[PATCH 0/5] use pinned_vm instead of locked_vm to account pinned pages

[Ira Weiny]

On Wed, Feb 13, 2019 at 11:00:06PM -0700, Jason Gunthorpe wrote:
> On Wed, Feb 13, 2019 at 05:53:14PM -0800, Ira Weiny wrote:
> > On Mon, Feb 11, 2019 at 03:54:47PM -0700, Jason Gunthorpe wrote:
> > > On Mon, Feb 11, 2019 at 05:44:32PM -0500, Daniel Jordan wrote:
> > > 
> > > > All five of these places, and probably some of Davidlohr's conversions,
> > > > probably want to be collapsed into a common helper in the core mm for
> > > > accounting pinned pages.  I tried, and there are several details that
> > > > likely need discussion, so this can be done as a follow-on.
> > > 
> > > I've wondered the same..
> > 
> > I'm really thinking this would be a nice way to ensure it gets cleaned up and
> > does not happen again.
> > 
> > Also, by moving it to the core we could better manage any user visible changes.
> > 
> > From a high level, pinned is a subset of locked so it seems like we need a 2
> > sets of helpers.
> > 
> > try_increment_locked_vm(...)
> > decrement_locked_vm(...)
> > 
> > try_increment_pinned_vm(...)
> > decrement_pinned_vm(...)
> > 
> > Where try_increment_pinned_vm() also increments locked_vm...  Of course this
> > may end up reverting the improvement of Davidlohr  Bueso's atomic work...  :-(
> > 
> > Furthermore it would seem better (although I don't know if at all possible) if
> > this were accounted for in core calls which tracked them based on how the pages
> > are being used so that drivers can't call try_increment_locked_vm() and then
> > pin the pages...  Thus getting the account wrong vs what actually happened.
> > 
> > And then in the end we can go back to locked_vm being the value checked against
> > RLIMIT_MEMLOCK.
> 
> Someone would need to understand the bug that was fixed by splitting
> them. 
>

My suggestion above assumes that splitting them is required/correct.  To be
fair I've not dug into if this is true or not, but I trust Christopher.

What I have found is this commit:

bc3e53f682d9 mm: distinguish between mlocked and pinned pages

I think that commit introduced the bug (for IB) which at the time may have been
"ok" because many users of IB at the time were HPC/MPI users and I don't think
MPI does a lot of _separate_ mlock operations so the count of locked_vm was
probably negligible.  Alternatively, the clusters I've worked on in the past
had compute nodes set with RLIMIT_MEMLOCK to 'unlimited' whilst running MPI
applications on compute nodes of a cluster...  :-/

I think what Christopher did was probably ok for the internal tracking but we
_should_ have had something which summed the 2 for RLIMIT_MEMLOCK checking at
that time to be 100% correct?  Christopher do you remember why you did not do
that?

[1] http://lkml.kernel.org/r/20130524140114.GK23650@twins.programming.kicks-ass.net

> 
> I think it had to do with double accounting pinned and mlocked pages
> and thus delivering a lower than expected limit to userspace.
> 
> vfio has this bug, RDMA does not. RDMA has a bug where it can
> overallocate locked memory, vfio doesn't.

Wouldn't vfio also be able to overallocate if the user had RDMA pinned pages?

I think the problem is that if the user calls mlock on a large range then both
vfio and RDMA could potentially overallocate even with this fix.  This was your
initial email to Daniel, I think...  And Alex's concern.

> 
> Really unclear how to fix this. The pinned/locked split with two
> buckets may be the right way.

Are you suggesting that we have 2 user limits?

Ira

> 
> Jason