Re: [Tee-dev] [PATCHv8 1/3] optee: use uuid for sysfs driver entry

[Daniel Thompson]

On Fri, Jun 26, 2020 at 10:40:41AM +0530, Sumit Garg wrote:
> On Fri, 26 Jun 2020 at 05:01, James Bottomley
> <James.Bottomley@xxxxxxxxxxxxxxxxxxxxx> wrote:
> >
> > On Thu, 2020-06-25 at 19:54 +0530, Sumit Garg wrote:
> > > On Wed, 24 Jun 2020 at 20:51, James Bottomley
> > > <James.Bottomley@xxxxxxxxxxxxxxxxxxxxx> wrote:
> > > >
> > > > On Wed, 2020-06-24 at 16:17 +0530, Sumit Garg wrote:
> > > > > Apologies for delay in my reply as I was busy with some other
> > > > > stuff.
> > > > >
> > > > > On Fri, 19 Jun 2020 at 20:30, James Bottomley
> > > > > <James.Bottomley@xxxxxxxxxxxxxxxxxxxxx> wrote:
> > > >
> > > > [...]
> > > > > > it's about consistency with what the kernel types mean.  When
> > > > > > some checker detects your using little endian operations on a
> > > > > > big endian structure (like in the prink for instance) they're
> > > > > > going to keep emailing you about it.
> > > > >
> > > > > As mentioned above, using different terminology is meant to cause
> > > > > more confusion than just difference in endianness which is
> > > > > manageable inside TEE.
> > > > >
> > > > > And I think it's safe to say that the kernel implements UUID in
> > > > > big endian format and thus uses %pUb whereas OP-TEE implements
> > > > > UUID in little endian format and thus uses %pUl.
> > > >
> > > > So what I think you're saying is that if we still had uuid_be and
> > > > uuid_le you'd use uuid_le, because that's exactly the structure
> > > > described in the docs.  But because we renamed
> > > >
> > > > uuid_be -> uuid_t
> > > > uuid_le -> guid_t
> > > >
> > > > You can't use guid_t as a kernel type because it has the wrong
> > > > name?
> > >
> > > Isn't the rename commit description [1] pretty clear about which is
> > > the true UUID type from Linux point of view?
> >
> > I don't think the kernel code takes a position on eternal verity, just
> > on logical or arithmetic truth.  We just have to deal with both LE and
> > BE UUIDs so we have appropriate types for them and the LE type is now
> > named guid_t.  They're both equally correct to use provided the use
> > case matches the designed one. So does the name really matter?
> 
> Yes it does. I guess I have provided enough reasoning for that. Also,
> the rename commit itself illustrates its importance and clarifies the
> use case for which they are meant to be used.
> 
> > If we
> > did
> >
> > #define uuid_le_t guid_t
> >
> > would you be happy? (not that the kernel cares about karmic emotional
> > states either ...)
> 
> It's not about me being happy but more about confusion and
> inconsistency it will bring.
> 
> IMO, either kernel should be opinionated about UUID endianness like
> currently it is:
> 
> uuid_t and its corresponding helpers (eg. UUID_INIT) follows BE format.
> 
> or support both endianness for UUID (no common type: uuid_t) like we
> had earlier prior to rename commit:
> 
> uuid_be_t and its corresponding helpers (eg. UUID_BE_INIT) follow BE format.
> uuid_le_t and its corresponding helpers (eg. UUID_LE_INIT) follow LE format.
> 
> But even if we consider later case as well, I am still not sure if we
> can switch to uuid_le_t as it's been part of TEE core ABI
> (open_session) where UUID is passed in BE format (see LE to BE
> conversion in TEE client [1] and vice-versa in OP-TEE OS [2]) and
> won't be a backwards compatible change.
> 
> [1] https://github.com/OP-TEE/optee_client/blob/master/libteec/src/tee_client_api.c#L595
> [2] https://github.com/OP-TEE/optee_os/blob/master/core/arch/arm/kernel/ree_fs_ta.c#L92

I'm struck that all references here are to code that does not run in
kernel space. Frankly on LKML I don't know if we should even *care* what
format UUIDs are stored in other address spaces.

We care about is the endianness of the UUID on the interface boundaries
into and out of the kernel[1] and we care that we use the correct kernel
type to describe it.

I understood from Jerome's post that the UUID that the kernel
manipulates are, in fact, big endian and that they should be called
uuid_t.

Is there more going on here or is that it?


Daniel.


[1] I guess we'd also like those interface boundaries to be reasonably
    sane but if they already exist and are the interface is consistent
    about endianness then I find it hard to see no reason for a
    gratuitous endianness change.