RDMA/bnxt_re: Add SRQ support for Broadcom adapters

Shared receive queue (SRQ) is defined as a pool of
receive buffers shared among multiple QPs which belong
to same protection domain in a given process context.
Use of SRQ reduces the memory foot print of IB applications.

Broadcom adapters support SRQ, adding code-changes to enable
shared receive queue.
Signed-off-by: Devesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

drivers/infiniband/hw/bnxt_re/ib_verbs.c

@@ -1027,6 +1027,7 @@ struct ib_qp *bnxt_re_create_qp(struct ib_pd *ib_pd,
 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
 	struct bnxt_re_qp *qp;
 	struct bnxt_re_cq *cq;
+	struct bnxt_re_srq *srq;
 	int rc, entries;
 
 	if ((qp_init_attr->cap.max_send_wr > dev_attr->max_qp_wqes) ||
@@ -1082,9 +1083,15 @@ struct ib_qp *bnxt_re_create_qp(struct ib_pd *ib_pd,
 	}
 
 	if (qp_init_attr->srq) {
-		dev_err(rdev_to_dev(rdev), "SRQ not supported");
-		rc = -ENOTSUPP;
+		srq = container_of(qp_init_attr->srq, struct bnxt_re_srq,
+				   ib_srq);
+		if (!srq) {
+			dev_err(rdev_to_dev(rdev), "SRQ not found");
+			rc = -EINVAL;
 			goto fail;
+		}
+		qp->qplib_qp.srq = &srq->qplib_srq;
+		qp->qplib_qp.rq.max_wqe = 0;
 	} else {
 		/* Allocate 1 more than what's provided so posting max doesn't
 		 * mean empty
@@ -1289,6 +1296,237 @@ static enum ib_mtu __to_ib_mtu(u32 mtu)
 	}
 }
 
