t->bt_mp, h, 0);
return (RET_ERROR);
}
if (F_ISSET(t, B_NODUPS)) {
if (NEXTINDEX(h) == 0) {
if (__bt_pdelete(t, h))
return (RET_ERROR);
} else
mpool_put(t->bt_mp, h, MPOOL_DIRTY);
return (RET_SUCCESS);
}
deleted = 1;
} while (e->index < NEXTINDEX(h) && __bt_cmp(t, key, e) == 0);
/* Check for right-hand edge of the page. */
if (e->index == NEXTINDEX(h))
redo = 1;
/* Delete from the key to the beginning of the page. */
while (e->index-- > 0) {
if (__bt_cmp(t, key, e) != 0)
break;
if (__bt_dleaf(t, key, h, e->index) == RET_ERROR) {
mpool_put(t->bt_mp, h, 0);
return (RET_ERROR);
}
if (e->index == 0)
redo = 1;
}
/* Check for an empty page. */
if (NEXTINDEX(h) == 0) {
if (__bt_pdelete(t, h))
return (RET_ERROR);
goto loop;
}
/* Put the page. */
mpool_put(t->bt_mp, h, MPOOL_DIRTY);
if (redo)
goto loop;
return (RET_SUCCESS);
}
/*
* __bt_pdelete --
* Delete a single page from the tree.
*
* Parameters:
* t: tree
* h: leaf page
*
* Returns:
* RET_SUCCESS, RET_ERROR.
*
* Side-effects:
* mpool_put's the page
*/
static int
__bt_pdelete(BTREE *t, PAGE *h)
{
BINTERNAL *bi;
PAGE *pg;
EPGNO *parent;
indx_t cnt, idx, *ip, offset;
u_int32_t nksize;
char *from;
/*
* Walk the parent page stack -- a LIFO stack of the pages that were
* traversed when we searched for the page where the delete occurred.
* Each stack entry is a page number and a page index offset. The
* offset is for the page traversed on the search. We've just deleted
* a page, so we have to delete the key from the parent page.
*
* If the delete from the parent page makes it empty, this process may
* continue all the way up the tree. We stop if we reach the root page
* (which is never deleted, it's just not worth the effort) or if the
* delete does not empty the page.
*/
while ((parent = BT_POP(t)) != NULL) {
/* Get the parent page. */
if ((pg = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
return (RET_ERROR);
idx = parent->index;
bi = GETBINTERNAL(pg, idx);
/* Free any overflow pages. */
if (bi->flags & P_BIGKEY &&
__ovfl_delete(t, bi->bytes) == RET_ERROR) {
mpool_put(t->bt_mp, pg, 0);
return (RET_ERROR);
}
/*
* Free the parent if it has only the one key and it's not the
* root page. If it's the rootpage, turn it back into an empty
* leaf page.
*/
if (NEXTINDEX(pg) == 1) {
if (pg->pgno == P_ROOT) {
pg->lower = BTDATAOFF;
pg->upper = t->bt_psize;
pg->flags = P_BLEAF;
} else {
if (__bt_relink(t, pg) || __bt_free(t, pg))
return (RET_ERROR);
continue;
}
} else {
/* Pack remaining key items at the end of the page. */
nksize = NBINTERNAL(bi->ksize);
from = (char *)pg + pg->upper;
memmove(from + nksize, from, (char *)bi - from);
pg->upper += nksize;
/* Adjust indices' offsets, shift the indices down. */
offset = pg->linp[idx];
for (cnt = idx, ip = &pg->linp[0]; cnt--; ++ip)
if (ip[0] < offset)
ip[0] += nksize;
for (cnt = NEXTINDEX(pg) - idx; --cnt; ++ip)
ip[0] = ip[1] < offset ? ip[1] + nksize : ip[1];
pg->lower -= sizeof(indx_t);
}
mpool_put(t->bt_mp, pg, MPOOL_DIRTY);
break;
}
/* Free the leaf page, as long as it wasn't the root. */
if (h->pgno == P_ROOT) {
mpool_put(t->bt_mp, h, MPOOL_DIRTY);
return (RET_SUCCESS);
}
return (__bt_relink(t, h) || __bt_free(t, h));
}
/*
* __REC_IPUT -- Add a recno item to the tree.
*
* Parameters:
* t: tree
* nrec: record number
* data: data
*
* Returns:
* RET_ERROR, RET_SUCCESS
*/
int
__rec_iput(BTREE *t, recno_t nrec, const DBT *data, u_int flags)
{
DBT tdata;
EPG *e;
PAGE *h;
indx_t idx, nxtindex;
pgno_t pg;
uint32_t nbytes;
int dflags, status;
char *dest, db[NOVFLSIZE];
/*
* If the data won't fit on a page, store it on indirect pages.
*
* XXX
* If the insert fails later on, these pages aren't recovered.
*/
if (data->size > t->bt_ovflsize) {
if (__ovfl_put(t, data, &pg) == RET_ERROR)
return (RET_ERROR);
tdata.data = db;
tdata.size = NOVFLSIZE;
*(pgno_t *)(void *)db = pg;
_DBFIT(data->size, uint32_t);
*(uint32_t *)(void *)(db + sizeof(pgno_t)) =
(uint32_t)data->size;
dflags = P_BIGDATA;
data = &tdata;
} else
dflags = 0;
/* __rec_search pins the returned page. */
if ((e = __rec_search(t, nrec,
nrec > t->bt_nrecs || flags == R_IAFTER || flags == R_IBEFORE ?
