mbox series

[v2,0/7] VIRTIO-IOMMU/HostIOMMUDevice: Fixes and page size mask rework

Message ID 20240701084957.1567641-1-eric.auger@redhat.com (mailing list archive)
Headers show
Series VIRTIO-IOMMU/HostIOMMUDevice: Fixes and page size mask rework | expand

Message

Eric Auger July 1, 2024, 8:48 a.m. UTC
The 2 first patches are fixes of
cf2647a76e ("virtio-iommu: Compute host reserved regions")
They can be taken separately of the rest.

Then the series uses the HostIOMMUDevice interface to fetch
information about the page size mask supported along the assigned
device and propagate it to the virtio-iommu. This is a similar
work as what was done in

VIRTIO-IOMMU/VFIO: Fix host iommu geometry handling series

but this time for the page size mask. Using this new
infrastructure allows to handle page size mask conflicts
earlier on device hotplug as opposed to current QEMU
abort:

qemu-system-aarch64: virtio-iommu virtio-iommu-memory-region-8-0
does not support frozen granule 0x10000
qemu: hardware error: vfio: DMA mapping failed, unable to continue

With this series the hotplug nicely fails.

Also this allows to remove hacks consisting in transiently enabling
IOMMU MRs to collect coldplugged device page size mask before machine
init. Those hacks were introduced by

94df5b2180d6 ("virtio-iommu: Fix 64kB host page size VFIO device
assignment")

The series can be found at:
https://github.com/eauger/qemu/tree/virtio-iommu-psmask-rework-v2

---
v1 -> v2:
- Don't update mask if the granule is frozen (Zhenzhong)
- Collected Zhenzhong's and Cédric's R-bs

Eric Auger (7):
  virtio-iommu: Fix error handling in
    virtio_iommu_set_host_iova_ranges()
  vfio-container-base: Introduce vfio_container_get_iova_ranges() helper
  HostIOMMUDevice : remove Error handle from get_iova_ranges callback
  HostIOMMUDevice: Introduce get_page_size_mask() callback
  virtio-iommu : Retrieve page size mask on
    virtio_iommu_set_iommu_device()
  memory: remove IOMMU MR iommu_set_page_size_mask() callback
  virtio-iommu: Revert transient enablement of IOMMU MR in bypass mode

 include/exec/memory.h                 |  38 --------
 include/hw/vfio/vfio-container-base.h |   9 ++
 include/sysemu/host_iommu_device.h    |  11 ++-
 hw/vfio/common.c                      |   8 --
 hw/vfio/container-base.c              |  15 ++++
 hw/vfio/container.c                   |  16 ++--
 hw/vfio/iommufd.c                     |  21 +++--
 hw/virtio/virtio-iommu.c              | 123 +++++++++++++-------------
 system/memory.c                       |  13 ---
 hw/virtio/trace-events                |   2 +-
 10 files changed, 119 insertions(+), 137 deletions(-)

Comments

Michael S. Tsirkin July 1, 2024, 9:14 p.m. UTC | #1
On Mon, Jul 01, 2024 at 10:48:52AM +0200, Eric Auger wrote:
> The 2 first patches are fixes of
> cf2647a76e ("virtio-iommu: Compute host reserved regions")
> They can be taken separately of the rest.
> 
> Then the series uses the HostIOMMUDevice interface to fetch
> information about the page size mask supported along the assigned
> device and propagate it to the virtio-iommu. This is a similar
> work as what was done in
> 
> VIRTIO-IOMMU/VFIO: Fix host iommu geometry handling series
> 
> but this time for the page size mask. Using this new
> infrastructure allows to handle page size mask conflicts
> earlier on device hotplug as opposed to current QEMU
> abort:
> 
> qemu-system-aarch64: virtio-iommu virtio-iommu-memory-region-8-0
> does not support frozen granule 0x10000
> qemu: hardware error: vfio: DMA mapping failed, unable to continue
> 
> With this series the hotplug nicely fails.
> 
> Also this allows to remove hacks consisting in transiently enabling
> IOMMU MRs to collect coldplugged device page size mask before machine
> init. Those hacks were introduced by
> 
> 94df5b2180d6 ("virtio-iommu: Fix 64kB host page size VFIO device
> assignment")
> 
> The series can be found at:
> https://github.com/eauger/qemu/tree/virtio-iommu-psmask-rework-v2


So presumably will be merged with rest of host iommu stuff.
For virtio bits:

