Re: [PATCH] slub: print just the function name in /sys/kernel/slab/*/ctor

On Tue, Aug 04, 2026 at 07:25:15PM +0200, Vlastimil Babka (SUSE) wrote:
> On 8/2/26 07:13, Harry Yoo wrote:
> > On Thu, Jul 30, 2026 at 04:15:26PM +0200, Vlastimil Babka (SUSE) wrote:
> >> On 7/27/26 07:27, Harry Yoo wrote:
> >> > 
> >> > 
> >> > On 7/25/26 10:28 PM, Alexey Dobriyan wrote:
> >> >> Sysfs "ctor" file prints offset/length of the cache's ctor function
> >> >> 
> >> >> 	$ sudo cat /sys/kernel/slab/bdev_cache/ctor
> >> >> 	init_once+0x0/0x10
> >> >> 
> >> >> This is not useful:
> >> >> 
> >> >> Offset will always be 0 because ctor is a function.
> >> >> 
> >> >> I'm not sure what ctor function size is doing here, it should be in
> >> >> /proc/kallsyms
> >> >> 
> >> >> Signed-off-by: Alexey Dobriyan <[email protected]>
> >> >> ---
> >> > 
> >> > I'm not convinced that changing this (without strong justification)
> >> > after exposing it to sysfs for 10+ years is worth the trouble.
> >> 
> >> Agreed. It's a pity this was exported in the first place. I can't see a
> >> benefit for anyone knowing what the function is called. Should have been at
> >> most a flag whether there's ctor or not.
> > 
> > Agreed.
> > 
> >> But possibly a justification is not to leak the function size, which might
> >> be theoretically (although unlikely) a hint to some attack.
> > 
> > ...but you need to be the owner (root) to read this :P 
> > 
> >> Well at least if somebody complains about getting broken, it's trivial to
> >> revert and we can hear about their usecase.
> > 
> > If we were to experiment with this, we'll change it to print a flag
> > rather than the ctor's name, right?
> 
> Yep, let's try that. Alexey?

I don't mind ctor name, it is internal implementation detail,
but offset/size are kind of useless.

My order of preferences is
1) delete +0x0/... part altogeter,
2) when someone complains, restore it as "+0x0/0x4" which is a minimum
   function size across all archs (or some other fixed fake size).

> 
> >> In-tree tools (slabinfo etc) were checked to work properly?
> > 
> > Looks like it doesn't read ctor at all.
> > 
> > Commit a87615b8f9e2 ("SLUB: slabinfo upgrade") added -o/--ops option
> > that should "Display of ctor / dtor etc.", but it reads "ops" attribute
> > that has never existed? Bit puzzled.
> 
> Might have been stale code from some previous revisions that included dtor.
> 
> >> >>  mm/slub.c |    2 +-
> >> >>  1 file changed, 1 insertion(+), 1 deletion(-)
> >> >> 
> >> >> --- a/mm/slub.c
> >> >> +++ b/mm/slub.c
> >> >> @@ -9149,7 +9149,7 @@ static ssize_t ctor_show(struct kmem_cache *s, char *buf)
> >> >>  {
> >> >>  	if (!s->ctor)
> >> >>  		return 0;
> >> >> -	return sysfs_emit(buf, "%pS\n", s->ctor);
> >> >> +	return sysfs_emit(buf, "%ps\n", s->ctor);
> >> >>  }
> >> >>  SLAB_ATTR_RO(ctor);