diff mbox

[qedr,09/10] qedr: Add GSI support

Message ID 1475682483-9878-10-git-send-email-Ram.Amrani@cavium.com (mailing list archive)
State Accepted
Headers show

Commit Message

Amrani, Ram Oct. 5, 2016, 3:48 p.m. UTC
Add support for GSI over light L2.

Signed-off-by: Rajesh Borundia <rajesh.borundia@cavium.com>
Signed-off-by: Ram Amrani <Ram.Amrani@cavium.com>
---
 drivers/infiniband/hw/qedr/Makefile  |   2 +-
 drivers/infiniband/hw/qedr/main.c    |   3 +
 drivers/infiniband/hw/qedr/qedr.h    |  15 +
 drivers/infiniband/hw/qedr/qedr_cm.c | 636 +++++++++++++++++++++++++++++++++++
 drivers/infiniband/hw/qedr/qedr_cm.h |  21 ++
 drivers/infiniband/hw/qedr/verbs.c   |  37 ++
 6 files changed, 713 insertions(+), 1 deletion(-)
 create mode 100644 drivers/infiniband/hw/qedr/qedr_cm.c
diff mbox

Patch

diff --git a/drivers/infiniband/hw/qedr/Makefile b/drivers/infiniband/hw/qedr/Makefile
index b10f2b1..ba7067c 100644
--- a/drivers/infiniband/hw/qedr/Makefile
+++ b/drivers/infiniband/hw/qedr/Makefile
@@ -1,3 +1,3 @@ 
 obj-$(CONFIG_INFINIBAND_QEDR) := qedr.o
 
-qedr-y := main.o verbs.o
+qedr-y := main.o verbs.o qedr_cm.o
diff --git a/drivers/infiniband/hw/qedr/main.c b/drivers/infiniband/hw/qedr/main.c
index d423f72..306a393 100644
--- a/drivers/infiniband/hw/qedr/main.c
+++ b/drivers/infiniband/hw/qedr/main.c
@@ -138,6 +138,9 @@  static int qedr_register_device(struct qedr_dev *dev)
 
 	dev->ibdev.query_pkey = qedr_query_pkey;
 
+	dev->ibdev.create_ah = qedr_create_ah;
+	dev->ibdev.destroy_ah = qedr_destroy_ah;
+
 	dev->ibdev.get_dma_mr = qedr_get_dma_mr;
 	dev->ibdev.dereg_mr = qedr_dereg_mr;
 	dev->ibdev.reg_user_mr = qedr_reg_user_mr;
diff --git a/drivers/infiniband/hw/qedr/qedr.h b/drivers/infiniband/hw/qedr/qedr.h
index 9e282fe..bc6ce29 100644
--- a/drivers/infiniband/hw/qedr/qedr.h
+++ b/drivers/infiniband/hw/qedr/qedr.h
@@ -55,6 +55,7 @@ 
 #define QEDR_MSG_RQ   "  RQ"
 #define QEDR_MSG_SQ   "  SQ"
 #define QEDR_MSG_QP   "  QP"
+#define QEDR_MSG_GSI  " GSI"
 
 #define QEDR_CQ_MAGIC_NUMBER   (0x11223344)
 
@@ -148,6 +149,10 @@  struct qedr_dev {
 	u8			num_hwfns;
 	uint			wq_multiplier;
 	u8			gsi_ll2_mac_address[ETH_ALEN];
+	int			gsi_qp_created;
+	struct qedr_cq		*gsi_sqcq;
+	struct qedr_cq		*gsi_rqcq;
+	struct qedr_qp		*gsi_qp;
 };
 
 #define QEDR_MAX_SQ_PBL			(0x8000)
@@ -246,6 +251,9 @@  struct qedr_cq {
 
 	u16 icid;
 
+	/* Lock to protect completion handler */
+	spinlock_t comp_handler_lock;
+
 	/* Lock to protect multiplem CQ's */
 	spinlock_t cq_lock;
 	u8 arm_flags;
@@ -292,6 +300,7 @@  struct qedr_qp_hwq_info {
 	u16 prod;
 	u16 cons;
 	u16 wqe_cons;
+	u16 gsi_cons;
 	u16 max_wr;
 
 	/* DB */
@@ -366,6 +375,7 @@  struct qedr_qp {
 		struct ib_sge sg_list[RDMA_MAX_SGE_PER_RQ_WQE];
 		u8 wqe_size;
 
+		u8 smac[ETH_ALEN];
 		u16 vlan_id;
 		int rc;
 	} *rqe_wr_id;
@@ -473,6 +483,11 @@  static inline struct qedr_qp *get_qedr_qp(struct ib_qp *ibqp)
 	return container_of(ibqp, struct qedr_qp, ibqp);
 }
 
