Re: [7.0-rc1] codetag: kernel warning "alloc_tag was not set" during boot

[Suren Baghdasaryan <surenb@google.com>]

On Tue, Feb 24, 2026 at 6:26 AM Harry Yoo <harry.yoo@oracle.com> wrote:
>
> On Tue, Feb 24, 2026 at 11:16:12PM +0900, Harry Yoo wrote:
> > Just sharing updates on this before going to bed...
> >
> > I was able to reproduce it on my machine I have a working fix
> > (nowhere close to upstream quality, though)
> >
> > The reason why we're seeing "alloc_tag not was set" is
> > because SLUB allocates empty sheaves for kmalloc
> > with __GFP_NO_OBJ_EXT and it doesn't teach memory profiling
> > how to handle this when such sheaves are freed.
> >
> > When __GFP_NO_OBJ_EXT is used in alloc_slab_obj_exts(), it later
> > avoids this "alloc_tag was not set" warning by marking alloc_tags
> > empty in free_slab_obj_exts(), just before freeing obj_exts.
> >
> > And we don't handle this when allocating freeing sheaves
> > that were allocated with __GFP_NO_OBJ_EXT.

Ah, thanks for the analyses! So, the fact that __alloc_empty_sheaf()
sets __GFP_NO_OBJ_EXT when allocating the sheaf leads to this issue
later on. Makes sense. I only wonder why I could not reproduce this...
Will try to find out later.

> >
> > Passing __GFP_NO_OBJ_EXT skips 1) allocation of obj_exts,
> > and 2) alloc_tag_add() even when obj_exts is already allocated,
> > and this confuses memory profiling later.
> >
> > I'm adding the fix I have now. (I guess Suren might have some preference
> > on how to solve it though)
> >
> >   1. mark obj_exts allocation failure when slab has no obj_exts and
> >      gfp flag has __GFP_NO_OBJ_EXT, so that a later obj_exts allocation
> >      will mark alloc_tags empty.
> >
> >   2. Set alloc_tag when obj_exts is allocated available,
> >      even when __GFP_NO_OBJ_EXT is set.
> >
> >      Because it's already allocated, we don't have to worry about
> >      recursive allocation.
>
> Wait, isn't it much simpler to just do mark_objexts_empty(sheaf);
> just before freeing sheaves, if s->flags has SLAB_KMALLOC?

I think that's the best fix and it makes logical sense.
alloc_slab_obj_exts() allocates obj_exts with __GFP_NO_OBJ_EXT and
then free_slab_obj_exts() does mark_objexts_empty() before freeing
them. Similarly,  __alloc_empty_sheaf() allocates sheaves with
__GFP_NO_OBJ_EXT (for SLAB_KMALLOC objects), therefore
free_empty_sheaf() should also call mark_objexts_empty(sheaf) before
freeing them. mark_objexts_empty() currently accepts only slabobj_ext
but I think we can change that. I'll take a stab at it.

>
> /me goes to bed anyway
>
> --
> Cheers,
> Harry / Hyeonggon