RDMA通过kernel-bypass和协议栈offload两大核心技术，实现了远高于传统TCP/IP的网络通信性能。尽管RDMA的性能要远好于TCP/IP，但目前RDMA的实际落地业务场景却寥寥无几，这其中制约RDMA技术大规模上线应用的主要原因有两点：
- 主流互联网公司普遍选择RoCE（RDMA over Converged Ethernet）作为RDMA部署方案，而RoCE本质上是RDMA over UDP，在网络上无法保证不丢包。因此RoCE部署方案需要额外的拥塞控制机制来保证底层的无损网络，如PFC、ECN等，这给大规模的上线部署带来挑战。而且目前各大厂商对硬件拥塞控制的支持均还不完善，存在兼容性问题。
- RDMA提供了完全不同于socket的编程接口，因此要想使用RDMA，需要对现有应用进行改造。而RDMA原生编程API（verbs/RDMA_CM）比较复杂，需要对RDMA技术有深入理解才能做好开发，学习成本较高。
为了降低应用程序的改造成本，决定研发一个RDMA通信库，该通信库直接基于ibvebrs和RDMA_CM，避免对其他第三方库的调用。
本文主要对rdma编程的事件通知机制进行归纳总结。
传统socket编程中通常采用IO复用技术（select、poll、epoll等）来实现事件通知机制，那么对于rdma是否可以同样基于IO复用技术来实现事件通知机制？答案是完全可以。
1. RDMA_CM API（For Connection）
在rdma编程时，可以直接通过RDMA_CM API来建立RDMA连接。
对rdma_create_id函数进行分析，其主要创建了rdma_cm_id对象，并将其注册到驱动中。
int rdma_create_id(struct rdma_event_channel *channel,
struct rdma_cm_id **id, void *context,
enum rdma_port_space ps)
{
enum ibv_qp_type qp_type = (ps == RDMA_PS_IPOIB || ps == RDMA_PS_UDP) ?
IBV_QPT_UD : IBV_QPT_RC;
ret = ucma_init(); //查询获取所有IB设备，存放在cma_dev_array全局数组中；检测是否支持AF_IB协议
struct cma_id_private *id_priv =
ucma_alloc_id(channel, context, ps, qp_type); //创建并初始化id_priv对象：若未创建rdma_event_channel，那么调用rdma_create_event_channel创建一个。
CMA_INIT_CMD_RESP(&cmd, sizeof cmd, CREATE_ID, &resp, sizeof resp);
cmd.uid = (uintptr_t) id_priv;
cmd.ps = ps;
cmd.qp_type = qp_type;
ret = write(id_priv->id.channel->fd, &cmd, sizeof cmd); //将id_priv相关信息注册到内核驱动中，不做过多分析
*id = &id_priv->id; //返回rdma_cm_id对象
}
rdma_cm_id数据结构定义如下：
struct rdma_cm_id {
struct ibv_context *verbs; //ibv_open_device
struct rdma_event_channel *channel; //rdma_create_event_channel创建；For Setup connection
void *context; //user specified context
struct ibv_qp *qp; //rdma_create_qp，底层调用的是ibv_create_qp
struct rdma_route route;
enum rdma_port_space ps; //RDMA_PS_IPOIB or RDMA_PS_UDP or RDMA_PS_TCP
uint8_t port_num; //port数目
struct rdma_cm_event *event; //rdma_cm相关的事件events
struct ibv_comp_channel *send_cq_channel; //ibv_create_comp_channel创建；For data transfer
struct ibv_cq *send_cq; //发送CQ，通常和recv_cq是同一个CQ
struct ibv_comp_channel *recv_cq_channel; //ibv_create_comp_channel创建；For data transfer
struct ibv_cq *recv_cq; //接收CQ，通常和send_cq是同一个CQ
struct ibv_srq *srq;
struct ibv_pd *pd; //ibv_open_device
enum ibv_qp_type qp_type; //IBV_QPT_RC or IBV_QPT_UD
};
在创建rdma_cm_id时，如果预先没有创建rdma_event_channel，那么需要调用rdma_create_event_channel函数。
struct rdma_event_channel *rdma_create_event_channel(void)
{
struct rdma_event_channel *channel;
if (ucma_init()) //通过static局部变量，保证只做一次初始化
return NULL;
channel = malloc(sizeof *channel); //创建rdma_event_channel
if (!channel)
return NULL;
channel->fd = open("/dev/infiniband/rdma_cm", O_RDWR | O_CLOEXEC); //可以看出rdma_event_channel本质上就是一个fd
