Re: [PATCH slab v5 5/6] slab: Reuse first bit for OBJEXTS_ALLOC_FAIL

[Alexei Starovoitov <alexei.starovoitov@gmail.com>]

On Fri, Sep 12, 2025 at 5:36 PM Suren Baghdasaryan <surenb@google.com> wrote:
>
> > > > Suren is
> > > > fixing the condition of VM_BUG_ON_PAGE() in slab_obj_exts(). With this
> > > > patch, I think, that condition will need to be changed again.
> > >
> > > That's orthogonal and I'm not convinced it's correct.
> > > slab_obj_exts() is doing the right thing. afaict.
> >
> > Currently we have
> >
> > VM_BUG_ON_PAGE(obj_exts && !(obj_exts & MEMCG_DATA_OBJEXTS))
> >
> > but it should be (before your patch) something like:
> >
> > VM_BUG_ON_PAGE(obj_exts && !(obj_exts & (MEMCG_DATA_OBJEXTS | OBJEXTS_ALLOC_FAIL)))
> >
> > After your patch, hmmm, the previous one would be right again and the
> > newer one will be the same as the previous due to aliasing. This patch
> > doesn't need to touch that VM_BUG. Older kernels will need to move to
> > the second condition though.
>
> Correct. Currently slab_obj_exts() will issue a warning when (obj_exts
> == OBJEXTS_ALLOC_FAIL), which is a perfectly valid state indicating
> that previous allocation of the vector failed due to memory
> exhaustion. Changing that warning to:
>
> VM_BUG_ON_PAGE(obj_exts && !(obj_exts & (MEMCG_DATA_OBJEXTS |
> OBJEXTS_ALLOC_FAIL)))
>
> will correctly avoid this warning and after your change will still
> work. (MEMCG_DATA_OBJEXTS | OBJEXTS_ALLOC_FAIL) when
> (MEMCG_DATA_OBJEXTS == OBJEXTS_ALLOC_FAIL) is technically unnecessary
> but is good for documenting the conditions we are checking.

I see what you mean. I feel the comment in slab_obj_exts()
that explains all that would be better long term than decipher
from C code. Both are fine, I guess.