RDMA/erdma: Add verbs implementation

The RDMA verbs implementation of erdma is divided into three files:
erdma_qp.c, erdma_cq.c, and erdma_verbs.c. Internal used functions and
datapath functions of QP/CQ are put in erdma_qp.c and erdma_cq.c, the rest
is in erdma_verbs.c.

This commit also fixes some static check warnings.

Link: https://lore.kernel.org/r/20220727014927.76564-8-chengyou@linux.alibaba.com
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Reported-by: Abaci Robot <abaci@linux.alibaba.com>
Signed-off-by: Yang Li <yang.lee@linux.alibaba.com>
Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Signed-off-by: Cheng Xu <chengyou@linux.alibaba.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>

drivers/infiniband/hw/erdma/erdma_cq.c

@@ -0,0 +1,205 @@
+// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
+
+/* Authors: Cheng Xu <chengyou@linux.alibaba.com> */
+/*          Kai Shen <kaishen@linux.alibaba.com> */
+/* Copyright (c) 2020-2022, Alibaba Group. */
+
+#include <rdma/ib_verbs.h>
+
+#include "erdma_hw.h"
+#include "erdma_verbs.h"
+
+static void *get_next_valid_cqe(struct erdma_cq *cq)
+{
+	__be32 *cqe = get_queue_entry(cq->kern_cq.qbuf, cq->kern_cq.ci,
+				      cq->depth, CQE_SHIFT);
+	u32 owner = FIELD_GET(ERDMA_CQE_HDR_OWNER_MASK,
+			      __be32_to_cpu(READ_ONCE(*cqe)));
+
+	return owner ^ !!(cq->kern_cq.ci & cq->depth) ? cqe : NULL;
+}
+
+static void notify_cq(struct erdma_cq *cq, u8 solcitied)
+{
+	u64 db_data =
+		FIELD_PREP(ERDMA_CQDB_IDX_MASK, (cq->kern_cq.notify_cnt)) |
+		FIELD_PREP(ERDMA_CQDB_CQN_MASK, cq->cqn) |
+		FIELD_PREP(ERDMA_CQDB_ARM_MASK, 1) |
+		FIELD_PREP(ERDMA_CQDB_SOL_MASK, solcitied) |
+		FIELD_PREP(ERDMA_CQDB_CMDSN_MASK, cq->kern_cq.cmdsn) |
+		FIELD_PREP(ERDMA_CQDB_CI_MASK, cq->kern_cq.ci);
+
+	*cq->kern_cq.db_record = db_data;
+	writeq(db_data, cq->kern_cq.db);
+}
+
+int erdma_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
+{
+	struct erdma_cq *cq = to_ecq(ibcq);
+	unsigned long irq_flags;
+	int ret = 0;
+
+	spin_lock_irqsave(&cq->kern_cq.lock, irq_flags);
+
+	notify_cq(cq, (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED);
+
+	if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && get_next_valid_cqe(cq))
+		ret = 1;
+
+	cq->kern_cq.notify_cnt++;
+
+	spin_unlock_irqrestore(&cq->kern_cq.lock, irq_flags);
+
+	return ret;
+}
+
+static const enum ib_wc_opcode wc_mapping_table[ERDMA_NUM_OPCODES] = {
+	[ERDMA_OP_WRITE] = IB_WC_RDMA_WRITE,
+	[ERDMA_OP_READ] = IB_WC_RDMA_READ,
+	[ERDMA_OP_SEND] = IB_WC_SEND,
+	[ERDMA_OP_SEND_WITH_IMM] = IB_WC_SEND,
+	[ERDMA_OP_RECEIVE] = IB_WC_RECV,
+	[ERDMA_OP_RECV_IMM] = IB_WC_RECV_RDMA_WITH_IMM,
+	[ERDMA_OP_RECV_INV] = IB_WC_RECV,
+	[ERDMA_OP_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
+	[ERDMA_OP_INVALIDATE] = IB_WC_LOCAL_INV,
+	[ERDMA_OP_RSP_SEND_IMM] = IB_WC_RECV,
+	[ERDMA_OP_SEND_WITH_INV] = IB_WC_SEND,
+	[ERDMA_OP_REG_MR] = IB_WC_REG_MR,
