RDMA编程实例rdma_cm API

RDMA编程基础
存储大师班 | RDMA简介与编程基础 -https://zhuanlan.zhihu.com/p/387549948

1. RDMA的学习环境搭建
RDMA需要专门的RDMA网卡或者InfiniBand卡才能使用，学习RDMA而又没有这些硬件设备，可以使用一个软件RDMA模拟环境，softiwarp ，
- 这是加载地址：https://github.com/zrlio/softiwarp
- 这是安装教程：http://www.reflectionsofthevoid.com/2011/03/how-to-install-soft-iwarp-on-ubuntu.html

更多的rdmacm实例：，
- https://github.com/tarickb/the-geek-in-the-corner
需要注意的是，这个例子里面缺省用的是IPv6连接，如果希望在IPv4环境下测试，需要先改代码用IPv4地址。

2. RDMA与socket的类比
和Socket连接类似，RDMA连接也分为可靠连接和不可靠连接。然而也不完全相同，Socket的可靠连接就是TCP连接，是流式的；不可靠连接也就是UDP，是消息式的。对于RDMA来说，无论是可靠连接和不可靠连接，都是消息式的。
编程角度看，RDMA代码也分为Server端，Client端，也有bind, listen, connect, accept,等动作，然而细节上仍有不少区别。

大家可以关注一下mellonx的vma，貌似可以直接用socket api通信，方便很多：
【RDMA】降低CPU除了RDMA (vbers）还是VMA ？|使用socket进行RDMA编程？_bandaoyu的note-CSDN博客
前言看介绍，像是mellonx针对其kernel bypass网卡（RDMA网卡）提供的一个lib库，该lib库对外提供socket api，使得用户的程序不需要修改就可以直接使用kernel bypass网卡（如RDMA网卡）。我们都知道RDMA 网卡目前使用的是rdma_cm和vbers api编程，和socket不一样，如果能用socket对RDMA编程，那确实是很大的利好。官网介绍什么是VMA？Mellanox Interconnect Community官方介绍：M
https://blog.csdn.net/bandaoyu/article/details/120726746

rdma_cm API说明：

https://linux.die.net/man/3/rdma_create_id （推荐）

https://www.ibm.com/docs/en/aix/7.2?topic=operations-rdma-listen （内容少）

rdma_cm API 管理连接（建立连接和销毁）+vbers api 管理收发

RDMA主机使用queue pairs（QP）进行通信；主机创建由发送队列SQ和接收队列RQ组成的QP，并使用verbs API将操作post 到这些队列。（所以rdma_cm是管理连接的，收发还是verbs API。）

3. RDMA服务器的代码流程
main（）
{
channel=rdma_create_event_channel
这一步是创建一个event channel，event channel是RDMA设备在操作完成后，或者有连接请求等事件发生时，用来通知应用程序的通道。其内部就是一个file descriptor, 因此可以进行poll等操作。

rdma_create_id(channel, **id，……)
这一步创建一个rdma_cm_id, 概念上等价与socket编程时的listen socket。

rdma_bind_addr(id,addr)
和socket编程一样，也要先绑定一个本地的地址和端口，以进行listen操作。

rdma_listen(id,block)
开始侦听客户端的连接请求

rdma_get_cm_event（channel，&event）
这个调用就是作用在第一步创建的event channel上面，要从event channel中获取一个事件。这是个阻塞调用，只有有事件时才会返回。在一切正常的情况下，函数返回时会得到一个 RDMA_CM_EVENT_CONNECT_REQUEST事件，也就是说，有客户端发起连接了。
在事件的参数里面，会有一个新的rdma_cm_id传入。这点和socket是不同的，socket只有在accept后才有新的socket fd创建。

on_event()

{

     on_connect_request()//RDMA_CM_EVENT_CONNECT_REQUEST

    {

        build_context()

      {

6.ibv_alloc_pd

创建一个protection domain。protection domain可以看作是一个内存保护单位，在内存区域和队列直接建立一个关联关系，防止未授权的访问。

7.ibv_create_comp_channel

和之前创建的event channel类似，这也是一个event channel，但只用来报告【完成队列】里面的事件。当【完成队列】里有新的任务完成时，就通过这个channel向应用程序报告。

8.ibv_create_cq

创建【完成队列】，创建时就指定使用第6步的channel。

}//--end build_context()

