Re: [PATCH 0/3] Add gup fast + longterm and use it in HFI1

[Jason Gunthorpe]

On Mon, Feb 11, 2019 at 09:14:56PM +0000, Weiny, Ira wrote:
> > 
> > On Mon, Feb 11, 2019 at 12:16:40PM -0800, ira.weiny@xxxxxxxxx wrote:
> > > From: Ira Weiny <ira.weiny@xxxxxxxxx>
> > >
> > > NOTE: This series depends on my clean up patch to remove the write
> > > parameter from gup_fast_permitted()[1]
> > >
> > > HFI1 uses get_user_pages_fast() due to it performance advantages.
> > > Like RDMA,
> > > HFI1 pages can be held for a significant time.  But
> > > get_user_pages_fast() does not protect against mapping of FS DAX pages.
> > 
> > If HFI1 can use the _fast varient, can't all the general RDMA stuff use it too?
> > 
> > What is the guidance on when fast vs not fast should be use?
> 
> Right now it can't because it holds mmap_sem across the call.  Once
> Shiraz's patches are accepted removing the umem->hugetlb flag I
> think we can change umem.c.

Okay, that make sense, we should change it when Shiraz's patches are
merged
 
> Also, it specifies FOLL_FORCE which can't currently be specified
> with gup fast.  One idea I had was to change get_user_pages_fast()
> to use gup_flags instead of a single write flag.  But that proved to
> be a very big cosmetic change across a lot of callers so I went this
> way.

I think you should do it.. :)

Jason