Re: [PATCH] iommu/intel-iommu: set as DUMMY_DEVICE_DOMAIN_INFO if no IOMMU

[Daniel Drake]

On Sat, Feb 8, 2020 at 2:30 PM Lu Baolu <baolu.lu@xxxxxxxxxxxxxxx> wrote:
> > The devices under segment 1 are fake devices produced by
> > intel-nvme-remap mentioned here https://lkml.org/lkml/2020/2/5/139
>
> Has this series been accepted?

Sadly not - we didn't find any consensus on the right approach, and
further conversation is hindered by the questionable hardware design
and lack of further communication from Intel in explaining it. It's
one of the exceptional cases where we carry a significant non-upstream
kernel change, because unfortunately most of the current day consumer
PCs we work with have this BIOS option on by default and hence
unmodified Linux can't access the storage devices. On the offchance
that you have some energy to bubble this up inside Intel please let me
know and we will talk more... :)

That said, this thread was indeed only opened since we thought we had
found a more general issue that would potentially affect other cases.
The issue described does seem to highlight a possible imperfection in
code flow, although it may also be reasonable to say that (without
crazy downstream patches at play) if intel_iommu_add_device() fails
then we have bigger problems to face anyway...

> Will this help here? https://www.spinics.net/lists/iommu/msg41300.html

Yes! Very useful info and a real improvement. We'll follow the same
approach here. That does solve the problem we were facing, although we
needed one more fixup which I've sent separately for your
consideration: iommu/vt-d: consider real PCI device when checking if
mapping is needed

Thanks!
Daniel