[PATCH 0/3] vfio: selftests: Automatically detect the IOMMU mode

[David Matlack <dmatlack@google.com>]

This series makes the VFIO selftests library automatically detect which
IOMMU mode to use, rather than hardcoding iommufd as the default, and
makes tests skip, rather than fail, when the IOMMU mode they need is not
available. This was suggested by Sean while reviewing the KVM IRQ test,
which currently has to probe for the IOMMU mode itself [1].

iommufd is an optional feature, and VFIO's iommufd support additionally
requires CONFIG_VFIO_DEVICE_CDEV. On otherwise functional kernels that
only support the legacy VFIO container, the VFIO selftests (and the KVM
IRQ test, when using a VFIO device) currently fail with an unhelpful
assert:

  tools/testing/selftests/vfio/lib/iommu.c:448: Assertion Failure

  Expression: iommu->iommufd > 0
  Observed: 0xffffffffffffffff > 0
  [errno: 2 - No such file or directory]

Patch 1 makes iommu_init(NULL) pick the most preferred IOMMU mode that is
available on the current system, and drops the global default_iommu_mode.
iommu_modes[] is now ordered by preference: iommufd, vfio_type1v2_iommu,
vfio_type1_iommu, iommufd_compat_type1v2, iommufd_compat_type1. A mode is
considered available if its container (or /dev/iommu) can be opened and,
for container modes, if VFIO_CHECK_EXTENSION reports support for the
mode's IOMMU type. iommufd mode additionally requires VFIO device cdevs.

Patch 2 adds iommu_mode_is_available() and
SKIP_IF_IOMMU_MODE_UNAVAILABLE(), and uses them to skip tests when the
IOMMU mode they need is not available. Tests skip via the kselftest
harness's SKIP() rather than iommu_init() calling ksft_exit_skip(), since
the latter produces malformed TAP output when called from within a
harness test.

Patch 3 converts the KVM IRQ test to use the VFIO selftests library to
pick the IOMMU mode, and drops its probe_iommu_type().

This series is based on vfio/next (vfio-v7.4-rc1).

[1] https://lore.kernel.org/kvm/aiIG2L_KdPwuxYA1@google.com/

Testing
-----------------------------------------------------------------------

Tested in a QEMU VM with an emulated igb device behind an emulated Intel
IOMMU, bound to vfio-pci, on a kernel with CONFIG_IOMMUFD=y,
CONFIG_VFIO_DEVICE_CDEV=y, and CONFIG_VFIO_CONTAINER=y. To simulate
kernels with fewer IOMMU modes available, /dev/iommu and /dev/vfio/vfio
were deleted in the VM:

 - All IOMMU modes available: All VFIO selftests pass. irq_test -d passes
   using iommufd.

 - Only the legacy VFIO container available (/dev/iommu deleted): All
   VFIO selftests pass, with the iommufd and iommufd_compat variants
   skipped instead of failing. irq_test -d passes using
   vfio_type1v2_iommu, and skips if -t iommufd is passed.

 - No IOMMU mode available (/dev/iommu and /dev/vfio/vfio deleted):
   vfio_pci_device_test, vfio_pci_driver_test, and irq_test -d all skip
   with a message explaining why, and with well-formed TAP output.

David Matlack (3):
  vfio: selftests: Automatically detect the IOMMU mode in iommu_init()
  vfio: selftests: Skip tests when the IOMMU mode is not available
  KVM: selftests: Use VFIO selftests library to pick IOMMU mode in IRQ
    test

 tools/testing/selftests/kvm/irq_test.c        | 19 +----
 .../vfio/lib/include/libvfio/iommu.h          | 30 +++++++-
 tools/testing/selftests/vfio/lib/iommu.c      | 75 ++++++++++++++++---
 .../vfio/vfio_dma_mapping_mmio_test.c         |  2 +
 .../selftests/vfio/vfio_dma_mapping_test.c    |  4 +
 .../vfio/vfio_pci_device_init_perf_test.c     |  2 +
 .../selftests/vfio/vfio_pci_device_test.c     |  8 +-
 .../selftests/vfio/vfio_pci_driver_test.c     |  9 ++-
 .../selftests/vfio/vfio_pci_sriov_uapi_test.c |  4 +
 9 files changed, 118 insertions(+), 35 deletions(-)


base-commit: 076d7e989cb69f6ded7d60f2a080eb909930061b
-- 
2.56.0.317.g454f8302e1-goog