[rdma-core,2/5] RDMA-CORE/erdma: Add userspace verbs implementation

Commit Message

Cheng Xu Dec. 24, 2021, 6:55 a.m. UTC

Imlementation of the erdma's 'struct verbs_context_ops' interface.
Due to doorbells may be drop by hardware in some situations, such as
hardware hot-upgrade, driver will keep the latest doorbell value of each
QP and CQ. We introduce the doorbell records to store the latest doorbell
values, and its allocation mechanism comes from dbre.c of mlx5.

Signed-off-by: Cheng Xu <[e-mail]>
---
 providers/erdma/erdma_db.c    | 110 ++++
 providers/erdma/erdma_verbs.c | 934 ++++++++++++++++++++++++++++++++++
 2 files changed, 1044 insertions(+)
 create mode 100644 providers/erdma/erdma_db.c
 create mode 100644 providers/erdma/erdma_verbs.c

Comments

Devesh Sharma Dec. 27, 2021, 6:29 a.m. UTC | #1

> -----Original Message-----
> From: Cheng Xu <[e-mail]>
> Sent: Friday, December 24, 2021 12:25 PM
> To: [e-mail]
> Cc: [e-mail]; [e-mail]; [e-mail];
> [e-mail]; [e-mail]
> Subject: [PATCH rdma-core 2/5] RDMA-CORE/erdma: Add userspace verbs
> implementation
> 
> Imlementation of the erdma's 'struct verbs_context_ops' interface.
> Due to doorbells may be drop by hardware in some situations, such as
> hardware hot-upgrade, driver will keep the latest doorbell value of each QP
> and CQ. We introduce the doorbell records to store the latest doorbell
> values, and its allocation mechanism comes from dbre.c of mlx5.
> 
> Signed-off-by: Cheng Xu <[e-mail]>
> ---
>  providers/erdma/erdma_db.c    | 110 ++++
>  providers/erdma/erdma_verbs.c | 934
> ++++++++++++++++++++++++++++++++++
>  2 files changed, 1044 insertions(+)
>  create mode 100644 providers/erdma/erdma_db.c  create mode 100644
> providers/erdma/erdma_verbs.c
> 
> diff --git a/providers/erdma/erdma_db.c b/providers/erdma/erdma_db.c
> new file mode 100644 index 00000000..83db76d1
> --- /dev/null
> +++ b/providers/erdma/erdma_db.c
> @@ -0,0 +1,110 @@
> +// SPDX-License-Identifier: GPL-2.0 or OpenIB.org BSD (MIT) See COPYING
> +file
> +
> +// Authors: Cheng Xu <[e-mail]> // Copyright (c)
> [phone], Alibaba Group.
> +
> +// Copyright (c) 2012 Mellanox Technologies, Inc.  All rights reserved.
> +
> +#include <inttypes.h>
> +#include <stdlib.h>
> +#include <util/util.h>
> +
> +#include "erdma.h"
> +#include "erdma_db.h"
> +
> +#define ERDMA_DBRECORDS_SIZE 16
> +
> +struct erdma_dbrecord_page {
> +	struct erdma_dbrecord_page *prev, *next;
> +	void *page_buf;
> +	int cnt;
> +	int used;
> +	unsigned long free[0];
> +};
> +
> +uint64_t *erdma_alloc_dbrecords(struct erdma_context *ctx) {
> +	struct erdma_dbrecord_page *page = NULL;
> +	uint64_t *db_records = NULL;
> +	int dbrecords_per_page, nlongs = 0, bits_perlong = (8 *
> sizeof(unsigned long));
> +	int i, j, rv;
> +
> +	pthread_mutex_lock(&ctx->dbrecord_pages_mutex);
> +
> +	for (page = ctx->dbrecord_pages; page; page = page->next)
> +		if (page->used < page->cnt)
> +			goto found;
> +
> +	dbrecords_per_page = ctx->page_size / ERDMA_DBRECORDS_SIZE;
> +	nlongs = align(dbrecords_per_page, bits_perlong) / bits_perlong;
> +	page = malloc(sizeof(*page) + nlongs * sizeof(unsigned long));
> +	if (!page)
> +		goto out;
> +
> +	rv = posix_memalign(&page->page_buf, ctx->page_size, ctx-
> >page_size);
> +	if (rv) {
> +		free(page);
> +		goto out;
> +	}
> +
> +	page->cnt = dbrecords_per_page;
> +	page->used = 0;
> +	for (i = 0; i < nlongs; i++)
> +		page->free[i] = ~0UL;
> +
> +	page->prev = NULL;
> +	page->next = ctx->dbrecord_pages;
> +	ctx->dbrecord_pages = page;
> +	if (page->next)
> +		page->next->prev = page;
> +
> +found:
> +	++page->used;
> +
> +	for (i = 0; !page->free[i]; ++i)
> +		;/* nothing */
Why?
> +
> +	j = ffsl(page->free[i]) - 1;
> +	page->free[i] &= ~(1UL << j);
> +
> +	db_records = page->page_buf + (i * bits_perlong + j) *
> +ERDMA_DBRECORDS_SIZE;
> +
> +out:
> +	pthread_mutex_unlock(&ctx->dbrecord_pages_mutex);
> +
> +	return db_records;
> +}
> +
> +void erdma_dealloc_dbrecords(struct erdma_context *ctx, uint64_t
> +*dbrecords) {
> +	struct erdma_dbrecord_page *page;
> +	int page_mask = ~(ctx->page_size - 1);
> +	int idx;
> +
> +	pthread_mutex_lock(&ctx->dbrecord_pages_mutex);
> +	for (page = ctx->dbrecord_pages; page; page = page->next)
> +		if (((uintptr_t)dbrecords & page_mask) == (uintptr_t)page-
> >page_buf)
> +			break;
> +
> +	if (!page)
> +		goto out;
> +
> +	idx = ((void *)dbrecords - page->page_buf) /
> ERDMA_DBRECORDS_SIZE;
> +	page->free[idx / (8 * sizeof(unsigned long))] |=
> +		1UL << (idx % (8 * sizeof(unsigned long)));
> +
> +	if (!--page->used) {
> +		if (page->prev)
> +			page->prev->next = page->next;
> +		else
> +			ctx->dbrecord_pages = page->next;
> +		if (page->next)
> +			page->next->prev = page->prev;
> +
> +		free(page->page_buf);
> +		free(page);
> +	}
> +
> +out:
> +	pthread_mutex_unlock(&ctx->dbrecord_pages_mutex);
> +}
> diff --git a/providers/erdma/erdma_verbs.c
> b/providers/erdma/erdma_verbs.c new file mode 100644 index
> 00000000..3c1c9769
> --- /dev/null
> +++ b/providers/erdma/erdma_verbs.c
> @@ -0,0 +1,934 @@
> +// SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
> +
> +// Authors: Cheng Xu <[e-mail]> // Copyright (c)
> [phone], Alibaba Group.
> +// Authors: Bernard Metzler <[e-mail]> // Copyright (c)
> [phone], IBM Corporation
> +
> +#include <ccan/minmax.h>
> +#include <endian.h>
> +#include <stdio.h>
> +#include <stdlib.h>
> +#include <sys/mman.h>
> +#include <sys/types.h>
> +#include <unistd.h>
> +#include <util/mmio.h>
> +#include <util/udma_barrier.h>
> +#include <util/util.h>
> +
> +#include "erdma.h"
> +#include "erdma_abi.h"
> +#include "erdma_db.h"
> +#include "erdma_hw.h"
> +#include "erdma_verbs.h"
> +
> +int erdma_query_device(struct ibv_context *ctx,
> +		       const struct ibv_query_device_ex_input *input,
> +		       struct ibv_device_attr_ex *attr, size_t attr_size) {
> +	struct ib_uverbs_ex_query_device_resp resp;
> +	size_t resp_size = sizeof(resp);
> +	uint64_t raw_fw_ver;
> +	unsigned int major, minor, sub_minor;
> +	int rv;
> +
> +	rv = ibv_cmd_query_device_any(ctx, input, attr, attr_size, &resp,
> &resp_size);
> +	if (rv)
> +		return rv;
> +
> +	raw_fw_ver = resp.base.fw_ver;
> +	major = (raw_fw_ver >> 32) & 0xffff;
> +	minor = (raw_fw_ver >> 16) & 0xffff;
> +	sub_minor = raw_fw_ver & 0xffff;
> +
> +	snprintf(attr->orig_attr.fw_ver, sizeof(attr->orig_attr.fw_ver),
> +		"%d.%d.%d", major, minor, sub_minor);
> +
> +	return 0;
> +}
> +
> +int erdma_query_port(struct ibv_context *ctx, uint8_t port, struct
> +ibv_port_attr *attr) {
> +	struct ibv_query_port cmd;
> +
> +	memset(&cmd, 0, sizeof(cmd));
> +
> +	return ibv_cmd_query_port(ctx, port, attr, &cmd, sizeof(cmd)); }
> +
> +int erdma_query_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr,
> +		   int attr_mask, struct ibv_qp_init_attr *init_attr) {
> +	struct ibv_query_qp cmd;
> +
> +	memset(&cmd, 0, sizeof(cmd));
> +
> +	return ibv_cmd_query_qp(qp, attr, attr_mask, init_attr, &cmd,
> +sizeof(cmd)); }
> +
> +struct ibv_pd *erdma_alloc_pd(struct ibv_context *ctx) {
> +	struct ibv_alloc_pd cmd;
> +	struct ib_uverbs_alloc_pd_resp resp;
> +	struct ibv_pd *pd;
> +
> +	memset(&cmd, 0, sizeof(cmd));
> +
> +	pd = calloc(1, sizeof(*pd));
> +	if (!pd)
> +		return NULL;
> +
> +	if (ibv_cmd_alloc_pd(ctx, pd, &cmd, sizeof(cmd), &resp,
> sizeof(resp))) {
> +		free(pd);
> +		return NULL;
> +	}
> +
> +	return pd;
> +}
> +
> +int erdma_free_pd(struct ibv_pd *pd)
> +{
> +	int rv;
> +
> +	rv = ibv_cmd_dealloc_pd(pd);
> +	if (rv)
> +		return rv;
> +
> +	free(pd);
> +	return 0;
> +}
> +
> +struct ibv_mr *erdma_reg_mr(struct ibv_pd *pd, void *addr, size_t len,
> +			    uint64_t hca_va, int access)
> +{
> +	struct ibv_reg_mr cmd;
> +	struct ib_uverbs_reg_mr_resp resp;
> +	struct verbs_mr *vmr;
> +	int ret;
> +
> +	vmr = calloc(1, sizeof(*vmr));
> +	if (!vmr)
> +		return NULL;
> +
> +	ret = ibv_cmd_reg_mr(pd, addr, len, hca_va, access, vmr, &cmd,
> sizeof(cmd),
> +		&resp, sizeof(resp));
> +	if (ret) {
> +		free(vmr);
> +		return NULL;
> +	}
> +
> +	return &vmr->ibv_mr;
> +}
> +
> +int erdma_dereg_mr(struct verbs_mr *vmr) {
> +	int ret;
> +
> +	ret = ibv_cmd_dereg_mr(vmr);
> +	if (ret)
> +		return ret;
> +
> +	free(vmr);
> +	return 0;
> +}
> +
> +int erdma_notify_cq(struct ibv_cq *ibcq, int solicited) {
> +	struct erdma_cq *cq = to_ecq(ibcq);
> +	uint64_t db_data;
> +	int ret;
> +
> +	ret = pthread_spin_lock(&cq->lock);
> +	if (ret)
> +		return ret;
> +
> +	db_data = FIELD_PREP(ERDMA_CQDB_EQN_MASK, cq-
> >comp_vector + 1) |
> +		FIELD_PREP(ERDMA_CQDB_CQN_MASK, cq->id) |
> +		FIELD_PREP(ERDMA_CQDB_ARM_MASK, 1) |
> +		FIELD_PREP(ERDMA_CQDB_SOL_MASK, solicited) |
> +		FIELD_PREP(ERDMA_CQDB_CMDSN_MASK, cq->cmdsn) |
> +		FIELD_PREP(ERDMA_CQDB_CI_MASK, cq->ci);
> +
> +	*(__le64 *)cq->db_record = htole64(db_data);
> +	udma_to_device_barrier();
> +	mmio_write64_le(cq->db, htole64(db_data));
> +
> +	pthread_spin_unlock(&cq->lock);
> +
> +	return ret;
> +}
> +
> +struct ibv_cq *erdma_create_cq(struct ibv_context *ctx, int num_cqe,
> +			       struct ibv_comp_channel *channel, int
> comp_vector) {
> +	struct erdma_context *ectx;
> +	struct erdma_cmd_create_cq cmd = {};
> +	struct erdma_cmd_create_cq_resp resp = {};
> +	struct erdma_cq *cq;
> +	uint64_t *db_records = NULL;
> +	int cq_size;
> +	int rv;
> +
> +	ectx = to_ectx(ctx);
> +
> +	cq = calloc(1, sizeof(*cq));
> +	if (!cq)
> +		return NULL;
> +
> +	if (num_cqe < 64)
> +		num_cqe = 64;
> +
> +	num_cqe = roundup_pow_of_two(num_cqe);
> +	cq_size = num_cqe * sizeof(struct erdma_cqe);
> +	cq_size = ERDMA_SIZE_TO_NPAGE(cq_size) <<
> ERDMA_PAGE_SHIFT;
> +
> +	rv = posix_memalign((void **)&cq->queue, ERDMA_PAGE_SIZE,
> cq_size);
> +	if (rv) {
> +		errno = rv;
> +		free(cq);
> +		return NULL;
> +	}
> +
> +	db_records = erdma_alloc_dbrecords(ectx);
> +	if (!db_records) {
> +		errno = ENOMEM;
> +		goto error_alloc;
> +	}
> +
> +	cmd.db_record_va = (__u64)db_records;
> +	cmd.qbuf_va = (uint64_t)cq->queue;
> +	cmd.qbuf_len = cq_size;
> +
> +	rv = ibv_cmd_create_cq(ctx, num_cqe, channel, comp_vector, &cq-
> >base_cq,
> +		&cmd.ibv_cmd, sizeof(cmd), &resp.ibv_resp, sizeof(resp));
> +	if (rv) {
> +		free(cq);
> +		errno = EIO;
> +		goto error_alloc;
> +	}
> +
> +	pthread_spin_init(&cq->lock, PTHREAD_PROCESS_PRIVATE);
> +
> +	*db_records = 0;
> +	cq->db_record = db_records;
> +
> +	cq->id = resp.cq_id;
> +	cq->depth = resp.num_cqe;
> +	cq->owner = 1;
> +
> +	cq->db = ectx->cdb;
> +	cq->db_offset = (cq->id & (ERDMA_PAGE_SIZE / ERDMA_CQDB_SIZE
