Re: [PATCH 3/3] x86: mm: add x86_64 support for page table check

[Pasha Tatashin <pasha.tatashin@soleen.com>]

On Thu, Dec 2, 2021 at 5:05 AM Jiri Slaby <jirislaby@kernel.org> wrote:
>
> On 02. 12. 21, 10:59, Jiri Slaby wrote:
> > On 01. 12. 21, 14:55, Pasha Tatashin wrote:
> >> On Wed, Dec 1, 2021 at 3:44 AM Jiri Slaby <jirislaby@kernel.org> wrote:
> >>>
> >>> On 01. 12. 21, 9:00, Jiri Slaby wrote:
> >>>> On 23. 11. 21, 22:48, Pasha Tatashin wrote:
> >>>>> Add page table check hooks into routines that modify user page tables.
> >>>>
> >>>> Hi,
> >>>>
> >>>> I bisected to this as this causes crashes during modules load:
> >>>
> >>> And it's not enough to unset CONFIG_PAGE_TABLE_CHECK_ENFORCED. I had to
> >>> unset CONFIG_PAGE_TABLE_CHECK completely to get rid of this.
> >>
> >> Hi,
> >>
> >> Thanks for reporting this. Seems like module load for some reasons
> >> does not like the static branches. However, I was not able to repro
> >> this. Could you please share your config and the module that you were
> >> loading?
> >
> > It's the openSUSE's -next config:
> > https://raw.githubusercontent.com/openSUSE/kernel-source/linux-next/config/x86_64/vanilla
> >
> >
> > But with CONFIG_IWLMEI=n (as that fails to link).
> >
> > One has to load i915, other modules (71 on my system) are apparently fine.
> >
> > i915 tries to patch your `page_table_check_disabled' which is defined as
> > `DEFINE_STATIC_KEY_TRUE_RO':
> >  > jump_label_add_module: key=__tracepoint_mmap_lock_released
> > (ffffffff93f36d88) mod=0000000000000000
> >  > jump_label_add_module: key=hugetlb_free_vmemmap_enabled_key
> > (ffffffff94873560) mod=0000000000000000
> >  > jump_label_add_module: key=devmap_managed_key (ffffffff94902700)
> > mod=0000000000000000
> >  > jump_label_add_module: key=page_table_check_disabled
> > (ffffffff939da760) mod=0000000000000000
> >  > BUG: unable to handle page fault for address: ffffffff939da768
>
> Provided you EXPORT page_table_check_disabled and it can be used from
> modules, it can be barely marked as RO. So the below indeed fixes the
> problem for me:
>
> --- a/mm/page_table_check.c
> +++ b/mm/page_table_check.c
> @@ -18,7 +18,7 @@ struct page_table_check {
>   static bool __page_table_check_enabled __initdata =
>
> IS_ENABLED(CONFIG_PAGE_TABLE_CHECK_ENFORCED);
>
> -DEFINE_STATIC_KEY_TRUE_RO(page_table_check_disabled);
> +DEFINE_STATIC_KEY_TRUE(page_table_check_disabled);

Makes sense, thanks! I will remove _RO in the next version.

Pasha