Re: [PATCH 3/3] ptdesc: Account page tables to memcgs again

[Shakeel Butt <shakeel.butt@linux.dev>]

On Wed, Feb 25, 2026 at 09:01:18PM +0000, Matthew Wilcox wrote:
> On Wed, Feb 25, 2026 at 08:55:54AM -0800, Shakeel Butt wrote:
> > >  #ifdef CONFIG_MEMCG
> > > -	unsigned long pt_memcg_data;
> > > +	struct mem_cgroup *pt_memcg;
> > 
> > This is kernel memory, so this would be struct obj_cgroup * instead of struct
> > mem_cgroup pointer. We will need something similar to __folio_objcg(), maybe
> > __ptdesc_objcg() and then call obj_cgroup_memcg() on it. Basically how
> > folio_memcg() handles the kernel memory.
> 
> Why would we want to do that instead of just stashing a pointer to the
> memcg in the ptdesc?

Not the memcg pointer but the objcg pointer, a bit background first though.

Underlying we are using alloc_pages_noprof(__GFP_ACCOUNT) and __free_pages() for
ptdesc, so allocation path looks like the following:

	alloc_pages_noprof(__GFP_ACCOUNT)
	...
		-> __alloc_frozen_pages_noprof(__GFP_ACCOUNT)
			-> __memcg_kmem_charge_page()
				-> page_set_objcg(page, objcg)

and page_set_objcg() is defined as 

static void page_set_objcg(struct page *page, const struct obj_cgroup *objcg)
{
	page->memcg_data = (unsigned long)objcg | MEMCG_DATA_KMEM;
}

page->memcg_data overlaps with ptdesc->pt_memcg_data, so we need remove
MEMCG_DATA_KMEM to get the objcg pointer.

If we want to store a pointer in struct ptdesc then we can't use the raw page
allocator/free functions. We have to allocate without __GFP_ACCOUNT and then do
the charging in __pagetable_ctor and uncharging in pagetable_dtor explicitly.

BTW we are trying to migrate from memcg pointers to objcg pointers in most of
the places due to zombie issue.

> I feel very stupid about the differences between
> all of these things and would dearly love to read some documentation to
> learn.

Unfortunetely we don't have a good documentation, just code.