Re: [RFC 00/18] vfio: Adopt iommufd

[Eric Auger]

Hi Shameer,

On 6/28/22 10:14, Shameerali Kolothum Thodi wrote:
>
>> -----Original Message-----
>> From: Yi Liu [mailto:yi.l.liu@intel.com]
>> Sent: 18 May 2022 15:01
>> To: zhangfei.gao@foxmail.com; Jason Gunthorpe <jgg@nvidia.com>;
>> Zhangfei Gao <zhangfei.gao@linaro.org>
>> Cc: eric.auger@redhat.com; Alex Williamson <alex.williamson@redhat.com>;
>> Shameerali Kolothum Thodi <shameerali.kolothum.thodi@huawei.com>;
>> cohuck@redhat.com; qemu-devel@nongnu.org;
>> david@gibson.dropbear.id.au; thuth@redhat.com; farman@linux.ibm.com;
>> mjrosato@linux.ibm.com; akrowiak@linux.ibm.com; pasic@linux.ibm.com;
>> jjherne@linux.ibm.com; jasowang@redhat.com; kvm@vger.kernel.org;
>> nicolinc@nvidia.com; eric.auger.pro@gmail.com; kevin.tian@intel.com;
>> chao.p.peng@intel.com; yi.y.sun@intel.com; peterx@redhat.com
>> Subject: Re: [RFC 00/18] vfio: Adopt iommufd
>>
>> On 2022/5/18 15:22, zhangfei.gao@foxmail.com wrote:
>>>
>>> On 2022/5/17 下午4:55, Yi Liu wrote:
>>>> Hi Zhangfei,
>>>>
>>>> On 2022/5/12 17:01, zhangfei.gao@foxmail.com wrote:
>>>>> Hi, Yi
>>>>>
>>>>> On 2022/5/11 下午10:17, zhangfei.gao@foxmail.com wrote:
>>>>>>
>>>>>> On 2022/5/10 下午10:08, Yi Liu wrote:
>>>>>>> On 2022/5/10 20:45, Jason Gunthorpe wrote:
>>>>>>>> On Tue, May 10, 2022 at 08:35:00PM +0800, Zhangfei Gao wrote:
>>>>>>>>> Thanks Yi and Eric,
>>>>>>>>> Then will wait for the updated iommufd kernel for the PCI MMIO
>> region.
>>>>>>>>> Another question,
>>>>>>>>> How to get the iommu_domain in the ioctl.
>>>>>>>> The ID of the iommu_domain (called the hwpt) it should be returned
>> by
>>>>>>>> the vfio attach ioctl.
>>>>>>> yes, hwpt_id is returned by the vfio attach ioctl and recorded in
>>>>>>> qemu. You can query page table related capabilities with this id.
>>>>>>>
>>>>>>>
>> https://lore.kernel.org/kvm/20220414104710.28534-16-yi.l.liu@intel.com/
>>>>>> Thanks Yi,
>>>>>>
>>>>>> Do we use iommufd_hw_pagetable_from_id in kernel?
>>>>>>
>>>>>> The qemu send hwpt_id via ioctl.
>>>>>> Currently VFIOIOMMUFDContainer has hwpt_list,
>>>>>> Which member is good to save hwpt_id, IOMMUTLBEntry?
>>>>> Can VFIOIOMMUFDContainer  have multi hwpt?
>>>> yes, it is possible
>>> Then how to get hwpt_id in map/unmap_notify(IOMMUNotifier *n,
>> IOMMUTLBEntry
>>> *iotlb)
>> in map/unmap, should use ioas_id instead of hwpt_id
>>
>>>>> Since VFIOIOMMUFDContainer has hwpt_list now.
>>>>> If so, how to get specific hwpt from map/unmap_notify in hw/vfio/as.c,
>>>>> where no vbasedev can be used for compare.
>>>>>
>>>>> I am testing with a workaround, adding VFIOIOASHwpt *hwpt in
>>>>> VFIOIOMMUFDContainer.
>>>>> And save hwpt when vfio_device_attach_container.
>>>>>
>>>>>> In kernel ioctl: iommufd_vfio_ioctl
>>>>>> @dev: Device to get an iommu_domain for
>>>>>> iommufd_hw_pagetable_from_id(struct iommufd_ctx *ictx, u32 pt_id,
>>>>>> struct device *dev)
>>>>>> But iommufd_vfio_ioctl seems no para dev?
>>>>> We can set dev=Null since IOMMUFD_OBJ_HW_PAGETABLE does not
>> need dev.
>>>>> iommufd_hw_pagetable_from_id(ictx, hwpt_id, NULL)
>>>> this is not good. dev is passed in to this function to allocate domain
>>>> and also check sw_msi things. If you pass in a NULL, it may even unable
>>>> to get a domain for the hwpt. It won't work I guess.
>>> The iommufd_hw_pagetable_from_id can be used for
>>> 1, allocate domain, which need para dev
>>> case IOMMUFD_OBJ_IOAS
>>> hwpt = iommufd_hw_pagetable_auto_get(ictx, ioas, dev);
>> this is used when attaching ioas.
>>
>>> 2. Just return allocated domain via hwpt_id, which does not need dev.
>>> case IOMMUFD_OBJ_HW_PAGETABLE:
>>> return container_of(obj, struct iommufd_hw_pagetable, obj);
>> yes, this would be the usage in nesting. you may check my below
>> branch. It's for nesting integration.
>>
>> https://github.com/luxis1999/iommufd/tree/iommufd-v5.18-rc4-nesting
>>
>>> By the way, any plan of the nested mode?
>> I'm working with Eric, Nic on it. Currently, I've got the above kernel
>> branch, QEMU side is also WIP.
> Hi Yi/Eric,
>
> I had a look at the above nesting kernel and Qemu branches and as mentioned
> in the cover letter it is not working on ARM yet.
>
> IIUC, to get it working via the iommufd the main thing is we need a way to 
> configure
> the phys SMMU in nested mode and setup the mappings for the stage 2. The
> Cache/PASID related changes looks more straight forward. 
>
> I had quite a few hacks to get it working on ARM, but still a WIP. So just 
> wondering
> do you guys have something that can be shared yet?

I am working on the respin based on latest iommufd kernel branches and
qemu RFC v2 but it is still WIP.

I will share as soon as possible.

Eric
>
> Please let me know.
>
> Thanks,
> Shameer