9.rdma_create_qp
创建一个queue pair, 一个queue pair包括一个发送queue和一个接收queue. 指定使用前面创建的cq作为完成队列。该qp创建时就指定关联到第6步创建的pd上。

10.ibv_reg_mr
注册内存区域。RDMA使用的内存，必须事先进行注册。这个是可以理解的，DMA的内存在边界对齐，能否被swap等方面，都有要求。

11.rdma_accept
至此，做好了全部的准备工作，可以调用accept接受客户端的这个请求了。 –：）长出一口气 ~~ 且慢，

}

//--end on_connect_request()

12.rdma_ack_cm_event
对于每个从event channel得到的事件，都要调用ack函数，否则会产生内存泄漏。这一步的ack是对应第5步的get。每一次get调用，都要有对应的ack调用。

13.rdma_get_cm_event
继续调用rdma_get_cm_event, 一切正常的话我们此时应该得到 RDMA_CM_EVENT_ESTABLISHED 事件，表示连接已经建立起来。不需要做额外的处理，直接rdma_ack_cm_event就行了

}//--end on_event()

终于可以开始进行数据传输了 ==== (如何传输下篇再说)

参考：http://10.165.104.246:8080/#/c/43882/

4. 关闭连接

断开连接
当rdma_get_cm_event返回RDMA_CM_EVENT_DISCONNECTED事件时，表示客户端断开了连接，server端要进行对应的清理。此时可以调用rdma_ack_cm_event释放事件资源。然后依次调用下面的函数，释放连接资源，内存资源，队列资源。

rdma_disconnect

rdma_destroy_qp

ibv_dereg_mr

rdma_destroy_id
释放同客户端连接的rdma_cm_id

rdma_destroy_id
释放用于侦听的rdma_cm_id

rdma_destroy_event_channel
释放 event channel

}
// end main

实例
源码地址- https://github.com/tarickb/the-geek-in-the-corner

用法
[root@localhost 01_basic-client-server]# ./server 
listening on port 42956.

client <server-address> <server-port>
 

Makefile
.PHONY: clean

CFLAGS  := -Wall -g
LDLIBS  := ${LDLIBS} -lrdmacm -libverbs -lpthread

APPS    := server client

all: ${APPS}

clean:
    rm -f ${APPS}

注意：makefile 没有-L 指定lib的路径，所以 -lrdmacm -libverbs -lpthread 对应的库 librdmacm.so libibverbs.so libpthread.so 应放在默认的路径下/usr/lib 或/usr/lib64

服务端server.c：

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <rdma/rdma_cma.h>
 
#define TEST_NZ(x) do { if ( (x)) die("error: " #x " failed (returned non-zero)." ); } while (0)
#define TEST_Z(x)  do { if (!(x)) die("error: " #x " failed (returned zero/null)."); } while (0)
 
const int BUFFER_SIZE = 1024;
 
struct context {
  struct ibv_context *ctx;
  struct ibv_pd *pd;
  struct ibv_cq *cq;
  struct ibv_comp_channel *comp_channel;
 
  pthread_t cq_poller_thread;
};
 
struct connection {
  struct ibv_qp *qp;
 
  struct ibv_mr *recv_mr;
  struct ibv_mr *send_mr;
 
  char *recv_region;
  char *send_region;
};
 
static void die(const char *reason);
 
static void build_context(struct ibv_context *verbs);
static void build_qp_attr(struct ibv_qp_init_attr *qp_attr);
static void * poll_cq(void *);
static void post_receives(struct connection *conn);
static void register_memory(struct connection *conn);
 
static void on_completion(struct ibv_wc *wc);
static int on_connect_request(struct rdma_cm_id *id);
static int on_connection(void *context);
static int on_disconnect(struct rdma_cm_id *id);
static int on_event(struct rdma_cm_event *event);
 
static struct context *s_ctx = NULL;
 
int main(int argc, char **argv)
{
#if _USE_IPV6
  struct sockaddr_in6 addr;
#else
  struct sockaddr_in addr;
#endif
  struct rdma_cm_event *event = NULL;
  struct rdma_cm_id *listener = NULL;
  struct rdma_event_channel *ec = NULL;
  uint16_t port = 0;
 
  memset(&addr, 0, sizeof(addr));
#if _USE_IPV6
  addr.sin6_family = AF_INET6;
#else
  addr.sin_family = AF_INET;
#endif
 