+	[ERDMA_OP_LOCAL_INV] = IB_WC_LOCAL_INV,
+	[ERDMA_OP_READ_WITH_INV] = IB_WC_RDMA_READ,
+};
+
+static const struct {
+	enum erdma_wc_status erdma;
+	enum ib_wc_status base;
+	enum erdma_vendor_err vendor;
+} map_cqe_status[ERDMA_NUM_WC_STATUS] = {
+	{ ERDMA_WC_SUCCESS, IB_WC_SUCCESS, ERDMA_WC_VENDOR_NO_ERR },
+	{ ERDMA_WC_GENERAL_ERR, IB_WC_GENERAL_ERR, ERDMA_WC_VENDOR_NO_ERR },
+	{ ERDMA_WC_RECV_WQE_FORMAT_ERR, IB_WC_GENERAL_ERR,
+	  ERDMA_WC_VENDOR_INVALID_RQE },
+	{ ERDMA_WC_RECV_STAG_INVALID_ERR, IB_WC_REM_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_RQE_INVALID_STAG },
+	{ ERDMA_WC_RECV_ADDR_VIOLATION_ERR, IB_WC_REM_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_RQE_ADDR_VIOLATION },
+	{ ERDMA_WC_RECV_RIGHT_VIOLATION_ERR, IB_WC_REM_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_RQE_ACCESS_RIGHT_ERR },
+	{ ERDMA_WC_RECV_PDID_ERR, IB_WC_REM_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_RQE_INVALID_PD },
+	{ ERDMA_WC_RECV_WARRPING_ERR, IB_WC_REM_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_RQE_WRAP_ERR },
+	{ ERDMA_WC_SEND_WQE_FORMAT_ERR, IB_WC_LOC_QP_OP_ERR,
+	  ERDMA_WC_VENDOR_INVALID_SQE },
+	{ ERDMA_WC_SEND_WQE_ORD_EXCEED, IB_WC_GENERAL_ERR,
+	  ERDMA_WC_VENDOR_ZERO_ORD },
+	{ ERDMA_WC_SEND_STAG_INVALID_ERR, IB_WC_LOC_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_SQE_INVALID_STAG },
+	{ ERDMA_WC_SEND_ADDR_VIOLATION_ERR, IB_WC_LOC_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_SQE_ADDR_VIOLATION },
+	{ ERDMA_WC_SEND_RIGHT_VIOLATION_ERR, IB_WC_LOC_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_SQE_ACCESS_ERR },
+	{ ERDMA_WC_SEND_PDID_ERR, IB_WC_LOC_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_SQE_INVALID_PD },
+	{ ERDMA_WC_SEND_WARRPING_ERR, IB_WC_LOC_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_SQE_WARP_ERR },
+	{ ERDMA_WC_FLUSH_ERR, IB_WC_WR_FLUSH_ERR, ERDMA_WC_VENDOR_NO_ERR },
+	{ ERDMA_WC_RETRY_EXC_ERR, IB_WC_RETRY_EXC_ERR, ERDMA_WC_VENDOR_NO_ERR },
+};
+
+#define ERDMA_POLLCQ_NO_QP 1
+
+static int erdma_poll_one_cqe(struct erdma_cq *cq, struct ib_wc *wc)
+{
+	struct erdma_dev *dev = to_edev(cq->ibcq.device);
+	u8 opcode, syndrome, qtype;
+	struct erdma_kqp *kern_qp;
+	struct erdma_cqe *cqe;
+	struct erdma_qp *qp;
+	u16 wqe_idx, depth;
+	u32 qpn, cqe_hdr;
+	u64 *id_table;
+	u64 *wqe_hdr;
+
+	cqe = get_next_valid_cqe(cq);
+	if (!cqe)
+		return -EAGAIN;
+
+	cq->kern_cq.ci++;
+
+	/* cqbuf should be ready when we poll */
+	dma_rmb();
+
+	qpn = be32_to_cpu(cqe->qpn);
+	wqe_idx = be32_to_cpu(cqe->qe_idx);
+	cqe_hdr = be32_to_cpu(cqe->hdr);
+
+	qp = find_qp_by_qpn(dev, qpn);
+	if (!qp)
+		return ERDMA_POLLCQ_NO_QP;
+
+	kern_qp = &qp->kern_qp;
+
+	qtype = FIELD_GET(ERDMA_CQE_HDR_QTYPE_MASK, cqe_hdr);
+	syndrome = FIELD_GET(ERDMA_CQE_HDR_SYNDROME_MASK, cqe_hdr);
+	opcode = FIELD_GET(ERDMA_CQE_HDR_OPCODE_MASK, cqe_hdr);
+
+	if (qtype == ERDMA_CQE_QTYPE_SQ) {
+		id_table = kern_qp->swr_tbl;
+		depth = qp->attrs.sq_size;
