Re: [PATCH 05/16] bus: mhi: core: Add support for ringing channel/event ring doorbells

[Jeffrey Hugo]

On 1/23/2020 5:00 AM, Manivannan Sadhasivam wrote:
> Hi Arnd,
>
> On Thu, Jan 23, 2020 at 12:39:06PM +0100, Arnd Bergmann wrote:
> > On Thu, Jan 23, 2020 at 12:19 PM Manivannan Sadhasivam
> > <manivannan.sadhasivam@xxxxxxxxxx> wrote:
> >
> > > +int __must_check mhi_read_reg(struct mhi_controller *mhi_cntrl,
> > > +                             void __iomem *base, u32 offset, u32 *out)
> > > +{
> > > +       u32 tmp = readl_relaxed(base + offset);
> > ....
> > > +void mhi_write_reg(struct mhi_controller *mhi_cntrl, void __iomem *base,
> > > +                  u32 offset, u32 val)
> > > +{
> > > +       writel_relaxed(val, base + offset);
> >
> > Please avoid using _relaxed accessors by default, and use the regular
> > ones instead. There are a number of things that can go wrong with
> > the relaxed version, so ideally each caller should have a comment
> > explaining why this instance is safe without the barriers and why it
> > matters to not have it.
> >
> > If there are performance critical callers of mhi_read_reg/mhi_write_reg,
> > you could add mhi_read_reg_relaxed/mhi_write_reg_relaxed for those
> > and apply the same rules there.
> >
> > Usually most mmio accesses are only needed for reconfiguration or
> > other slow paths.
>
> Fair point. I'll defer to readl/writel APIs and I also need to add
> le32_to_cpu/cpu_to_le32 to them.

I would expect we would be using these in the "hot" path.

I'm a bit confused, I thought the convention was to put a comment why a 
barrier was necessary, now we should be putting a comment why a barrier 
is not necessary?


--
Jeffrey Hugo
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.