On Sat, Jul 8, 2023 at 2:18 PM Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > > On Sat, 8 Jul 2023 at 12:12, Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote: > > > > kernel/fork.c | 1 + > > 1 file changed, 1 insertion(+) > > I ended up editing your explanation a lot. > > I'm not convinced that the bug has much to do with the delayed tlb flushing. > > I think it's more fundamental than some tlb coherence issue: our VM > copying simply expects to not have any unrelated concurrent page fault > activity, and various random internal data structures simply rely on > that. > > I made up an example that I'm not sure is relevant to any of the > particular failures, but that I think is a non-TLB case: the parent > 'vma->anon_vma' chain is copied by dup_mmap() in anon_vma_fork(), and > it's possible that the parent vma didn't have any anon_vma associated > with it at that point. > > But a concurrent page fault to the same vma - even *before* the page > tables have been copied, and when the TLB is still entirely coherent - > could then cause a anon_vma_prepare() on that parent vma, and > associate one of the pages with that anon-vma. > > Then the page table copy happens, and that page gets marked read-only > again, and is added to both the parent and the child vma's, but the > child vma never got associated with the parents new anon_vma, because > it didn't exist when anon_vma_fork() happened. > > Does this ever happen? I have no idea. But it would seem to be an > example that really has nothing to do with any TLB state, and is just > simply "we cannot handle concurrent page faults while we're busy > copying the mm". > > Again - maybe I messed up, but it really feels like the missing > vma_start_write() was more fundamental, and not some "TLB coherency" > issue. Sounds plausible. I'll try to use the reproducer to verify if that's indeed happening here. It's likely there are multiple problematic scenarios due to this missing lock though. Thanks, Suren. > > Linus
