uapi: Prevent redefinition of struct iphdr

As with struct ethhdr, the musl libc provides its own definition of the
iphdr struct.  This guard ensures software like net-tools builds correctly
on the musl libc.

The __UAPI_DEF_IPHDR definition is in ip.h itself to prevent the issue in
commit da360299b673 ("uapi/if_ether.h: move __UAPI_DEF_ETHHDR libc define")
from being seen here.

Signed-off-by: A. Wilcox <AWilcox@Wilcox-Tech.com>
---
 include/uapi/linux/ip.h | 8 ++++++++
 1 file changed, 8 insertions(+)

From: "A. Wilcox" <AWilcox@Wilcox-Tech.com>
Date: Sun, 22 Dec 2019 00:02:27 -0600

> @@ -83,6 +83,13 @@
>  
>  #define IPV4_BEET_PHMAXLEN 8
>  
> +/* Allow libcs to deactivate this - musl has its own copy in <netinet/ip.h> */
> +
> +#ifndef __UAPI_DEF_IPHDR
> +#define __UAPI_DEF_IPHDR	1
> +#endif

How is this a musl-only problem?  I see that glibc also defines struct iphdr
in netinet/ip.h, so why doesn't it also suffer from this?

I find it really strange that this, therefore, only happens for musl
and we haven't had thousands of reports of this conflict with glibc
over the years.

I want an explanation, and suitably appropriate adjustments to the commit
message and comments of this change.

Thank you.

On Wed, Dec 25, 2019 at 04:34:11PM -0800, David Miller wrote:
> From: "A. Wilcox" <AWilcox@Wilcox-Tech.com>
> Date: Sun, 22 Dec 2019 00:02:27 -0600
> 
> > @@ -83,6 +83,13 @@
> >  
> >  #define IPV4_BEET_PHMAXLEN 8
> >  
> > +/* Allow libcs to deactivate this - musl has its own copy in <netinet/ip.h> */
> > +
> > +#ifndef __UAPI_DEF_IPHDR
> > +#define __UAPI_DEF_IPHDR	1
> > +#endif
> 
> How is this a musl-only problem?

I don't think it is, unless glibc's includes linux/ip.h to get the
definition, which does not seem to be the case -- at least not on the
Debian system I had handy to check on.

> I see that glibc also defines struct iphdr
> in netinet/ip.h, so why doesn't it also suffer from this?

Maybe it does.

> I find it really strange that this, therefore, only happens for musl
> and we haven't had thousands of reports of this conflict with glibc
> over the years.

It's possible that there's software that's including just one of the
headers conditional on __GLIBC__, and including both otherwise, or
something like that. Arguably this should be considered unsupported
usage; there are plenty of headers where that doesn't work and
shouldn't be expected to.

> I want an explanation, and suitably appropriate adjustments to the commit
> message and comments of this change.

Agreed. Commit messages should not imply that something is a
musl-specific workaround when it's generally the right thing to do.

Rich

From: Rich Felker <dalias@libc.org>
Date: Wed, 25 Dec 2019 20:05:15 -0500

> On Wed, Dec 25, 2019 at 04:34:11PM -0800, David Miller wrote:
>> I find it really strange that this, therefore, only happens for musl
>> and we haven't had thousands of reports of this conflict with glibc
>> over the years.
> 
> It's possible that there's software that's including just one of the
> headers conditional on __GLIBC__, and including both otherwise, or
> something like that. Arguably this should be considered unsupported
> usage; there are plenty of headers where that doesn't work and
> shouldn't be expected to.

I don't buy that, this is waaaaaay too common a header to use.

Please investigate.

On Thu, Dec 26, 2019, at 04:49, David Miller wrote:
> From: Rich Felker <dalias@libc.org>
> Date: Wed, 25 Dec 2019 20:05:15 -0500
> 
> > On Wed, Dec 25, 2019 at 04:34:11PM -0800, David Miller wrote:
> >> I find it really strange that this, therefore, only happens for musl
> >> and we haven't had thousands of reports of this conflict with glibc
> >> over the years.
> > 
> > It's possible that there's software that's including just one of the
> > headers conditional on __GLIBC__, and including both otherwise, or
> > something like that. Arguably this should be considered unsupported
> > usage; there are plenty of headers where that doesn't work and
> > shouldn't be expected to.
> 
> I don't buy that, this is waaaaaay too common a header to use.

In case of net-tools, only <linux/ip.h> is included, and never <netinet/ip.h> directly. Chances are in musl the indirect include tree happens to be different and conflicting, while in glibc it is not.

> 
> Please investigate.
>

Daniel

On Thu, Dec 26, 2019 at 12:13:37PM +0100, Daniel Kolesa wrote:
> On Thu, Dec 26, 2019, at 04:49, David Miller wrote:
> > From: Rich Felker <dalias@libc.org>
> > Date: Wed, 25 Dec 2019 20:05:15 -0500
> > 
> > > On Wed, Dec 25, 2019 at 04:34:11PM -0800, David Miller wrote:
> > >> I find it really strange that this, therefore, only happens for musl
> > >> and we haven't had thousands of reports of this conflict with glibc
> > >> over the years.
> > > 
> > > It's possible that there's software that's including just one of the
> > > headers conditional on __GLIBC__, and including both otherwise, or
> > > something like that. Arguably this should be considered unsupported
> > > usage; there are plenty of headers where that doesn't work and
> > > shouldn't be expected to.
> > 
> > I don't buy that, this is waaaaaay too common a header to use.
> 
> In case of net-tools, only <linux/ip.h> is included, and never
> <netinet/ip.h> directly. Chances are in musl the indirect include
> tree happens to be different and conflicting, while in glibc it is
> not.

musl has no indirect inclusion of netinet/ip.h from standard headers,
but does include it from netinet/ip_icmp.h. It seems glibc only does
this conditional on __USE_MISC, which doesn't make much sense to me
since this is not a standardized header with namespace rules, but
normally __USE_MISC is defined anyway on glibc so I kinda doubt this
is the difference.

Any other ideas?

Rich