[PATCH RFC 00/10] hw/vfio, x86/iommu: IOMMUFD Dirty Tracking

Joao Martins posted 10 patches 1 year, 12 months ago
accel/kvm/kvm-all.c            |  12 +++
hmp-commands.hx                |   5 +-
hw/i386/amd_iommu.c            |  76 +++++++++++++-
hw/i386/amd_iommu.h            |  11 +-
hw/i386/intel_iommu.c          | 103 +++++++++++++++++--
hw/i386/intel_iommu_internal.h |   4 +
hw/i386/trace-events           |   4 +
hw/iommufd/iommufd.c           |  63 ++++++++++++
hw/iommufd/trace-events        |   3 +
hw/vfio/container.c            |  20 +++-
hw/vfio/iommufd.c              | 179 ++++++++++++++++++++++++++++++++-
hw/vfio/trace-events           |   3 +-
include/exec/memory.h          |  10 +-
include/hw/i386/intel_iommu.h  |   1 +
include/hw/iommufd/iommufd.h   |   6 ++
linux-headers/linux/iommufd.h  |  78 ++++++++++++++
migration/dirtyrate.c          |  59 ++++++++---
migration/dirtyrate.h          |   1 +
qapi/migration.json            |  15 +++
softmmu/memory.c               |   5 +
20 files changed, 618 insertions(+), 40 deletions(-)
[PATCH RFC 00/10] hw/vfio, x86/iommu: IOMMUFD Dirty Tracking
Posted by Joao Martins 1 year, 12 months ago
This series expands IOMMUFD series from Yi and Eric into
supporting IOMMU Dirty Tracking. It adds both the emulated x86
IOMMUs, as well as IOMMUFD support that exercises said
emulation (or H/W).

It is organized into:

* Patches 1 - 4: x86 IOMMU emulation that performs dirty tracking,
useful for a nested guest exercising the IOMMUFD kernel counterpart
for coverage and development. The caching of the PASID PTE root flags
and AMD DTE is not exactly the cleanest, still wondering about a
better way without having to lookup context/DTE prior to IOTLB lookup.

* Patches 5 - 9: IOMMUFD backend support for dirty tracking;
Sadly this wasn't exactly tested with VF Live Migration, but it adds
(last 2 patches) a way to measure dirty rate via HMP and QMP.

The IOMMUFD kernel patches go into extent on what each of the ioctls
do, albeit the workflow is relatively simple:

 1) Toggling dirty tracking via HWPT_SET_DIRTY or get -ENOTSUPP otherwise
    which voids any attempt of walking IO pagetables

 2) Read-and-clear of Dirty IOVAs

 3) Unmap vIOMMU-backed IOVAs and fetch its dirty state right-before-unmap.

The series is still a WIP. The intend is to exercise the kernel side[0]
aside from the selftests that provide some coverage.

Comments, feedback, as always appreciated.

	Joao

[0] https://lore.kernel.org/linux-iommu/20220428210933.3583-1-joao.m.martins@oracle.com/

Joao Martins (10):
  amd-iommu: Cache PTE/DTE info in IOTLB
  amd-iommu: Access/Dirty bit support
  intel-iommu: Cache PASID entry flags
  intel_iommu: Second Stage Access Dirty bit support
  linux-headers: import iommufd.h hwpt extensions
  vfio/iommufd: Add HWPT_SET_DIRTY support
  vfio/iommufd: Add HWPT_GET_DIRTY_IOVA support
  vfio/iommufd: Add IOAS_UNMAP_DIRTY support
  migration/dirtyrate: Expand dirty_bitmap to be tracked separately for
    devices
  hw/vfio: Add nr of dirty pages to tracepoints

 accel/kvm/kvm-all.c            |  12 +++
 hmp-commands.hx                |   5 +-
 hw/i386/amd_iommu.c            |  76 +++++++++++++-
 hw/i386/amd_iommu.h            |  11 +-
 hw/i386/intel_iommu.c          | 103 +++++++++++++++++--
 hw/i386/intel_iommu_internal.h |   4 +
 hw/i386/trace-events           |   4 +
 hw/iommufd/iommufd.c           |  63 ++++++++++++
 hw/iommufd/trace-events        |   3 +
 hw/vfio/container.c            |  20 +++-
 hw/vfio/iommufd.c              | 179 ++++++++++++++++++++++++++++++++-
 hw/vfio/trace-events           |   3 +-
 include/exec/memory.h          |  10 +-
 include/hw/i386/intel_iommu.h  |   1 +
 include/hw/iommufd/iommufd.h   |   6 ++
 linux-headers/linux/iommufd.h  |  78 ++++++++++++++
 migration/dirtyrate.c          |  59 ++++++++---
 migration/dirtyrate.h          |   1 +
 qapi/migration.json            |  15 +++
 softmmu/memory.c               |   5 +
 20 files changed, 618 insertions(+), 40 deletions(-)

-- 
2.17.2