> - 1)) * ERDMA_CQDB_SIZE;
> +	cq->db += cq->db_offset;
> +
> +	cq->comp_vector = comp_vector;
> +
> +	return &cq->base_cq;
> +
> +error_alloc:
> +	if (db_records)
> +		erdma_dealloc_dbrecords(ectx, db_records);
> +
> +	if (cq->queue)
> +		free(cq->queue);
> +
> +	free(cq);
> +
> +	return NULL;
> +}
> +
> +int erdma_destroy_cq(struct ibv_cq *base_cq) {
> +	struct erdma_cq *cq = to_ecq(base_cq);
> +	struct erdma_context *ctx = to_ectx(base_cq->context);
> +	int rv;
> +
> +	pthread_spin_lock(&cq->lock);
> +	rv = ibv_cmd_destroy_cq(base_cq);
> +	if (rv) {
> +		pthread_spin_unlock(&cq->lock);
> +		errno = EIO;
> +		return rv;
> +	}
> +	pthread_spin_destroy(&cq->lock);
> +
> +	if (cq->db_record)
> +		erdma_dealloc_dbrecords(ctx, cq->db_record);
> +
> +	if (cq->queue)
> +		free(cq->queue);
> +
> +	free(cq);
> +
> +	return 0;
> +}
> +
> +static void __erdma_alloc_dbs(struct erdma_qp *qp, struct erdma_context
> +*ctx) {
> +	uint32_t qpn = qp->id;
> +
> +	if (ctx->sdb_type == ERDMA_SDB_PAGE) {
> +		/* qpn[4:0] as the index in this db page. */
> +		qp->sq.db = ctx->sdb + (qpn & 31) * ERDMA_SQDB_SIZE;
> +	} else if (ctx->sdb_type == ERDMA_SDB_ENTRY) {
> +		/* for type 'ERDMA_SDB_ENTRY', each uctx has 2 dwqe,
> totally takes 256Bytes. */
> +		qp->sq.db = ctx->sdb + ctx->sdb_offset * 256;
Generally we use macros to define hard-coded integers. E.g 256 should use a macro.
> +	} else {
> +		/* qpn[4:0] as the index in this db page. */
> +		qp->sq.db = ctx->sdb + (qpn & 31) * ERDMA_SQDB_SIZE;
> +	}
> +
> +	/* qpn[6:0] as the index in this rq db page. */
> +	qp->rq.db = ctx->rdb + (qpn & 127) * ERDMA_RQDB_SPACE_SIZE; }
> +
> +struct ibv_qp *erdma_create_qp(struct ibv_pd *pd, struct
> +ibv_qp_init_attr *attr) {
> +	struct erdma_cmd_create_qp cmd = {};
> +	struct erdma_cmd_create_qp_resp resp = {};
> +	struct erdma_qp *qp;
> +	struct ibv_context *base_ctx = pd->context;
> +	struct erdma_context *ctx = to_ectx(base_ctx);
> +	uint64_t *db_records  = NULL;
> +	int rv, tbl_idx, tbl_off;
> +	int sq_size = 0, rq_size = 0, total_bufsize = 0;
> +
> +	memset(&cmd, 0, sizeof(cmd));
> +	memset(&resp, 0, sizeof(resp));
No need of memset due to declaration step.
> +
> +	qp = calloc(1, sizeof(*qp));
> +	if (!qp)
> +		return NULL;
> +
> +	sq_size = roundup_pow_of_two(attr->cap.max_send_wr *
> MAX_WQEBB_PER_SQE) << SQEBB_SHIFT;
> +	sq_size = align(sq_size, ctx->page_size);
> +	rq_size = align(roundup_pow_of_two(attr->cap.max_recv_wr) <<
> RQE_SHIFT, ctx->page_size);
> +	total_bufsize = sq_size + rq_size;
> +	rv = posix_memalign(&qp->qbuf, ctx->page_size, total_bufsize);
> +	if (rv || !qp->qbuf) {
> +		errno = ENOMEM;
> +		goto error_alloc;
> +	}
> +
> +	db_records = erdma_alloc_dbrecords(ctx);
> +	if (!db_records) {
> +		errno = ENOMEM;
> +		goto error_alloc;
> +	}
> +
> +	cmd.db_record_va = (__u64)db_records;
> +	cmd.qbuf_va = (__u64)qp->qbuf;
> +	cmd.qbuf_len = (__u32)total_bufsize;
> +
> +	rv = ibv_cmd_create_qp(pd, &qp->base_qp, attr, &cmd.ibv_cmd,
> +		sizeof(cmd), &resp.ibv_resp, sizeof(resp));
> +	if (rv) {
> +		errno = EIO;
> +		goto error_alloc;
> +	}
> +
> +	qp->id = resp.qp_id;
> +
> +	pthread_mutex_lock(&ctx->qp_table_mutex);
> +	tbl_idx = qp->id >> ERDMA_QP_TABLE_SHIFT;
> +	tbl_off = qp->id & ERDMA_QP_TABLE_MASK;
> +
> +	if (ctx->qp_table[tbl_idx].refcnt == 0) {
> +		ctx->qp_table[tbl_idx].table = calloc(ERDMA_PAGE_SIZE,
> sizeof(struct erdma_qp *));
> +		if (!ctx->qp_table[tbl_idx].table) {
> +			errno = ENOMEM;
> +			goto fail;
> +		}
> +	}
> +
> +	/* exist qp */
> +	if (ctx->qp_table[tbl_idx].table[tbl_off]) {
> +		errno = EBUSY;
> +		goto fail;
> +	}
> +
> +	ctx->qp_table[tbl_idx].table[tbl_off] = qp;
> +	ctx->qp_table[tbl_idx].refcnt++;
> +	pthread_mutex_unlock(&ctx->qp_table_mutex);
> +
> +	qp->sq.qbuf = qp->qbuf;
> +	qp->rq.qbuf = qp->qbuf + resp.rq_offset;
> +	qp->sq.depth = resp.num_sqe;
> +	qp->rq.depth = resp.num_rqe;
> +	qp->sq_sig_all = attr->sq_sig_all;
> +	qp->sq.size = resp.num_sqe * SQEBB_SIZE;
> +	qp->rq.size = resp.num_rqe * sizeof(struct erdma_rqe);
> +
> +	/* doorbell allocation. */
> +	__erdma_alloc_dbs(qp, ctx);
> +
> +	pthread_spin_init(&qp->sq_lock, PTHREAD_PROCESS_PRIVATE);
> +	pthread_spin_init(&qp->rq_lock, PTHREAD_PROCESS_PRIVATE);
> +
> +	*db_records = 0;
> +	*(db_records + 1) = 0;
> +	qp->db_records = db_records;
> +	qp->sq.db_record = db_records;
> +	qp->rq.db_record = db_records + 1;
> +
> +	qp->rq.wr_tbl = calloc(qp->rq.depth, sizeof(uint64_t));
> +	if (!qp->rq.wr_tbl)
> +		goto fail;
> +
> +	qp->sq.wr_tbl = calloc(qp->sq.depth, sizeof(uint64_t));
> +	if (!qp->sq.wr_tbl)
> +		goto fail;
> +
> +
> +	return &qp->base_qp;
> +fail:
> +	if (qp->sq.wr_tbl)
> +		free(qp->sq.wr_tbl);
> +
> +	if (qp->rq.wr_tbl)
> +		free(qp->rq.wr_tbl);
> +
> +	ibv_cmd_destroy_qp(&qp->base_qp);
> +
> +error_alloc:
> +	if (db_records)
> +		erdma_dealloc_dbrecords(ctx, db_records);
> +
> +	if (qp->qbuf)
> +		free(qp->qbuf);
> +
> +	free(qp);
> +
> +	return NULL;
> +}
> +
> +int erdma_modify_qp(struct ibv_qp *base_qp, struct ibv_qp_attr *attr,
> +int attr_mask) {
> +	struct ibv_modify_qp cmd;
> +	struct erdma_qp *qp = to_eqp(base_qp);
> +	int rv;
> +
> +	memset(&cmd, 0, sizeof(cmd));
> +
> +	pthread_spin_lock(&qp->sq_lock);
> +	pthread_spin_lock(&qp->rq_lock);
> +
> +	rv = ibv_cmd_modify_qp(base_qp, attr, attr_mask, &cmd,
> sizeof(cmd));
> +
> +	pthread_spin_unlock(&qp->rq_lock);
> +	pthread_spin_unlock(&qp->sq_lock);
> +
> +	return rv;
> +}
> +
> +int erdma_destroy_qp(struct ibv_qp *base_qp) {
> +	struct erdma_qp *qp = to_eqp(base_qp);
> +	struct ibv_context *base_ctx = base_qp->pd->context;
> +	struct erdma_context *ctx = to_ectx(base_ctx);
> +	int rv, tbl_idx, tbl_off;
> +
> +	pthread_spin_lock(&qp->sq_lock);
> +	pthread_spin_lock(&qp->rq_lock);
Why to hold these?
> +
> +	pthread_mutex_lock(&ctx->qp_table_mutex);
> +	tbl_idx = qp->id >> ERDMA_QP_TABLE_SHIFT;
> +	tbl_off = qp->id & ERDMA_QP_TABLE_MASK;
> +
> +	ctx->qp_table[tbl_idx].table[tbl_off] = NULL;
> +	ctx->qp_table[tbl_idx].refcnt--;
> +
> +	if (ctx->qp_table[tbl_idx].refcnt == 0) {
> +		free(ctx->qp_table[tbl_idx].table);
> +		ctx->qp_table[tbl_idx].table = NULL;
> +	}
> +
> +	pthread_mutex_unlock(&ctx->qp_table_mutex);
> +
> +	rv = ibv_cmd_destroy_qp(base_qp);
> +	if (rv) {
> +		pthread_spin_unlock(&qp->rq_lock);
> +		pthread_spin_unlock(&qp->sq_lock);
> +		return rv;
> +	}
> +	pthread_spin_destroy(&qp->rq_lock);
> +	pthread_spin_destroy(&qp->sq_lock);
> +
> +	if (qp->db_records)
> +		erdma_dealloc_dbrecords(ctx, qp->db_records);
> +
> +	if (qp->qbuf)
> +		free(qp->qbuf);
> +
> +	free(qp);
> +
> +	return 0;
> +}
> +
> +static int erdma_push_one_sqe(struct erdma_qp *qp, struct ibv_send_wr
> +*wr, uint16_t *sq_pi) {
> +	uint16_t tmp_pi = *sq_pi;
> +	void *sqe;
> +	uint64_t sqe_hdr;
> +	struct erdma_write_sqe *write_sqe;
> +	struct erdma_send_sqe *send_sqe;
> +	struct erdma_readreq_sqe *read_sqe;
> +	uint32_t wqe_size = 0;
> +	__le32 *length_field = NULL;
> +	struct erdma_sge *sgl_base = NULL;
> +	uint32_t i, bytes = 0;
> +	uint32_t sgl_off, sgl_idx, wqebb_cnt, opcode;
> +
> +	sqe = get_sq_wqebb(qp, tmp_pi);
> +	/* Clear the first 8Byte of the wqe hdr. */
> +	*(uint64_t *)sqe = 0;
> +
> +	qp->sq.wr_tbl[tmp_pi & (qp->sq.depth - 1)] = wr->wr_id;
> +
> +	sqe_hdr = FIELD_PREP(ERDMA_SQE_HDR_QPN_MASK, qp->id) |
> +		FIELD_PREP(ERDMA_SQE_HDR_CE_MASK, wr->send_flags &
> IBV_SEND_SIGNALED ? 1 : 0) |
> +		FIELD_PREP(ERDMA_SQE_HDR_CE_MASK, qp->sq_sig_all) |
> +		FIELD_PREP(ERDMA_SQE_HDR_SE_MASK, wr->send_flags &
> IBV_SEND_SOLICITED ? 1 : 0) |
> +		FIELD_PREP(ERDMA_SQE_HDR_FENCE_MASK, wr-
> >send_flags & IBV_SEND_FENCE ? 1 : 0) |
> +		FIELD_PREP(ERDMA_SQE_HDR_INLINE_MASK, wr-
> >send_flags &
> +IBV_SEND_INLINE ? 1 : 0);
> +
> +	switch (wr->opcode) {
> +	case IBV_WR_RDMA_WRITE:
> +	case IBV_WR_RDMA_WRITE_WITH_IMM:
> +		opcode = wr->opcode == IBV_WR_RDMA_WRITE ?
> +			ERDMA_OP_WRITE :
> ERDMA_OP_WRITE_WITH_IMM;
> +		sqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_OPCODE_MASK,
> opcode);
> +		write_sqe = (struct erdma_write_sqe *)sqe;
> +		write_sqe->imm_data = wr->imm_data;
> +		write_sqe->sink_stag = htole32(wr->wr.rdma.rkey);
> +		write_sqe->sink_to_low = htole32(wr-
> >wr.rdma.remote_addr & 0xFFFFFFFF);
> +		write_sqe->sink_to_high = htole32((wr-
> >wr.rdma.remote_addr >> 32) &
> +0xFFFFFFFF);
> +
> +		length_field = &write_sqe->length;
> +		sgl_base = get_sq_wqebb(qp, tmp_pi + 1);
> +		/* sgl is in next wqebb. */
> +		sgl_off = 0;
> +		sgl_idx = tmp_pi + 1;
> +		wqe_size = sizeof(struct erdma_write_sqe);
> +
> +		break;
> +	case IBV_WR_SEND:
> +	case IBV_WR_SEND_WITH_IMM:
> +		opcode = wr->opcode == IBV_WR_SEND ? ERDMA_OP_SEND
> : ERDMA_OP_SEND_WITH_IMM;
> +		sqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_OPCODE_MASK,
> opcode);
> +		send_sqe = (struct erdma_send_sqe *)sqe;
> +		send_sqe->imm_data = wr->imm_data;
> +
> +		length_field = &send_sqe->length;
> +		sgl_base = (void *)send_sqe;
> +		/* sgl is in the half of current wqebb */
> +		sgl_off = 16;
> +		sgl_idx = tmp_pi;
> +		wqe_size = sizeof(struct erdma_send_sqe);
> +
> +		break;
> +	case IBV_WR_RDMA_READ:
> +		sqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_OPCODE_MASK,
> ERDMA_OP_READ);
> +
> +		read_sqe = (struct erdma_readreq_sqe *)sqe;
> +
> +		read_sqe->sink_to_low = htole32(wr->sg_list->addr &
> 0xFFFFFFFF);
> +		read_sqe->sink_to_high = htole32((wr->sg_list->addr >> 32)
> & 0xFFFFFFFF);
> +		read_sqe->sink_stag = htole32(wr->sg_list->lkey);
> +		read_sqe->length = htole32(wr->sg_list->length);
> +
> +		struct erdma_sge *sgl = (struct erdma_sge
> *)get_sq_wqebb(qp, tmp_pi +
> +1);
> +
> +		sgl->laddr = htole64(wr->wr.rdma.remote_addr);
> +		sgl->length = htole32(wr->sg_list->length);
> +		sgl->lkey = htole32(wr->wr.rdma.rkey);
> +
> +		wqe_size = sizeof(struct erdma_readreq_sqe);
> +
> +		goto out;
> +	default:
> +		return -EINVAL;
> +	}
> +
> +	if (wr->send_flags & IBV_SEND_INLINE) {
> +		char *data = (char *)sgl_base;
> +		uint32_t remain_size;
> +		uint32_t copy_size;
> +		uint32_t data_off;
> +
> +		i = 0;
> +		bytes = 0;
> +
> +		/* Allow more than ERDMA_MAX_SGE, since content copied
> here */
> +		while (i < wr->num_sge) {
> +			bytes += wr->sg_list[i].length;
> +			if (bytes > (int)ERDMA_MAX_INLINE)
> +				return -EINVAL;
> +
> +			remain_size = wr->sg_list[i].length;
> +			data_off = 0;
> +
> +			while (1) {
> +				copy_size = min(remain_size, SQEBB_SIZE -
> sgl_off);
> +				memcpy(data + sgl_off,
> +					(void *)(uintptr_t)wr->sg_list[i].addr
> + data_off,
