Re: [PATCH 2/3] iommu/hisilicon: Add hi6220 iommu driver

[chenfeng <puck.chen@hisilicon.com>]

On 2015/10/14 20:18, Joerg Roedel wrote:
> On Thu, Oct 08, 2015 at 03:45:47PM +0800, Chen Feng wrote:
>> +static int hi6220_smmu_attach_dev(struct iommu_domain *domain,
>> +				  struct device *dev)
>> +{
>> +	struct hi6220_domain *m_domain = to_hi6220_domain(domain);
>> +
>> +	smmu_domain_prepare(m_domain);
>> +	dev->archdata.iommu = &m_domain->smmu_dev->iova_allocator;
>> +
>> +	return 0;
>> +}
> 
> What happens when you attach devices behind different smmus to one
> domain? Will that overwrite the smmu_dev pointer in the domain?
> 

Since the smmu master use the same pagetable, the master can use the iova_allocator
in smmu driver to allocate iova address.

In this way, the different master can get the different io address.

>> +static size_t hi6220_smmu_unmap(struct iommu_domain *domain, unsigned long iova,
>> +				size_t size)
>> +{
>> +	struct hi6220_domain *m_domain = to_hi6220_domain(domain);
>> +	size_t page_size = m_domain->smmu_dev->page_size;
>> +	struct hi6220_smmu *smmu_dev = m_domain->smmu_dev;
>> +	int *page_table = (unsigned int *)smmu_dev->pgtable_virt;
>> +
>> +	if (size != page_size) {
>> +		pr_err("unmap size error, only support %zd\n", page_size);
>> +		return 0;
>> +	}
>> +
>> +	__clear_smmu_pte(page_table + IOVA_PFN(iova));
>> +
>> +	return page_size;
>> +}
> 
> Don't you need a call to __invalid_smmu_tlb here too?

Thanks, I will fix this next version.
> 
> 
> 
> 	Joerg
> 
> 
> .
>