Re: [PATCH v2 3/3] crypto: hisilicon/qm - defining the device isolation strategy

[yekai(A)]

On 2022/6/14 21:29, Greg KH wrote:
> On Tue, Jun 14, 2022 at 09:24:08PM +0800, yekai(A) wrote:
> > > >  struct hisi_qm {
> > > >  	enum qm_hw_ver ver;
> > > >  	enum qm_fun_type fun_type;
> > > > @@ -335,6 +341,9 @@ struct hisi_qm {
> > > >  	struct qm_shaper_factor *factor;
> > > >  	u32 mb_qos;
> > > >  	u32 type_rate;
> > > > +	struct list_head uacce_hw_errs;
> > > > +	atomic_t uacce_ref; /* reference of the uacce */
> > >
> > > That is not how reference counts work, sorry.  Please use 'struct kref'
> > > for a real reference count, never roll your own.
> > >
> > > thanks,
> > >
> > > greg k-h
> > > .
> >
> > this atomic_t reference is lightweight than 'struct kref',
>
> It's the same size, why would it be "lighter"?  Why do you need it to be
> lighter, what performance issue is there with a kref?
>
> > this reference
> > means whether the task is running. So would it be better to use atomic_t
> > reference?
>
> I do not know, as "running or not running" is a state, not a count or a
> reference.  why does this have to be atomic at all?
>
> thanks,
>
> greg k-h
> .

I will use 'qm_state' instead of reference count by zhangfei Gao's opinion.

Thanks
Kai