RDMA/mlx5: Separate to user/kernel create QP flows
authorLeon Romanovsky <leonro@mellanox.com>
Mon, 27 Apr 2020 15:46:28 +0000 (18:46 +0300)
committerJason Gunthorpe <jgg@mellanox.com>
Thu, 30 Apr 2020 21:45:44 +0000 (18:45 -0300)
The kernel and user create QP flows have very little common code,
separate them to simplify the future work of creating per-type
create_*_qp() functions.

Link: https://lore.kernel.org/r/20200427154636.381474-29-leon@kernel.org
Reviewed-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
drivers/infiniband/hw/mlx5/qp.c

index 8890c17..d506100 100644 (file)
@@ -911,12 +911,12 @@ static int adjust_bfregn(struct mlx5_ib_dev *dev,
                                bfregn % MLX5_NON_FP_BFREGS_PER_UAR;
 }
 
-static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
-                         struct mlx5_ib_qp *qp, struct ib_udata *udata,
-                         struct ib_qp_init_attr *attr, u32 **in,
-                         struct mlx5_ib_create_qp_resp *resp, int *inlen,
-                         struct mlx5_ib_qp_base *base,
-                         struct mlx5_ib_create_qp *ucmd)
+static int _create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
+                          struct mlx5_ib_qp *qp, struct ib_udata *udata,
+                          struct ib_qp_init_attr *attr, u32 **in,
+                          struct mlx5_ib_create_qp_resp *resp, int *inlen,
+                          struct mlx5_ib_qp_base *base,
+                          struct mlx5_ib_create_qp *ucmd)
 {
        struct mlx5_ib_ucontext *context;
        struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
@@ -1083,11 +1083,10 @@ static void *get_sq_edge(struct mlx5_ib_wq *sq, u32 idx)
        return fragment_end + MLX5_SEND_WQE_BB;
 }
 
-static int create_kernel_qp(struct mlx5_ib_dev *dev,
-                           struct ib_qp_init_attr *init_attr,
-                           struct mlx5_ib_qp *qp,
-                           u32 **in, int *inlen,
-                           struct mlx5_ib_qp_base *base)
+static int _create_kernel_qp(struct mlx5_ib_dev *dev,
+                            struct ib_qp_init_attr *init_attr,
+                            struct mlx5_ib_qp *qp, u32 **in, int *inlen,
+                            struct mlx5_ib_qp_base *base)
 {
        int uar_index;
        void *qpc;
@@ -1987,11 +1986,11 @@ static int create_xrc_tgt_qp(struct mlx5_ib_dev *dev,
        return 0;
 }
 
-static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
-                           struct ib_qp_init_attr *init_attr,
-                           struct mlx5_ib_create_qp *ucmd,
-                           struct ib_udata *udata, struct mlx5_ib_qp *qp,
-                           u32 uidx)
+static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
+                         struct ib_qp_init_attr *init_attr,
+                         struct mlx5_ib_create_qp *ucmd,
+                         struct ib_udata *udata, struct mlx5_ib_qp *qp,
+                         u32 uidx)
 {
        struct mlx5_ib_resources *devr = &dev->devr;
        int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
@@ -2032,28 +2031,15 @@ static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
                return err;
        }
 
-       if (udata) {
-               __u32 max_wqes = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
+       if (ucmd->rq_wqe_shift != qp->rq.wqe_shift ||
+           ucmd->rq_wqe_count != qp->rq.wqe_cnt)
+               return -EINVAL;
 
-               mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n",
-                           ucmd->sq_wqe_count);
-               if (ucmd->rq_wqe_shift != qp->rq.wqe_shift ||
-                   ucmd->rq_wqe_count != qp->rq.wqe_cnt) {
-                       mlx5_ib_dbg(dev, "invalid rq params\n");
-                       return -EINVAL;
-               }
-               if (ucmd->sq_wqe_count > max_wqes) {
-                       mlx5_ib_dbg(
-                               dev,
-                               "requested sq_wqe_count (%d) > max allowed (%d)\n",
-                               ucmd->sq_wqe_count, max_wqes);
-                       return -EINVAL;
-               }
-               err = create_user_qp(dev, pd, qp, udata, init_attr, &in, &resp,
-                                    &inlen, base, ucmd);
-       } else
-               err = create_kernel_qp(dev, init_attr, qp, &in, &inlen, base);
+       if (ucmd->sq_wqe_count > (1 << MLX5_CAP_GEN(mdev, log_max_qp_sz)))
+               return -EINVAL;
 
+       err = _create_user_qp(dev, pd, qp, udata, init_attr, &in, &resp, &inlen,
+                             base, ucmd);
        if (err)
                return err;
 
@@ -2064,12 +2050,7 @@ static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
 
        MLX5_SET(qpc, qpc, st, mlx5_st);
        MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
-
-       if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
-               MLX5_SET(qpc, qpc, pd, to_mpd(pd ? pd : devr->p0)->pdn);
-       else
-               MLX5_SET(qpc, qpc, latency_sensitive, 1);
-
+       MLX5_SET(qpc, qpc, pd, to_mpd(pd)->pdn);
 
        if (qp->flags_en & MLX5_QP_FLAG_SIGNATURE)
                MLX5_SET(qpc, qpc, wq_signature, 1);
@@ -2145,10 +2126,6 @@ static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
        if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
                MLX5_SET(qpc, qpc, user_index, uidx);
 
