RE: [PATCH 2/3] mm/gup: Introduce get_user_pages_fast_longterm()

[Weiny, Ira]

> 
> On Mon, Feb 11, 2019 at 2:07 PM Jason Gunthorpe <jgg@xxxxxxxx> wrote:
> >
> > On Mon, Feb 11, 2019 at 01:52:38PM -0800, Ira Weiny wrote:
> > > On Mon, Feb 11, 2019 at 01:39:12PM -0800, John Hubbard wrote:
> > > > On 2/11/19 1:26 PM, Ira Weiny wrote:
> > > > > On Mon, Feb 11, 2019 at 01:13:56PM -0800, John Hubbard wrote:
> > > > >> On 2/11/19 12:39 PM, Jason Gunthorpe wrote:
> > > > >>> On Mon, Feb 11, 2019 at 12:16:42PM -0800, ira.weiny@xxxxxxxxx
> wrote:
> > > > >>>> From: Ira Weiny <ira.weiny@xxxxxxxxx>
> > > > >> [...]
> > > > >> It seems to me that the longterm vs. short-term is of questionable
> value.
> > > > >
> > > > > This is exactly why I did not post this before.  I've been
> > > > > waiting our other discussions on how GUP pins are going to be
> > > > > handled to play out.  But with the netdev thread today[1] it
> > > > > seems like we need to make sure we have a "safe" fast variant
> > > > > for a while.  Introducing FOLL_LONGTERM seemed like the cleanest
> > > > > way to do that even if we will not need the distinction in the
> > > > > future...  :-(
> > > >
> > > > Yes, I agree. Below...
> > > >
> > > > > [...]
> > > > > This is also why I did not change the get_user_pages_longterm
> > > > > because we could be ripping this all out by the end of the
> > > > > year...  (I hope. :-)
> > > > >
> > > > > So while this does "pollute" the GUP family of calls I'm hoping
> > > > > it is not forever.
> > > > >
> > > > > Ira
> > > > >
> > > > > [1] https://lkml.org/lkml/2019/2/11/1789
> > > > >
> > > >
> > > > Yes, and to be clear, I think your patchset here is fine. It is
> > > > easy to find the FOLL_LONGTERM callers if and when we want to
> > > > change anything. I just think also it's appopriate to go a bit further, and
> use FOLL_LONGTERM all by itself.
> > > >
> > > > That's because in either design outcome, it's better that way:
> > > >
> > > > is just right. The gup API already has _fast and non-fast
> > > > variants, and once you get past a couple, you end up with a
> > > > multiplication of names that really work better as flags. We're there.
> > > >
> > > > the _longterm API variants.
> > >
> > > Fair enough.   But to do that correctly I think we will need to convert
> > > get_user_pages_fast() to use flags as well.  I have a version of
> > > this series which includes a patch does this, but the patch touched
> > > a lot of subsystems and a couple of different architectures...[1]
> >
> > I think this should be done anyhow, it is trouble the two basically
> > identical interfaces have different signatures. This already caused a
> > bug in vfio..
> >
> > I also wonder if someone should think about making fast into a flag
> > too..
> >
> > But I'm not sure when fast should be used vs when it shouldn't :(
> 
> Effectively fast should always be used just in case the user cares about
> performance. It's just that it may fail and need to fall back to requiring the
> vma.
> 
> Personally I thought RDMA memory registration is a one-time / upfront slow
> path so that non-fast-GUP is tolerable.
> 
> The workloads that *need* it are O_DIRECT users that can't tolerate a vma
> lookup on every I/O.

There are some users who need to [un]register memory more often.  While not in the strict fast path these users would like the registrations to occur as fast as possible.  I don't personally have the results but our OPA team did do performance tests on the GUP vs GUP fast and for the hfi1 case fast was better.  I don't have any reason to believe that regular RDMA users would not also benefit.

Ira