Re: [PATCH v6 08/12] PCI: liveupdate: Inherit ACS flags in incoming preserved devices

[Pranjal Shrivastava <praan@google.com>]

On Tue, Jun 09, 2026 at 11:40:39AM -0700, David Matlack wrote:
> On Tue, Jun 9, 2026 at 10:20 AM Pranjal Shrivastava <praan@google.com> wrote:
> >
> > On Mon, Jun 08, 2026 at 09:56:41PM +0000, David Matlack wrote:
> > > On 2026-06-07 08:37 PM, Pranjal Shrivastava wrote:
> > > > On Fri, May 22, 2026 at 08:24:06PM +0000, David Matlack wrote:
> > > > > Inherit Access Control Services (ACS) flags on all incoming preserved
> > > > > devices (endpoints and upstream bridges) during a Live Update.
> > > > >
> > > > > Inheriting ACS flags avoids changing routing rules while memory
> > > > > transactions are in flight from preserved devices. This is also strictly
> > > > > necessary to ensure that IOMMU group assignments do not change across
> > > > > a Live Update for preserved devices, as changing ACS configurations can
> > > > > split or merge IOMMU groups.
> > > > >
> > > > > Cache the inherited ACS controls established by the previous kernel in
> > > > > struct pci_dev so that ACS controls do not change after a reset
> > > > > (pci_restore_state() calls pci_enable_acs()).
> > > > >
> > > > > To simplify ACS inheritance, reject preserving any devices that require
> > > > > quirks to enable ACS as those quirks would also have to take Live Update
> > > > > into account.
> > > > >
> > > > > Signed-off-by: David Matlack <dmatlack@google.com>
> > > > > ---
> > > > >  drivers/pci/liveupdate.c       | 68 ++++++++++++++++++++++++++++++++++
> > > > >  drivers/pci/liveupdate.h       | 11 ++++++
> > > > >  drivers/pci/pci.c              |  5 +++
> > > > >  drivers/pci/pci.h              |  5 +++
> > > > >  drivers/pci/quirks.c           |  7 ++++
> > > > >  include/linux/pci_liveupdate.h |  6 +++
> > > > >  6 files changed, 102 insertions(+)
> > > > >
> > > >
> > > > [...]
> > > >
> > > > >
> > > > > +void pci_liveupdate_init_acs(struct pci_dev *dev)
> > > > > +{
> > > > > + guard(rwsem_read)(&pci_liveupdate.rwsem);
> > > > > +
> > > > > + if (!dev->acs_cap || !dev->liveupdate.incoming)
> > > > > +         return;
> > > > > +
> > > > > + pci_read_config_word(dev, dev->acs_cap + PCI_ACS_CTRL, &dev->liveupdate.acs_ctrl);
> > > >
> > > > I might be thinking out loud here, but as an attacker, this motivates me
> > > > to somehow hack the EP FW to mis-report the PCI_ACS_CTRL register across
> > > > a liveupdate to fool the incoming kernel. If the FW feeds a 0, it silently
> > > > strips ACS protections.
> > > >
> > > > Should we also serialize ACS state in ser somehow to ensure we aren't
> > > > fooled by something like this?
> > >
> > > What does "EP FW" mean?
> >
> > I was referring to the Endpoint Firmware (basically any SW running on
> > a downstream device)
> >
> > >
> > > Does such an attacker even need Live Update to attack the system? It
> > > seems like such an attacker could route TLPs in whatever malicious way
> > > they want regardless of Live Update.
> > >
> >
> > I agree that compromised PCIe devices are a menace anyway. But I was
> > talking about the potential window opened up by Live Update here,
> > suppose we have Device A & B assigned to 2 different VMs (implying they
> > are in separate IOMMU groups because the switch set ACS_RR = 1).
> >
> > Now, the attacker has an opportunity with Liveupdate, since the devices
> > are already assigned, if *somehow* it flips a bit like ACS_RR, the
> > incoming kernel might see both the devices in the same IOMMU group.
> > Who detects this case and what happens if this happens if the devices
> > are kept assigned to these VMs?
> 
> I suspect that would be caught during the restore of the iommufds to
> which those devices are attached.
> 
> The kernel would attempt to restore each device into a separate domain
> (since that's how they were preserved before the Live Update) but that
> will fail because they are in the same group now. Even if one of the
> devices was not preserved, that will still cause a failure when a user
> tries to start using that device (e.g. to try to attach it to a
> different VM).

Yes, IOMMU would eventually catch-up but what about the DMAs that were
done already? Say to an NVMe disk? We'll have to wipe the entire disk in
such a case? Also, we wouldn't know the offending device..

If such situations aren't a problem, then I guess it's fine.

Thanks,
Praan