> +					copy_size);
> +				remain_size -= copy_size;
> +
> +				/* Update sgl_offset. */
> +				sgl_idx += ((sgl_off + copy_size) >>
> SQEBB_SHIFT);
> +				sgl_off = (sgl_off + copy_size) & (SQEBB_SIZE
> - 1);
> +				data_off += copy_size;
> +				data = get_sq_wqebb(qp, sgl_idx);
> +
> +				if (!remain_size)
> +					break;
> +			};
> +
> +			i++;
> +		}
> +
> +		*length_field = htole32(bytes);
> +		wqe_size += bytes;
> +		sqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_SGL_LEN_MASK,
> bytes);
> +	} else {
> +		char *sgl = (char *)sgl_base;
> +
> +		if (wr->num_sge > ERDMA_MAX_SEND_SGE)
> +			return -EINVAL;
> +
> +		i = 0;
> +		bytes = 0;
> +
> +		while (i < wr->num_sge) {
> +			bytes += wr->sg_list[i].length;
> +			memcpy(sgl + sgl_off, &wr->sg_list[i], sizeof(struct
> ibv_sge));
> +
> +			if (sgl_off == 0)
> +				*(uint32_t *)(sgl + 28) = qp->id;
> +
> +			sgl_idx += (sgl_off == sizeof(struct ibv_sge) ? 1 : 0);
> +			sgl = get_sq_wqebb(qp, sgl_idx);
> +			sgl_off = sizeof(struct ibv_sge) - sgl_off;
> +
> +			i++;
> +		}
> +
> +		*length_field = htole32(bytes);
> +		sqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_SGL_LEN_MASK,
> wr->num_sge);
> +		wqe_size += wr->num_sge * sizeof(struct ibv_sge);
> +	}
> +
> +out:
> +	wqebb_cnt = SQEBB_COUNT(wqe_size);
> +	assert(wqebb_cnt <= MAX_WQEBB_PER_SQE);
> +	sqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_WQEBB_CNT_MASK,
> wqebb_cnt - 1);
> +	sqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_WQEBB_INDEX_MASK,
> tmp_pi +
> +wqebb_cnt);
> +
> +	*(__le64 *)sqe = htole64(sqe_hdr);
> +	*sq_pi = tmp_pi + wqebb_cnt;
> +
> +	return 0;
> +}
> +
> +int erdma_post_send(struct ibv_qp *base_qp, struct ibv_send_wr *wr,
> +struct ibv_send_wr **bad_wr) {
> +	struct erdma_qp *qp = to_eqp(base_qp);
> +	uint16_t sq_pi;
> +	int new_sqe = 0, rv = 0;
> +
> +	*bad_wr = NULL;
> +
> +	if (base_qp->state == IBV_QPS_ERR) {
Post_send is allowed in Error state. Thus the check is redundant.
> +		*bad_wr = wr;
> +		return -EIO;
> +	}
> +
> +	pthread_spin_lock(&qp->sq_lock);
> +
> +	sq_pi = qp->sq.pi;
> +
> +	while (wr) {
> +		if ((uint16_t)(sq_pi - qp->sq.ci) >= qp->sq.depth) {
> +			rv = -ENOMEM;
> +			*bad_wr = wr;
> +			break;
> +		}
> +
> +		rv = erdma_push_one_sqe(qp, wr, &sq_pi);
> +		if (rv) {
> +			*bad_wr = wr;
> +			break;
> +		}
> +
> +		new_sqe++;
> +		wr = wr->next;
> +	}
> +
> +	if (new_sqe) {
> +		qp->sq.pi = sq_pi;
> +		__kick_sq_db(qp, sq_pi); /* normal doorbell. */
> +	}
> +
> +	pthread_spin_unlock(&qp->sq_lock);
> +
> +	return rv;
> +}
> +
> +static int push_recv_wqe(struct erdma_qp *qp, struct ibv_recv_wr *wr) {
> +	uint16_t rq_pi = qp->rq.pi;
> +	uint16_t idx = rq_pi & (qp->rq.depth - 1);
> +	struct erdma_rqe *rqe = (struct erdma_rqe *)qp->rq.qbuf + idx;
> +
> +	if ((uint16_t)(rq_pi - qp->rq.ci) == qp->rq.depth)
> +		return -ENOMEM;
> +
> +	rqe->qe_idx = htole16(rq_pi + 1);
> +	rqe->qpn = htole32(qp->id);
> +	qp->rq.wr_tbl[idx] = wr->wr_id;
> +
> +	if (wr->num_sge == 0) {
> +		rqe->length = 0;
> +	} else if (wr->num_sge == 1) {
> +		rqe->stag = htole32(wr->sg_list[0].lkey);
> +		rqe->to = htole64(wr->sg_list[0].addr);
> +		rqe->length = htole32(wr->sg_list[0].length);
> +	} else {
> +		return -EINVAL;
> +	}
> +
> +	*(__le64 *)qp->rq.db_record = *(__le64 *)rqe;
> +	udma_to_device_barrier();
> +	mmio_write64_le(qp->rq.db, *(__le64 *)rqe);
> +
> +	qp->rq.pi = rq_pi + 1;
> +
> +	return 0;
> +}
> +
> +int erdma_post_recv(struct ibv_qp *base_qp, struct ibv_recv_wr *wr,
> +struct ibv_recv_wr **bad_wr) {
> +	struct erdma_qp *qp = to_eqp(base_qp);
> +	int ret = 0;
> +
> +	if (base_qp->state == IBV_QPS_ERR) {
> +		*bad_wr = wr;
> +		return -EIO;
> +	}
> +
> +	pthread_spin_lock(&qp->rq_lock);
> +
> +	while (wr) {
> +		ret = push_recv_wqe(qp, wr);
> +		if (ret) {
> +			*bad_wr = wr;
> +			break;
> +		}
> +
> +		wr = wr->next;
> +	}
> +
> +	pthread_spin_unlock(&qp->rq_lock);
> +
> +	return ret;
> +}
> +
> +
> +void erdma_cq_event(struct ibv_cq *ibcq) {
> +	struct erdma_cq *cq = to_ecq(ibcq);
> +
> +	cq->cmdsn++;
> +}
> +
> +static const struct {
> +	enum erdma_opcode erdma;
> +	enum ibv_wc_opcode base;
> +} map_cqe_opcode[ERDMA_NUM_OPCODES] = {
> +	{ ERDMA_OP_WRITE, IBV_WC_RDMA_WRITE },
> +	{ ERDMA_OP_READ, IBV_WC_RDMA_READ },
> +	{ ERDMA_OP_SEND, IBV_WC_SEND },
> +	{ ERDMA_OP_SEND_WITH_IMM, IBV_WC_SEND },
> +	{ ERDMA_OP_RECEIVE, IBV_WC_RECV },
> +	{ ERDMA_OP_RECV_IMM, IBV_WC_RECV_RDMA_WITH_IMM },
> +	{ ERDMA_OP_RECV_INV, IBV_WC_LOCAL_INV},  /* can not appear
> */
> +	{ ERDMA_OP_REQ_ERR, IBV_WC_RECV },  /* can not appear */
> +	{ ERDNA_OP_READ_RESPONSE, IBV_WC_RECV }, /* can not appear
> */
> +	{ ERDMA_OP_WRITE_WITH_IMM, IBV_WC_RDMA_WRITE },
> +	{ ERDMA_OP_RECV_ERR, IBV_WC_RECV_RDMA_WITH_IMM }, /*
> can not appear */
> +	{ ERDMA_OP_INVALIDATE, IBV_WC_LOCAL_INV },
> +	{ ERDMA_OP_RSP_SEND_IMM, IBV_WC_RECV },
> +	{ ERDMA_OP_SEND_WITH_INV, IBV_WC_SEND },
> +	{ ERDMA_OP_REG_MR, IBV_WC_RECV }, /* can not appear */
> +	{ ERDMA_OP_LOCAL_INV, IBV_WC_LOCAL_INV },
> +	{ ERDMA_OP_READ_WITH_INV, IBV_WC_RDMA_READ }, };
> +
> +static const struct {
> +	enum erdma_wc_status erdma;
> +	enum ibv_wc_status base;
> +	enum erdma_vendor_err vendor;
> +} map_cqe_status[ERDMA_NUM_WC_STATUS] = {
> +	{ ERDMA_WC_SUCCESS, IBV_WC_SUCCESS,
> ERDMA_WC_VENDOR_NO_ERR },
> +	{ ERDMA_WC_GENERAL_ERR, IBV_WC_GENERAL_ERR,
> ERDMA_WC_VENDOR_NO_ERR },
> +	{ ERDMA_WC_RECV_WQE_FORMAT_ERR, IBV_WC_GENERAL_ERR,
> ERDMA_WC_VENDOR_INVALID_RQE },
> +	{ ERDMA_WC_RECV_STAG_INVALID_ERR,
> +		IBV_WC_REM_ACCESS_ERR,
> ERDMA_WC_VENDOR_RQE_INVALID_STAG },
> +	{ ERDMA_WC_RECV_ADDR_VIOLATION_ERR,
> +		IBV_WC_REM_ACCESS_ERR,
> ERDMA_WC_VENDOR_RQE_ADDR_VIOLATION },
> +	{ ERDMA_WC_RECV_RIGHT_VIOLATION_ERR,
> +		IBV_WC_REM_ACCESS_ERR,
> ERDMA_WC_VENDOR_RQE_ACCESS_RIGHT_ERR },
> +	{ ERDMA_WC_RECV_PDID_ERR, IBV_WC_REM_ACCESS_ERR,
> ERDMA_WC_VENDOR_RQE_INVALID_PD },
> +	{ ERDMA_WC_RECV_WARRPING_ERR, IBV_WC_REM_ACCESS_ERR,
> ERDMA_WC_VENDOR_RQE_WRAP_ERR },
> +	{ ERDMA_WC_SEND_WQE_FORMAT_ERR,
> IBV_WC_LOC_QP_OP_ERR, ERDMA_WC_VENDOR_INVALID_SQE },
> +	{ ERDMA_WC_SEND_WQE_ORD_EXCEED, IBV_WC_GENERAL_ERR,
> ERDMA_WC_VENDOR_ZERO_ORD },
> +	{ ERDMA_WC_SEND_STAG_INVALID_ERR,
> +		IBV_WC_LOC_ACCESS_ERR,
> ERDMA_WC_VENDOR_SQE_INVALID_STAG },
> +	{ ERDMA_WC_SEND_ADDR_VIOLATION_ERR,
> +		IBV_WC_LOC_ACCESS_ERR,
> ERDMA_WC_VENDOR_SQE_ADDR_VIOLATION },
> +	{ ERDMA_WC_SEND_RIGHT_VIOLATION_ERR,
> +		IBV_WC_LOC_ACCESS_ERR,
> ERDMA_WC_VENDOR_SQE_ACCESS_ERR },
> +	{ ERDMA_WC_SEND_PDID_ERR, IBV_WC_LOC_ACCESS_ERR,
> ERDMA_WC_VENDOR_SQE_INVALID_PD },
> +	{ ERDMA_WC_SEND_WARRPING_ERR, IBV_WC_LOC_ACCESS_ERR,
> ERDMA_WC_VENDOR_SQE_WARP_ERR },
> +	{ ERDMA_WC_FLUSH_ERR, IBV_WC_WR_FLUSH_ERR,
> ERDMA_WC_VENDOR_NO_ERR },
> +	{ ERDMA_WC_RETRY_EXC_ERR, IBV_WC_RETRY_EXC_ERR,
> ERDMA_WC_VENDOR_NO_ERR
> +}, };
> +
> +#define ERDMA_POLLCQ_NO_QP      (-1)
> +#define ERDMA_POLLCQ_DUP_COMP   (-2)
> +#define ERDMA_POLLCQ_WRONG_IDX  (-3)
> +
> +static int __erdma_poll_one_cqe(struct erdma_context *ctx, struct
> erdma_cq *cq,
> +				struct erdma_cqe *cqe, uint32_t cqe_hdr,
> struct ibv_wc *wc) {
> +	struct erdma_qp *qp;
> +	uint64_t *qeidx2wrid = NULL;
> +	uint32_t qpn = be32toh(cqe->qpn);
> +	uint16_t depth = 0;
> +	uint64_t *sqe_hdr;
> +	uint16_t wqe_idx = be32toh(cqe->qe_idx);
> +	uint16_t old_ci, new_ci;
> +	uint32_t opcode = FIELD_GET(ERDMA_CQE_HDR_OPCODE_MASK,
> cqe_hdr);
> +	uint32_t syndrome =
> FIELD_GET(ERDMA_CQE_HDR_SYNDROME_MASK, cqe_hdr);
> +
> +	int tbl_idx = qpn >> ERDMA_QP_TABLE_SHIFT;
> +	int tbl_off = qpn & ERDMA_QP_TABLE_MASK;
> +
> +	if (!ctx->qp_table[tbl_idx].table || !ctx-
> >qp_table[tbl_idx].table[tbl_off])
> +		return ERDMA_POLLCQ_NO_QP;
> +
> +	qp = ctx->qp_table[tbl_idx].table[tbl_off];
> +
> +	if (FIELD_GET(ERDMA_CQE_HDR_QTYPE_MASK, cqe_hdr) ==
> ERDMA_CQE_QTYPE_SQ) {
> +		qeidx2wrid = qp->sq.wr_tbl;
> +		depth = qp->sq.depth;
> +		sqe_hdr = (uint64_t *)get_sq_wqebb(qp, wqe_idx);
> +		old_ci = qp->sq.ci;
> +		new_ci = wqe_idx +
> FIELD_GET(ERDMA_SQE_HDR_WQEBB_CNT_MASK, *sqe_hdr)
> ++ 1;
> +
> +		if ((uint16_t)(new_ci - old_ci) > depth)
> +			return ERDMA_POLLCQ_WRONG_IDX;
> +		else if (new_ci == old_ci)
> +			return ERDMA_POLLCQ_DUP_COMP;
> +
> +		qp->sq.ci = new_ci;
> +	} else {
> +		qeidx2wrid = qp->rq.wr_tbl;
> +		depth = qp->rq.depth;
> +		qp->rq.ci++;
> +	}
> +
> +	wc->wr_id = qeidx2wrid[wqe_idx & (depth - 1)];
> +	wc->byte_len = be32toh(cqe->size);
> +	wc->wc_flags = 0;
> +
> +	if (opcode == ERDMA_OP_RECV_IMM) {
> +		wc->opcode = IBV_WC_RECV_RDMA_WITH_IMM;
> +		wc->imm_data = htobe32(le32toh(cqe->imm_data));
> +		wc->wc_flags |= IBV_WC_WITH_IMM;
> +	} else if (opcode == ERDMA_OP_RSP_SEND_IMM) {
> +		wc->opcode = IBV_WC_RECV;
> +		wc->imm_data = htobe32(le32toh(cqe->imm_data));
> +		wc->wc_flags |= IBV_WC_WITH_IMM;
> +	} else {
> +		wc->opcode = map_cqe_opcode[opcode].base;
> +	}
> +
> +	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_num = qpn;
> +
> +	return 0;
> +}
> +
> +int erdma_poll_cq(struct ibv_cq *ibcq, int num_entries, struct ibv_wc
> +*wc) {
> +	struct erdma_cq *cq = to_ecq(ibcq);
> +	struct erdma_context *ctx = to_ectx(ibcq->context);
> +	int new = 0;
> +	struct erdma_cqe *cqe;
> +	int owner;
> +	uint32_t ci;
> +	uint32_t depth_mask = cq->depth - 1;
> +	uint32_t hdr;
> +	int i, ret;
> +
> +	pthread_spin_lock(&cq->lock);
> +
> +	owner = cq->owner;
> +	ci = cq->ci;
> +
> +	for (i = 0; i < num_entries; i++) {
> +		cqe = &cq->queue[ci & depth_mask];
> +		hdr = be32toh(cqe->hdr);
> +
> +		if (FIELD_GET(ERDMA_CQE_HDR_OWNER_MASK, hdr) !=
> owner)
> +			break;
> +
> +		udma_from_device_barrier();
> +
> +		ret = __erdma_poll_one_cqe(ctx, cq, cqe, hdr, wc);
> +
> +		ci++;
> +		if ((ci & depth_mask) == 0)
> +			owner = !owner;
> +
> +		if (ret)
> +			continue;
> +
> +		wc++;
> +		new++;
> +	}
> +
> +	cq->owner = owner;
> +	cq->ci = ci;
> +
> +	pthread_spin_unlock(&cq->lock);
> +
> +	return new;
> +}
> +
> +void erdma_free_context(struct ibv_context *ibv_ctx) {
> +	struct erdma_context *ctx = to_ectx(ibv_ctx);
> +	int i;
> +
> +	munmap(ctx->sdb, ERDMA_PAGE_SIZE);
> +	munmap(ctx->rdb, ERDMA_PAGE_SIZE);
> +	munmap(ctx->cdb, ERDMA_PAGE_SIZE);
> +
> +	pthread_mutex_lock(&ctx->qp_table_mutex);
> +	for (i = 0; i < ERDMA_QP_TABLE_SIZE; ++i) {
> +		if (ctx->qp_table[i].refcnt)
> +			free(ctx->qp_table[i].table);
> +	}
> +
> +	pthread_mutex_unlock(&ctx->qp_table_mutex);
> +	pthread_mutex_destroy(&ctx->qp_table_mutex);
> +
> +	verbs_uninit_context(&ctx->ibv_ctx);
> +	free(ctx);
> +}
> --
> 2.27.0

