Re: [hurd-amd64] ibus test failures

[Samuel Thibault]

Diego Nieto Cid, le mar. 31 déc. 2024 02:47:35 +0000, a ecrit:
> On Tue, Dec 31, 2024 at 02:02:53AM +0000, Diego Nieto Cid wrote:
> >
> > Here are the results for `hurd-amd64`:
> > 
> >     $ cc -I/usr/include/glib-2.0 -I/usr/lib/x86_64-gnu/glib-2.0/include \
> >         -pthread  cred_size.c  -lgio-2.0 -lgobject-2.0 -lglib-2.0 \
> >         -o cred_size
> > 
> >     $ ./cred_size
> >     sizeof (int): 4
> >     sizeof(struct cmsgcred): 84
> >     CMSG_LEN(sizeof (int)): 20
> >     CMSG_LEN(g_socket_control_message_get_size (scm)): 100
> > 
> > and here the results for `hurd-i386`:
> > 
> >     $ ./cred_size
> >     sizeof (int): 4
> >     sizeof(struct cmsgcred): 84
> >     CMSG_LEN(sizeof (int)): 16
> >     CMSG_LEN(g_socket_control_message_get_size (scm)): 96
> >
> 
> I think the usage of size_t is what breaks the computation on
> hurd-amd64 (8 bytes while on i386 is 4):
> 
>     #define CMSG_ALIGN(len) (((len) + sizeof (size_t) - 1) \
>                                & (size_t) ~(sizeof (size_t) - 1))

That's how it is defined in Linux too. CMSG_ALIGN really needs to be
aligned so.

>     #define CMSG_SPACE(len) (CMSG_ALIGN (len) \
>                              + CMSG_ALIGN (sizeof (struct cmsghdr)))
>     #define CMSG_LEN(len)   (CMSG_ALIGN (sizeof (struct cmsghdr)) + (len))

This alignment is also required.

> I'm not sure how to fix it. Should the expectations of GLib be updated
> to be aligned? Like below:
> 
> - if (size != G_CREDENTIALS_NATIVE_SIZE)
> + if (size != CMSG_ALIGN(G_CREDENTIALS_NATIVE_SIZE))
> 
> Or should the macros be updated to use a 4 bytes size_t equivalent?

I'd rather say it's our socket layer which is aligning cmsg_len while
it shouldn't, since it's CMSG_NXTHDR that aligns it to get to the next
header.

Samuel