On Sun, Oct 4, 2020 at 1:24 AM Jason Gunthorpe <jgg@xxxxxxxx> wrote:
> On Sat, Oct 03, 2020 at 03:52:32PM -0700, John Hubbard wrote:
> > On 10/3/20 2:45 AM, Daniel Vetter wrote:
> > > On Sat, Oct 3, 2020 at 12:39 AM John Hubbard <jhubbard@xxxxxxxxxx> wrote:
> > > >
> > > > On 10/2/20 10:53 AM, Daniel Vetter wrote:
> > > > > For $reasons I've stumbled over this code and I'm not sure the change
> > > > > to the new gup functions in 55a650c35fea ("mm/gup: frame_vector:
> > > > > convert get_user_pages() --> pin_user_pages()") was entirely correct.
> > > > >
> > > > > This here is used for long term buffers (not just quick I/O) like
> > > > > RDMA, and John notes this in his patch. But I thought the rule for
> > > > > these is that they need to add FOLL_LONGTERM, which John's patch
> > > > > didn't do.
> > > >
> > > > Yep. The earlier gup --> pup conversion patches were intended to not
> > > > have any noticeable behavior changes, and FOLL_LONGTERM, with it's
> > > > special cases and such, added some risk that I wasn't ready to take
> > > > on yet. Also, FOLL_LONGTERM rules are only *recently* getting firmed
> > > > up. So there was some doubt at least in my mind, about which sites
> > > > should have it.
> > > >
> > > > But now that we're here, I think it's really good that you've brought
> > > > this up. It's definitely time to add FOLL_LONGTERM wherever it's missing.
> > >
> > > So should I keep this patch, or will it collide with a series you're working on?
> >
> > It doesn't collide with anything on my end yet, because I've been slow to
> > pick up on the need for changing callsites to add FOLL_LONGTERM. :)
> >
> > And it looks like that's actually a problem, because:
> >
> > >
> > > Also with the firmed up rules, correct that I can also drop the
> > > vma_is_fsdax check when the FOLL_LONGTERM flag is set?
> >
> > That's the right direction to go *in general*, but I see that the
> > pin_user_pages code is still a bit stuck in the past. And this patch
> > won't actually work, with or without that vma_is_fsdax() check.
> > Because:
> >
> > get_vaddr_frames(FOLL_LONGTERM)
> > pin_user_pages_locked()
> > if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
> > return -EINVAL;
>
> There is no particular reason this code needs to have the mm sem at
> that point.
>
> It should call pin_user_pages_fast() and only if that fails get the mmap
> lock and extract the VMA to do broken hackery.
Yeah I think that works. I tried understanding gup.c code a bit more,
and it looks like FOLL_LONGTERM only works for the pup_fast variant
right now? All others seem to have this comment that it's somehow
incompatible with FAULT_FLAG_ALLOW_RETRY and daxfs. But grepping
around for that didn't show up anything, at least not nearby dax code.
For my understanding of all this, what's the hiccup there?
For plans, I only started this for a bit of my own learning, but I
think I'll respin with the following changes:
- convert exynos and habanalabs to pin_user_pages_fast directly,
instead of going through this frame-vector detour
- move the locking and convert get_vaddr_frames to pup_fast as Jason suggested
- hack up some truly gross rfc to plug the follow_pfn hole
Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