Reviewed-by: Michael S. Tsirkin <mst@redhat.com>



> ---
> v1 -> v2:
> - Don't update mask if the granule is frozen (Zhenzhong)
> - Collected Zhenzhong's and Cédric's R-bs
> 
> Eric Auger (7):
>   virtio-iommu: Fix error handling in
>     virtio_iommu_set_host_iova_ranges()
>   vfio-container-base: Introduce vfio_container_get_iova_ranges() helper
>   HostIOMMUDevice : remove Error handle from get_iova_ranges callback
>   HostIOMMUDevice: Introduce get_page_size_mask() callback
>   virtio-iommu : Retrieve page size mask on
>     virtio_iommu_set_iommu_device()
>   memory: remove IOMMU MR iommu_set_page_size_mask() callback
>   virtio-iommu: Revert transient enablement of IOMMU MR in bypass mode
> 
>  include/exec/memory.h                 |  38 --------
>  include/hw/vfio/vfio-container-base.h |   9 ++
>  include/sysemu/host_iommu_device.h    |  11 ++-
>  hw/vfio/common.c                      |   8 --
>  hw/vfio/container-base.c              |  15 ++++
>  hw/vfio/container.c                   |  16 ++--
>  hw/vfio/iommufd.c                     |  21 +++--
>  hw/virtio/virtio-iommu.c              | 123 +++++++++++++-------------
>  system/memory.c                       |  13 ---
>  hw/virtio/trace-events                |   2 +-
>  10 files changed, 119 insertions(+), 137 deletions(-)
> 
> -- 
> 2.41.0
Cédric Le Goater July 1, 2024, 9:14 p.m. UTC | #2
On 7/1/24 10:48 AM, Eric Auger wrote:
> The 2 first patches are fixes of
> cf2647a76e ("virtio-iommu: Compute host reserved regions")
> They can be taken separately of the rest.
> 
> Then the series uses the HostIOMMUDevice interface to fetch
> information about the page size mask supported along the assigned
> device and propagate it to the virtio-iommu. This is a similar
> work as what was done in
> 
> VIRTIO-IOMMU/VFIO: Fix host iommu geometry handling series
> 
> but this time for the page size mask. Using this new
> infrastructure allows to handle page size mask conflicts
> earlier on device hotplug as opposed to current QEMU
> abort:
> 
> qemu-system-aarch64: virtio-iommu virtio-iommu-memory-region-8-0
> does not support frozen granule 0x10000
> qemu: hardware error: vfio: DMA mapping failed, unable to continue
> 
> With this series the hotplug nicely fails.
> 
> Also this allows to remove hacks consisting in transiently enabling
> IOMMU MRs to collect coldplugged device page size mask before machine
> init. Those hacks were introduced by
> 
> 94df5b2180d6 ("virtio-iommu: Fix 64kB host page size VFIO device
> assignment")
> 
> The series can be found at:
> https://github.com/eauger/qemu/tree/virtio-iommu-psmask-rework-v2
> 
> ---
> v1 -> v2:
> - Don't update mask if the granule is frozen (Zhenzhong)
> - Collected Zhenzhong's and Cédric's R-bs
> 
> Eric Auger (7):
>    virtio-iommu: Fix error handling in
>      virtio_iommu_set_host_iova_ranges()
>    vfio-container-base: Introduce vfio_container_get_iova_ranges() helper
>    HostIOMMUDevice : remove Error handle from get_iova_ranges callback
>    HostIOMMUDevice: Introduce get_page_size_mask() callback
>    virtio-iommu : Retrieve page size mask on
>      virtio_iommu_set_iommu_device()
>    memory: remove IOMMU MR iommu_set_page_size_mask() callback
>    virtio-iommu: Revert transient enablement of IOMMU MR in bypass mode
> 
>   include/exec/memory.h                 |  38 --------
>   include/hw/vfio/vfio-container-base.h |   9 ++
>   include/sysemu/host_iommu_device.h    |  11 ++-
>   hw/vfio/common.c                      |   8 --
>   hw/vfio/container-base.c              |  15 ++++
>   hw/vfio/container.c                   |  16 ++--
>   hw/vfio/iommufd.c                     |  21 +++--
>   hw/virtio/virtio-iommu.c              | 123 +++++++++++++-------------
>   system/memory.c                       |  13 ---
>   hw/virtio/trace-events                |   2 +-
>   10 files changed, 119 insertions(+), 137 deletions(-)
> 


Applied to vfio-next.

Thanks,

C.