Re: [PATCH v2 2/6] iommufd: Add iommu page fault uapi data

[Baolu Lu <baolu.lu@linux.intel.com>]

On 12/1/23 11:14 PM, Jason Gunthorpe wrote:
> On Thu, Oct 26, 2023 at 10:49:26AM +0800, Lu Baolu wrote:
> 
>> + * @IOMMU_HWPT_ALLOC_IOPF_CAPABLE: User is capable of handling IO page faults.
> 
> This does not seem like the best name?
> 
> Probably like this given my remark in the cover letter:
> 
> --- a/include/uapi/linux/iommufd.h
> +++ b/include/uapi/linux/iommufd.h
> @@ -359,6 +359,7 @@ struct iommu_vfio_ioas {
>   enum iommufd_hwpt_alloc_flags {
>          IOMMU_HWPT_ALLOC_NEST_PARENT = 1 << 0,
>          IOMMU_HWPT_ALLOC_DIRTY_TRACKING = 1 << 1,
> +       IOMMU_HWPT_IOPFD_FD_VALID = 1 << 2,
>   };
>   
>   /**
> @@ -440,6 +441,7 @@ struct iommu_hwpt_alloc {
>          __u32 data_type;
>          __u32 data_len;
>          __aligned_u64 data_uptr;
> +       __s32 iopf_fd;
>   };
>   #define IOMMU_HWPT_ALLOC _IO(IOMMUFD_TYPE, IOMMUFD_CMD_HWPT_ALLOC)

Yes. Agreed.

>> @@ -679,6 +688,62 @@ struct iommu_dev_data_arm_smmuv3 {
>>   	__u32 sid;
>>   };
>>   
>> +/**
>> + * struct iommu_hwpt_pgfault - iommu page fault data
>> + * @size: sizeof(struct iommu_hwpt_pgfault)
>> + * @flags: Combination of IOMMU_PGFAULT_FLAGS_ flags.
>> + *  - PASID_VALID: @pasid field is valid
>> + *  - LAST_PAGE: the last page fault in a group
>> + *  - PRIV_DATA: @private_data field is valid
>> + *  - RESP_NEEDS_PASID: the page response must have the same
>> + *                      PASID value as the page request.
>> + * @dev_id: id of the originated device
>> + * @pasid: Process Address Space ID
>> + * @grpid: Page Request Group Index
>> + * @perm: requested page permissions (IOMMU_PGFAULT_PERM_* values)
>> + * @addr: page address
>> + * @private_data: device-specific private information
>> + */
>> +struct iommu_hwpt_pgfault {
>> +	__u32 size;
>> +	__u32 flags;
>> +#define IOMMU_PGFAULT_FLAGS_PASID_VALID		(1 << 0)
>> +#define IOMMU_PGFAULT_FLAGS_LAST_PAGE		(1 << 1)
>> +#define IOMMU_PGFAULT_FLAGS_PRIV_DATA		(1 << 2)
>> +#define IOMMU_PGFAULT_FLAGS_RESP_NEEDS_PASID	(1 << 3)
>> +	__u32 dev_id;
>> +	__u32 pasid;
>> +	__u32 grpid;
>> +	__u32 perm;
>> +#define IOMMU_PGFAULT_PERM_READ			(1 << 0)
>> +#define IOMMU_PGFAULT_PERM_WRITE		(1 << 1)
>> +#define IOMMU_PGFAULT_PERM_EXEC			(1 << 2)
>> +#define IOMMU_PGFAULT_PERM_PRIV			(1 << 3)
>> +	__u64 addr;
>> +	__u64 private_data[2];
>> +};
> 
> This mixed #define is not the style, these should be in enums,
> possibly with kdocs
> 
> Use __aligned_u64 also

Sure.

> 
>> +
>> +/**
>> + * struct iommu_hwpt_response - IOMMU page fault response
>> + * @size: sizeof(struct iommu_hwpt_response)
>> + * @flags: Must be set to 0
>> + * @hwpt_id: hwpt ID of target hardware page table for the response
>> + * @dev_id: device ID of target device for the response
>> + * @pasid: Process Address Space ID
>> + * @grpid: Page Request Group Index
>> + * @code: response code. The supported codes include:
>> + *        0: Successful; 1: Response Failure; 2: Invalid Request.
>> + */
>> +struct iommu_hwpt_page_response {
>> +	__u32 size;
>> +	__u32 flags;
>> +	__u32 hwpt_id;
>> +	__u32 dev_id;
>> +	__u32 pasid;
>> +	__u32 grpid;
>> +	__u32 code;
>> +};
> 
> Is it OK to have the user pass in all this detailed information? Is it
> a security problem if the user lies? Ie shouldn't we only ack page
> faults we actually have outstanding?
> 
> IOW should iommu_hwpt_pgfault just have a 'response_cookie' generated
> by the kernel that should be placed here? The kernel would keep track
> of all this internal stuff?

The iommu core has already kept the outstanding faults that have been
awaiting a response. So even if the user lies about a fault, the kernel
does not send the wrong respond message to the device. {device_id,
grpid, code} is just enough from the user. This means the user wants to
respond to the @grpid fault from @device with the @code result.

Best regards,
baolu