if (channel->fd < 0) {
goto err;
}
return channel;
err:
free(channel);
return NULL;
}
rdma_event_channel的定义如下：
struct rdma_event_channel {
int fd;
}
1.1 RDMA_CM原生事件通知实现(in block way)
static int cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event);
ret = rdma_get_cm_event(channel, &event); //阻塞操作，直到有rdma_cm event发生才返回
if (!ret) {
ret = cma_handler(event->id, event); //处理事件
rdma_ack_cm_event(event); //ack event
}
static int cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event) {
int ret = 0;
switch (event->event)
{
case RDMA_CM_EVENT_ADDR_RESOLVED:
ret = addr_handler(cma_id->context);
break;
case RDMA_CM_EVENT_MULTICAST_JOIN:
ret = join_handler(cma_id->context, &event->param.ud);
break;
case RDMA_CM_EVENT_ADDR_ERROR:
case RDMA_CM_EVENT_ROUTE_ERROR:
case RDMA_CM_EVENT_MULTICAST_ERROR:
printf("mckey: event: %s, error: %d\n", rdma_event_str(event->event), event->status); connect_error();
ret = event->status;
break;
case RDMA_CM_EVENT_DEVICE_REMOVAL:
/* Cleanup will occur after test completes. */
break;
default:
break;
}
可以看出，RDMA_CM的fd所侦测的都是建立连接相关的event，其不涉及数据传输相关的event，所以rdma_cm event只用于通知建连相关事件
enum rdma_cm_event_type {
RDMA_CM_EVENT_ADDR_RESOLVED,
RDMA_CM_EVENT_ADDR_ERROR,
RDMA_CM_EVENT_ROUTE_RESOLVED,
RDMA_CM_EVENT_ROUTE_ERROR,
RDMA_CM_EVENT_CONNECT_REQUEST,
RDMA_CM_EVENT_CONNECT_RESPONSE,
RDMA_CM_EVENT_CONNECT_ERROR,
RDMA_CM_EVENT_UNREACHABLE,
RDMA_CM_EVENT_REJECTED,
RDMA_CM_EVENT_ESTABLISHED,
RDMA_CM_EVENT_DISCONNECTED,
RDMA_CM_EVENT_DEVICE_REMOVAL,
RDMA_CM_EVENT_MULTICAST_JOIN,
RDMA_CM_EVENT_MULTICAST_ERROR,
RDMA_CM_EVENT_ADDR_CHANGE,
RDMA_CM_EVENT_TIMEWAIT_EXIT
};
1.2 IO复用poll/epoll（in non-block way）
rdma_cm fd不同于传统socket fd，其只会向上抛POLLIN事件，表示有rdma_cm event事件发生，具体event类型需要通过rdma_get_cm_event来获取。
/* change the blocking mode of the completion channel */
flags = fcntl(cm_id->channel->fd, F_GETFL);
rc = fcntl(cm_id->channel->fd, F_SETFL, flags | O_NONBLOCK); //设置rdma_cm fd为NONBLOCK
if (rc < 0) {
fprintf(stderr, "Failed to change file descriptor of Completion Event Channel\n");
return -1;
}
struct pollfd my_pollfd;
int ms_timeout = 10;
/*
* poll the channel until it has an event and sleep ms_timeout
* milliseconds between any iteration
*/
my_pollfd.fd = cm_id->channel->fd;
my_pollfd.events = POLLIN; //只需要监听POLLIN事件，POLLIN事件意味着有rdma_cm event发生
my_pollfd.revents = 0;
do {
rc = poll(&my_pollfd, 1, ms_timeout); //非阻塞操作，有事件或者超时时返回
} while (rc == 0);
/* 注意：poll监听到有事件发生，只意味着有rdma_cm event事件发生，但具体event仍然需要通过rdma_get_cm_event来获取。*/
ret = rdma_get_cm_event(channel, &event);
if (!ret) {
ret = cma_handler(event->id, event); //处理收到的事件
