Re: [PATCH v8 14/29] iommufd/viommu: Add IOMMUFD_CMD_HW_QUEUE_ALLOC ioctl

[Vasant Hegde <vasant.hegde@amd.com>]

Hi ,


On 7/5/2025 6:43 AM, Nicolin Chen wrote:
> Introduce a new IOMMUFD_CMD_HW_QUEUE_ALLOC ioctl for user space to allocate
> a HW QUEUE object for a vIOMMU specific HW-accelerated queue, e.g.:
>  - NVIDIA's Virtual Command Queue
>  - AMD vIOMMU's Command Buffer, Event Log Buffers, and PPR Log Buffers
> 
> Since this is introduced with NVIDIA's VCMDQs that access the guest memory
> in the physical address space, add an iommufd_hw_queue_alloc_phys() helper
> that will create an access object to the queue memory in the IOAS, to avoid
> the mappings of the guest memory from being unmapped, during the life cycle
> of the HW queue object.
> 
> AMD's HW will need an hw_queue_init op that is mutually exclusive with the
> hw_queue_init_phys op, and their case will bypass the access part, i.e. no
> iommufd_hw_queue_alloc_phys() call.

Thanks. We will implement hw_queue_init[_iova] to support AMD driver and fixup
iommufd_hw_queue_alloc_ioctl(). Is that the correct understanding?

-Vasant

> 
> Reviewed-by: Pranjal Shrivastava <praan@google.com>
> Reviewed-by: Kevin Tian <kevin.tian@intel.com>
> Reviewed-by: Lu Baolu <baolu.lu@linux.intel.com>
> Signed-off-by: Nicolin Chen <nicolinc@nvidia.com>
> --->  drivers/iommu/iommufd/iommufd_private.h |   2 +
>  include/linux/iommufd.h                 |   1 +
>  include/uapi/linux/iommufd.h            |  33 +++++
>  drivers/iommu/iommufd/main.c            |   6 +
>  drivers/iommu/iommufd/viommu.c          | 177 ++++++++++++++++++++++++
>  5 files changed, 219 insertions(+)
> 
> diff --git a/drivers/iommu/iommufd/iommufd_private.h b/drivers/iommu/iommufd/iommufd_private.h
> index 06b8c2e2d9e6..dcd609573244 100644
> --- a/drivers/iommu/iommufd/iommufd_private.h
> +++ b/drivers/iommu/iommufd/iommufd_private.h
> @@ -652,6 +652,8 @@ int iommufd_viommu_alloc_ioctl(struct iommufd_ucmd *ucmd);
>  void iommufd_viommu_destroy(struct iommufd_object *obj);
>  int iommufd_vdevice_alloc_ioctl(struct iommufd_ucmd *ucmd);
>  void iommufd_vdevice_destroy(struct iommufd_object *obj);
> +int iommufd_hw_queue_alloc_ioctl(struct iommufd_ucmd *ucmd);
> +void iommufd_hw_queue_destroy(struct iommufd_object *obj);
>  
>  #ifdef CONFIG_IOMMUFD_TEST
>  int iommufd_test(struct iommufd_ucmd *ucmd);
> diff --git a/include/linux/iommufd.h b/include/linux/iommufd.h
> index f13f3ca6adb5..ce4011a2fc27 100644
> --- a/include/linux/iommufd.h
> +++ b/include/linux/iommufd.h
> @@ -123,6 +123,7 @@ struct iommufd_vdevice {
>  struct iommufd_hw_queue {
>  	struct iommufd_object obj;
>  	struct iommufd_viommu *viommu;
> +	struct iommufd_access *access;
>  
>  	u64 base_addr; /* in guest physical address space */
>  	size_t length;
