Thread: Should libpq set close-on-exec flag on its socket?

It was suggested to me off-list that libpq should do
"fcntl(fd, F_SETFD, FD_CLOEXEC)" on the socket connecting to the server.
This would prevent any child program from accidentally or maliciously
interfering with the connection.  It would also prevent people from
deliberately turning over a connection to a child; I'm not sure that
that's useful, but I'm not sure it's useless either.

Comments, opinions?
        regards, tom lane

On Thu, 21 Oct 2004, Tom Lane wrote:

> It was suggested to me off-list that libpq should do
> "fcntl(fd, F_SETFD, FD_CLOEXEC)" on the socket connecting to the server.
> This would prevent any child program from accidentally or maliciously
> interfering with the connection.

Either way that the lib sets it, the client can alter the setting itself
by issuing a new SETFD command. I would not have expected it to be set 
but it is probably a good idea for most clients (and for most file 
descriptors).

-- 
/Dennis Björklund

Dennis Bjorklund <db@zigo.dhs.org> writes:
> On Thu, 21 Oct 2004, Tom Lane wrote:
>> It was suggested to me off-list that libpq should do
>> "fcntl(fd, F_SETFD, FD_CLOEXEC)" on the socket connecting to the server.
>> This would prevent any child program from accidentally or maliciously
>> interfering with the connection.

> Either way that the lib sets it, the client can alter the setting itself
> by issuing a new SETFD command.

That's a fair point, and certainly passing it down to the child
intentionally wouldn't be a common case.  I'll put the change in.
        regards, tom lane

Tom Lane wrote:
> Dennis Bjorklund <db@zigo.dhs.org> writes:
> > On Thu, 21 Oct 2004, Tom Lane wrote:
> >> It was suggested to me off-list that libpq should do
> >> "fcntl(fd, F_SETFD, FD_CLOEXEC)" on the socket connecting to the server.
> >> This would prevent any child program from accidentally or maliciously
> >> interfering with the connection.
> 
> > Either way that the lib sets it, the client can alter the setting itself
> > by issuing a new SETFD command.
> 
> That's a fair point, and certainly passing it down to the child
> intentionally wouldn't be a common case.  I'll put the change in.

Since program authors who would care about this one way or another
probably won't be expecting this behavior, it should also be
documented reasonably well -- something which I'm rather sure you were
going to do anyway.




-- 
Kevin Brown                          kevin@sysexperts.com

On Thu, Oct 21, 2004 at 02:10:48PM -0400, Tom Lane wrote:
> It was suggested to me off-list that libpq should do
> "fcntl(fd, F_SETFD, FD_CLOEXEC)" on the socket connecting to the server.
> This would prevent any child program from accidentally or maliciously
> interfering with the connection.  It would also prevent people from
> deliberately turning over a connection to a child; I'm not sure that
> that's useful, but I'm not sure it's useless either.
> 
> Comments, opinions?

This is a very good idea.  We've had problems with Perl programs that
call other scripts (over an exec boundary) and end up with unnecessary
DBD::Pg file handles hanging around.  This would be good to prevent
that.

-Dom