Cheng Xu Dec. 27, 2021, 7:59 a.m. UTC | #2

On 12/27/21 2:29 PM, Devesh Sharma wrote:
> 
> 

<...>

>> +	++page->used;
>> +
>> +	for (i = 0; !page->free[i]; ++i)
>> +		;/* nothing */
> Why?

page->free[i] is a 64bits bitmap, all zeros means the corespoding
entries are all used, then we need to traverse next. If stopped, then
we get a valid entry to use.

>> +
>> +	j = ffsl(page->free[i]) - 1;
>> +	page->free[i] &= ~(1UL << j);
>> +
>> +	db_records = page->page_buf + (i * bits_perlong + j) *
>> +ERDMA_DBRECORDS_SIZE;
>> +
>> +out:
>> +	pthread_mutex_unlock(&ctx->dbrecord_pages_mutex);
>> +
>> +	return db_records;
>> +}
>> +
>> +void erdma_dealloc_dbrecords(struct erdma_context *ctx, uint64_t
>> +*dbrecords) {
>> +	struct erdma_dbrecord_page *page;
>> +	int page_mask = ~(ctx->page_size - 1);
>> +	int idx;
>> +
>> +	pthread_mutex_lock(&ctx->dbrecord_pages_mutex);
>> +	for (page = ctx->dbrecord_pages; page; page = page->next)
>> +		if (((uintptr_t)dbrecords & page_mask) == (uintptr_t)page-
>>> page_buf)
>> +			break;
>> +
>> +	if (!page)
>> +		goto out;
>> +
>> +	idx = ((void *)dbrecords - page->page_buf) /
>> ERDMA_DBRECORDS_SIZE;
>> +	page->free[idx / (8 * sizeof(unsigned long))] |=
>> +		1UL << (idx % (8 * sizeof(unsigned long)));
>> +
>> +	if (!--page->used) {
>> +		if (page->prev)
>> +			page->prev->next = page->next;
>> +		else
>> +			ctx->dbrecord_pages = page->next;
>> +		if (page->next)
>> +			page->next->prev = page->prev;
>> +
>> +		free(page->page_buf);
>> +		free(page);
>> +	}
>> +
>> +out:
>> +	pthread_mutex_unlock(&ctx->dbrecord_pages_mutex);
>> +}

<...>

>> +static void __erdma_alloc_dbs(struct erdma_qp *qp, struct erdma_context
>> +*ctx) {
>> +	uint32_t qpn = qp->id;
>> +
>> +	if (ctx->sdb_type == ERDMA_SDB_PAGE) {
>> +		/* qpn[4:0] as the index in this db page. */
>> +		qp->sq.db = ctx->sdb + (qpn & 31) * ERDMA_SQDB_SIZE;
>> +	} else if (ctx->sdb_type == ERDMA_SDB_ENTRY) {
>> +		/* for type 'ERDMA_SDB_ENTRY', each uctx has 2 dwqe,
>> totally takes 256Bytes. */
>> +		qp->sq.db = ctx->sdb + ctx->sdb_offset * 256;
> Generally we use macros to define hard-coded integers. E.g 256 should use a macro.

OK, will fix.

>> +	} else {
>> +		/* qpn[4:0] as the index in this db page. */
>> +		qp->sq.db = ctx->sdb + (qpn & 31) * ERDMA_SQDB_SIZE;
>> +	}
>> +
>> +	/* qpn[6:0] as the index in this rq db page. */
>> +	qp->rq.db = ctx->rdb + (qpn & 127) * ERDMA_RQDB_SPACE_SIZE; }
>> +
>> +struct ibv_qp *erdma_create_qp(struct ibv_pd *pd, struct
>> +ibv_qp_init_attr *attr) {
>> +	struct erdma_cmd_create_qp cmd = {};
>> +	struct erdma_cmd_create_qp_resp resp = {};
>> +	struct erdma_qp *qp;
>> +	struct ibv_context *base_ctx = pd->context;
>> +	struct erdma_context *ctx = to_ectx(base_ctx);
>> +	uint64_t *db_records  = NULL;
>> +	int rv, tbl_idx, tbl_off;
>> +	int sq_size = 0, rq_size = 0, total_bufsize = 0;
>> +
>> +	memset(&cmd, 0, sizeof(cmd));
>> +	memset(&resp, 0, sizeof(resp));
> No need of memset due to declaration step.

OK, will fix.

>> +
>> +	qp = calloc(1, sizeof(*qp));
>> +	if (!qp)
>> +		return NULL;
>> +
>> +	sq_size = roundup_pow_of_two(attr->cap.max_send_wr *
>> MAX_WQEBB_PER_SQE) << SQEBB_SHIFT;
>> +	sq_size = align(sq_size, ctx->page_size);
>> +	rq_size = align(roundup_pow_of_two(attr->cap.max_recv_wr) <<
>> RQE_SHIFT, ctx->page_size);
>> +	total_bufsize = sq_size + rq_size;
>> +	rv = posix_memalign(&qp->qbuf, ctx->page_size, total_bufsize);
>> +	if (rv || !qp->qbuf) {
>> +		errno = ENOMEM;
>> +		goto error_alloc;
>> +	}

<...>

>> +
>> +int erdma_destroy_qp(struct ibv_qp *base_qp) {
>> +	struct erdma_qp *qp = to_eqp(base_qp);
>> +	struct ibv_context *base_ctx = base_qp->pd->context;
>> +	struct erdma_context *ctx = to_ectx(base_ctx);
>> +	int rv, tbl_idx, tbl_off;
>> +
>> +	pthread_spin_lock(&qp->sq_lock);
>> +	pthread_spin_lock(&qp->rq_lock);
> Why to hold these?

Here, we are destroying the qp resources, such as the queue buffers.
we want to avoid race condition with post_send and post_recv.

>> +
>> +	pthread_mutex_lock(&ctx->qp_table_mutex);
>> +	tbl_idx = qp->id >> ERDMA_QP_TABLE_SHIFT;
>> +	tbl_off = qp->id & ERDMA_QP_TABLE_MASK;
>> +
>> +	ctx->qp_table[tbl_idx].table[tbl_off] = NULL;
>> +	ctx->qp_table[tbl_idx].refcnt--;
>> +
>> +	if (ctx->qp_table[tbl_idx].refcnt == 0) {
>> +		free(ctx->qp_table[tbl_idx].table);
>> +		ctx->qp_table[tbl_idx].table = NULL;
>> +	}
>> +
>> +	pthread_mutex_unlock(&ctx->qp_table_mutex);
>> +
>> +	rv = ibv_cmd_destroy_qp(base_qp);
>> +	if (rv) {
>> +		pthread_spin_unlock(&qp->rq_lock);
>> +		pthread_spin_unlock(&qp->sq_lock);
>> +		return rv;
>> +	}
>> +	pthread_spin_destroy(&qp->rq_lock);
>> +	pthread_spin_destroy(&qp->sq_lock);
>> +
>> +	if (qp->db_records)
>> +		erdma_dealloc_dbrecords(ctx, qp->db_records);
>> +
>> +	if (qp->qbuf)
>> +		free(qp->qbuf);
>> +
>> +	free(qp);
>> +
>> +	return 0;
>> +}
>> +

<...>

>> +
>> +int erdma_post_send(struct ibv_qp *base_qp, struct ibv_send_wr *wr,
>> +struct ibv_send_wr **bad_wr) {
>> +	struct erdma_qp *qp = to_eqp(base_qp);
>> +	uint16_t sq_pi;
>> +	int new_sqe = 0, rv = 0;
>> +
>> +	*bad_wr = NULL;
>> +
>> +	if (base_qp->state == IBV_QPS_ERR) {
> Post_send is allowed in Error state. Thus the check is redundant.

