[U-Boot] [PATCH] efi_loader: Patch non-runtime code out at ExitBootServices already

[Alexander Graf]

On 26.01.19 10:31, Heinrich Schuchardt wrote:
> On 1/25/19 10:43 PM, Ard Biesheuvel wrote:
>> On Fri, 25 Jan 2019 at 19:21, Heinrich Schuchardt <xypron.glpk at gmx.de> wrote:
>>>
>>> On 1/23/19 6:33 PM, Alexander Graf wrote:
>>>> While discussing something compeltely different, Ard pointed out
>>>> that it might be legal to omit calling SetVirtualAddressMap altogether.
>>>>
>>>> While that sounds great, we currently rely on that call to remove
>>>> all function pointers to code that we do not support outside of
>>>> boot services.
>>>>
>>>> So let's patch out those bits already on the call to ExitBootServices,
>>>> so that we can successfully run even when an OS chooses to omit
>>>> any call to SetVirtualAddressMap.
>>>
>>> Such an operating system would not be allowed to use any virtual
>>> addresses at any time because runtime drivers would not be informed
>>> about the mapping.
>>>
>>
>> No. The OS would be permitted to invoke the runtime services at their
>> original offset.
>>
>> For arm64, this is trivially achievable, since we already use userland
>> mappings for the runtime services. On 32-bit architectures, it is more
>> complicated, since the boot time mapping of peripherals and/or memory
>> may conflict with the kernel's layout of the address space (e.g., if
>> your NOR flash lives above 0xc0000000, you *have* to remap it for the
>> runtime services to be able to use it).
>>
>>
>>
>>> Does such an operating system exist?
>>> Or is this only a hypothetical case?
>>>
>>>>
>>>> Reported-by: Ard Biesheuvel <ard.biesheuvel at linaro.org>
>>>> Signed-off-by: Alexander Graf <agraf at suse.de>
>>>
>>> I am missing a description of the side effects of the change, e.g.
>>>
>>> Our EFI unit tests call the Reset() service.
>>>
>>> So Python tests will fail on systems that do not implement Reset() in a
>>> runtime compatible manner.
>>>
>>
>> All runtime services (except SetVirtualAddresMap() and
>> ConvertPointer(), obviously) may be used at runtime with or without
>> installing a virtual address mapping.
>>
> 
> With this patch in place Travis CI testing fails:
> 
> After 'bootefi selftest' this output is not ejected:
> m = u_boot_console.p.expect(['resetting', 'U-Boot'])
> 
> And after the reset U-Boot crashes (observed for qemu-arm64_defconfig).

I just tried this out on the x86_64 variant. There reset does not occur,
because at the point in time when we trigger the reset,
ExitBootServices() has already been called. I would call that
intentional - we patch the reset function away now because x86 doesn't
provide an explicit efi enabled reset function.

The arm breakage is related to a missing icache flush. I'll fix that one up.


Thanks,

Alex