Re: [PATCH] mm/mmn: prevent unpaired invalidate_start and invalidate_end with non-blocking

[Jason Gunthorpe <jgg@mellanox.com>]

On Thu, Aug 08, 2019 at 10:18:27AM +0200, Michal Hocko wrote:
> On Wed 07-08-19 19:16:32, Jason Gunthorpe wrote:
> > Many users of the mmu_notifier invalidate_range callbacks maintain
> > locking/counters/etc on a paired basis and have long expected that
> > invalidate_range start/end are always paired.
> > 
> > The recent change to add non-blocking notifiers breaks this assumption
> > when multiple notifiers are present in the list as an EAGAIN return from a
> > later notifier causes all earlier notifiers to get their
> > invalidate_range_end() skipped.
> > 
> > During the development of non-blocking each user was audited to be sure
> > they can skip their invalidate_range_end() if their start returns -EAGAIN,
> > so the only place that has a problem is when there are multiple
> > subscriptions.
> > 
> > Due to the RCU locking we can't reliably generate a subset of the linked
> > list representing the notifiers already called, and generate an
> > invalidate_range_end() pairing.
> > 
> > Rather than design an elaborate fix, for now, just block non-blocking
> > requests early on if there are multiple subscriptions.
> 
> Which means that the oom path cannot really release any memory for
> ranges covered by these notifiers which is really unfortunate because
> that might cover a lot of memory. Especially when the particular range
> might not be tracked at all, right?

Yes, it is a very big hammer to avoid a bug where the locking schemes
get corrupted and the impacted drivers deadlock.

If you really don't like it then we have to push ahead on either an
rcu-safe undo algorithm or some locking thing. I've been looking at
the locking thing, so we can wait a bit more and see. 

At least it doesn't seem urgent right now as nobody is reporting
hitting this bug, but we are moving toward cases where a process will
have 4 notififers (amdgpu kfd, hmm, amd iommu, RDMA ODP), so the
chance is higher

> If a different fix is indeed too elaborate then make sure to let users
> known that there is a restriction in place and dump something useful
> into the kernel log.

The 'simple' alternative I see is to use a rcu safe undo algorithm,
such as sorting the hlist. This is not so much code, but it is tricky
stuff.

Jason