+/* Shared Receive Queues */
+int bnxt_re_destroy_srq(struct ib_srq *ib_srq)
+{
+	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
+					       ib_srq);
+	struct bnxt_re_dev *rdev = srq->rdev;
+	struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq;
+	struct bnxt_qplib_nq *nq = NULL;
+	int rc;
+
+	if (qplib_srq->cq)
+		nq = qplib_srq->cq->nq;
+	rc = bnxt_qplib_destroy_srq(&rdev->qplib_res, qplib_srq);
+	if (rc) {
+		dev_err(rdev_to_dev(rdev), "Destroy HW SRQ failed!");
+		return rc;
+	}
+
+	if (srq->umem && !IS_ERR(srq->umem))
+		ib_umem_release(srq->umem);
+	kfree(srq);
+	atomic_dec(&rdev->srq_count);
+	if (nq)
+		nq->budget--;
+	return 0;
+}
+
+static int bnxt_re_init_user_srq(struct bnxt_re_dev *rdev,
+				 struct bnxt_re_pd *pd,
+				 struct bnxt_re_srq *srq,
+				 struct ib_udata *udata)
+{
+	struct bnxt_re_srq_req ureq;
+	struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq;
+	struct ib_umem *umem;
+	int bytes = 0;
+	struct ib_ucontext *context = pd->ib_pd.uobject->context;
+	struct bnxt_re_ucontext *cntx = container_of(context,
+						     struct bnxt_re_ucontext,
+						     ib_uctx);
+	if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
+		return -EFAULT;
+
+	bytes = (qplib_srq->max_wqe * BNXT_QPLIB_MAX_RQE_ENTRY_SIZE);
+	bytes = PAGE_ALIGN(bytes);
+	umem = ib_umem_get(context, ureq.srqva, bytes,
+			   IB_ACCESS_LOCAL_WRITE, 1);
+	if (IS_ERR(umem))
+		return PTR_ERR(umem);
+
+	srq->umem = umem;
+	qplib_srq->nmap = umem->nmap;
+	qplib_srq->sglist = umem->sg_head.sgl;
+	qplib_srq->srq_handle = ureq.srq_handle;
+	qplib_srq->dpi = &cntx->dpi;
+
+	return 0;
+}
+
+struct ib_srq *bnxt_re_create_srq(struct ib_pd *ib_pd,
+				  struct ib_srq_init_attr *srq_init_attr,
+				  struct ib_udata *udata)
+{
+	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
+	struct bnxt_re_dev *rdev = pd->rdev;
+	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
+	struct bnxt_re_srq *srq;
+	struct bnxt_qplib_nq *nq = NULL;
+	int rc, entries;
+
+	if (srq_init_attr->attr.max_wr >= dev_attr->max_srq_wqes) {
+		dev_err(rdev_to_dev(rdev), "Create CQ failed - max exceeded");
+		rc = -EINVAL;
+		goto exit;
+	}
+
+	if (srq_init_attr->srq_type != IB_SRQT_BASIC) {
+		rc = -ENOTSUPP;
+		goto exit;
+	}
+
+	srq = kzalloc(sizeof(*srq), GFP_KERNEL);
+	if (!srq) {
+		rc = -ENOMEM;
+		goto exit;
+	}
+	srq->rdev = rdev;
+	srq->qplib_srq.pd = &pd->qplib_pd;
+	srq->qplib_srq.dpi = &rdev->dpi_privileged;
+	/* Allocate 1 more than what's provided so posting max doesn't
+	 * mean empty
+	 */
+	entries = roundup_pow_of_two(srq_init_attr->attr.max_wr + 1);
+	if (entries > dev_attr->max_srq_wqes + 1)
+		entries = dev_attr->max_srq_wqes + 1;
+
+	srq->qplib_srq.max_wqe = entries;
+	srq->qplib_srq.max_sge = srq_init_attr->attr.max_sge;
+	srq->qplib_srq.threshold = srq_init_attr->attr.srq_limit;
+	srq->srq_limit = srq_init_attr->attr.srq_limit;
+	srq->qplib_srq.eventq_hw_ring_id = rdev->nq[0].ring_id;
+	nq = &rdev->nq[0];
+
+	if (udata) {
+		rc = bnxt_re_init_user_srq(rdev, pd, srq, udata);
+		if (rc)
+			goto fail;
+	}
+
+	rc = bnxt_qplib_create_srq(&rdev->qplib_res, &srq->qplib_srq);
+	if (rc) {
+		dev_err(rdev_to_dev(rdev), "Create HW SRQ failed!");
+		goto fail;
+	}
+
+	if (udata) {
+		struct bnxt_re_srq_resp resp;
+
+		resp.srqid = srq->qplib_srq.id;
+		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
+		if (rc) {
+			dev_err(rdev_to_dev(rdev), "SRQ copy to udata failed!");
+			bnxt_qplib_destroy_srq(&rdev->qplib_res,
+					       &srq->qplib_srq);
+			goto exit;
+		}
+	}
+	if (nq)
+		nq->budget++;
+	atomic_inc(&rdev->srq_count);
+
+	return &srq->ib_srq;
+
+fail:
+	if (udata && srq->umem && !IS_ERR(srq->umem)) {
+		ib_umem_release(srq->umem);
+		srq->umem = NULL;
+	}
+
+	kfree(srq);
+exit:
+	return ERR_PTR(rc);
+}
+
+int bnxt_re_modify_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr,
+		       enum ib_srq_attr_mask srq_attr_mask,
+		       struct ib_udata *udata)
+{
+	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
+					       ib_srq);
+	struct bnxt_re_dev *rdev = srq->rdev;
+	int rc;
+
+	switch (srq_attr_mask) {
+	case IB_SRQ_MAX_WR:
+		/* SRQ resize is not supported */
+		break;
+	case IB_SRQ_LIMIT:
+		/* Change the SRQ threshold */
+		if (srq_attr->srq_limit > srq->qplib_srq.max_wqe)
+			return -EINVAL;
+
+		srq->qplib_srq.threshold = srq_attr->srq_limit;
+		rc = bnxt_qplib_modify_srq(&rdev->qplib_res, &srq->qplib_srq);
+		if (rc) {
+			dev_err(rdev_to_dev(rdev), "Modify HW SRQ failed!");
+			return rc;
+		}
+		/* On success, update the shadow */
+		srq->srq_limit = srq_attr->srq_limit;
+		/* No need to Build and send response back to udata */
+		break;
+	default:
+		dev_err(rdev_to_dev(rdev),
+			"Unsupported srq_attr_mask 0x%x", srq_attr_mask);
+		return -EINVAL;
+	}
+	return 0;
+}
+
+int bnxt_re_query_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr)
+{
+	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
+					       ib_srq);
+	struct bnxt_re_srq tsrq;
+	struct bnxt_re_dev *rdev = srq->rdev;
+	int rc;
+
+	/* Get live SRQ attr */
+	tsrq.qplib_srq.id = srq->qplib_srq.id;
+	rc = bnxt_qplib_query_srq(&rdev->qplib_res, &tsrq.qplib_srq);
+	if (rc) {
+		dev_err(rdev_to_dev(rdev), "Query HW SRQ failed!");
+		return rc;
+	}
+	srq_attr->max_wr = srq->qplib_srq.max_wqe;
+	srq_attr->max_sge = srq->qplib_srq.max_sge;
+	srq_attr->srq_limit = tsrq.qplib_srq.threshold;
+
+	return 0;
+}
+
+int bnxt_re_post_srq_recv(struct ib_srq *ib_srq, struct ib_recv_wr *wr,
+			  struct ib_recv_wr **bad_wr)
+{
+	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
+					       ib_srq);
+	struct bnxt_qplib_swqe wqe;
+	unsigned long flags;
+	int rc = 0, payload_sz = 0;
+
+	spin_lock_irqsave(&srq->lock, flags);
+	while (wr) {
+		/* Transcribe each ib_recv_wr to qplib_swqe */
+		wqe.num_sge = wr->num_sge;
+		payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe.sg_list,
+					       wr->num_sge);
+		wqe.wr_id = wr->wr_id;
+		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
+
+		rc = bnxt_qplib_post_srq_recv(&srq->qplib_srq, &wqe);
+		if (rc) {
+			*bad_wr = wr;
+			break;
+		}
+		wr = wr->next;
+	}
+	spin_unlock_irqrestore(&srq->lock, flags);
+
+	return rc;
+}
 static int bnxt_re_modify_shadow_qp(struct bnxt_re_dev *rdev,
 				    struct bnxt_re_qp *qp1_qp,
 				    int qp_attr_mask)

drivers/infiniband/hw/bnxt_re/ib_verbs.h

@@ -68,6 +68,15 @@ struct bnxt_re_ah {
 	struct bnxt_qplib_ah	qplib_ah;
 };
 
+struct bnxt_re_srq {
+	struct bnxt_re_dev	*rdev;
+	u32			srq_limit;
+	struct ib_srq		ib_srq;
+	struct bnxt_qplib_srq	qplib_srq;
+	struct ib_umem		*umem;
+	spinlock_t		lock;		/* protect srq */
+};
+
 struct bnxt_re_qp {
 	struct list_head	list;
 	struct bnxt_re_dev	*rdev;
@@ -165,6 +174,16 @@ struct ib_ah *bnxt_re_create_ah(struct ib_pd *pd,
 int bnxt_re_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr);
 int bnxt_re_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr);
 int bnxt_re_destroy_ah(struct ib_ah *ah);
+struct ib_srq *bnxt_re_create_srq(struct ib_pd *pd,
+				  struct ib_srq_init_attr *srq_init_attr,
+				  struct ib_udata *udata);
+int bnxt_re_modify_srq(struct ib_srq *srq, struct ib_srq_attr *srq_attr,
+		       enum ib_srq_attr_mask srq_attr_mask,
+		       struct ib_udata *udata);
+int bnxt_re_query_srq(struct ib_srq *srq, struct ib_srq_attr *srq_attr);
+int bnxt_re_destroy_srq(struct ib_srq *srq);
+int bnxt_re_post_srq_recv(struct ib_srq *srq, struct ib_recv_wr *recv_wr,
+			  struct ib_recv_wr **bad_recv_wr);
 struct ib_qp *bnxt_re_create_qp(struct ib_pd *pd,
 				struct ib_qp_init_attr *qp_init_attr,
 				struct ib_udata *udata);

drivers/infiniband/hw/bnxt_re/main.c

@@ -588,6 +588,12 @@ static int bnxt_re_register_ib(struct bnxt_re_dev *rdev)
 	ibdev->query_ah			= bnxt_re_query_ah;
 	ibdev->destroy_ah		= bnxt_re_destroy_ah;
 