+		wqe_hdr = get_queue_entry(qp->kern_qp.sq_buf, wqe_idx,
+					  qp->attrs.sq_size, SQEBB_SHIFT);
+		kern_qp->sq_ci =
+			FIELD_GET(ERDMA_SQE_HDR_WQEBB_CNT_MASK, *wqe_hdr) +
+			wqe_idx + 1;
+	} else {
+		id_table = kern_qp->rwr_tbl;
+		depth = qp->attrs.rq_size;
+	}
+	wc->wr_id = id_table[wqe_idx & (depth - 1)];
+	wc->byte_len = be32_to_cpu(cqe->size);
+
+	wc->wc_flags = 0;
+
+	wc->opcode = wc_mapping_table[opcode];
+	if (opcode == ERDMA_OP_RECV_IMM || opcode == ERDMA_OP_RSP_SEND_IMM) {
+		wc->ex.imm_data = cpu_to_be32(le32_to_cpu(cqe->imm_data));
+		wc->wc_flags |= IB_WC_WITH_IMM;
+	} else if (opcode == ERDMA_OP_RECV_INV) {
+		wc->ex.invalidate_rkey = be32_to_cpu(cqe->inv_rkey);
+		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
+	}
+
+	if (syndrome >= ERDMA_NUM_WC_STATUS)
+		syndrome = ERDMA_WC_GENERAL_ERR;
+
+	wc->status = map_cqe_status[syndrome].base;
+	wc->vendor_err = map_cqe_status[syndrome].vendor;
+	wc->qp = &qp->ibqp;
+
+	return 0;
+}
+
+int erdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
+{
+	struct erdma_cq *cq = to_ecq(ibcq);
+	unsigned long flags;
+	int npolled, ret;
+
+	spin_lock_irqsave(&cq->kern_cq.lock, flags);
+
+	for (npolled = 0; npolled < num_entries;) {
+		ret = erdma_poll_one_cqe(cq, wc + npolled);
+
+		if (ret == -EAGAIN) /* no received new CQEs. */
+			break;
+		else if (ret) /* ignore invalid CQEs. */
+			continue;
+
+		npolled++;
+	}
+
+	spin_unlock_irqrestore(&cq->kern_cq.lock, flags);
+
+	return npolled;
+}

drivers/infiniband/hw/erdma/erdma_qp.c

@@ -0,0 +1,566 @@
+// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
+
+/* Authors: Cheng Xu <chengyou@linux.alibaba.com> */
+/*          Kai Shen <kaishen@linux.alibaba.com> */
+/* Copyright (c) 2020-2021, Alibaba Group */
+/* Authors: Bernard Metzler <bmt@zurich.ibm.com> */
+/* Copyright (c) 2008-2019, IBM Corporation */
+
+#include <linux/errno.h>
+#include <linux/pci.h>
+#include <linux/scatterlist.h>
+#include <linux/types.h>
+
+#include <rdma/ib_user_verbs.h>
+#include <rdma/ib_verbs.h>
+
+#include "erdma.h"
+#include "erdma_cm.h"
+#include "erdma_verbs.h"
+
+void erdma_qp_llp_close(struct erdma_qp *qp)
+{
+	struct erdma_qp_attrs qp_attrs;
+
+	down_write(&qp->state_lock);
+
+	switch (qp->attrs.state) {
+	case ERDMA_QP_STATE_RTS:
+	case ERDMA_QP_STATE_RTR:
+	case ERDMA_QP_STATE_IDLE:
+	case ERDMA_QP_STATE_TERMINATE:
+		qp_attrs.state = ERDMA_QP_STATE_CLOSING;
+		erdma_modify_qp_internal(qp, &qp_attrs, ERDMA_QP_ATTR_STATE);
+		break;
+	case ERDMA_QP_STATE_CLOSING:
+		qp->attrs.state = ERDMA_QP_STATE_IDLE;
+		break;
+	default:
+		break;
+	}
+
+	if (qp->cep) {
+		erdma_cep_put(qp->cep);
+		qp->cep = NULL;
+	}
+
+	up_write(&qp->state_lock);
+}
+
+struct ib_qp *erdma_get_ibqp(struct ib_device *ibdev, int id)
+{
+	struct erdma_qp *qp = find_qp_by_qpn(to_edev(ibdev), id);
+
+	if (qp)
+		return &qp->ibqp;
+
+	return NULL;
+}
+
+static int erdma_modify_qp_state_to_rts(struct erdma_qp *qp,
