Re: [PATCH] drm/msm/adreno: De-spaghettify the use of memory barriers

[Will Deacon <will@kernel.org>]

On Thu, Jun 06, 2024 at 02:03:24PM +0200, Konrad Dybcio wrote:
> On 4.06.2024 4:40 PM, Will Deacon wrote:
> > On Thu, May 16, 2024 at 01:55:26PM -0500, Andrew Halaney wrote:
> >> On Thu, May 16, 2024 at 08:20:05PM GMT, Akhil P Oommen wrote:
> >>> On Thu, May 16, 2024 at 08:15:34AM -0500, Andrew Halaney wrote:
> >>>> If I understand correctly, you don't need any memory barrier.
> >>>> writel()/readl()'s are ordered to the same endpoint. That goes for all
> >>>> the reordering/barrier comments mentioned below too.
> >>>>
> >>>> device-io.rst:
> >>>>
> >>>>     The read and write functions are defined to be ordered. That is the
> >>>>     compiler is not permitted to reorder the I/O sequence. When the ordering
> >>>>     can be compiler optimised, you can use __readb() and friends to
> >>>>     indicate the relaxed ordering. Use this with care.
> >>>>
> >>>> memory-barriers.txt:
> >>>>
> >>>>      (*) readX(), writeX():
> >>>>
> >>>> 	    The readX() and writeX() MMIO accessors take a pointer to the
> >>>> 	    peripheral being accessed as an __iomem * parameter. For pointers
> >>>> 	    mapped with the default I/O attributes (e.g. those returned by
> >>>> 	    ioremap()), the ordering guarantees are as follows:
> >>>>
> >>>> 	    1. All readX() and writeX() accesses to the same peripheral are ordered
> >>>> 	       with respect to each other. This ensures that MMIO register accesses
> >>>> 	       by the same CPU thread to a particular device will arrive in program
> >>>> 	       order.
> >>>>
> >>>
> >>> In arm64, a writel followed by readl translates to roughly the following
> >>> sequence: dmb_wmb(), __raw_writel(), __raw_readl(), dmb_rmb(). I am not
> >>> sure what is stopping compiler from reordering  __raw_writel() and __raw_readl()
> >>> above? I am assuming iomem cookie is ignored during compilation.
> >>
> >> It seems to me that is due to some usage of volatile there in
> >> __raw_writel() etc, but to be honest after reading about volatile and
> >> some threads from gcc mailing lists, I don't have a confident answer :)
> >>
> >>>
> >>> Added Will to this thread if he can throw some light on this.
> >>
> >> Hopefully Will can school us.
> > 
> > The ordering in this case is ensured by the memory attributes used for
> > ioremap(). When an MMIO region is mapped using Device-nGnRE attributes
> > (as it the case for ioremap()), the "nR" part means "no reordering", so
> > readX() and writeX() to that region are ordered wrt each other.
> > 
> > Note that guarantee _doesn't_ apply to other flavours of ioremap(), so
> > e.g. ioremap_wc() won't give you the ordering.
> > 
> > Hope that helps,
> 
> Just to make sure I'm following, would mapping things as nGnRnE effectively
> get rid of write buffering, perhaps being a way of debugging whether that
> in particular is causing issues (at the cost of speed)?

I think the "nE" part is just a hint, so it will depend on how the
hardware has been built. On top of that, you'll still need something
like a DSB to force the CPU to wait for the write response.

Will