On Thu, May 16, 2019 at 12:51:34AM -0400, Tom Lane wrote:
> Hm, should we separate the cleanup of the root pgsql_tmp/ from the rest of
> what RemovePgTempFiles does? I'm still feeling like we should be trying
> to launch the syslogger as soon as possible.
Indeed, this should be safe, the syslogger is not going to touch any
of the other temp files anyway. If we begin to do that, we had better
update the comments at the top of RemovePgTempFiles()'s call, and
expose RemovePgTempFilesInDir().
Anyway, I have to admit that I am not really enthusiastic about
relaxing the assertions for the win32 and the sysv interfaces, and do
that inconsistently on top of it.
As far as my first investigations have gone, assertions in
pgwin32_ReserveSharedMemoryRegion() and
PGSharedMemoryReAttach():sysv_shmem.c would need relaxing, but we rely
on that since a7e5878, which is quite some time ago. And, actually,
the failure for pgwin32_ReserveSharedMemoryRegion() should never
happen at all, because we need to call reset_shared() before starting
the syslogger as well, no? If we want to be able to log the creation
socket messages with the logging collector, it seems to me that we
would need first to remove the dependency with the port number in
PGSharedMemoryCreate() to find a free IPC key. Thoughts?
--
Michael