/*
* __curjoin_iter_bump --
* Called to advance the iterator to the next endpoint, which may in turn
* advance to the next entry.
*/
static int
__curjoin_iter_bump(WT_CURSOR_JOIN_ITER *iter)
{
WT_CURSOR_JOIN_ENTRY *entry;
WT_SESSION_IMPL *session;
session = iter->session;
iter->positioned = false;
entry = iter->entry;
if (entry->subjoin == NULL && iter->is_equal &&
++iter->end_pos < iter->end_count) {
WT_RET(__wt_cursor_dup_position(
entry->ends[iter->end_pos].cursor, iter->cursor));
return (0);
}
iter->end_pos = iter->end_count = iter->end_skip = 0;
if (entry->subjoin != NULL && entry->subjoin->iter != NULL)
WT_RET(__curjoin_iter_close_all(entry->subjoin->iter));
if (++iter->entry_pos >= iter->entry_count) {
iter->entry = NULL;
return (0);
}
iter->entry = ++entry;
if (entry->subjoin != NULL) {
WT_RET(__curjoin_iter_init(session, entry->subjoin,
&iter->child));
return (0);
}
WT_RET(__curjoin_iter_set_entry(iter, iter->entry_pos));
return (0);
}
/*
* __curjoin_entry_iter_init --
* Initialize an iteration for the index managed by a join entry.
*
*/
static int
__curjoin_entry_iter_init(WT_SESSION_IMPL *session, WT_CURSOR_JOIN *cjoin,
WT_CURSOR_JOIN_ENTRY *entry, WT_CURSOR_JOIN_ITER **iterp)
{
WT_CURSOR *newcur;
WT_CURSOR *to_dup;
WT_DECL_RET;
const char *raw_cfg[] = { WT_CONFIG_BASE(
session, WT_SESSION_open_cursor), "raw", NULL };
const char *def_cfg[] = { WT_CONFIG_BASE(
session, WT_SESSION_open_cursor), NULL };
const char *uri, **config;
char *uribuf;
WT_CURSOR_JOIN_ITER *iter;
size_t size;
iter = NULL;
uribuf = NULL;
to_dup = entry->ends[0].cursor;
uri = to_dup->uri;
if (F_ISSET((WT_CURSOR *)cjoin, WT_CURSTD_RAW))
config = &raw_cfg[0];
else
config = &def_cfg[0];
if (cjoin->projection != NULL) {
size = strlen(uri) + strlen(cjoin->projection) + 1;
WT_ERR(__wt_calloc(session, size, 1, &uribuf));
snprintf(uribuf, size, "%s%s", uri, cjoin->projection);
uri = uribuf;
}
WT_ERR(__wt_open_cursor(session, uri, (WT_CURSOR *)cjoin, config,
&newcur));
WT_ERR(__wt_cursor_dup_position(to_dup, newcur));
WT_ERR(__wt_calloc_one(session, &iter));
iter->cjoin = cjoin;
iter->session = session;
iter->entry = entry;
iter->cursor = newcur;
iter->advance = false;
*iterp = iter;
if (0) {
err: __wt_free(session, iter);
}
__wt_free(session, uribuf);
return (ret);
}
/*
* __curjoin_entry_iter_reset --
* Reset an iteration to the starting point.
*
*/
static int
__curjoin_entry_iter_reset(WT_CURSOR_JOIN_ITER *iter)
{
WT_DECL_RET;
if (iter->advance) {
WT_ERR(iter->cursor->reset(iter->cursor));
WT_ERR(__wt_cursor_dup_position(
iter->cjoin->entries[0].ends[0].cursor, iter->cursor));
iter->advance = false;
iter->entry->stats.actual_count = 0;
}
err: return (ret);
}
/*
* __curjoin_init_bloom --
* Populate Bloom filters
*/
static int
__curjoin_init_bloom(WT_SESSION_IMPL *session, WT_CURSOR_JOIN *cjoin,
WT_CURSOR_JOIN_ENTRY *entry, WT_BLOOM *bloom)
{
WT_COLLATOR *collator;
WT_CURSOR *c;
WT_CURSOR_JOIN_ENDPOINT *end, *endmax;
WT_DECL_ITEM(uribuf);
WT_DECL_RET;
WT_ITEM curkey, curvalue;
size_t size;
u_int skip;
int cmp;
const char *uri;
const char *raw_cfg[] = { WT_CONFIG_BASE(
session, WT_SESSION_open_cursor), "raw", NULL };
c = NULL;
skip = 0;
if (entry->index != NULL)
/*
* Open the raw index. We're avoiding any references
* to the main table, they may be expensive.
