On Thu, Nov 17, 2022 at 03:54:21PM -0800, Dave Hansen wrote:
> On 11/17/22 15:23, Marco Elver wrote:
> > Yes - it's the 'level != PG_LEVEL_4K'.
>
> That plus the bisect made it pretty easy to find, thanks for the effort!
>
> Could you double-check that the attached patch fixes it? It seemed to
> for me.
>
> The issue was that the new "No changes, easy!" check in the suspect
> commit didn't check the cpa->force_split option. It didn't split down
> to 4k and then all hell broke loose.
>
> Oh, and I totally misread the kfence ability to tolerate partial TLB
> flushes. Sorry for the noise there!
> diff --git a/arch/x86/mm/pat/set_memory.c b/arch/x86/mm/pat/set_memory.c
> index 220361ceb997..9b4e2ad957f6 100644
> --- a/arch/x86/mm/pat/set_memory.c
> +++ b/arch/x86/mm/pat/set_memory.c
> @@ -1727,7 +1727,8 @@ static int __change_page_attr_set_clr(struct cpa_data *cpa, int primary)
> /*
> * No changes, easy!
> */
> - if (!(pgprot_val(cpa->mask_set) | pgprot_val(cpa->mask_clr)))
> + if (!(pgprot_val(cpa->mask_set) | pgprot_val(cpa->mask_clr))
> + && !cpa->force_split)
(operators go at the end of the previous line)
> return ret;
>
> while (rempages) {
Urgh.. sorry about that.