  TEST_Z(ec = rdma_create_event_channel());
  TEST_NZ(rdma_create_id(ec, &listener, NULL, RDMA_PS_TCP));
  TEST_NZ(rdma_bind_addr(listener, (struct sockaddr *)&addr));
  TEST_NZ(rdma_listen(listener, 10)); /* backlog=10 is arbitrary */
 
  port = ntohs(rdma_get_src_port(listener)); //rdma_get_src_port 返回listener对应的tcp 端口
 
  printf("listening on port %d.\n", port);
 
  while (rdma_get_cm_event(ec, &event) == 0) {
    struct rdma_cm_event event_copy;
 
    memcpy(&event_copy, event, sizeof(*event));
    rdma_ack_cm_event(event);
 
    if (on_event(&event_copy))
      break;
  }
 
  rdma_destroy_id(listener);
  rdma_destroy_event_channel(ec);
 
  return 0;
}
 
void die(const char *reason)
{
  fprintf(stderr, "%s\n", reason);
  exit(EXIT_FAILURE);
}
 
void build_context(struct ibv_context *verbs)
{
  if (s_ctx) {
    if (s_ctx->ctx != verbs)
      die("cannot handle events in more than one context.");
 
    return;
  }
 
  s_ctx = (struct context *)malloc(sizeof(struct context));
 
  s_ctx->ctx = verbs;
 
  TEST_Z(s_ctx->pd = ibv_alloc_pd(s_ctx->ctx));
  TEST_Z(s_ctx->comp_channel = ibv_create_comp_channel(s_ctx->ctx));
  TEST_Z(s_ctx->cq = ibv_create_cq(s_ctx->ctx, 10, NULL, s_ctx->comp_channel, 0)); /* cqe=10 is arbitrary */
  TEST_NZ(ibv_req_notify_cq(s_ctx->cq, 0)); #完成完成队列与完成通道的关联
 
  TEST_NZ(pthread_create(&s_ctx->cq_poller_thread, NULL, poll_cq, NULL));
}
 
void build_qp_attr(struct ibv_qp_init_attr *qp_attr)
{
  memset(qp_attr, 0, sizeof(*qp_attr));
 
  qp_attr->send_cq = s_ctx->cq;
  qp_attr->recv_cq = s_ctx->cq;
  qp_attr->qp_type = IBV_QPT_RC;
 
  qp_attr->cap.max_send_wr = 10;
  qp_attr->cap.max_recv_wr = 10;
  qp_attr->cap.max_send_sge = 1;
  qp_attr->cap.max_recv_sge = 1;
}
 
void * poll_cq(void *ctx)
{
  struct ibv_cq *cq;
  struct ibv_wc wc;
 
  while (1) {
    TEST_NZ(ibv_get_cq_event(s_ctx->comp_channel, &cq, &ctx));
    ibv_ack_cq_events(cq, 1);
    TEST_NZ(ibv_req_notify_cq(cq, 0));
 
    while (ibv_poll_cq(cq, 1, &wc))
      on_completion(&wc);
  }
 
  return NULL;
}
 
void post_receives(struct connection *conn)
{
  struct ibv_recv_wr wr, *bad_wr = NULL;
  struct ibv_sge sge;
 
  wr.wr_id = (uintptr_t)conn;
  wr.next = NULL;
  wr.sg_list = &sge;
  wr.num_sge = 1;
 
  sge.addr = (uintptr_t)conn->recv_region;
  sge.length = BUFFER_SIZE;
  sge.lkey = conn->recv_mr->lkey;
 
  TEST_NZ(ibv_post_recv(conn->qp, &wr, &bad_wr));
}
 
void register_memory(struct connection *conn)
{
  conn->send_region = malloc(BUFFER_SIZE);
  conn->recv_region = malloc(BUFFER_SIZE);
 
  TEST_Z(conn->send_mr = ibv_reg_mr(
    s_ctx->pd,
    conn->send_region,
    BUFFER_SIZE,
    0));
 
