Re: [RFC 2/2] RDMA/irdma: Add pinned revocable dmabuf support

[Jason Gunthorpe <jgg@ziepe.ca>]

On Mon, Feb 23, 2026 at 07:53:33PM +0000, Jacob Moroni wrote:

> +static void irdma_umem_dmabuf_revoke(void *priv)
> +{
> +	struct irdma_mr *iwmr = priv;
> +	int err;
> +
> +	iwmr->revoked = true;
> +
> +	if (!iwmr->is_hwreg)
> +		return;
> +
> +	/* Invalidate the key in hardware. This does not actually release the
> +	 * key for potential reuse - that only occurs when the region is fully
> +	 * deregistered.
> +	 */
> +	err = irdma_hwdereg_mr(&iwmr->ibmr);
> +	if (err) {
> +		struct irdma_device *iwdev = to_iwdev(iwmr->ibmr.device);
> +
> +		ibdev_err(&iwdev->ibdev, "dmabuf mr invalidate failed %d", err);
> +		if (!iwdev->rf->reset) {
> +			iwdev->rf->reset = true;
> +			iwdev->rf->gen_ops.request_reset(iwdev->rf);
> +		}
> +	}
> +}
> +
>  static struct ib_mr *irdma_reg_user_mr_dmabuf(struct ib_pd *pd, u64 start,
>  					      u64 len, u64 virt,
>  					      int fd, int access,
> @@ -3607,31 +3627,45 @@ static struct ib_mr *irdma_reg_user_mr_dmabuf(struct ib_pd *pd, u64 start,
>  	if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size)
>  		return ERR_PTR(-EINVAL);
>  
> -	umem_dmabuf = ib_umem_dmabuf_get_pinned(pd->device, start, len, fd, access);
> +	iwmr = kzalloc_obj(*iwmr);
> +	if (!iwmr)
> +		return ERR_PTR(-ENOMEM);
> +
> +	umem_dmabuf =
> +		ib_umem_dmabuf_get_pinned_revocable(pd->device, start, len, fd,
> +						    access,
> +						    irdma_umem_dmabuf_revoke,
> +						    iwmr);
>  	if (IS_ERR(umem_dmabuf)) {
>  		ibdev_dbg(&iwdev->ibdev, "Failed to get dmabuf umem[%pe]\n",
>  			  umem_dmabuf);
> +		kfree(iwmr);
>  		return ERR_CAST(umem_dmabuf);
>  	}
>  
> -	iwmr = irdma_alloc_iwmr(&umem_dmabuf->umem, pd, virt, IRDMA_MEMREG_TYPE_MEM);
> -	if (IS_ERR(iwmr)) {
> -		err = PTR_ERR(iwmr);
> +	err = irdma_init_iwmr(iwmr, &umem_dmabuf->umem, pd, virt,
> +			      IRDMA_MEMREG_TYPE_MEM);

Is it OK to call irdma_hwdereg_mr() before this? Seems really sketchy

I think if revoke is being used you have to use a protocol where the
dmabuf reservation lock is left held for the caller to complete setup
so you don't have revoke races.

I guess this is some ib_umem_dmabuf_get_pinned_revocable_and_lock()
pattern

Maybe it can be some two step process:

 ib_umem_dmabuf_get_pinned_locked()

 [..]
 ib_umem_dmabuf_set_pinned_revocable()

 dma_buf__resv_unlock()


?

Then you don't need to restructure all the iwmr allocation (which
should be a seperated patch too)

Jason