On Thu, Dec 15, 2022 at 09:38:42AM -0500, Steven Sistare wrote: > > I'm not sure this is quite, right, or at least the comment isn't quite > > right.. > > > > The issue is that the vfio_dma does not store the mm that provided the > > pages. By going through the task every time it allows the mm to change > > under its feet which of course can corrupt the assumed accounting in > > various ways. > > > > To fix this, the mm should be kept, as you did above. All the code > > that is touching the task to get the mm should be dropped. The only > > purpose of the task is to check the rlimit. > > Yes. While developing my "redo" series I tried it that way, but did not post > that version. Functionally it should be equivalent to this series, but I can > code it again to see if it looks cleaner. The big reason to split is that this existing bug pre-dates the vaddr work and should probably be backported further. It can be triggered just by doing exec.. Jason