> diff --git a/include/uapi/linux/iommufd.h b/include/uapi/linux/iommufd.h
> index 640a8b5147c2..55459b9eee31 100644
> --- a/include/uapi/linux/iommufd.h
> +++ b/include/uapi/linux/iommufd.h
> @@ -56,6 +56,7 @@ enum {
>  	IOMMUFD_CMD_VDEVICE_ALLOC = 0x91,
>  	IOMMUFD_CMD_IOAS_CHANGE_PROCESS = 0x92,
>  	IOMMUFD_CMD_VEVENTQ_ALLOC = 0x93,
> +	IOMMUFD_CMD_HW_QUEUE_ALLOC = 0x94,
>  };
>  
>  /**
> @@ -1156,4 +1157,36 @@ enum iommu_hw_queue_type {
>  	IOMMU_HW_QUEUE_TYPE_DEFAULT = 0,
>  };
>  
> +/**
> + * struct iommu_hw_queue_alloc - ioctl(IOMMU_HW_QUEUE_ALLOC)
> + * @size: sizeof(struct iommu_hw_queue_alloc)
> + * @flags: Must be 0
> + * @viommu_id: Virtual IOMMU ID to associate the HW queue with
> + * @type: One of enum iommu_hw_queue_type
> + * @index: The logical index to the HW queue per virtual IOMMU for a multi-queue
> + *         model
> + * @out_hw_queue_id: The ID of the new HW queue
> + * @nesting_parent_iova: Base address of the queue memory in the guest physical
> + *                       address space
> + * @length: Length of the queue memory
> + *
> + * Allocate a HW queue object for a vIOMMU-specific HW-accelerated queue, which
> + * allows HW to access a guest queue memory described using @nesting_parent_iova
> + * and @length.
> + *
> + * A vIOMMU can allocate multiple queues, but it must use a different @index per
> + * type to separate each allocation, e.g.
> + *     Type1 HW queue0, Type1 HW queue1, Type2 HW queue0, ...
> + */
> +struct iommu_hw_queue_alloc {
> +	__u32 size;
> +	__u32 flags;
> +	__u32 viommu_id;
> +	__u32 type;
> +	__u32 index;
> +	__u32 out_hw_queue_id;
> +	__aligned_u64 nesting_parent_iova;
> +	__aligned_u64 length;
> +};
> +#define IOMMU_HW_QUEUE_ALLOC _IO(IOMMUFD_TYPE, IOMMUFD_CMD_HW_QUEUE_ALLOC)
>  #endif
> diff --git a/drivers/iommu/iommufd/main.c b/drivers/iommu/iommufd/main.c
> index 778694d7c207..4e8dbbfac890 100644
> --- a/drivers/iommu/iommufd/main.c
> +++ b/drivers/iommu/iommufd/main.c
> @@ -354,6 +354,7 @@ union ucmd_buffer {
>  	struct iommu_destroy destroy;
>  	struct iommu_fault_alloc fault;
>  	struct iommu_hw_info info;
> +	struct iommu_hw_queue_alloc hw_queue;
>  	struct iommu_hwpt_alloc hwpt;
>  	struct iommu_hwpt_get_dirty_bitmap get_dirty_bitmap;
>  	struct iommu_hwpt_invalidate cache;
> @@ -396,6 +397,8 @@ static const struct iommufd_ioctl_op iommufd_ioctl_ops[] = {
>  		 struct iommu_fault_alloc, out_fault_fd),
>  	IOCTL_OP(IOMMU_GET_HW_INFO, iommufd_get_hw_info, struct iommu_hw_info,
>  		 __reserved),
> +	IOCTL_OP(IOMMU_HW_QUEUE_ALLOC, iommufd_hw_queue_alloc_ioctl,
> +		 struct iommu_hw_queue_alloc, length),
>  	IOCTL_OP(IOMMU_HWPT_ALLOC, iommufd_hwpt_alloc, struct iommu_hwpt_alloc,
>  		 __reserved),
>  	IOCTL_OP(IOMMU_HWPT_GET_DIRTY_BITMAP, iommufd_hwpt_get_dirty_bitmap,
