Re: [PATCH v2 4/5] mm: make vma cache SLAB_TYPESAFE_BY_RCU

[Suren Baghdasaryan <surenb@google.com>]

On Tue, Nov 12, 2024 at 10:52 PM Hugh Dickins <hughd@google.com> wrote:
>
> On Tue, 12 Nov 2024, Suren Baghdasaryan wrote:
> > On Tue, Nov 12, 2024 at 9:08 PM Hugh Dickins <hughd@google.com> wrote:
> > > On Tue, 12 Nov 2024, Suren Baghdasaryan wrote:
> > > >
> > > > Thinking about this some more, I don't think this works. I'm relying
> > > > on vma_start_read() to stabilize the vma, however the lock I'm taking
> > > > is part of the vma which can be reused from under us. So, the lock I'm
> > > > taking might be reinitialized after I take the lock...
> > > > I need to figure out a way to stabilize the vma in some other manner
> > > > before taking this lock.
> > >
> > > (I'm not paying attention and following the patches, I just happened
> > > to notice this remark: forgive me if I'm out of context and have
> > > misunderstood, but hope this might help:)
> > >
> > > But this is exactly the problem SLAB_TYPESAFE_BY_RCU was invented for.
> > > You just have to be careful that the locks are initialized only when the
> > > slab is first created (allocated from buddy), not reinitialized whenever
> > > a new object is allocated from that slab.
> >
> > Hi Hugh!
> > I'm looking into SLAB_TYPESAFE_BY_RCU implementation and trying to
> > figure out if initializing the lock in the ctor() of the cache as
> > mentioned in the comment here:
> > https://elixir.bootlin.com/linux/v6.12-rc7/source/include/linux/slab.h#L127
> > would help my case. I assume that's what you are hinting here?
>
> Yes, if I'm "hinting", it's because offhand I forget the right names:
> "ctor", yes, that sounds right.

Just wanted to make sure I understood you correctly. Thanks for confirmation.

>
> Just grep around for examples of how it is used: there must be plenty
> now. but anon_vma is what it was first used for.

Yeah, there are plenty of examples now.

>
> But given the title of this patch, I'm surprised it's new to you.

Thinking about issues arising from possible object reuse is indeed new
to me, that's why I missed the lock reinitialization issue. I think I
know how to fix that now.
Thanks,
Suren.

>
> Hugh