  TEST_Z(conn->recv_mr = ibv_reg_mr(
    s_ctx->pd,
    conn->recv_region,
    BUFFER_SIZE,
    IBV_ACCESS_LOCAL_WRITE));
}
 
void on_completion(struct ibv_wc *wc)
{
  if (wc->status != IBV_WC_SUCCESS)
    die("on_completion: status is not IBV_WC_SUCCESS.");
 
  if (wc->opcode & IBV_WC_RECV) {
    struct connection *conn = (struct connection *)(uintptr_t)wc->wr_id;
 
    printf("received message: %s\n", conn->recv_region);
 
  } else if (wc->opcode == IBV_WC_SEND) {
    printf("send completed successfully.\n");
  }
}
 
int on_connect_request(struct rdma_cm_id *id)
{
  struct ibv_qp_init_attr qp_attr;
  struct rdma_conn_param cm_params;
  struct connection *conn;
 
  printf("received connection request.\n");
 
  build_context(id->verbs);
  build_qp_attr(&qp_attr);
 
  TEST_NZ(rdma_create_qp(id, s_ctx->pd, &qp_attr));
 
  id->context = conn = (struct connection *)malloc(sizeof(struct connection));
  conn->qp = id->qp;
 
  register_memory(conn);
  post_receives(conn);
 
  memset(&cm_params, 0, sizeof(cm_params));
  TEST_NZ(rdma_accept(id, &cm_params));
 
  return 0;
}
 
int on_connection(void *context)
{
  struct connection *conn = (struct connection *)context;
  struct ibv_send_wr wr, *bad_wr = NULL;
  struct ibv_sge sge;
 
  snprintf(conn->send_region, BUFFER_SIZE, "message from passive/server side with pid %d", getpid());
 
  printf("connected. posting send...\n");
 
  memset(&wr, 0, sizeof(wr));
 
  wr.opcode = IBV_WR_SEND;
  wr.sg_list = &sge;
  wr.num_sge = 1;
  wr.send_flags = IBV_SEND_SIGNALED;
 
  sge.addr = (uintptr_t)conn->send_region;
  sge.length = BUFFER_SIZE;
  sge.lkey = conn->send_mr->lkey;
 
  TEST_NZ(ibv_post_send(conn->qp, &wr, &bad_wr));
 
  return 0;
}
 
int on_disconnect(struct rdma_cm_id *id)
{
  struct connection *conn = (struct connection *)id->context;
 
  printf("peer disconnected.\n");
 
  rdma_destroy_qp(id);
 
  ibv_dereg_mr(conn->send_mr);
  ibv_dereg_mr(conn->recv_mr);
 
  free(conn->send_region);
  free(conn->recv_region);
 
  free(conn);
 
  rdma_destroy_id(id);
 
  return 0;
}
 
int on_event(struct rdma_cm_event *event)
{
  int r = 0;
 
  if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST)
    r = on_connect_request(event->id);
  else if (event->event == RDMA_CM_EVENT_ESTABLISHED)
    r = on_connection(event->id->context);
  else if (event->event == RDMA_CM_EVENT_DISCONNECTED)
    r = on_disconnect(event->id);
  else
    die("on_event: unknown event.");
 
  return r;
}

客户端client.c：

#include <netdb.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <rdma/rdma_cma.h>
 
#define TEST_NZ(x) do { if ( (x)) die("error: " #x " failed (returned non-zero)." ); } while (0)
#define TEST_Z(x)  do { if (!(x)) die("error: " #x " failed (returned zero/null)."); } while (0)
 
const int BUFFER_SIZE = 1024;
const int TIMEOUT_IN_MS = 500; /* ms */
 
struct context {
  struct ibv_context *ctx;
  struct ibv_pd *pd;
  struct ibv_cq *cq;
  struct ibv_comp_channel *comp_channel;
 
  pthread_t cq_poller_thread;
};
 
struct connection {
  struct rdma_cm_id *id;
  struct ibv_qp *qp;
 
  struct ibv_mr *recv_mr;
  struct ibv_mr *send_mr;
 
  char *recv_region;
  char *send_region;
 
  int num_completions;
};
 
static void die(const char *reason);
 
static void build_context(struct ibv_context *verbs);
static void build_qp_attr(struct ibv_qp_init_attr *qp_attr);
static void * poll_cq(void *);
static void post_receives(struct connection *conn);
static void register_memory(struct connection *conn);
 
static int on_addr_resolved(struct rdma_cm_id *id);
static void on_completion(struct ibv_wc *wc);
static int on_connection(void *context);
static int on_disconnect(struct rdma_cm_id *id);
static int on_event(struct rdma_cm_event *event);
static int on_route_resolved(struct rdma_cm_id *id);
 
static struct context *s_ctx = NULL;
 
int main(int argc, char **argv)
{
  struct addrinfo *addr;
  struct rdma_cm_event *event = NULL;
  struct rdma_cm_id *conn= NULL;
  struct rdma_event_channel *ec = NULL;
 
  if (argc != 3)
    die("usage: client <server-address> <server-port>");
 
