On Wed, 27 Nov 2024 15:55:00 +0100 Laurent Vivier wrote:

On 27/11/2024 15:48, Stefano Brivio wrote:

...

On Wed, 27 Nov 2024 15:41:30 +0100 Laurent Vivier wrote:

...

there is a conflict between netinet/if_ether.h provided by musl and linux/if_ether.h provided by the linux headers:

In file included from passt.h:185, from tcp_vu.c:21: /usr/include/netinet/if_ether.h:115:8: error: redefinition of 'struct ethhdr' 115 | struct ethhdr { | ^~~~~~ In file included from /usr/include/linux/virtio_net.h:32, from tcp_vu.c:17: /usr/include/linux/if_ether.h:173:8: note: originally defined here 173 | struct ethhdr { | ^~~~~~

The kernel headers provide a flag to disable the definition in this case, __UAPI_DEF_ETHHDR (see /usr/include/linux/if_ether.h comment).

Signed-off-by: Laurent Vivier --- Makefile | 3 +++ 1 file changed, 3 insertions(+)

diff --git a/Makefile b/Makefile index cb7448079de5..2aa56ada65fd 100644 --- a/Makefile +++ b/Makefile @@ -33,6 +33,9 @@ FLAGS += $(FORTIFY_FLAG) -O2 -pie -fPIE FLAGS += -DPAGE_SIZE=$(shell getconf PAGE_SIZE) FLAGS += -DVERSION=\"$(VERSION)\" FLAGS += -DDUAL_STACK_SOCKETS=$(DUAL_STACK_SOCKETS) +ifeq (musl, $(word 4,$(subst -, ,$(TARGET)))) +FLAGS += -D__UAPI_DEF_ETHHDR=0 +endif

Uh oh, I just solved this locally (I was about to push, but I stopped in time) by simply including instead of from passt.h.

I'm not sure what's the best solution: mine is simpler, but I'm not really fond of including Linux-specific stuff if there's no need for it, so I'd slightly prefer yours.

On the other hand, it's not just for musl. If somebody tries to build this against, say, uClibc-ng (no reports about that yet), they will probably run into trouble as well.

In fact, in /usr/include/linux/if_ether.h they say:

/* allow libcs like musl to deactivate this, glibc does not implement this. */

So I think musl should set the value and then I found:

#define __UAPI_DEF_ETHHDR 0

in /usr/include/netinet/if_ether.h

Ah, I hadn't checked that yet.

So I think the solution is to include netinet/if_ether.h before linux/if_ether.h

Just in tcp_vu.c you mean? That's where we include now. Any advantage over simply including linux/if_ether.h in passt.h as I did?

...

Or should we just unconditionally #define __UAPI_DEF_ETHHDR=0 just before including ?

If you do that, I think it will not work anymore with glibc

Right, discarded that option. -- Stefano

On Wed, 27 Nov 2024 16:37:05 +0100 Laurent Vivier wrote:

On 27/11/2024 16:05, Stefano Brivio wrote:

...

On Wed, 27 Nov 2024 15:55:00 +0100 Laurent Vivier wrote:

...

On 27/11/2024 15:48, Stefano Brivio wrote:

...

On Wed, 27 Nov 2024 15:41:30 +0100 Laurent Vivier wrote:

...

there is a conflict between netinet/if_ether.h provided by musl and linux/if_ether.h provided by the linux headers:

In file included from passt.h:185, from tcp_vu.c:21: /usr/include/netinet/if_ether.h:115:8: error: redefinition of 'struct ethhdr' 115 | struct ethhdr { | ^~~~~~ In file included from /usr/include/linux/virtio_net.h:32, from tcp_vu.c:17: /usr/include/linux/if_ether.h:173:8: note: originally defined here 173 | struct ethhdr { | ^~~~~~

The kernel headers provide a flag to disable the definition in this case, __UAPI_DEF_ETHHDR (see /usr/include/linux/if_ether.h comment).

Signed-off-by: Laurent Vivier --- Makefile | 3 +++ 1 file changed, 3 insertions(+)

diff --git a/Makefile b/Makefile index cb7448079de5..2aa56ada65fd 100644 --- a/Makefile +++ b/Makefile @@ -33,6 +33,9 @@ FLAGS += $(FORTIFY_FLAG) -O2 -pie -fPIE FLAGS += -DPAGE_SIZE=$(shell getconf PAGE_SIZE) FLAGS += -DVERSION=\"$(VERSION)\" FLAGS += -DDUAL_STACK_SOCKETS=$(DUAL_STACK_SOCKETS) +ifeq (musl, $(word 4,$(subst -, ,$(TARGET)))) +FLAGS += -D__UAPI_DEF_ETHHDR=0 +endif

Uh oh, I just solved this locally (I was about to push, but I stopped in time) by simply including instead of from passt.h.

I'm not sure what's the best solution: mine is simpler, but I'm not really fond of including Linux-specific stuff if there's no need for it, so I'd slightly prefer yours.

On the other hand, it's not just for musl. If somebody tries to build this against, say, uClibc-ng (no reports about that yet), they will probably run into trouble as well.

In fact, in /usr/include/linux/if_ether.h they say:

/* allow libcs like musl to deactivate this, glibc does not implement this. */

So I think musl should set the value and then I found:

#define __UAPI_DEF_ETHHDR 0

in /usr/include/netinet/if_ether.h

Ah, I hadn't checked that yet.

...

So I think the solution is to include netinet/if_ether.h before linux/if_ether.h

Just in tcp_vu.c you mean? That's where we include now.

Any advantage over simply including linux/if_ether.h in passt.h as I did?

passt.h remains generic by not including linux specific headers.

Something like that:

diff --git a/tcp_vu.c b/tcp_vu.c index 05e2d1d..943b6d9 100644 --- a/tcp_vu.c +++ b/tcp_vu.c @@ -13,7 +13,7 @@ #include

#include - +#include #include

#include "util.h" diff --git a/vu_common.c b/vu_common.c index 0763ec0..299b5a3 100644 --- a/vu_common.c +++ b/vu_common.c @@ -8,6 +8,7 @@ #include #include #include +#include #include

#include "util.h"

Okay, that works as well, just tested on glibc and musl. I'll simply fold it in 7/9 so that we don't break bisection with musl. -- Stefano