> @@ -559,6 +562,9 @@ static const struct iommufd_object_ops iommufd_object_ops[] = {
>  	[IOMMUFD_OBJ_FAULT] = {
>  		.destroy = iommufd_fault_destroy,
>  	},
> +	[IOMMUFD_OBJ_HW_QUEUE] = {
> +		.destroy = iommufd_hw_queue_destroy,
> +	},
>  	[IOMMUFD_OBJ_HWPT_PAGING] = {
>  		.destroy = iommufd_hwpt_paging_destroy,
>  		.abort = iommufd_hwpt_paging_abort,
> diff --git a/drivers/iommu/iommufd/viommu.c b/drivers/iommu/iommufd/viommu.c
> index 081ee6697a11..00641204efb2 100644
> --- a/drivers/iommu/iommufd/viommu.c
> +++ b/drivers/iommu/iommufd/viommu.c
> @@ -201,3 +201,180 @@ int iommufd_vdevice_alloc_ioctl(struct iommufd_ucmd *ucmd)
>  	iommufd_put_object(ucmd->ictx, &viommu->obj);
>  	return rc;
>  }
> +
> +static void iommufd_hw_queue_destroy_access(struct iommufd_ctx *ictx,
> +					    struct iommufd_access *access,
> +					    u64 base_iova, size_t length)
> +{
> +	iommufd_access_unpin_pages(access, base_iova, length);
> +	iommufd_access_detach_internal(access);
> +	iommufd_access_destroy_internal(ictx, access);
> +}
> +
> +void iommufd_hw_queue_destroy(struct iommufd_object *obj)
> +{
> +	struct iommufd_hw_queue *hw_queue =
> +		container_of(obj, struct iommufd_hw_queue, obj);
> +
> +	if (hw_queue->destroy)
> +		hw_queue->destroy(hw_queue);
> +	if (hw_queue->access)
> +		iommufd_hw_queue_destroy_access(hw_queue->viommu->ictx,
> +						hw_queue->access,
> +						hw_queue->base_addr,
> +						hw_queue->length);
> +	if (hw_queue->viommu)
> +		refcount_dec(&hw_queue->viommu->obj.users);
> +}
> +
> +/*
> + * When the HW accesses the guest queue via physical addresses, the underlying
> + * physical pages of the guest queue must be contiguous. Also, for the security
> + * concern that IOMMUFD_CMD_IOAS_UNMAP could potentially remove the mappings of
> + * the guest queue from the nesting parent iopt while the HW is still accessing
> + * the guest queue memory physically, such a HW queue must require an access to
> + * pin the underlying pages and prevent that from happening.
> + */
> +static struct iommufd_access *
> +iommufd_hw_queue_alloc_phys(struct iommu_hw_queue_alloc *cmd,
> +			    struct iommufd_viommu *viommu, phys_addr_t *base_pa)
> +{
> +	struct iommufd_access *access;
> +	struct page **pages;
> +	size_t max_npages;
> +	size_t length;
> +	u64 offset;
> +	size_t i;
> +	int rc;
> +
> +	offset =
> +		cmd->nesting_parent_iova - PAGE_ALIGN(cmd->nesting_parent_iova);
> +	/* DIV_ROUND_UP(offset + cmd->length, PAGE_SIZE) */
> +	if (check_add_overflow(offset, cmd->length, &length))
> +		return ERR_PTR(-ERANGE);
> +	if (check_add_overflow(length, PAGE_SIZE - 1, &length))
> +		return ERR_PTR(-ERANGE);
> +	max_npages = length / PAGE_SIZE;
> +
> +	/*
> +	 * Use kvcalloc() to avoid memory fragmentation for a large page array.