+static inline struct qedr_ah *get_qedr_ah(struct ib_ah *ibah)
+{
+	return container_of(ibah, struct qedr_ah, ibah);
+}
+
 static inline struct qedr_mr *get_qedr_mr(struct ib_mr *ibmr)
 {
 	return container_of(ibmr, struct qedr_mr, ibmr);
diff --git a/drivers/infiniband/hw/qedr/qedr_cm.c b/drivers/infiniband/hw/qedr/qedr_cm.c
new file mode 100644
index 0000000..2af1395
--- /dev/null
+++ b/drivers/infiniband/hw/qedr/qedr_cm.c
@@ -0,0 +1,636 @@ 
+/* QLogic qedr NIC Driver
+ * Copyright (c) 2015-2016  QLogic Corporation
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and /or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include <linux/dma-mapping.h>
+#include <linux/crc32.h>
+#include <linux/iommu.h>
+#include <net/ip.h>
+#include <net/ipv6.h>
+#include <net/udp.h>
+
+#include <rdma/ib_verbs.h>
+#include <rdma/ib_user_verbs.h>
+#include <rdma/iw_cm.h>
+#include <rdma/ib_umem.h>
+#include <rdma/ib_addr.h>
+#include <rdma/ib_cache.h>
+
+#include "qedr_hsi.h"
+#include <linux/qed/qed_if.h>
+#include <linux/qed/qed_roce_if.h>
+#include "qedr.h"
+#include "qedr_hsi.h"
+#include "verbs.h"
+#include <rdma/qedr-abi.h>
+#include "qedr_hsi.h"
+#include "qedr_cm.h"
+
+void qedr_inc_sw_gsi_cons(struct qedr_qp_hwq_info *info)
+{
+	info->gsi_cons = (info->gsi_cons + 1) % info->max_wr;
+}
+
+void qedr_store_gsi_qp_cq(struct qedr_dev *dev, struct qedr_qp *qp,
+			  struct ib_qp_init_attr *attrs)
+{
+	dev->gsi_qp_created = 1;
+	dev->gsi_sqcq = get_qedr_cq(attrs->send_cq);
+	dev->gsi_rqcq = get_qedr_cq(attrs->recv_cq);
+	dev->gsi_qp = qp;
+}
+
+void qedr_ll2_tx_cb(void *_qdev, struct qed_roce_ll2_packet *pkt)
+{
+	struct qedr_dev *dev = (struct qedr_dev *)_qdev;
+	struct qedr_cq *cq = dev->gsi_sqcq;
+	struct qedr_qp *qp = dev->gsi_qp;
+	unsigned long flags;
+
+	DP_DEBUG(dev, QEDR_MSG_GSI,
+		 "LL2 TX CB: gsi_sqcq=%p, gsi_rqcq=%p, gsi_cons=%d, ibcq_comp=%s\n",
+		 dev->gsi_sqcq, dev->gsi_rqcq, qp->sq.gsi_cons,
+		 cq->ibcq.comp_handler ? "Yes" : "No");
+
+	dma_free_coherent(&dev->pdev->dev, pkt->header.len, pkt->header.vaddr,
+			  pkt->header.baddr);
+	kfree(pkt);
+
+	spin_lock_irqsave(&qp->q_lock, flags);
+	qedr_inc_sw_gsi_cons(&qp->sq);
+	spin_unlock_irqrestore(&qp->q_lock, flags);
+
+	if (cq->ibcq.comp_handler) {
+		spin_lock_irqsave(&cq->comp_handler_lock, flags);
+		(*cq->ibcq.comp_handler) (&cq->ibcq, cq->ibcq.cq_context);
+		spin_unlock_irqrestore(&cq->comp_handler_lock, flags);
+	}
+}
+
+static inline void qedr_print_hex_dump(const char *str, void *buf, size_t size)
+{
+	print_hex_dump(KERN_INFO, str, DUMP_PREFIX_OFFSET, 32, 1, buf,
+		       size, false);
+}
+
+void qedr_ll2_rx_cb(void *_dev, struct qed_roce_ll2_packet *pkt,
+		    struct qed_roce_ll2_rx_params *params)
+{
+	struct qedr_dev *dev = (struct qedr_dev *)_dev;
+	struct qedr_cq *cq = dev->gsi_rqcq;
+	struct qedr_qp *qp = dev->gsi_qp;
+	unsigned long flags;
+
+	spin_lock_irqsave(&qp->q_lock, flags);
+
+	qp->rqe_wr_id[qp->rq.gsi_cons].rc = params->rc;
+	qp->rqe_wr_id[qp->rq.gsi_cons].vlan_id = params->vlan_id;
+	qp->rqe_wr_id[qp->rq.gsi_cons].sg_list[0].length = pkt->payload[0].len;
+	ether_addr_copy(qp->rqe_wr_id[qp->rq.gsi_cons].smac, params->smac);
+
+	qedr_inc_sw_gsi_cons(&qp->rq);
+
+	spin_unlock_irqrestore(&qp->q_lock, flags);
+
+	if (cq->ibcq.comp_handler) {
+		spin_lock_irqsave(&cq->comp_handler_lock, flags);
+		(*cq->ibcq.comp_handler) (&cq->ibcq, cq->ibcq.cq_context);
+		spin_unlock_irqrestore(&cq->comp_handler_lock, flags);
+	}
+}
+
+static void qedr_destroy_gsi_cq(struct qedr_dev *dev,
+				struct ib_qp_init_attr *attrs)
+{
