Re: [PATCH 11/12] swiotlb: merge swiotlb-xen initialization into swiotlb

[Stefano Stabellini <sstabellini@xxxxxxxxxx>]

On Fri, 4 Mar 2022, Christoph Hellwig wrote:
> On Thu, Mar 03, 2022 at 02:49:29PM -0800, Stefano Stabellini wrote:
> > On Thu, 3 Mar 2022, Christoph Hellwig wrote:
> > > On Wed, Mar 02, 2022 at 05:25:10PM -0800, Stefano Stabellini wrote:
> > > > Thinking more about it we actually need to drop the xen_initial_domain()
> > > > check otherwise some cases won't be functional (Dom0 not 1:1 mapped, or
> > > > DomU 1:1 mapped).
> > > 
> > > Hmm, but that would be the case even before this series, right?
> > 
> > Before this series we only have the xen_swiotlb_detect() check in
> > xen_mm_init, we don't have a second xen_initial_domain() check.
> > 
> > The issue is that this series is adding one more xen_initial_domain()
> > check in xen_mm_init.
> 
> In current mainline xen_mm_init calls xen_swiotlb_init unconditionally.
> But xen_swiotlb_init then calls xen_swiotlb_fixup after allocating
> the memory, which in turn calls xen_create_contiguous_region.
> xen_create_contiguous_region fails with -EINVAL for the
> !xen_initial_domain() and thus caues xen_swiotlb_fixup and
> xen_swiotlb_init to unwind and return -EINVAL.
> 
> So as far as I can tell there is no change in behavior, but maybe I'm
> missing something subtle?

You are right.

The xen_initial_domain() check in xen_create_contiguous_region() is
wrong and we should get rid of it. It is a leftover from before the
xen_swiotlb_detect rework.

We could either remove it or change it into another xen_swiotlb_detect()
check.

Feel free to add the patch to your series or fold it with another patch
or rework it as you prefer. Thanks for spotting this!

---

arm/xen: don't check for xen_initial_domain() in xen_create_contiguous_region

It used to be that Linux enabled swiotlb-xen when running a dom0 on ARM.
Since f5079a9a2a31 "xen/arm: introduce XENFEAT_direct_mapped and
XENFEAT_not_direct_mapped", Linux detects whether to enable or disable
swiotlb-xen based on the new feature flags: XENFEAT_direct_mapped and
XENFEAT_not_direct_mapped.

However, there is still a leftover xen_initial_domain() check in
xen_create_contiguous_region. Remove the check as
xen_create_contiguous_region is only called by swiotlb-xen during
initialization. If xen_create_contiguous_region is called, we know Linux
is running 1:1 mapped so there is no need for additional checks.

Also update the in-code comment.

Signed-off-by: Stefano Stabellini <stefano.stabellini@xxxxxxxxxx>


diff --git a/arch/arm/xen/mm.c b/arch/arm/xen/mm.c
index a7e54a087b80..28c207060253 100644
--- a/arch/arm/xen/mm.c
+++ b/arch/arm/xen/mm.c
@@ -122,10 +122,7 @@ int xen_create_contiguous_region(phys_addr_t pstart, 
unsigned int order,
                                 unsigned int address_bits,
                                 dma_addr_t *dma_handle)
 {
-       if (!xen_initial_domain())
-               return -EINVAL;
-
-       /* we assume that dom0 is mapped 1:1 for now */
+       /* the domain is 1:1 mapped to use swiotlb-xen */
        *dma_handle = pstart;
        return 0;
 }