mbox series

[0/2] Two enhancements to iommu_at[de]tach_device_pasid()

Message ID 20240327125433.248946-1-yi.l.liu@intel.com (mailing list archive)
Headers show
Series Two enhancements to iommu_at[de]tach_device_pasid() | expand

Message

Yi Liu March 27, 2024, 12:54 p.m. UTC 
There are minor mistakes in the iommu set_dev_pasid() and remove_dev_pasid()
paths. The set_dev_pasid() path updates the group->pasid_array first, and
then may call into remove_dev_pasid() in error handling when there are devices
within the group failed to set_dev_pasid. The remove_dev_pasid() callbacks of
the underlying iommu drivers get domain for pasid from the group->pasid_array.
So the remove_dev_pasid() callback may get a wrong domain in the set_dev_pasid()
path. [1] Even the group is singleton, the existing code logic would have
unnecessary warnings in the error handling of the set_dev_pasid() path.

The fix is passing the domain to be detached to the remove_dev_pasid()
callback, and only undo the set_dev_pasid result on the devices that have
succeeded.

[1] https://lore.kernel.org/linux-iommu/20240320123803.GD159172@nvidia.com/

Regards,
	Yi Liu

Yi Liu (2):
  iommu: Pass domain to remove_dev_pasid() op
  iommu: Undo pasid attachment only for the devices that have succeeded

 drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c |  9 ++-----
 drivers/iommu/intel/iommu.c                 | 11 +++------
 drivers/iommu/iommu.c                       | 27 ++++++++++++++-------
 include/linux/iommu.h                       |  3 ++-
 4 files changed, 25 insertions(+), 25 deletions(-)