SINSERT : SEARCH)) == NULL)
return (RET_ERROR);
h = e->page;
idx = e->index;
/*
* Add the specified key/data pair to the tree. The R_IAFTER and
* R_IBEFORE flags insert the key after/before the specified key.
*
* Pages are split as required.
*/
switch (flags) {
case R_IAFTER:
++idx;
break;
case R_IBEFORE:
break;
default:
if (nrec < t->bt_nrecs &&
__rec_dleaf(t, h, (uint32_t)idx) == RET_ERROR) {
mpool_put(t->bt_mp, h, 0);
return (RET_ERROR);
}
break;
}
/*
* If not enough room, split the page. The split code will insert
* the key and data and unpin the current page. If inserting into
* the offset array, shift the pointers up.
*/
nbytes = NRLEAFDBT(data->size);
if ((uint32_t) (h->upper - h->lower) < nbytes + sizeof(indx_t)) {
status = __bt_split(t, h, NULL, data, dflags, nbytes,
(uint32_t)idx);
if (status == RET_SUCCESS)
++t->bt_nrecs;
return (status);
}
if (idx < (nxtindex = NEXTINDEX(h)))
memmove(h->linp + idx + 1, h->linp + idx,
(nxtindex - idx) * sizeof(indx_t));
h->lower += sizeof(indx_t);
h->linp[idx] = h->upper -= nbytes;
dest = (char *)(void *)h + h->upper;
WR_RLEAF(dest, data, dflags);
++t->bt_nrecs;
F_SET(t, B_MODIFIED);
mpool_put(t->bt_mp, h, MPOOL_DIRTY);
return (RET_SUCCESS);
}
/*
* __bt_stkacq --
* Acquire a stack so we can delete a cursor entry.
*
* Parameters:
* t: tree
* hp: pointer to current, pinned PAGE pointer
* c: pointer to the cursor
*
* Returns:
* 0 on success, 1 on failure
*/
static int
__bt_stkacq(BTREE *t, PAGE **hp, CURSOR *c)
{
BINTERNAL *bi;
EPG *e;
EPGNO *parent;
PAGE *h;
indx_t idx;
pgno_t pgno;
recno_t nextpg, prevpg;
int exact, level;
/*
* Find the first occurrence of the key in the tree. Toss the
* currently locked page so we don't hit an already-locked page.
*/
h = *hp;
mpool_put(t->bt_mp, h, 0);
if ((e = __bt_search(t, &c->key, &exact)) == NULL)
return (1);
h = e->page;
/* See if we got it in one shot. */
if (h->pgno == c->pg.pgno)
goto ret;
/*
* Move right, looking for the page. At each move we have to move
* up the stack until we don't have to move to the next page. If
* we have to change pages at an internal level, we have to fix the
* stack back up.
*/
while (h->pgno != c->pg.pgno) {
if ((nextpg = h->nextpg) == P_INVALID)
break;
mpool_put(t->bt_mp, h, 0);
/* Move up the stack. */
for (level = 0; (parent = BT_POP(t)) != NULL; ++level) {
/* Get the parent page. */
if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
return (1);
/* Move to the next index. */
if (parent->index != NEXTINDEX(h) - 1) {
idx = parent->index + 1;
BT_PUSH(t, h->pgno, idx);
break;
}
mpool_put(t->bt_mp, h, 0);
}
/* Restore the stack. */
while (level--) {
/* Push the next level down onto the stack. */
bi = GETBINTERNAL(h, idx);
pgno = bi->pgno;
BT_PUSH(t, pgno, 0);
/* Lose the currently pinned page. */
mpool_put(t->bt_mp, h, 0);
/* Get the next level down. */
if ((h = mpool_get(t->bt_mp, pgno, 0)) == NULL)
return (1);
idx = 0;
}
mpool_put(t->bt_mp, h, 0);
if ((h = mpool_get(t->bt_mp, nextpg, 0)) == NULL)
return (1);
}
if (h->pgno == c->pg.pgno)
goto ret;
/* Reacquire the original stack. */
mpool_put(t->bt_mp, h, 0);
if ((e = __bt_search(t, &c->key, &exact)) == NULL)
return (1);
h = e->page;
/*
* Move left, looking for the page. At each move we have to move
* up the stack until we don't have to change pages to move to the
* next page. If we have to change pages at an internal level, we
* have to fix the stack back up.
*/
while (h->pgno != c->pg.pgno) {
if ((prevpg = h->prevpg) == P_INVALID)
break;
mpool_put(t->bt_mp, h, 0);
/* Move up the stack. */
for (level = 0; (parent = BT_POP(t)) != NULL; ++level) {
/* Get the parent page. */
if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
return (1);
/* Move to the next index. */
if (parent->index != 0) {
idx = parent->index - 1;
BT_PUSH(t, h->pgno, idx);
break;
}
mpool_put(t->bt_mp, h, 0);
}
/* Restore the stack. */
while (level--) {
/* Push the next level down onto the stack. */
bi = GETBINTERNAL(h, idx);
pgno = bi->pgno;
/* Lose the currently pinned page. */
mpool_put(t->bt_mp, h, 0);
/* Get the next level down. */
if ((h = mpool_get(t->bt_mp, pgno, 0)) == NULL)
return (1);
idx = NEXTINDEX(h) - 1;
BT_PUSH(t, pgno, idx);
}
mpool_put(t->bt_mp, h, 0);
if ((h = mpool_get(t->bt_mp, prevpg, 0)) == NULL)
return (1);
}
ret: mpool_put(t->bt_mp, h, 0);
return ((*hp = mpool_get(t->bt_mp, c->pg.pgno, 0)) == NULL);
}
/*
* __bt_seqadvance --
* Advance the sequential scan.
