Re: [PATCH] PoC: Rust binding for QAPI (qemu-ga only, for now)

[Paolo Bonzini <pbonzini@redhat.com>]

On 11/09/20 16:00, Marc-André Lureau wrote:
>     - from_qemu_none should be a "to_*" or "constructor" conversion (I used
>     new_from_foreign)
> 
> new_ prefix is not very rusty.

Right, I have changed it to with_foreign so now there is
{as,with,to,into}_foreign, plus unsafe_{from,into}.

These two could even be renamed to from_foreign and into_native at the
cost of making the trait less general purpose.  This way we have the
typical Rust names: as_* for Borrowed->Borrowed, with_*/to_* for
Borrowed->Owned, from_*/into_* for Owned->Owned.

> However, the memory allocator (or the stack) may not be compatible
> with the one used in C.

Hmm, that's a good point.  The simplest solution might be just to get
rid of IntoForeign, it's just an optimization.

> from_raw() is common, and takes ownership.

from_raw()/into_raw() would be equivalent to
into_foreign()/from_foreign().  However as you point out the allocators
are different, so it's a good idea IMHO to separate
into_raw()/from_raw() for the Rust allocator from
into_foreign()/from_foreign() for the libc allocator.

> I would need to modify this PoC for example

Yes of course.  Can you just try splitting the PoC in multiple patches?
 That should also make it easier to review, so far all I did was
comparing against glib-rs.

> But I must say I feel quite comfortable with the glib approach. It
> would be nice to have some feedback from glib-rs maintainers about your
> proposal.

QAPI is not tied to glib-rs, so I don't think qemu-ga will need to use
glib-rs.  I think either we use glib-rs, or if we are to roll our own we
should not be tied to the naming.  We don't use GObject introspection,
so none/full means nothing to most QEMU developers (and to Rust
developers too).

There are other things I don't like very much in glib-rs, for example
the use of tuples and public fields and the somewhat messy usage of
*const/*mut (I tried to be stricter on that).

Paolo