Re: inserts into partitioned table may cause crash

(2018/02/28 17:36), Amit Langote wrote:
> I've run into what seems to be a bug in ExecInsert() that causes a crash
> when inserting multiple rows into a partitioned table that each go into
> different partitions with different tuple descriptors.  Crash occurs if
> ExecInsert() returns without resetting estate->es_result_relation_info
> back to the root table's resultRelInfo.  For example, if a BR trigger on a
> partition returns NULL for a row.
>
> Crashing example:
>
> create table p (a int, b text) partition by list (a);
> create table p12 (b text, a int);
>
> -- p12 has different tuple descriptor than p
> alter table p attach partition p12 for values in (1, 2);
>
> create table p4 partition of p for values in (4);
>
> create function blackhole () returns trigger as $$ begin return NULL; end;
> $$ language plpgsql;
> create trigger blackhole before insert on p12 for each row execute
> procedure blackhole();
>
> insert into p values (1, 'b'), (4, 'a');
> server closed the connection unexpectedly
>     This probably means the server terminated abnormally
>     before or while processing the request.
> The connection to the server was lost. Attempting reset: Failed.
>
> Crash is caused because we enter into ExecFindPartition with p12's
> resultRelInfo as if it correponds to the root table.  That happens because
> we didn't reset estate->es_result_relation_info, which had been set to
> p12's resultRelInfo to point back to the original resultRelInfo (that is,
> p's) before returning like below:
>
>         slot = ExecIRInsertTriggers(estate, resultRelInfo, slot);
>
>          if (slot == NULL)       /* "do nothing" */
>              return NULL;
>
> There are other places where we prematurely return like that.
>
> Attached a patch to fix that, which would need to be back-patched to 10.

Good catch!  Will review.

Best regards,
Etsuro Fujita