Re: [PATCH 1/3] rust: implement `kernel::sync::Refcount`

[Alice Ryhl <aliceryhl@google.com>]

On Sat, Oct 5, 2024 at 4:26 PM Gary Guo <gary@garyguo.net> wrote:
>
> On Sat, 5 Oct 2024 14:31:06 +0100
> Gary Guo <gary@garyguo.net> wrote:
>
> > On Sat, 5 Oct 2024 09:40:53 +0200
> > Greg KH <gregkh@linuxfoundation.org> wrote:
> >
> > > On Fri, Oct 04, 2024 at 04:52:22PM +0100, Gary Guo wrote:
> > > > This is a wrapping layer of `include/linux/refcount.h`. Currently only
> > > > the most basic operations (read/set/inc/dec/dec_and_test) are implemented,
> > > > additional methods can be implemented when they are needed.
> > > >
> > > > Currently the kernel refcount has already been used in `Arc`, however it
> > > > calls into FFI directly.
> > > >
> > > > Cc: Will Deacon <will@kernel.org>
> > > > Cc: Peter Zijlstra <peterz@infradead.org>
> > > > Cc: Boqun Feng <boqun.feng@gmail.com>
> > > > Cc: Mark Rutland <mark.rutland@arm.com>
> > > > Signed-off-by: Gary Guo <gary@garyguo.net>
> > > > ---
> > > >  rust/helpers/refcount.c      | 15 ++++++
> > > >  rust/kernel/sync.rs          |  2 +
> > > >  rust/kernel/sync/refcount.rs | 94 ++++++++++++++++++++++++++++++++++++
> > > >  3 files changed, 111 insertions(+)
> > > >  create mode 100644 rust/kernel/sync/refcount.rs
> > > >
> > > > diff --git a/rust/helpers/refcount.c b/rust/helpers/refcount.c
> > > > index f47afc148ec3..39649443426b 100644
> > > > --- a/rust/helpers/refcount.c
> > > > +++ b/rust/helpers/refcount.c
> > > > @@ -8,11 +8,26 @@ refcount_t rust_helper_REFCOUNT_INIT(int n)
> > > >   return (refcount_t)REFCOUNT_INIT(n);
> > > >  }
> > > >
> > > > +unsigned int rust_helper_refcount_read(refcount_t *r)
> > > > +{
> > > > + return refcount_read(r);
> > > > +}
> > >
> > > Reading a refcount is almost always a wrong thing to do (it can change
> > > right after you read it), and I don't see any of the later patches in
> > > this series use this call, so can you just drop this?
> > >
> > > thanks,
> > >
> > > greg k-h
> >
> > I originally introduced this thinking I can replace Andreas's atomic
> > 2->0 operation with a read + set, but ended up couldn't do it.
> >
> > The refcount read is still useful to determine if the current value is
> > 1 -- in fact, `Arc::into_unique_or_drop` could use this rather than
> > decrementing the refcount and then incrementing it again (just doing a
> > refcount read would be much better codegen-wise than the current
> > behaviour). I'll probably make this change in the next version of the
> > series.
>
> Actually `into_unique_or_drop` can't use this because it needs to avoid
> running destructor when it races with other threads. The semantics for
> that function is better reflected with `refcount_dec_not_one`, which
> I'll introduce in v2, and I'll drop `read` in v2.

Ah, I did not know that C had a refcount_dec_not_one. Yeah, that's
exactly what into_unique_or_drop does.

Though I don't know if the cmpxchg loop is really more efficient than
just doing an infallible decrement like I do right now?

Alice