*/
uri = entry->index->source;
else {
/*
* For joins on the main table, we just need the primary
* key for comparison, we don't need any values.
*/
size = strlen(cjoin->table->iface.name) + 3;
WT_ERR(__wt_scr_alloc(session, size, &uribuf));
WT_ERR(__wt_buf_fmt(session, uribuf, "%s()",
cjoin->table->iface.name));
uri = uribuf->data;
}
WT_ERR(__wt_open_cursor(session, uri, &cjoin->iface, raw_cfg, &c));
/* Initially position the cursor if necessary. */
endmax = &entry->ends[entry->ends_next];
if ((end = &entry->ends[0]) < endmax) {
if (F_ISSET(end, WT_CURJOIN_END_GT) ||
WT_CURJOIN_END_RANGE(end) == WT_CURJOIN_END_EQ) {
WT_ERR(__wt_cursor_dup_position(end->cursor, c));
if (WT_CURJOIN_END_RANGE(end) == WT_CURJOIN_END_GE)
skip = 1;
} else if (F_ISSET(end, WT_CURJOIN_END_LT)) {
if ((ret = c->next(c)) == WT_NOTFOUND)
goto done;
WT_ERR(ret);
} else
WT_PANIC_ERR(session, EINVAL,
"fatal error in join cursor position state");
}
collator = (entry->index == NULL) ? NULL : entry->index->collator;
while (ret == 0) {
WT_ERR(c->get_key(c, &curkey));
entry->stats.iterated++;
if (entry->index != NULL) {
/*
* Repack so it's comparable to the
* reference endpoints.
*/
WT_ERR(__wt_struct_repack(session,
c->key_format,
(entry->repack_format != NULL ?
entry->repack_format : entry->index->idxkey_format),
&c->key, &curkey));
}
for (end = &entry->ends[skip]; end < endmax; end++) {
WT_ERR(__wt_compare(session, collator, &curkey,
&end->key, &cmp));
if (F_ISSET(entry, WT_CURJOIN_ENTRY_DISJUNCTION)) {
/* if condition satisfied, insert immediately */
switch (WT_CURJOIN_END_RANGE(end)) {
case WT_CURJOIN_END_EQ:
if (cmp == 0)
goto insert;
break;
case WT_CURJOIN_END_GT:
if (cmp > 0) {
/* skip this check next time */
skip = entry->ends_next;
goto insert;
}
break;
case WT_CURJOIN_END_GE:
if (cmp >= 0)
goto insert;
break;
case WT_CURJOIN_END_LT:
if (cmp < 0)
goto insert;
break;
case WT_CURJOIN_END_LE:
if (cmp <= 0)
goto insert;
break;
}
} else if (!F_ISSET(end, WT_CURJOIN_END_LT)) {
if (cmp < 0 || (cmp == 0 &&
!F_ISSET(end, WT_CURJOIN_END_EQ)))
goto advance;
if (cmp > 0) {
if (F_ISSET(end, WT_CURJOIN_END_GT))
skip = 1;
else
goto done;
}
} else {
if (cmp > 0 || (cmp == 0 &&
!F_ISSET(end, WT_CURJOIN_END_EQ)))
goto done;
}
}
/*
* Either it's a disjunction that hasn't satisfied any
* condition, or it's a conjunction that has satisfied all
* conditions.
