Re: [PATCH] gfs2: Fix mmap + page fault deadlocks

[Linus Torvalds <torvalds@linux-foundation.org>]

On Thu, Jul 1, 2021 at 1:43 PM Andreas Gruenbacher <agruenba@redhat.com> wrote:
>
> here's another attempt at fixing the mmap + page fault deadlocks we're
> seeing on gfs2.  Still not ideal because get_user_pages_fast ignores the
> current->pagefault_disabled flag

Of course get_user_pages_fast() ignores the pagefault_disabled flag,
because it doesn't do any page faults.

If you don't want to fall back to the "maybe do IO" case, you should
use the FOLL_FAST_ONLY flag - or get_user_pages_fast_only(), which
does that itself.

> For getting get_user_pages_fast changed to fix this properly, I'd need
> help from the memory management folks.

I really don't think you need anything at all from the mm people,
because we already support that whole "fast only" case.

Also, I have to say that I think the direct-IO code is fundamentally
mis-designed. Why it is doing the page lookup _during_ the IO is a
complete mystery to me. Why wasn't that done ahead of time before the
filesystem took the locks it needed?

So what the direct-IO code _should_ do is to turn an ITER_IOVEC into a
ITER_KVEC by doing the page lookup ahead of time, and none of these
issues should even exist, and then the whole pagefault_disabled and/or
FOLL_FAST_ONLY would be a complete non-issue.

Is there any reason why that isn't what it does (other than historical baggage)?

               Linus