*
* Parameters:
* t: tree
* flags: R_NEXT, R_PREV
*
* Side effects:
* Pins the page the new key/data record is on.
*
* Returns:
* RET_ERROR, RET_SUCCESS or RET_SPECIAL if there's no next key.
*/
static int
__bt_seqadv(BTREE *t, EPG *ep, int flags)
{
CURSOR *c;
PAGE *h;
indx_t idx = 0; /* pacify gcc */
pgno_t pg;
int exact;
/*
* There are a couple of states that we can be in. The cursor has
* been initialized by the time we get here, but that's all we know.
*/
c = &t->bt_cursor;
/*
* The cursor was deleted where there weren't any duplicate records,
* so the key was saved. Find out where that key would go in the
* current tree. It doesn't matter if the returned key is an exact
* match or not -- if it's an exact match, the record was added after
* the delete so we can just return it. If not, as long as there's
* a record there, return it.
*/
if (F_ISSET(c, CURS_ACQUIRE))
return (__bt_first(t, &c->key, ep, &exact));
/* Get the page referenced by the cursor. */
if ((h = mpool_get(t->bt_mp, c->pg.pgno, 0)) == NULL)
return (RET_ERROR);
/*
* Find the next/previous record in the tree and point the cursor at
* it. The cursor may not be moved until a new key has been found.
*/
switch (flags) {
case R_NEXT: /* Next record. */
/*
* The cursor was deleted in duplicate records, and moved
* forward to a record that has yet to be returned. Clear
* that flag, and return the record.
*/
if (F_ISSET(c, CURS_AFTER))
goto usecurrent;
idx = c->pg.index;
if (++idx == NEXTINDEX(h)) {
pg = h->nextpg;
mpool_put(t->bt_mp, h, 0);
if (pg == P_INVALID)
return (RET_SPECIAL);
if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
return (RET_ERROR);
idx = 0;
}
break;
case R_PREV: /* Previous record. */
/*
* The cursor was deleted in duplicate records, and moved
* backward to a record that has yet to be returned. Clear
* that flag, and return the record.
*/
if (F_ISSET(c, CURS_BEFORE)) {
usecurrent: F_CLR(c, CURS_AFTER | CURS_BEFORE);
ep->page = h;
ep->index = c->pg.index;
return (RET_SUCCESS);
}
idx = c->pg.index;
if (idx == 0) {
pg = h->prevpg;
mpool_put(t->bt_mp, h, 0);
if (pg == P_INVALID)
return (RET_SPECIAL);
if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
return (RET_ERROR);
idx = NEXTINDEX(h) - 1;
} else
--idx;
break;
}
ep->page = h;
ep->index = idx;
return (RET_SUCCESS);
}
/*
* __bt_search --
* Search a btree for a key.
*
* Parameters:
* t: tree to search
* key: key to find
* exactp: pointer to exact match flag
*
* Returns:
* The EPG for matching record, if any, or the EPG for the location
* of the key, if it were inserted into the tree, is entered into
* the bt_cur field of the tree. A pointer to the field is returned.
*/
EPG *
__bt_search(BTREE *t, const DBT *key, int *exactp)
{
PAGE *h;
indx_t base, idx, lim;
pgno_t pg;
int cmp;
BT_CLR(t);
for (pg = P_ROOT;;) {
if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
return (NULL);
/* Do a binary search on the current page. */
t->bt_cur.page = h;
for (base = 0, lim = NEXTINDEX(h); lim; lim >>= 1) {
t->bt_cur.index = idx = base + (lim >> 1);
if ((cmp = __bt_cmp(t, key, &t->bt_cur)) == 0) {
if (h->flags & P_BLEAF) {
*exactp = 1;
return (&t->bt_cur);
}
goto next;
}
if (cmp > 0) {
base = idx + 1;
--lim;
}
}
/*
* If it's a leaf page, we're almost done. If no duplicates
* are allowed, or we have an exact match, we're done. Else,
* it's possible that there were matching keys on this page,
* which later deleted, and we're on a page with no matches
* while there are matches on other pages. If at the start or
* end of a page, check the adjacent page.
*/
if (h->flags & P_BLEAF) {
if (!F_ISSET(t, B_NODUPS)) {
if (base == 0 &&
h->prevpg != P_INVALID &&
__bt_sprev(t, h, key, exactp))
return (&t->bt_cur);
if (base == NEXTINDEX(h) &&
h->nextpg != P_INVALID &&
__bt_snext(t, h, key, exactp))
return (&t->bt_cur);
}
*exactp = 0;
t->bt_cur.index = base;
return (&t->bt_cur);
}
/*
* No match found. Base is the smallest index greater than
* key and may be zero or a last + 1 index. If it's non-zero,
* decrement by one, and record the internal page which should
* be a parent page for the key. If a split later occurs, the
* inserted page will be to the right of the saved page.