*/
if (F_ISSET(entry, WT_CURJOIN_ENTRY_DISJUNCTION))
goto advance;
insert:
if (entry->index != NULL) {
curvalue.data =
(unsigned char *)curkey.data + curkey.size;
WT_ASSERT(session, c->key.size > curkey.size);
curvalue.size = c->key.size - curkey.size;
}
else
WT_ERR(c->get_key(c, &curvalue));
__wt_bloom_insert(bloom, &curvalue);
entry->stats.bloom_insert++;
advance:
if ((ret = c->next(c)) == WT_NOTFOUND)
break;
}
done:
WT_ERR_NOTFOUND_OK(ret);
err: if (c != NULL)
WT_TRET(c->close(c));
__wt_scr_free(session, &uribuf);
return (ret);
}
/*
* __curjoin_iter_set_entry --
* Set the current entry for an iterator.
*/
static int
__curjoin_iter_set_entry(WT_CURSOR_JOIN_ITER *iter, u_int entry_pos)
{
WT_CURSOR *c, *to_dup;
WT_CURSOR_JOIN *cjoin, *topjoin;
WT_CURSOR_JOIN_ENTRY *entry;
WT_DECL_RET;
WT_SESSION_IMPL *session;
size_t size;
const char *raw_cfg[] = { WT_CONFIG_BASE(
iter->session, WT_SESSION_open_cursor), "raw", NULL };
const char *def_cfg[] = { WT_CONFIG_BASE(
iter->session, WT_SESSION_open_cursor), NULL };
const char **config;
char *uri;
session = iter->session;
cjoin = iter->cjoin;
uri = NULL;
entry = iter->entry = &cjoin->entries[entry_pos];
iter->positioned = false;
iter->entry_pos = entry_pos;
iter->end_pos = 0;
iter->is_equal = (entry->ends_next == 1 &&
WT_CURJOIN_END_RANGE(&entry->ends[0]) == WT_CURJOIN_END_EQ);
iter->end_skip = (entry->ends_next > 0 &&
WT_CURJOIN_END_RANGE(&entry->ends[0]) == WT_CURJOIN_END_GE) ? 1 : 0;
iter->end_count = WT_MIN(1, entry->ends_next);
if (F_ISSET(cjoin, WT_CURJOIN_DISJUNCTION)) {
iter->entry_count = cjoin->entries_next;
if (iter->is_equal)
iter->end_count = entry->ends_next;
} else
iter->entry_count = 1;
WT_ASSERT(iter->session, iter->entry_pos < iter->entry_count);
entry->stats.iterated = 0;
if (entry->subjoin == NULL) {
for (topjoin = iter->cjoin; topjoin->parent != NULL;
topjoin = topjoin->parent)
;
to_dup = entry->ends[0].cursor;
if (F_ISSET((WT_CURSOR *)topjoin, WT_CURSTD_RAW))
config = &raw_cfg[0];
else
config = &def_cfg[0];
size = strlen(to_dup->internal_uri) + 3;
WT_ERR(__wt_calloc(session, size, 1, &uri));
WT_ERR(__wt_snprintf(uri, size, "%s()", to_dup->internal_uri));
if ((c = iter->cursor) == NULL || strcmp(c->uri, uri) != 0) {
iter->cursor = NULL;
if (c != NULL)
WT_ERR(c->close(c));
WT_ERR(__wt_open_cursor(session, uri,
(WT_CURSOR *)topjoin, config, &iter->cursor));
}
WT_ERR(__wt_cursor_dup_position(to_dup, iter->cursor));
} else if (iter->cursor != NULL) {
WT_ERR(iter->cursor->close(iter->cursor));
iter->cursor = NULL;
}
err: __wt_free(session, uri);
return (ret);
}
/*
* __curjoin_init_bloom --
* Populate Bloom filters
*/
static int
__curjoin_init_bloom(WT_SESSION_IMPL *session, WT_CURSOR_JOIN *cjoin,
WT_CURSOR_JOIN_ENTRY *entry, WT_BLOOM *bloom)
{
WT_COLLATOR *collator;
WT_CURSOR *c;
WT_CURSOR_INDEX *cindex;
WT_CURSOR_JOIN_ENDPOINT *end, *endmax;
WT_DECL_RET;
WT_DECL_ITEM(uribuf);
WT_ITEM curkey, curvalue, *k;
WT_TABLE *maintable;
const char *raw_cfg[] = { WT_CONFIG_BASE(
session, WT_SESSION_open_cursor), "raw", NULL };
const char *mainkey_str, *p;
void *allocbuf;
size_t mainkey_len, size;
u_int i;
int cmp, skip;
c = NULL;
allocbuf = NULL;
skip = 0;
if (entry->index != NULL) {
/*
* Open a cursor having a projection of the keys of the
* index we're comparing against. Open it raw, we're
* going to compare it to the raw keys of the
* reference cursors.
