Re: [PATCH V4 mlx5-next 11/15] RDMA/core: Get xmit slave for LAG

[Jason Gunthorpe <jgg@mellanox.com>]

On Wed, Apr 22, 2020 at 11:39:47AM +0300, Maor Gottlieb wrote:
> Add a call to rdma_lag_get_ah_roce_slave when
> Address handle is created.
> Low driver can use it to select the QP's affinity port.

Lower
 
> Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
> Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
>  drivers/infiniband/core/verbs.c | 44 ++++++++++++++++++++++-----------
>  1 file changed, 30 insertions(+), 14 deletions(-)
> 
> diff --git a/drivers/infiniband/core/verbs.c b/drivers/infiniband/core/verbs.c
> index 56a71337112c..a0d60376ba6b 100644
> +++ b/drivers/infiniband/core/verbs.c
> @@ -50,6 +50,7 @@
>  #include <rdma/ib_cache.h>
>  #include <rdma/ib_addr.h>
>  #include <rdma/rw.h>
> +#include <rdma/lag.h>
>  
>  #include "core_priv.h"
>  #include <trace/events/rdma_core.h>
> @@ -554,8 +555,14 @@ struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr,
>  	if (ret)
>  		return ERR_PTR(ret);
>  
> -	ah = _rdma_create_ah(pd, ah_attr, flags, NULL);
> +	ret = rdma_lag_get_ah_roce_slave(pd->device, ah_attr);
> +	if (ret) {
> +		rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
> +		return ERR_PTR(ret);
> +	}
>  
> +	ah = _rdma_create_ah(pd, ah_attr, flags, NULL);
> +	rdma_lag_put_ah_roce_slave(ah_attr);
>  	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
>  	return ah;
>  }
> @@ -1638,6 +1645,25 @@ static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
>  					  &old_sgid_attr_av);
>  		if (ret)
>  			return ret;
> +
> +		if (attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
> +		    is_qp_type_connected(qp)) {
> +			/*
> +			 * If the user provided the qp_attr then we have to
> +			 * resolve it. Kerne users have to provide already

Kernel

> +			 * resolved rdma_ah_attr's.
> +			 */
> +			if (udata) {
> +				ret = ib_resolve_eth_dmac(qp->device,
> +							  &attr->ah_attr);
> +				if (ret)
> +					goto out_av;
> +			}
> +			ret = rdma_lag_get_ah_roce_slave(qp->device,
> +							 &attr->ah_attr);
> +			if (ret)
> +				goto out_av;
> +		}
>  	}
>  	if (attr_mask & IB_QP_ALT_PATH) {
>  		/*
> @@ -1664,18 +1690,6 @@ static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
>  		}
>  	}
>  
> -	/*
> -	 * If the user provided the qp_attr then we have to resolve it. Kernel
> -	 * users have to provide already resolved rdma_ah_attr's
> -	 */
> -	if (udata && (attr_mask & IB_QP_AV) &&
> -	    attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
> -	    is_qp_type_connected(qp)) {
> -		ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
> -		if (ret)
> -			goto out;
> -	}

Why did this need to move up?

> -
>  	if (rdma_ib_or_roce(qp->device, port)) {
>  		if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
>  			dev_warn(&qp->device->dev,
> @@ -1717,8 +1731,10 @@ static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
>  	if (attr_mask & IB_QP_ALT_PATH)
>  		rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av);
>  out_av:
> -	if (attr_mask & IB_QP_AV)
> +	if (attr_mask & IB_QP_AV) {
> +		rdma_lag_put_ah_roce_slave(&attr->ah_attr);

This seems wwrong why doesn't rdma_unfill_sgid_attr do this?

Jason