+	struct qed_rdma_destroy_cq_in_params iparams;
+	struct qed_rdma_destroy_cq_out_params oparams;
+	struct qedr_cq *cq;
+
+	cq = get_qedr_cq(attrs->send_cq);
+	iparams.icid = cq->icid;
+	dev->ops->rdma_destroy_cq(dev->rdma_ctx, &iparams, &oparams);
+	dev->ops->common->chain_free(dev->cdev, &cq->pbl);
+
+	cq = get_qedr_cq(attrs->recv_cq);
+	/* if a dedicated recv_cq was used, delete it too */
+	if (iparams.icid != cq->icid) {
+		iparams.icid = cq->icid;
+		dev->ops->rdma_destroy_cq(dev->rdma_ctx, &iparams, &oparams);
+		dev->ops->common->chain_free(dev->cdev, &cq->pbl);
+	}
+}
+
+static inline int qedr_check_gsi_qp_attrs(struct qedr_dev *dev,
+					  struct ib_qp_init_attr *attrs)
+{
+	if (attrs->cap.max_recv_sge > QEDR_GSI_MAX_RECV_SGE) {
+		DP_ERR(dev,
+		       " create gsi qp: failed. max_recv_sge is larger the max %d>%d\n",
+		       attrs->cap.max_recv_sge, QEDR_GSI_MAX_RECV_SGE);
+		return -EINVAL;
+	}
+
+	if (attrs->cap.max_recv_wr > QEDR_GSI_MAX_RECV_WR) {
+		DP_ERR(dev,
+		       " create gsi qp: failed. max_recv_wr is too large %d>%d\n",
+		       attrs->cap.max_recv_wr, QEDR_GSI_MAX_RECV_WR);
+		return -EINVAL;
+	}
+
+	if (attrs->cap.max_send_wr > QEDR_GSI_MAX_SEND_WR) {
+		DP_ERR(dev,
+		       " create gsi qp: failed. max_send_wr is too large %d>%d\n",
+		       attrs->cap.max_send_wr, QEDR_GSI_MAX_SEND_WR);
+		return -EINVAL;
+	}
+
+	return 0;
+}
+
+struct ib_qp *qedr_create_gsi_qp(struct qedr_dev *dev,
+				 struct ib_qp_init_attr *attrs,
+				 struct qedr_qp *qp)
+{
+	struct qed_roce_ll2_params ll2_params;
+	int rc;
+
+	rc = qedr_check_gsi_qp_attrs(dev, attrs);
+	if (rc)
+		return ERR_PTR(rc);
+
+	/* configure and start LL2 */
+	memset(&ll2_params, 0, sizeof(ll2_params));
+	ll2_params.max_tx_buffers = attrs->cap.max_send_wr;
+	ll2_params.max_rx_buffers = attrs->cap.max_recv_wr;
+	ll2_params.cbs.tx_cb = qedr_ll2_tx_cb;
+	ll2_params.cbs.rx_cb = qedr_ll2_rx_cb;
+	ll2_params.cb_cookie = (void *)dev;
+	ll2_params.mtu = dev->ndev->mtu;
+	ether_addr_copy(ll2_params.mac_address, dev->ndev->dev_addr);
+	rc = dev->ops->roce_ll2_start(dev->cdev, &ll2_params);
+	if (rc) {
+		DP_ERR(dev, "create gsi qp: failed on ll2 start. rc=%d\n", rc);
+		return ERR_PTR(rc);
+	}
+
+	/* create QP */
+	qp->ibqp.qp_num = 1;
+	qp->rq.max_wr = attrs->cap.max_recv_wr;
+	qp->sq.max_wr = attrs->cap.max_send_wr;
+
+	qp->rqe_wr_id = kcalloc(qp->rq.max_wr, sizeof(*qp->rqe_wr_id),
+				GFP_KERNEL);
+	if (!qp->rqe_wr_id) {
+		DP_ERR(dev, "create gsi qp: failed on rqe_wr_id allocation\n");
+		goto err;
+	}
+	qp->wqe_wr_id = kcalloc(qp->sq.max_wr, sizeof(*qp->wqe_wr_id),
+				GFP_KERNEL);
+	if (!qp->wqe_wr_id) {
+		DP_ERR(dev, "create gsi qp: failed on wqe_wr_id allocation\n");
+		goto err;
+	}
+
+	qedr_store_gsi_qp_cq(dev, qp, attrs);
+	ether_addr_copy(dev->gsi_ll2_mac_address, dev->ndev->dev_addr);
+
+	/* the GSI CQ is handled by the driver so remove it from the FW */
+	qedr_destroy_gsi_cq(dev, attrs);
+	dev->gsi_rqcq->cq_type = QEDR_CQ_TYPE_GSI;
+	dev->gsi_rqcq->cq_type = QEDR_CQ_TYPE_GSI;
+
+	DP_DEBUG(dev, QEDR_MSG_GSI, "created GSI QP %p\n", qp);
+
+	return &qp->ibqp;
+
+err:
+	kfree(qp->rqe_wr_id);
+
+	rc = dev->ops->roce_ll2_stop(dev->cdev);
+	if (rc)
+		DP_ERR(dev, "create gsi qp: failed destroy on create\n");
+
+	return ERR_PTR(-ENOMEM);
+}
+
+int qedr_destroy_gsi_qp(struct qedr_dev *dev)
+{
+	int rc;
+
+	rc = dev->ops->roce_ll2_stop(dev->cdev);
+	if (rc)
+		DP_ERR(dev, "destroy gsi qp: failed (rc=%d)\n", rc);
+	else
+		DP_DEBUG(dev, QEDR_MSG_GSI, "destroy gsi qp: success\n");
+
+	return rc;
+}
+
+#define QEDR_MAX_UD_HEADER_SIZE	(100)
+#define QEDR_GSI_QPN		(1)
