Heikki Linnakangas <hlinnakangas@vmware.com> writes:
> On 05/16/2014 01:28 PM, Andres Freund wrote:
>> Presumably it's because the type is a fixed width type with a length of
>> 6. I guess it's allocating stuff sizeof(macaddr) but then passes the
>> result around as a Datum which doesn't work well.
I've not tried valgrinding to check, but for macaddr I suspect the blame
can be laid at the feet of gbt_num_compress, which is creating a datum of
actual size 2 * tinfo->size (ie, 12 bytes for macaddr). This is later
stored into an index column of declared type gbtreekey16, so the tuple
assembly code is expecting the datum to have size 16.
AFAICS there is no direct connection between the sizes the C code knows
about and the sizes of the datatypes declared as STORAGE options in
btree_gist--1.0.sql. Probably a reasonable fix would be to add a field
to gbtree_ninfo that specifies the index datum size, and then make
gbt_num_compress palloc0 that size rather than 2 * tinfo->size.
> The complaints seem to be coming from all the varlen data types, too,
> not just macaddr:
Dunno what's the problem for the varlena types, but that's a completely
separate code path.
regards, tom lane