+					struct erdma_qp_attrs *attrs,
+					enum erdma_qp_attr_mask mask)
+{
+	int ret;
+	struct erdma_dev *dev = qp->dev;
+	struct erdma_cmdq_modify_qp_req req;
+	struct tcp_sock *tp;
+	struct erdma_cep *cep = qp->cep;
+	struct sockaddr_storage local_addr, remote_addr;
+
+	if (!(mask & ERDMA_QP_ATTR_LLP_HANDLE))
+		return -EINVAL;
+
+	if (!(mask & ERDMA_QP_ATTR_MPA))
+		return -EINVAL;
+
+	ret = getname_local(cep->sock, &local_addr);
+	if (ret < 0)
+		return ret;
+
+	ret = getname_peer(cep->sock, &remote_addr);
+	if (ret < 0)
+		return ret;
+
+	qp->attrs.state = ERDMA_QP_STATE_RTS;
+
+	tp = tcp_sk(qp->cep->sock->sk);
+
+	erdma_cmdq_build_reqhdr(&req.hdr, CMDQ_SUBMOD_RDMA,
+				CMDQ_OPCODE_MODIFY_QP);
+
+	req.cfg = FIELD_PREP(ERDMA_CMD_MODIFY_QP_STATE_MASK, qp->attrs.state) |
+		  FIELD_PREP(ERDMA_CMD_MODIFY_QP_CC_MASK, qp->attrs.cc) |
+		  FIELD_PREP(ERDMA_CMD_MODIFY_QP_QPN_MASK, QP_ID(qp));
+
+	req.cookie = be32_to_cpu(qp->cep->mpa.ext_data.cookie);
+	req.dip = to_sockaddr_in(remote_addr).sin_addr.s_addr;
+	req.sip = to_sockaddr_in(local_addr).sin_addr.s_addr;
+	req.dport = to_sockaddr_in(remote_addr).sin_port;
+	req.sport = to_sockaddr_in(local_addr).sin_port;
+
+	req.send_nxt = tp->snd_nxt;
+	/* rsvd tcp seq for mpa-rsp in server. */
+	if (qp->attrs.qp_type == ERDMA_QP_PASSIVE)
+		req.send_nxt += MPA_DEFAULT_HDR_LEN + qp->attrs.pd_len;
+	req.recv_nxt = tp->rcv_nxt;
+
+	return erdma_post_cmd_wait(&dev->cmdq, (u64 *)&req, sizeof(req), NULL,
+				   NULL);
+}
+
+static int erdma_modify_qp_state_to_stop(struct erdma_qp *qp,
+					 struct erdma_qp_attrs *attrs,
+					 enum erdma_qp_attr_mask mask)
+{
+	struct erdma_dev *dev = qp->dev;
+	struct erdma_cmdq_modify_qp_req req;
+
+	qp->attrs.state = attrs->state;
+
+	erdma_cmdq_build_reqhdr(&req.hdr, CMDQ_SUBMOD_RDMA,
+				CMDQ_OPCODE_MODIFY_QP);
+
+	req.cfg = FIELD_PREP(ERDMA_CMD_MODIFY_QP_STATE_MASK, attrs->state) |
+		  FIELD_PREP(ERDMA_CMD_MODIFY_QP_QPN_MASK, QP_ID(qp));
+
+	return erdma_post_cmd_wait(&dev->cmdq, (u64 *)&req, sizeof(req), NULL,
+				   NULL);
+}
+
+int erdma_modify_qp_internal(struct erdma_qp *qp, struct erdma_qp_attrs *attrs,
+			     enum erdma_qp_attr_mask mask)
+{
+	int drop_conn, ret = 0;
+
+	if (!mask)
+		return 0;
+
+	if (!(mask & ERDMA_QP_ATTR_STATE))
+		return 0;
+
+	switch (qp->attrs.state) {
+	case ERDMA_QP_STATE_IDLE:
+	case ERDMA_QP_STATE_RTR:
+		if (attrs->state == ERDMA_QP_STATE_RTS) {
+			ret = erdma_modify_qp_state_to_rts(qp, attrs, mask);
+		} else if (attrs->state == ERDMA_QP_STATE_ERROR) {
+			qp->attrs.state = ERDMA_QP_STATE_ERROR;
+			if (qp->cep) {
+				erdma_cep_put(qp->cep);
+				qp->cep = NULL;
+			}
+			ret = erdma_modify_qp_state_to_stop(qp, attrs, mask);
+		}
+		break;
+	case ERDMA_QP_STATE_RTS:
+		drop_conn = 0;
+
+		if (attrs->state == ERDMA_QP_STATE_CLOSING) {
+			ret = erdma_modify_qp_state_to_stop(qp, attrs, mask);
+			drop_conn = 1;