rdma_ack_cm_event(event); //ack event
}
2. verbs API（For data transfer）
从上一节可以看出，RDMA_CM中的fd只涉及建连相关的事件，其无法获取数据传输相关的事件。
对于RDMA传输，数据传输是由NIC硬件完成的，完全不需要CPU参与。网卡硬件完成数据传输后，会向CQ（completion queue中）提交一个cqe，用于描述数据传输完成情况。
struct ibv_cq *ibv_create_cq(struct ibv_context *context, int cqe,
void *cq_context, struct ibv_comp_channel *channel, int comp_vector)
# 作用：创建CQ，每个QP都有对应的send cq和recv cq。
# 一个CQ可以被同一个QP的send queue和recv queue共享，也可以被多个不同的QP共享
# 注意：CQ仅仅只是一个queue，其本身没有built-in的事件通知机制。如果想要增加事件通知机制，那么需要指定channel对象。
verbs API提供了创建ibv_comp_channel的编程接口:
struct ibv_comp_channel *ibv_create_comp_channel(struct ibv_context *context)
# 作用：创建completion channel，用于向user通知有新的completion queue event（cqe）已经被写入CQ中。
struct ibv_comp_channel {
struct ibv_context *context;
int fd;
int refcnt;
};$
2.1 Verbs原生事件通知实现（in block way）
struct ibv_context *context;
struct ibv_cq *cq;
void *ev_ctx = NULL; /* can be initialized with other values for the CQ context */
/* Create a CQ, which is associated with a Completion Event Channel */
cq = ibv_create_cq(ctx, 1, ev_ctx, channel, 0);
if (!cq) {
fprintf(stderr, "Failed to create CQ\n");
return -1;
}
/* Request notification before any completion can be created (to prevent races) */
ret = ibv_req_notify_cq(cq, 0);
if (ret) {
fprintf(stderr, "Couldn't request CQ notification\n");
return -1;
}
/* The following code will be called each time you need to read a Work Completion */
struct ibv_cq *ev_cq;
void *ev_ctx;
int ret;
int ne;
/* Wait for the Completion event */
ret = ibv_get_cq_event(channel, &ev_cq, &ev_ctx); //阻塞函数，直到有cqe发生才返回，ev_cq指向发生cqe的CQ
if (ret) {
fprintf(stderr, "Failed to get CQ event\n");
return -1;
}
/* Ack the event */
ibv_ack_cq_events(ev_cq, 1);
/* Request notification upon the next completion event */
ret = ibv_req_notify_cq(ev_cq, 0);
if (ret) {
fprintf(stderr, "Couldn't request CQ notification\n");
return -1;
}
/* Empty the CQ: poll all of the completions from the CQ (if any exist) */
do {
ne = ibv_poll_cq(cq, 1, &wc);
if (ne < 0) {
fprintf(stderr, "Failed to poll completions from the CQ: ret = %d\n",
ne);
return -1;
}
/* there may be an extra event with no completion in the CQ */
if (ne == 0)
continue;
if (wc.status != IBV_WC_SUCCESS) {
fprintf(stderr, "Completion with status 0x%x was found\n",
wc.status);
return -1;
}
} while (ne);
2.2 IO复用poll/epoll（in non-block way）
利用fcntl设置channel->fd的属性为non-block，然后就可以用poll/epoll/select等来监听channel->fd的POLLIN事件，POLLIN事件意味着有新的completion queue event被填入CQ中。user程序在被唤醒后，无需像传统socket那样进行read/write操作(因为data已经直接DMA到用户态缓存中)，而是需要做poll_cq操作，对每一个cqe进行解析处理。
struct ibv_context *context;
struct ibv_cq *cq;
