Virtualization - May 2013 - [Qemu-devel] [PATCH v2 00/11] qemu: use virtio linux headers in portable code

"Michael S. Tsirkin" <mst at redhat.com> writes:
> On Sun, May 26, 2013 at 07:55:25PM -0500, Anthony Liguori wrote:
>> "Michael S. Tsirkin" <mst at redhat.com> writes:
>> 
>> > On Sun, May 26, 2013 at 03:49:53PM -0500, Anthony Liguori wrote:
>> >> Paolo Bonzini <pbonzini at redhat.com> writes:
>> >> 
>> >> > Il 26/05/2013 22:02, Michael S. Tsirkin ha scritto:
>> >> >> > My fault.  I should have looked at linux/types.h
(actually asm-generic/).
>> >> >> 
>> >> >> Not really, __uX appear in the headers that were
posted.
>> >> 
>> >> Which is a problem because this is a reserved namespace in
C99.
>> >> 
>> >> >> What I'm saying is - a chance of a conflict is
very remote,
>> >> >> if it happens it's a build failure so easy to
debug.
>> >> >
>> >> > I'm sure that others will complain, :) but you can go
ahead.
>> >> 
>> >> I think we should clean up the virtio headers in the kernel
first.
>> >> Seems like a good thing to do given the standardization effort
too.
>> >> There are lots of headers in uapi that use the C99 int types
>> >
>> > I found 4:
>> > $ grep -l uint include/uapi/linux/*h
>> > include/uapi/linux/dm-log-userspace.h
>> > include/uapi/linux/fuse.h
>> > include/uapi/linux/jffs2.h
>> > include/uapi/linux/pps.h
>> > include/uapi/linux/rds.h
>> > include/uapi/linux/sctp.h
>> >
>> > That's not really lots.
>> >
>> >> so there
>> >> doesn't seem to be a reason they couldn't be used in
virtio.  fuse.h
>> >> even has a:
>> >> 
>> >>     #ifdef __KERNEL__
>> >>     #include <linux/types.h>
>> >>     #else
>> >>     #include <stdint.h>
>> >>     #endif
>> >> 
>> >> Which seems like a reasonable thing to do.
>> >
>> > In kernel, we really want to use things like endian-ness
>> > checks (and, we really should have them in userspace too!).
>> > So we want __le32 and other kernel goodies
>> > like that. stdint.h won't cut it.
>> 
>> With the spec being standardized, I think having a stand alone set of
>> headers is a good thing.
>
> Sure, that's possible. We'll have to find some way to
> preserve the endian-ness annotations, I think.
> And then import them into kernel/qemu with some script, converting
> to kernel/qemu style and annotations?
See below.
>>  Licensing is problematic here too.
>> 
>> If virtio headers depend on other Linux headers, then it really
doesn't
>> matter if they are BSD licensed if you need a GPL header (like
>> linux/if_ether.h).
>> 
>> Now, we can certainly debate the copyrightability of these defines and
>> what have you but if the headers are meant to be 1) consumed outside
the
>> kernel 2) licensed under a different license than the general kernel
>> then depending on kernel goodies is the wrong strategy.
>
> Well specifically if_ether.h says GPLv2+ so it's OK for QEMU.
> Do you mean for some other non GPL app?
Ignore QEMU for the moment.

The headers say they are BSD licensed... but they include a GPLv2+
header.  Doesn't make a lot of sense, does it?

Again, I think it's something pretty basic here.  Either (1) these are
kernel/userspace headers that are meant to be consumed through libc or
whatever (2) these are kernel-private headers not to be consumed outside
of the kernel (3) these are headers meant to be copied widely and used
as a reference implementation of virtio.

If (1) or (2), copying them into QEMU via a magic sanitizing script is
wrong.  We should just keep doing what we're doing.

If (3), then we should clean up the virtio headers to be license clean
and usable outside of the kernel.  We shouldn't try to solve this
problem in QEMU (via scripts) if we can just solve it in the kernel.

Regards,

Anthony Liguori
>
>> >> The only other kernel dependency is linux/if_ether.h to define
>> >> ETH_ALEN.  But it seems completely reasonable to define a
>> >> VIRTIO_NET_MAC_ALEN or something like that.
>> >
>> > Ugh. Not really reasonable IMO. We also use ETH_P_IP in code,
>> > would like to get rid of redefining that too.
>> > But we can have our own linux/if_ether.h for non-linux hosts,
>> > just with a
>> > couple of macros like that, it's not a big deal.
>> 
>> See above.
>> 
>> Regards,
>> 
>> Anthony Liguori
>> 
>> >
>> >> This would make the virtio headers completely stand alone and
includable
>> >> in userspace (without violating C99).
>> >> 
>> >> Perhaps it's even worth moving the headers from uapi/linux
to
>> >> uapi/virtio.  Rusty, what do you think?
>> >> 
>> >> Regards,
>> >> 
>> >> Anhtony Liguori
>> >> 
>> >> >
>> >> > Paolo

On Mon, May 27, 2013 at 11:14:47AM -0500, Anthony Liguori
wrote:
> > Well specifically if_ether.h says GPLv2+ so it's OK for QEMU.
> > Do you mean for some other non GPL app?
> 
> Ignore QEMU for the moment.
> 
> The headers say they are BSD licensed... but they include a GPLv2+
> header.
Above is a bit misleading.  To be precise, they don't include a GPLv2+
header. One header merely includes this line:
#include <linux/if_ether.h>

-- 
MST

Anthony Liguori <anthony at codemonkey.ws> writes:
> The headers say they are BSD licensed... but they include a GPLv2+
> header.  Doesn't make a lot of sense, does it?
It makes perfect sense: you're overthinking it.  It just means that
copying the BSD headers outside Linux is encouraged.

And it's clearly nonsensical to claim the GPL on kernel headers means
that userspace needs to be GPL.  So please ignore this licensing
red-herring.

And we'll bikeshed the headers in the standard when we have to :)  They
certainly don't need to be cut&paste into the kernel sources.

Cheers,
Rusty.