+		} else if (attrs->state == ERDMA_QP_STATE_TERMINATE) {
+			qp->attrs.state = ERDMA_QP_STATE_TERMINATE;
+			ret = erdma_modify_qp_state_to_stop(qp, attrs, mask);
+			drop_conn = 1;
+		} else if (attrs->state == ERDMA_QP_STATE_ERROR) {
+			ret = erdma_modify_qp_state_to_stop(qp, attrs, mask);
+			qp->attrs.state = ERDMA_QP_STATE_ERROR;
+			drop_conn = 1;
+		}
+
+		if (drop_conn)
+			erdma_qp_cm_drop(qp);
+
+		break;
+	case ERDMA_QP_STATE_TERMINATE:
+		if (attrs->state == ERDMA_QP_STATE_ERROR)
+			qp->attrs.state = ERDMA_QP_STATE_ERROR;
+		break;
+	case ERDMA_QP_STATE_CLOSING:
+		if (attrs->state == ERDMA_QP_STATE_IDLE) {
+			qp->attrs.state = ERDMA_QP_STATE_IDLE;
+		} else if (attrs->state == ERDMA_QP_STATE_ERROR) {
+			ret = erdma_modify_qp_state_to_stop(qp, attrs, mask);
+			qp->attrs.state = ERDMA_QP_STATE_ERROR;
+		} else if (attrs->state != ERDMA_QP_STATE_CLOSING) {
+			return -ECONNABORTED;
+		}
+		break;
+	default:
+		break;
+	}
+
+	return ret;
+}
+
+static void erdma_qp_safe_free(struct kref *ref)
+{
+	struct erdma_qp *qp = container_of(ref, struct erdma_qp, ref);
+
+	complete(&qp->safe_free);
+}
+
+void erdma_qp_put(struct erdma_qp *qp)
+{
+	WARN_ON(kref_read(&qp->ref) < 1);
+	kref_put(&qp->ref, erdma_qp_safe_free);
+}
+
+void erdma_qp_get(struct erdma_qp *qp)
+{
+	kref_get(&qp->ref);
+}
+
+static int fill_inline_data(struct erdma_qp *qp,
+			    const struct ib_send_wr *send_wr, u16 wqe_idx,
+			    u32 sgl_offset, __le32 *length_field)
+{
+	u32 remain_size, copy_size, data_off, bytes = 0;
+	char *data;
+	int i = 0;
+
+	wqe_idx += (sgl_offset >> SQEBB_SHIFT);
+	sgl_offset &= (SQEBB_SIZE - 1);
+	data = get_queue_entry(qp->kern_qp.sq_buf, wqe_idx, qp->attrs.sq_size,
+			       SQEBB_SHIFT);
+
+	while (i < send_wr->num_sge) {
+		bytes += send_wr->sg_list[i].length;
+		if (bytes > (int)ERDMA_MAX_INLINE)
+			return -EINVAL;
+
+		remain_size = send_wr->sg_list[i].length;
+		data_off = 0;
+
+		while (1) {
+			copy_size = min(remain_size, SQEBB_SIZE - sgl_offset);
+
+			memcpy(data + sgl_offset,
+			       (void *)(uintptr_t)send_wr->sg_list[i].addr +
+				       data_off,
+			       copy_size);
+			remain_size -= copy_size;
+			data_off += copy_size;
+			sgl_offset += copy_size;
+			wqe_idx += (sgl_offset >> SQEBB_SHIFT);
+			sgl_offset &= (SQEBB_SIZE - 1);
+
+			data = get_queue_entry(qp->kern_qp.sq_buf, wqe_idx,
+					       qp->attrs.sq_size, SQEBB_SHIFT);
+			if (!remain_size)
+				break;
+		}
+
+		i++;
+	}
+	*length_field = cpu_to_le32(bytes);
+
+	return bytes;
+}
+
+static int fill_sgl(struct erdma_qp *qp, const struct ib_send_wr *send_wr,
+		    u16 wqe_idx, u32 sgl_offset, __le32 *length_field)
+{
+	int i = 0;
+	u32 bytes = 0;
+	char *sgl;
+
+	if (send_wr->num_sge > qp->dev->attrs.max_send_sge)
+		return -EINVAL;
+
+	if (sgl_offset & 0xF)
+		return -EINVAL;
+
+	while (i < send_wr->num_sge) {
+		wqe_idx += (sgl_offset >> SQEBB_SHIFT);
+		sgl_offset &= (SQEBB_SIZE - 1);
