On Fri, Oct 13, 2023 at 9:08 AM Peter Xu <peterx@xxxxxxxxxx> wrote:
>
> On Fri, Oct 13, 2023 at 11:56:31AM +0200, David Hildenbrand wrote:
> > Hi Peter,
>
> Hi, David,
>
> >
> > > I used to have the same thought with David on whether we can simplify the
> > > design to e.g. limit it to single mm. Then I found that the trickiest is
> > > actually patch 1 together with the anon_vma manipulations, and the problem
> > > is that's not avoidable even if we restrict the api to apply on single mm.
> > >
> > > What else we can benefit from single mm? One less mmap read lock, but
> > > probably that's all we can get; IIUC we need to keep most of the rest of
> > > the code, e.g. pgtable walks, double pgtable lockings, etc.
> >
> > No existing mechanisms move anon pages between unrelated processes, that
> > naturally makes me nervous if we're doing it "just because we can".
>
> IMHO that's also the potential, when guarded with userfaultfd descriptor
> being shared between two processes.
>
> See below with more comment on the raised concerns.
>
> >
> > >
> > > Actually, even though I have no solid clue, but I had a feeling that there
> > > can be some interesting way to leverage this across-mm movement, while
> > > keeping things all safe (by e.g. elaborately requiring other proc to create
> > > uffd and deliver to this proc).
> >
> > Okay, but no real use cases yet.
>
> I can provide a "not solid" example. I didn't mention it because it's
> really something that just popped into my mind when thinking cross-mm, so I
> never discussed with anyone yet nor shared it anywhere.
>
> Consider VM live upgrade in a generic form (e.g., no VFIO), we can do that
> very efficiently with shmem or hugetlbfs, but not yet anonymous. We can do
> extremely efficient postcopy live upgrade now with anonymous if with REMAP.
>
> Basically I see it a potential way of moving memory efficiently especially
> with thp.
>
> >
> > >
> > > Considering Andrea's original version already contains those bits and all
> > > above, I'd vote that we go ahead with supporting two MMs.
> >
> > You can do nasty things with that, as it stands, on the upstream codebase.
> >
> > If you pin the page in src_mm and move it to dst_mm, you successfully broke
> > an invariant that "exclusive" means "no other references from other
> > processes". That page is marked exclusive but it is, in fact, not exclusive.
>
> It is still exclusive to the dst mm? I see your point, but I think you're
> taking exclusiveness altogether with pinning, and IMHO that may not be
> always necessary?
>
> >
> > Once you achieved that, you can easily have src_mm not have MMF_HAS_PINNED,
>
> (I suppose you meant dst_mm here)
>
> > so you can just COW-share that page. Now you successfully broke the
> > invariant that COW-shared pages must not be pinned. And you can even trigger
> > VM_BUG_ONs, like in sanity_check_pinned_pages().
>
> Yeah, that's really unfortunate. But frankly, I don't think it's the fault
> of this new feature, but the rest.
>
> Let's imagine if the MMF_HAS_PINNED wasn't proposed as a per-mm flag, but
> per-vma, which I don't see why we can't because it's simply a hint so far.
> Then if we apply the same rule here, UFFDIO_REMAP won't even work for
> single-mm as long as cross-vma. Then UFFDIO_REMAP as a whole feature will
> be NACKed simply because of this..
>
> And I don't think anyone can guarantee a per-vma MMF_HAS_PINNED can never
> happen, or any further change to pinning solution that may affect this. So
> far it just looks unsafe to remap a pin page to me.
>
> I don't have a good suggestion here if this is a risk.. I'd think it risky
> then to do REMAP over pinned pages no matter cross-mm or single-mm. It
> means probably we just rule them out: folio_maybe_dma_pinned() may not even
> be enough to be safe with fast-gup. We may need page_needs_cow_for_dma()
> with proper write_protect_seq no matter cross-mm or single-mm?
>
> >
> > Can it all be fixed? Sure, with more complexity. For something without clear
> > motivation, I'll have to pass.
>
> I think what you raised is a valid concern, but IMHO it's better fixed no
> matter cross-mm or single-mm. What do you think?
>
> In general, pinning lose its whole point here to me for an userspace either
> if it DONTNEEDs it or REMAP it. What would be great to do here is we unpin
> it upon DONTNEED/REMAP/whatever drops the page, because it loses its
> coherency anyway, IMHO.
>
> >
> > Once there is real demand, we can revisit it and explore what else we would
> > have to take care of (I don't know how memcg behaves when moving between
> > completely unrelated processes, maybe that works as expected, I don't know
> > and I have no time to spare on reviewing features with no real use cases)
> > and announce it as a new feature.
>
> Good point. memcg is probably needed..
>
> So you reminded me to do a more thorough review against zap/fault paths, I
> think what's missing are (besides page pinning):
>
> - mem_cgroup_charge()/mem_cgroup_uncharge():
>
> (side note: I think folio_throttle_swaprate() is only for when
> allocating new pages, so not needed here)
>
> - check_stable_address_space() (under pgtable lock)
>
> - tlb flush
>
> Hmm???????????????? I can't see anywhere we did tlb flush, batched or
> not, either single-mm or cross-mm should need it. Is this missing?
>
IIUC, ptep_clear_flush() flushes tlb entry. So I think we are doing
unbatched flushing. Possibly a nice performance improvement later on
would be to try doing it batched. Suren can throw more light on it.
One thing I was wondering is don't we need cache flush for the src
pages? mremap's move_page_tables() does it. IMHO, it's required here
as well.
> >
> >
> > Note: that (with only reading the documentation) it also kept me wondering
> > how the MMs are even implied from
> >
> > struct uffdio_move {
> > __u64 dst; /* Destination of move */
> > __u64 src; /* Source of move */
> > __u64 len; /* Number of bytes to move */
> > __u64 mode; /* Flags controlling behavior of move */
> > __s64 move; /* Number of bytes moved, or negated error */
> > };
> >
> > That probably has to be documented as well, in which address space dst and
> > src reside.
>
> Agreed, some better documentation will never hurt. Dst should be in the mm
> address space that was bound to the userfault descriptor. Src should be in
> the current mm address space.
>
> Thanks,
>
> --
> Peter Xu
>
