Re: [Qemu-devel] [PATCH 0/4] exec: address space translation cleanups

[Peter Xu]

On Tue, May 16, 2017 at 06:51:03PM +0200, Maxime Coquelin wrote:
> Hi Peter,
> 
> On 05/16/2017 03:24 PM, Maxime Coquelin wrote:
> >
> >
> >On 05/15/2017 10:50 AM, Peter Xu wrote:
> >>The problem is that, address_space_get_iotlb_entry() shares a lot with
> >>address_space_translate(). This patch tries to abstract the
> >>shared elements.
> >>
> >>Originally, this work is derived from discussion from VT-d passthrough
> >>series discussions [1]. But for sure we can just see this series as a
> >>standalone cleanup. So I posted it separately here.
> >>
> >>Smoke tests are done with general VM boots, IOs, especially with vhost
> >>dmar configurations.
> >>
> >>I believe with current series I can throw away the old patch [1],
> >>which may be good. But before that, please kindly review. Thanks.
> >
> >I faced the problem the old patch fixes when declaring and attaching an
> >IOMMU device, but booting the kernel with intel_iommu=off.
> >
> >I tested again with patches 1 & 4 of your series, and I confirm it fixes
> >the issue:
> >Tested-by: Maxime Coquelin <address@hidden>
> 
> I did some more testing with my "vhost-user IOMMU" setup, and the series
> actually breaks with IOMMU device attached, and intel_iommu=on.
> 
> The main difference with the previous passing test is the guest RAM
> size. In the working setup, it is 2G of 2M hugepages, vs. 4G of 2M
> hugepages in the failing one. Note that I also reproduce with vhost-kernel
> backend.
> 
> The error happens in the first vhost_device_iotlb_miss() call:
> qemu-system-x86_64: Fail to lookup the translated address b5d7c000
> 
> I don't have the root cause yet, I'll keep you updated.

Maxime,

Thanks a lot for help testing this series!

I reproduced this problem, and this is not a problem obvious enough
for me. Let me investigate as well.

-- 
Peter Xu