On Wed, Oct 12, 2022 at 10:55:57AM -0400, Steven Sistare wrote:
On 10/12/2022 10:40 AM, Jason Gunthorpe wrote:
> On Wed, Oct 12, 2022 at 09:50:53AM -0400, Steven Sistare wrote:
>
>>> Anyhow, I think this conversation has convinced me there is no way to
>>> fix VFIO_DMA_UNMAP_FLAG_VADDR. I'll send a patch reverting it due to
>>> it being a security bug, basically.
>>
>> Please do not. Please give me the courtesy of time to develop a replacement
>> before we delete it. Surely you can make progress on other opens areas of
iommufd
>> without needing to delete this immediately.
>
> I'm not worried about iommufd, I'm worried about shipping kernels with
> a significant security problem backed into them.
>
> As we cannot salvage this interface it should quickly deleted so that
> it doesn't cause any incidents.
>
> It will not effect your ability to create a replacement.
I am not convinced we cannot salvage the interface, and indeed I might want to reuse
parts of it, and you are over-stating the risk of a feature that is already in
millions of kernels and has been for years. Deleting it all before having a
replacement hurts the people like myself who are continuing to develop and test
live update in qemu on the latest kernels.
I think this is a mistake, as I'm convinced it cannot be salvaged, but
we could instead guard it by a config symbol and CONFIG_EXPERIMENTAL.
Jason