Re: [PATCH v2] docs/misra: document the expected sizes of integer types

[Jan Beulich <jbeulich@xxxxxxxx>]

On 19.03.2024 04:37, Stefano Stabellini wrote:
> On Mon, 18 Mar 2024, Jan Beulich wrote:
>> On 16.03.2024 01:07, Stefano Stabellini wrote:
>>> On Fri, 15 Mar 2024, Jan Beulich wrote:
>>>> On 14.03.2024 23:17, Stefano Stabellini wrote:
>>>>> Xen makes assumptions about the size of integer types on the various
>>>>> architectures. Document these assumptions.
>>>>
>>>> My prior reservation wrt exact vs minimum sizes remains.
>>>
>>> We have to specify the exact size. In practice the size is predetermined
>>> and exact with all our supported compilers given a architecture.
>>
>> But that's not the purpose of this document; if it was down to what
>> compilers offer, we could refer to compiler documentation (and iirc we
>> already do for various aspects). The purpose of this document, aiui,
>> is to document assumption we make in hypervisor code. And those should
>> be >=, not ==.
> 
> Well... I guess the two of us are making different assumptions then :-)
> 
> Which is the reason why documenting assumptions is so important. More at
> the bottom.
> 
> 
>>> Most importantly, unfortunately we use non-fixed-size integer types in
>>> C hypercall entry points and public ABIs. In my opinion, that is not
>>> acceptable.
>>
>> The problem is that I can't see the reason for you thinking so. The C
>> entry points sit past assembly code doing (required to do) necessary
>> adjustments, if any. If there was no assembly layer, whether to use
>> fixed with types for such parameters would depend on what the
>> architecture guarantees.
> 
> This could be the source of the disagreement. I see the little assembly
> code as not important, I consider it just like a little trampoline to
> me. As we describe the hypercalls in C header files, I consider the C
> functions the "official" hypercall entry points.

Why would that be? Any code we execute in Xen is relevant.

> Also, as this is an ABI, I consider mandatory to use clear width
> definitions of all the types (whether with this document or with
> fixed-width types, and fixed-width types are clearer and better) in both
> the C header files that describe the ABI interfaces, as well as the C
> entry points that corresponds to it. E.g. I think we have to use
> the same types in both do_sched_op and the hypercall description in
> xen/include/public/sched.h

There are two entirely separate aspects to the ABI: One is what we
document towards consumers of it. The other is entirely internal, i.e.
an implementation detail - how we actually consume the data.
Documenting fixed-width types towards consumers is probably okay,
albeit (see below) imo still not strictly necessary (for being
needlessly limiting).

>> As to public ABIs - that's structure definitions, and I agree we ought
>> to uniformly use fixed-width types there. We largely do; a few things
>> still require fixing.
> 
> +1
> 
> 
>>> We have two options:
>>>
>>> 1) we go with this document, and we clarify that even if we specify
>>>   "unsigned int", we actually mean a 32-bit integer
>>>
>>> 2) we change all our public ABIs and C hypercall entry points to use
>>>    fixed-size types (e.g. s/unsigned int/uint32_t/g)
>>>
>>> 2) is preferred because it is clearer but it is more work. So I went
>>> with 1). I also thought you would like 1) more.
>>
>> For ABIs (i.e. structures) we ought to be making that change anyway.
>> Leaving basic types in there is latently buggy.
> 
> I am glad we agree :-)
> 
> It is just that I also consinder the C hypercall entry points as part of
> the ABI
> 
> 
>> I'm happy to see a document like this added, for the purpose described
>> above. But to me 1) and 2) and largely independent of one another.
> 
> Good that you are also happy with a document like this.
> 
> The remaining question is: what about the rest of the C functions in Xen
> that are certainly not part of an ABI?

As per above - anything internal isn't part of the ABI, C entry points
for hypercall handlers included. All we need to ensure is that we consume
the data according to what the ABI sets forth.

To use wording from George when he criticized my supposed lack of actual
arguments: While there's nothing technically wrong with using fixed
width types there (or in fact everywhere), there's also nothing technically
wrong with using plain C types there and almost everywhere else (ABI
structures excluded). With both technically equal, ./CODING_STYLE has the
only criteria to pick between the two. IOW that's what I view wrong in
George's argumentation: Demanding that I provide technical arguments when
the desire to use fixed width types for the purpose under discussion also
isn't backed by any.

> Those are less critical, still this document should apply uniformily to
> them too. I don't understand why you are making the >= width assumption
> you mentioned at the top of the file when actually it is impossible to
> exercise or test this assumption on any compiler or any architecture
> that works with Xen. If it cannot be enabled, it hasn't been tested, and
> it probably won't work.

Hmm, yes, that's one way to look at it. My perspective is different though:
By writing down assumptions that are more strict than necessary, we'd be
excluding ports to environments meeting the >= assumption, but not meeting
the == one. Unless of course you can point me at any place where - not
just by mistake / by being overly lax - we truly depend on the == that you
want to put in place. IOW yes, there likely would need to be adjustments
to code if such a port was to happen. Yet we shouldn't further harden
requirements that were never meant to be there.

Note that by writing down anything more strict than necessary, you'd also
encourage people to further wrongly treat e.g. uint32_t and unsigned int
as identical. Such wrong assumptions had been a severe hindrance in doing
ports from 32- to 64-bit processors some 20 years ago. I would have hoped
that we'd learn from such mistakes.

Jan