+	ibdev->create_srq		= bnxt_re_create_srq;
+	ibdev->modify_srq		= bnxt_re_modify_srq;
+	ibdev->query_srq		= bnxt_re_query_srq;
+	ibdev->destroy_srq		= bnxt_re_destroy_srq;
+	ibdev->post_srq_recv		= bnxt_re_post_srq_recv;
+
 	ibdev->create_qp		= bnxt_re_create_qp;
 	ibdev->modify_qp		= bnxt_re_modify_qp;
 	ibdev->query_qp			= bnxt_re_query_qp;
@@ -689,10 +695,10 @@ static struct bnxt_re_dev *bnxt_re_dev_add(struct net_device *netdev,
 	return rdev;
 }
 
-static int bnxt_re_aeq_handler(struct bnxt_qplib_rcfw *rcfw,
-			       struct creq_func_event *aeqe)
+static int bnxt_re_handle_unaffi_async_event(struct creq_func_event
+					     *unaffi_async)
 {
-	switch (aeqe->event) {
+	switch (unaffi_async->event) {
 	case CREQ_FUNC_EVENT_EVENT_TX_WQE_ERROR:
 		break;
 	case CREQ_FUNC_EVENT_EVENT_TX_DATA_ERROR:
@@ -721,6 +727,93 @@ static int bnxt_re_aeq_handler(struct bnxt_qplib_rcfw *rcfw,
 	return 0;
 }
 
+static int bnxt_re_handle_qp_async_event(struct creq_qp_event *qp_event,
+					 struct bnxt_re_qp *qp)
+{
+	struct ib_event event;
+
+	memset(&event, 0, sizeof(event));
+	if (qp->qplib_qp.srq) {
+		event.device = &qp->rdev->ibdev;
+		event.element.qp = &qp->ib_qp;
+		event.event = IB_EVENT_QP_LAST_WQE_REACHED;
+	}
+
+	if (event.device && qp->ib_qp.event_handler)
+		qp->ib_qp.event_handler(&event, qp->ib_qp.qp_context);
+
+	return 0;
+}
+
+static int bnxt_re_handle_affi_async_event(struct creq_qp_event *affi_async,
+					   void *obj)
+{
+	int rc = 0;
+	u8 event;
+
+	if (!obj)
+		return rc; /* QP was already dead, still return success */
+
+	event = affi_async->event;
+	if (event == CREQ_QP_EVENT_EVENT_QP_ERROR_NOTIFICATION) {
+		struct bnxt_qplib_qp *lib_qp = obj;
+		struct bnxt_re_qp *qp = container_of(lib_qp, struct bnxt_re_qp,
+						     qplib_qp);
+		rc = bnxt_re_handle_qp_async_event(affi_async, qp);
+	}
+	return rc;
+}
+
+static int bnxt_re_aeq_handler(struct bnxt_qplib_rcfw *rcfw,
+			       void *aeqe, void *obj)
+{
+	struct creq_qp_event *affi_async;
+	struct creq_func_event *unaffi_async;
+	u8 type;
+	int rc;
+
+	type = ((struct creq_base *)aeqe)->type;
+	if (type == CREQ_BASE_TYPE_FUNC_EVENT) {
+		unaffi_async = aeqe;
+		rc = bnxt_re_handle_unaffi_async_event(unaffi_async);
+	} else {
+		affi_async = aeqe;
+		rc = bnxt_re_handle_affi_async_event(affi_async, obj);
+	}
+
+	return rc;
+}
+
+static int bnxt_re_srqn_handler(struct bnxt_qplib_nq *nq,
+				struct bnxt_qplib_srq *handle, u8 event)
+{
+	struct bnxt_re_srq *srq = container_of(handle, struct bnxt_re_srq,
+					       qplib_srq);
+	struct ib_event ib_event;
+	int rc = 0;
+
+	if (!srq) {
+		dev_err(NULL, "%s: SRQ is NULL, SRQN not handled",
+			ROCE_DRV_MODULE_NAME);
+		rc = -EINVAL;
+		goto done;
+	}
+	ib_event.device = &srq->rdev->ibdev;
+	ib_event.element.srq = &srq->ib_srq;
+	if (event == NQ_SRQ_EVENT_EVENT_SRQ_THRESHOLD_EVENT)
+		ib_event.event = IB_EVENT_SRQ_LIMIT_REACHED;
+	else
+		ib_event.event = IB_EVENT_SRQ_ERR;
+
+	if (srq->ib_srq.event_handler) {
+		/* Lock event_handler? */
+		(*srq->ib_srq.event_handler)(&ib_event,
+					     srq->ib_srq.srq_context);
+	}
+done:
+	return rc;
+}
+
 static int bnxt_re_cqn_handler(struct bnxt_qplib_nq *nq,
 			       struct bnxt_qplib_cq *handle)
 {
@@ -763,7 +856,8 @@ static int bnxt_re_init_res(struct bnxt_re_dev *rdev)
 		rc = bnxt_qplib_enable_nq(rdev->en_dev->pdev, &rdev->nq[i - 1],
 					  i - 1, rdev->msix_entries[i].vector,
 					  rdev->msix_entries[i].db_offset,
-					  &bnxt_re_cqn_handler, NULL);
+					  &bnxt_re_cqn_handler,
+					  &bnxt_re_srqn_handler);
 
