Re: [Patch] PR 31228, Fix close-on-exec race.

[David Daney <ddaney at avtrex dot com>]

David Daney wrote:
> Tom Tromey wrote:
> > > > > > > "David" == David Daney <ddaney@avtrex.com> writes:
> >
> > David> I hope someone can think of a better way to fix the race.  If
> > David> only there were some way to set the default value of the
> > David> FD_CLOEXEC flag...
> >
> > We could wrap each "open + set FD_CLOEXEC" in a mutex, and do the same
> > for the fork.
>
> That does not work in a blocking socket accept situation.  It would 
> prevent Runtime.exec() from starting the subprocess until a connection 
> arrived.

OK, I will respond to myself:

Make the accept non-blocking, and do a select/poll on the ServerSocket. 
 That way the accept would never block and you could use a mutex.

This is workable.  I think there are only 4 or 5 places file descriptors 
are obtained in libgcj.

This makes me think that ld.so is probably broken also.  When we do 
dlopen ld.so opens the library file and does several mmaps before 
closing the file descriptor.  These could leak out as well.

David Daney