Re: [RFC 2/5] rust: sync: Arc: Introduces ArcInner::count()

[Boqun Feng]

On Fri, Feb 03, 2023 at 08:38:25AM +0100, Greg KH wrote:
> On Thu, Feb 02, 2023 at 11:25:08PM -0800, Boqun Feng wrote:
> > On Fri, Feb 03, 2023 at 06:22:15AM +0100, Greg KH wrote:
> > > On Thu, Feb 02, 2023 at 10:47:12PM +0100, Miguel Ojeda wrote:
> > > > On Thu, Feb 2, 2023 at 5:52 PM Boqun Feng <boqun.feng@xxxxxxxxx> wrote:
> > > > >
> > > > > As I said, I'm open to remove the printing of the refcount, and if you
> > > > > and Peter think maybe it's OK to do that after the explanation above,
> > > > 
> > > > Perhaps part of the confusion came from the overloaded "safe" term.
> > > > 
> > > > When Gary and Boqun used the term "safe", they meant it in the Rust
> > > > sense, i.e. calling the method will not allow to introduce undefined
> > > > behavior. While I think Peter and Greg are using the term to mean
> > > > something different.
> > > 
> > > Yes, I mean it in a "this is not giving you the value you think you are
> > > getting and you can not rely on it for anything at all as it is going to
> > > be incorrect" meaning.
> > > 
> > > Which in kernel code means "this is not something you should do".
> > > 
> > 
> > Now what really confuses me is why kref_read() is safe..
> 
> It isn't, and I hate it and it should be removed from the kernel
> entirely.  But the scsi and drm developers seem to insist that "their
> locking model ensures it will be safe to use" and I lost that argument
> :(
> 
> > or how this is different than kref_read().
> 
> It isn't, but again, I don't like that and do not agree it should be
> used as it is almost always a sign that the logic in the code is
> incorrect.
> 
> > Needless to say that ArcInner::count() can guarantee not reading 0
> 
> How?  Because you have an implicit reference on it already?  If so, then
> why does reading from it matter at all, as if you have a reference, you
> know it isn't 0, and that's all that you can really care about.  You
> don't care about any number other than 0 for a reference count, as by
> definition, that's what a reference count does :)
> 

Fair enough!

> > (because of the type invariants) but kref_read() cannot..
> 
> I totally agree with you.  Let's not mirror bad decisions of legacy
> subsystems in the kernel written in C with new designs in Rust please.
> 

Roger that, will remove this in the next version ;-)

Regards,
Boqun

> thanks,
> 
> greg k-h