 		if (rc) {
 			dev_err(rdev_to_dev(rdev),

drivers/infiniband/hw/bnxt_re/qplib_fp.c

@@ -52,6 +52,7 @@
 
 static void bnxt_qplib_arm_cq_enable(struct bnxt_qplib_cq *cq);
 static void __clean_cq(struct bnxt_qplib_cq *cq, u64 qp);
+static void bnxt_qplib_arm_srq(struct bnxt_qplib_srq *srq, u32 arm_type);
 
 static void bnxt_qplib_cancel_phantom_processing(struct bnxt_qplib_qp *qp)
 {
@@ -278,6 +279,7 @@ static void bnxt_qplib_service_nq(unsigned long data)
 	struct nq_base *nqe, **nq_ptr;
 	struct bnxt_qplib_cq *cq;
 	int num_cqne_processed = 0;
+	int num_srqne_processed = 0;
 	u32 sw_cons, raw_cons;
 	u16 type;
 	int budget = nq->budget;
@@ -320,6 +322,26 @@ static void bnxt_qplib_service_nq(unsigned long data)
 			spin_unlock_bh(&cq->compl_lock);
 			break;
 		}
+		case NQ_BASE_TYPE_SRQ_EVENT:
+		{
+			struct nq_srq_event *nqsrqe =
+						(struct nq_srq_event *)nqe;
+
+			q_handle = le32_to_cpu(nqsrqe->srq_handle_low);
+			q_handle |= (u64)le32_to_cpu(nqsrqe->srq_handle_high)
+				     << 32;
+			bnxt_qplib_arm_srq((struct bnxt_qplib_srq *)q_handle,
+					   DBR_DBR_TYPE_SRQ_ARMENA);
+			if (!nq->srqn_handler(nq,
+					      (struct bnxt_qplib_srq *)q_handle,
+					      nqsrqe->event))
+				num_srqne_processed++;
+			else
+				dev_warn(&nq->pdev->dev,
+					 "QPLIB: SRQ event 0x%x not handled",
+					 nqsrqe->event);
+			break;
+		}
 		case NQ_BASE_TYPE_DBQ_EVENT:
 			break;
 		default:
@@ -384,16 +406,18 @@ int bnxt_qplib_enable_nq(struct pci_dev *pdev, struct bnxt_qplib_nq *nq,
 			 int (*cqn_handler)(struct bnxt_qplib_nq *nq,
 					    struct bnxt_qplib_cq *),
 			 int (*srqn_handler)(struct bnxt_qplib_nq *nq,
-					     void *, u8 event))
+					     struct bnxt_qplib_srq *,
+					     u8 event))
 {
 	resource_size_t nq_base;
 	int rc = -1;
 
 	nq->pdev = pdev;
 	nq->vector = msix_vector;
-
+	if (cqn_handler)
 		nq->cqn_handler = cqn_handler;
 
+	if (srqn_handler)
 		nq->srqn_handler = srqn_handler;
 
 	tasklet_init(&nq->worker, bnxt_qplib_service_nq, (unsigned long)nq);
@@ -468,6 +492,238 @@ int bnxt_qplib_alloc_nq(struct pci_dev *pdev, struct bnxt_qplib_nq *nq)
 	return 0;
 }
 
