Re: [PATCH 7/8] iommu/amd: Add support for nested domain allocation

[Nicolin Chen <nicolinc@nvidia.com>]

On Wed, Aug 20, 2025 at 11:30:08AM +0000, Suravee Suthikulpanit wrote:
> diff --git a/drivers/iommu/amd/iommu.c b/drivers/iommu/amd/iommu.c
> index 46682c8ba28d..ea790a8997ee 100644
> --- a/drivers/iommu/amd/iommu.c
> +++ b/drivers/iommu/amd/iommu.c
> @@ -2616,6 +2616,7 @@ amd_iommu_domain_alloc_paging_flags(struct device *dev, u32 flags,
>  				    const struct iommu_user_data *user_data)
>  
>  {
> +	struct iommu_domain *dom = ERR_PTR(-EOPNOTSUPP);
>  	struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev);
>  	struct amd_iommu *iommu = get_amd_iommu_from_dev(dev);
>  	const u32 supported_flags = IOMMU_HWPT_ALLOC_DIRTY_TRACKING |
> @@ -2626,29 +2627,31 @@ amd_iommu_domain_alloc_paging_flags(struct device *dev, u32 flags,
>  	if ((flags & ~supported_flags) || user_data || !is_nest_parent_supported(flags))
>  		return ERR_PTR(-EOPNOTSUPP);
>  
> -	pr_debug("%s: IOMMU devid=%#x, flags=%#x\n", __func__, dev_data->devid, flags);
> +	pr_debug("%s: IOMMU devid=%#x, flags=%#x, supported_flags=%#x\n", __func__, dev_data->devid, flags, supported_flags);
>  
>  	switch (flags & supported_flags) {
>  	case IOMMU_HWPT_ALLOC_DIRTY_TRACKING:
>  	case IOMMU_HWPT_ALLOC_DIRTY_TRACKING | IOMMU_HWPT_ALLOC_NEST_PARENT:
>  	case IOMMU_HWPT_ALLOC_NEST_PARENT:
>  		/* Allocate domain with v1 page table for dirty tracking */
> -		if (!amd_iommu_hd_support(iommu))
> -			break;
> -		return do_iommu_domain_alloc(dev, flags, PD_MODE_V1);
> +		if (amd_iommu_hd_support(iommu))
> +			dom = do_iommu_domain_alloc(dev, flags, PD_MODE_V1);
> +		break;
>  	case IOMMU_HWPT_ALLOC_PASID:
>  		/* Allocate domain with v2 page table if IOMMU supports PASID. */
> -		if (!amd_iommu_pasid_supported())
> -			break;
> -		return do_iommu_domain_alloc(dev, flags, PD_MODE_V2);
> +		if (amd_iommu_pasid_supported())
> +			dom = do_iommu_domain_alloc(dev, flags, PD_MODE_V2);
> +		break;
>  	case 0:
>  		/* If nothing specific is required use the kernel commandline default */
> -		return do_iommu_domain_alloc(dev, 0, amd_iommu_pgtable);
> +		dom = do_iommu_domain_alloc(dev, 0, amd_iommu_pgtable);
> +		break;
>  	default:
>  		pr_err("%s: Unhandled flag : 0x%x\n", __func__, flags);
>  		break;
>  	}
> -	return ERR_PTR(-EOPNOTSUPP);
> +
> +	return dom;

These seem better to be a preparatory patch.

> @@ -3113,6 +3116,7 @@ const struct iommu_ops amd_iommu_ops = {
>  	.release_domain = &release_domain,
>  	.identity_domain = &identity_domain.domain,
>  	.domain_alloc_paging_flags = amd_iommu_domain_alloc_paging_flags,
> +	.domain_alloc_nested = amd_iommu_domain_alloc_nested,
>  	.domain_alloc_sva = amd_iommu_domain_alloc_sva,
>  	.probe_device = amd_iommu_probe_device,
>  	.release_device = amd_iommu_release_device,

This will be an HWPT-based nesting support, v.s. vIOMMU-based.