Does this have specification? We didn't find the description in IB
specification.
ERDMA uses iWarp, and in this specification (link: [1]), it says that
two actions should be taken: "post wqe, and then flush it" or "return
an immediate error" when post WR in ERROR state. After modify qp to err,
our hardware will flush all the wqes, thus for the newly posted wr, we
return an error immediately.
Also, I review other providers' implementation, ocrdma/efa/bnxt_re also
don't allow post_send in ERROR state. Does this can have a little
difference depended on different HCAs?

Thanks,
Cheng Xu

[1] 
https://datatracker.ietf.org/doc/html/draft-hilland-rddp-verbs-00#section-6.2.4



>> +		*bad_wr = wr;
>> +		return -EIO;
>> +	}
>> +
>> +	pthread_spin_lock(&qp->sq_lock);
>> +
>> +	sq_pi = qp->sq.pi;
>> +
>> +	while (wr) {
>> +		if ((uint16_t)(sq_pi - qp->sq.ci) >= qp->sq.depth) {
>> +			rv = -ENOMEM;
>> +			*bad_wr = wr;
>> +			break;
>> +		}
>> +

<...>

>> --
>> 2.27.0

Devesh Sharma Dec. 28, 2021, 12:17 p.m. UTC | #3

> -----Original Message-----
> From: Cheng Xu <[e-mail]>
> Sent: Monday, December 27, 2021 1:29 PM
> To: Devesh Sharma <[e-mail]>; [e-mail]
> Cc: [e-mail]; [e-mail]; [e-mail];
> [e-mail]
> Subject: [External] : Re: [PATCH rdma-core 2/5] RDMA-CORE/erdma: Add
> userspace verbs implementation
> 
> 
> 
> On 12/27/21 2:29 PM, Devesh Sharma wrote:
> >
> >
> 
> <...>
> 
> >> +	++page->used;
> >> +
> >> +	for (i = 0; !page->free[i]; ++i)
> >> +		;/* nothing */
> > Why?
> 
> page->free[i] is a 64bits bitmap, all zeros means the corespoding
> entries are all used, then we need to traverse next. If stopped, then we get a
> valid entry to use.
The loop could be a candidate for compiler optimization?
> 
> >> +
> >> +	j = ffsl(page->free[i]) - 1;
> >> +	page->free[i] &= ~(1UL << j);
> >> +
> >> +	db_records = page->page_buf + (i * bits_perlong + j) *
> >> +ERDMA_DBRECORDS_SIZE;
> >> +
> >> +out:
> >> +	pthread_mutex_unlock(&ctx->dbrecord_pages_mutex);
> >> +
> >> +	return db_records;
> >> +}
> >> +
> >> +void erdma_dealloc_dbrecords(struct erdma_context *ctx, uint64_t
> >> +*dbrecords) {
> >> +	struct erdma_dbrecord_page *page;
> >> +	int page_mask = ~(ctx->page_size - 1);
> >> +	int idx;
> >> +
> >> +	pthread_mutex_lock(&ctx->dbrecord_pages_mutex);
> >> +	for (page = ctx->dbrecord_pages; page; page = page->next)
> >> +		if (((uintptr_t)dbrecords & page_mask) == (uintptr_t)page-
> >>> page_buf)
> >> +			break;
> >> +
> >> +	if (!page)
> >> +		goto out;
> >> +
> >> +	idx = ((void *)dbrecords - page->page_buf) /
> >> ERDMA_DBRECORDS_SIZE;
> >> +	page->free[idx / (8 * sizeof(unsigned long))] |=
> >> +		1UL << (idx % (8 * sizeof(unsigned long)));
> >> +
> >> +	if (!--page->used) {
> >> +		if (page->prev)
> >> +			page->prev->next = page->next;
> >> +		else
> >> +			ctx->dbrecord_pages = page->next;
> >> +		if (page->next)
> >> +			page->next->prev = page->prev;
> >> +
> >> +		free(page->page_buf);
> >> +		free(page);
> >> +	}
> >> +
> >> +out:
> >> +	pthread_mutex_unlock(&ctx->dbrecord_pages_mutex);
> >> +}
> 
> <...>
> 
> >> +static void __erdma_alloc_dbs(struct erdma_qp *qp, struct
> >> +erdma_context
> >> +*ctx) {
> >> +	uint32_t qpn = qp->id;
> >> +
> >> +	if (ctx->sdb_type == ERDMA_SDB_PAGE) {
> >> +		/* qpn[4:0] as the index in this db page. */
> >> +		qp->sq.db = ctx->sdb + (qpn & 31) * ERDMA_SQDB_SIZE;
> >> +	} else if (ctx->sdb_type == ERDMA_SDB_ENTRY) {
> >> +		/* for type 'ERDMA_SDB_ENTRY', each uctx has 2 dwqe,
> >> totally takes 256Bytes. */
> >> +		qp->sq.db = ctx->sdb + ctx->sdb_offset * 256;
> > Generally we use macros to define hard-coded integers. E.g 256 should use
> a macro.
> 
> OK, will fix.
> 
> >> +	} else {
> >> +		/* qpn[4:0] as the index in this db page. */
> >> +		qp->sq.db = ctx->sdb + (qpn & 31) * ERDMA_SQDB_SIZE;
> >> +	}
> >> +
> >> +	/* qpn[6:0] as the index in this rq db page. */
> >> +	qp->rq.db = ctx->rdb + (qpn & 127) * ERDMA_RQDB_SPACE_SIZE; }
> >> +
> >> +struct ibv_qp *erdma_create_qp(struct ibv_pd *pd, struct
> >> +ibv_qp_init_attr *attr) {
> >> +	struct erdma_cmd_create_qp cmd = {};
> >> +	struct erdma_cmd_create_qp_resp resp = {};
> >> +	struct erdma_qp *qp;
> >> +	struct ibv_context *base_ctx = pd->context;
> >> +	struct erdma_context *ctx = to_ectx(base_ctx);
> >> +	uint64_t *db_records  = NULL;
> >> +	int rv, tbl_idx, tbl_off;
> >> +	int sq_size = 0, rq_size = 0, total_bufsize = 0;
> >> +
> >> +	memset(&cmd, 0, sizeof(cmd));
> >> +	memset(&resp, 0, sizeof(resp));
> > No need of memset due to declaration step.
> 
> OK, will fix.
> 
> >> +
> >> +	qp = calloc(1, sizeof(*qp));
> >> +	if (!qp)
> >> +		return NULL;
> >> +
> >> +	sq_size = roundup_pow_of_two(attr->cap.max_send_wr *
> >> MAX_WQEBB_PER_SQE) << SQEBB_SHIFT;
> >> +	sq_size = align(sq_size, ctx->page_size);
> >> +	rq_size = align(roundup_pow_of_two(attr->cap.max_recv_wr) <<
> >> RQE_SHIFT, ctx->page_size);
> >> +	total_bufsize = sq_size + rq_size;
> >> +	rv = posix_memalign(&qp->qbuf, ctx->page_size, total_bufsize);
> >> +	if (rv || !qp->qbuf) {
> >> +		errno = ENOMEM;
> >> +		goto error_alloc;
> >> +	}
> 
> <...>
> 
> >> +
> >> +int erdma_destroy_qp(struct ibv_qp *base_qp) {
> >> +	struct erdma_qp *qp = to_eqp(base_qp);
> >> +	struct ibv_context *base_ctx = base_qp->pd->context;
> >> +	struct erdma_context *ctx = to_ectx(base_ctx);
> >> +	int rv, tbl_idx, tbl_off;
> >> +
> >> +	pthread_spin_lock(&qp->sq_lock);
> >> +	pthread_spin_lock(&qp->rq_lock);
> > Why to hold these?
> 
> Here, we are destroying the qp resources, such as the queue buffers.
> we want to avoid race condition with post_send and post_recv.
Application should make sure no one is accessing the QP when qp-destroy is called. Probably you can be easy on these locks.
> 
> >> +
> >> +	pthread_mutex_lock(&ctx->qp_table_mutex);
> >> +	tbl_idx = qp->id >> ERDMA_QP_TABLE_SHIFT;
> >> +	tbl_off = qp->id & ERDMA_QP_TABLE_MASK;
> >> +
> >> +	ctx->qp_table[tbl_idx].table[tbl_off] = NULL;
> >> +	ctx->qp_table[tbl_idx].refcnt--;
> >> +
> >> +	if (ctx->qp_table[tbl_idx].refcnt == 0) {
> >> +		free(ctx->qp_table[tbl_idx].table);
> >> +		ctx->qp_table[tbl_idx].table = NULL;
> >> +	}
> >> +
> >> +	pthread_mutex_unlock(&ctx->qp_table_mutex);
> >> +
> >> +	rv = ibv_cmd_destroy_qp(base_qp);
> >> +	if (rv) {
> >> +		pthread_spin_unlock(&qp->rq_lock);
> >> +		pthread_spin_unlock(&qp->sq_lock);
> >> +		return rv;
> >> +	}
> >> +	pthread_spin_destroy(&qp->rq_lock);
> >> +	pthread_spin_destroy(&qp->sq_lock);
> >> +
> >> +	if (qp->db_records)
> >> +		erdma_dealloc_dbrecords(ctx, qp->db_records);
> >> +
> >> +	if (qp->qbuf)
> >> +		free(qp->qbuf);
> >> +
> >> +	free(qp);
> >> +
> >> +	return 0;
> >> +}
> >> +
> 
> <...>
> 
> >> +
> >> +int erdma_post_send(struct ibv_qp *base_qp, struct ibv_send_wr *wr,
> >> +struct ibv_send_wr **bad_wr) {
> >> +	struct erdma_qp *qp = to_eqp(base_qp);
> >> +	uint16_t sq_pi;
> >> +	int new_sqe = 0, rv = 0;
> >> +
> >> +	*bad_wr = NULL;
> >> +
> >> +	if (base_qp->state == IBV_QPS_ERR) {
> > Post_send is allowed in Error state. Thus the check is redundant.
> 
> Does this have specification? We didn't find the description in IB
> specification.
> ERDMA uses iWarp, and in this specification (link: [1]), it says that two actions
> should be taken: "post wqe, and then flush it" or "return an immediate
> error" when post WR in ERROR state. After modify qp to err, our hardware
> will flush all the wqes, thus for the newly posted wr, we return an error
> immediately.
> Also, I review other providers' implementation, ocrdma/efa/bnxt_re also
> don't allow post_send in ERROR state. Does this can have a little difference
> depended on different HCAs?
Well its indeed. You can omit it for now.
> 
> Thanks,
> Cheng Xu
> 
> [1]
> https://urldefense.com/v3/__https://datatracker.ietf.org/doc/html/draft-
> hilland-rddp-verbs-00*section-
> 6.2.4__;Iw!!ACWV5N9M2RV99hQ!fcFF2OWy2n6miTv9UwPpXs1y1RbO4pxYs
> XOCc59hTAL0Z4nagUt0Z2Aaht8jX0alE_TU$
> 
> 
> 
> >> +		*bad_wr = wr;
> >> +		return -EIO;
> >> +	}
> >> +
> >> +	pthread_spin_lock(&qp->sq_lock);
> >> +
> >> +	sq_pi = qp->sq.pi;
> >> +
> >> +	while (wr) {
> >> +		if ((uint16_t)(sq_pi - qp->sq.ci) >= qp->sq.depth) {
> >> +			rv = -ENOMEM;
> >> +			*bad_wr = wr;
> >> +			break;
> >> +		}
> >> +
> 
> <...>
> 
> >> --
> >> 2.27.0

