On Tue, May 24, 2016 at 06:36:02PM -0400, Rich Felker wrote:
> On Tue, May 24, 2016 at 03:07:35PM -0700, William Ahern wrote:
> > GCC 6.1 more aggressively decomposes aggregate assignments into a series of
> > scalar member assignments. This has uncovered an issue with glibc's layout
> > of struct sockaddr_storage, which has a padding hole from offsets 2 to 8,
> > precisely where .sin_port and .sin_addr are in struct sockaddr_in.
> > 
> >   https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71120
> > 
> > musl shares this same issue. Specifically, the __ss_align member with an
> > 8-byte alignment on LP64 archs. You can track the glibc resolution at
> > 
> >   https://sourceware.org/bugzilla/show_bug.cgi?id=20111
> > 
> > Or not track it. Reasonable folks can disagree regarding many aspects of
> > this issue, but I thought it worthwhile to bring to people's attention.
> 
> I maintain that it's a bug (violation of effective type rules) for a
> program to attempt to copy sockaddr types using sockaddr_storage, but
> this is a nasty application bug to track down (usually silent
> breakage) that's worth avoiding since it's easy. Does the attached
> patch work?
> 
> I don't think we should even consider the sorts of may_alias hacks
> glibc/gcc folks are discussing, though. There's already a gcc option
> for compiling broken code like that; it's called -fno-strict-aliasing.
> 
> Rich

> diff --git a/include/sys/socket.h b/include/sys/socket.h
> index 6788375..d2bd5df 100644
> --- a/include/sys/socket.h
> +++ b/include/sys/socket.h
> @@ -286,7 +286,7 @@ struct sockaddr
>  
>  struct sockaddr_storage
>  {
> -	sa_family_t ss_family;
> +	sa_family_t ss_family, __ss_family_pad;
>  	unsigned long __ss_align;
>  	char __ss_padding[128-2*sizeof(unsigned long)];
>  };

This is wrong for 64-bit archs; new version attached.

Rich