Hi, On Thu, Oct 24, 2019 at 11:10:54AM -0400, Andrea Arcangeli wrote: > Hello, > > On Thu, Oct 24, 2019 at 12:02:59PM +0300, Mike Rapoport wrote: > > That's no the reason that UFFD_FEATURE_EVENT_FORK does not show up in > > Debian code search, CRIU simply is not there. Debian packages CRIU only in > > experimental and I believe that's not indexed by the code search. > > > > As for the limitations, the races were fixed, I just forgot to update the > > wiki. As for the supported memory types and COW pages, these only affect > > efficiency of post-copy, but not the correctness. > > That's what I was hoping for. If the wiki information is stale and > there are no races it is totally plausible that it's being actively > used in production so we need to fix the kernel bug. I was just > checking because I wasn't sure anymore of the status after I read the > wiki. > > If the CRIU initialization code that issues the uffd syscall runs as > global root the ABI breaking permission check from Andy sounds the > simplest for a short term fix, because it will be unnoticed by any > production usage with CIRU --lazy-pages. In general, criu can run as non-root, but such use of criu has limitations, so allowing criu --lazy-pages only for root sounds reasonable as a short term solution. > Then later we could add a UFFD_FEATURE_EVENT_FORK2 that will not > require root permission. Agree. > Thanks, > Andrea > -- Sincerely yours, Mike.