*/
idx = base ? base - 1 : base;
next: BT_PUSH(t, h->pgno, idx);
pg = GETBINTERNAL(h, idx)->pgno;
mpool_put(t->bt_mp, h, 0);
}
}
/*
* __BT_SPLIT -- Split the tree.
*
* Parameters:
* t: tree
* sp: page to split
* key: key to insert
* data: data to insert
* flags: BIGKEY/BIGDATA flags
* ilen: insert length
* skip: index to leave open
*
* Returns:
* RET_ERROR, RET_SUCCESS
*/
int
__bt_split(BTREE *t, PAGE *sp, const DBT *key, const DBT *data, int flags,
size_t ilen, u_int32_t argskip)
{
BINTERNAL *bi;
BLEAF *bl, *tbl;
DBT a, b;
EPGNO *parent;
PAGE *h, *l, *r, *lchild, *rchild;
indx_t nxtindex;
u_int16_t skip;
u_int32_t n, nbytes, nksize;
int parentsplit;
char *dest;
/*
* Split the page into two pages, l and r. The split routines return
* a pointer to the page into which the key should be inserted and with
* skip set to the offset which should be used. Additionally, l and r
* are pinned.
*/
skip = argskip;
h = sp->pgno == P_ROOT ?
bt_root(t, sp, &l, &r, &skip, ilen) :
bt_page(t, sp, &l, &r, &skip, ilen);
if (h == NULL)
return (RET_ERROR);
/*
* Insert the new key/data pair into the leaf page. (Key inserts
* always cause a leaf page to split first.)
*/
h->linp[skip] = h->upper -= ilen;
dest = (char *)h + h->upper;
if (F_ISSET(t, R_RECNO))
WR_RLEAF(dest, data, flags)
else
WR_BLEAF(dest, key, data, flags)
/* If the root page was split, make it look right. */
if (sp->pgno == P_ROOT &&
(F_ISSET(t, R_RECNO) ?
bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
goto err2;
/*
* Now we walk the parent page stack -- a LIFO stack of the pages that
* were traversed when we searched for the page that split. Each stack
* entry is a page number and a page index offset. The offset is for
* the page traversed on the search. We've just split a page, so we
* have to insert a new key into the parent page.
*
* If the insert into the parent page causes it to split, may have to
* continue splitting all the way up the tree. We stop if the root
* splits or the page inserted into didn't have to split to hold the
* new key. Some algorithms replace the key for the old page as well
* as the new page. We don't, as there's no reason to believe that the
* first key on the old page is any better than the key we have, and,
* in the case of a key being placed at index 0 causing the split, the
* key is unavailable.
*
* There are a maximum of 5 pages pinned at any time. We keep the left
* and right pages pinned while working on the parent. The 5 are the
* two children, left parent and right parent (when the parent splits)
* and the root page or the overflow key page when calling bt_preserve.
* This code must make sure that all pins are released other than the
* root page or overflow page which is unlocked elsewhere.
*/
while ((parent = BT_POP(t)) != NULL) {
lchild = l;
rchild = r;
/* Get the parent page. */
if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
goto err2;
/*
* The new key goes ONE AFTER the index, because the split
* was to the right.
*/
skip = parent->index + 1;
/*
* Calculate the space needed on the parent page.
*
* Prefix trees: space hack when inserting into BINTERNAL
* pages. Retain only what's needed to distinguish between
* the new entry and the LAST entry on the page to its left.
* If the keys compare equal, retain the entire key. Note,
* we don't touch overflow keys, and the entire key must be
* retained for the next-to-left most key on the leftmost
* page of each level, or the search will fail. Applicable
* ONLY to internal pages that have leaf pages as children.
* Further reduction of the key between pairs of internal
* pages loses too much information.
*/
switch (rchild->flags & P_TYPE) {
case P_BINTERNAL:
bi = GETBINTERNAL(rchild, 0);
nbytes = NBINTERNAL(bi->ksize);
break;
case P_BLEAF:
bl = GETBLEAF(rchild, 0);
nbytes = NBINTERNAL(bl->ksize);
if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
(h->prevpg != P_INVALID || skip > 1)) {
tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
a.size = tbl->ksize;
a.data = tbl->bytes;
b.size = bl->ksize;
b.data = bl->bytes;
nksize = t->bt_pfx(&a, &b);
n = NBINTERNAL(nksize);
if (n < nbytes) {
#ifdef STATISTICS
bt_pfxsaved += nbytes - n;
#endif
nbytes = n;
} else
nksize = 0;
} else
nksize = 0;
break;
case P_RINTERNAL:
case P_RLEAF:
nbytes = NRINTERNAL;
break;
default:
abort();
}
/* Split the parent page if necessary or shift the indices. */
if ((u_int32_t)(h->upper - h->lower) < nbytes + sizeof(indx_t)) {
sp = h;
h = h->pgno == P_ROOT ?
