Re: [PATCH v3 3/6] iommu: add ARM short descriptor page table allocator.

[Will Deacon <will.deacon-5wv7dgnIgG8@public.gmane.org>]

Hello Yong,

On Mon, Sep 14, 2015 at 01:25:00PM +0100, Yong Wu wrote:
> On Tue, 2015-07-21 at 18:11 +0100, Will Deacon wrote:
> > > +       ret = _arm_short_map(data, iova, paddr, pgdprot, pteprot, large);
> > > +
> > > +       tlb->tlb_add_flush(iova, size, true, data->iop.cookie);
> > > +       tlb->tlb_sync(data->iop.cookie);
> > 
> > In _arm_short_map, it looks like you can only go from invalid -> valid,
> > so why do you need to flush the TLB here?
> 
> Hi Will,
>    Here is about flush-tlb after map iova, I have deleted it in v4
> following this suggestion. But We meet a problem about it.

Ok.

> Take a example with JPEG. the test steps is:
> a).JPEG HW decode a picture with the source iova,like 0xfd780000.
> b).JPEG HW decode done, It will unmap the iova(write 0 in pagetable and
> flush tlb).
> c).JPEG HW decode the second picture, whose source iova is also
> 0xfd780000.
>    Then our HW maybe fail due to it will auto prefetch, It may prefecth
> between the step b) and c). then the HW may fetch the pagetable content
> which has been unmapped in step b). then the HW will get the iova's
> physical address is 0, It will translation fault!

Oh no! So-called "negative caching" is certainly prohibited by the ARM
architecture, but if you've built it then we can probably work around it
as an additional quirk. I assume the prefetcher stops prefetching when
it sees an invalid descriptor?

>     So I think our HW need flush-tlb after map iova. Could we add a
> QUIRK like "IO_PGTABLE_QUIRK_AUTO_PREFETCH_ENABLE" for it?
> If it's not allowed, we will have to add this in our internal function
> mtk_iommu_map of mtk_iommu.c.

Actually, this type of quirk is ringing bells with me (I think another
IOMMU needed something similar in the past), so maybe just add
IO_PGTABLE_QUIRK_TLBI_ON_MAP?

Will