+/* SRQ */
+static void bnxt_qplib_arm_srq(struct bnxt_qplib_srq *srq, u32 arm_type)
+{
+	struct bnxt_qplib_hwq *srq_hwq = &srq->hwq;
+	struct dbr_dbr db_msg = { 0 };
+	void __iomem *db;
+	u32 sw_prod = 0;
+
+	/* Ring DB */
+	sw_prod = (arm_type == DBR_DBR_TYPE_SRQ_ARM) ? srq->threshold :
+		   HWQ_CMP(srq_hwq->prod, srq_hwq);
+	db_msg.index = cpu_to_le32((sw_prod << DBR_DBR_INDEX_SFT) &
+				   DBR_DBR_INDEX_MASK);
+	db_msg.type_xid = cpu_to_le32(((srq->id << DBR_DBR_XID_SFT) &
+					DBR_DBR_XID_MASK) | arm_type);
+	db = (arm_type == DBR_DBR_TYPE_SRQ_ARMENA) ?
+		srq->dbr_base : srq->dpi->dbr;
+	wmb(); /* barrier before db ring */
+	__iowrite64_copy(db, &db_msg, sizeof(db_msg) / sizeof(u64));
+}
+
+int bnxt_qplib_destroy_srq(struct bnxt_qplib_res *res,
+			   struct bnxt_qplib_srq *srq)
+{
+	struct bnxt_qplib_rcfw *rcfw = res->rcfw;
+	struct cmdq_destroy_srq req;
+	struct creq_destroy_srq_resp resp;
+	u16 cmd_flags = 0;
+	int rc;
+
+	RCFW_CMD_PREP(req, DESTROY_SRQ, cmd_flags);
+
+	/* Configure the request */
+	req.srq_cid = cpu_to_le32(srq->id);
+
+	rc = bnxt_qplib_rcfw_send_message(rcfw, (void *)&req,
+					  (void *)&resp, NULL, 0);
+	if (rc)
+		return rc;
+
+	bnxt_qplib_free_hwq(res->pdev, &srq->hwq);
+	kfree(srq->swq);
+	return 0;
+}
+
+int bnxt_qplib_create_srq(struct bnxt_qplib_res *res,
+			  struct bnxt_qplib_srq *srq)
+{
+	struct bnxt_qplib_rcfw *rcfw = res->rcfw;
+	struct cmdq_create_srq req;
+	struct creq_create_srq_resp resp;
+	struct bnxt_qplib_pbl *pbl;
+	u16 cmd_flags = 0;
+	int rc, idx;
+
+	srq->hwq.max_elements = srq->max_wqe;
+	rc = bnxt_qplib_alloc_init_hwq(res->pdev, &srq->hwq, srq->sglist,
+				       srq->nmap, &srq->hwq.max_elements,
+				       BNXT_QPLIB_MAX_RQE_ENTRY_SIZE, 0,
+				       PAGE_SIZE, HWQ_TYPE_QUEUE);
+	if (rc)
+		goto exit;
+
+	srq->swq = kcalloc(srq->hwq.max_elements, sizeof(*srq->swq),
+			   GFP_KERNEL);
+	if (!srq->swq)
+		goto fail;
+
+	RCFW_CMD_PREP(req, CREATE_SRQ, cmd_flags);
+
+	/* Configure the request */
+	req.dpi = cpu_to_le32(srq->dpi->dpi);
+	req.srq_handle = cpu_to_le64(srq);
+
+	req.srq_size = cpu_to_le16((u16)srq->hwq.max_elements);
+	pbl = &srq->hwq.pbl[PBL_LVL_0];
+	req.pg_size_lvl = cpu_to_le16((((u16)srq->hwq.level &
+				      CMDQ_CREATE_SRQ_LVL_MASK) <<
+				      CMDQ_CREATE_SRQ_LVL_SFT) |
+				      (pbl->pg_size == ROCE_PG_SIZE_4K ?
+				       CMDQ_CREATE_SRQ_PG_SIZE_PG_4K :
+				       pbl->pg_size == ROCE_PG_SIZE_8K ?
+				       CMDQ_CREATE_SRQ_PG_SIZE_PG_8K :
+				       pbl->pg_size == ROCE_PG_SIZE_64K ?
+				       CMDQ_CREATE_SRQ_PG_SIZE_PG_64K :
+				       pbl->pg_size == ROCE_PG_SIZE_2M ?
+				       CMDQ_CREATE_SRQ_PG_SIZE_PG_2M :
+				       pbl->pg_size == ROCE_PG_SIZE_8M ?
+				       CMDQ_CREATE_SRQ_PG_SIZE_PG_8M :
+				       pbl->pg_size == ROCE_PG_SIZE_1G ?
+				       CMDQ_CREATE_SRQ_PG_SIZE_PG_1G :
+				       CMDQ_CREATE_SRQ_PG_SIZE_PG_4K));
+	req.pbl = cpu_to_le64(pbl->pg_map_arr[0]);
+	req.pd_id = cpu_to_le32(srq->pd->id);
+	req.eventq_id = cpu_to_le16(srq->eventq_hw_ring_id);
+
+	rc = bnxt_qplib_rcfw_send_message(rcfw, (void *)&req,
+					  (void *)&resp, NULL, 0);
+	if (rc)
+		goto fail;
+
+	spin_lock_init(&srq->lock);
+	srq->start_idx = 0;
+	srq->last_idx = srq->hwq.max_elements - 1;
+	for (idx = 0; idx < srq->hwq.max_elements; idx++)
+		srq->swq[idx].next_idx = idx + 1;
+	srq->swq[srq->last_idx].next_idx = -1;
+
+	srq->id = le32_to_cpu(resp.xid);
+	srq->dbr_base = res->dpi_tbl.dbr_bar_reg_iomem;
+	if (srq->threshold)
+		bnxt_qplib_arm_srq(srq, DBR_DBR_TYPE_SRQ_ARMENA);
+	srq->arm_req = false;
+
+	return 0;
+fail:
+	bnxt_qplib_free_hwq(res->pdev, &srq->hwq);
+	kfree(srq->swq);
+exit:
+	return rc;
+}
+
+int bnxt_qplib_modify_srq(struct bnxt_qplib_res *res,
+			  struct bnxt_qplib_srq *srq)
+{
+	struct bnxt_qplib_hwq *srq_hwq = &srq->hwq;
+	u32 sw_prod, sw_cons, count = 0;
+
+	sw_prod = HWQ_CMP(srq_hwq->prod, srq_hwq);
+	sw_cons = HWQ_CMP(srq_hwq->cons, srq_hwq);
+
+	count = sw_prod > sw_cons ? sw_prod - sw_cons :
+				    srq_hwq->max_elements - sw_cons + sw_prod;
+	if (count > srq->threshold) {
+		srq->arm_req = false;
+		bnxt_qplib_arm_srq(srq, DBR_DBR_TYPE_SRQ_ARM);
+	} else {
+		/* Deferred arming */
+		srq->arm_req = true;
+	}
+
+	return 0;
+}
+
+int bnxt_qplib_query_srq(struct bnxt_qplib_res *res,
+			 struct bnxt_qplib_srq *srq)
+{
+	struct bnxt_qplib_rcfw *rcfw = res->rcfw;
+	struct cmdq_query_srq req;
+	struct creq_query_srq_resp resp;
+	struct bnxt_qplib_rcfw_sbuf *sbuf;
+	struct creq_query_srq_resp_sb *sb;
+	u16 cmd_flags = 0;
+	int rc = 0;
+
+	RCFW_CMD_PREP(req, QUERY_SRQ, cmd_flags);
+	req.srq_cid = cpu_to_le32(srq->id);
+
+	/* Configure the request */
+	sbuf = bnxt_qplib_rcfw_alloc_sbuf(rcfw, sizeof(*sb));
+	if (!sbuf)
+		return -ENOMEM;
+	sb = sbuf->sb;
+	rc = bnxt_qplib_rcfw_send_message(rcfw, (void *)&req, (void *)&resp,
+					  (void *)sbuf, 0);
+	srq->threshold = le16_to_cpu(sb->srq_limit);
+	bnxt_qplib_rcfw_free_sbuf(rcfw, sbuf);
+
+	return rc;
+}
+
+int bnxt_qplib_post_srq_recv(struct bnxt_qplib_srq *srq,
+			     struct bnxt_qplib_swqe *wqe)
+{
+	struct bnxt_qplib_hwq *srq_hwq = &srq->hwq;
+	struct rq_wqe *srqe, **srqe_ptr;
+	struct sq_sge *hw_sge;
+	u32 sw_prod, sw_cons, count = 0;
+	int i, rc = 0, next;
+
+	spin_lock(&srq_hwq->lock);
+	if (srq->start_idx == srq->last_idx) {
+		dev_err(&srq_hwq->pdev->dev, "QPLIB: FP: SRQ (0x%x) is full!",
+			srq->id);
+		rc = -EINVAL;
+		spin_unlock(&srq_hwq->lock);
+		goto done;
+	}
+	next = srq->start_idx;
+	srq->start_idx = srq->swq[next].next_idx;
+	spin_unlock(&srq_hwq->lock);
+
+	sw_prod = HWQ_CMP(srq_hwq->prod, srq_hwq);
+	srqe_ptr = (struct rq_wqe **)srq_hwq->pbl_ptr;
+	srqe = &srqe_ptr[RQE_PG(sw_prod)][RQE_IDX(sw_prod)];
+	memset(srqe, 0, BNXT_QPLIB_MAX_RQE_ENTRY_SIZE);
+	/* Calculate wqe_size16 and data_len */
+	for (i = 0, hw_sge = (struct sq_sge *)srqe->data;
+	     i < wqe->num_sge; i++, hw_sge++) {
+		hw_sge->va_or_pa = cpu_to_le64(wqe->sg_list[i].addr);
+		hw_sge->l_key = cpu_to_le32(wqe->sg_list[i].lkey);
+		hw_sge->size = cpu_to_le32(wqe->sg_list[i].size);
+	}
+	srqe->wqe_type = wqe->type;
+	srqe->flags = wqe->flags;
+	srqe->wqe_size = wqe->num_sge +
+			((offsetof(typeof(*srqe), data) + 15) >> 4);
+	srqe->wr_id[0] = cpu_to_le32((u32)next);
+	srq->swq[next].wr_id = wqe->wr_id;
+
+	srq_hwq->prod++;
+
+	spin_lock(&srq_hwq->lock);
+	sw_prod = HWQ_CMP(srq_hwq->prod, srq_hwq);
+	/* retaining srq_hwq->cons for this logic
+	 * actually the lock is only required to
+	 * read srq_hwq->cons.
+	 */
+	sw_cons = HWQ_CMP(srq_hwq->cons, srq_hwq);
+	count = sw_prod > sw_cons ? sw_prod - sw_cons :
+				    srq_hwq->max_elements - sw_cons + sw_prod;
+	spin_unlock(&srq_hwq->lock);
+	/* Ring DB */
+	bnxt_qplib_arm_srq(srq, DBR_DBR_TYPE_SRQ);
+	if (srq->arm_req == true && count > srq->threshold) {
+		srq->arm_req = false;
+		bnxt_qplib_arm_srq(srq, DBR_DBR_TYPE_SRQ_ARM);
+	}
+done:
+	return rc;
+}
+
 /* QP */
 int bnxt_qplib_create_qp1(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp)
 {
@@ -736,6 +992,12 @@ int bnxt_qplib_create_qp(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp)
 				 pbl->pg_size == ROCE_PG_SIZE_1G ?
 					CMDQ_CREATE_QP_RQ_PG_SIZE_PG_1G :
 				 CMDQ_CREATE_QP_RQ_PG_SIZE_PG_4K);
+	} else {
+		/* SRQ */
+		if (qp->srq) {
+			qp_flags |= CMDQ_CREATE_QP_QP_FLAGS_SRQ_USED;
+			req.srq_cid = cpu_to_le32(qp->srq->id);
+		}
 	}
 