bt_root(t, h, &l, &r, &skip, nbytes) :
bt_page(t, h, &l, &r, &skip, nbytes);
if (h == NULL)
goto err1;
parentsplit = 1;
} else {
if (skip < (nxtindex = NEXTINDEX(h)))
memmove(h->linp + skip + 1, h->linp + skip,
(nxtindex - skip) * sizeof(indx_t));
h->lower += sizeof(indx_t);
parentsplit = 0;
}
/* Insert the key into the parent page. */
switch (rchild->flags & P_TYPE) {
case P_BINTERNAL:
h->linp[skip] = h->upper -= nbytes;
dest = (char *)h + h->linp[skip];
memmove(dest, bi, nbytes);
((BINTERNAL *)dest)->pgno = rchild->pgno;
break;
case P_BLEAF:
h->linp[skip] = h->upper -= nbytes;
dest = (char *)h + h->linp[skip];
WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
rchild->pgno, bl->flags & P_BIGKEY);
memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
if (bl->flags & P_BIGKEY) {
pgno_t pgno;
memcpy(&pgno, bl->bytes, sizeof(pgno));
if (bt_preserve(t, pgno) == RET_ERROR)
goto err1;
}
break;
case P_RINTERNAL:
/*
* Update the left page count. If split
* added at index 0, fix the correct page.
*/
if (skip > 0)
dest = (char *)h + h->linp[skip - 1];
else
dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
((RINTERNAL *)dest)->nrecs = rec_total(lchild);
((RINTERNAL *)dest)->pgno = lchild->pgno;
/* Update the right page count. */
h->linp[skip] = h->upper -= nbytes;
dest = (char *)h + h->linp[skip];
((RINTERNAL *)dest)->nrecs = rec_total(rchild);
((RINTERNAL *)dest)->pgno = rchild->pgno;
break;
case P_RLEAF:
/*
* Update the left page count. If split
* added at index 0, fix the correct page.
*/
if (skip > 0)
dest = (char *)h + h->linp[skip - 1];
else
dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild);
((RINTERNAL *)dest)->pgno = lchild->pgno;
/* Update the right page count. */
h->linp[skip] = h->upper -= nbytes;
dest = (char *)h + h->linp[skip];
((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild);
((RINTERNAL *)dest)->pgno = rchild->pgno;
break;
default:
abort();
}
/* Unpin the held pages. */
if (!parentsplit) {
mpool_put(t->bt_mp, h, MPOOL_DIRTY);
break;
}
/* If the root page was split, make it look right. */
if (sp->pgno == P_ROOT &&
(F_ISSET(t, R_RECNO) ?
bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
goto err1;
mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
}
/* Unpin the held pages. */
mpool_put(t->bt_mp, l, MPOOL_DIRTY);
mpool_put(t->bt_mp, r, MPOOL_DIRTY);
/* Clear any pages left on the stack. */
return (RET_SUCCESS);
/*
* If something fails in the above loop we were already walking back
* up the tree and the tree is now inconsistent. Nothing much we can
* do about it but release any memory we're holding.
*/
err1: mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
err2: mpool_put(t->bt_mp, l, 0);
mpool_put(t->bt_mp, r, 0);
__dbpanic(t->bt_dbp);
return (RET_ERROR);
}
/*
* BT_PSPLIT -- Do the real work of splitting the page.
*
* Parameters:
* t: tree
* h: page to be split
* l: page to put lower half of data
* r: page to put upper half of data
* pskip: pointer to index to leave open
* ilen: insert length
*
* Returns:
* Pointer to page in which to insert.
*/
static PAGE *
bt_psplit(BTREE *t, PAGE *h, PAGE *l, PAGE *r, indx_t *pskip, size_t ilen)
{
BINTERNAL *bi;
BLEAF *bl;
CURSOR *c;
RLEAF *rl;
PAGE *rval;
void *src;
indx_t full, half, nxt, off, skip, top, used;
u_int32_t nbytes;
int bigkeycnt, isbigkey;
/*
* Split the data to the left and right pages. Leave the skip index
* open. Additionally, make some effort not to split on an overflow
* key. This makes internal page processing faster and can save
* space as overflow keys used by internal pages are never deleted.
*/
bigkeycnt = 0;
skip = *pskip;
full = t->bt_psize - BTDATAOFF;
half = full / 2;
used = 0;
for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
if (skip == off) {
nbytes = ilen;
isbigkey = 0; /* XXX: not really known. */
} else
switch (h->flags & P_TYPE) {
case P_BINTERNAL:
src = bi = GETBINTERNAL(h, nxt);
nbytes = NBINTERNAL(bi->ksize);
isbigkey = bi->flags & P_BIGKEY;
break;
case P_BLEAF:
src = bl = GETBLEAF(h, nxt);
nbytes = NBLEAF(bl);
isbigkey = bl->flags & P_BIGKEY;
break;
case P_RINTERNAL:
src = GETRINTERNAL(h, nxt);
nbytes = NRINTERNAL;
isbigkey = 0;
break;
case P_RLEAF:
src = rl = GETRLEAF(h, nxt);
nbytes = NRLEAF(rl);
isbigkey = 0;
break;
default:
abort();
}
/*
* If the key/data pairs are substantial fractions of the max
* possible size for the page, it's possible to get situations
* where we decide to try and copy too much onto the left page.
* Make sure that doesn't happen.
