Re: 5.3 boot regression caused by 5.3 TPM changes

[Hans de Goede <hdegoede@redhat.com>]

Hi,

On 05-08-19 18:01, Ard Biesheuvel wrote:
> On Sun, 4 Aug 2019 at 19:12, Hans de Goede <hdegoede@redhat.com> wrote:
>>
>> Hi,
>>
>> On 04-08-19 17:33, Ard Biesheuvel wrote:
>>> Hi Hans,
>>>
>>> On Sun, 4 Aug 2019 at 13:00, Hans de Goede <hdegoede@redhat.com> wrote:
>>>>
>>>> Hi All,
>>>>
>>>> While testing 5.3-rc2 on an Irbis TW90 Intel Cherry Trail based
>>>> tablet I noticed that it does not boot on this device.
>>>>
>>>> A git bisect points to commit 166a2809d65b ("tpm: Don't duplicate
>>>> events from the final event log in the TCG2 log")
>>>>
>>>> And I can confirm that reverting just that single commit makes
>>>> the TW90 boot again.
>>>>
>>>> This machine uses AptIO firmware with base component versions
>>>> of: UEFI 2.4 PI 1.3. I've tried to reproduce the problem on
>>>> a Teclast X80 Pro which is also CHT based and also uses AptIO
>>>> firmware with the same base components. But it does not reproduce
>>>> there. Neither does the problem reproduce on a CHT tablet using
>>>> InsideH20 based firmware.
>>>>
>>>> Note that these devices have a software/firmware TPM-2.0
>>>> implementation, they do not have an actual TPM chip.
>>>>
>>>> Comparing TPM firmware setting between the 2 AptIO based
>>>> tablets the settings are identical, but the troublesome
>>>> TW90 does have some more setting then the X80, it has
>>>> the following settings which are not shown on the X80:
>>>>
>>>> Active PCR banks:           SHA-1         (read only)
>>>> Available PCR banks:        SHA-1,SHA256  (read only)
>>>> TPM2.0 UEFI SPEC version:   TCG_2         (other possible setting: TCG_1_2
>>>> Physical Presence SPEC ver: 1.2           (other possible setting: 1.3)
>>>>
>>>> I have the feeling that at least the first 2 indicate that
>>>> the previous win10 installation has actually used the
>>>> TPM, where as on the X80 the TPM is uninitialized.
>>>> Note this is just a hunch I could be completely wrong.
>>>>
>>>> I would be happy to run any commands to try and debug this
>>>> or to build a kernel with some patches to gather more info.
>>>>
>>>> Note any kernel patches to printk some debug stuff need
>>>> to be based on 5.3 with 166a2809d65b reverted, without that
>>>> reverted the device will not boot, and thus I cannot collect
>>>> logs without it reverted.
>>>>
>>>
>>> Are you booting a 64-bit kernel on 32-bit firmware?
>>
>> Yes you are right, I must say that this is somewhat surprising
>> most Cherry Trail devices do use 64 bit firmware (where as Bay Trail
>> typically uses 32 bit). But I just checked efibootmgr output and it
>> says it is booting: \EFI\FEDORA\SHIMIA32.EFI so yeah 32 bit firmware.
>>
>> Recent Fedora releases take care of this so seamlessly I did not
>> even realize...
>>
> 
> OK, so we'll have to find out how this patch affects 64-bit code
> running on 32-bit firmware.

I was not sure this really is a 32 bit firmware issue, as I believed
I saw 5.3 running fine on other 32 bit firmware devices, so I tried
this on another device with 32 bit UEFI and you're right this is a
32 bit issue.

> The only EFI call in that patch is
> get_config_table(), which is not actually a EFI boot service call but
> a EFI stub helper that parses the config table array in the EFI system
> table.

Well get_efi_config_table() is a new function in 5.3, introduced
specifically for the 166a2809d65b ("tpm: Don't duplicate events from the
final event log in the TCG2 log") commit.

It was introduced in commit 82d736ac56d7 ("Abstract out support for
locating an EFI config table") and after taking a good look at this
commit I'm pretty confident to say that the new get_efi_config_table()
function is the problem, as it is broken in multiple ways.

In itself the new get_efi_config_table() is just factoring out some
code used in drivers/firmware/efi/libstub/fdt.c into a new helper
for reuse and not making any functional changes to the factored out
code.

The problem is that the old code which it factors out contains a number
of assumptions which are true in the get_fdt() context from which it
was called but are not true when used in more generic code as is done
from the 166a2809d65b ("tpm: Don't duplicate events from the
final event log in the TCG2 log") commit.

There are 2 problems with the new get_efi_config_table() function:

1) sys_table->tables contains a physical address, we cannot just
cast that to a pointer and deref it, it needs to be early_memremap-ed
and then we deref the mapping. I'm somewhat amazed that this works
at all for the 64 bit case, but apparently it does.

2) sys_table->tables points to either an array of either
efi_config_table_64_t structd or an array of efi_config_table_32_t
structs.  efi_config_table_t is a generic type for storing information
when parsing it should NOT be used for reading the actual tables
as they come from the firmware when parsing! Now efi_config_table_t
happens to be an exact match for efi_config_table_64_t when building
an x86_64 kernel, so this happens to work for the 64 bit firmware case.

The properway to deal with this would be to use the existing
efi_config_parse_tables() functionality from drivers/firmware/efi/efi.c
by adding entry for the LINUX_EFI_TPM_FINAL_LOG_GUID to the
common_tables[] array in drivers/firmware/efi/efi.c and make that
entry store the table address (if found) in a new efi.final_log
member.

I'm not sure how important this functionality is to have in 5.3.

I will try to come up with a fix for this using efi_config_parse_tables()
but it might be better to just revert for 5.3 .

Regards,

Hans