We can't just free it, because has the GUID index list as a child, and
these are shared by the new dn list (from the subtransaction we are
committing). But if the dn list is long and the main transaction is
long-lived, we can save a lot of memory by turning this dn list into
an almost empty node in the talloc tree. This returns us to roughly
the situation we had prior to the last commit.
For example, with the repro.sh script on bug 15590 in indexes mode
with 10000 rules, The last 3 commits use this much memory at the end
of an unusually large transaction:
full talloc report on 'struct ldb_context' (total
4012222 bytes in 90058 blocks)
full talloc report on 'struct ldb_context' (total
2405482219 bytes in 90058 blocks)
full talloc report on 'struct ldb_context' (total
4282195 bytes in 90058 blocks)
That is, the last commit increased usage 500 fold, and this commit
brings it back to normal.
BUG: https://bugzilla.samba.org/show_bug.cgi?id=15590
Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Stefan Metzmacher <metze@samba.org>
(cherry picked from commit
1bf9ede94f0a6b41fb18e880e59a8e390f8c21d3)
*
* So we don't free the index_in_top_level dn list yet,
* because we are (probably) borrowing most of its
- * children.
+ * children. But we can save memory by discarding the
+ * values and keeping it as an almost empty talloc
+ * node.
*/
+ talloc_realloc(index_in_top_level,
+ index_in_top_level->dn, struct ldb_val *, 1);
index_in_top_level->dn
= talloc_steal(index_in_top_level,
index_in_subtransaction->dn);