*/
if ((skip <= off && used + nbytes + sizeof(indx_t) >= full) ||
nxt == top - 1) {
--off;
break;
}
/* Copy the key/data pair, if not the skipped index. */
if (skip != off) {
++nxt;
l->linp[off] = l->upper -= nbytes;
memmove((char *)l + l->upper, src, nbytes);
}
used += nbytes + sizeof(indx_t);
if (used >= half) {
if (!isbigkey || bigkeycnt == 3)
break;
else
++bigkeycnt;
}
}
/*
* Off is the last offset that's valid for the left page.
* Nxt is the first offset to be placed on the right page.
*/
l->lower += (off + 1) * sizeof(indx_t);
/*
* If splitting the page that the cursor was on, the cursor has to be
* adjusted to point to the same record as before the split. If the
* cursor is at or past the skipped slot, the cursor is incremented by
* one. If the cursor is on the right page, it is decremented by the
* number of records split to the left page.
*/
c = &t->bt_cursor;
if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
if (c->pg.index >= skip)
++c->pg.index;
if (c->pg.index < nxt) /* Left page. */
c->pg.pgno = l->pgno;
else { /* Right page. */
c->pg.pgno = r->pgno;
c->pg.index -= nxt;
}
}
/*
* If the skipped index was on the left page, just return that page.
* Otherwise, adjust the skip index to reflect the new position on
* the right page.
*/
if (skip <= off) {
skip = MAX_PAGE_OFFSET;
rval = l;
} else {
rval = r;
*pskip -= nxt;
}
for (off = 0; nxt < top; ++off) {
if (skip == nxt) {
++off;
skip = MAX_PAGE_OFFSET;
}
switch (h->flags & P_TYPE) {
case P_BINTERNAL:
src = bi = GETBINTERNAL(h, nxt);
nbytes = NBINTERNAL(bi->ksize);
break;
case P_BLEAF:
src = bl = GETBLEAF(h, nxt);
nbytes = NBLEAF(bl);
break;
case P_RINTERNAL:
src = GETRINTERNAL(h, nxt);
nbytes = NRINTERNAL;
break;
case P_RLEAF:
src = rl = GETRLEAF(h, nxt);
nbytes = NRLEAF(rl);
break;
default:
abort();
}
++nxt;
r->linp[off] = r->upper -= nbytes;
memmove((char *)r + r->upper, src, nbytes);
}
r->lower += off * sizeof(indx_t);
/* If the key is being appended to the page, adjust the index. */
if (skip == top)
r->lower += sizeof(indx_t);
return (rval);
}
/*
* __bt_search --
* Search a btree for a key.
*
* Parameters:
* t: tree to search
* key: key to find
* exactp: pointer to exact match flag
*
* Returns:
* The EPG for matching record, if any, or the EPG for the location
* of the key, if it were inserted into the tree, is entered into
* the bt_cur field of the tree. A pointer to the field is returned.
*/
EPG *
__bt_search_st(BTREE *t,const DBT *key,int *exactp)
{
/*
* 1.Read from root of the btree in NTT
* 2.reconstruct the node
*/
PAGE *h=NULL; /* h is a logical B-Tree node, either a disk mode node or a virtual node in memory construct by log */
indx_t base, index, lim;
pgno_t pg; //node id of the page
int cmp;
BT_CLR(t); /* @mx it initializes t->bt_sp */
err_debug(("Searh Btree"));
for (pg = P_ROOT;;) {
err_debug(("~^"));
err_debug(("Read Node %ud",pg));
h = read_node(t,pg);
//__bt_dpage(h);
err_debug(("~$End Read"));
if(h==NULL)
return (NULL);
/* ??? not so clear about the binary search */
/* Do a binary search on the current page. */
t->bt_cur.page = h;
for (base = 0, lim = NEXTINDEX(h); lim; lim >>= 1) {
t->bt_cur.index = index = base + (lim >> 1);
if ((cmp = __bt_cmp(t, key, &t->bt_cur)) == 0) {
if (h->flags & P_BLEAF) {
*exactp = 1;
err_debug(("End Search"));
return (&t->bt_cur);
}
goto next;
}
if (cmp > 0) {
base = index + 1;
--lim;
}
}
/*
* If it's a leaf page, we're almost done. If no duplicates
* are allowed, or we have an exact match, we're done. Else,
* it's possible that there were matching keys on this page,
* which later deleted, and we're on a page with no matches
* while there are matches on other pages. If at the start or
* end of a page, check the adjacent page.
*
* TODO: what about this condition for log mode ?
*/
if (h->flags & P_BLEAF) {
#if 0
if (!F_ISSET(t, B_NODUPS)) {
if (base == 0 &&
h->prevpg != P_INVALID &&
__bt_sprev(t, h, key, exactp))
err_debug(("End Search"));
return (&t->bt_cur);
if (base == NEXTINDEX(h) &&
h->nextpg != P_INVALID &&
__bt_snext(t, h, key, exactp))
err_debug(("End Search\n"));
return (&t->bt_cur);
}
#endif
*exactp = 0;
t->bt_cur.index = base;
err_debug(("End Search"));
return (&t->bt_cur);
}
/*
* No match found. Base is the smallest index greater than
* key and may be zero or a last + 1 index. If it's non-zero,
* decrement by one, and record the internal page which should
* be a parent page for the key. If a split later occurs, the
* inserted page will be to the right of the saved page.