+		sgl = get_queue_entry(qp->kern_qp.sq_buf, wqe_idx,
+				      qp->attrs.sq_size, SQEBB_SHIFT);
+
+		bytes += send_wr->sg_list[i].length;
+		memcpy(sgl + sgl_offset, &send_wr->sg_list[i],
+		       sizeof(struct ib_sge));
+
+		sgl_offset += sizeof(struct ib_sge);
+		i++;
+	}
+
+	*length_field = cpu_to_le32(bytes);
+	return 0;
+}
+
+static int erdma_push_one_sqe(struct erdma_qp *qp, u16 *pi,
+			      const struct ib_send_wr *send_wr)
+{
+	u32 wqe_size, wqebb_cnt, hw_op, flags, sgl_offset;
+	u32 idx = *pi & (qp->attrs.sq_size - 1);
+	enum ib_wr_opcode op = send_wr->opcode;
+	struct erdma_readreq_sqe *read_sqe;
+	struct erdma_reg_mr_sqe *regmr_sge;
+	struct erdma_write_sqe *write_sqe;
+	struct erdma_send_sqe *send_sqe;
+	struct ib_rdma_wr *rdma_wr;
+	struct erdma_mr *mr;
+	__le32 *length_field;
+	u64 wqe_hdr, *entry;
+	struct ib_sge *sge;
+	u32 attrs;
+	int ret;
+
+	entry = get_queue_entry(qp->kern_qp.sq_buf, idx, qp->attrs.sq_size,
+				SQEBB_SHIFT);
+
+	/* Clear the SQE header section. */
+	*entry = 0;
+
+	qp->kern_qp.swr_tbl[idx] = send_wr->wr_id;
+	flags = send_wr->send_flags;
+	wqe_hdr = FIELD_PREP(
+		ERDMA_SQE_HDR_CE_MASK,
+		((flags & IB_SEND_SIGNALED) || qp->kern_qp.sig_all) ? 1 : 0);
+	wqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_SE_MASK,
+			      flags & IB_SEND_SOLICITED ? 1 : 0);
+	wqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_FENCE_MASK,
+			      flags & IB_SEND_FENCE ? 1 : 0);
+	wqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_INLINE_MASK,
+			      flags & IB_SEND_INLINE ? 1 : 0);
+	wqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_QPN_MASK, QP_ID(qp));
+
+	switch (op) {
+	case IB_WR_RDMA_WRITE:
+	case IB_WR_RDMA_WRITE_WITH_IMM:
+		hw_op = ERDMA_OP_WRITE;
+		if (op == IB_WR_RDMA_WRITE_WITH_IMM)
+			hw_op = ERDMA_OP_WRITE_WITH_IMM;
+		wqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_OPCODE_MASK, hw_op);
+		rdma_wr = container_of(send_wr, struct ib_rdma_wr, wr);
+		write_sqe = (struct erdma_write_sqe *)entry;
+
+		write_sqe->imm_data = send_wr->ex.imm_data;
+		write_sqe->sink_stag = cpu_to_le32(rdma_wr->rkey);
+		write_sqe->sink_to_h =
+			cpu_to_le32(upper_32_bits(rdma_wr->remote_addr));
+		write_sqe->sink_to_l =
+			cpu_to_le32(lower_32_bits(rdma_wr->remote_addr));
+
+		length_field = &write_sqe->length;
+		wqe_size = sizeof(struct erdma_write_sqe);
+		sgl_offset = wqe_size;
+		break;
+	case IB_WR_RDMA_READ:
+	case IB_WR_RDMA_READ_WITH_INV:
+		read_sqe = (struct erdma_readreq_sqe *)entry;
+		if (unlikely(send_wr->num_sge != 1))
+			return -EINVAL;
+		hw_op = ERDMA_OP_READ;
+		if (op == IB_WR_RDMA_READ_WITH_INV) {
+			hw_op = ERDMA_OP_READ_WITH_INV;
+			read_sqe->invalid_stag =
+				cpu_to_le32(send_wr->ex.invalidate_rkey);
+		}
+
+		wqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_OPCODE_MASK, hw_op);
+		rdma_wr = container_of(send_wr, struct ib_rdma_wr, wr);
+		read_sqe->length = cpu_to_le32(send_wr->sg_list[0].length);
