Re: [PATCH 08/11] rust/block: Add driver module

[Paolo Bonzini <pbonzini@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 1214 bytes --]

Il mer 12 feb 2025, 18:32 Kevin Wolf <kwolf@redhat.com> ha scritto:

> > > +        mut buf: MaybeUninit<T>,
> >
> > I think Rust doesn't guarantee no copies here, so maybe this could be
>
> Do you think that in practice the compiler won't optimise the copy away?
>

It's possiblr that it does not, because it has to build the io::Result and
stick the result of assume_init() in there. It all depends on the amount of
inlining perhaps?

I think Box<MaybeUninit>> is the only way to guarantee no copies
(assume_init for Box was only stabilized recently but it can be emulated
with Box::into_raw and Box::from_raw).

>    pub async fn read_uninit<T: SizedIoBuffer>(
> >        &self,
> >        offset: u64,
> >        buf: &mut MaybeUninit<T>,
> >     ) -> io::Result<&mut T>
> >
> > using assume_init_mut().
>
> Are you sure that callers are ok with only getting a &mut T rather than
> an owned T?
>

The one you have would need to be adjusted but it would work.

Another possibility by the way is to have "pub async fn read_obj<T:
SizedIoBuffer>(&self, offset: u64) -> io::Result<T>" and hide the usage of
MaybeUninit inside the function... That one doesn't even try to avoid
copies though.

Paolo



> Kevin
>
>

[-- Attachment #2: Type: text/html, Size: 2424 bytes --]