*/
index = base ? base - 1 : base;
next: BT_PUSH(t, pg, index);
pg = GETBINTERNAL(h, index)->pgno;
Mpool_put(t->bt_mp, h, 0);
}
}
/*
* BT_PAGE -- Split a non-root page of a btree.
*
* Parameters:
* t: tree
* h: root page
* lp: pointer to left page pointer
* rp: pointer to right page pointer
* skip: pointer to index to leave open
* ilen: insert length
*
* Returns:
* Pointer to page in which to insert or NULL on error.
*/
static PAGE *
bt_page(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
{
PAGE *l, *r, *tp;
pgno_t npg;
#ifdef STATISTICS
++bt_split;
#endif
/* Put the new right page for the split into place. */
if ((r = __bt_new(t, &npg)) == NULL)
return (NULL);
r->pgno = npg;
r->lower = BTDATAOFF;
r->upper = t->bt_psize;
r->nextpg = h->nextpg;
r->prevpg = h->pgno;
r->flags = h->flags & P_TYPE;
/*
* If we're splitting the last page on a level because we're appending
* a key to it (skip is NEXTINDEX()), it's likely that the data is
* sorted. Adding an empty page on the side of the level is less work
* and can push the fill factor much higher than normal. If we're
* wrong it's no big deal, we'll just do the split the right way next
* time. It may look like it's equally easy to do a similar hack for
* reverse sorted data, that is, split the tree left, but it's not.
* Don't even try.
*/
if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
#ifdef STATISTICS
++bt_sortsplit;
#endif
h->nextpg = r->pgno;
r->lower = BTDATAOFF + sizeof(indx_t);
*skip = 0;
*lp = h;
*rp = r;
return (r);
}
/* Put the new left page for the split into place. */
if ((l = (PAGE *)calloc(1, t->bt_psize)) == NULL) {
mpool_put(t->bt_mp, r, 0);
return (NULL);
}
l->pgno = h->pgno;
l->nextpg = r->pgno;
l->prevpg = h->prevpg;
l->lower = BTDATAOFF;
l->upper = t->bt_psize;
l->flags = h->flags & P_TYPE;
/* Fix up the previous pointer of the page after the split page. */
if (h->nextpg != P_INVALID) {
if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
free(l);
/* XXX mpool_free(t->bt_mp, r->pgno); */
return (NULL);
}
tp->prevpg = r->pgno;
mpool_put(t->bt_mp, tp, MPOOL_DIRTY);
}
/*
* Split right. The key/data pairs aren't sorted in the btree page so
* it's simpler to copy the data from the split page onto two new pages
* instead of copying half the data to the right page and compacting
* the left page in place. Since the left page can't change, we have
* to swap the original and the allocated left page after the split.
*/
tp = bt_psplit(t, h, l, r, skip, ilen);
/* Move the new left page onto the old left page. */
memmove(h, l, t->bt_psize);
if (tp == l)
tp = h;
free(l);
*lp = h;
*rp = r;
return (tp);
}
/*
* BT_STAT -- Gather/print the tree statistics
*
* Parameters:
* dbp: pointer to the DB
*/
void
__bt_stat(DB *dbp)
{
extern u_long bt_cache_hit, bt_cache_miss, bt_pfxsaved, bt_rootsplit;
extern u_long bt_sortsplit, bt_split;
BTREE *t;
PAGE *h;
pgno_t i, pcont, pinternal, pleaf;
u_long ifree, lfree, nkeys;
int levels;
t = dbp->internal;
pcont = pinternal = pleaf = 0;
nkeys = ifree = lfree = 0;
for (i = P_ROOT;
(h = mpool_get(t->bt_mp, i, MPOOL_IGNOREPIN)) != NULL; ++i)
switch (h->flags & P_TYPE) {
case P_BINTERNAL:
case P_RINTERNAL:
++pinternal;
ifree += h->upper - h->lower;
break;
case P_BLEAF:
case P_RLEAF:
++pleaf;
lfree += h->upper - h->lower;
nkeys += NEXTINDEX(h);
break;
case P_OVERFLOW:
++pcont;
break;
}
/* Count the levels of the tree. */
for (i = P_ROOT, levels = 0 ;; ++levels) {
h = mpool_get(t->bt_mp, i, MPOOL_IGNOREPIN);
if (h->flags & (P_BLEAF|P_RLEAF)) {
if (levels == 0)
levels = 1;
break;
}
i = F_ISSET(t, R_RECNO) ?
GETRINTERNAL(h, 0)->pgno :
GETBINTERNAL(h, 0)->pgno;
}
(void)fprintf(stderr, "%d level%s with %lu keys",
levels, levels == 1 ? "" : "s", nkeys);
if (F_ISSET(t, R_RECNO))
(void)fprintf(stderr, " (%u header count)", t->bt_nrecs);
(void)fprintf(stderr,
"\n%u pages (leaf %u, internal %u, overflow %u)\n",
pinternal + pleaf + pcont, pleaf, pinternal, pcont);
(void)fprintf(stderr, "%lu cache hits, %lu cache misses\n",
bt_cache_hit, bt_cache_miss);
(void)fprintf(stderr, "%lu splits (%lu root splits, %lu sort splits)\n",
bt_split, bt_rootsplit, bt_sortsplit);
pleaf *= t->bt_psize - BTDATAOFF;
if (pleaf)
(void)fprintf(stderr,
"%.0f%% leaf fill (%lu bytes used, %lu bytes free)\n",
((double)(pleaf - lfree) / pleaf) * 100,
pleaf - lfree, lfree);
pinternal *= t->bt_psize - BTDATAOFF;
if (pinternal)
(void)fprintf(stderr,
"%.0f%% internal fill (%lu bytes used, %lu bytes free\n",
((double)(pinternal - ifree) / pinternal) * 100,
pinternal - ifree, ifree);
if (bt_pfxsaved)
(void)fprintf(stderr, "prefix checking removed %lu bytes.\n",
bt_pfxsaved);
}
/*
* __bt_first --
* Find the first entry.
