目录
前言
- rdma_cm依赖ib_verbs
- ib_verbs和rdma_cm 都是Mellanox公司提供的两个动态链接库。(这两个库的API ，queue pair (QP) 类似于 TCP的sockets)
- rdma_cm，连接管理器库，包含了对ib_verbs的封装或具体化，通过verbs API 对硬件进行访问。
例如：
rdma_post_send = ibv_post_send(qp,wr.opcode=IBV_WR_SEND,bad_wr) ?
rdma_post_read = ibv_post_send(qp,wr.opcode=IBV_WR_RDMA_READ ,bad_wr) ?
rdma_post_write = ibv_post_send(qp,wr.opcode=IBV_WR_RDMA_WRITE,bad_wr) ?
libibverbs和librdmacm的区别？
在infiniband/verbs.h中，定义了ibv_post_send()和ibv_post_recv()操作，分别表示，将wr发布到SQ和RQ中，至于是什么操作(send or write/read)，和wr中的opcode有关。
对ibv_post_send()来说，对应的是struct ibv_send_wr，其中有opcode，表示操作码，有SEND/WRITE/READ等。
对于ibv_post_recv()来说，对应的是struct ibv_recv_wr，没有操作码，因为只有接收一个动作，所以不需要定义其它的操作码。但是发送来说，有三类。
在rdma/rdma_verbs.h中，有rdma_post_send()，rdma_post_recv()，rdma_post_read()，rdma_post_write()。
rdma_post_send()：把wr发布到QP的SQ中，需要mr
rdma_post_recv()：把wr发布到QP的RQ中，需要mr
rdma_post_read()：把wr发布到QP的SQ中，执行RDMA READ操作，需要远程地址和rkey，以及本地存储地址和长度，以及mr
rdma_post_write()：把wr发布到QP的SQ中，RDMA WRITE操作，需要远程的被写入地址和rkey，以及本地要发送数据的地址和长度，以及mr
所以rdma/rdma_verbs.h中的四种通信函数其实和infiniband/verbs.h中的两种方法是一致的。
ibv_post_send()对应rdma_post_send()、rdma_post_read()、rdma_post_write()，ibv_post_recv()对应rdma_post_recv()。
客户端和服务端操作
Rdma Verbs
The rdma_cm supports the full range of verbs available through the libibverbs library and interfaces. However, it also provides wrapper functions for some of the more commonly used verbs funcationality. The full set of abstracted verb calls are:rdma_reg_msgs - register an array of buffers for sending and receiving
rdma_reg_read - registers a buffer for RDMA read operations
rdma_reg_write - registers a buffer for RDMA write operations
rdma_dereg_mr - deregisters a memory region
rdma_post_recv - post a buffer to receive a message
rdma_post_send - post a buffer to send a message
rdma_post_read - post an RDMA to read data into a buffer
rdma_post_write - post an RDMA to send data from a buffer
rdma_post_recvv - post a vector of buffers to receive a message
rdma_post_sendv - post a vector of buffers to send a message
rdma_post_readv - post a vector of buffers to receive an RDMA read
rdma_post_writev - post a vector of buffers to send an RDMA write
rdma_post_ud_send - post a buffer to send a message on a UD QP
rdma_get_send_comp - get completion status for a send or RDMA operation
rdma_get_recv_comp - get information about a completed receive
Client Operation
This section provides a general overview of the basic operation for the active, or client, side of communication. This flow assume asynchronous operation with low level call details shown. For synchronous operation, calls to rdma_create_event_channel, rdma_get_cm_event, rdma_ack_cm_event, and rdma_destroy_event_channel would be eliminated. Abstracted calls, such as rdma_create_ep encapsulate serveral of these calls under a single API. Users may also refer to the example applications for code samples. A general connection flow would be:rdma_getaddrinfo
retrieve address information of the destination
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
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
An almost identical process is used to setup unreliable datagram (UD) communication between nodes. No actual connection is formed between QPs however, so disconnection is not needed.
Although this example shows the client initiating the disconnect, either side of a connection may initiate the disconnect.
Server Operation
This section provides a general overview of the basic operation for the passive, or server, side of communication. A general connection flow would be: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_bind_addr
set the local port number to listen on
rdma_listen
begin listening for connection requests
rdma_get_cm_event
wait for RDMA_CM_EVENT_CONNECT_REQUEST event with a new rdma_cm_id
rdma_create_qp
allocate a QP for the communication on the new rdma_cm_id
rdma_accept
accept the connection request
rdma_ack_cm_event
ack event
rdma_get_cm_event
wait for RDMA_CM_EVENT_ESTABLISHED event
rdma_ack_cm_event
ack event
Perform data transfers over connection
rdma_get_cm_event
wait for RDMA_CM_EVENT_DISCONNECTED event
rdma_ack_cm_event
ack event
rdma_disconnect
tear-down connection
rdma_destroy_qp
destroy the QP
rdma_destroy_id
release the connected rdma_cm_id
rdma_destroy_id
release the listening rdma_cm_id
rdma_destroy_event_channel
release the event channel
Return Codes
= 0
success
= -1
error - see errno for more details
Most librdmacm functions return 0 to indicate success, and a -1 return value to indicate failure. If a function operates asynchronously, a return value of 0 means that the operation was successfully started. The operation could still complete in error; users should check the status of the related event. If the return value is -1, then errno will contain additional information regarding the reason for the failure.
Prior versions of the library would return -errno and not set errno for some cases related to ENOMEM, ENODEV, ENODATA, EINVAL, and EADDRNOTAVAIL codes. Applications that want to check these codes and have compatability with prior library versions must manually set errno to the negative of the return code if it is < -1.
