On 2022-09-02 18:21, Jason Gunthorpe wrote:
On Fri, Sep 02, 2022 at 01:11:09PM -0400, Matthew Rosato wrote:
On 9/1/22 4:37 PM, Jason Gunthorpe wrote:
On Thu, Sep 01, 2022 at 12:14:24PM -0400, Matthew Rosato wrote:
On 9/1/22 6:25 AM, Robin Murphy wrote:
On 2022-08-31 21:12, Matthew Rosato wrote:
With commit fa7e9ecc5e1c ("iommu/s390: Tolerate repeat attach_dev
calls") s390-iommu is supposed to handle dynamic switching between IOMMU
domains and the DMA API handling. However, this commit does not
sufficiently handle the case where the device is released via a call
to the release_device op as it may occur at the same time as an opposing
attach_dev or detach_dev since the group mutex is not held over
release_device. This was observed if the device is deconfigured during a
small window during vfio-pci initialization and can result in WARNs and
potential kernel panics.
Hmm, the more I think about it, something doesn't sit right about this whole situation... release_device is called via the notifier from device_del() after the device has been removed from its parent bus and largely dismantled; it should definitely not still have a driver bound by that point, so how is VFIO doing things that manage to race at all?
Robin.
So, I generally have seen the issue manifest as one of the calls
into the iommu core from __vfio_group_unset_container
(e.g. iommu_deatch_group via vfio_type1_iommu) failing with a WARN.
This happens when the vfio group fd is released, which could be
coming e.g. from a userspace ioctl VFIO_GROUP_UNSET_CONTAINER.
AFAICT there's nothing serializing the notion of calling into the
iommu core here against a device that is simultaneously going
through release_device (because we don't enter release_device with
the group mutex held), resulting in unpredictable behavior between
the dueling attach_dev/detach_dev and the release_device for
s390-iommu at least.
Oh, this is a vfio bug.
I've been running with your diff applied today on s390 and this
indeed fixes the issue by preventing the detach-after-release coming
out of vfio.
Heh, I'm shocked it worked at all
I've been trying to understand Robin's latest remarks because maybe I
don't really understand your situation right.
That was really just me thinking out loud to guess at how it must be
happening - I wasn't sure whether VFIO is actually intended to allow
that or not, so if not then by all means let's look at fixing that, but
as I say I think we're only seeing it provoke a problem at the driver
level because of 9ac8545199a1, and fixing VFIO doesn't fix that in
general. And conversely if we *can* fix that properly at the IOMMU API
level then the current VFIO behaviour should become benign again anyway.
IMHO this is definately a VFIO bug, because in a single-device group
we must not allow the domain to remain attached past remove(). Or more
broadly we shouldn't be holding ownership of a group without also
having a driver attached.
FWIW I was assuming it might be fine for a VFIO user to hold the group
open if they expect the device to come back again and re-bind (for
example, perhaps over some reconfiguration that requires turning SR-IOV
off and on again?)
Cheers,
Robin.
