Re: [PATCH V5 8/8] mm/slab: place slabobj_ext metadata in unused space within s->size

[Hao Li <hao.li@linux.dev>]

On Thu, Jan 08, 2026 at 07:44:47PM +0900, Harry Yoo wrote:
> On Thu, Jan 08, 2026 at 05:52:27PM +0800, Hao Li wrote:
> > On Thu, Jan 08, 2026 at 05:41:00PM +0900, Harry Yoo wrote:
> > > On Thu, Jan 08, 2026 at 01:52:09PM +0800, Hao Li wrote:
> > > > On Mon, Jan 05, 2026 at 05:02:30PM +0900, Harry Yoo wrote:
> > > > > When a cache has high s->align value and s->object_size is not aligned
> > > > > to it, each object ends up with some unused space because of alignment.
> > > > > If this wasted space is big enough, we can use it to store the
> > > > > slabobj_ext metadata instead of wasting it.
> > > > 
> > > > Hi, Harry,
> > > 
> > > Hi Hao,
> > > 
> > > > When we save obj_ext in s->size space, it seems that slab_ksize() might
> > > > be missing the corresponding handling.
> > > 
> > > Oops.
> > > 
> > > > It still returns s->size, which could cause callers of slab_ksize()
> > > > to see unexpected data (i.e. obj_ext), or even overwrite the obj_ext data.
> > > 
> > > Yes indeed.
> > > Great point, thanks!
> > > 
> > > I'll fix it by checking if the slab has obj_exts within the object
> > > layout and returning s->object_size if so.
> > 
> > Makes sense - I think there's one more nuance worth capturing.
> > slab_ksize() seems to compute the maximum safe size by applying layout
> > constraints from most-restrictive to least-restrictive:
> > redzones/poison/KASAN clamp it to object_size, tail metadata
> > (SLAB_TYPESAFE_BY_RCU / SLAB_STORE_USER) clamps it to inuse, and only
> > when nothing metadata lives does it return s->size.
> 
> Waaaait, SLAB_TYPESAFE_BY_RCU isn't the only case where we put freelist
> pointer after the object.
> 
> What about caches with constructor?
> We do place it after object, but slab_ksize() may return s->size? 

That's a really good question - thanks for calling it out. I took
another look at the code, and the comment for ksize() notes that it's
only meant to be used with kmalloc()-family allocations; those objects
don't have a ctor pointer. So as long as callers stick to that contract,
I think we should be fine and don't need to worry too much about this
case.

-- 
Thanks,
Hao