Re: [PATCH v2 3/4] iommu/mediatek: Use writel for TLB range invalidation

[Will Deacon]

On Sat, Oct 12, 2019 at 02:23:47PM +0800, Yong Wu wrote:
> On Fri, 2019-10-11 at 17:29 +0100, Will Deacon wrote:
> > On Wed, Oct 09, 2019 at 09:19:02PM +0800, Yong Wu wrote:
> > > Use writel for the register F_MMU_INV_RANGE which is for triggering the
> > > HW work. We expect all the setting(iova_start/iova_end...) have already
> > > been finished before F_MMU_INV_RANGE.
> > > 
> > > Signed-off-by: Anan.Sun <anan.sun@xxxxxxxxxxxx>
> > > Signed-off-by: Yong Wu <yong.wu@xxxxxxxxxxxx>
> > > ---
> > > This is a improvement rather than fixing a issue.
> > > ---
> > >  drivers/iommu/mtk_iommu.c | 3 +--
> > >  1 file changed, 1 insertion(+), 2 deletions(-)
> > > 
> > > diff --git a/drivers/iommu/mtk_iommu.c b/drivers/iommu/mtk_iommu.c
> > > index 24a13a6..607f92c 100644
> > > --- a/drivers/iommu/mtk_iommu.c
> > > +++ b/drivers/iommu/mtk_iommu.c
> > > @@ -187,8 +187,7 @@ static void mtk_iommu_tlb_add_flush(unsigned long iova, size_t size,
> > >  		writel_relaxed(iova, data->base + REG_MMU_INVLD_START_A);
> > >  		writel_relaxed(iova + size - 1,
> > >  			       data->base + REG_MMU_INVLD_END_A);
> > > -		writel_relaxed(F_MMU_INV_RANGE,
> > > -			       data->base + REG_MMU_INVALIDATE);
> > > +		writel(F_MMU_INV_RANGE, data->base + REG_MMU_INVALIDATE);
> > 
> > I don't understand this change.
> > 
> > Why is it an "improvement" and which accesses are you ordering with the
> > writel?
> 
> The register(F_MMU_INV_RANGE) will trigger HW to begin flush range. HW
> expect the other register iova_start/end/flush_type always is ready
> before trigger. thus I'd like use writel to guarantee the previous
> register has been finished.

Given that these are all MMIO writes to the same device, then
writel_relaxed() should give you the ordering you need. If you look at
memory_barriers.txt, it says:

  | they [readX_relaxed() and writeX_relaxed()] are still guaranteed to
  | be ordered with respect to other accesses from the same CPU thread
  | to the same peripheral when operating on __iomem pointers mapped
  | with the default I/O attributes.

> I didn't see the writel_relaxed cause some error in practice, we only
> think writel is necessary here in theory. so call it "improvement".

Ok, but I don't think it's needed in this case.

Will