Re: [PATCH v5 14/29] iommufd/viommu: Add IOMMUFD_CMD_HW_QUEUE_ALLOC ioctl

[Nicolin Chen <nicolinc@nvidia.com>]

On Fri, May 30, 2025 at 01:14:55PM -0300, Jason Gunthorpe wrote:
> Put a size_t in the driver ops:
> 
>  size_t size_viommu;
>  size_t size_hw_queue;
> 
> Have the driver set it via a macro like INIT_RDMA_OBJ_SIZE
> 
> #define INIT_RDMA_OBJ_SIZE(ib_struct, drv_struct, member)                      \
> 	.size_##ib_struct =                                                    \
> 		(sizeof(struct drv_struct) +                                   \
> 		 BUILD_BUG_ON_ZERO(offsetof(struct drv_struct, member)) +      \
> 		 BUILD_BUG_ON_ZERO(                                            \
> 			 !__same_type(((struct drv_struct *)NULL)->member,     \
> 				      struct ib_struct)))
> 
> Which proves the core structure is at the front.
> 
> Then the core code can allocate the object along with enough space for
> the driver and call a driver function to init the driver portion of
> the already allocated object.

I found that the size_viommu or size_hw_queue might not work using
a static macro as that RDMA one does:

 - The size in vIOMMU case is type dependent. E.g. smmuv3 driver
   uses one iommu_ops to support two types: vSMMU and vCMDQ

 - Changing to a type-indexed size array would eventually result
   some driver having a big size array, as the type number grows

I came up with two alternatives:

1) Define a get_viommu_size(unsigned int type) op: use a similar
   macro in the driver function to return with:

#define VIOMMU_STRUCT_SIZE(ib_struct, drv_struct, member)                      \
	    (sizeof(drv_struct) +                                              \
 		 BUILD_BUG_ON_ZERO(offsetof(drv_struct, member)) +             \
 		 BUILD_BUG_ON_ZERO(!__same_type(((drv_struct *)NULL)->member,  \
 						ib_struct)))

	if (type == SMMU)
		return VIOMMU_STRUCT_SIZE(
			struct arm_vsmmu, struct iommufd_viommu, core);
	return 0;

2) Let core allocate with sizeof(struct iommufd_viommu), then let
   driver krealloc during the viommu_init op call:

	viommu = kzalloc(sizeof(struct iommufd_viommu), ...);
	...
	viommu = ops->viommu_init(viommu, dev, parent_dom, type);

I am guessing that you may prefer 1 over 2? Or any better idea?

Thanks
Nicolin