vfio-pci currently goes through some pretty nasty locking algorithms
since commit abafbc551fdd ("vfio-pci: Invalidate mmaps and block MMIO
access on disabled memory") was added to invalidate and re-fault mmaps
to device MMIO around cases where device memory is disabled. This
series greatly simplifies that by making use of an address space on
the vfio device file descriptor, as suggested by Jason Gunthorpe.
This allows us to use unmap_mapping_range() on the device fd to zap
such mappings, and by creating a vma-to-pfn callback, we can implement
a reverse function to restore all mappings.
This series was originally part of a larger series which also added a
callback to get a vfio device from a vma, which allows the IOMMU
backend to limit pfnmaps to vfio device memory. The long term goal
is to implement the vma-to-pfn for all vfio device drivers to enable
this in the IOMMU backend and proceed with a mechanism to also
invalidate DMA mappings to device memory while disabled.
Given my slow progress towards that longer goal, I'd like to get this
in as an interim cleanup as it seems worthwhile on its own. I'll
intend to rework this on top of Jason's device_open/close series.
Thanks,
Alex
---
Alex Williamson (7):
vfio: Create vfio_fs_type with inode per device
vfio: Export unmap_mapping_range() wrapper
vfio/pci: Use vfio_device_unmap_mapping_range()
vfio,vfio-pci: Add vma to pfn callback
mm/interval_tree.c: Export vma interval tree iterators
vfio: Add vfio_device_io_remap_mapping_range()
vfio/pci: Remove map-on-fault behavior
drivers/vfio/pci/vfio_pci.c | 279 +++++++---------------------
drivers/vfio/pci/vfio_pci_config.c | 8 +-
drivers/vfio/pci/vfio_pci_private.h | 5 +-
drivers/vfio/vfio.c | 69 ++++++-
include/linux/vfio.h | 10 +
mm/interval_tree.c | 3 +
6 files changed, 156 insertions(+), 218 deletions(-)
