[PATCH v3 5/6] iommu/mediatek: Add mt8173 IOMMU driver

[Will Deacon]

On Mon, Jul 27, 2015 at 05:24:31AM +0100, Yong Wu wrote:
> On Fri, 2015-07-24 at 17:55 +0100, Will Deacon wrote:
> > On Fri, Jul 24, 2015 at 06:43:13AM +0100, Yong Wu wrote:
> > > On Tue, 2015-07-21 at 15:59 +0100, Will Deacon wrote:
> > > > On Thu, Jul 16, 2015 at 10:04:34AM +0100, Yong Wu wrote:
> > > > > +static void mtk_iommu_tlb_flush_all(void *cookie)
> > > > > +{
> > > > > +       struct mtk_iommu_domain *domain = cookie;
> > > > > +       void __iomem *base;
> > > > > +
> > > > > +       base = domain->data->base;
> > > > > +       writel(F_INVLD_EN1 | F_INVLD_EN0, base + REG_MMU_INV_SEL);
> > > > > +       writel(F_ALL_INVLD, base + REG_MMU_INVALIDATE);
> > > > 
> > > > This needs to be synchronous, so you probably want to call
> > > > mtk_iommu_tlb_sync at the end.
> > > 
> > > From our spec, we have to wait until HW done after tlb flush range.
> > > But it don't need wait after tlb flush all.
> > > so It isn't necessary to add mtk_iommu_tlb_sync in tlb_flush_all here.
> > 
> > Okey doke, but I'm surprised you don't need a subsequent DSB or read-back.
> > What if the writel is buffered on the way to the IOMMU?
> 
> Then I change to this:
>  //==========
>     writel_relaxed(F_INVLD_EN1 | F_INVLD_EN0, base + REG_MMU_INV_SEL);
>     writel_relaxed(F_ALL_INVLD, base + REG_MMU_INVALIDATE);    
>     dsb(ishst);
> //===========
>     dsb or mb(). which one is better here?

I think you should use mb();

Will