On Wed, Feb 16, 2022 at 12:06 AM Michal Hocko <mhocko@xxxxxxxx> wrote: > > On Tue 15-02-22 15:02:54, Suren Baghdasaryan wrote: > > On Tue, Feb 15, 2022 at 12:05 PM Michal Hocko <mhocko@xxxxxxxx> wrote: > > > > > > One thing I was considering is to check agains ref counte overflo (a > > > deep process chain with many vmas could grow really high. ref_count > > > interface doesn't provide any easy way to check for overflows as far as > > > I could see from a quick glance so I gave up there but the logic would > > > be really straightforward. We just create a new anon_vma_name with the same > > > content and use it when duplicating if the usage grow really > > > (arbitrarily) high. > > > > I went over proposed changes. I see a couple small required fixes > > (resetting the name to NULL seems to be missing and I think > > dup_vma_anon_name needs some tweaking) but overall quite > > straight-forward. > > OK, great that this makes sense to you. As I've said I didn't really go > into details, not even dared to boot that to test. So it will very > likely need some more work but I do not expect this to grow much. > > > I'll post a separate patch to do this refactoring. > > The original patch is fixing the UAF issue, so I don't want to mix it > > with refactoring. Please let me know if you see an issue with > > separating it that way. > > Well, I am not sure TBH. Look at diffstats. Your fix > 2 files changed, 63 insertions(+), 17 deletions(-) > the refactoring which should fix this and potentially others that might > be still lurking there (because mixing shared pointers and their internal > objects just begs for problems) is > 7 files changed, 63 insertions(+), 86 deletions(-) > > more files touched for sure but the net result is much more clear and a > much more code removed. > The overflow logic would make it bigger but I guess the existing scheme > needs it as well. Ok, I'll see how to slice it after it's complete and tested. Thanks for the input! > > I would also claim that both approaches are really painful to review > because the existing model spreads into several areas and it is not > really clear you caught them all just by staring into the diff so both > will be rather painful to backport to older kernels. Fortunately this > would be only 5.17. > -- > Michal Hocko > SUSE Labs