  TEST_NZ(getaddrinfo(argv[1], argv[2], NULL, &addr));
 
  TEST_Z(ec = rdma_create_event_channel());
  TEST_NZ(rdma_create_id(ec, &conn, NULL, RDMA_PS_TCP));
  TEST_NZ(rdma_resolve_addr(conn, NULL, addr->ai_addr, TIMEOUT_IN_MS));
 
  freeaddrinfo(addr);
 
  while (rdma_get_cm_event(ec, &event) == 0) {
    struct rdma_cm_event event_copy;
 
    memcpy(&event_copy, event, sizeof(*event));
    rdma_ack_cm_event(event);
 
    if (on_event(&event_copy))
      break;
  }
 
  rdma_destroy_event_channel(ec);
 
  return 0;
}
 
void die(const char *reason)
{
  fprintf(stderr, "%s\n", reason);
  exit(EXIT_FAILURE);
}
 
void build_context(struct ibv_context *verbs)
{
  if (s_ctx) {
    if (s_ctx->ctx != verbs)
      die("cannot handle events in more than one context.");
 
    return;
  }
 
  s_ctx = (struct context *)malloc(sizeof(struct context));
 
  s_ctx->ctx = verbs;
 
  TEST_Z(s_ctx->pd = ibv_alloc_pd(s_ctx->ctx));
  TEST_Z(s_ctx->comp_channel = ibv_create_comp_channel(s_ctx->ctx));
  TEST_Z(s_ctx->cq = ibv_create_cq(s_ctx->ctx, 10, NULL, s_ctx->comp_channel, 0)); /* cqe=10 is arbitrary */
  TEST_NZ(ibv_req_notify_cq(s_ctx->cq, 0));
 
  TEST_NZ(pthread_create(&s_ctx->cq_poller_thread, NULL, poll_cq, NULL));
}
 
void build_qp_attr(struct ibv_qp_init_attr *qp_attr)
{
  memset(qp_attr, 0, sizeof(*qp_attr));
 
  qp_attr->send_cq = s_ctx->cq;
  qp_attr->recv_cq = s_ctx->cq;
  qp_attr->qp_type = IBV_QPT_RC;
 
  qp_attr->cap.max_send_wr = 10;
  qp_attr->cap.max_recv_wr = 10;
  qp_attr->cap.max_send_sge = 1;
  qp_attr->cap.max_recv_sge = 1;
}
 
void * poll_cq(void *ctx)
{
  struct ibv_cq *cq;
  struct ibv_wc wc;
 
  while (1) {
    TEST_NZ(ibv_get_cq_event(s_ctx->comp_channel, &cq, &ctx));
    ibv_ack_cq_events(cq, 1);
    TEST_NZ(ibv_req_notify_cq(cq, 0));
 
    while (ibv_poll_cq(cq, 1, &wc))
      on_completion(&wc);
  }
 
  return NULL;
}
 
void post_receives(struct connection *conn)
{
  struct ibv_recv_wr wr, *bad_wr = NULL;
  struct ibv_sge sge;
 
  wr.wr_id = (uintptr_t)conn;
  wr.next = NULL;
  wr.sg_list = &sge;
  wr.num_sge = 1;
 
  sge.addr = (uintptr_t)conn->recv_region;
  sge.length = BUFFER_SIZE;
  sge.lkey = conn->recv_mr->lkey;
 
  TEST_NZ(ibv_post_recv(conn->qp, &wr, &bad_wr));
}
 
void register_memory(struct connection *conn)
{
  conn->send_region = malloc(BUFFER_SIZE);
  conn->recv_region = malloc(BUFFER_SIZE);
 
  TEST_Z(conn->send_mr = ibv_reg_mr(
    s_ctx->pd, 
    conn->send_region, 
    BUFFER_SIZE, 
    0));
 