+static inline int qedr_gsi_build_header(struct qedr_dev *dev,
+					struct qedr_qp *qp,
+					struct ib_send_wr *swr,
+					struct ib_ud_header *udh,
+					int *roce_mode)
+{
+	bool has_vlan = false, has_grh_ipv6 = true;
+	struct ib_ah_attr *ah_attr = &get_qedr_ah(ud_wr(swr)->ah)->attr;
+	struct ib_global_route *grh = &ah_attr->grh;
+	union ib_gid sgid;
+	int send_size = 0;
+	u16 vlan_id = 0;
+	u16 ether_type;
+	struct ib_gid_attr sgid_attr;
+	int rc;
+	int ip_ver = 0;
+
+	bool has_udp = false;
+	int i;
+
+	send_size = 0;
+	for (i = 0; i < swr->num_sge; ++i)
+		send_size += swr->sg_list[i].length;
+
+	rc = ib_get_cached_gid(qp->ibqp.device, ah_attr->port_num,
+			       grh->sgid_index, &sgid, &sgid_attr);
+	if (rc) {
+		DP_ERR(dev,
+		       "gsi post send: failed to get cached GID (port=%d, ix=%d)\n",
+		       ah_attr->port_num, grh->sgid_index);
+		return rc;
+	}
+
+	vlan_id = rdma_vlan_dev_vlan_id(sgid_attr.ndev);
+	if (vlan_id < VLAN_CFI_MASK)
+		has_vlan = true;
+	if (sgid_attr.ndev)
+		dev_put(sgid_attr.ndev);
+
+	if (!memcmp(&sgid, &zgid, sizeof(sgid))) {
+		DP_ERR(dev, "gsi post send: GID not found GID index %d\n",
+		       ah_attr->grh.sgid_index);
+		return -ENOENT;
+	}
+
+	has_udp = (sgid_attr.gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP);
+	if (!has_udp) {
+		/* RoCE v1 */
+		ether_type = ETH_P_ROCE;
+		*roce_mode = ROCE_V1;
+	} else if (ipv6_addr_v4mapped((struct in6_addr *)&sgid)) {
+		/* RoCE v2 IPv4 */
+		ip_ver = 4;
+		ether_type = ETH_P_IP;
+		has_grh_ipv6 = false;
+		*roce_mode = ROCE_V2_IPV4;
+	} else {
+		/* RoCE v2 IPv6 */
+		ip_ver = 6;
+		ether_type = ETH_P_IPV6;
+		*roce_mode = ROCE_V2_IPV6;
+	}
+
+	rc = ib_ud_header_init(send_size, false, true, has_vlan,
+			       has_grh_ipv6, ip_ver, has_udp, 0, udh);
+	if (rc) {
+		DP_ERR(dev, "gsi post send: failed to init header\n");
+		return rc;
+	}
+
+	/* ENET + VLAN headers */
+	ether_addr_copy(udh->eth.dmac_h, ah_attr->dmac);
+	ether_addr_copy(udh->eth.smac_h, dev->ndev->dev_addr);
+	if (has_vlan) {
+		udh->eth.type = htons(ETH_P_8021Q);
+		udh->vlan.tag = htons(vlan_id);
+		udh->vlan.type = htons(ether_type);
+	} else {
+		udh->eth.type = htons(ether_type);
+	}
+
+	/* BTH */
+	udh->bth.solicited_event = !!(swr->send_flags & IB_SEND_SOLICITED);
+	udh->bth.pkey = QEDR_ROCE_PKEY_DEFAULT;
+	udh->bth.destination_qpn = htonl(ud_wr(swr)->remote_qpn);
+	udh->bth.psn = htonl((qp->sq_psn++) & ((1 << 24) - 1));
+	udh->bth.opcode = IB_OPCODE_UD_SEND_ONLY;
+
+	/* DETH */
+	udh->deth.qkey = htonl(0x80010000);
+	udh->deth.source_qpn = htonl(QEDR_GSI_QPN);
+
+	if (has_grh_ipv6) {
+		/* GRH / IPv6 header */
+		udh->grh.traffic_class = grh->traffic_class;
+		udh->grh.flow_label = grh->flow_label;
+		udh->grh.hop_limit = grh->hop_limit;
+		udh->grh.destination_gid = grh->dgid;
+		memcpy(&udh->grh.source_gid.raw, &sgid.raw,
+		       sizeof(udh->grh.source_gid.raw));
+	} else {
+		/* IPv4 header */
+		u32 ipv4_addr;
+
+		udh->ip4.protocol = IPPROTO_UDP;
+		udh->ip4.tos = htonl(ah_attr->grh.flow_label);
+		udh->ip4.frag_off = htons(IP_DF);
+		udh->ip4.ttl = ah_attr->grh.hop_limit;
+
+		ipv4_addr = qedr_get_ipv4_from_gid(sgid.raw);
+		udh->ip4.saddr = ipv4_addr;
+		ipv4_addr = qedr_get_ipv4_from_gid(ah_attr->grh.dgid.raw);
+		udh->ip4.daddr = ipv4_addr;
+		/* note: checksum is calculated by the device */
+	}
+
+	/* UDP */
+	if (has_udp) {
+		udh->udp.sport = htons(QEDR_ROCE_V2_UDP_SPORT);
+		udh->udp.dport = htons(ROCE_V2_UDP_DPORT);
+		udh->udp.csum = 0;
+		/* UDP length is untouched hence is zero */
+	}
+	return 0;
+}
+