> +	 * Set __GFP_NOWARN to avoid syzkaller blowups
> +	 */
> +	pages = kvcalloc(max_npages, sizeof(*pages), GFP_KERNEL | __GFP_NOWARN);
> +	if (!pages)
> +		return ERR_PTR(-ENOMEM);
> +
> +	access = iommufd_access_create_internal(viommu->ictx);
> +	if (IS_ERR(access)) {
> +		rc = PTR_ERR(access);
> +		goto out_free;
> +	}
> +
> +	rc = iommufd_access_attach_internal(access, viommu->hwpt->ioas);
> +	if (rc)
> +		goto out_destroy;
> +
> +	rc = iommufd_access_pin_pages(access, cmd->nesting_parent_iova,
> +				      cmd->length, pages, 0);
> +	if (rc)
> +		goto out_detach;
> +
> +	/* Validate if the underlying physical pages are contiguous */
> +	for (i = 1; i < max_npages; i++) {
> +		if (page_to_pfn(pages[i]) == page_to_pfn(pages[i - 1]) + 1)
> +			continue;
> +		rc = -EFAULT;
> +		goto out_unpin;
> +	}
> +
> +	*base_pa = page_to_pfn(pages[0]) << PAGE_SHIFT;
> +	kfree(pages);
> +	return access;
> +
> +out_unpin:
> +	iommufd_access_unpin_pages(access, cmd->nesting_parent_iova,
> +				   cmd->length);
> +out_detach:
> +	iommufd_access_detach_internal(access);
> +out_destroy:
> +	iommufd_access_destroy_internal(viommu->ictx, access);
> +out_free:
> +	kfree(pages);
> +	return ERR_PTR(rc);
> +}
> +
> +int iommufd_hw_queue_alloc_ioctl(struct iommufd_ucmd *ucmd)
> +{
> +	struct iommu_hw_queue_alloc *cmd = ucmd->cmd;
> +	struct iommufd_hw_queue *hw_queue;
> +	struct iommufd_viommu *viommu;
> +	struct iommufd_access *access;
> +	size_t hw_queue_size;
> +	phys_addr_t base_pa;
> +	u64 last;
> +	int rc;
> +
> +	if (cmd->flags || cmd->type == IOMMU_HW_QUEUE_TYPE_DEFAULT)
> +		return -EOPNOTSUPP;
> +	if (!cmd->length)
> +		return -EINVAL;
> +	if (check_add_overflow(cmd->nesting_parent_iova, cmd->length - 1,
> +			       &last))
> +		return -EOVERFLOW;
> +
> +	viommu = iommufd_get_viommu(ucmd, cmd->viommu_id);
> +	if (IS_ERR(viommu))
> +		return PTR_ERR(viommu);
> +
> +	if (!viommu->ops || !viommu->ops->get_hw_queue_size ||
> +	    !viommu->ops->hw_queue_init_phys) {
> +		rc = -EOPNOTSUPP;
> +		goto out_put_viommu;
> +	}
> +
> +	hw_queue_size = viommu->ops->get_hw_queue_size(viommu, cmd->type);
> +	if (!hw_queue_size) {
> +		rc = -EOPNOTSUPP;
> +		goto out_put_viommu;
> +	}
> +
> +	/*
> +	 * It is a driver bug for providing a hw_queue_size smaller than the
> +	 * core HW queue structure size
> +	 */
> +	if (WARN_ON_ONCE(hw_queue_size < sizeof(*hw_queue))) {
> +		rc = -EOPNOTSUPP;
> +		goto out_put_viommu;
> +	}
> +
> +	hw_queue = (struct iommufd_hw_queue *)_iommufd_object_alloc_ucmd(
> +		ucmd, hw_queue_size, IOMMUFD_OBJ_HW_QUEUE);
> +	if (IS_ERR(hw_queue)) {
> +		rc = PTR_ERR(hw_queue);
> +		goto out_put_viommu;
> +	}
> +
> +	access = iommufd_hw_queue_alloc_phys(cmd, viommu, &base_pa);
> +	if (IS_ERR(access)) {
> +		rc = PTR_ERR(access);
> +		goto out_put_viommu;
> +	}
> +
> +	hw_queue->viommu = viommu;
> +	refcount_inc(&viommu->obj.users);
> +	hw_queue->access = access;
> +	hw_queue->type = cmd->type;
> +	hw_queue->length = cmd->length;
> +	hw_queue->base_addr = cmd->nesting_parent_iova;
> +
> +	rc = viommu->ops->hw_queue_init_phys(hw_queue, cmd->index, base_pa);
> +	if (rc)
> +		goto out_put_viommu;
> +
> +	cmd->out_hw_queue_id = hw_queue->obj.id;
> +	rc = iommufd_ucmd_respond(ucmd, sizeof(*cmd));
> +
> +out_put_viommu:
> +	iommufd_put_object(ucmd->ictx, &viommu->obj);
> +	return rc;
> +}