On Mon, Jun 23, 2014 at 03:39:32PM -0700, Jesse Gross wrote:
> On Fri, Jun 13, 2014 at 3:28 PM, Ben Pfaff <b...@nicira.com> wrote:
> > Until now, the OVS source tree has had a whole maze of header files that
> > make "#include <linux/openvswitch.h>" work OK regardless of platform, but
> > this confuses everyone new to the tree, at first glance, and is difficult
> > to understand at second glance too.
> >
> > This commit moves the core of the Netlink definitions that were in
> > include/linux/openvswitch.h into a new header
> > include/odp-netlink-internal.h, and then adds two wrappers:
> >
> > * datapath/linux/compat/include/linux/openvswitch.h: This wrapper
> > allows the Linux datapath code to continue using "#include
> > <linux/openvswitch.h>".
> >
> > * lib/odp-netlink.h: Userspace code uses this wrapper.
> >
> > This change allows all of the include/linux/* files to be deleted.
> >
> > Signed-off-by: Ben Pfaff <b...@nicira.com>
>
> I'm somewhat worried that diverging our copy of linux/openvswitch.h
> will make things harder to maintain, especially since we've tried to
> reduce differences between upstream and the out-of-tree module as much
> as possible.
>
> I was imagining that we would model this as if we were compiling
> purely against the upstream kernel and openvswitch.h was an exported
> header (since it is, we just happen to have a compatibility backport).
> In that case, we wouldn't be able to change the original header and
> presumably Linux-specific code would translate to something
> OVS-internal.
I've been starting to think about this again since (as you pointed out
elsewhere) it came up in the hyperv port.
I'm not too sure what you're envisioning in the model that you're
thinking of. Here are my thoughts on how to do this in terms of
backports:
* We'd need some base version of the whole header file, because
not every build environment would have any version at all
(given that we build on non-Linux and often system Linux
header files are for kernels older than the one actually in
use).
* We'd need to detect the presence of numerous "enum" values and
define them if not present (or we could unconditionally
#define them as backports whether present or not, since using
#defines as overlays for enums is mostly harmless).
* We'd need to detect the presence of numerous structs and
define them if not present.
It might be more work to maintain a set of backports than a slightly
different header file. (It's definitely more work than what we're doing
now.)
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
