Re: [PATCH RFC v2 00/13] IOMMUFD Generic interface

> Do we have a solution to this? > > If not I would like to make a patch removing VFIO_DMA_UNMAP_FLAG_VADDR > > Aside from the approach to use the FD, another idea is to just use > fork. > > qemu would do something like > > .. stop all container ioctl activity .. > fork() > ioctl(CHANGE_MM) // switch all maps to this mm > .. signal parent.. > .. wait parent.. > exit(0) > .. wait child .. > exec() > ioctl(CHANGE_MM) // switch all maps to this mm > ..signal child.. > waitpid(childpid) > > This way the kernel is never left without a page provider for the > maps, the dummy mm_struct belonging to the fork will serve that role > for the gap. > > And the above is only required if we have mdevs, so we could imagine > userspace optimizing it away for, eg vfio-pci only cases. > > It is not as efficient as using a FD backing, but this is super easy > to implement in the kernel. I propose to avoid deadlock for mediated devices as follows. Currently, an mdev calling vfio_pin_pages blocks in vfio_wait while VFIO_DMA_UNMAP_FLAG_VADDR is asserted. * In vfio_wait, I will maintain a list of waiters, each list element consisting of (task, mdev, close_flag=false). * When the vfio device descriptor is closed, vfio_device_fops_release will notify the vfio_iommu driver, which will find the mdev on the waiters list, set elem->close_flag=true, and call wake_up_process for the task.