+static inline int qedr_gsi_build_packet(struct qedr_dev *dev,
+					struct qedr_qp *qp,
+					struct ib_send_wr *swr,
+					struct qed_roce_ll2_packet **p_packet)
+{
+	u8 ud_header_buffer[QEDR_MAX_UD_HEADER_SIZE];
+	struct qed_roce_ll2_packet *packet;
+	struct pci_dev *pdev = dev->pdev;
+	int roce_mode, header_size;
+	struct ib_ud_header udh;
+	int i, rc;
+
+	*p_packet = NULL;
+
+	rc = qedr_gsi_build_header(dev, qp, swr, &udh, &roce_mode);
+	if (rc)
+		return rc;
+
+	header_size = ib_ud_header_pack(&udh, &ud_header_buffer);
+
+	packet = kzalloc(sizeof(*packet), GFP_ATOMIC);
+	if (!packet) {
+		DP_ERR(dev, "gsi post send: failed to allocate packet\n");
+		return -ENOMEM;
+	}
+
+	packet->header.vaddr = dma_alloc_coherent(&pdev->dev, header_size,
+						  &packet->header.baddr,
+						  GFP_ATOMIC);
+	if (!packet->header.vaddr) {
+		DP_ERR(dev,
+		       "cm header build: failed to allocate buffer for header\n");
+		kfree(packet);
+		return -ENOMEM;
+	}
+
+	if (ether_addr_equal(udh.eth.smac_h, udh.eth.dmac_h))
+		packet->tx_dest = QED_ROCE_LL2_TX_DEST_NW;
+	else
+		packet->tx_dest = QED_ROCE_LL2_TX_DEST_LB;
+
+	packet->roce_mode = roce_mode;
+	memcpy(packet->header.vaddr, ud_header_buffer, header_size);
+	packet->header.len = header_size;
+	packet->n_seg = swr->num_sge;
+	for (i = 0; i < packet->n_seg; i++) {
+		packet->payload[i].baddr = swr->sg_list[i].addr;
+		packet->payload[i].len = swr->sg_list[i].length;
+	}
+
+	*p_packet = packet;
+
+	return 0;
+}
+
+int qedr_gsi_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
+		       struct ib_send_wr **bad_wr)
+{
+	struct qed_roce_ll2_packet *pkt = NULL;
+	struct qedr_qp *qp = get_qedr_qp(ibqp);
+	struct qed_roce_ll2_tx_params params;
+	struct qedr_dev *dev = qp->dev;
+	unsigned long flags;
+	int rc;
+
+	if (qp->state != QED_ROCE_QP_STATE_RTS) {
+		*bad_wr = wr;
+		DP_ERR(dev,
+		       "gsi post recv: failed to post rx buffer. state is %d and not QED_ROCE_QP_STATE_RTS\n",
+		       qp->state);
+		return -EINVAL;
+	}
+
+	if (wr->num_sge > RDMA_MAX_SGE_PER_SQ_WQE) {
+		DP_ERR(dev, "gsi post send: num_sge is too large (%d>%d)\n",
+		       wr->num_sge, RDMA_MAX_SGE_PER_SQ_WQE);
+		rc = -EINVAL;
+		goto err;
+	}
+
+	if (wr->opcode != IB_WR_SEND) {
+		DP_ERR(dev,
+		       "gsi post send: failed due to unsupported opcode %d\n",
+		       wr->opcode);
+		rc = -EINVAL;
+		goto err;
+	}
+
+	memset(&params, 0, sizeof(params));
+
+	spin_lock_irqsave(&qp->q_lock, flags);
+
+	rc = qedr_gsi_build_packet(dev, qp, wr, &pkt);
+	if (rc) {
+		spin_unlock_irqrestore(&qp->q_lock, flags);
+		goto err;
+	}
+
+	rc = dev->ops->roce_ll2_tx(dev->cdev, pkt, &params);
+	if (!rc) {
+		qp->wqe_wr_id[qp->sq.prod].wr_id = wr->wr_id;
+		qedr_inc_sw_prod(&qp->sq);
+		DP_DEBUG(qp->dev, QEDR_MSG_GSI,
+			 "gsi post send: opcode=%d, in_irq=%ld, irqs_disabled=%d, wr_id=%llx\n",
+			 wr->opcode, in_irq(), irqs_disabled(), wr->wr_id);
+	} else {
+		if (rc == QED_ROCE_TX_HEAD_FAILURE) {
+			/* TX failed while posting header - release resources */
+			dma_free_coherent(&dev->pdev->dev, pkt->header.len,
+					  pkt->header.vaddr, pkt->header.baddr);
+			kfree(pkt);
+		} else if (rc == QED_ROCE_TX_FRAG_FAILURE) {
+			/* NTD since TX failed while posting a fragment. We will
+			 * release the resources on TX callback
+			 */
+		}
+
+		DP_ERR(dev, "gsi post send: failed to transmit (rc=%d)\n", rc);
+		rc = -EAGAIN;
+		*bad_wr = wr;
+	}
+
+	spin_unlock_irqrestore(&qp->q_lock, flags);
+
+	if (wr->next) {
+		DP_ERR(dev,
+		       "gsi post send: failed second WR. Only one WR may be passed at a time\n");
+		*bad_wr = wr->next;
+		rc = -EINVAL;
