Re: [PATCH 3/5] ARM: trusted_foundations: do not use naked function

[Stephen Warren <swarren@wwwdotorg.org>]

On 03/21/2018 09:26 AM, Dmitry Osipenko wrote:
> On 21.03.2018 17:09, Stefan Agner wrote:
>> On 21.03.2018 13:13, Robin Murphy wrote:
>>> On 20/03/18 23:02, Stefan Agner wrote:
>>>> As documented in GCC naked functions should only use Basic asm
>>>> syntax. The Extended asm or mixture of Basic asm and "C" code is
>>>> not guaranteed. Currently this works because it was hard coded
>>>> to follow and check GCC behavior for arguments and register
>>>> placement.
>>>>
>>>> Furthermore with clang using parameters in Extended asm in a
>>>> naked function is not supported:
>>>>     arch/arm/firmware/trusted_foundations.c:47:10: error: parameter
>>>>             references not allowed in naked functions
>>>>                   : "r" (type), "r" (arg1), "r" (arg2)
>>>>                          ^
>>>>
>>>> Use a regular function to be more portable. This aligns also with
>>>> the other smc call implementations e.g. in qcom_scm-32.c and
>>>> bcm_kona_smc.c.
>>>>
>>>> Additionally also make sure all callee-saved registers get saved
>>>> as it has been done before.
>>>>
>>>> Signed-off-by: Stefan Agner <stefan@agner.ch>
>>>> ---
>>>>    arch/arm/firmware/trusted_foundations.c | 12 +++++++-----
>>>>    1 file changed, 7 insertions(+), 5 deletions(-)
>>>>
>>>> diff --git a/arch/arm/firmware/trusted_foundations.c b/arch/arm/firmware/trusted_foundations.c
>>>> index 3fb1b5a1dce9..426d732e6591 100644
>>>> --- a/arch/arm/firmware/trusted_foundations.c
>>>> +++ b/arch/arm/firmware/trusted_foundations.c
>>>> @@ -31,21 +31,23 @@
>>>>      static unsigned long cpu_boot_addr;
>>>>    -static void __naked tf_generic_smc(u32 type, u32 arg1, u32 arg2)
>>>> +static void tf_generic_smc(u32 type, u32 arg1, u32 arg2)
>>>>    {
>>>> +	register u32 r0 asm("r0") = type;
>>>> +	register u32 r1 asm("r1") = arg1;
>>>> +	register u32 r2 asm("r2") = arg2;
>>>> +
>>>>    	asm volatile(
>>>>    		".arch_extension	sec\n\t"
>>>> -		"stmfd	sp!, {r4 - r11, lr}\n\t"
>>>>    		__asmeq("%0", "r0")
>>>>    		__asmeq("%1", "r1")
>>>>    		__asmeq("%2", "r2")
>>>>    		"mov	r3, #0\n\t"
>>>>    		"mov	r4, #0\n\t"
>>>>    		"smc	#0\n\t"
>>>> -		"ldmfd	sp!, {r4 - r11, pc}"
>>>>    		:
>>>> -		: "r" (type), "r" (arg1), "r" (arg2)
>>>> -		: "memory");
>>>> +		: "r" (r0), "r" (r1), "r" (r2)
>>>> +		: "memory", "r3", "r4", "r5", "r6", "r7", "r8", "r9", "r10");
>>>
>>> I may be missing a subtlety, but it looks like we no longer have a
>>> guarantee that r11 will be caller-saved as it was previously. I don't
>>> know the Trusted Foundations ABI to say whether that matters or not,
>>> but if it is the case that it never needed preserving anyway, that
>>> might be worth calling out in the commit message.
>>
>> Adding r11 (fp) to the clobber list causes an error when using gcc and
>> CONFIG_FRAME_POINTER=y:
>> arch/arm/firmware/trusted_foundations.c: In function ‘tf_generic_smc’:
>> arch/arm/firmware/trusted_foundations.c:51:1: error: fp cannot be used
>> in asm here
>>
>> Not sure what ABI Trusted Foundations follow.
>>
>> [adding Stephen, Thierry and Dmitry]
>> Maybe someone more familiar with NVIDIA Tegra SoCs can help?
>>
>> When CONFIG_FRAME_POINTER=y fp gets saved anyway. So we could add r11 to
>> clobber list ifndef CONFIG_FRAME_POINTER...
> 
> I have no idea about TF ABI either. Looking at the downstream kernel code, r4 -
> r12 should be saved. I've CC'd Alexandre as he is the author of the original
> patch and may still remember the details.
> 
> I'm also wondering why original code doesn't have r3 in the clobber list and why
> r3 is set to '0', downstream sets it to the address of SP and on return from SMC
> r3 contains the address of SP which should be restored. I'm now wondering how
> SMC calling worked for me at all on T30, maybe it didn't..

I don't know what the ABI for ATF is. I assume it's documented in the 
ATF, PSCI, or similar specification, or ATF source code. Hence, I don't 
know whether ATF restores fp/r11.

My guess is that r3/r4 are set to 0 because they're defined as inputs by 
the SMC/ATF ABI, yet nothing the kernel does needed that many 
parameters, so they're hard-coded to 0 (to ensure they're set to 
something predictable) rather than also being parameters to 
tf_generic_smc().

The original code used to save/restore a lot of registers, including 
r11/fp. Can't we side-step the issue of including/not-including r11/fp 
in the clobber list by not removing those stmfd/ldmfd assembly instructions?