| Message ID | 20240716094619.1713905-1-eric.auger@redhat.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | VIRTIO-IOMMU/VFIO: Revert IOMMUDevice clear and fix hotunplug | expand |
On 7/16/24 11:45, Eric Auger wrote: > 1b889d6e39c3 ("virtio-iommu: Clear IOMMUDevice when VFIO > device is unplugged" fixes the VFIO hotplug/hotunplug/hotplug > sequence by clearing the IOMMUDevice which backs the VFIO device. > However this brings other troubles such as a memory corruption. > > Even when fixing some cleanups that were missed on the first > attempt the memory corruption still exists if the IOMMUDevice is > freed. Until we understand the exact cause let's make things simpler: > let the backing IOMMUDevice survive the unplug as what is done > on intel iommu for instance. Clean up/reset resources that would > prevent the device from being hotplugged again (probe_done is removed, > [host_]resv_regions are reset). By doing this we also rule out the > use case of aliased BDFs which is known to be not functional with > virtio-iommu (missing guest kernel support) and the virtio-iommu device > implementation is not ready either. > > This series can be found at: > https://github.com/eauger/qemu/tree/virtio_iommu_device_clear_fixes_v1 > > > Eric Auger (6): > Revert "virtio-iommu: Clear IOMMUDevice when VFIO device is unplugged" > virtio-iommu: Remove probe_done > virtio-iommu: Free [host_]resv_ranges on unset_iommu_devices > virtio-iommu: Remove the end point on detach > hw/vfio/common: Add vfio_listener_region_del_iommu trace event > virtio-iommu: Add trace point on > virtio_iommu_detach_endpoint_from_domain > > include/hw/virtio/virtio-iommu.h | 1 - > hw/vfio/common.c | 3 +- > hw/virtio/virtio-iommu.c | 88 +++++++++++--------------------- > hw/vfio/trace-events | 3 +- > hw/virtio/trace-events | 1 + > 5 files changed, 35 insertions(+), 61 deletions(-) > Tested-by: Cédric Le Goater <clg@redhat.com> with a vfio-pci device (mlx5 VF) and a virtio-net-pci device. Thanks, C.
Hi Cédric, On 7/16/24 16:02, Cédric Le Goater wrote: > On 7/16/24 11:45, Eric Auger wrote: >> 1b889d6e39c3 ("virtio-iommu: Clear IOMMUDevice when VFIO >> device is unplugged" fixes the VFIO hotplug/hotunplug/hotplug >> sequence by clearing the IOMMUDevice which backs the VFIO device. >> However this brings other troubles such as a memory corruption. >> >> Even when fixing some cleanups that were missed on the first >> attempt the memory corruption still exists if the IOMMUDevice is >> freed. Until we understand the exact cause let's make things simpler: >> let the backing IOMMUDevice survive the unplug as what is done >> on intel iommu for instance. Clean up/reset resources that would >> prevent the device from being hotplugged again (probe_done is removed, >> [host_]resv_regions are reset). By doing this we also rule out the >> use case of aliased BDFs which is known to be not functional with >> virtio-iommu (missing guest kernel support) and the virtio-iommu device >> implementation is not ready either. >> >> This series can be found at: >> https://github.com/eauger/qemu/tree/virtio_iommu_device_clear_fixes_v1 >> >> >> Eric Auger (6): >> Revert "virtio-iommu: Clear IOMMUDevice when VFIO device is >> unplugged" >> virtio-iommu: Remove probe_done >> virtio-iommu: Free [host_]resv_ranges on unset_iommu_devices >> virtio-iommu: Remove the end point on detach >> hw/vfio/common: Add vfio_listener_region_del_iommu trace event >> virtio-iommu: Add trace point on >> virtio_iommu_detach_endpoint_from_domain >> >> include/hw/virtio/virtio-iommu.h | 1 - >> hw/vfio/common.c | 3 +- >> hw/virtio/virtio-iommu.c | 88 +++++++++++--------------------- >> hw/vfio/trace-events | 3 +- >> hw/virtio/trace-events | 1 + >> 5 files changed, 35 insertions(+), 61 deletions(-) >> > > > Tested-by: Cédric Le Goater <clg@redhat.com> > > with a vfio-pci device (mlx5 VF) and a virtio-net-pci device. Many Thanks! Eric > > Thanks, > > C. >
1b889d6e39c3 ("virtio-iommu: Clear IOMMUDevice when VFIO device is unplugged" fixes the VFIO hotplug/hotunplug/hotplug sequence by clearing the IOMMUDevice which backs the VFIO device. However this brings other troubles such as a memory corruption. Even when fixing some cleanups that were missed on the first attempt the memory corruption still exists if the IOMMUDevice is freed. Until we understand the exact cause let's make things simpler: let the backing IOMMUDevice survive the unplug as what is done on intel iommu for instance. Clean up/reset resources that would prevent the device from being hotplugged again (probe_done is removed, [host_]resv_regions are reset). By doing this we also rule out the use case of aliased BDFs which is known to be not functional with virtio-iommu (missing guest kernel support) and the virtio-iommu device implementation is not ready either. This series can be found at: https://github.com/eauger/qemu/tree/virtio_iommu_device_clear_fixes_v1 Eric Auger (6): Revert "virtio-iommu: Clear IOMMUDevice when VFIO device is unplugged" virtio-iommu: Remove probe_done virtio-iommu: Free [host_]resv_ranges on unset_iommu_devices virtio-iommu: Remove the end point on detach hw/vfio/common: Add vfio_listener_region_del_iommu trace event virtio-iommu: Add trace point on virtio_iommu_detach_endpoint_from_domain include/hw/virtio/virtio-iommu.h | 1 - hw/vfio/common.c | 3 +- hw/virtio/virtio-iommu.c | 88 +++++++++++--------------------- hw/vfio/trace-events | 3 +- hw/virtio/trace-events | 1 + 5 files changed, 35 insertions(+), 61 deletions(-)