*/
maintable = ((WT_CURSOR_TABLE *)entry->main)->table;
mainkey_str = maintable->colconf.str + 1;
for (p = mainkey_str, i = 0;
p != NULL && i < maintable->nkey_columns; i++)
p = strchr(p + 1, ',');
WT_ASSERT(session, p != 0);
mainkey_len = WT_PTRDIFF(p, mainkey_str);
size = strlen(entry->index->name) + mainkey_len + 3;
WT_ERR(__wt_scr_alloc(session, size, &uribuf));
WT_ERR(__wt_buf_fmt(session, uribuf, "%s(%.*s)",
entry->index->name, (int)mainkey_len, mainkey_str));
} else {
/*
* For joins on the main table, we just need the primary
* key for comparison, we don't need any values.
*/
size = strlen(cjoin->table->name) + 3;
WT_ERR(__wt_scr_alloc(session, size, &uribuf));
WT_ERR(__wt_buf_fmt(session, uribuf, "%s()",
cjoin->table->name));
}
WT_ERR(__wt_open_cursor(
session, uribuf->data, &cjoin->iface, raw_cfg, &c));
/* Initially position the cursor if necessary. */
endmax = &entry->ends[entry->ends_next];
if ((end = &entry->ends[0]) < endmax &&
F_ISSET(end, WT_CURJOIN_END_GE)) {
WT_ERR(__wt_cursor_dup_position(end->cursor, c));
if (end->flags == WT_CURJOIN_END_GE)
skip = 1;
}
collator = (entry->index == NULL) ? NULL : entry->index->collator;
while (ret == 0) {
WT_ERR(c->get_key(c, &curkey));
if (entry->index != NULL) {
cindex = (WT_CURSOR_INDEX *)c;
if (cindex->index->extractor == NULL) {
/*
* Repack so it's comparable to the
* reference endpoints.
*/
k = &cindex->child->key;
WT_ERR(__wt_struct_repack(session,
cindex->child->key_format,
entry->main->value_format, k, &curkey,
&allocbuf));
} else
curkey = cindex->child->key;
}
for (end = &entry->ends[skip]; end < endmax; end++) {
WT_ERR(__wt_compare(session, collator, &curkey,
&end->key, &cmp));
if (!F_ISSET(end, WT_CURJOIN_END_LT)) {
if (cmp < 0 || (cmp == 0 &&
!F_ISSET(end, WT_CURJOIN_END_EQ)))
goto advance;
if (cmp > 0) {
if (F_ISSET(end, WT_CURJOIN_END_GT))
skip = 1;
else
goto done;
}
} else {
if (cmp > 0 || (cmp == 0 &&
!F_ISSET(end, WT_CURJOIN_END_EQ)))
goto done;
}
}
if (entry->index != NULL)
WT_ERR(c->get_value(c, &curvalue));
else
WT_ERR(c->get_key(c, &curvalue));
WT_ERR(__wt_bloom_insert(bloom, &curvalue));
entry->stats.actual_count++;
advance:
if ((ret = c->next(c)) == WT_NOTFOUND)
break;
}
done:
WT_ERR_NOTFOUND_OK(ret);
err: if (c != NULL)
WT_TRET(c->close(c));
__wt_scr_free(session, &uribuf);
__wt_free(session, allocbuf);
return (ret);
}
