Re: [PATCH 1/5] mm: Introduce mm_struct.has_pinned

[Jason Gunthorpe]

On Fri, Sep 25, 2020 at 02:06:59PM -0700, Linus Torvalds wrote:
> On Fri, Sep 25, 2020 at 12:56 PM Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> >
> > And honestly, since this is all getting fairly late in the rc, and it
> > took longer than I thought, I think we should do the GFP_ATOMIC
> > approach for now - not great, but since it only triggers for this case
> > that really should never happen anyway, I think it's probably the best
> > thing for 5.9, and we can improve on things later.
> 
> I'm not super-happy with this patch, but I'm throwing it out anyway, in case
> 
>  (a) somebody can test it - I don't have any test cases

It looks like it will work and resolve the RDMA case that triggered
this discussion. I will send it to our testers, should hear back
around Monday.

They previously said Peter's v1 patch worked, expecting the same here,
unless something unexpected hits the extra pre-conditions.

Though, we might hit the THP case and find it fails...

>  (b) somebody can find issues and improve on it

The THP hunks from Peter's series looked pretty straightforward, I'd
include at least the PMD one.

As a tiny optimization, the preconditions in copy_normal_page() could
order the atomics last to try and reduce the atomics done per fork.

> I'm happy to take Peter's series too, this is more of an alternative
> simplified version to keep the discussion going.

I don't completely grok the consequences of the anon_vma check. We
can exclude file backed mappings as they are broken for pinning
anyhow, so what is left that could be MAP_PRIVATE of a non-anon_vma?

Feels obscure, probably OK. If something does break userspace could
use MAP_SHARED and be fixed.

Otherwise, I do prefer Peter's version because of the GFP_KERNEL. To
touch on your other email..

It was my hope we could move away from the "This should never
happen". From a RDMA POV this idea was sort of managable years ago,
but now I have folks writing data science/ML software in Python that
deep under the libraries use RDMA and has pinned pages. It was a
Python program that detected this regression.

Having all that "just work" regardless of what foolish stuff happens
in the Python layer is very appealing.

Jason