On Thu, Nov 09, 2023 at 04:04:39PM +0000, jeff.xie@xxxxxxxxx wrote:
> November 9, 2023 at 11:36 PM, "Matthew Wilcox" <willy@xxxxxxxxxxxxx> wrote:
> > But we want that anyway (or at least I do). You're right that vmalloc
> > pages are not marked as being vmalloc pages and don't contain the
> > information about which vmalloc area they belong to. I've talked about
> > ways we can add that information to folios in the past, but I have a lot
> > of other projects I'm working on. Are you interested in doing that?
> >
>
> Certainly, I'm willing to give it a try. If a folio can include vmalloc information
> or more information, this is great. I may need to understand the background of why
> you proposed this method in the past.
I can't find the proposal now, but it's basically this:
Turn PG_slab into a PG_kernel. If PG_kernel is set, then other flags
change their meaning. Flags that should be reusable: writeback,
referenced, uptodate, lru, active, workingset, private, reclaim
One of those flags gets reused to be the new slab. So, eg
folio_test_slab() becomes:
return (folio->flags & (PG_kernel | PG_slab)) == (PG_kernel | PG_slab);
Now we have somewhere that we can use for PG_vmalloc (also PG_reserved
can become a PG_kernel sub-flag, freeing up a page flag).
We'd need to change some helpers. eg folio_mapping() currently does:
if (unlikely(folio_test_slab(folio)))
return NULL;
and that should be:
if (unlikely(folio_test_kernel(folio)))
return NULL;
With that in place, we can reuse the folio->mapping space to point to
the struct vm_struct that allocated it.
This isn't an easy project and will require a lot of testing. It has
some upsides, like freeing up a page flag.
