On 10/07/2015 09:27 AM, Michael Paquier wrote:
> On Wed, Oct 7, 2015 at 7:51 AM, Michael Paquier
> <michael.paquier@gmail.com> wrote:
>> On Wed, Oct 7, 2015 at 7:43 AM, Michael Paquier
>> <michael.paquier@gmail.com> wrote:
>>> On Wed, Oct 7, 2015 at 5:58 AM, Robert Haas wrote:
>>>> On Sat, Oct 3, 2015 at 7:38 AM, Michael Paquier wrote:
>>>> It seems that these days 'make check' creates a directory in /tmp
>>>> called /tmp/pg_regress-RANDOMSTUFF. Listening on TCP ports is
>>>> disabled, and the socket goes inside this directory with a name like
>>>> .s.PGSQL.PORT. You can connect with psql -h
>>>> /tmp/pg_regress-RANDOMSTUFF -p PORT, but not over TCP. This basically
>>>> removes the risk of TCP port number collisions, as well as the risk of
>>>> your temporary instance being hijacked by a malicious user on the same
>>>> machine.
>>>
>>> Right, that's for example /var/folders/ on OSX, and this is defined
>>> once per test run via $tempdir_short. PGHOST is set to that as well.
>>
>> Er, mistake here. That's actually once per standard_initdb, except
>> that all the tests I have included in my patch run it just once to
>> create a master node. It seems that it would be wiser to set one
>> socket dir per node then, remove the port assignment stuff, and use
>> tempdir_short as a key to define a node as well as in the connection
>> string to this node. I'll update the patch later today...
>
> So, my conclusion regarding multiple calls of make_master is that we
> should not allow to do it. On Unix/Linux we could have a separate unix
> socket directory for each node, but not on Windows where
> listen_addresses is set to look after 127.0.0.1. On Unix/Linux, PGHOST
> is set by the master node to a tempdir once and for all. Hence, to
> make the code more consistent, I think that we should keep the port
> lookup machinery here. An updated patch is attached.
>
If parallel tests are supported, get_free_port is still racy even
with last_port_found because it's:
1) process-local.
2) even if it were shared, there's the race window between the
available-check and the listen() I mentioned upthread.
If parallel tests are explicitly disallowed, a comment to that
effect (and a note on things known to break) might help someone
down the road.
Also, the removal of poll_query_until from pg_rewind looks suspiciously
like a copy-paste gone bad. Pardon if I'm missing something.
Amir