Patch

diff --git a/providers/erdma/erdma_db.c b/providers/erdma/erdma_db.c
new file mode 100644
index 00000000..83db76d1
--- /dev/null
+++ b/providers/erdma/erdma_db.c
@@ -0,0 +1,110 @@ 
+// SPDX-License-Identifier: GPL-2.0 or OpenIB.org BSD (MIT) See COPYING file
+
+// Authors: Cheng Xu <[e-mail]>
+// Copyright (c) 2020-2021, Alibaba Group.
+
+// Copyright (c) 2012 Mellanox Technologies, Inc.  All rights reserved.
+
+#include <inttypes.h>
+#include <stdlib.h>
+#include <util/util.h>
+
+#include "erdma.h"
+#include "erdma_db.h"
+
+#define ERDMA_DBRECORDS_SIZE 16
+
+struct erdma_dbrecord_page {
+	struct erdma_dbrecord_page *prev, *next;
+	void *page_buf;
+	int cnt;
+	int used;
+	unsigned long free[0];
+};
+
+uint64_t *erdma_alloc_dbrecords(struct erdma_context *ctx)
+{
+	struct erdma_dbrecord_page *page = NULL;
+	uint64_t *db_records = NULL;
+	int dbrecords_per_page, nlongs = 0, bits_perlong = (8 * sizeof(unsigned long));
+	int i, j, rv;
+
+	pthread_mutex_lock(&ctx->dbrecord_pages_mutex);
+
+	for (page = ctx->dbrecord_pages; page; page = page->next)
+		if (page->used < page->cnt)
+			goto found;
+
+	dbrecords_per_page = ctx->page_size / ERDMA_DBRECORDS_SIZE;
+	nlongs = align(dbrecords_per_page, bits_perlong) / bits_perlong;
+	page = malloc(sizeof(*page) + nlongs * sizeof(unsigned long));
+	if (!page)
+		goto out;
+
+	rv = posix_memalign(&page->page_buf, ctx->page_size, ctx->page_size);
+	if (rv) {
+		free(page);
+		goto out;
+	}
+
+	page->cnt = dbrecords_per_page;
+	page->used = 0;
+	for (i = 0; i < nlongs; i++)
+		page->free[i] = ~0UL;
+
+	page->prev = NULL;
+	page->next = ctx->dbrecord_pages;
+	ctx->dbrecord_pages = page;
+	if (page->next)
+		page->next->prev = page;
+
+found:
+	++page->used;
+
+	for (i = 0; !page->free[i]; ++i)
+		;/* nothing */
+
+	j = ffsl(page->free[i]) - 1;
+	page->free[i] &= ~(1UL << j);
+
+	db_records = page->page_buf + (i * bits_perlong + j) * ERDMA_DBRECORDS_SIZE;
+
+out:
+	pthread_mutex_unlock(&ctx->dbrecord_pages_mutex);
+
+	return db_records;
+}
+
+void erdma_dealloc_dbrecords(struct erdma_context *ctx, uint64_t *dbrecords)
+{
+	struct erdma_dbrecord_page *page;
+	int page_mask = ~(ctx->page_size - 1);
+	int idx;
+
+	pthread_mutex_lock(&ctx->dbrecord_pages_mutex);
+	for (page = ctx->dbrecord_pages; page; page = page->next)
+		if (((uintptr_t)dbrecords & page_mask) == (uintptr_t)page->page_buf)
+			break;
+
+	if (!page)
+		goto out;
+
+	idx = ((void *)dbrecords - page->page_buf) / ERDMA_DBRECORDS_SIZE;
+	page->free[idx / (8 * sizeof(unsigned long))] |=
+		1UL << (idx % (8 * sizeof(unsigned long)));
+
+	if (!--page->used) {
+		if (page->prev)
+			page->prev->next = page->next;
+		else
+			ctx->dbrecord_pages = page->next;
+		if (page->next)
+			page->next->prev = page->prev;
+
+		free(page->page_buf);
+		free(page);
+	}
+
+out:
+	pthread_mutex_unlock(&ctx->dbrecord_pages_mutex);
+}
diff --git a/providers/erdma/erdma_verbs.c b/providers/erdma/erdma_verbs.c
new file mode 100644
index 00000000..3c1c9769
--- /dev/null
+++ b/providers/erdma/erdma_verbs.c
@@ -0,0 +1,934 @@ 
+// SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
+
+// Authors: Cheng Xu <[e-mail]>
+// Copyright (c) 2020-2021, Alibaba Group.
+// Authors: Bernard Metzler <[e-mail]>
+// Copyright (c) 2008-2019, IBM Corporation
+
+#include <ccan/minmax.h>
+#include <endian.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <sys/mman.h>
+#include <sys/types.h>
+#include <unistd.h>
+#include <util/mmio.h>
+#include <util/udma_barrier.h>
+#include <util/util.h>
+
+#include "erdma.h"
+#include "erdma_abi.h"
+#include "erdma_db.h"
+#include "erdma_hw.h"
+#include "erdma_verbs.h"
+
+int erdma_query_device(struct ibv_context *ctx,
+		       const struct ibv_query_device_ex_input *input,
+		       struct ibv_device_attr_ex *attr, size_t attr_size)
+{
+	struct ib_uverbs_ex_query_device_resp resp;
+	size_t resp_size = sizeof(resp);
+	uint64_t raw_fw_ver;
+	unsigned int major, minor, sub_minor;
+	int rv;
+
+	rv = ibv_cmd_query_device_any(ctx, input, attr, attr_size, &resp, &resp_size);
+	if (rv)
+		return rv;
+
+	raw_fw_ver = resp.base.fw_ver;
+	major = (raw_fw_ver >> 32) & 0xffff;
+	minor = (raw_fw_ver >> 16) & 0xffff;
+	sub_minor = raw_fw_ver & 0xffff;
+
+	snprintf(attr->orig_attr.fw_ver, sizeof(attr->orig_attr.fw_ver),
+		"%d.%d.%d", major, minor, sub_minor);
+
+	return 0;
+}
+
+int erdma_query_port(struct ibv_context *ctx, uint8_t port, struct ibv_port_attr *attr)
+{
+	struct ibv_query_port cmd;
+
+	memset(&cmd, 0, sizeof(cmd));
+
+	return ibv_cmd_query_port(ctx, port, attr, &cmd, sizeof(cmd));
+}
+
+int erdma_query_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr,
+		   int attr_mask, struct ibv_qp_init_attr *init_attr)
+{
+	struct ibv_query_qp cmd;
+
+	memset(&cmd, 0, sizeof(cmd));
+
+	return ibv_cmd_query_qp(qp, attr, attr_mask, init_attr, &cmd, sizeof(cmd));
+}
+
+struct ibv_pd *erdma_alloc_pd(struct ibv_context *ctx)
+{
+	struct ibv_alloc_pd cmd;
+	struct ib_uverbs_alloc_pd_resp resp;
+	struct ibv_pd *pd;
+
+	memset(&cmd, 0, sizeof(cmd));
+
+	pd = calloc(1, sizeof(*pd));
+	if (!pd)
+		return NULL;
+
+	if (ibv_cmd_alloc_pd(ctx, pd, &cmd, sizeof(cmd), &resp, sizeof(resp))) {
+		free(pd);
+		return NULL;
+	}
+
+	return pd;
+}
+
+int erdma_free_pd(struct ibv_pd *pd)
+{
+	int rv;
+
+	rv = ibv_cmd_dealloc_pd(pd);
+	if (rv)
+		return rv;
+
+	free(pd);
+	return 0;
+}
+
+struct ibv_mr *erdma_reg_mr(struct ibv_pd *pd, void *addr, size_t len,
+			    uint64_t hca_va, int access)
+{
+	struct ibv_reg_mr cmd;
+	struct ib_uverbs_reg_mr_resp resp;
+	struct verbs_mr *vmr;
+	int ret;
+
+	vmr = calloc(1, sizeof(*vmr));
+	if (!vmr)
+		return NULL;
+
+	ret = ibv_cmd_reg_mr(pd, addr, len, hca_va, access, vmr, &cmd, sizeof(cmd),
+		&resp, sizeof(resp));
+	if (ret) {
+		free(vmr);
+		return NULL;
+	}
+
+	return &vmr->ibv_mr;
+}
+
+int erdma_dereg_mr(struct verbs_mr *vmr)
+{
+	int ret;
+
+	ret = ibv_cmd_dereg_mr(vmr);
+	if (ret)
+		return ret;
+
+	free(vmr);
+	return 0;
+}
+
+int erdma_notify_cq(struct ibv_cq *ibcq, int solicited)
+{
+	struct erdma_cq *cq = to_ecq(ibcq);
+	uint64_t db_data;
+	int ret;
+
+	ret = pthread_spin_lock(&cq->lock);
+	if (ret)
+		return ret;
+
+	db_data = FIELD_PREP(ERDMA_CQDB_EQN_MASK, cq->comp_vector + 1) |
+		FIELD_PREP(ERDMA_CQDB_CQN_MASK, cq->id) |
+		FIELD_PREP(ERDMA_CQDB_ARM_MASK, 1) |
+		FIELD_PREP(ERDMA_CQDB_SOL_MASK, solicited) |
+		FIELD_PREP(ERDMA_CQDB_CMDSN_MASK, cq->cmdsn) |
+		FIELD_PREP(ERDMA_CQDB_CI_MASK, cq->ci);
+
+	*(__le64 *)cq->db_record = htole64(db_data);
+	udma_to_device_barrier();
+	mmio_write64_le(cq->db, htole64(db_data));
+
+	pthread_spin_unlock(&cq->lock);
+
+	return ret;
+}
+
+struct ibv_cq *erdma_create_cq(struct ibv_context *ctx, int num_cqe,
+			       struct ibv_comp_channel *channel, int comp_vector)
+{
+	struct erdma_context *ectx;
+	struct erdma_cmd_create_cq cmd = {};
+	struct erdma_cmd_create_cq_resp resp = {};
+	struct erdma_cq *cq;
+	uint64_t *db_records = NULL;
+	int cq_size;
+	int rv;
+
+	ectx = to_ectx(ctx);
+
+	cq = calloc(1, sizeof(*cq));
+	if (!cq)
+		return NULL;
+
+	if (num_cqe < 64)
+		num_cqe = 64;
+
+	num_cqe = roundup_pow_of_two(num_cqe);
+	cq_size = num_cqe * sizeof(struct erdma_cqe);
+	cq_size = ERDMA_SIZE_TO_NPAGE(cq_size) << ERDMA_PAGE_SHIFT;
+
+	rv = posix_memalign((void **)&cq->queue, ERDMA_PAGE_SIZE, cq_size);
+	if (rv) {
+		errno = rv;
+		free(cq);
+		return NULL;
+	}
+
+	db_records = erdma_alloc_dbrecords(ectx);
+	if (!db_records) {
+		errno = ENOMEM;
+		goto error_alloc;
+	}
+
+	cmd.db_record_va = (__u64)db_records;
+	cmd.qbuf_va = (uint64_t)cq->queue;
+	cmd.qbuf_len = cq_size;
+
+	rv = ibv_cmd_create_cq(ctx, num_cqe, channel, comp_vector, &cq->base_cq,
+		&cmd.ibv_cmd, sizeof(cmd), &resp.ibv_resp, sizeof(resp));
+	if (rv) {
+		free(cq);
+		errno = EIO;
+		goto error_alloc;
+	}
+
+	pthread_spin_init(&cq->lock, PTHREAD_PROCESS_PRIVATE);
+
+	*db_records = 0;
+	cq->db_record = db_records;
+
+	cq->id = resp.cq_id;
+	cq->depth = resp.num_cqe;
+	cq->owner = 1;
+
+	cq->db = ectx->cdb;
+	cq->db_offset = (cq->id & (ERDMA_PAGE_SIZE / ERDMA_CQDB_SIZE - 1)) * ERDMA_CQDB_SIZE;
+	cq->db += cq->db_offset;
+
+	cq->comp_vector = comp_vector;
+
+	return &cq->base_cq;
+
+error_alloc:
+	if (db_records)
+		erdma_dealloc_dbrecords(ectx, db_records);
+
+	if (cq->queue)
+		free(cq->queue);
+
+	free(cq);
+
+	return NULL;
+}
+
+int erdma_destroy_cq(struct ibv_cq *base_cq)
+{
+	struct erdma_cq *cq = to_ecq(base_cq);
+	struct erdma_context *ctx = to_ectx(base_cq->context);
+	int rv;
+
+	pthread_spin_lock(&cq->lock);
+	rv = ibv_cmd_destroy_cq(base_cq);
+	if (rv) {
+		pthread_spin_unlock(&cq->lock);
+		errno = EIO;
+		return rv;
+	}
+	pthread_spin_destroy(&cq->lock);
+
+	if (cq->db_record)
+		erdma_dealloc_dbrecords(ctx, cq->db_record);
+
+	if (cq->queue)
+		free(cq->queue);
+
+	free(cq);
+
+	return 0;
+}
+
+static void __erdma_alloc_dbs(struct erdma_qp *qp, struct erdma_context *ctx)
+{
+	uint32_t qpn = qp->id;
+
+	if (ctx->sdb_type == ERDMA_SDB_PAGE) {
+		/* qpn[4:0] as the index in this db page. */
+		qp->sq.db = ctx->sdb + (qpn & 31) * ERDMA_SQDB_SIZE;
+	} else if (ctx->sdb_type == ERDMA_SDB_ENTRY) {
