Re: [PATCH 1/2] afs: Move UUID struct to linux/uuid.h

[David Howells <dhowells@redhat.com>]

Arnd Bergmann <arnd@arndb.de> wrote:

> Looks good to me, but I wonder if this part:
> 
>                 r = call->request;
> -               r->time_low                     = ntohl(b[0]);
> -               r->time_mid                     = ntohl(b[1]);
> -               r->time_hi_and_version          = ntohl(b[2]);
> +               r->time_low                     = b[0];
> +               r->time_mid                     = htons(ntohl(b[1]));
> +               r->time_hi_and_version          = htons(ntohl(b[2]));
>                 r->clock_seq_hi_and_reserved    = ntohl(b[3]);
>                 r->clock_seq_low                = ntohl(b[4]);
>  
> should be considered a bugfix and split out into a
> separate patch.

I changed the definitions in the struct from u16/u32 to __be16/__be32 so it's
not a bugfix.

For some reason, rather than specifying UUIDs as just a 16-octet field, the
AFS protocol breaks the UUID down into pieces and converts them into 32-bit
fields (apparently signed in some places:-/).

> From what I understand about the mess in UUID formats, the time fields can
> either be big-endian (as defined) or little-endian (for all things
> Microsoft),

RFC 4122 specified that the multi-octet fields are stored MSB-first.

> and you are changing the representation from CPU-specific to big-endian,
> which makes it different for x86 and most ARM at least.

In-kernel, not in the protocol.

The problem is that you can't do what you put in your suggested patch and just
copy the UUID produced by the generate_random_uuid() over the afs_uuid struct
since that puts the version in the wrong place.

David