+		read_sqe->sink_stag = cpu_to_le32(send_wr->sg_list[0].lkey);
+		read_sqe->sink_to_l =
+			cpu_to_le32(lower_32_bits(send_wr->sg_list[0].addr));
+		read_sqe->sink_to_h =
+			cpu_to_le32(upper_32_bits(send_wr->sg_list[0].addr));
+
+		sge = get_queue_entry(qp->kern_qp.sq_buf, idx + 1,
+				      qp->attrs.sq_size, SQEBB_SHIFT);
+		sge->addr = rdma_wr->remote_addr;
+		sge->lkey = rdma_wr->rkey;
+		sge->length = send_wr->sg_list[0].length;
+		wqe_size = sizeof(struct erdma_readreq_sqe) +
+			   send_wr->num_sge * sizeof(struct ib_sge);
+
+		goto out;
+	case IB_WR_SEND:
+	case IB_WR_SEND_WITH_IMM:
+	case IB_WR_SEND_WITH_INV:
+		send_sqe = (struct erdma_send_sqe *)entry;
+		hw_op = ERDMA_OP_SEND;
+		if (op == IB_WR_SEND_WITH_IMM) {
+			hw_op = ERDMA_OP_SEND_WITH_IMM;
+			send_sqe->imm_data = send_wr->ex.imm_data;
+		} else if (op == IB_WR_SEND_WITH_INV) {
+			hw_op = ERDMA_OP_SEND_WITH_INV;
+			send_sqe->invalid_stag =
+				cpu_to_le32(send_wr->ex.invalidate_rkey);
+		}
+		wqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_OPCODE_MASK, hw_op);
+		length_field = &send_sqe->length;
+		wqe_size = sizeof(struct erdma_send_sqe);
+		sgl_offset = wqe_size;
+
+		break;
+	case IB_WR_REG_MR:
+		wqe_hdr |=
+			FIELD_PREP(ERDMA_SQE_HDR_OPCODE_MASK, ERDMA_OP_REG_MR);
+		regmr_sge = (struct erdma_reg_mr_sqe *)entry;
+		mr = to_emr(reg_wr(send_wr)->mr);
+
+		mr->access = ERDMA_MR_ACC_LR |
+			     to_erdma_access_flags(reg_wr(send_wr)->access);
+		regmr_sge->addr = cpu_to_le64(mr->ibmr.iova);
+		regmr_sge->length = cpu_to_le32(mr->ibmr.length);
+		regmr_sge->stag = cpu_to_le32(mr->ibmr.lkey);
+		attrs = FIELD_PREP(ERDMA_SQE_MR_MODE_MASK, 0) |
+			FIELD_PREP(ERDMA_SQE_MR_ACCESS_MASK, mr->access) |
+			FIELD_PREP(ERDMA_SQE_MR_MTT_CNT_MASK,
+				   mr->mem.mtt_nents);
+
+		if (mr->mem.mtt_nents < ERDMA_MAX_INLINE_MTT_ENTRIES) {
+			attrs |= FIELD_PREP(ERDMA_SQE_MR_MTT_TYPE_MASK, 0);
+			/* Copy SGLs to SQE content to accelerate */
+			memcpy(get_queue_entry(qp->kern_qp.sq_buf, idx + 1,
+					       qp->attrs.sq_size, SQEBB_SHIFT),
+			       mr->mem.mtt_buf, MTT_SIZE(mr->mem.mtt_nents));
+			wqe_size = sizeof(struct erdma_reg_mr_sqe) +
+				   MTT_SIZE(mr->mem.mtt_nents);
+		} else {
+			attrs |= FIELD_PREP(ERDMA_SQE_MR_MTT_TYPE_MASK, 1);
+			wqe_size = sizeof(struct erdma_reg_mr_sqe);
+		}
+
+		regmr_sge->attrs = cpu_to_le32(attrs);
+		goto out;
+	case IB_WR_LOCAL_INV:
+		wqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_OPCODE_MASK,
+				      ERDMA_OP_LOCAL_INV);
+		regmr_sge = (struct erdma_reg_mr_sqe *)entry;
+		regmr_sge->stag = cpu_to_le32(send_wr->ex.invalidate_rkey);
+		wqe_size = sizeof(struct erdma_reg_mr_sqe);
+		goto out;
+	default:
+		return -EOPNOTSUPP;
+	}
+
+	if (flags & IB_SEND_INLINE) {
+		ret = fill_inline_data(qp, send_wr, idx, sgl_offset,
+				       length_field);
+		if (ret < 0)
+			return -EINVAL;
+		wqe_size += ret;