+		/* for type 'ERDMA_SDB_ENTRY', each uctx has 2 dwqe, totally takes 256Bytes. */
+		qp->sq.db = ctx->sdb + ctx->sdb_offset * 256;
+	} else {
+		/* qpn[4:0] as the index in this db page. */
+		qp->sq.db = ctx->sdb + (qpn & 31) * ERDMA_SQDB_SIZE;
+	}
+
+	/* qpn[6:0] as the index in this rq db page. */
+	qp->rq.db = ctx->rdb + (qpn & 127) * ERDMA_RQDB_SPACE_SIZE;
+}
+
+struct ibv_qp *erdma_create_qp(struct ibv_pd *pd, struct ibv_qp_init_attr *attr)
+{
+	struct erdma_cmd_create_qp cmd = {};
+	struct erdma_cmd_create_qp_resp resp = {};
+	struct erdma_qp *qp;
+	struct ibv_context *base_ctx = pd->context;
+	struct erdma_context *ctx = to_ectx(base_ctx);
+	uint64_t *db_records  = NULL;
+	int rv, tbl_idx, tbl_off;
+	int sq_size = 0, rq_size = 0, total_bufsize = 0;
+
+	memset(&cmd, 0, sizeof(cmd));
+	memset(&resp, 0, sizeof(resp));
+
+	qp = calloc(1, sizeof(*qp));
+	if (!qp)
+		return NULL;
+
+	sq_size = roundup_pow_of_two(attr->cap.max_send_wr * MAX_WQEBB_PER_SQE) << SQEBB_SHIFT;
+	sq_size = align(sq_size, ctx->page_size);
+	rq_size = align(roundup_pow_of_two(attr->cap.max_recv_wr) << RQE_SHIFT, ctx->page_size);
+	total_bufsize = sq_size + rq_size;
+	rv = posix_memalign(&qp->qbuf, ctx->page_size, total_bufsize);
+	if (rv || !qp->qbuf) {
+		errno = ENOMEM;
+		goto error_alloc;
+	}
+
+	db_records = erdma_alloc_dbrecords(ctx);
+	if (!db_records) {
+		errno = ENOMEM;
+		goto error_alloc;
+	}
+
+	cmd.db_record_va = (__u64)db_records;
+	cmd.qbuf_va = (__u64)qp->qbuf;
+	cmd.qbuf_len = (__u32)total_bufsize;
+
+	rv = ibv_cmd_create_qp(pd, &qp->base_qp, attr, &cmd.ibv_cmd,
+		sizeof(cmd), &resp.ibv_resp, sizeof(resp));
+	if (rv) {
+		errno = EIO;
+		goto error_alloc;
+	}
+
+	qp->id = resp.qp_id;
+
+	pthread_mutex_lock(&ctx->qp_table_mutex);
+	tbl_idx = qp->id >> ERDMA_QP_TABLE_SHIFT;
+	tbl_off = qp->id & ERDMA_QP_TABLE_MASK;
+
+	if (ctx->qp_table[tbl_idx].refcnt == 0) {
+		ctx->qp_table[tbl_idx].table = calloc(ERDMA_PAGE_SIZE, sizeof(struct erdma_qp *));
+		if (!ctx->qp_table[tbl_idx].table) {
+			errno = ENOMEM;
+			goto fail;
+		}
+	}
+
+	/* exist qp */
+	if (ctx->qp_table[tbl_idx].table[tbl_off]) {
+		errno = EBUSY;
+		goto fail;
+	}
+
+	ctx->qp_table[tbl_idx].table[tbl_off] = qp;
+	ctx->qp_table[tbl_idx].refcnt++;
+	pthread_mutex_unlock(&ctx->qp_table_mutex);
+
+	qp->sq.qbuf = qp->qbuf;
+	qp->rq.qbuf = qp->qbuf + resp.rq_offset;
+	qp->sq.depth = resp.num_sqe;
+	qp->rq.depth = resp.num_rqe;
+	qp->sq_sig_all = attr->sq_sig_all;
+	qp->sq.size = resp.num_sqe * SQEBB_SIZE;
+	qp->rq.size = resp.num_rqe * sizeof(struct erdma_rqe);
+
+	/* doorbell allocation. */
+	__erdma_alloc_dbs(qp, ctx);
+
+	pthread_spin_init(&qp->sq_lock, PTHREAD_PROCESS_PRIVATE);
+	pthread_spin_init(&qp->rq_lock, PTHREAD_PROCESS_PRIVATE);
+
+	*db_records = 0;
+	*(db_records + 1) = 0;
+	qp->db_records = db_records;
+	qp->sq.db_record = db_records;
+	qp->rq.db_record = db_records + 1;
+
+	qp->rq.wr_tbl = calloc(qp->rq.depth, sizeof(uint64_t));
+	if (!qp->rq.wr_tbl)
+		goto fail;
+
+	qp->sq.wr_tbl = calloc(qp->sq.depth, sizeof(uint64_t));
+	if (!qp->sq.wr_tbl)
+		goto fail;
+
+
+	return &qp->base_qp;
+fail:
+	if (qp->sq.wr_tbl)
+		free(qp->sq.wr_tbl);
+
+	if (qp->rq.wr_tbl)
+		free(qp->rq.wr_tbl);
+
+	ibv_cmd_destroy_qp(&qp->base_qp);
+
+error_alloc:
+	if (db_records)
+		erdma_dealloc_dbrecords(ctx, db_records);
+
+	if (qp->qbuf)
+		free(qp->qbuf);
+
+	free(qp);
+
+	return NULL;
+}
+
+int erdma_modify_qp(struct ibv_qp *base_qp, struct ibv_qp_attr *attr, int attr_mask)
+{
+	struct ibv_modify_qp cmd;
+	struct erdma_qp *qp = to_eqp(base_qp);
+	int rv;
+
+	memset(&cmd, 0, sizeof(cmd));
+
+	pthread_spin_lock(&qp->sq_lock);
+	pthread_spin_lock(&qp->rq_lock);
+
+	rv = ibv_cmd_modify_qp(base_qp, attr, attr_mask, &cmd, sizeof(cmd));
+
+	pthread_spin_unlock(&qp->rq_lock);
+	pthread_spin_unlock(&qp->sq_lock);
+
+	return rv;
+}
+
+int erdma_destroy_qp(struct ibv_qp *base_qp)
+{
+	struct erdma_qp *qp = to_eqp(base_qp);
+	struct ibv_context *base_ctx = base_qp->pd->context;
+	struct erdma_context *ctx = to_ectx(base_ctx);
+	int rv, tbl_idx, tbl_off;
+
+	pthread_spin_lock(&qp->sq_lock);
+	pthread_spin_lock(&qp->rq_lock);
+
+	pthread_mutex_lock(&ctx->qp_table_mutex);
+	tbl_idx = qp->id >> ERDMA_QP_TABLE_SHIFT;
+	tbl_off = qp->id & ERDMA_QP_TABLE_MASK;
+
+	ctx->qp_table[tbl_idx].table[tbl_off] = NULL;
+	ctx->qp_table[tbl_idx].refcnt--;
+
+	if (ctx->qp_table[tbl_idx].refcnt == 0) {
+		free(ctx->qp_table[tbl_idx].table);
+		ctx->qp_table[tbl_idx].table = NULL;
+	}
+
+	pthread_mutex_unlock(&ctx->qp_table_mutex);
+
+	rv = ibv_cmd_destroy_qp(base_qp);
+	if (rv) {
+		pthread_spin_unlock(&qp->rq_lock);
+		pthread_spin_unlock(&qp->sq_lock);
+		return rv;
+	}
+	pthread_spin_destroy(&qp->rq_lock);
+	pthread_spin_destroy(&qp->sq_lock);
+
+	if (qp->db_records)
+		erdma_dealloc_dbrecords(ctx, qp->db_records);
+
+	if (qp->qbuf)
+		free(qp->qbuf);
+
+	free(qp);
+
+	return 0;
+}
+
+static int erdma_push_one_sqe(struct erdma_qp *qp, struct ibv_send_wr *wr, uint16_t *sq_pi)
+{
+	uint16_t tmp_pi = *sq_pi;
+	void *sqe;
+	uint64_t sqe_hdr;
+	struct erdma_write_sqe *write_sqe;
+	struct erdma_send_sqe *send_sqe;
+	struct erdma_readreq_sqe *read_sqe;
+	uint32_t wqe_size = 0;
+	__le32 *length_field = NULL;
+	struct erdma_sge *sgl_base = NULL;
+	uint32_t i, bytes = 0;
+	uint32_t sgl_off, sgl_idx, wqebb_cnt, opcode;
+
+	sqe = get_sq_wqebb(qp, tmp_pi);
+	/* Clear the first 8Byte of the wqe hdr. */
+	*(uint64_t *)sqe = 0;
+
+	qp->sq.wr_tbl[tmp_pi & (qp->sq.depth - 1)] = wr->wr_id;
+
+	sqe_hdr = FIELD_PREP(ERDMA_SQE_HDR_QPN_MASK, qp->id) |
+		FIELD_PREP(ERDMA_SQE_HDR_CE_MASK, wr->send_flags & IBV_SEND_SIGNALED ? 1 : 0) |
+		FIELD_PREP(ERDMA_SQE_HDR_CE_MASK, qp->sq_sig_all) |
+		FIELD_PREP(ERDMA_SQE_HDR_SE_MASK, wr->send_flags & IBV_SEND_SOLICITED ? 1 : 0) |
+		FIELD_PREP(ERDMA_SQE_HDR_FENCE_MASK, wr->send_flags & IBV_SEND_FENCE ? 1 : 0) |
+		FIELD_PREP(ERDMA_SQE_HDR_INLINE_MASK, wr->send_flags & IBV_SEND_INLINE ? 1 : 0);
+
+	switch (wr->opcode) {
+	case IBV_WR_RDMA_WRITE:
+	case IBV_WR_RDMA_WRITE_WITH_IMM:
+		opcode = wr->opcode == IBV_WR_RDMA_WRITE ?
+			ERDMA_OP_WRITE : ERDMA_OP_WRITE_WITH_IMM;
+		sqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_OPCODE_MASK, opcode);
+		write_sqe = (struct erdma_write_sqe *)sqe;
+		write_sqe->imm_data = wr->imm_data;
+		write_sqe->sink_stag = htole32(wr->wr.rdma.rkey);
+		write_sqe->sink_to_low = htole32(wr->wr.rdma.remote_addr & 0xFFFFFFFF);
+		write_sqe->sink_to_high = htole32((wr->wr.rdma.remote_addr >> 32) & 0xFFFFFFFF);
+
+		length_field = &write_sqe->length;
+		sgl_base = get_sq_wqebb(qp, tmp_pi + 1);
+		/* sgl is in next wqebb. */
+		sgl_off = 0;
+		sgl_idx = tmp_pi + 1;
+		wqe_size = sizeof(struct erdma_write_sqe);
+
+		break;
+	case IBV_WR_SEND:
+	case IBV_WR_SEND_WITH_IMM:
+		opcode = wr->opcode == IBV_WR_SEND ? ERDMA_OP_SEND : ERDMA_OP_SEND_WITH_IMM;
+		sqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_OPCODE_MASK, opcode);
+		send_sqe = (struct erdma_send_sqe *)sqe;
+		send_sqe->imm_data = wr->imm_data;
+
+		length_field = &send_sqe->length;
+		sgl_base = (void *)send_sqe;
+		/* sgl is in the half of current wqebb */
+		sgl_off = 16;
+		sgl_idx = tmp_pi;
+		wqe_size = sizeof(struct erdma_send_sqe);
+
+		break;
+	case IBV_WR_RDMA_READ:
+		sqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_OPCODE_MASK, ERDMA_OP_READ);
+
+		read_sqe = (struct erdma_readreq_sqe *)sqe;
+
+		read_sqe->sink_to_low = htole32(wr->sg_list->addr & 0xFFFFFFFF);
+		read_sqe->sink_to_high = htole32((wr->sg_list->addr >> 32) & 0xFFFFFFFF);
+		read_sqe->sink_stag = htole32(wr->sg_list->lkey);
+		read_sqe->length = htole32(wr->sg_list->length);
+
+		struct erdma_sge *sgl = (struct erdma_sge *)get_sq_wqebb(qp, tmp_pi + 1);
+
+		sgl->laddr = htole64(wr->wr.rdma.remote_addr);
+		sgl->length = htole32(wr->sg_list->length);
+		sgl->lkey = htole32(wr->wr.rdma.rkey);
+
+		wqe_size = sizeof(struct erdma_readreq_sqe);
+
+		goto out;
+	default:
+		return -EINVAL;
+	}
+
+	if (wr->send_flags & IBV_SEND_INLINE) {
+		char *data = (char *)sgl_base;
+		uint32_t remain_size;
+		uint32_t copy_size;
+		uint32_t data_off;
+
+		i = 0;
+		bytes = 0;
+
+		/* Allow more than ERDMA_MAX_SGE, since content copied here */
+		while (i < wr->num_sge) {
+			bytes += wr->sg_list[i].length;
+			if (bytes > (int)ERDMA_MAX_INLINE)
+				return -EINVAL;
+
+			remain_size = wr->sg_list[i].length;
+			data_off = 0;
+
+			while (1) {
+				copy_size = min(remain_size, SQEBB_SIZE - sgl_off);
+				memcpy(data + sgl_off,
+					(void *)(uintptr_t)wr->sg_list[i].addr + data_off,
+					copy_size);
+				remain_size -= copy_size;
+
+				/* Update sgl_offset. */
+				sgl_idx += ((sgl_off + copy_size) >> SQEBB_SHIFT);
+				sgl_off = (sgl_off + copy_size) & (SQEBB_SIZE - 1);
+				data_off += copy_size;
+				data = get_sq_wqebb(qp, sgl_idx);
+
+				if (!remain_size)
+					break;
+			};
+
+			i++;
+		}
+
+		*length_field = htole32(bytes);
+		wqe_size += bytes;
+		sqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_SGL_LEN_MASK, bytes);
+	} else {
+		char *sgl = (char *)sgl_base;
+
+		if (wr->num_sge > ERDMA_MAX_SEND_SGE)
