[PATCH] efi_loader: fix use after free in efi_exit() with tcg2

[Vincent Stehlé]

(Updated Kojima-san's e-mail address.)

On Tue, Jan 27, 2026 at 09:58:33AM +0200, Ilias Apalodimas wrote:
> Hi Vincent,

Hi Ilias,

Thank you for having a look!
My comments below.

> On Fri, 23 Jan 2026 at 14:58, Vincent Stehlé <vincent.stehle at arm.com> wrote:
> >
> > The efi_exit() function frees the loaded image memory by calling
> > efi_delete_image(). However, when CONFIG_EFI_TCG2_PROTOCOL is enabled, the
> > image_obj->image_type structure member is accessed after the memory has
> > been freed.
> >
> > Fix this by keeping a copy of image_type, as is already done for exit_jmp.
> >
> > Fixes: 8fc4e0b4273a ("efi_loader: add boot variable measurement")
> > Signed-off-by: Vincent Stehlé <vincent.stehle at arm.com>
> > Cc: Heinrich Schuchardt <xypron.glpk at gmx.de>
> > Cc: Ilias Apalodimas <ilias.apalodimas at linaro.org>
> > Cc: Tom Rini <trini at konsulko.com>
> > Cc: Masahisa Kojima <masahisa.kojima at linaro.org>
> > ---
> >
> > Hi,
> >
> > This can be verified with sandbox_defconfig + CONFIG_VALGRIND=y and the
> > following command:
> >
> >   valgrind --suppressions=scripts/u-boot.supp \
> >     ./u-boot -T -c "setenv efi_selftest start image return; \
> >                     bootefi selftest"
> >
> > Best regards,
> > Vincent.
> >
> >  lib/efi_loader/efi_boottime.c | 6 ++++--
> >  1 file changed, 4 insertions(+), 2 deletions(-)
> >
> > diff --git a/lib/efi_loader/efi_boottime.c b/lib/efi_loader/efi_boottime.c
> > index ddc935d2240..0b3f2fd276c 100644
> > --- a/lib/efi_loader/efi_boottime.c
> > +++ b/lib/efi_loader/efi_boottime.c
> > @@ -3453,6 +3453,7 @@ static efi_status_t EFIAPI efi_exit(efi_handle_t image_handle,
> >         struct efi_loaded_image_obj *image_obj =
> >                 (struct efi_loaded_image_obj *)image_handle;
> >         jmp_buf *exit_jmp;
> > +       u16 image_type;
> >
> >         EFI_ENTRY("%p, %ld, %zu, %p", image_handle, exit_status,
> >                   exit_data_size, exit_data);
> > @@ -3496,13 +3497,14 @@ static efi_status_t EFIAPI efi_exit(efi_handle_t image_handle,
> >         }
> >         /* efi_delete_image() frees image_obj. Copy before the call. */
> >         exit_jmp = image_obj->exit_jmp;
> > +       image_type = image_obj->image_type;
> >         *image_obj->exit_status = exit_status;
> > -       if (image_obj->image_type == IMAGE_SUBSYSTEM_EFI_APPLICATION ||
> > +       if (image_type == IMAGE_SUBSYSTEM_EFI_APPLICATION ||
> >             exit_status != EFI_SUCCESS)
> >                 efi_delete_image(image_obj, loaded_image_protocol);
> >
> >         if (IS_ENABLED(CONFIG_EFI_TCG2_PROTOCOL)) {
> > -               if (image_obj->image_type == IMAGE_SUBSYSTEM_EFI_APPLICATION) {
> > +               if (image_type == IMAGE_SUBSYSTEM_EFI_APPLICATION) {
> 
> This looks ok, but I don't remember the TCG spec on top of my head.
> Can't we just move the measurement before deleting the handle?

I think this can work indeed, and it is more elegant; I will send a v2, thanks!

If I read the specification[1] correctly, the measurements sequence in 8.2.4 is
the following:

  f. Execute UEFI application
     i. If additional loads
          measure
  g. Measure EV_EFI_ACTION "Returning ..."
  h. If select next
       goto a
  i. If ExitBootServices
       measure EV_EFI_ACTION "Exit Boot Services ..."

Everything efi_exit() is doing is happening in step "g", therefore moving code
around efi_delete_image() should have no effect on the measurements.

Best regards,
Vincent.

[1] https://trustedcomputinggroup.org/wp-content/uploads/PC-Client-Platform-Firmware-Profile-Version-1.06-Revision-52_pub.pdf

> 
> Thanks
> /Ilias
> >                         ret = efi_tcg2_measure_efi_app_exit();
> >                         if (ret != EFI_SUCCESS)
> >                                 log_debug("tcg2 measurement fails (0x%lx)\n",
> > --
> > 2.51.0
> >