RDMA编程-2 client端流程

在上一篇建立RDMA侦听部分讲了RDMA 通信时server端需要进行的操作，这一篇我们来说一下客户端的流程。

##client端流程
RDMA编程时的client端和socket编程也是大体类似的，只是有一些跟设备相关的细节，包括地址解析，queue pair, memory region 等要自己显式处理。

rdma_getaddrinfo
获取服务器的地址信息
rdma_create_event_channel
create channel to receive events
rdma_create_id
allocate an rdma_cm_id, this is conceptually similar to a socket
rdma_resolve_addr
obtain a local RDMA device to reach the remote address
rdma_get_cm_event
wait for RDMA_CM_EVENT_ADDR_RESOLVED event
rdma_ack_cm_event
ack event
rdma_create_qp
allocate a QP for the communication
rdma_resolve_route
determine the route to the remote address
rdma_get_cm_event
wait for RDMA_CM_EVENT_ROUTE_RESOLVED event
rdma_ack_cm_event
ack event
rdma_connect
connect to the remote server
rdma_get_cm_event
wait for RDMA_CM_EVENT_ESTABLISHED event
rdma_ack_cm_event
ack event
ibv_post_send
Perform data transfers over connection
rdma_disconnect
tear-down connection
rdma_get_cm_event
wait for RDMA_CM_EVENT_DISCONNECTED event
rdma_ack_cm_event
ack event
rdma_destroy_qp
destroy the QP
rdma_destroy_id
release the rdma_cm_id
rdma_destroy_event_channel
release the event channel

下面是一个完整的client代码，这个代码可以配合建立RDMA侦听里的server端代码进行测试。

测试时需要注意，不要使用使用localhost/127.0.0.1这样的回环地址去连接server端，回环地址会绕过RDMA网卡，因此无法进行RDMA通信。

//compile :  gcc rdma-client.cpp -libverbs -lrdmacm -o rdma-client

#define _ISOC11_SOURCE

#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#include <infiniband/verbs.h>
#include <rdma/rdma_cma.h>
#include <sys/types.h>
#include <sys/socket.h>
#include <netdb.h>


struct rdma_event_channel* ec;
struct rdma_cm_id* cm_id;
struct ibv_pd* pd;
struct ibv_context* rdma_cm_context;
void* recv_buf;
#define RECV_BUF_SIZE 4096
struct ibv_mr* recv_mr;



int on_route_resolved(struct rdma_cm_id* id)
{
	int rc = 0;
	struct rdma_conn_param cm_params;
	struct ibv_qp_init_attr qp_attr={0};
	struct PfRdmaConnection* conn = (struct PfRdmaConnection*)id->context;
	
	struct ibv_context* rdma_context = id->verbs;
	pd = ibv_alloc_pd(rdma_context);
	recv_buf = aligned_alloc(4096, RECV_BUF_SIZE);
	recv_mr = ibv_reg_mr(pd, recv_buf, RECV_BUF_SIZE, IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE);	
	
	struct ibv_comp_channel* comp_channel = ibv_create_comp_channel(rdma_context);
	struct ibv_cq* cq = ibv_create_cq(rdma_context, 512, NULL, comp_channel, 0);
	ibv_req_notify_cq(cq, 0);


	
    qp_attr.send_cq = cq;
    qp_attr.recv_cq = cq;
    qp_attr.qp_type = IBV_QPT_RC;
    qp_attr.cap.max_send_wr = 512;
    qp_attr.cap.max_recv_wr = 512;
    qp_attr.cap.max_send_sge = 1;
    qp_attr.cap.max_recv_sge = 1;
	rc = rdma_create_qp(id, pd, &qp_attr);
	if (rc)
	{
		perror("create_qp failed, errno:%d");
        return rc;
	}
	

	memset(&cm_params, 0, sizeof(cm_params));

	cm_params.private_data = NULL;
	cm_params.private_data_len = 0;
	cm_params.responder_resources = (uint8_t)8;
	cm_params.initiator_depth = (uint8_t)8;
	cm_params.retry_count = 7;
	cm_params.rnr_retry_count = 7;
	rc = rdma_connect(id, &cm_params);
	if(rc)
	{
		perror("rdma_connect failed, errno");
		return rc;
	}
	return 0;
}