 	if (qp->rcq)
@@ -2067,6 +2329,16 @@ static int bnxt_qplib_cq_process_req(struct bnxt_qplib_cq *cq,
 	return rc;
 }
 
+static void bnxt_qplib_release_srqe(struct bnxt_qplib_srq *srq, u32 tag)
+{
+	spin_lock(&srq->hwq.lock);
+	srq->swq[srq->last_idx].next_idx = (int)tag;
+	srq->last_idx = (int)tag;
+	srq->swq[srq->last_idx].next_idx = -1;
+	srq->hwq.cons++; /* Support for SRQE counter */
+	spin_unlock(&srq->hwq.lock);
+}
+
 static int bnxt_qplib_cq_process_res_rc(struct bnxt_qplib_cq *cq,
 					struct cq_res_rc *hwcqe,
 					struct bnxt_qplib_cqe **pcqe,
@@ -2074,6 +2346,7 @@ static int bnxt_qplib_cq_process_res_rc(struct bnxt_qplib_cq *cq,
 {
 	struct bnxt_qplib_qp *qp;
 	struct bnxt_qplib_q *rq;
+	struct bnxt_qplib_srq *srq;
 	struct bnxt_qplib_cqe *cqe;
 	u32 wr_id_idx;
 	int rc = 0;
@@ -2101,15 +2374,33 @@ static int bnxt_qplib_cq_process_res_rc(struct bnxt_qplib_cq *cq,
 
 	wr_id_idx = le32_to_cpu(hwcqe->srq_or_rq_wr_id) &
 				CQ_RES_RC_SRQ_OR_RQ_WR_ID_MASK;
+	if (cqe->flags & CQ_RES_RC_FLAGS_SRQ_SRQ) {
+		srq = qp->srq;
+		if (!srq)
+			return -EINVAL;
+		if (wr_id_idx > srq->hwq.max_elements) {
+			dev_err(&cq->hwq.pdev->dev,
+				"QPLIB: FP: CQ Process RC ");
+			dev_err(&cq->hwq.pdev->dev,
+				"QPLIB: wr_id idx 0x%x exceeded SRQ max 0x%x",
+				wr_id_idx, srq->hwq.max_elements);
+			return -EINVAL;
+		}
+		cqe->wr_id = srq->swq[wr_id_idx].wr_id;
+		bnxt_qplib_release_srqe(srq, wr_id_idx);
+		cqe++;
+		(*budget)--;
+		*pcqe = cqe;
+	} else {
 		rq = &qp->rq;
 		if (wr_id_idx > rq->hwq.max_elements) {
-		dev_err(&cq->hwq.pdev->dev, "QPLIB: FP: CQ Process RC ");
+			dev_err(&cq->hwq.pdev->dev,
+				"QPLIB: FP: CQ Process RC ");
 			dev_err(&cq->hwq.pdev->dev,
 				"QPLIB: wr_id idx 0x%x exceeded RQ max 0x%x",
 				wr_id_idx, rq->hwq.max_elements);
 			return -EINVAL;
 		}
-
 		cqe->wr_id = rq->swq[wr_id_idx].wr_id;
 		cqe++;
 		(*budget)--;
@@ -2123,6 +2414,7 @@ static int bnxt_qplib_cq_process_res_rc(struct bnxt_qplib_cq *cq,
 			__bnxt_qplib_add_flush_qp(qp);
 			bnxt_qplib_unlock_buddy_cq(qp, cq);
 		}
+	}
 
