Hi, In a recent discussion [1], an unsafe locking pattern was discovered. Initially, it was believed that the locking issue is related to calling copy_from_user()/copy_to_user() under state_mutex [2], but it turns out that it's a generic issue that has impact on all drivers that are not implementing a "deferred reset".
This is a first attempt to simplify the migration device state locking, following a suggestion from Jason to try and solve the issue on core side rather than duplicating deferred reset in all of the drivers. Introduce a dedicated .migration_reset_state() callback called outside vdev->memory_lock to break the locking dependency chain and convert all of the drivers to use it. [1] https://lore.kernel.org/intel-xe/7dtl5qum4mfgjosj2mkfqu5u5tu7p2roi2et3env4lhrccmiqi@asemffaeeflr/ [2] https://lore.kernel.org/kvm/[email protected]/ Michał Winiarski (6): vfio: Introduce .migration_reset_state() callback hisi_acc_vfio_pci: Use .migration_reset_state() callback vfio/pds: Use .migration_reset_state() callback vfio/qat: Use .migration_reset_state() callback vfio/mlx5: Use .migration_reset_state() callback vfio/virtio: Use .migration_reset_state() callback .../vfio/pci/hisilicon/hisi_acc_vfio_pci.c | 7 +- drivers/vfio/pci/mlx5/cmd.c | 15 ++-- drivers/vfio/pci/mlx5/cmd.h | 3 - drivers/vfio/pci/mlx5/main.c | 59 ++++----------- drivers/vfio/pci/pds/pci_drv.c | 10 --- drivers/vfio/pci/pds/vfio_dev.c | 12 ++++ drivers/vfio/pci/qat/main.c | 36 +++++----- drivers/vfio/pci/vfio_pci_core.c | 25 ++++++- drivers/vfio/pci/virtio/common.h | 3 - drivers/vfio/pci/virtio/main.c | 1 - drivers/vfio/pci/virtio/migrate.c | 71 ++++++------------- include/linux/vfio.h | 4 ++ 12 files changed, 101 insertions(+), 145 deletions(-) -- 2.51.2