If AMD wants to enable its Command/Event Buffers, I think this
should follow the vIOMMU model instead.

> +// SPDX-License-Identifier: GPL-2.0-only
> +/*
> + * Copyright (C) 2023 Advanced Micro Devices, Inc.
> + */
> +
> +#define pr_fmt(fmt)     "AMD-Vi: " fmt
> +#define dev_fmt(fmt)    pr_fmt(fmt)
> +
> +#include <linux/iommu.h>
> +#include <uapi/linux/iommufd.h>
> +
> +#include "amd_iommu.h"
> +#include "amd_iommu_types.h"

It seems that you already included the uapi header in "amd_iommu.h".

> +static int udata_to_iommu_hwpt_amd_v2(const struct iommu_user_data *user_data,
> +				       struct iommu_hwpt_amd_v2 *hwpt)
> +{
> +	if (!user_data)
> +		return -EINVAL;
> +
> +	if (user_data->type != IOMMU_HWPT_DATA_AMD_V2)
> +		return -EOPNOTSUPP;
> +

iommu_copy_struct_from_user() internally checks these two already.

> +	return iommu_copy_struct_from_user(hwpt, user_data,
> +					   IOMMU_HWPT_DATA_AMD_V2,
> +					   dte);
> +}
> +
> +struct iommu_domain *
> +amd_iommu_domain_alloc_nested(struct device *dev, struct iommu_domain *parent,
> +			      u32 flags, const struct iommu_user_data *user_data)
> +{
> +	int ret;
> +	struct iommu_hwpt_amd_v2 hwpt;
> +	struct protection_domain *pdom;
> +
> +	if (parent->ops != amd_iommu_ops.default_domain_ops)
> +		return ERR_PTR(-EINVAL);
> +
> +	ret = udata_to_iommu_hwpt_amd_v2(user_data, &hwpt);
> +	if (ret)
> +		return ERR_PTR(ret);
> +
> +	pdom = kzalloc(sizeof(*pdom), GFP_KERNEL);
> +	if (IS_ERR(pdom))
> +		return ERR_PTR(-ENOMEM);
> +
> +	pdom->id = amd_iommu_pdom_id_alloc();
> +	if (!pdom->id)
> +		goto out_err;

This seems incorrect. amd_iommu_pdom_id_alloc() is a wrapper of the
ida_alloc_range() that would return -ENOMEM or -ENOSPC on failure.

Also, -EINVAL in out_err isn't nice to replace either of them.

So, I think this should be:

	if (pdom->id <= 0) {
		ret = pdom->id;
		goto out_err;
	}

> +
> +	pr_debug("%s: Allocating nested domain with parent domid=%#x\n",
> +		 __func__, to_pdomain(parent)->id);
> +
> +	spin_lock_init(&pdom->lock);
> +	INIT_LIST_HEAD(&pdom->dev_list);
> +	INIT_LIST_HEAD(&pdom->dev_data_list);
> +	xa_init(&pdom->iommu_array);
> +
> +	pdom->pd_mode = PD_MODE_V2;
> +	pdom->iop.pgtbl.cfg.amd.nid = NUMA_NO_NODE;
> +	pdom->parent = to_pdomain(parent);
> +	pdom->domain.ops = &nested_domain_ops;
> +	pdom->domain.type = IOMMU_DOMAIN_NESTED;
> +	pdom->domain.geometry.aperture_start = 0;
> +	pdom->domain.geometry.aperture_end = ((1ULL << PM_LEVEL_SHIFT(amd_iommu_gpt_level)) - 1);
> +	pdom->domain.geometry.force_aperture = true;
> +	pdom->domain.pgsize_bitmap = pdom->iop.pgtbl.cfg.pgsize_bitmap;
> +	memcpy(&pdom->guest_hwpt, &hwpt, sizeof(struct iommu_hwpt_amd_v2));

How about just hold a "struct dev_table_entry guest_dte" in the
pdom, instead of holding a uAPI structure?

Nicolin