[PATCH] efi_loader: Disable devices before handing over control

[Ilias Apalodimas]

Hi Mark, 

On Wed, Oct 21, 2020 at 03:21:51PM +0200, Mark Kettenis wrote:
> > From: Ilias Apalodimas <ilias.apalodimas at linaro.org>
> > Date: Wed, 21 Oct 2020 15:42:02 +0300
> > 
> > Hi Heinrich,
> > 
> > 
> > On Wed, 21 Oct 2020 at 15:35, Heinrich Schuchardt <xypron.glpk at gmx.de> wrote:
> > >
> > > On 21.10.20 13:41, Ilias Apalodimas wrote:
> > > > Hi Heinrich,
> > > >
> > > > On Wed, Oct 21, 2020 at 12:17:29PM +0200, Heinrich Schuchardt wrote:
> > > >> On 10/21/20 9:32 AM, Ilias Apalodimas wrote:
> > > >>> U-Boot Driver Model is supposed to remove devices with either
> > > >>> DM_REMOVE_ACTIVE_DMA or DM_REMOVE_OS_PREPARE flags set, before exiting.
> > > >>> Our bootm command does that by explicitly calling calling
> > > >>> "dm_remove_devices_flags(DM_REMOVE_ACTIVE_ALL);" and we also disable any
> > > >>> USB devices.
> > > >>>
> > > >>> The EFI equivalent is doing none of those at the moment. As a result
> > > >>> probing an fTPM driver now renders it unusable in Linux. During our
> > > >>> (*probe) callback we open a session with OP-TEE, which is supposed to
> > > >>> close with our (*remove) callback. Since the (*remove) is never called,
> > > >>> once we boot into Linux and try to probe the device again we are getting
> > > >>> a busy error response. We also never free
> > > >>>
> > > >>> So let's fix this by mimicking what bootm does and disconnect devices
> > > >>> when efi_exit_boot_services() is called. Note that for the OP-TEE case
> > > >>> and in particular any subsequent bootloader that wants to use a device
> > > >>> (e.g GRUB) will need to call exit_boot_services() in order to close the
> > > >>> session.
> > > >>
> > > >> Hello Ilias,
> > > >>
> > > >> thanks for the patch. Adding the function calls looks correct to me,
> > > >
> > > > Well the only doubt I have is what if GRUB has to extend some PCRs before
> > > > calling Linux? Any idea if it's currently calling ExitBootSevices?
> > > > I was considering if it would be a better idea to call the device unbinding during
> > > > some kind of "exit" from U-boot's EFI code. (i.e before StartImage)
> > >
> > > ExitBootServices() is called by the Linux EFI stub in function
> > > allocate_new_fdt_and_exit_boot().
> > >
> > > If GRUB would call ExitBootServices(), it would not be able to launch
> > > the EFI stub via StartImage().
> > 
> > Yea that's my point. So with the current patch, you won't be able to
> > access the fTPM driver from GRUB
> > (or any other EFI application) until the Linux EFI stub calls exit
> > boot services. Maybe calling those 2 functions in
> > StartImage is a better idea?
> 
> Shouldn't an EFI application (such as GRUB) be using EFI protocols to
> access the TPM?

Yea I already responded to myself on the previous mail. We don't have support
for the EFI protocol in U-Boot (yet), but cleaning up in exit boot services makes
more sense. So I'll fix the compilation error and send a V2 with the DM removal 
as-is. GRUB will be able to access the TPM once we add the EFI protocols in U-boot.

Cheers
/Ilias