On Tue, Feb 15, 2022 at 11:54 PM Michal Hocko <mhocko@xxxxxxxx> wrote: > > On Tue 15-02-22 12:19:22, Suren Baghdasaryan wrote: > > After exit_mmap frees all vmas in the mm, mm->mmap needs to be reset, > > otherwise it points to a vma that was freed and when reused leads to > > a use-after-free bug. > > OK, so I have dived into this again. exit_mmap doesn't reset mmap > indeed. That doesn't really matter for _oom victims_. Both the oom reaper and > mrelease do check for MMF_OOM_SKIP before calling __oom_reap_task_mm. > exit_mmap still sets MMF_OOM_SKIP before taking the mmap_lock for oom > victims so those paths should be still properly synchronized. I have > proposed to get rid of this > http://lkml.kernel.org/r/YbHIaq9a0CtqRulE@xxxxxxxxxxxxxx but we haven't > agreed on that. > > mrelease path is broken because it doesn't mark the process oom_victim > and so the MMF_OOM_SKIP synchronization doesn't work. So we really need > this. > > I would propose to rephrase the changelog to be more specific because I > do not want to remember all those details later on. > What about > " > oom reaping (__oom_reap_task_mm) relies on a 2 way synchronization with > exit_mmap. First it relies on the mmap_lock to exclude from unlock > path[1], page tables tear down (free_pgtables) and vma destruction. > This alone is not sufficient because mm->mmap is never reset. For > historical reasons[2] the lock is taken there is also MMF_OOM_SKIP set > for oom victims before. > > The oom reaper only ever looks at oom victims so the whole scheme works > properly but process_mrelease can opearate on any task (with fatal > signals pending) which doesn't really imply oom victims. That means that > the MMF_OOM_SKIP part of the synchronization doesn't work and it can > see a task after the whole address space has been demolished and > traverse an already released mm->mmap list. This leads to use after free > as properly caught up by KASAN report. > > Fix the issue by reseting mm->mmap so that MMF_OOM_SKIP synchronization > is not needed anymore. The MMF_OOM_SKIP is not removed from exit_mmap > yet but it acts mostly as an optimization now. > > [1] 27ae357fa82b ("mm, oom: fix concurrent munlock and oom reaper unmap, > v3") > [2] 212925802454 ("mm: oom: let oom_reap_task and exit_mmap run > concurrently") > " This changelog is very detailed and I support switching to it. Andrew, please let me know if I should re-post the patch with this description or you will just amend the one in your tree. > > I really have to say that I hate how complex this has grown in the name > of optimizations. This has backfired several times already resulting in > 2 security issues. I really hope to get read any note of the oom reaper > from exit_mmap. Agree. I want to take another stab at removing __oom_reap_task_mm from exit_mmap. Now that Hugh made changes to mlock mechanisms and __oom_reap_task_mm does not skip locked vmas I think that should be possible. Planning to look into that sometimes next week. > > > Reported-by: syzbot+2ccf63a4bd07cf39cab0@xxxxxxxxxxxxxxxxxxxxxxxxx > > Suggested-by: Michal Hocko <mhocko@xxxxxxxx> > > Signed-off-by: Suren Baghdasaryan <surenb@xxxxxxxxxx> > > Acked-by: Michal Hocko <mhocko@xxxxxxxx> Thanks! > > Thanks! > > --- > > mm/mmap.c | 1 + > > 1 file changed, 1 insertion(+) > > > > diff --git a/mm/mmap.c b/mm/mmap.c > > index 1e8fdb0b51ed..d445c1b9d606 100644 > > --- a/mm/mmap.c > > +++ b/mm/mmap.c > > @@ -3186,6 +3186,7 @@ void exit_mmap(struct mm_struct *mm) > > vma = remove_vma(vma); > > cond_resched(); > > } > > + mm->mmap = NULL; > > mmap_write_unlock(mm); > > vm_unacct_memory(nr_accounted); > > } > > -- > > 2.35.1.265.g69c8d7142f-goog > > -- > Michal Hocko > SUSE Labs
