Re: [PATCH v2 1/2] Clean stale environment pointer in finish_command()

[Jeff King <peff@peff.net>]

On Mon, Nov 10, 2014 at 01:44:24PM -0800, Junio C Hamano wrote:

> Johannes Schindelin <johannes.schindelin@gmx.de> writes:
> 
> > In start_command(), unset "env" fields are initialized via "env_array". In
> > finish_command(), the "env_array" is cleared, therefore the "env" field
> > will point to free()d data.
> >
> > However, start_command() will set "env" to env_array.argv only if "env"
> > was unset to begin with, and if it was already set, the caller will need
> > the original value. Therefore, we need to be very careful only to reset
> > "env" in finish_command() when it has been initialized in start_command().
> 
> Hmph.  Does the same observation apply to cmd->argv that is
> initialied to point to cmd->args.argv only when it is unset?

Yes, they behave exactly the same (I think Dscho just doesn't hit it in
his patch because he assigns argv manually).

I don't have a real problem with going in this direction as a safety
measure, but I am not sure that it is safe to reuse a child_process
after finish_command in general, without an intervening
child_process_init. For instance, calling start_command will convert a
"child_process.in" value of "-1" instead a pipe, and overwrite that "-1"
with the descriptor of the pipe. A subsequent use of the same
child_process struct will ask the second child to use that pipe (the
write-half of the pipe, mind you) as its stdin, which is nonsensical.

So I think you are much better off just using two child_process structs
(or a single one and reinitializing in between calls).

-Peff