Re: [hurd-amd64] ibus test failures

[Samuel Thibault]

Diego Nieto Cid, le mer. 01 janv. 2025 21:03:23 +0000, a ecrit:
> On Wed, Jan 01, 2025 at 08:07:04PM +0000, Diego Nieto Cid wrote:
> > On Wed, Jan 01, 2025 at 06:51:47PM +0100, Samuel Thibault wrote:
> > > > 
> > > >     /* Structure used for storage of ancillary data object information. 
> > > >  */
> > > >     struct cmsghdr
> > > >       {
> > > >         socklen_t cmsg_len;         /* Length of data in cmsg_data plus 
> > > > length
> > > >                                        of cmsghdr structure.  */
> > > 
> > > That's the correct type according to posix:
> > > 
> > > ./basedefs/sys_socket.h.html:<tt>socklen_t  cmsg_len   </tt> Data byte 
> > > count, including the <b>cmsghdr</b>. <tt>
> > >
> > 
> > Ok, but we are giving unaligned data pointers with the current 
> > implementation
> > of CMSG_DATA.
> 
> On the other hand, nothing says the pointer shall be aligned:
> 
>     CMSG_DATA(cmsg) [1]
> 
>     If the argument is a pointer to a cmsghdr structure, this macro
>     shall return an unsigned character pointer to the data array
>     associated with the cmsghdr structure.

But on various archs, unaligned 64b-value access will SIGBUS.

> [1] 
> https://pubs.opengroup.org/onlinepubs/9699919799/basedefs/sys_socket.h.html
> 
> and GLib computation is relying on undefined behaviour.

No, it's CMSG_DATA and CMSG_LEN which are incoherent. CMSG_LEN shouldn't
be aligning on size_t if CMSG_DATA was just using the __cmsg_data field.
The sys/socket.h header has to be fixed one way or the other. Better
align ourself on the BSD behavior.

Samuel