2023-07-23 22:10:57

by Hugh Dickins

Subject: [PATCH mm-hotfixes] mm/pagewalk: fix EFI_PGT_DUMP of espfix area

Booting x86_64 with CONFIG_EFI_PGT_DUMP=y shows messages of the form
"mm/pgtable-generic.c:53: bad pmd (____ptrval____)(8000000100077061)".

EFI_PGT_DUMP dumps all of efi_mm, including the espfix area, which is
set up with pmd entries which fit the pmd_bad() check: so 0d940a9b270b
warns and clears those entries, which would ruin running Win16 binaries.

The failing pte_offset_map() stopped such a kernel from even booting,
until a few commits later be872f83bf57 changed the pagewalk to tolerate
that: but it needs to be even more careful, to not spoil those entries.

I might have preferred to change init_espfix_ap() not to use "bad" pmd
entries; or to leave them out of the efi_mm dump. But there is great
value in staying away from there, and a pagewalk check of address
against TASK_SIZE may protect from other such aberrations too.

Reported-by: Mikhail Gavrilov <[email protected]>
Closes: https://lore.kernel.org/linux-mm/CABXGCsN3JqXckWO=V7p=FhPU1tK03RE1w9UE6xL5Y86SMk209w@mail.gmail.com/
Fixes: 0d940a9b270b ("mm/pgtable: allow pte_offset_map[_lock]() to fail")
Fixes: be872f83bf57 ("mm/pagewalk: walk_pte_range() allow for pte_offset_map()")
Signed-off-by: Hugh Dickins <[email protected]>
---
mm/pagewalk.c | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/mm/pagewalk.c b/mm/pagewalk.c
index 64437105fe0d..2022333805d3 100644
--- a/mm/pagewalk.c
+++ b/mm/pagewalk.c
@@ -48,8 +48,11 @@ static int walk_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
if (walk->no_vma) {
/*
* pte_offset_map() might apply user-specific validation.
+ * Indeed, on x86_64 the pmd entries set up by init_espfix_ap()
+ * fit its pmd_bad() check (_PAGE_NX set and _PAGE_RW clear),
+ * and CONFIG_EFI_PGT_DUMP efi_mm goes so far as to walk them.
*/
- if (walk->mm == &init_mm)
+ if (walk->mm == &init_mm || addr >= TASK_SIZE)
pte = pte_offset_kernel(pmd, addr);
else
pte = pte_offset_map(pmd, addr);
--
2.35.3

2023-07-24 09:09:31

by Mikhail Gavrilov

Subject: Re: [PATCH mm-hotfixes] mm/pagewalk: fix EFI_PGT_DUMP of espfix area

On Mon, Jul 24, 2023 at 2:17 AM Hugh Dickins <[email protected]> wrote:
>
> Booting x86_64 with CONFIG_EFI_PGT_DUMP=y shows messages of the form
> "mm/pgtable-generic.c:53: bad pmd (____ptrval____)(8000000100077061)".
>
> EFI_PGT_DUMP dumps all of efi_mm, including the espfix area, which is
> set up with pmd entries which fit the pmd_bad() check: so 0d940a9b270b
> warns and clears those entries, which would ruin running Win16 binaries.
>
> The failing pte_offset_map() stopped such a kernel from even booting,
> until a few commits later be872f83bf57 changed the pagewalk to tolerate
> that: but it needs to be even more careful, to not spoil those entries.
>
> I might have preferred to change init_espfix_ap() not to use "bad" pmd
> entries; or to leave them out of the efi_mm dump. But there is great
> value in staying away from there, and a pagewalk check of address
> against TASK_SIZE may protect from other such aberrations too.
>
> Reported-by: Mikhail Gavrilov <[email protected]>
> Closes: https://lore.kernel.org/linux-mm/CABXGCsN3JqXckWO=V7p=FhPU1tK03RE1w9UE6xL5Y86SMk209w@mail.gmail.com/
> Fixes: 0d940a9b270b ("mm/pgtable: allow pte_offset_map[_lock]() to fail")
> Fixes: be872f83bf57 ("mm/pagewalk: walk_pte_range() allow for pte_offset_map()")
> Signed-off-by: Hugh Dickins <[email protected]>
> ---
> mm/pagewalk.c | 5 ++++-
> 1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/mm/pagewalk.c b/mm/pagewalk.c
> index 64437105fe0d..2022333805d3 100644
> --- a/mm/pagewalk.c
> +++ b/mm/pagewalk.c
> @@ -48,8 +48,11 @@ static int walk_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
> if (walk->no_vma) {
> /*
> * pte_offset_map() might apply user-specific validation.
> + * Indeed, on x86_64 the pmd entries set up by init_espfix_ap()
> + * fit its pmd_bad() check (_PAGE_NX set and _PAGE_RW clear),
> + * and CONFIG_EFI_PGT_DUMP efi_mm goes so far as to walk them.
> */
> - if (walk->mm == &init_mm)
> + if (walk->mm == &init_mm || addr >= TASK_SIZE)
> pte = pte_offset_kernel(pmd, addr);
> else
> pte = pte_offset_map(pmd, addr);
> --
> 2.35.3
>

Thanks,
I confirm with this patch "bad pmd" went from kernel logs on all my machines.
Tested-by: Mikhail Gavrilov <[email protected]>

--
Best Regards,
Mike Gavrilov.