+	}
+
+	return rc;
+
+err:
+	*bad_wr = wr;
+	return rc;
+}
+
+int qedr_gsi_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
+		       struct ib_recv_wr **bad_wr)
+{
+	struct qedr_dev *dev = get_qedr_dev(ibqp->device);
+	struct qedr_qp *qp = get_qedr_qp(ibqp);
+	struct qed_roce_ll2_buffer buf;
+	unsigned long flags;
+	int status = 0;
+	int rc;
+
+	if ((qp->state != QED_ROCE_QP_STATE_RTR) &&
+	    (qp->state != QED_ROCE_QP_STATE_RTS)) {
+		*bad_wr = wr;
+		DP_ERR(dev,
+		       "gsi post recv: failed to post rx buffer. state is %d and not QED_ROCE_QP_STATE_RTR/S\n",
+		       qp->state);
+		return -EINVAL;
+	}
+
+	memset(&buf, 0, sizeof(buf));
+
+	spin_lock_irqsave(&qp->q_lock, flags);
+
+	while (wr) {
+		if (wr->num_sge > QEDR_GSI_MAX_RECV_SGE) {
+			DP_ERR(dev,
+			       "gsi post recv: failed to post rx buffer. too many sges %d>%d\n",
+			       wr->num_sge, QEDR_GSI_MAX_RECV_SGE);
+			goto err;
+		}
+
+		buf.baddr = wr->sg_list[0].addr;
+		buf.len = wr->sg_list[0].length;
+
+		rc = dev->ops->roce_ll2_post_rx_buffer(dev->cdev, &buf, 0, 1);
+		if (rc) {
+			DP_ERR(dev,
+			       "gsi post recv: failed to post rx buffer (rc=%d)\n",
+			       rc);
+			goto err;
+		}
+
+		memset(&qp->rqe_wr_id[qp->rq.prod], 0,
+		       sizeof(qp->rqe_wr_id[qp->rq.prod]));
+		qp->rqe_wr_id[qp->rq.prod].sg_list[0] = wr->sg_list[0];
+		qp->rqe_wr_id[qp->rq.prod].wr_id = wr->wr_id;
+
+		qedr_inc_sw_prod(&qp->rq);
+
+		wr = wr->next;
+	}
+
+	spin_unlock_irqrestore(&qp->q_lock, flags);
+
+	return status;
+err:
+	spin_unlock_irqrestore(&qp->q_lock, flags);
+	*bad_wr = wr;
+	return -ENOMEM;
+}
+
+int qedr_gsi_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
+{
+	struct qedr_dev *dev = get_qedr_dev(ibcq->device);
+	struct qedr_cq *cq = get_qedr_cq(ibcq);
+	struct qedr_qp *qp = dev->gsi_qp;
+	unsigned long flags;
+	int i = 0;
+
+	spin_lock_irqsave(&cq->cq_lock, flags);
+
+	while (i < num_entries && qp->rq.cons != qp->rq.gsi_cons) {
+		memset(&wc[i], 0, sizeof(*wc));
+
+		wc[i].qp = &qp->ibqp;
+		wc[i].wr_id = qp->rqe_wr_id[qp->rq.cons].wr_id;
+		wc[i].opcode = IB_WC_RECV;
+		wc[i].pkey_index = 0;
+		wc[i].status = (qp->rqe_wr_id[qp->rq.cons].rc) ?
+		    IB_WC_GENERAL_ERR : IB_WC_SUCCESS;
+		/* 0 - currently only one recv sg is supported */
+		wc[i].byte_len = qp->rqe_wr_id[qp->rq.cons].sg_list[0].length;
+		wc[i].wc_flags |= IB_WC_GRH | IB_WC_IP_CSUM_OK;
+		ether_addr_copy(wc[i].smac, qp->rqe_wr_id[qp->rq.cons].smac);
+		wc[i].wc_flags |= IB_WC_WITH_SMAC;
+		if (qp->rqe_wr_id[qp->rq.cons].vlan_id) {
+			wc[i].wc_flags |= IB_WC_WITH_VLAN;
+			wc[i].vlan_id = qp->rqe_wr_id[qp->rq.cons].vlan_id;
+		}
+
+		qedr_inc_sw_cons(&qp->rq);
+		i++;
+	}
+
+	while (i < num_entries && qp->sq.cons != qp->sq.gsi_cons) {
+		memset(&wc[i], 0, sizeof(*wc));
+
+		wc[i].qp = &qp->ibqp;
+		wc[i].wr_id = qp->wqe_wr_id[qp->sq.cons].wr_id;
+		wc[i].opcode = IB_WC_SEND;
+		wc[i].status = IB_WC_SUCCESS;
+
+		qedr_inc_sw_cons(&qp->sq);
+		i++;
+	}
+
+	spin_unlock_irqrestore(&cq->cq_lock, flags);
+
+	DP_DEBUG(dev, QEDR_MSG_GSI,
+		 "gsi poll_cq: requested entries=%d, actual=%d, qp->rq.cons=%d, qp->rq.gsi_cons=%x, qp->sq.cons=%d, qp->sq.gsi_cons=%d, qp_num=%d\n",
+		 num_entries, i, qp->rq.cons, qp->rq.gsi_cons, qp->sq.cons,
+		 qp->sq.gsi_cons, qp->ibqp.qp_num);
+
+	return i;
+}
diff --git a/drivers/infiniband/hw/qedr/qedr_cm.h b/drivers/infiniband/hw/qedr/qedr_cm.h
index b8a8b76..9ba6e15 100644
--- a/drivers/infiniband/hw/qedr/qedr_cm.h
+++ b/drivers/infiniband/hw/qedr/qedr_cm.h
@@ -32,9 +32,30 @@ 
 #ifndef LINUX_QEDR_CM_H_
 #define LINUX_QEDR_CM_H_
 
