Re: [PATCH 04/13] Kbuild: Rust support

[Willy Tarreau <w@1wt.eu>]

On Fri, Apr 16, 2021 at 11:39:00PM +0200, Miguel Ojeda wrote:
> On Fri, Apr 16, 2021 at 10:58 PM Willy Tarreau <w@1wt.eu> wrote:
> >
> > No, two:
> >   - ok in %rax (seems like it's "!ok" technically speaking since it
> >     returns 1 on !ok and 0 on ok)
> >   - foo_or_err in %rdx
> 
> Yes, but that is the implementation -- conceptually you only have one
> or the other, and Rust won't allow you to use the wrong one.

OK so for unions you always pass two values along the whole chain, a
selector and the value itself.

But my point remains that the point of extreme care is at the interface
with the rest of the kernel because there is a change of semantics
there.

> > However then I'm bothered because Miguel's example showed that regardless
> > of OK, EINVAL was always returned in foo_or_err, so maybe it's just
> > because his example was not well chosen but it wasn't very visible from
> > the source:
> 
> That is the optimizer being fancy since the error can be put
> unconditionally in `rdx`.

Yes that's what I understood as well. I just didn't know that it had
to be seen as a union.

On Fri, Apr 16, 2021 at 11:19:18PM +0200, Miguel Ojeda wrote:
> On Fri, Apr 16, 2021 at 10:22 PM Willy Tarreau <w@1wt.eu> wrote:
> >
> > So it simply does the equivalent of:
> >
> >   struct result {
> >      int status;
> >      int error;
> >   };
> 
> Not exactly, it is more like a tagged union, as Connor mentioned.
> 
> However, and this is the critical bit: it is a compile-time error to
> access the inactive variants (in safe code). In C, it is on you to
> keep track which one is the current one.

Sure but as I said most often (due to API or ABI inheritance), both
are already exclusive and stored as ranges. Returning 1..4095 for
errno or a pointer including NULL for a success doesn't shock me at
all.

Along thes lines I hardly see how you'd tag pointers by manipulating
their lower unused bits. That's something important both for memory
usage and performance (supports atomic opts).

> >      kill_foo();   // only for rust, C doesn't need it
> 
> Please note that `kill_foo()` is not needed in Rust -- it was an
> example of possible cleanup (since Al mentioned resources/cleanup)
> using RAII.

Yep but I kept it just to have comparable output code since in C
you'd simply use "goto leave" and not have this function call to
do the cleanup.

Willy