int main(int argc, char** argv)
{
	
	int port=10121;
	
	
	int rc = 0;
	if(argc != 2){
		fprintf(stderr, "Usage: rdma-client <server_ip>\n");
		return 1;
	}
	char* ip_str=argv[1];
	struct addrinfo* addr;
	rc = getaddrinfo(ip_str, NULL, NULL, &addr);
	if(rc)
	{
		perror("getaddrinfo failed, ");
		return rc;
	}
	((struct sockaddr_in*)addr->ai_addr)->sin_port = htons((uint16_t)port);
	printf("connecting to server ip:%s...\n", ip_str);
	
	ec = rdma_create_event_channel();
	if(ec == NULL)
	{
		perror("rdma_create_event_channel failed, errno:");
		return errno;
	}
	rc = rdma_create_id(ec, &cm_id, NULL, RDMA_PS_TCP);
	if(rc)
	{
		perror("rdma_create_id failed, errno");
		return rc;
	}
	
	rc = rdma_resolve_addr(cm_id, NULL, addr->ai_addr, 500);
	if(rc)
	{
		perror("rdma_resolve_addr failed, errno:");
		return rc;
	}
	
	printf("Begin process cm event...\n");
	struct rdma_cm_event *event = NULL;
	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);
		switch(event_copy.event)
		{
			case RDMA_CM_EVENT_ADDR_RESOLVED:
				printf("get event RDMA_CM_EVENT_ADDR_RESOLVED\n");
				rc = rdma_resolve_route(event_copy.id, 2000);
				if(rc)
				{
					perror("rdma_resolve_route failed, errno:");
					return rc;
				}
				break;
			case RDMA_CM_EVENT_ROUTE_RESOLVED:
				printf("get event RDMA_CM_EVENT_ROUTE_RESOLVED\n");
				on_route_resolved(event_copy.id);
				break;
			case RDMA_CM_EVENT_ESTABLISHED:
				printf("get event RDMA_CM_EVENT_ESTABLISHED\n");
				{
					struct ibv_send_wr wr, *bad_wr = NULL;
					struct ibv_sge sge;


					memset(&wr, 0, sizeof(wr));
					wr.wr_id = (uint64_t)1;
					wr.next = NULL;
					wr.sg_list = &sge;
					wr.num_sge = 1;
					wr.opcode = IBV_WR_SEND;
					wr.send_flags = IBV_SEND_SIGNALED;
					strcpy((char*)recv_buf, "HelloWorld"); 
					sge.addr = (uint64_t)recv_buf;
					sge.length = strlen((char*)recv_buf);
					sge.lkey = recv_mr->lkey;
					printf("send data to server\n");
					rc = ibv_post_send(cm_id->qp, &wr, &bad_wr);
					if (rc)
					{
						fprintf(stderr, "ibv_post_send failed, rc:%d\n", rc);
						return rc;
					}
					return 0;
				}
				break;
			case RDMA_CM_EVENT_DISCONNECTED:
				printf("get event RDMA_CM_EVENT_DISCONNECTED\n");
				
				break;
			default:
				break;
		}
	}
	freeaddrinfo(addr);
	return 0;
}

通过这两篇文章，读者应该可以快速的建立一个自己的RDMA通信程序了。在上面的例子里面，server端使用了IBV_RECEIVE操作，client端使用了IBV_SEND操作。这个意义上还是和socket编程很像的。虽然这样操作也避免了类似TCP协议栈在CPU上的开销，但这远不是RDMA的精髓。RDMA总共提供了4个操作，分成两类：

双边API (two-sided), 这类API的操作需要server, client双方配合，需要一边先post_recv，然后另外一边post_send，两个操作完成一次传输。这类操作就是我们例子里用到的
- IBV_SEND
- IBV_RECEIVE
单边API (one-sided)，这类API只需要server或者client一边调用API，就可以从另一边的内存里面读写数据。包括
- IBV_READ
- IBV_WRITE

要想用好RDMA，就要将这两类API配合使用好。往往需要结合业务设计出最合适的协议模型。也欢迎大家到我的开源存储项目PureFlash ，这是一个RDMA在实际项目应用的具体例子。