Re: [PATCH 0/5] bio: Direct IO: convert to pin_user_pages_fast()

[John Hubbard]

On 8/24/20 7:07 PM, Al Viro wrote:
> On Tue, Aug 25, 2020 at 02:54:28AM +0100, Al Viro wrote:
> > On Fri, Aug 21, 2020 at 09:20:54PM -0700, John Hubbard wrote:
> >
> > > Direct IO behavior:
> > >
> > >      ITER_IOVEC:
> > >          pin_user_pages_fast();
> > >          break;
> > >
> > >      ITER_KVEC:    // already elevated page refcount, leave alone
> > >      ITER_BVEC:    // already elevated page refcount, leave alone
> > >      ITER_PIPE:    // just, no :)
> >
> > Why?  What's wrong with splice to O_DIRECT file?
>
> Sorry - s/to/from/, obviously.
>
> To spell it out: consider generic_file_splice_read() behaviour when
> the source had been opened with O_DIRECT; you will get a call of
> ->read_iter() into ITER_PIPE destination.  And it bloody well
> will hit iov_iter_get_pages() on common filesystems, to pick the
> pages we want to read into.
>
> So... what's wrong with having that "pin" primitive making sure
> the pages are there and referenced by the pipe?

(our emails crossed) OK, yes, let me hook that up. I was just unaware
of that flow, I'll go off and figure it out.

Thanks for looking at this!

thanks,
--
John Hubbard
NVIDIA