void *ev_ctx = NULL; /* can be initialized with other values for the CQ context */
/* Create a CQ, which is associated with a Completion Event Channel */
cq = ibv_create_cq(ctx, 1, ev_ctx, channel, 0);
if (!cq) {
fprintf(stderr, "Failed to create CQ\n");
return -1;
}
/* Request notification before any completion can be created (to prevent races) */
ret = ibv_req_notify_cq(cq, 0);
if (ret) {
fprintf(stderr, "Couldn't request CQ notification\n");
return -1;
}
/* The following code will be called only once, after the Completion Event Channel
was created，to change the blocking mode of the completion channel */
int flags = fcntl(channel->fd, F_GETFL);
rc = fcntl(channel->fd, F_SETFL, flags | O_NONBLOCK);
if (rc < 0) {
fprintf(stderr, "Failed to change file descriptor of Completion Event Channel\n");
return -1;
}
/* The following code will be called each time you need to read a Work Completion */
struct pollfd my_pollfd;
struct ibv_cq *ev_cq;
void *ev_ctx;
int ne;
int ms_timeout = 10;
/*
* poll the channel until it has an event and sleep ms_timeout
* milliseconds between any iteration
*/
my_pollfd.fd = channel->fd;
my_pollfd.events = POLLIN; //只需要监听POLLIN事件，POLLIN事件意味着有新的cqe发生
my_pollfd.revents = 0;
do {
rc = poll(&my_pollfd, 1, ms_timeout); //非阻塞函数，有cqe事件或超时时退出
} while (rc == 0);
if (rc < 0) {
fprintf(stderr, "poll failed\n");
return -1;
}
ev_cq = cq;
/* Wait for the completion event */
ret = ibv_get_cq_event(channel, &ev_cq, &ev_ctx); //获取completion queue event。对于epoll水平触发模式，必须要执行ibv_get_cq_event并将该cqe取出，否则会不断重复唤醒epoll
if (ret) {
fprintf(stderr, "Failed to get cq_event\n");
return -1;
}
/* Ack the event */
ibv_ack_cq_events(ev_cq, 1); //ack cqe
/* Request notification upon the next completion event */
ret = ibv_req_notify_cq(ev_cq, 0);
if (ret) {
fprintf(stderr, "Couldn't request CQ notification\n");
return -1;
}
/* Empty the CQ: poll all of the completions from the CQ (if any exist) */
do {
ne = ibv_poll_cq(cq, 1, &wc);
if (ne < 0) {
fprintf(stderr, "Failed to poll completions from the CQ: ret = %d\n",
ne);
return -1;
}
/* there may be an extra event with no completion in the CQ */
if (ne == 0)
continue;
if (wc.status != IBV_WC_SUCCESS) {
fprintf(stderr, "Completion with status 0x%x was found\n",
wc.status);
return -1;
}
} while (ne);
3. rpoll实现（rsocket）
rsocket是附在rdma_cm库中的一个子模块，提供了完全类似于socket接口的rdma调用。此处主要对rpoll的实现进行分析。
rpoll同时支持对rdma fd和正常socket fd进行监听，但对于rdma fd，其目前仅支持四种事件：POLLIN、POLLOUT、POLLHUP、POLLERR。
* Note that we may receive events on an rsocket that may not be reported
* to the user (e.g. connection events or credit updates). Process those
* events, then return to polling until we find ones of interest.
