[PATCH v2 5/7] arm: efi: split zImage code and data into separate PE/COFF sections

[Ard Biesheuvel]

On 8 September 2017 at 15:57, Ard Biesheuvel <ard.biesheuvel at linaro.org> wrote:
> On 8 September 2017 at 15:56, Gregory CLEMENT
> <gregory.clement at free-electrons.com> wrote:
>> Hi Ard,
>>
>>  On ven., sept. 08 2017, Ard Biesheuvel <ard.biesheuvel at linaro.org> wrote:
>>
>>> On 8 September 2017 at 15:33, Gregory CLEMENT
>>> <gregory.clement at free-electrons.com> wrote:
>>>> Hi Ard,
>>>>
>>>>  On ven., sept. 08 2017, Ard Biesheuvel <ard.biesheuvel at linaro.org> wrote:
>>>>
>>>>> On 8 September 2017 at 14:54, Ard Biesheuvel <ard.biesheuvel at linaro.org> wrote:
>>>>>> On 8 September 2017 at 14:50, Gregory CLEMENT
>>>>>> <gregory.clement at free-electrons.com> wrote:
>>>>>>> Hi Ard,
>>>>>>>
>>>>>>>  On jeu., juin 29 2017, Ard Biesheuvel <ard.biesheuvel at linaro.org> wrote:
>>>>>>>
>>>>>>>> To prevent unintended modifications to the kernel text (malicious or
>>>>>>>> otherwise) while running the EFI stub, describe the kernel image as
>>>>>>>> two separate sections: a .text section with read-execute permissions,
>>>>>>>> covering .text, .rodata, .piggytext and the GOT sections (which the
>>>>>>>> stub does not care about anyway), and a .data section with read-write
>>>>>>>> permissions, covering .data and .bss.
>>>>>>>>
>>>>>>>> This relies on the firmware to actually take the section permission
>>>>>>>> flags into account, but this is something that is currently being
>>>>>>>> implemented in EDK2, which means we will likely start seeing it in
>>>>>>>> the wild between one and two years from now.
>>>>>>>
>>>>>>> This patch had been merged in mainline yesterday and now prevent the
>>>>>>> Marvell Armada 370 and the Armada XP based SoC to boot. I also suspect
>>>>>>> that more Socs are impacted because the number of boot fail exploded
>>>>>>> according to kci:
>>>>>>> https://kernelci.org/boot/all/job/mainline/branch/master/kernel/v4.13-8899-g8dc5b3a6cb2f/
>>>>>>>
>>>>>>
>>>>>> Ouch.
>>>>>>
>>>>>>> I found this patch after bisecting (I can provide the bisect log if
>>>>>>> needed).
>>>>>>>
>>>>>>> The kernel failed to boot only if CONFIG_EFI is enabled so it occurs in
>>>>>>> multi_v7_defconfig but not with mvebu_v7_defconfig.
>>>>>>>
>>>>>>> Currently the solution is to revert this patch.
>>>>>>>
>>>>>>> Have you a better option?
>>>>>>>
>>>>>>
>>>>>> I will investigate.
>>>>>
>>>>> I cannot reproduce this on QEMU or my Beaglebone white. I have tried a
>>>>> locally built zImage as well as the one built by kernelci.
>>>>>
>>>>> Could you please try whether this fixes things? It does not explain
>>>>> anything but it will help me figure out what is going on (hopefully)
>>>>
>>>> I've just tested this change and it didn't fix anything.
>>>>
>>>> Gregory
>>>>
>>>>>
>>>>>
>>>>> diff --git a/arch/arm/boot/compressed/efi-header.S
>>>>> b/arch/arm/boot/compressed/efi-header.S
>>>>> index c94a88ae834d..671a6e5b7b99 100644
>>>>> --- a/arch/arm/boot/compressed/efi-header.S
>>>>> +++ b/arch/arm/boot/compressed/efi-header.S
>>>>> @@ -127,7 +127,7 @@ section_table:
>>>>>
>>>>>                 .set    section_count, (. - section_table) / 40
>>>>>
>>>>> -               .align  12
>>>>> +               .align  9
>>>>>  __efi_start:
>>>>>  #endif
>>>>>                 .endm
>>>>
>>>
>>>
>>> How about this?
>>
>> It fixed the bug! (I tested with and without your previous patch and it
>> worked in both case)
>>
>> When you will send your patch, you can add my:
>> Tested-by: Gregory CLEMENT <gregory.clement at free-electrons.com>
>>

Would you mind checking whether this fixes the issue as well?

diff --git a/arch/arm/boot/compressed/piggy.S b/arch/arm/boot/compressed/piggy.S
index f72088495f43..5d52c556dd32 100644
--- a/arch/arm/boot/compressed/piggy.S
+++ b/arch/arm/boot/compressed/piggy.S
@@ -1,5 +1,6 @@
        .section .piggydata,#alloc
        .globl  input_data
+       .align  2
 input_data:
        .incbin "arch/arm/boot/compressed/piggy_data"
        .globl  input_data_end

There may be other reasons than ksymtab entries that could result in
piggydata ending up unaligned in the decompressor (which is what
caused the issue before)
This is a better fix, because it addresses the root cause.