|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [PATCH v2] EFI: avoid use-after-free when chaining config files
On 18.11.2026 08:50, Anandu Padinhare wrote:
> When following "chain=", efi_start() frees the buffer holding the
> current config file (if it was allocated) and only afterwards converts
> the chained file's name, which points into that buffer, with s2w().
> The section name, when taken from "default=", also points there and is
> used for the subsequent lookups in the chained file.
>
> read_file() may well get back the very pages just freed (as observed
> with OVMF), so the section name may end up referring to arbitrary text
> of the chained file, selecting the wrong section (or none). With
> firmware overwriting freed memory (e.g. Ubuntu's OVMF 2025.11, which
> fills it with 0xAF), even the file name is garbage, and loading the
> chained file fails with "Open failed for ...: Invalid parameter".
>
> Convert the file name before freeing the buffer, and duplicate a
> section name taken from "default=" the first time round. A
> configuration relying on the stale pointer happening to refer to the
> chained file's own "default=" value will now use the section named in
> the original file instead.
>
> Fixes: 81da4016bfeb ("x86-64/EFI: allow chaining of config files")
> Assisted-by: Claude
> Signed-off-by: Anandu Padinhare <Anandu.Padinhare@xxxxxxxxxxxxxx>
> ---
> v2:
> - Introduce _strdup() and use it (Jan).
> - Duplicate the section name only once, and only when it was taken from
> "default=" (Jan).
>
> Tested in QEMU booting xen.efi directly (x86 with OVMF and a PV Dom0,
> Arm64 with AAVMF up to Dom0 construction): chaining with and without
> "default=", across several levels and (x86 only) with the section given
> on the command line; booting without "chain=" is unaffected.
>
> A pre-existing issue is left for a separate change: "chain=" has never
> detected cycles, so a config file chaining to itself makes Xen loop
> silently (unless the firmware garbles the freed name first, as Ubuntu's
> OVMF does). I can send a patch limiting the chaining depth, if so
> desired.
>
> xen/common/efi/boot.c | 29 ++++++++++++++++++++++++++++-
> 1 file changed, 28 insertions(+), 1 deletion(-)
>
> diff --git a/xen/common/efi/boot.c b/xen/common/efi/boot.c
> index 89b5831691..6db3573030 100644
> --- a/xen/common/efi/boot.c
> +++ b/xen/common/efi/boot.c
> @@ -285,6 +285,17 @@ static char *__init w2s(const union string *str)
> return str->s;
> }
>
> +static char *__init _strdup(const char *s)
> +{
> + size_t len = strlen(s) + 1;
> + void *ptr;
> +
> + if ( efi_bs->AllocatePool(EfiLoaderData, len, &ptr) != EFI_SUCCESS )
> + return NULL;
> +
> + return memcpy(ptr, s, len);
> +}
While I agree with the use of size_t and char here, the use of void (vs VOID)
is a little questionable. The EFI maintainers will want to judge.
> @@ -1441,6 +1452,7 @@ void EFIAPI __init noreturn efi_start(EFI_HANDLE
> ImageHandle,
> UINTN gop_mode = ~0;
> EFI_GRAPHICS_OUTPUT_PROTOCOL *gop = NULL;
> union string section = { NULL }, name;
> + bool section_in_cfg = false;
This doesn't need to live at function scope, it can move into the more narrow
if()'s one.
Then:
Reviewed-by: Jan Beulich <jbeulich@xxxxxxxx>
(Perhaps the adjustment(s) could be made while committing if no other need for
a v3 arises.)
Jan
|
![]() |
Lists.xenproject.org is hosted with RackSpace, monitoring our |