These days drivers with state should use cdev_device_add() and cdev_device_del() to manage the cdev and sysfs lifetime. This simple pattern ties all the state (vfio, dev, and cdev) together in one memory structure and uses container_of() to navigate between the layers. This is a followup to the discussion here: https://lore.kernel.org/kvm/20210921155705.GN327412@xxxxxxxxxx/ This builds on Christoph's work to revise how the vfio_group works and is against the latest VFIO tree. This is on github: https://github.com/jgunthorpe/linux/commits/vfio_group_cdev v3: - Streamline vfio_group_find_or_alloc() - Remove vfio_group_try_get() and just opencode the refcount_inc_not_zero v2: https://lore.kernel.org/r/0-v2-fd9627d27b2b+26c-vfio_group_cdev_jgg@xxxxxxxxxx - Remove comment before iommu_group_unregister_notifier() - Add comment explaining what the WARN_ONs vfio_group_put() do - Fix error logic around vfio_create_group() in patch 3 - Add horizontal whitespace - Clarify comment is refering to group->users v1: https://lore.kernel.org/r/0-v1-fba989159158+2f9b-vfio_group_cdev_jgg@xxxxxxxxxx Cc: Liu Yi L <yi.l.liu@xxxxxxxxx> Cc: "Tian, Kevin" <kevin.tian@xxxxxxxxx> Cc: Christoph Hellwig <hch@xxxxxx> Signed-off-by: Jason Gunthorpe <jgg@xxxxxxxxxx> Jason Gunthorpe (5): vfio: Delete vfio_get/put_group from vfio_iommu_group_notifier() vfio: Do not open code the group list search in vfio_create_group() vfio: Don't leak a group reference if the group already exists vfio: Use a refcount_t instead of a kref in the vfio_group vfio: Use cdev_device_add() instead of device_create() drivers/vfio/vfio.c | 381 +++++++++++++++++--------------------------- 1 file changed, 149 insertions(+), 232 deletions(-) base-commit: d9a0cd510c3383b61db6f70a84e0c3487f836a63 -- 2.33.0