+		wqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_SGL_LEN_MASK, ret);
+	} else {
+		ret = fill_sgl(qp, send_wr, idx, sgl_offset, length_field);
+		if (ret)
+			return -EINVAL;
+		wqe_size += send_wr->num_sge * sizeof(struct ib_sge);
+		wqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_SGL_LEN_MASK,
+				      send_wr->num_sge);
+	}
+
+out:
+	wqebb_cnt = SQEBB_COUNT(wqe_size);
+	wqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_WQEBB_CNT_MASK, wqebb_cnt - 1);
+	*pi += wqebb_cnt;
+	wqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_WQEBB_INDEX_MASK, *pi);
+
+	*entry = wqe_hdr;
+
+	return 0;
+}
+
+static void kick_sq_db(struct erdma_qp *qp, u16 pi)
+{
+	u64 db_data = FIELD_PREP(ERDMA_SQE_HDR_QPN_MASK, QP_ID(qp)) |
+		      FIELD_PREP(ERDMA_SQE_HDR_WQEBB_INDEX_MASK, pi);
+
+	*(u64 *)qp->kern_qp.sq_db_info = db_data;
+	writeq(db_data, qp->kern_qp.hw_sq_db);
+}
+
+int erdma_post_send(struct ib_qp *ibqp, const struct ib_send_wr *send_wr,
+		    const struct ib_send_wr **bad_send_wr)
+{
+	struct erdma_qp *qp = to_eqp(ibqp);
+	int ret = 0;
+	const struct ib_send_wr *wr = send_wr;
+	unsigned long flags;
+	u16 sq_pi;
+
+	if (!send_wr)
+		return -EINVAL;
+
+	spin_lock_irqsave(&qp->lock, flags);
+	sq_pi = qp->kern_qp.sq_pi;
+
+	while (wr) {
+		if ((u16)(sq_pi - qp->kern_qp.sq_ci) >= qp->attrs.sq_size) {
+			ret = -ENOMEM;
+			*bad_send_wr = send_wr;
+			break;
+		}
+
+		ret = erdma_push_one_sqe(qp, &sq_pi, wr);
+		if (ret) {
+			*bad_send_wr = wr;
+			break;
+		}
+		qp->kern_qp.sq_pi = sq_pi;
+		kick_sq_db(qp, sq_pi);
+
+		wr = wr->next;
+	}
+	spin_unlock_irqrestore(&qp->lock, flags);
+
+	return ret;
+}
+
+static int erdma_post_recv_one(struct erdma_qp *qp,
+			       const struct ib_recv_wr *recv_wr)
+{
+	struct erdma_rqe *rqe =
+		get_queue_entry(qp->kern_qp.rq_buf, qp->kern_qp.rq_pi,
+				qp->attrs.rq_size, RQE_SHIFT);
+
+	rqe->qe_idx = cpu_to_le16(qp->kern_qp.rq_pi + 1);
+	rqe->qpn = cpu_to_le32(QP_ID(qp));
+
+	if (recv_wr->num_sge == 0) {
+		rqe->length = 0;
+	} else if (recv_wr->num_sge == 1) {
+		rqe->stag = cpu_to_le32(recv_wr->sg_list[0].lkey);
+		rqe->to = cpu_to_le64(recv_wr->sg_list[0].addr);
+		rqe->length = cpu_to_le32(recv_wr->sg_list[0].length);
+	} else {
+		return -EINVAL;
+	}
+
+	*(u64 *)qp->kern_qp.rq_db_info = *(u64 *)rqe;
+	writeq(*(u64 *)rqe, qp->kern_qp.hw_rq_db);
+
+	qp->kern_qp.rwr_tbl[qp->kern_qp.rq_pi & (qp->attrs.rq_size - 1)] =
+		recv_wr->wr_id;
+	qp->kern_qp.rq_pi++;
+
+	return 0;
+}
+
+int erdma_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *recv_wr,
+		    const struct ib_recv_wr **bad_recv_wr)
+{
+	const struct ib_recv_wr *wr = recv_wr;
+	struct erdma_qp *qp = to_eqp(ibqp);
+	unsigned long flags;
+	int ret;
+
+	spin_lock_irqsave(&qp->lock, flags);
+
+	while (wr) {
+		ret = erdma_post_recv_one(qp, wr);
+		if (ret) {
+			*bad_recv_wr = wr;
+			break;
+		}
+		wr = wr->next;
+	}
+
+	spin_unlock_irqrestore(&qp->lock, flags);
+	return ret;
+}