+#define QEDR_GSI_MAX_RECV_WR	(4096)
+#define QEDR_GSI_MAX_SEND_WR	(4096)
+
+#define QEDR_GSI_MAX_RECV_SGE	(1)	/* LL2 FW limitation */
+
+#define ETH_P_ROCE		(0x8915)
+#define QEDR_ROCE_V2_UDP_SPORT	(0000)
+
 static inline u32 qedr_get_ipv4_from_gid(u8 *gid)
 {
 	return *(u32 *)(void *)&gid[12];
 }
 
+/* RDMA CM */
+int qedr_gsi_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc);
+int qedr_gsi_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
+		       struct ib_recv_wr **bad_wr);
+int qedr_gsi_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
+		       struct ib_send_wr **bad_wr);
+struct ib_qp *qedr_create_gsi_qp(struct qedr_dev *dev,
+				 struct ib_qp_init_attr *attrs,
+				 struct qedr_qp *qp);
+void qedr_store_gsi_qp_cq(struct qedr_dev *dev,
+			  struct qedr_qp *qp, struct ib_qp_init_attr *attrs);
+int qedr_destroy_gsi_qp(struct qedr_dev *dev);
+void qedr_inc_sw_gsi_cons(struct qedr_qp_hwq_info *info);
 #endif
diff --git a/drivers/infiniband/hw/qedr/verbs.c b/drivers/infiniband/hw/qedr/verbs.c
index cd7c8cd..378eafa 100644
--- a/drivers/infiniband/hw/qedr/verbs.c
+++ b/drivers/infiniband/hw/qedr/verbs.c
@@ -1511,6 +1511,15 @@  struct ib_qp *qedr_create_qp(struct ib_pd *ibpd,
 
