Re: [PATCH v2] iommu/rockchip: Add missing set_platform_dma_ops callback

[Steven Price <steven.price@arm.com>]

On 24/03/2023 13:16, Baolu Lu wrote:
> On 2023/3/24 19:11, Steven Price wrote:
>> Similar to exynos, we need a set_platform_dma_ops() callback for proper
>> operation on ARM 32 bit after recent changes in the IOMMU framework
>> (detach ops removal). But also the use of a NULL domain is confusing.
>>
>> Rework the code to have a singleton rk_identity_domain which is assigned
>> to domain when using an identity mapping rather than "detaching". This
>> makes the code easier to reason about.
>>
>> Signed-off-by: Steven Price <steven.price@arm.com>
>> ---
>> Changes since v1[1]:
>>
>>   * Reworked the code to avoid a NULL domain, instead a singleton
>>     rk_identity_domain is used instead. The 'detach' language is no
>>     longer used.
>>
>> [1]
>> https://lore.kernel.org/r/20230315164152.333251-1-steven.price%40arm.com
>>
>>   drivers/iommu/rockchip-iommu.c | 50 ++++++++++++++++++++++++++--------
>>   1 file changed, 39 insertions(+), 11 deletions(-)
>>
>> diff --git a/drivers/iommu/rockchip-iommu.c
>> b/drivers/iommu/rockchip-iommu.c
>> index f30db22ea5d7..437541004994 100644
>> --- a/drivers/iommu/rockchip-iommu.c
>> +++ b/drivers/iommu/rockchip-iommu.c
>> @@ -124,6 +124,7 @@ struct rk_iommudata {
>>     static struct device *dma_dev;
>>   static const struct rk_iommu_ops *rk_ops;
>> +static struct iommu_domain rk_identity_domain;
>>     static inline void rk_table_flush(struct rk_iommu_domain *dom,
>> dma_addr_t dma,
>>                     unsigned int count)
>> @@ -980,26 +981,27 @@ static int rk_iommu_enable(struct rk_iommu *iommu)
>>       return ret;
>>   }
>>   -static void rk_iommu_detach_device(struct iommu_domain *domain,
>> -                   struct device *dev)
>> +static int rk_iommu_identity_attach(struct iommu_domain
>> *identity_domain,
>> +                    struct device *dev)
>>   {
>>       struct rk_iommu *iommu;
>> -    struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
>> +    struct rk_iommu_domain *rk_domain;
>>       unsigned long flags;
>>       int ret;
>>         /* Allow 'virtual devices' (eg drm) to detach from domain */
>>       iommu = rk_iommu_from_dev(dev);
>>       if (!iommu)
>> -        return;
>> +        return -ENODEV;
>> +
>> +    rk_domain = to_rk_domain(iommu->domain);
>>         dev_dbg(dev, "Detaching from iommu domain\n");
>>   -    /* iommu already detached */
>> -    if (iommu->domain != domain)
>> -        return;
>> +    if (iommu->domain == identity_domain)
>> +        return 0;
>>   -    iommu->domain = NULL;
>> +    iommu->domain = identity_domain;
> 
> Where did identity_domain come from? Is it rk_identity_domain?

It's a parameter of the function. In the case of the call in
rk_iommu_attach_device() then, yes, it's rk_identity_domain. But this
function is also the "attach_dev" callback of "rk_identity_ops".

I'll admit this is cargo-culted from Jason's example:

https://lore.kernel.org/linux-iommu/ZBnef7g7GCxogPNz@ziepe.ca/

Steve

> Best regards,
> baolu