*/
int rpoll(struct pollfd *fds, nfds_t nfds, int timeout)
{
struct timeval s, e;
struct pollfd *rfds;
uint32_t poll_time = 0;
int ret;
do {
ret = rs_poll_check(fds, nfds); //主动轮询查看是否有event发生
if (ret || !timeout) //如果有event发生或者timeout为0，直接返回
return ret;
if (!poll_time)
gettimeofday(&s, NULL);
gettimeofday(&e, NULL);
poll_time = (e.tv_sec - s.tv_sec) * 1000000 +
(e.tv_usec - s.tv_usec) + 1;
} while (poll_time <= polling_time); //尝试轮询polling_time时间，该时间内如果有event发生，那么直接返回，否则进入后续逻辑
rfds = rs_fds_alloc(nfds); //创建新的pollfd数组rfds，用于添加到原生poll中。
if (!rfds)
return ERR(ENOMEM);
do {
ret = rs_poll_arm(rfds, fds, nfds); //对所有verbs fd进行arm操作，并将待监听事件全部改为POLLIN
if (ret)
break;
ret = poll(rfds, nfds, timeout); //调用OS原生poll
if (ret <= 0)
break;
ret = rs_poll_events(rfds, fds, nfds); //将cqe或rdma_cm event转化为具体event
} while (!ret);
rpoll中调用rs_poll_check进行轮询，查看是否有event发生。
static int rs_poll_check(struct pollfd *fds, nfds_t nfds)
{
struct rsocket *rs;
int i, cnt = 0;
for (i = 0; i < nfds; i++) {
rs = idm_lookup(&idm, fds[i].fd); //根据fd找到对应的rsocket对象
if (rs)
fds[i].revents = rs_poll_rs(rs, fds[i].events, 1, rs_poll_all);
//查看rsocket fd是否有event发生，手动向上抛事件
else
poll(&fds[i], 1, 0); //普通fd，非阻塞poll一次，查询是否有event发生
if (fds[i].revents)
cnt++;
}
return cnt;
}
static int rs_poll_rs(struct rsocket *rs, int events,
int nonblock, int (*test)(struct rsocket *rs))
{
struct pollfd fds;
short revents;
int ret;
check_cq:
if ((rs->type == SOCK_STREAM) && ((rs->state & rs_connected) ||
(rs->state == rs_disconnected) || (rs->state & rs_error))) {
rs_process_cq(rs, nonblock, test); //调用ibv_poll_cq遍历cqe
//对于send cqe，可以在处理函数中将发送缓存重新放回到内存池中，
//对于recv cqe，可以在处理函数中更新可读数据length和addr等
revents = 0;
if ((events & POLLIN) && rs_conn_have_rdata(rs)) //接收缓存有数据，抛POLLIN
事件
revents |= POLLIN;
if ((events & POLLOUT) && rs_can_send(rs)) //发送缓存可写，抛POLLOUT事件
revents |= POLLOUT;
if (!(rs->state & rs_connected)) {
if (rs->state == rs_disconnected)
revents |= POLLHUP; //断开连接，抛POLLHUP事件
else
revents |= POLLERR; //抛POLLERR事件
}
return revents;
} else if (rs->type == SOCK_DGRAM) { //UDP相关逻辑，不关注
ds_process_cqs(rs, nonblock, test);
revents = 0;
if ((events & POLLIN) && rs_have_rdata(rs))
revents |= POLLIN;
if ((events & POLLOUT) && ds_can_send(rs))
revents |= POLLOUT;
return revents;
}
if (rs->state == rs_listening) { //rmda_cm fd
fds.fd = rs->cm_id->channel->fd;
fds.events = events; //此处没有将要监听的事件设置为POLLIN，why？
