Re: [PATCH 3/3] s390: supplement for ptdesc conversion

[Qi Zheng]

On 2024/3/27 03:48, Vishal Moola wrote:
> On Mon, Mar 04, 2024 at 07:07:20PM +0800, Qi Zheng wrote:
> > --- a/arch/s390/mm/gmap.c
> > +++ b/arch/s390/mm/gmap.c
> > @@ -206,9 +206,11 @@ static void gmap_free(struct gmap *gmap)
> >   
> >   	/* Free additional data for a shadow gmap */
> >   	if (gmap_is_shadow(gmap)) {
> > +		struct ptdesc *ptdesc;
> > +
> >   		/* Free all page tables. */
> > -		list_for_each_entry_safe(page, next, &gmap->pt_list, lru)
> > -			page_table_free_pgste(page);
> > +		list_for_each_entry_safe(ptdesc, next, &gmap->pt_list, pt_list)
> > +			page_table_free_pgste(ptdesc);
>
> An important note: ptdesc allocation/freeing is different than the
> standard alloc_pages()/free_pages() routines architectures are used to.
> Are we sure we don't have memory leaks here?
>
> We always allocate and free ptdescs as compound pages; for page table
> struct pages, most archictectures do not. s390 has CRST_ALLOC_ORDER
> pagetables, meaning if we free anything using the ptdesc api, we better
> be sure it was allocated using the ptdesc api as well.

According to the code inspection, all ptdescs added to the pmap->pt_list
are allocated via page_table_alloc_pgste().

> Like you, I don't have a s390 to test on, so hopefully some s390 expert
> can chime in to let us know if we need a fix for this.

Yes, hope so!