Re: [PATCH] conntrack: -L doesn't take a value, so don't discard one (same for -IUDGEFA)

[Pablo Neira Ayuso <pablo@netfilter.org>]

On Thu, Sep 26, 2024 at 12:33:00PM +0200, Pablo Neira Ayuso wrote:
> On Thu, Sep 26, 2024 at 10:28:58AM +0200, наб wrote:
> > On Wed, Sep 25, 2024 at 10:32:59PM +0200, Pablo Neira Ayuso wrote:
> > > On Wed, Sep 25, 2024 at 05:11:01PM +0200, Ahelenia Ziemiańska wrote:
> > > > On Wed, Sep 25, 2024 at 04:53:49PM +0200, Pablo Neira Ayuso wrote:
> > > > > On Tue, Sep 03, 2024 at 04:53:46PM +0200, Ahelenia Ziemiańska wrote:
> > > > > > On Tue, Sep 03, 2024 at 10:22:09AM +0200, Pablo Neira Ayuso wrote:
> > > > > > > On Tue, Sep 03, 2024 at 04:16:21AM +0200, Ahelenia Ziemiańska wrote:
> > > > > > > > The manual says
> > > > > > > >    COMMANDS
> > > > > > > >        These options specify the particular operation to perform.
> > > > > > > >        Only one of them can be specified at any given time.
> > > > > > > > 
> > > > > > > >        -L --dump
> > > > > > > >               List connection tracking or expectation table
> > > > > > > > 
> > > > > > > > So, naturally, "conntrack -Lo extended" should work,
> > > > > > > > but it doesn't, it's equivalent to "conntrack -L",
> > > > > > > > and you need "conntrack -L -o extended".
> > > > > > > > This violates user expectations (borne of the Utility Syntax Guidelines)
> > > > > > > > and contradicts the manual.
> > > > > > > > 
> > > > > > > > optarg is unused, anyway. Unclear why any of these were :: at all?
> > > > > > > Because this supports:
> > > > > > >         -L
> > > > > > >         -L conntrack
> > > > > > >         -L expect
> > > > > > Well that's not what :: does, though; we realise this, right?
> > > > > > 
> > > > > > "L::" means that getopt() will return
> > > > > >   "-L", "conntrack" -> 'L',optarg=NULL
> > > > > >   "-Lconntrack"     -> 'L',optarg="conntrack"
> > > > > > and the parser for -L (&c.) doesn't... use optarg.
> > > > > Are you sure it does not use optarg?
> > > > > 
> > > > > static unsigned int check_type(int argc, char *argv[])
> > > > > {
> > > > >         const char *table = get_optional_arg(argc, argv);
> > > > > 
> > > > > and get_optional_arg() uses optarg.
> > > > This I've missed, but actually my diagnosis still holds:
> > > >   static unsigned int check_type(int argc, char *argv[])
> > > >   {
> > > >   	const char *table = get_optional_arg(argc, argv);
> > > >   
> > > >   	/* default to conntrack subsystem if nothing has been specified. */
> > > >   	if (table == NULL)
> > > >   		return CT_TABLE_CONNTRACK;
> > > > 
> > > >   static char *get_optional_arg(int argc, char *argv[])
> > > >   {
> > > >   	char *arg = NULL;
> > > >   
> > > >   	/* Nasty bug or feature in getopt_long ?
> > > >   	 * It seems that it behaves badly with optional arguments.
> > > >   	 * Fortunately, I just stole the fix from iptables ;) */
> > > >   	if (optarg)
> > > >   		return arg;
> > > > 
> > > > So, if you say -Lanything, then
> > > >   optarg=anything
> > > >   get_optional_arg=(null)
> > > > (notice that it says "return arg;", not "return optarg;",
> > > >  i.e. this is "return NULL").
> > > > 
> > > > It /doesn't/ use optarg, because it explicitly treats an optarg as no optarg.
> > > > 
> > > > It's unclear to me what the comment is referencing,
> > > > but I'm assuming some sort of confusion with what :: does?
> > > > Anyway, that if(){ can be removed now, since it can never be taken now.
> > > Then, this breaks:
> > > # conntrack -Lexpect
> > > conntrack v1.4.9 (conntrack-tools): Bad parameter `xpect'
> > > Try `conntrack -h' or 'conntrack --help' for more information.
> > > 
> > > Maybe your patch needs an extension to deal with this case too?
> > 
> > This doesn't "break", this is equivalent to conntrack -L -e xpect.
> > It's now correct. This was the crux of the patch, actually.
> > 
> > Compare the manual:
> >   SYNOPSIS
> >     conntrack -L [table] [options] [-z]
> >   COMMANDS
> >     -L --dump     List connection tracking or expectation table
> >   PARAMETERS
> >     -e, --event-mask [ALL|NEW|UPDATES|DESTROY][,...]
> >                   Set the bitmask of events that are to be generated by the in-kernel ctnetlink event code.  Using this parameter, you can reduce the event messages  generated
> >                   by the kernel to the types that you are actually interested in.  This option can only be used in conjunction with "-E, --event".
> > 
> > Previously, it /was/ broken: conntrack -Lexpect was as-if --dump=expect
> > (also not legal since --dump doesn't take an argument),
> > and the "expect" was ignored, so it was equivalent to conntrack -L.
> > You can trivially validate this by running an older version.
> > 
> > (Well, --dump=expect /is/ accepted. And ignored.
> >  So fix that too with s/optional_argument/no_argument/ (or s/2/0/).
> >  I didn't actually look at the longopts before.)
> > 
> > > The issue that I'm observing is that
> > >   # conntrack -Lconntrack
> > > now optarg is NULL after your patch, so 'conntrack' is ignored, so it
> > > falls back to list the conntrack table.
> > 
> > What do you mean "now". That shit was always ignored.
> > You can read trace the calls yourself if you don't believe my analysis.
> > Now it behaves as-documented (-L -c onntrack).
> > 
> > And, per
> >                 case 'c':
> >                         options |= opt2type[c];
> >                         nfct_set_attr_u32(tmpl->ct,
> >                                           opt2attr[c],
> >                                           strtoul(optarg, NULL, 0));
> >                         break;
> > -c onntrack is equivalent to -c 0.
> > This is also obviously wrong.
> > 
> > I will repeat this and you can confirm this once more
> > (or refer back to my analysis above):
> > for all of -LIUDGEFA, an optional parameter was accepted, and always discarded.
> > It now isn't, and behaves as-expected per the USG
> > ("the USG" is an annoying way to say "how getopt() works".
> > 
> > > Regarding your question, this parser is old and I shamelessly took it
> > > from the original iptables to make syntax similar.
> > So you have someone to blame it on when it turns out to be dysfunctional.
> > But you also have a huge parser that doesn't work.
> > Win some/lose some, I suppose.
> 
> Your stuff breaks existing behaviour. I will revert and leave it as is.
> 
> There is a risk of breaking existing applications.
> 
> You can use the word shit, dysfunctional, and keep augment your
> wording as many times as you want, but that does not change my point.

So either fix it is a backward compatible way or there will be no fix.