Re: [PATCH] hw/virtio/vhost: Disable IOTLB callbacks when IOMMU gets disabled

[Peter Xu]

On Mon, Jan 27, 2025 at 08:44:50AM +0800, Jason Wang wrote:
> On Sun, Jan 26, 2025 at 3:56 PM Jason Wang <jasowang@redhat.com> wrote:
> >
> > On Sat, Jan 25, 2025 at 12:42 AM Peter Xu <peterx@redhat.com> wrote:
> > >
> > > Hello, Jason, Eric,
> > >
> > > On Fri, Jan 24, 2025 at 11:30:56AM +0800, Jason Wang wrote:
> > > > It might be because neither virtio bus nor virtio-net provides a
> > > > shutdown method.
> > > >
> > > > There used to be requests to provide those to unbreak the kexec.
> > > >
> > > > A quick try might be to provide a .driver.shutdown to
> > > > virtio_net_driver structure and reset the device there as a start.
> > >
> > > I didn't check virtio driver path, but if that's missing it's reasonable 
> > > to
> > > support it indeed.
> > >
> > > OTOH, even with that, vhost can still hit such DMA issue if it's a
> > > hard-reset, am I right?  IOW, when using QMP command "system-reset".  If 
> > > my
> > > memory is correct, that's the problem I was working on the VFIO series,
> > > rather than a clean reboot.  And that won't give guest driver chance to 
> > > run
> > > anything, IIUC.
> >
> > Yes.
> >
> > >
> > > I am wildly suspecting a VT-d write to GCMD to disable it can also appear
> > > if there's a hard reset, then when bootloading the VM the bios (or 
> > > whatever
> > > firmware at early stage) may want to make sure the VT-d device is
> > > completely off by writting to GCMD. But that's a pure guess.. and that may
> > > or may not matter much on how we fix this problem.
> > >
> > > IOW, I suspect we need to fix both of them,
> > >
> > >   (a) for soft-reset, by making sure drivers properly quiesce DMAs
> > >   proactively when VM gracefully shuts down.
> > >
> > >   (b) for hard-reset, by making sure QEMU reset in proper order.
> > >
> > > One thing to mention is for problem (b) VFIO used to have an extra
> > > challenge on !FLR devices, I discussed it in patch 4's comment there.
> > > Quotting from patch 4 of series:
> > >
> > > https://lore.kernel.org/all/20240117091559.144730-1-peterx@redhat.com
> > >
> > >      * (1) Device depth-first reset hierachy doesn't yet work for vIOMMUs
> > >      *     (reference: resettable_cold_reset_fn())
> > >      *
> > >      *     Currently, vIOMMU devices are created as normal '-device'
> > >      *     cmdlines.  It means in many ways it has the same attributes 
> > > with
> > >      *     most of the rest devices, even if the rest devices should
> > >      *     logically be under control of the vIOMMU unit.
> > >      *
> > >      *     One side effect of it is vIOMMU devices will be currently put
> > >      *     randomly under qdev tree hierarchy, which is the source of
> > >      *     device reset ordering in current QEMU (depth-first traversal).
> > >      *     It means vIOMMU now can be reset before some devices.  For 
> > > fully
> > >      *     emulated devices that's not a problem, because the traversal
> > >      *     holds BQL for the whole process.  However it is a problem if 
> > > DMA
> > >      *     can happen without BQL, like VFIO, vDPA or remote device 
> > > process.
> > >      *
> > >      *     TODO: one ideal solution can be that we make vIOMMU the parent
> > >      *     of the whole pci host bridge.  Hence vIOMMU can be reset after
> > >      *     all the devices are reset and quiesced.
> > >      *
> > >      * (2) Some devices register its own reset functions
> > >      *
> > >      *     Even if above issue solved, if devices register its own reset
> > >      *     functions for some reason via QEMU reset hooks, vIOMMU can 
> > > still
> > >      *     be reset before the device. One example is vfio_reset_handler()
> > >      *     where FLR is not supported on the device.
> > >      *
> > >      *     TODO: merge relevant reset functions into the device tree reset
> > >      *     framework.
> > >
> > > So maybe vhost doesn't have problem (2) listed above, and maybe it means
> > > it's still worthwhile thinking more about problem (1), which is to change
> > > the QOM tree to provide a correct topology representation when vIOMMU is
> > > present: so far it should be still a pretty much orphaned object there.. 
> > > if
> > > QEMU relies on QOM tree topology for reset order, we may need to move it 
> > > to
> > > the right place sooner or later.
> >
> > Sounds like a non-trivial task, so for a hard reset, maybe we can
> > proceed with Eric's proposal to deal with the reset before the device
> > stops.

The major challenge when I was working on that (as far as I can still
remember..): some devices are created at early stage of QEMU startup, which
can happen before the vIOMMU is created and realized.  Then it can be
challenging to re-parent those devices to be childrens of the vIOMMU, or we
may need a way to create vIOMMU always earlier than those..

> 
> Btw, I actually meant to break the assumption that vhost needs to be
> enabled/disabled after/before vIOMMU. This only works for virtio-net /
> vhost. From the view of vhost, it would work similar to _F_LOG_ALL
> (where there's no assumption on the order of enabling/disabling dirty
> page tracking and device start/stop).

Yes, we can go for a lightweight solution.

Thanks,

-- 
Peter Xu