Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Date:   Thu, 22 Mar 2018 13:43:02 +0100
From:   Stefan Agner <stefan@agner.ch>
To:     Robin Murphy <robin.murphy@arm.com>
Cc:     Stephen Warren <swarren@wwwdotorg.org>,
        Dmitry Osipenko <digetx@gmail.com>, swarren@nvidia.com,
        thierry.reding@gmail.com,
        Alexandre Courbot <acourbot@chromium.org>,
        nicolas.pitre@linaro.org, keescook@chromium.org, arnd@arndb.de,
        ard.biesheuvel@linaro.org, marc.zyngier@arm.com,
        linux@armlinux.org.uk, linux-kernel@vger.kernel.org,
        mka@chromium.org, linux-arm-kernel@lists.infradead.org,
        Bernhard.Rosenkranzer@linaro.org
Subject: Re: [PATCH 3/5] ARM: trusted_foundations: do not use naked function
In-Reply-To: <bd09fb3a-b8d5-995c-3a0c-271468f51d4e@arm.com>
References: <20180320230206.25289-1-stefan@agner.ch>
 <20180320230206.25289-4-stefan@agner.ch>
 <124e16c9-b8ca-9c7e-ade5-b033eed76e14@arm.com>
 <302cfd1a4324e064cd4f189e4e0ffc21@agner.ch>
 <03de5241-fb69-9097-5020-f9843482318d@gmail.com>
 <1dd52edb-412c-2d26-7e6a-d567695a89fe@wwwdotorg.org>
 <d8b3053d-eeae-4dbc-75e5-b2e07ea58f3d@arm.com>
 <d8adef8767543f8cd0d88546a1940a66@agner.ch>
 <bd09fb3a-b8d5-995c-3a0c-271468f51d4e@arm.com>
Message-ID: <8158625159419db87906989ac541c7d3@agner.ch>
User-Agent: Roundcube Webmail/1.3.4
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On 22.03.2018 12:48, Robin Murphy wrote:
> On 21/03/18 21:41, Stefan Agner wrote:
>> On 21.03.2018 18:16, Robin Murphy wrote:
>>> On 21/03/18 16:40, Stephen Warren wrote:
>>>> 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.
>>>
>>> Oops, I think we're starting to diverge here - "ATF" (as in "Arm
>>> Trusted Firmware") does implement the ARM SMCCC, which more or less
>>> just follows the regular procedure call standard in terms of register
>>> saving. The "TF" in question here is "Trusted Foundations" from
>>> Trusted Logic (who apparently don't exist any more) which is
>>> explicitly called out in the header as having its own nonstandard
>>> calling convention. I guess newer Tegras are using the former, whereas
>>> the older ones used the latter.
>>>
>>
>> What do you mean by "called out in the header as having its own
>> nonstandard"?
> 
> Specifically, the comment in
> arch/arm/include/asm/trusted_foundations.h which says:
> 
>   "The calls are completely specific to Trusted Foundations, and do
>    *not* follow the SMC calling convention or the PSCI standard."
> 

Oh didn't notice that. Thanks for pointing out.

>> It is unclear what ABI is used, I just inferred from the fact that
>> register have been saved before that it might use a nonstandard calling
>> convention.
>>
>> Tegra 4i/TK1 and newer seem to use something called Trusted Little
>> Kernel.
>>
>>>> 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?
>>>
>>> That might be reasonable - fiddling with a C function's stack inside
>>> an asm is a bit grim, but for this case I can't see that it would mess
>>> with unwinding etc. or otherwise go wrong any more than the existing
>>> code, and I doubt the slight efficiency hit from having to change the
>>> "pop the LR straight into the PC" idiom matters much.
>>
>> Sounds reasonable, I guess in that case we can also omit all the
>> additional register in the clobber list.
> 
> Yeah, you should only need to specify clobbers for any registers which
> are neither used as arguments nor explicitly preserved - looking at
> the layout of the code, it seems unlikely that the compiler would have
> anything live in r3 or r12 across the call (since the scope for
> inlining is pretty trivial), but there's no harm in being strictly
> correct :)

So something like this?

-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");
 }


--
Stefan