Re: PCI pass-through problem for SN570 NVME SSD

[Jason Andryuk <jandryuk@xxxxxxxxx>]

On Mon, Jul 4, 2022 at 11:57 AM Roger Pau Monné <roger.pau@xxxxxxxxxx> wrote:
>
> On Mon, Jul 04, 2022 at 11:44:14PM +0800, G.R. wrote:
> > On Mon, Jul 4, 2022 at 11:33 PM Roger Pau Monné <roger.pau@xxxxxxxxxx> 
> > wrote:
> > >
> > > Right, so hvmloader attempts to place a BAR from 05:00.0 and a BAR
> > > from 00:17.0 into the same page, which is not that good behavior.  It
> > > might be sensible to attempt to share the page if both BARs belong to
> > > the same device, but not if they belong to different devices.
> > >
> > > I think the following patch:
> > >
> > > https://lore.kernel.org/xen-devel/20200117110811.43321-1-roger.pau@xxxxxxxxxx/
> > >
> > > Might help with this.
> > >
> > > Thanks, Roger.
> > I suppose this patch has been released in a newer XEN version that I
> > can pick up if I decide to upgrade?
> > Which version would it be?
> >
> > On the other hand, according to the other experiment I did, this may
> > not be the only issue related to this device.
> > Still not sure if the device or the SW stack is faulty this time...
>
> I don't think this patch has been applied to any release, adding Jason
> who I think was also interested in the fix and might provide more
> info.

Roger wrote the above patch after I tried to upstream a Qubes QEMU
patch 
https://lore.kernel.org/xen-devel/20190311180216.18811-7-jandryuk@xxxxxxxxx/.
The patch rounded up BAR sizes for passed through devices which
ensured they couldn't share a page.  But Roger rightfully pointed out
that changing the BAR size is incorrect, and hvmloader could just
enforce a minimum alignment.  However, nothing prevents a guest from
relocating the BARs again.

I tested Roger's patch, but Qubes and OpenXT have kept using the QEMU
patch.  When I added the QEMU patch to OpenXT, I wrote in the commit
message that it fixed probing an e1000e nic


Regards,
Jason