 done:
 	return rc;
@@ -2135,6 +2427,7 @@ static int bnxt_qplib_cq_process_res_ud(struct bnxt_qplib_cq *cq,
 {
 	struct bnxt_qplib_qp *qp;
 	struct bnxt_qplib_q *rq;
+	struct bnxt_qplib_srq *srq;
 	struct bnxt_qplib_cqe *cqe;
 	u32 wr_id_idx;
 	int rc = 0;
@@ -2165,11 +2458,31 @@ static int bnxt_qplib_cq_process_res_ud(struct bnxt_qplib_cq *cq,
 				  hwcqe->src_qp_high_srq_or_rq_wr_id) &
 				 CQ_RES_UD_SRC_QP_HIGH_MASK) >> 8);
 
+	if (cqe->flags & CQ_RES_RC_FLAGS_SRQ_SRQ) {
+		srq = qp->srq;
+		if (!srq)
+			return -EINVAL;
+
+		if (wr_id_idx > srq->hwq.max_elements) {
+			dev_err(&cq->hwq.pdev->dev,
+				"QPLIB: FP: CQ Process UD ");
+			dev_err(&cq->hwq.pdev->dev,
+				"QPLIB: wr_id idx 0x%x exceeded SRQ max 0x%x",
+				wr_id_idx, srq->hwq.max_elements);
+			return -EINVAL;
+		}
+		cqe->wr_id = srq->swq[wr_id_idx].wr_id;
+		bnxt_qplib_release_srqe(srq, wr_id_idx);
+		cqe++;
+		(*budget)--;
+		*pcqe = cqe;
+	} else {
 		rq = &qp->rq;
 		if (wr_id_idx > rq->hwq.max_elements) {
-		dev_err(&cq->hwq.pdev->dev, "QPLIB: FP: CQ Process UD ");
 			dev_err(&cq->hwq.pdev->dev,
-			"QPLIB: wr_id idx %#x exceeded RQ max %#x",
+				"QPLIB: FP: CQ Process UD ");
+			dev_err(&cq->hwq.pdev->dev,
+				"QPLIB: wr_id idx 0x%x exceeded RQ max 0x%x",
 				wr_id_idx, rq->hwq.max_elements);
 			return -EINVAL;
 		}
@@ -2187,6 +2500,7 @@ static int bnxt_qplib_cq_process_res_ud(struct bnxt_qplib_cq *cq,
 			__bnxt_qplib_add_flush_qp(qp);
 			bnxt_qplib_unlock_buddy_cq(qp, cq);
 		}
+	}
 done:
 	return rc;
 }
@@ -2217,6 +2531,7 @@ static int bnxt_qplib_cq_process_res_raweth_qp1(struct bnxt_qplib_cq *cq,
 {
 	struct bnxt_qplib_qp *qp;
 	struct bnxt_qplib_q *rq;
+	struct bnxt_qplib_srq *srq;
 	struct bnxt_qplib_cqe *cqe;
 	u32 wr_id_idx;
 	int rc = 0;
@@ -2255,14 +2570,36 @@ static int bnxt_qplib_cq_process_res_raweth_qp1(struct bnxt_qplib_cq *cq,
 	cqe->raweth_qp1_flags2 = le32_to_cpu(hwcqe->raweth_qp1_flags2);
 	cqe->raweth_qp1_metadata = le32_to_cpu(hwcqe->raweth_qp1_metadata);
 
+	if (cqe->flags & CQ_RES_RAWETH_QP1_FLAGS_SRQ_SRQ) {
+		srq = qp->srq;
+		if (!srq) {
+			dev_err(&cq->hwq.pdev->dev,
+				"QPLIB: FP: SRQ used but not defined??");
+			return -EINVAL;
+		}
+		if (wr_id_idx > srq->hwq.max_elements) {
+			dev_err(&cq->hwq.pdev->dev,
+				"QPLIB: FP: CQ Process Raw/QP1 ");
+			dev_err(&cq->hwq.pdev->dev,
+				"QPLIB: wr_id idx 0x%x exceeded SRQ max 0x%x",
+				wr_id_idx, srq->hwq.max_elements);
+			return -EINVAL;
+		}
+		cqe->wr_id = srq->swq[wr_id_idx].wr_id;
+		bnxt_qplib_release_srqe(srq, wr_id_idx);
+		cqe++;
+		(*budget)--;
+		*pcqe = cqe;
+	} else {
 		rq = &qp->rq;
 		if (wr_id_idx > rq->hwq.max_elements) {
-		dev_err(&cq->hwq.pdev->dev, "QPLIB: FP: CQ Process Raw/QP1 RQ wr_id ");
-		dev_err(&cq->hwq.pdev->dev, "QPLIB: ix 0x%x exceeded RQ max 0x%x",
+			dev_err(&cq->hwq.pdev->dev,
+				"QPLIB: FP: CQ Process Raw/QP1 RQ wr_id ");
+			dev_err(&cq->hwq.pdev->dev,
+				"QPLIB: ix 0x%x exceeded RQ max 0x%x",
 				wr_id_idx, rq->hwq.max_elements);
 			return -EINVAL;
 		}
-
 		cqe->wr_id = rq->swq[wr_id_idx].wr_id;
 		cqe++;
 		(*budget)--;
@@ -2276,6 +2613,7 @@ static int bnxt_qplib_cq_process_res_raweth_qp1(struct bnxt_qplib_cq *cq,
 			__bnxt_qplib_add_flush_qp(qp);
 			bnxt_qplib_unlock_buddy_cq(qp, cq);
 		}
+	}
 
