RE: [HACKERS] refcnt leak ?

[Hiroshi Inoue]

> -----Original Message----- > From: Tom Lane [mailto:[EMAIL PROTECTED]] > > Hiroshi Inoue <[EMAIL PROTECTED]> writes: > > While examining recursive use of catalog cache,I found > > a refcnt leak of relations. > > After further investigation,I found that the following seems > > to be the cause. > > > [ in EndAppend() in nodeAppend.c ] > > appendstate-> as_result_relation_info_list = NIL; > Oh no,my point isn't on this line but on the line estate->es_result_relation_info = NULL; As the comment says,it depends on the assumption that estate->es_result_relation_info points to one of the node of appendstate->as_result_relation_info_list(before set to NIL). However ISTM appendstate->as_result_relation_info _list is for inheritance and in the case "INSERT INTO .. SELECT .. FROM .." it's not used. > That doesn't look like a problem to me --- the result relations *have* > been closed, just above this line. > > > BTW,doesn't EndAppend() neglect to call ExecCloseIndices() > > for RelationInfos of appendstate->as_result_relation_info_list ? > > Comparing nodeAppend to EndPlan(), I think you are right --- each > resultinfo should have ExecCloseIndices applied too, in the loop just > above the line you quote. This did not use to be a problem because > Append plans were readonly, but now that we have UPDATE/DELETE on > inheritance hierarchies, there's a missing step here. Was your test > query of that kind? > I first changed this part but rd_refcnt leak didn't disappaear. I have no refcnt leak example which is caused due to this flaw(?). After that I found " estate->es_result_relation_info = NULL; " in EndAppend() . I changed it to not do so when appendstate-> as_result_relation_info_list is NIL and rd_refcnt leak disappeared. Regards. Hiroshi Inoue