*
* Parameters:
* t: the tree
* key: the key
* erval: return EPG
* exactp: pointer to exact match flag
*
* Returns:
* The first entry in the tree greater than or equal to key,
* or RET_SPECIAL if no such key exists.
*/
static int
__bt_first(BTREE *t, const DBT *key, EPG *erval, int *exactp)
{
PAGE *h;
EPG *ep, save;
pgno_t pg;
/*
* Find any matching record; __bt_search pins the page.
*
* If it's an exact match and duplicates are possible, walk backwards
* in the tree until we find the first one. Otherwise, make sure it's
* a valid key (__bt_search may return an index just past the end of a
* page) and return it.
*/
if ((ep = __bt_search(t, key, exactp)) == NULL)
return (0);
if (*exactp) {
if (F_ISSET(t, B_NODUPS)) {
*erval = *ep;
return (RET_SUCCESS);
}
/*
* Walk backwards, as long as the entry matches and there are
* keys left in the tree. Save a copy of each match in case
* we go too far.
*/
save = *ep;
h = ep->page;
do {
if (save.page->pgno != ep->page->pgno) {
mpool_put(t->bt_mp, save.page, 0);
save = *ep;
} else
save.index = ep->index;
/*
* Don't unpin the page the last (or original) match
* was on, but make sure it's unpinned if an error
* occurs.
*/
if (ep->index == 0) {
if (h->prevpg == P_INVALID)
break;
if (h->pgno != save.page->pgno)
mpool_put(t->bt_mp, h, 0);
if ((h = mpool_get(t->bt_mp,
h->prevpg, 0)) == NULL)
return (RET_ERROR);
ep->page = h;
ep->index = NEXTINDEX(h);
}
--ep->index;
} while (__bt_cmp(t, key, ep) == 0);
/*
* Reach here with the last page that was looked at pinned,
* which may or may not be the same as the last (or original)
* match page. If it's not useful, release it.
*/
if (h->pgno != save.page->pgno)
mpool_put(t->bt_mp, h, 0);
*erval = save;
return (RET_SUCCESS);
}
/* If at the end of a page, find the next entry. */
if (ep->index == NEXTINDEX(ep->page)) {
h = ep->page;
pg = h->nextpg;
mpool_put(t->bt_mp, h, 0);
if (pg == P_INVALID)
return (RET_SPECIAL);
if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
return (RET_ERROR);
ep->index = 0;
ep->page = h;
}
*erval = *ep;
return (RET_SUCCESS);
}
/*
* __BT_BPGIN, __BT_BPGOUT --
* Convert host-specific number layout to/from the host-independent
* format stored on disk.
*
* Parameters:
* t: tree
* pg: page number
* h: page to convert
*/
void
__bt_pgin(void *t, pgno_t pg, void *pp)
{
PAGE *h;
indx_t i, top;
u_char flags;
char *p;
if (!F_ISSET(((BTREE *)t), B_NEEDSWAP))
return;
if (pg == P_META) {
mswap(pp);
return;
}
h = pp;
M_32_SWAP(h->pgno);
M_32_SWAP(h->prevpg);
M_32_SWAP(h->nextpg);
M_32_SWAP(h->flags);
M_16_SWAP(h->lower);
M_16_SWAP(h->upper);
top = NEXTINDEX(h);
if ((h->flags & P_TYPE) == P_BINTERNAL)
for (i = 0; i < top; i++) {
M_16_SWAP(h->linp[i]);
p = (char *)GETBINTERNAL(h, i);
P_32_SWAP(p);
p += sizeof(u_int32_t);
P_32_SWAP(p);
p += sizeof(pgno_t);
if (*(u_char *)p & P_BIGKEY) {
p += sizeof(u_char);
P_32_SWAP(p);
p += sizeof(pgno_t);
P_32_SWAP(p);
}
}
else if ((h->flags & P_TYPE) == P_BLEAF)
for (i = 0; i < top; i++) {
M_16_SWAP(h->linp[i]);
p = (char *)GETBLEAF(h, i);
P_32_SWAP(p);
p += sizeof(u_int32_t);
P_32_SWAP(p);
p += sizeof(u_int32_t);
flags = *(u_char *)p;
if (flags & (P_BIGKEY | P_BIGDATA)) {
p += sizeof(u_char);
if (flags & P_BIGKEY) {
P_32_SWAP(p);
p += sizeof(pgno_t);
P_32_SWAP(p);
}
if (flags & P_BIGDATA) {
p += sizeof(u_int32_t);
P_32_SWAP(p);
p += sizeof(pgno_t);
P_32_SWAP(p);
}
}
}
}