-       /* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
-       if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO)
-               MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
-
        if (qp->flags & IB_QP_CREATE_PCI_WRITE_END_PADDING &&
            init_attr->qp_type != IB_QPT_RAW_PACKET) {
                MLX5_SET(qpc, qpc, end_padding_mode,
@@ -2200,6 +2177,133 @@ err_create:
        return err;
 }
 
+static int create_kernel_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
+                           struct ib_qp_init_attr *attr, struct mlx5_ib_qp *qp,
+                           u32 uidx)
+{
+       struct mlx5_ib_resources *devr = &dev->devr;
+       int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
+       struct mlx5_core_dev *mdev = dev->mdev;
+       struct mlx5_ib_cq *send_cq;
+       struct mlx5_ib_cq *recv_cq;
+       unsigned long flags;
+       struct mlx5_ib_qp_base *base;
+       int mlx5_st;
+       void *qpc;
+       u32 *in;
+       int err;
+
+       mutex_init(&qp->mutex);
+       spin_lock_init(&qp->sq.lock);
+       spin_lock_init(&qp->rq.lock);
+
+       mlx5_st = to_mlx5_st(qp->type);
+       if (mlx5_st < 0)
+               return -EINVAL;
+
+       if (attr->sq_sig_type == IB_SIGNAL_ALL_WR)
+               qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
+
+       base = &qp->trans_qp.base;
+
+       qp->has_rq = qp_has_rq(attr);
+       err = set_rq_size(dev, &attr->cap, qp->has_rq, qp, NULL);
+       if (err) {
+               mlx5_ib_dbg(dev, "err %d\n", err);
+               return err;
+       }
+
+       err = _create_kernel_qp(dev, attr, qp, &in, &inlen, base);
+       if (err)
+               return err;
+
+       if (is_sqp(attr->qp_type))
+               qp->port = attr->port_num;
+
+       qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
+
+       MLX5_SET(qpc, qpc, st, mlx5_st);
+       MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
+
+       if (attr->qp_type != MLX5_IB_QPT_REG_UMR)
+               MLX5_SET(qpc, qpc, pd, to_mpd(pd ? pd : devr->p0)->pdn);
+       else
+               MLX5_SET(qpc, qpc, latency_sensitive, 1);
+
+
+       if (qp->flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
+               MLX5_SET(qpc, qpc, block_lb_mc, 1);
+
+       if (qp->rq.wqe_cnt) {
+               MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
+               MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
+       }
+
+       MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, attr));
+
+       if (qp->sq.wqe_cnt)
+               MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
+       else
+               MLX5_SET(qpc, qpc, no_sq, 1);
+
+       if (attr->srq) {
+               MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x0)->xrcdn);
+               MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
+                        to_msrq(attr->srq)->msrq.srqn);
+       } else {
+               MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
+               MLX5_SET(qpc, qpc, srqn_rmpn_xrqn,
+                        to_msrq(devr->s1)->msrq.srqn);
+       }
+
+       if (attr->send_cq)
+               MLX5_SET(qpc, qpc, cqn_snd, to_mcq(attr->send_cq)->mcq.cqn);
+
+       if (attr->recv_cq)
+               MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(attr->recv_cq)->mcq.cqn);
+
+       MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
+
+       /* 0xffffff means we ask to work with cqe version 0 */
+       if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
+               MLX5_SET(qpc, qpc, user_index, uidx);
+
+       /* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
+       if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO)
+               MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
+
+       err = mlx5_core_create_qp(dev, &base->mqp, in, inlen);
+       kvfree(in);
+       if (err)
+               goto err_create;
+
+       base->container_mibqp = qp;
+       base->mqp.event = mlx5_ib_qp_event;
+
+       get_cqs(qp->type, attr->send_cq, attr->recv_cq,
+               &send_cq, &recv_cq);
+       spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
+       mlx5_ib_lock_cqs(send_cq, recv_cq);
+       /* Maintain device to QPs access, needed for further handling via reset
+        * flow
+        */
+       list_add_tail(&qp->qps_list, &dev->qp_list);
+       /* Maintain CQ to QPs access, needed for further handling via reset flow
+        */
+       if (send_cq)
+               list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
+       if (recv_cq)
+               list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
+       mlx5_ib_unlock_cqs(send_cq, recv_cq);
+       spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
+
+       return 0;
+
+err_create:
+       destroy_qp_kernel(dev, qp);
+       return err;
+}
+
 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
        __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
 {
@@ -2695,7 +2799,7 @@ static int create_raw_qp(struct ib_pd *pd, struct mlx5_ib_qp *qp,
        if (attr->rwq_ind_tbl)
                return create_rss_raw_qp_tir(pd, qp, attr, ucmd, udata);
 
-       return create_qp_common(dev, pd, attr, ucmd, udata, qp, uidx);
+       return create_user_qp(dev, pd, attr, ucmd, udata, qp, uidx);
 }
 
 static int check_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
@@ -2819,8 +2923,11 @@ struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
                err = create_xrc_tgt_qp(dev, init_attr, qp, udata, uidx);
                break;
        default:
-               err = create_qp_common(dev, pd, init_attr, ucmd, udata, qp,
-                                      uidx);
+               if (udata)
+                       err = create_user_qp(dev, pd, init_attr, ucmd, udata,
+                                            qp, uidx);
+               else
+                       err = create_kernel_qp(dev, pd, init_attr, qp, uidx);
        }
        if (err) {
                mlx5_ib_dbg(dev, "create_qp failed %d\n", err);