On Tue, Jan 09, 2024 at 02:34:10PM +0100, Linus Walleij wrote:
> On Tue, Jan 9, 2024 at 2:24 PM Mukesh Ojha <quic_mojha@xxxxxxxxxxx> wrote:
> > On 1/9/2024 6:44 PM, Linus Walleij wrote:
> > > On Mon, Jan 8, 2024 at 4:28 PM Mukesh Ojha <quic_mojha@xxxxxxxxxxx> wrote:
> > >
> > >> It was realized by Srinivas K. that there is a need of
> > >> read-modify-write scm exported function so that it can
> > >> be used by multiple clients.
> > >>
> > >> Let's introduce qcom_scm_io_rmw() which masks out the bits
> > >> and write the passed value to that bit-offset.
> > > (...)
> > >> +int qcom_scm_io_rmw(phys_addr_t addr, unsigned int mask, unsigned int val)
> > >> +{
> > >> + unsigned int old, new;
> > >> + int ret;
> > >> +
> > >> + if (!__scm)
> > >> + return -EINVAL;
> > >> +
> > >> + spin_lock(&__scm->lock);
> > >> + ret = qcom_scm_io_readl(addr, &old);
> > >> + if (ret)
> > >> + goto unlock;
> > >> +
> > >> + new = (old & ~mask) | (val & mask);
> > >> +
> > >> + ret = qcom_scm_io_writel(addr, new);
> > >> +unlock:
> > >> + spin_unlock(&__scm->lock);
> > >> + return ret;
> > >> +}
> > >> +EXPORT_SYMBOL_GPL(qcom_scm_io_rmw);
> > >
> > > This looks a lot like you are starting to re-invent regmaps
> > > regmap_update_bits().
> > >
> > > If you are starting to realize you need more and more of
> > > regmap, why not use regmap and its functions?
> >
> > I think, this discussion has happened already ..
> >
> > https://lore.kernel.org/lkml/CACRpkdb95V5GC81w8fiuLfx_V1DtWYpO33FOfMnArpJeC9SDQA@xxxxxxxxxxxxxx/
>
> That discussion ended with:
>
> [Bjorn]
> > We'd still need qcom_scm_io_readl() and qcom_scm_io_writel() exported to
> > implement the new custom regmap implementation - and the struct
> > regmap_config needed in just pinctrl-msm alone would be larger than the
> > one function it replaces.
>
> When you add more and more accessors the premise starts to
> change, and it becomes more and more of a reimplementation.
>
> It may be time to actually fix this.
>
Thought I had replied to this already, did we discuss this previously as
well?
My concern with expressing this as a regmap is that from the provider's
point of view, the regmap would span the entire 32-bit address space.
I'm guessing that there's something on the other side limiting what
subregions are actually accessible for each platform/firmware
configuration, but I'm not convinced that regmap a good abstraction...
Regards,
Bjorn
> Yours,
> Linus Walleij