 	qedr_set_qp_init_params(dev, qp, pd, attrs);
 
+	if (attrs->qp_type == IB_QPT_GSI) {
+		if (udata) {
+			DP_ERR(dev,
+			       "create qp: unexpected udata when creating GSI QP\n");
+			goto err0;
+		}
+		return qedr_create_gsi_qp(dev, attrs, qp);
+	}
+
 	memset(&in_params, 0, sizeof(in_params));
 
 	if (udata) {
@@ -2077,6 +2086,8 @@  int qedr_destroy_qp(struct ib_qp *ibqp)
 		rc = dev->ops->rdma_destroy_qp(dev->rdma_ctx, qp->qed_qp);
 		if (rc)
 			return rc;
+	} else {
+		qedr_destroy_gsi_qp(dev);
 	}
 
 	if (ibqp->uobject && ibqp->uobject->context) {
@@ -2092,6 +2103,23 @@  int qedr_destroy_qp(struct ib_qp *ibqp)
 	return rc;
 }
 
+struct ib_ah *qedr_create_ah(struct ib_pd *ibpd, struct ib_ah_attr *attr)
+{
+	struct qedr_ah *ah = kzalloc(sizeof(*ah), GFP_ATOMIC);
+
+	ah->attr = *attr;
+
+	return &ah->ibah;
+}
+
+int qedr_destroy_ah(struct ib_ah *ibah)
+{
+	struct qedr_ah *ah = get_qedr_ah(ibah);
+
+	kfree(ah);
+	return 0;
+}
+
 static void free_mr_info(struct qedr_dev *dev, struct mr_info *info)
 {
 	struct qedr_pbl *pbl, *tmp;
@@ -2943,6 +2971,9 @@  int qedr_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 
 	*bad_wr = NULL;
 
+	if (qp->qp_type == IB_QPT_GSI)
+		return qedr_gsi_post_send(ibqp, wr, bad_wr);
+
 	spin_lock_irqsave(&qp->q_lock, flags);
 
 	if ((qp->state == QED_ROCE_QP_STATE_RESET) ||
@@ -2999,6 +3030,9 @@  int qedr_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 	unsigned long flags;
 	int status = 0;
 
+	if (qp->qp_type == IB_QPT_GSI)
+		return qedr_gsi_post_recv(ibqp, wr, bad_wr);
+
 	spin_lock_irqsave(&qp->q_lock, flags);
 
 	if ((qp->state == QED_ROCE_QP_STATE_RESET) ||
@@ -3425,6 +3459,9 @@  int qedr_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
 	int update = 0;
 	int done = 0;
 
+	if (cq->cq_type == QEDR_CQ_TYPE_GSI)
+		return qedr_gsi_poll_cq(ibcq, num_entries, wc);
+
 	spin_lock_irqsave(&cq->cq_lock, flags);
 	old_cons = qed_chain_get_cons_idx_u32(&cq->pbl);
 	while (num_entries && is_valid_cqe(cq, cqe)) {