fds.revents = 0;
poll(&fds, 1, 0); //直接poll一次，然后返回
return fds.revents;
}
if (rs->state & rs_opening) {
ret = rs_do_connect(rs);
if (ret && (errno == EINPROGRESS)) {
errno = 0;
} else {
goto check_cq;
}
}
if (rs->state == rs_connect_error) {
revents = 0;
if (events & POLLOUT)
revents |= POLLOUT;
if (events & POLLIN)
revents |= POLLIN;
revents |= POLLERR;
return revents;
}
return 0;
}
当主动轮询polling_time时间后，如果仍然没有event发生，且尚未超时，那么就需要调用rs_poll_arm函数，其主要作用有两点：1）对所有verbs fd进行arm操作(ibv_notify_cq_event)；2）将所有rdma相关事件全部修改为监听POLLIN事件，然后丢给原生poll函数去监听。
static int rs_poll_arm(struct pollfd *rfds, struct pollfd *fds, nfds_t nfds)
{
struct rsocket *rs;
int i;
for (i = 0; i < nfds; i++) {
rs = idm_lookup(&idm, fds[i].fd);
if (rs) { // rdma相关fd
fds[i].revents = rs_poll_rs(rs, fds[i].events, 0, rs_is_cq_armed);
if (fds[i].revents)
return 1;
if (rs->type == SOCK_STREAM) {
if (rs->state >= rs_connected)
rfds[i].fd = rs->cm_id->recv_cq_channel->fd; //verbs fd，用于通知data传输event
else
rfds[i].fd = rs->cm_id->channel->fd; //rdma_cm fd，用于通知connect event
} else {
rfds[i].fd = rs->epfd;
}
rfds[i].events = POLLIN; //所有监听事件全部改为POLLIN
} else { //普通fd
rfds[i].fd = fds[i].fd;
rfds[i].events = fds[i].events;
}
rfds[i].revents = 0;
}
return 0;
}
原生poll在超时时间内如果监听到有事件发生，那么调用rs_poll_events函数。
static int rs_poll_events(struct pollfd *rfds, struct pollfd *fds, nfds_t nfds)
{
struct rsocket *rs;
int i, cnt = 0;
for (i = 0; i < nfds; i++) {
if (!rfds[i].revents) //没有事件发生，跳过
continue;
rs = idm_lookup(&idm, fds[i].fd);
if (rs) {
fastlock_acquire(&rs->cq_wait_lock);
if (rs->type == SOCK_STREAM)
rs_get_cq_event(rs); //调用ibv_get_cq_event
else
ds_get_cq_event(rs);
fastlock_release(&rs->cq_wait_lock);
fds[i].revents = rs_poll_rs(rs, fds[i].events, 1, rs_poll_all); //手动向上抛事件
} else {
fds[i].revents = rfds[i].revents; //普通fd，直接向上抛事件
}
if (fds[i].revents)
cnt++;
}
return cnt;
}
总结来看，对于rpoll实现，主要分两个步骤:
- 主动遍历轮询polling_time时间，查看是否有event发生；
- 如果polling_time时间内没有event发生，那么将verbs/rdma_cm fd直接注册到OS原生poll中，并将待监听事件改为POLLIN，然后调用原生poll。如果poll监听到verbs/rdma_cm fd的事件，这只意味着有cqe事件或rdma_cm事件发生，不能直接返回给用户，需要额外进行逻辑判断，以确定究竟是否要向上抛事件，以及抛什么事件。
4. 总结
对于rdma编程，目前主流实现是利用rdma_cm来建立连接，然后利用verbs来传输数据。
rdma_cm和ibverbs分别会创建一个fd，这两个fd的分工不同。rdma_cm fd主要用于通知建连相关的事件，verbs fd则主要通知有新的cqe发生。当直接对rdma_cm fd进行poll/epoll监听时，此时只能监听到POLLIN事件，这意味着有rdma_cm事件发生。当直接对verbs fd进行poll/epoll监听时，同样只能监听到POLLIN事件，这意味着有新的cqe。
)
