Re: [PATCH RFC 12/26] vfio/pci: Change the DMA-buf exporter to use mapping_type

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

On 2/18/26 08:11, Jason Gunthorpe wrote:
> Simple conversion to add a match_mapping() callback that offers an
> exporter SGT mapping type. Later patches will add a physical address
> exporter so go straight to adding the match_mapping() function.
> 
> The check for attachment->peer2peer is replaced with setting
> exporter_requires_p2p=true. VFIO always uses MMIO memory.
> 
> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
> ---
>   drivers/vfio/pci/vfio_pci_dmabuf.c | 31 +++++++++++++++++++++++++-----
>   1 file changed, 26 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/vfio/pci/vfio_pci_dmabuf.c b/drivers/vfio/pci/vfio_pci_dmabuf.c
> index d4d0f7d08c53e2..c7addef5794abf 100644
> --- a/drivers/vfio/pci/vfio_pci_dmabuf.c
> +++ b/drivers/vfio/pci/vfio_pci_dmabuf.c
> @@ -25,9 +25,6 @@ static int vfio_pci_dma_buf_attach(struct dma_buf *dmabuf,
>   {
>   	struct vfio_pci_dma_buf *priv = dmabuf->priv;
>   
> -	if (!attachment->peer2peer)
> -		return -EOPNOTSUPP;
> -
>   	if (priv->revoked)
>   		return -ENODEV;
>   
> @@ -75,11 +72,35 @@ static void vfio_pci_dma_buf_release(struct dma_buf *dmabuf)
>   	kfree(priv);
>   }
>   
> -static const struct dma_buf_ops vfio_pci_dmabuf_ops = {
> -	.attach = vfio_pci_dma_buf_attach,
> +static const struct dma_buf_mapping_sgt_exp_ops vfio_pci_dma_buf_sgt_ops = {
>   	.map_dma_buf = vfio_pci_dma_buf_map,
>   	.unmap_dma_buf = vfio_pci_dma_buf_unmap,
> +};
> +
> +static int vfio_pci_dma_buf_match_mapping(struct dma_buf_match_args *args)
> +{
> +	struct vfio_pci_dma_buf *priv = args->dmabuf->priv;
> +	struct dma_buf_mapping_match sgt_match[1];
> +
> +	dma_resv_assert_held(priv->dmabuf->resv);

My understanding of this lock assertion is that priv and the underlying
priv->vdev are accessed within this function. Therefore, the lock is
necessary to protect them. Do I understand it right?

However, callers - for example, dma_buf_mapping_attach() - do not
acquire dma_resv_lock() before calling this function. So kernel traces
will always be triggered.

> +
> +	/*
> +	 * Once we pass vfio_pci_dma_buf_cleanup() the dmabuf will never be
> +	 * usable again.
> +	 */
> +	if (!priv->vdev)
> +		return -ENODEV;
> +
> +	sgt_match[0] = DMA_BUF_EMAPPING_SGT_P2P(&vfio_pci_dma_buf_sgt_ops,
> +						priv->vdev->pdev);
> +
> +	return dma_buf_match_mapping(args, sgt_match, ARRAY_SIZE(sgt_match));
> +}
> +
> +static const struct dma_buf_ops vfio_pci_dmabuf_ops = {
> +	.attach = vfio_pci_dma_buf_attach,
>   	.release = vfio_pci_dma_buf_release,
> +	.match_mapping = vfio_pci_dma_buf_match_mapping,
>   };
>   
>   /*

Thanks,
baolu