[Intel-wired-lan] [tnguy-net-queue:dev-queue 2/12] include/linux/build_bug.h:78:41: error: static assertion failed: "sizeof(__uapi_uuid_t) == sizeof(uuid_t) && __alignof__(__uapi_uuid_t) == __alignof__(uuid_t)"

[Dan Williams]

Jason Gunthorpe wrote:
> On Thu, Apr 10, 2025 at 10:48:00AM -0700, Dan Williams wrote:
> > Jason Gunthorpe wrote:
> > > On Wed, Apr 09, 2025 at 07:01:29PM -0700, Dan Williams wrote:
> > > 
> > > > diff --git a/include/uapi/cxl/features.h b/include/uapi/cxl/features.h
> > > > index dd8874860cec..06a1ae3f3fd0 100644
> > > > --- a/include/uapi/cxl/features.h
> > > > +++ b/include/uapi/cxl/features.h
> > > > @@ -14,7 +14,8 @@ typedef unsigned char __uapi_uuid_t[16];
> > > >  #ifdef __KERNEL__
> > > >  #include <linux/uuid.h>
> > > >  static_assert(sizeof(__uapi_uuid_t) == sizeof(uuid_t) &&
> > > > -	      __alignof__(__uapi_uuid_t) == __alignof__(uuid_t));
> > > > +	      __alignof__(struct { __uapi_uuid_t uuid; }) ==
> > > > +		      __alignof__(struct { uuid_t uuid; }));
> > > 
> > > Really? I'm surprised that the struct wrapper increases the
> > > alignment..
> > 
> > I was also surprised that gcc has different rules for this alignment
> > across compilers. Empirically this change solves the assertion, but I
> > admit this was just the result of tinkering until the error goes away.
> 
> Hurm.
> 
> So the reason to have the align check is to ensure that when it is
> embedded in the structs it doesn't change the layout of the struct.
> 
> The only use is this:
> 
> struct cxl_feat_entry {
>         uuid_t uuid;
>         __le16 id;
> 
> Which is fine, but if instead it was:
> 
> struct cxl_feat_entry {
>         __le16 id;
>         uuid_t uuid;
> 
> Then you have a problem as the align by 1 version will have no
> padding while the align by 4 will have 2 bytes padding.
> 
> Wrapping in a struct for the static_assert does not prevent the above
> issue..
> 
> So maybe the answer is to drop the alignof check entirely and add a
> comment explaining that uuid must only be placed on 4 byte aligned
> offsets and all structs must have explicit padding.

Yeah, that's better. Ignore my v2, will respin with a v3.

The other mitigating factor is that all of the usage of __uapi_uuid_t in
the header is in __packed structures.