Re: [PATCH rdma-next v2] RDMA: Explicitly pass in the dma_device to ib_register_device

[Christoph Hellwig <hch@infradead.org>]

> +	WARN_ON(!IS_ENABLED(CONFIG_DMA_VIRT_OPS) && !dma_device);
> +	if (!dma_device) {
>  		/*
> -		 * The caller did not provide custom DMA operations. Use the
> -		 * DMA mapping operations of the parent device.
> +		 * If the caller does not provide a DMA capable device then the
> +		 * IB device will be used. In this case the caller should fully
> +		 * setup the ibdev for DMA. This usually means using
> +		 * dma_virt_ops.
>  		 */
> +		device->dev.dma_ops = &dma_virt_ops;
> +		dma_device = &device->dev;

The lack of the if probably means this will fail to link now when
CONFIG_DMA_VIRT_OPS is not set.  This also seems to not remove the
dma_virt_ops assignment in the callers.

> index f904bb34477a..2f117ac11c8b 100644
> --- a/drivers/infiniband/sw/rdmavt/vt.c
> +++ b/drivers/infiniband/sw/rdmavt/vt.c
> @@ -581,7 +581,11 @@ int rvt_register_device(struct rvt_dev_info *rdi)
>  	spin_lock_init(&rdi->n_cqs_lock);
> 
>  	/* DMA Operations */
> -	rdi->ibdev.dev.dma_ops = rdi->ibdev.dev.dma_ops ? : &dma_virt_ops;
> +	rdi->ibdev.dev.dma_ops = &dma_virt_ops;
> +	rdi->ibdev.dev.dma_parms = rdi->ibdev.dev.parent->dma_parms;
> +	rdi->ibdev.dev.dma_mask = rdi->ibdev.dev.parent->dma_mask;

This copies the dma_mask pointer, which seems completely bogus.