+			return -EINVAL;
+
+		i = 0;
+		bytes = 0;
+
+		while (i < wr->num_sge) {
+			bytes += wr->sg_list[i].length;
+			memcpy(sgl + sgl_off, &wr->sg_list[i], sizeof(struct ibv_sge));
+
+			if (sgl_off == 0)
+				*(uint32_t *)(sgl + 28) = qp->id;
+
+			sgl_idx += (sgl_off == sizeof(struct ibv_sge) ? 1 : 0);
+			sgl = get_sq_wqebb(qp, sgl_idx);
+			sgl_off = sizeof(struct ibv_sge) - sgl_off;
+
+			i++;
+		}
+
+		*length_field = htole32(bytes);
+		sqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_SGL_LEN_MASK, wr->num_sge);
+		wqe_size += wr->num_sge * sizeof(struct ibv_sge);
+	}
+
+out:
+	wqebb_cnt = SQEBB_COUNT(wqe_size);
+	assert(wqebb_cnt <= MAX_WQEBB_PER_SQE);
+	sqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_WQEBB_CNT_MASK, wqebb_cnt - 1);
+	sqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_WQEBB_INDEX_MASK, tmp_pi + wqebb_cnt);
+
+	*(__le64 *)sqe = htole64(sqe_hdr);
+	*sq_pi = tmp_pi + wqebb_cnt;
+
+	return 0;
+}
+
+int erdma_post_send(struct ibv_qp *base_qp, struct ibv_send_wr *wr, struct ibv_send_wr **bad_wr)
+{
+	struct erdma_qp *qp = to_eqp(base_qp);
+	uint16_t sq_pi;
+	int new_sqe = 0, rv = 0;
+
+	*bad_wr = NULL;
+
+	if (base_qp->state == IBV_QPS_ERR) {
+		*bad_wr = wr;
+		return -EIO;
+	}
+
+	pthread_spin_lock(&qp->sq_lock);
+
+	sq_pi = qp->sq.pi;
+
+	while (wr) {
+		if ((uint16_t)(sq_pi - qp->sq.ci) >= qp->sq.depth) {
+			rv = -ENOMEM;
+			*bad_wr = wr;
+			break;
+		}
+
+		rv = erdma_push_one_sqe(qp, wr, &sq_pi);
+		if (rv) {
+			*bad_wr = wr;
+			break;
+		}
+
+		new_sqe++;
+		wr = wr->next;
+	}
+
+	if (new_sqe) {
+		qp->sq.pi = sq_pi;
+		__kick_sq_db(qp, sq_pi); /* normal doorbell. */
+	}
+
+	pthread_spin_unlock(&qp->sq_lock);
+
+	return rv;
+}
+
+static int push_recv_wqe(struct erdma_qp *qp, struct ibv_recv_wr *wr)
+{
+	uint16_t rq_pi = qp->rq.pi;
+	uint16_t idx = rq_pi & (qp->rq.depth - 1);
+	struct erdma_rqe *rqe = (struct erdma_rqe *)qp->rq.qbuf + idx;
+
+	if ((uint16_t)(rq_pi - qp->rq.ci) == qp->rq.depth)
+		return -ENOMEM;
+
+	rqe->qe_idx = htole16(rq_pi + 1);
+	rqe->qpn = htole32(qp->id);
+	qp->rq.wr_tbl[idx] = wr->wr_id;
+
+	if (wr->num_sge == 0) {
+		rqe->length = 0;
+	} else if (wr->num_sge == 1) {
+		rqe->stag = htole32(wr->sg_list[0].lkey);
+		rqe->to = htole64(wr->sg_list[0].addr);
+		rqe->length = htole32(wr->sg_list[0].length);
+	} else {
+		return -EINVAL;
+	}
+
+	*(__le64 *)qp->rq.db_record = *(__le64 *)rqe;
+	udma_to_device_barrier();
+	mmio_write64_le(qp->rq.db, *(__le64 *)rqe);
+
+	qp->rq.pi = rq_pi + 1;
+
+	return 0;
+}
+
+int erdma_post_recv(struct ibv_qp *base_qp, struct ibv_recv_wr *wr, struct ibv_recv_wr **bad_wr)
+{
+	struct erdma_qp *qp = to_eqp(base_qp);
+	int ret = 0;
+
+	if (base_qp->state == IBV_QPS_ERR) {
+		*bad_wr = wr;
+		return -EIO;
+	}
+
+	pthread_spin_lock(&qp->rq_lock);
+
+	while (wr) {
+		ret = push_recv_wqe(qp, wr);
+		if (ret) {
+			*bad_wr = wr;
+			break;
+		}
+
+		wr = wr->next;
+	}
+
+	pthread_spin_unlock(&qp->rq_lock);
+
+	return ret;
+}
+
+
+void erdma_cq_event(struct ibv_cq *ibcq)
+{
+	struct erdma_cq *cq = to_ecq(ibcq);
+
+	cq->cmdsn++;
+}
+
+static const struct {
+	enum erdma_opcode erdma;
+	enum ibv_wc_opcode base;
+} map_cqe_opcode[ERDMA_NUM_OPCODES] = {
+	{ ERDMA_OP_WRITE, IBV_WC_RDMA_WRITE },
+	{ ERDMA_OP_READ, IBV_WC_RDMA_READ },
+	{ ERDMA_OP_SEND, IBV_WC_SEND },
+	{ ERDMA_OP_SEND_WITH_IMM, IBV_WC_SEND },
+	{ ERDMA_OP_RECEIVE, IBV_WC_RECV },
+	{ ERDMA_OP_RECV_IMM, IBV_WC_RECV_RDMA_WITH_IMM },
+	{ ERDMA_OP_RECV_INV, IBV_WC_LOCAL_INV},  /* can not appear */
+	{ ERDMA_OP_REQ_ERR, IBV_WC_RECV },  /* can not appear */
+	{ ERDNA_OP_READ_RESPONSE, IBV_WC_RECV }, /* can not appear */
+	{ ERDMA_OP_WRITE_WITH_IMM, IBV_WC_RDMA_WRITE },
+	{ ERDMA_OP_RECV_ERR, IBV_WC_RECV_RDMA_WITH_IMM }, /* can not appear */
+	{ ERDMA_OP_INVALIDATE, IBV_WC_LOCAL_INV },
+	{ ERDMA_OP_RSP_SEND_IMM, IBV_WC_RECV },
+	{ ERDMA_OP_SEND_WITH_INV, IBV_WC_SEND },
+	{ ERDMA_OP_REG_MR, IBV_WC_RECV }, /* can not appear */
+	{ ERDMA_OP_LOCAL_INV, IBV_WC_LOCAL_INV },
+	{ ERDMA_OP_READ_WITH_INV, IBV_WC_RDMA_READ },
+};
+
+static const struct {
+	enum erdma_wc_status erdma;
+	enum ibv_wc_status base;
+	enum erdma_vendor_err vendor;
+} map_cqe_status[ERDMA_NUM_WC_STATUS] = {
+	{ ERDMA_WC_SUCCESS, IBV_WC_SUCCESS, ERDMA_WC_VENDOR_NO_ERR },
+	{ ERDMA_WC_GENERAL_ERR, IBV_WC_GENERAL_ERR, ERDMA_WC_VENDOR_NO_ERR },
+	{ ERDMA_WC_RECV_WQE_FORMAT_ERR, IBV_WC_GENERAL_ERR, ERDMA_WC_VENDOR_INVALID_RQE },
+	{ ERDMA_WC_RECV_STAG_INVALID_ERR,
+		IBV_WC_REM_ACCESS_ERR, ERDMA_WC_VENDOR_RQE_INVALID_STAG },
+	{ ERDMA_WC_RECV_ADDR_VIOLATION_ERR,
+		IBV_WC_REM_ACCESS_ERR, ERDMA_WC_VENDOR_RQE_ADDR_VIOLATION },
+	{ ERDMA_WC_RECV_RIGHT_VIOLATION_ERR,
+		IBV_WC_REM_ACCESS_ERR, ERDMA_WC_VENDOR_RQE_ACCESS_RIGHT_ERR },
+	{ ERDMA_WC_RECV_PDID_ERR, IBV_WC_REM_ACCESS_ERR, ERDMA_WC_VENDOR_RQE_INVALID_PD },
+	{ ERDMA_WC_RECV_WARRPING_ERR, IBV_WC_REM_ACCESS_ERR, ERDMA_WC_VENDOR_RQE_WRAP_ERR },
+	{ ERDMA_WC_SEND_WQE_FORMAT_ERR, IBV_WC_LOC_QP_OP_ERR, ERDMA_WC_VENDOR_INVALID_SQE },
+	{ ERDMA_WC_SEND_WQE_ORD_EXCEED, IBV_WC_GENERAL_ERR, ERDMA_WC_VENDOR_ZERO_ORD },
+	{ ERDMA_WC_SEND_STAG_INVALID_ERR,
+		IBV_WC_LOC_ACCESS_ERR, ERDMA_WC_VENDOR_SQE_INVALID_STAG },
+	{ ERDMA_WC_SEND_ADDR_VIOLATION_ERR,
+		IBV_WC_LOC_ACCESS_ERR, ERDMA_WC_VENDOR_SQE_ADDR_VIOLATION },
+	{ ERDMA_WC_SEND_RIGHT_VIOLATION_ERR,
+		IBV_WC_LOC_ACCESS_ERR, ERDMA_WC_VENDOR_SQE_ACCESS_ERR },
+	{ ERDMA_WC_SEND_PDID_ERR, IBV_WC_LOC_ACCESS_ERR, ERDMA_WC_VENDOR_SQE_INVALID_PD },
+	{ ERDMA_WC_SEND_WARRPING_ERR, IBV_WC_LOC_ACCESS_ERR, ERDMA_WC_VENDOR_SQE_WARP_ERR },
+	{ ERDMA_WC_FLUSH_ERR, IBV_WC_WR_FLUSH_ERR, ERDMA_WC_VENDOR_NO_ERR },
+	{ ERDMA_WC_RETRY_EXC_ERR, IBV_WC_RETRY_EXC_ERR, ERDMA_WC_VENDOR_NO_ERR },
+};
+
+#define ERDMA_POLLCQ_NO_QP      (-1)
+#define ERDMA_POLLCQ_DUP_COMP   (-2)
+#define ERDMA_POLLCQ_WRONG_IDX  (-3)
+
+static int __erdma_poll_one_cqe(struct erdma_context *ctx, struct erdma_cq *cq,
+				struct erdma_cqe *cqe, uint32_t cqe_hdr, struct ibv_wc *wc)
+{
+	struct erdma_qp *qp;
+	uint64_t *qeidx2wrid = NULL;
+	uint32_t qpn = be32toh(cqe->qpn);
+	uint16_t depth = 0;
+	uint64_t *sqe_hdr;
+	uint16_t wqe_idx = be32toh(cqe->qe_idx);
+	uint16_t old_ci, new_ci;
+	uint32_t opcode = FIELD_GET(ERDMA_CQE_HDR_OPCODE_MASK, cqe_hdr);
+	uint32_t syndrome = FIELD_GET(ERDMA_CQE_HDR_SYNDROME_MASK, cqe_hdr);
+
+	int tbl_idx = qpn >> ERDMA_QP_TABLE_SHIFT;
+	int tbl_off = qpn & ERDMA_QP_TABLE_MASK;
+
+	if (!ctx->qp_table[tbl_idx].table || !ctx->qp_table[tbl_idx].table[tbl_off])
+		return ERDMA_POLLCQ_NO_QP;
+
+	qp = ctx->qp_table[tbl_idx].table[tbl_off];
+
+	if (FIELD_GET(ERDMA_CQE_HDR_QTYPE_MASK, cqe_hdr) == ERDMA_CQE_QTYPE_SQ) {
+		qeidx2wrid = qp->sq.wr_tbl;
+		depth = qp->sq.depth;
+		sqe_hdr = (uint64_t *)get_sq_wqebb(qp, wqe_idx);
+		old_ci = qp->sq.ci;
+		new_ci = wqe_idx + FIELD_GET(ERDMA_SQE_HDR_WQEBB_CNT_MASK, *sqe_hdr) + 1;
+
+		if ((uint16_t)(new_ci - old_ci) > depth)
+			return ERDMA_POLLCQ_WRONG_IDX;
+		else if (new_ci == old_ci)
+			return ERDMA_POLLCQ_DUP_COMP;
+
+		qp->sq.ci = new_ci;
+	} else {
+		qeidx2wrid = qp->rq.wr_tbl;
+		depth = qp->rq.depth;
+		qp->rq.ci++;
+	}
+
+	wc->wr_id = qeidx2wrid[wqe_idx & (depth - 1)];
+	wc->byte_len = be32toh(cqe->size);
+	wc->wc_flags = 0;
+
+	if (opcode == ERDMA_OP_RECV_IMM) {
+		wc->opcode = IBV_WC_RECV_RDMA_WITH_IMM;
+		wc->imm_data = htobe32(le32toh(cqe->imm_data));
+		wc->wc_flags |= IBV_WC_WITH_IMM;
+	} else if (opcode == ERDMA_OP_RSP_SEND_IMM) {
+		wc->opcode = IBV_WC_RECV;
+		wc->imm_data = htobe32(le32toh(cqe->imm_data));
+		wc->wc_flags |= IBV_WC_WITH_IMM;
+	} else {
+		wc->opcode = map_cqe_opcode[opcode].base;
+	}
+
+	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_num = qpn;
+
+	return 0;
+}
+
+int erdma_poll_cq(struct ibv_cq *ibcq, int num_entries, struct ibv_wc *wc)
+{
+	struct erdma_cq *cq = to_ecq(ibcq);
+	struct erdma_context *ctx = to_ectx(ibcq->context);
+	int new = 0;
+	struct erdma_cqe *cqe;
+	int owner;
+	uint32_t ci;
+	uint32_t depth_mask = cq->depth - 1;
+	uint32_t hdr;
+	int i, ret;
+
+	pthread_spin_lock(&cq->lock);
+
+	owner = cq->owner;
+	ci = cq->ci;
+
+	for (i = 0; i < num_entries; i++) {
+		cqe = &cq->queue[ci & depth_mask];
+		hdr = be32toh(cqe->hdr);
+
+		if (FIELD_GET(ERDMA_CQE_HDR_OWNER_MASK, hdr) != owner)
+			break;
+
+		udma_from_device_barrier();
+
+		ret = __erdma_poll_one_cqe(ctx, cq, cqe, hdr, wc);
+
+		ci++;
+		if ((ci & depth_mask) == 0)
+			owner = !owner;
+
+		if (ret)
+			continue;
+
+		wc++;
+		new++;
+	}
+
+	cq->owner = owner;
+	cq->ci = ci;
+
+	pthread_spin_unlock(&cq->lock);
+
+	return new;
+}
+
+void erdma_free_context(struct ibv_context *ibv_ctx)
+{
+	struct erdma_context *ctx = to_ectx(ibv_ctx);
+	int i;
+
+	munmap(ctx->sdb, ERDMA_PAGE_SIZE);
+	munmap(ctx->rdb, ERDMA_PAGE_SIZE);
+	munmap(ctx->cdb, ERDMA_PAGE_SIZE);
+
+	pthread_mutex_lock(&ctx->qp_table_mutex);
+	for (i = 0; i < ERDMA_QP_TABLE_SIZE; ++i) {
+		if (ctx->qp_table[i].refcnt)
+			free(ctx->qp_table[i].table);
+	}
+
+	pthread_mutex_unlock(&ctx->qp_table_mutex);
+	pthread_mutex_destroy(&ctx->qp_table_mutex);
+
+	verbs_uninit_context(&ctx->ibv_ctx);
+	free(ctx);
+}