Re: git: c93db4abf454 - main - udp: call UDP methods from UDP over IPv6 directly

From: Gleb Smirnoff <glebius_at_freebsd.org>
Date: Wed, 17 Aug 2022 16:57:22 UTC
On Wed, Aug 17, 2022 at 09:37:44AM -0700, John Baldwin wrote:
J> On 8/17/22 9:19 AM, Gleb Smirnoff wrote:
J> >    John,
J> > 
J> > On Wed, Aug 17, 2022 at 09:04:08AM -0700, John Baldwin wrote:
J> > J> > diff --git a/sys/netinet6/udp6_usrreq.c b/sys/netinet6/udp6_usrreq.c
J> > J> > index a7bdfce97707..6a3ac2abd90b 100644
J> > J> > --- a/sys/netinet6/udp6_usrreq.c
J> > J> > +++ b/sys/netinet6/udp6_usrreq.c
J> > J> > @@ -131,12 +131,18 @@ VNET_DEFINE(int, zero_checksum_port) = 0;
J> > J> >   SYSCTL_INT(_net_inet6_udp6, OID_AUTO, rfc6935_port, CTLFLAG_VNET | CTLFLAG_RW,
J> > J> >       &VNET_NAME(zero_checksum_port), 0,
J> > J> >       "Zero UDP checksum allowed for traffic to/from this port.");
J> > J> > +
J> > J> > +
J> > J> > +/* netinet/udp_usrreqs.c */
J> > J> > +pr_abort_t	udp_abort;
J> > J> > +pr_disconnect_t	udp_disconnect;
J> > J> > +pr_send_t	udp_send;
J> > J> > +
J> > J>
J> > J> Oof, can you please put these in a header instead?  <netinet/udp_var.h> would seem to be
J> > J> a good candidate.
J> > 
J> > I actually don't want to share these functions to anybody.  Sometimes we are
J> > in a situation when two compilation units are historically separated,
J> > while they need to call into each other functions, which by design are private.
J> 
J> Putting them in a header doesn't really change that.  Anyone can add a local
J> prototype and call it.  Even better, they can add a _wrong_ prototype and
J> call it with the wrong arguments.   Or more likely, if the original function changes
J> then previously working code breaks and the compiler won't catch this.  In this
J> specific case you mostly sidestep that due to using a typedef, but in general the
J> wrong prototype problem is the reason to store prototypes in headers.
J> 
J> I'm not sure that "I don't trust other developers" is a valid reason to avoid the
J> use of headers, or at least I'm not sure it outweighs "code breaks without
J> the compiler noticing in the future".  I'd rather signpost it in the header
J> clarifying when it is allowed to be used (e.g. /* Shared between udp*_usrreq.c. */
J> or the like)

I'm already working on the change you suggested and I'm already in dependency hell.

Since the typedefs are declared in protosw.h that would require anybody who included
udp_var.h also include protosw.h. And we have lots of files that include udp_var.h
for good or for no reason. Of course I don't want to sprinkle protosw.h include all
around.

Suggestion by Jessica about udp_usrreq_private.h seems to be more viable, but is
it worth to create yet another include to share a function between 2 files, when
we are sure that there would never be a 3 file to include it?

-- 
Gleb Smirnoff