Re: _libvirt, $opt_args, and noglob

[Daniel Shahaf <d.s@daniel.shahaf.name>]

See the penultimate paragraph for a question about whether ${(q)} should
quote leading '=' signs to guard against the EQUALS option.

Marko Myllynen wrote on Wed, Sep 07, 2016 at 12:53:13 +0300:
> Hi,
> 
> On 2016-09-07 02:46, Daniel Shahaf wrote:
> > Daniel Shahaf wrote on Sun, Sep 04, 2016 at 18:26:22 +0000:
> >> In the special case of _libvirt, simply adding noglob to the
> >> _call_program arguments would probably fix this.
> > 
> > Done, but the last line touches the sudo invocation from Oliver's
> > gain-privileges patch.  Oliver: shall I wait with committing this until
> > you've committed gain-privileges?
> > 
> > @@ -224,7 +224,7 @@ case $state in
> >        local srv ; (( ${(k)words[(I)--server]} > 0 )) && srv=${words[1+${(k)words[(I)--server]}]}
> >        [[ -z $srv ]] && return 1
> >        [[ -n ${srv//[[:alnum:]]} ]] && return 1
> > -      _wanted clients expl client compadd ${=${${(f):-"$(sudo virt-admin ${(Q)conn_opt} srv-clients-list --server $srv 2>/dev/null)"}/ [a-z]*}//[^0-9]} && return 0
> > +      _wanted clients expl client compadd ${=${${(f):-"$(noglob sudo virt-admin ${(Q)conn_opt} srv-clients-list --server $srv 2>/dev/null)"}/ [a-z]*}//[^0-9]} && return 0
> >      fi
> >      [[ -z $_cache_virt_admin_cmd_opts[$cmd] ]] && \
> >        _cache_virt_admin_cmd_opts[$cmd]=${(M)${${${${=${(f)"$(_call_program virt-admin virt-admin help $cmd 2>&1)"}}/\[}/\]}/\;}:#-[-0-9A-Za-z]*}
> > ]]]
> 
> Do we need any of (Q)'s there any more after your addition of:
> 
>   uri=${uri//(#m)\\([\\:])/${MATCH[2]}} # opt_args elements are colon-escaped
> 
> Based on a quick test looks like they could be dropped (or
> perhaps even changed to (q))?

_call_program expects its argument to be (q)'d.  (The interface is like
'eval', not like execve(2).)

The values of $opt_args are *already* (q)'d.  If the command line is
«virsh -c foo\\bar» then the value of $pot_args[-c] includes two
backslashes, not one as virsh(1)'s argv[2] would see.¹

Therefore, I think the correct way to pass values derived from $opt_args
— this includes $conn_opt — to _call_program is to use neither (q) nor (Q).

That is the general answer.  For the specific case of $conn_opt, the URI
format validator ensures it has nothing that needs quoting, apart from
'?' and '=', and the patch in the grandparent message removes the need
for '?' to be quoted.  However, there's no easy to to quote '=' when
it's the first character of a word, and _call_program does honour the
EQUALS option:
.
    % _f() { compadd - $(_call_program x 'echo =true') }
    % f <TAB>
    <becomes>
    % f /bin/true

Again, this is harmless as far as _libvirt is concerned, but it might
not be so in other callsites.  I think the right general solution to
this would be to write ${(q)${(Q)conn_opts}} in the _call_program
invocation, and teach the (q) flag to escape leading equals signs.

Anyway, back to Marko's question: I assume even if the EQUALS option did
affect the command executed, that wouldn't cause any harm, so I think
the primary reason to get _libvirt's quoting right is in case somebody
ever uses _libvirt as a basis for some other completion function.
That's part of why I committed the (b) patch, too.

Cheers,

Daniel

¹ This assumes that the word being completed, $CURRENT, is not the
argument to -c, in which cae $opt_args[-c] could have an opening quote
but no matching closing quote.