 done:
 	return rc;

drivers/infiniband/hw/bnxt_re/qplib_fp.h

@@ -39,6 +39,27 @@
 #ifndef __BNXT_QPLIB_FP_H__
 #define __BNXT_QPLIB_FP_H__
 
+struct bnxt_qplib_srq {
+	struct bnxt_qplib_pd		*pd;
+	struct bnxt_qplib_dpi		*dpi;
+	void __iomem			*dbr_base;
+	u64				srq_handle;
+	u32				id;
+	u32				max_wqe;
+	u32				max_sge;
+	u32				threshold;
+	bool				arm_req;
+	struct bnxt_qplib_cq		*cq;
+	struct bnxt_qplib_hwq		hwq;
+	struct bnxt_qplib_swq		*swq;
+	struct scatterlist		*sglist;
+	int				start_idx;
+	int				last_idx;
+	u32				nmap;
+	u16				eventq_hw_ring_id;
+	spinlock_t			lock; /* protect SRQE link list */
+};
+
 struct bnxt_qplib_sge {
 	u64				addr;
 	u32				lkey;
@@ -79,6 +100,7 @@ static inline u32 get_psne_idx(u32 val)
 
 struct bnxt_qplib_swq {
 	u64				wr_id;
+	int				next_idx;
 	u8				type;
 	u8				flags;
 	u32				start_psn;
@@ -418,12 +440,10 @@ struct bnxt_qplib_nq {
 	u16			ring_id;
 	void __iomem		*bar_reg_iomem;
 
-	int				(*cqn_handler)
-						(struct bnxt_qplib_nq *nq,
+	int			(*cqn_handler)(struct bnxt_qplib_nq *nq,
 					       struct bnxt_qplib_cq *cq);
-	int				(*srqn_handler)
-						(struct bnxt_qplib_nq *nq,
-						 void *srq,
+	int			(*srqn_handler)(struct bnxt_qplib_nq *nq,
+						struct bnxt_qplib_srq *srq,
 						u8 event);
 	struct workqueue_struct	*cqn_wq;
 	char			name[32];
@@ -441,8 +461,18 @@ int bnxt_qplib_enable_nq(struct pci_dev *pdev, struct bnxt_qplib_nq *nq,
 			 int (*cqn_handler)(struct bnxt_qplib_nq *nq,
 					    struct bnxt_qplib_cq *cq),
 			 int (*srqn_handler)(struct bnxt_qplib_nq *nq,
-					     void *srq,
+					     struct bnxt_qplib_srq *srq,
 					     u8 event));
+int bnxt_qplib_create_srq(struct bnxt_qplib_res *res,
+			  struct bnxt_qplib_srq *srq);
+int bnxt_qplib_modify_srq(struct bnxt_qplib_res *res,
+			  struct bnxt_qplib_srq *srq);
+int bnxt_qplib_query_srq(struct bnxt_qplib_res *res,
+			 struct bnxt_qplib_srq *srq);
+int bnxt_qplib_destroy_srq(struct bnxt_qplib_res *res,
+			   struct bnxt_qplib_srq *srq);
+int bnxt_qplib_post_srq_recv(struct bnxt_qplib_srq *srq,
+			     struct bnxt_qplib_swqe *wqe);
 int bnxt_qplib_create_qp1(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp);
 int bnxt_qplib_create_qp(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp);
 int bnxt_qplib_modify_qp(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp);

drivers/infiniband/hw/bnxt_re/qplib_rcfw.c

@@ -616,7 +616,7 @@ int bnxt_qplib_enable_rcfw_channel(struct pci_dev *pdev,
 				   int msix_vector,
 				   int cp_bar_reg_off, int virt_fn,
 				   int (*aeq_handler)(struct bnxt_qplib_rcfw *,
-						      struct creq_func_event *))
+						      void *, void *))
 {
 	resource_size_t res_base;
 	struct cmdq_init init;

drivers/infiniband/hw/bnxt_re/qplib_rcfw.h

@@ -167,7 +167,7 @@ struct bnxt_qplib_rcfw {
 #define FIRMWARE_TIMED_OUT		3
 	wait_queue_head_t	waitq;
 	int			(*aeq_handler)(struct bnxt_qplib_rcfw *,
-					       struct creq_func_event *);
+					       void *, void *);
 	u32			seq_num;
 
 	/* Bar region info */
@@ -199,9 +199,8 @@ int bnxt_qplib_enable_rcfw_channel(struct pci_dev *pdev,
 				   struct bnxt_qplib_rcfw *rcfw,
 				   int msix_vector,
 				   int cp_bar_reg_off, int virt_fn,
-				   int (*aeq_handler)
-					(struct bnxt_qplib_rcfw *,
-					 struct creq_func_event *));
+				   int (*aeq_handler)(struct bnxt_qplib_rcfw *,
+						      void *aeqe, void *obj));
 
 struct bnxt_qplib_rcfw_sbuf *bnxt_qplib_rcfw_alloc_sbuf(
 				struct bnxt_qplib_rcfw *rcfw,

include/uapi/rdma/bnxt_re-abi.h

@@ -82,6 +82,15 @@ struct bnxt_re_qp_resp {
 	__u32 rsvd;
 };
 
+struct bnxt_re_srq_req {
+	__u64 srqva;
+	__u64 srq_handle;
+};
+
+struct bnxt_re_srq_resp {
+	__u32 srqid;
+};
+
 enum bnxt_re_shpg_offt {
 	BNXT_RE_BEG_RESV_OFFT	= 0x00,
 	BNXT_RE_AVID_OFFT	= 0x10,