  TEST_Z(conn->recv_mr = ibv_reg_mr(
    s_ctx->pd, 
    conn->recv_region, 
    BUFFER_SIZE, 
    IBV_ACCESS_LOCAL_WRITE));
}
 
int on_addr_resolved(struct rdma_cm_id *id)
{
  struct ibv_qp_init_attr qp_attr;
  struct connection *conn;
 
  printf("address resolved.\n");
 
  build_context(id->verbs);
  build_qp_attr(&qp_attr);
 
  TEST_NZ(rdma_create_qp(id, s_ctx->pd, &qp_attr));
 
  id->context = conn = (struct connection *)malloc(sizeof(struct connection));
 
  conn->id = id;
  conn->qp = id->qp;
  conn->num_completions = 0;
 
  register_memory(conn);
  post_receives(conn);
 
  TEST_NZ(rdma_resolve_route(id, TIMEOUT_IN_MS));
 
  return 0;
}
 
void on_completion(struct ibv_wc *wc)
{
  struct connection *conn = (struct connection *)(uintptr_t)wc->wr_id;
 
  if (wc->status != IBV_WC_SUCCESS)
    die("on_completion: status is not IBV_WC_SUCCESS.");
 
  if (wc->opcode & IBV_WC_RECV)
    printf("received message: %s\n", conn->recv_region);
  else if (wc->opcode == IBV_WC_SEND)
    printf("send completed successfully.\n");
  else
    die("on_completion: completion isn't a send or a receive.");
 
  if (++conn->num_completions == 2)
    rdma_disconnect(conn->id);
}
 
int on_connection(void *context)
{
  struct connection *conn = (struct connection *)context;
  struct ibv_send_wr wr, *bad_wr = NULL;
  struct ibv_sge sge;
 
  snprintf(conn->send_region, BUFFER_SIZE, "message from active/client side with pid %d", getpid());
 
  printf("connected. posting send...\n");
 
  memset(&wr, 0, sizeof(wr));
 
  wr.wr_id = (uintptr_t)conn;
  wr.opcode = IBV_WR_SEND;
  wr.sg_list = &sge;
  wr.num_sge = 1;
  wr.send_flags = IBV_SEND_SIGNALED;
 
  sge.addr = (uintptr_t)conn->send_region;
  sge.length = BUFFER_SIZE;
  sge.lkey = conn->send_mr->lkey;
 
  TEST_NZ(ibv_post_send(conn->qp, &wr, &bad_wr));
 
  return 0;
}
 
int on_disconnect(struct rdma_cm_id *id)
{
  struct connection *conn = (struct connection *)id->context;
 
  printf("disconnected.\n");
 
  rdma_destroy_qp(id);
 
  ibv_dereg_mr(conn->send_mr);
  ibv_dereg_mr(conn->recv_mr);
 
  free(conn->send_region);
  free(conn->recv_region);
 
  free(conn);
 
  rdma_destroy_id(id);
 
  return 1; /* exit event loop */
}
 
int on_event(struct rdma_cm_event *event)
{
  int r = 0;
 
  if (event->event == RDMA_CM_EVENT_ADDR_RESOLVED)
    r = on_addr_resolved(event->id);
  else if (event->event == RDMA_CM_EVENT_ROUTE_RESOLVED)
    r = on_route_resolved(event->id);
  else if (event->event == RDMA_CM_EVENT_ESTABLISHED)
    r = on_connection(event->id->context);
  else if (event->event == RDMA_CM_EVENT_DISCONNECTED)
    r = on_disconnect(event->id);
  else
    die("on_event: unknown event.");
 
  return r;
}
 
int on_route_resolved(struct rdma_cm_id *id)
{
  struct rdma_conn_param cm_params;
 
  printf("route resolved.\n");
 
  memset(&cm_params, 0, sizeof(cm_params));
  TEST_NZ(rdma_connect(id, &cm_params));
 
  return 0;
}

更多讲解教程
InfiniBand, Verbs, RDMA | https://thegeekinthecorner.wordpress.com/category/infiniband-verbs-rdma/

RDMA read and write with IB verbs | https://thegeekinthecorner.wordpress.com/2010/09/28/rdma-read-and-write-with-ib-verbs/

http://www.hpcadvisorycouncil.com/pdf/building-an-rdma-capable-application-with-ib-verbs.pdf