RDMA/bnxt_re: Avoid an extra hwrm per MR creation

Firmware now have a new mr registration command where
both MR allocation and registration can be done in a
single hwrm command. Driver has to issue this new hwrm
command whenever the support flag is set. This reduces
the number of hwrm issued per MR creation and speed up
the MR creation.
Signed-off-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Link: https://patch.msgid.link/1725256351-12751-4-git-send-email-selvin.xavier@broadcom.comSigned-off-by: Leon Romanovsky <leon@kernel.org>

drivers/infiniband/hw/bnxt_re/ib_verbs.c

@@ -518,14 +518,18 @@ static int bnxt_re_create_fence_mr(struct bnxt_re_pd *pd)
 	mr->qplib_mr.pd = &pd->qplib_pd;
 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
 	mr->qplib_mr.access_flags = __from_ib_access_flags(mr_access_flags);
-	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
-	if (rc) {
-		ibdev_err(&rdev->ibdev, "Failed to alloc fence-HW-MR\n");
-		goto fail;
-	}
+	if (!_is_alloc_mr_unified(rdev->dev_attr.dev_cap_flags)) {
+		rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
+		if (rc) {
+			ibdev_err(&rdev->ibdev, "Failed to alloc fence-HW-MR\n");
+			goto fail;
+		}
 
-	/* Register MR */
-	mr->ib_mr.lkey = mr->qplib_mr.lkey;
+		/* Register MR */
+		mr->ib_mr.lkey = mr->qplib_mr.lkey;
+	} else {
+		mr->qplib_mr.flags = CMDQ_REGISTER_MR_FLAGS_ALLOC_MR;
+	}
 	mr->qplib_mr.va = (u64)(unsigned long)fence->va;
 	mr->qplib_mr.total_size = BNXT_RE_FENCE_BYTES;
 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, NULL,
@@ -4101,14 +4105,18 @@ static struct ib_mr *__bnxt_re_user_reg_mr(struct ib_pd *ib_pd, u64 length, u64
 	mr->qplib_mr.access_flags = __from_ib_access_flags(mr_access_flags);
 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_MR;
 
-	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
-	if (rc) {
-		ibdev_err(&rdev->ibdev, "Failed to allocate MR rc = %d", rc);
-		rc = -EIO;
-		goto free_mr;
+	if (!_is_alloc_mr_unified(rdev->dev_attr.dev_cap_flags)) {
+		rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
+		if (rc) {
+			ibdev_err(&rdev->ibdev, "Failed to allocate MR rc = %d", rc);
+			rc = -EIO;
+			goto free_mr;
+		}
+		/* The fixed portion of the rkey is the same as the lkey */
+		mr->ib_mr.rkey = mr->qplib_mr.rkey;
+	} else {
+		mr->qplib_mr.flags = CMDQ_REGISTER_MR_FLAGS_ALLOC_MR;
 	}
-	/* The fixed portion of the rkey is the same as the lkey */
-	mr->ib_mr.rkey = mr->qplib_mr.rkey;
 	mr->ib_umem = umem;
 	mr->qplib_mr.va = virt_addr;
 	mr->qplib_mr.total_size = length;

drivers/infiniband/hw/bnxt_re/qplib_res.h

@@ -565,4 +565,9 @@ static inline u8 bnxt_qplib_dbr_pacing_en(struct bnxt_qplib_chip_ctx *cctx)
 	return cctx->modes.dbr_pacing;
 }
 
+static inline bool _is_alloc_mr_unified(u16 dev_cap_flags)
+{
+	return dev_cap_flags & CREQ_QUERY_FUNC_RESP_SB_MR_REGISTER_ALLOC;
+}
+
 #endif /* __BNXT_QPLIB_RES_H__ */

drivers/infiniband/hw/bnxt_re/qplib_sp.c

@@ -659,6 +659,9 @@ int bnxt_qplib_reg_mr(struct bnxt_qplib_res *res, struct bnxt_qplib_mrw *mr,
 	req.access = (mr->access_flags & 0xFFFF);
 	req.va = cpu_to_le64(mr->va);
 	req.key = cpu_to_le32(mr->lkey);
+	if (_is_alloc_mr_unified(res->dattr->dev_cap_flags))
+		req.key = cpu_to_le32(mr->pd->id);
+	req.flags = cpu_to_le16(mr->flags);
 	req.mr_size = cpu_to_le64(mr->total_size);
 
 	bnxt_qplib_fill_cmdqmsg(&msg, &req, &resp, NULL, sizeof(req),
@@ -667,6 +670,11 @@ int bnxt_qplib_reg_mr(struct bnxt_qplib_res *res, struct bnxt_qplib_mrw *mr,
 	if (rc)
 		goto fail;
 
+	if (_is_alloc_mr_unified(res->dattr->dev_cap_flags)) {
+		mr->lkey = le32_to_cpu(resp.xid);
+		mr->rkey = mr->lkey;
+	}
+
 	return 0;
 
 fail:

drivers/infiniband/hw/bnxt_re/qplib_sp.h

@@ -117,6 +117,7 @@ struct bnxt_qplib_mrw {
 	u64				va;
 	u64				total_size;
 	u32				npages;
+	u16				flags;
 	u64				mr_handle;
 	struct bnxt_qplib_hwq		hwq;
 };