Re: [nft PATCH 1/2] monitor: Rewrite SETELEM callback

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

On Tue, Jul 18, 2017 at 11:17:26AM +0200, Phil Sutter wrote:
> On Tue, Jul 18, 2017 at 11:09:37AM +0200, Pablo Neira Ayuso wrote:
> > On Tue, Jul 18, 2017 at 11:05:16AM +0200, Phil Sutter wrote:
> > > On Mon, Jul 17, 2017 at 07:16:29PM +0200, Pablo Neira Ayuso wrote:
> > > > On Mon, Jul 17, 2017 at 06:41:14PM +0200, Phil Sutter wrote:
> > > > > On Mon, Jul 17, 2017 at 06:30:18PM +0200, Pablo Neira Ayuso wrote:
> > > > > > On Mon, Jul 17, 2017 at 05:06:05PM +0200, Phil Sutter wrote:
> > > > > > [...]
> > > > > > > +static int netlink_events_setelem_newgen_cb(const struct nlmsghdr *nlh,
> > > > > > > +					    int type,
> > > > > > > +					    struct netlink_mon_handler *monh)
> > > > > > > +{
> > > > > > > +	setelem_cache_print_default(monh);
> > > > > > > +
> > > > > > > +	return MNL_CB_OK;
> > > > > > >  }
> > > > > > 
> > > > > > I would really like we don't rely on newgen for this. If there is no
> > > > > > way to catch a case with the existing way we represent this, then we
> > > > > > probably need to fix things from the kernel.
> > > > > > 
> > > > > > Before we follow that patch, I would like to understand what corner
> > > > > > case is pushing us to use the newgen event.
> > > > > 
> > > > > It is required for half-open ranges occurring at the end of the
> > > > > transaction: For those, we only get a single element without
> > > > > EXPR_F_INTERVAL_END flag set. Since this could also be the first part of
> > > > > a regular range, monitor has to wait for what's next - which is in doubt
> > > > > only the NEWGEN message.
> > > > > 
> > > > > Maybe we could introduce a new flag to mark these?
> > > > 
> > > > Right, I think we need the new flag indeed, only for userspace.
> > > > 
> > > > Would you propose one and the specific semantics for it?
> > > 
> > > My current PoC passes the additional flag as userdata attribute so the
> > > kernel won't reject the element due to unknown flag. Is that fine with
> > > you? I'm trying to avoid changing the kernel so the solution is
> > > backwards compatible.
> > 
> > I suggest you add a new flag to SET_ELEM instead. Userdata area usage
> > is exclusive to userspace.
> 
> You mean nft_set_elem_flags? The new flag will indeed be used by
> userspace only: It is set when creating a half-open range and not used
> by the kernel at all.

That's fine indeed.