On Sat, Mar 4, 2017 at 10:32 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Without having actually looked at this patch, I would say that if it added
> a direct call of fopen() to backend-side code, that was already the wrong
> thing. Almost always, AllocateFile() would be a better choice, not only
> because it's tied into transaction abort, but also because it knows how to
> release virtual FDs in event of ENFILE/EMFILE errors. If there is some
> convincing reason why you shouldn't use AllocateFile(), then a safe
> cleanup pattern would be to have the fclose() in a PG_CATCH stanza.
I think that my previous remarks on this issue were simply muddled
thinking. The SQL-callable function pg_current_logfile() does use
AllocateFile(), so the ERROR which may occur afterward if the file is
corrupted is no problem. The syslogger, on the other hand, uses
logfile_open() to open the file, but it's careful not to throw an
ERROR while the file is open, just like other code which runs in the
syslogger. So now I think there's no bug here.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company