[PATCH 1/6] iommu/qcom: Use the asid read from device-tree if specified

[AngeloGioacchino Del Regno]

Il 03/06/22 20:03, Konrad Dybcio ha scritto:
> 
> 
> On 31.05.2022 22:57, Rob Clark wrote:
>> On Tue, May 31, 2022 at 9:19 AM Will Deacon <will at kernel.org> wrote:
>>>
>>> On Tue, May 31, 2022 at 09:15:22AM -0700, Rob Clark wrote:
>>>> On Tue, May 31, 2022 at 8:46 AM Will Deacon <will at kernel.org> wrote:
>>>>>
>>>>> On Fri, May 27, 2022 at 11:28:56PM +0200, Konrad Dybcio wrote:
>>>>>> From: AngeloGioacchino Del Regno <angelogioacchino.delregno at somainline.org>
>>>>>>
>>>>>> As specified in this driver, the context banks are 0x1000 apart.
>>>>>> Problem is that sometimes the context number (our asid) does not
>>>>>> match this logic and we end up using the wrong one: this starts
>>>>>> being a problem in the case that we need to send TZ commands
>>>>>> to do anything on a specific context.
>>>>>
>>>>> I don't understand this. The ASID is a software construct, so it shouldn't
>>>>> matter what we use. If it does matter, then please can you explain why? The
>>>>> fact that the context banks are 0x1000 apart seems unrelated.
>>>>
>>>> I think the connection is that mapping from ctx bank to ASID is 1:1
>>>
>>> But in what sense? How is the ASID used beyond a tag in the TLB? The commit
>>> message hints at "TZ commands" being a problem.
>>>
>>> I'm not doubting that this is needed to make the thing work, I just don't
>>> understand why.
>>
>> (disclaimer, it has been quite a while since I've looked at the smmu
>> setup with earlier tz, ie. things that use qcom_iommu, but from
>> memory...)
>>
>> We cannot actually assign the context banks ourselves, so in the dt
>> bindings the "ASID" is actually the context bank index.
> I think so.
> 
>    I don't
>> remember exactly if this was a limitation of the tz interface, or
>> result of not being able to program the smmu's global registers
>> ourselves.
> 
> As far as I understand, it's the latter, as changing the defaults is not allowed by the security policy on consumer devices.
> 
> Qualcomm arbitrarily chose some numbers that may or may have not aligned with their usual index-is-offset-divided-by-0x1000 and hardcoded them in the BSP, and now the secure side (if required, and well, it is..) expects precisely that configuration.
> 
> 
> Konrad
> 

I can confirm that it's the latter, as described by Konrad.
The inability of programming the global registers from Linux is due to the
hypervisor disallowing that (in different ways depending on the SoC's firmware
but with the same outcome: AP reset by HYP).

Cheers,
Angelo