On Thu, Aug 10, 2023 at 3:16 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> On Thu, Aug 10, 2023 at 06:24:15AM +0000, Suren Baghdasaryan wrote:
> > Ok, I think I found the issue. wp_page_shared() ->
> > fault_dirty_shared_page() can drop mmap_lock (see the comment saying
> > "Drop the mmap_lock before waiting on IO, if we can...", therefore we
> > have to ensure we are not doing this under per-VMA lock.
>
> ... or we could change maybe_unlock_mmap_for_io() the same way
> that we changed folio_lock_or_retry():
>
> +++ b/mm/internal.h
> @@ -706,7 +706,7 @@ static inline struct file *maybe_unlock_mmap_for_io(struct vm_fault *vmf,
> if (fault_flag_allow_retry_first(flags) &&
> !(flags & FAULT_FLAG_RETRY_NOWAIT)) {
> fpin = get_file(vmf->vma->vm_file);
> - mmap_read_unlock(vmf->vma->vm_mm);
> + release_fault_lock(vmf);
> }
> return fpin;
> }
>
> What do you think?
This is very tempting... Let me try that and see if anything explodes,
but yes, this would be ideal.
>
> > I think what happens is that this path is racing with another page
> > fault which took mmap_lock for read. fault_dirty_shared_page()
> > releases this lock which was taken by another page faulting thread and
> > that thread generates an assertion when it finds out the lock it just
> > took got released from under it.
>
> I'm confused that our debugging didn't catch this earlier. lockdep
> should always catch this.
Maybe this condition is rare enough?
