Re: mm: delay rmap removal until after TLB flush

[David Hildenbrand]

On 31.10.22 19:43, Linus Torvalds wrote:
> Updated subject line, and here's the link to the original discussion
> for new people:
>
>      https://lore.kernel.org/all/B88D3073-440A-41C7-95F4-895D3F657EF2@xxxxxxxxx/
>
> On Mon, Oct 31, 2022 at 10:28 AM Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> > Ok. At that point we no longer have the pte or the virtual address, so
> > it's not going to be exactly the same debug output.
> >
> > But I think it ends up being fairly natural to do
> >
> >          VM_WARN_ON_ONCE_PAGE(page_mapcount(page) < 0, page);
> >
> > instead, and I've fixed that last patch up to do that.
>
> Ok, so I've got a fixed set of patches based on the feedback from
> PeterZ, and also tried to do the s390 updates for this blindly, and
> pushed them out into a git branch:
>
>      https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/log/?h=mmu_gather-race-fix
>
> If people really want to see the patches in email again, I can do
> that, but most of you already have, and the changes are either trivial
> fixes or the s390 updates.
>
> For the s390 people that I've now added to the participant list maybe
> the git tree is fine - and the fundamental explanation of the problem
> is in that top-most commit (with the three preceding commits being
> prep-work). Or that link to the thread about this all.
>
> That top-most commit is also where I tried to fix things up for s390
> that uses its own non-gathering TLB flush due to
> CONFIG_MMU_GATHER_NO_GATHER.
>
> NOTE NOTE NOTE! Unlike my regular git branch, this one may end up
> rebased etc for further comments and fixes. So don't consider that
> stable, it's still more of an RFC branch.
>
> At a minimum I'll update it with Ack's etc, assuming I get those, and
> my s390 changes are entirely untested and probably won't work.
>
> As far as I can tell, s390 doesn't actually *have* the problem that
> causes this change, because of its synchronous TLB flush, but it
> obviously needs to deal with the change of rmap zapping logic.
>
> Also added a few people who are explicitly listed as being mmu_gather
> maintainers. Maybe people saw the discussion on the linux-mm list, but
> let's make it explicit.
>
> Do people have any objections to this approach, or other suggestions?
>
> I do *not* consider this critical, so it's a "queue for 6.2" issue for me.
>
> It probably makes most sense to queue in the -MM tree (after the thing
> is acked and people agree), but I can keep that branch alive too and
> just deal with it all myself as well.
>
> Anybody?

Happy to see that we're still decrementing the mapcount before 
decrementingthe refcount, I was briefly concerned.

I was not able to come up quickly with something that would be 
fundamentally wrong here, but devil is in the detail.

Some minor things could be improved IMHO (ENCODE_PAGE_BITS naming is 
unfortunate, TLB_ZAP_RMAP could be a __bitwise type, using VM_WARN_ON 
instead of VM_BUG_ON).

I agree that 6.2 is good enough and that upstreaming this via the -MM 
tree would be a good way to move forward.

--
Thanks,

David / dhildenb