Re: [PATCH] iommu/rockchip: Add missing set_platform_dma_ops callback

[Steven Price]

On 22/03/2023 17:36, Jason Gunthorpe wrote:
> On Wed, Mar 22, 2023 at 04:04:25PM +0000, Steven Price wrote:
>> On 22/03/2023 15:16, Jason Gunthorpe wrote:
>>> On Wed, Mar 22, 2023 at 03:08:41PM +0000, Steven Price wrote:
>>>> @@ -1035,8 +1055,9 @@ static int rk_iommu_attach_device(struct iommu_domain *domain,
>>>>  	if (iommu->domain == domain)
>>>>  		return 0;
>>>>  
>>>> -	if (iommu->domain)
>>>> -		rk_iommu_detach_device(iommu->domain, dev);
>>>> +	ret = rk_iommu_identity_attach(&rk_identity_domain, dev);
>>>> +	if (ret)
>>>> +		return ret;
>>>
>>>>  
>>>>  	iommu->domain = domain;
>>>>  
>>>> @@ -1049,8 +1070,6 @@ static int rk_iommu_attach_device(struct iommu_domain *domain,
>>>>  		return 0;
>>>>  
>>>>  	ret = rk_iommu_enable(iommu);
>>>> -	if (ret)
>>>> -		rk_iommu_detach_device(iommu->domain, dev);
>>>
>>> I think this still needs error handling, it should put it back to the
>>> identity domain and return an error code if it fails to attach to the
>>> requested domain.
>>
>> What confused me here is that there's already a call to
>> rk_iommu_identity_attach() just above. But I can obviously add a...
> 
> I don't know this driver at all, but to me it looks like this is
> perhaps undoing a partially failed rk_iommu_enable() since it doesn't
> seem to enetirely fix itself. Ie it zeros the INT_MASK and DTE_ADDR
> 
> Maybe it would be better to put that error cleanup direclty into
> enable and just move the iommu->domain assignment to after enable
> success.

While I agree this would be better - I don't feel I understand the
driver enough to have confidence in doing this. And I don't know how to
trigger the error conditions to test this either.

>>        if (ret)
>>                rk_iommu_identity_attach(&rk_identity_domain, dev);
>>
>> ... in here. But I don't know how to handle an error from
>> rk_iommu_identity_attach() at this point. Does it need handling - is a
>> WARN_ON sufficient?
> 
> WARN_ON should be fine, that is kind of hacky, it would be better to
> organize things so there is an identity attach function that cannot
> fail, ie pre-assumes all the validation is done alread.y

As the code currently stands rk_iommu_identity_attach can fail for
exactly one reason: if rk_iommu_from_dev() fails. And since that check
is already done in rk_iommu_attach_device() this cannot fail (baring
memory corruption etc). So I'll stick to WARN_ON for now.

>>
>>> It should also initlaize iommu->domain to the identity domain when the
>>> iommu struct is allocated. The iommu->domain should never be
>>> NULL. identity domain means the IOMMU is turned off which was
>>> previously called "detached".
>>
>> I presume you mean in rk_iommu_probe()?
> 
> It would be best if it was setup at allocation time so in
> rk_iommu_of_xlate() before dev_iommu_priv_set()

I've already put an assignment in rk_iommu_of_xlate() just before
dev_iommu_priv_set().

Steve