Rdma Verbs

RDMA stack
├── Application (libibverbs)
├── Kernel RDMA driver (mlx5, rdma_rxe)
├── NIC/HCA hardware
└── Fabric (InfiniBand or RoCE/Ethernet)

Operation types
├── Two-sided: Send/Recv (both sides participate)
└── One-sided: RDMA Read/Write (remote CPU not involved)

# List RDMA devices
ibv_devices
ibv_devinfo

# RoCE link status
rdma link show
ibstat

# Perftest prerequisites
modprobe ib_umad

#include <infiniband/verbs.h>
#include <stdio.h>
#include <stdlib.h>

int main(void) {
    int num_devices;
    struct ibv_device **dev_list = ibv_get_device_list(&num_devices);
    if (!dev_list || num_devices == 0) {
        fprintf(stderr, "No RDMA devices\n");
        return 1;
    }

    struct ibv_context *ctx = ibv_open_device(dev_list[0]);
    struct ibv_pd *pd = ibv_alloc_pd(ctx);

    char buf[4096];
    struct ibv_mr *mr = ibv_reg_mr(pd, buf, sizeof(buf),
        IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE);

    struct ibv_cq *cq = ibv_create_cq(ctx, 10, NULL, NULL, 0);

    struct ibv_qp_init_attr qp_attr = {
        .send_cq = cq,
        .recv_cq = cq,
        .cap = { .max_send_wr = 10, .max_recv_wr = 10,
                 .max_send_sge = 1, .max_recv_sge = 1 },
        .qp_type = IBV_QPT_RC,
    };
    struct ibv_qp *qp = ibv_create_qp(pd, &qp_attr);

    printf("QP num %u, MR lkey %u rkey %u\n",
           qp->qp_num, mr->lkey, mr->rkey);

    ibv_destroy_qp(qp);
    ibv_dereg_mr(mr);
    ibv_destroy_cq(cq);
    ibv_dealloc_pd(pd);
    ibv_close_device(ctx);
    ibv_free_device_list(dev_list);
    return 0;
}

gcc -o rdma_setup rdma_setup.c -libverbs

// Simplified: exchange via TCP socket before RDMA
struct qp_info {
    uint16_t lid;
    uint32_t qpn;
    uint32_t psn;
    union ibv_gid gid;
};

// Modify QP to INIT → RTR → RTS states
struct ibv_qp_attr attr = { .qp_state = IBV_QPS_INIT, ... };
ibv_modify_qp(qp, &attr, IBV_QP_STATE | IBV_QP_PKEY_INDEX | ...);
// RTR: set path_mtu, dest_qp_num, rq_psn, ah_attr
// RTS: set sq_psn, timeout, retry_cnt

#include <rdma/rdma_cma.h>
// rdma_create_event_channel, rdma_connect, rdma_accept

// Post receive
struct ibv_recv_wr recv_wr = {}, *bad_wr;
struct ibv_sge recv_sge = { .addr = (uint64_t)recv_buf, .length = 4096, .lkey = mr->lkey };
recv_wr.wr_id = 1;
recv_wr.sg_list = &recv_sge;
recv_wr.num_sge = 1;
ibv_post_recv(qp, &recv_wr, &bad_wr);

// Post send
struct ibv_send_wr send_wr = {}, *bad_send;
struct ibv_sge send_sge = { .addr = (uint64_t)send_buf, .length = msg_len, .lkey = mr->lkey };
send_wr.wr_id = 2;
send_wr.opcode = IBV_WR_SEND;
send_wr.sg_list = &send_sge;
send_wr.num_sge = 1;
ibv_post_send(qp, &send_wr, &bad_send);

// Poll completion
struct ibv_wc wc;
while (ibv_poll_cq(cq, 1, &wc) == 0);
if (wc.status != IBV_WC_SUCCESS)
    fprintf(stderr, "WC error: %s\n", ibv_wc_status_str(wc.status));

struct ibv_send_wr wr = {}, *bad;
struct ibv_sge sge = { .addr = (uint64_t)local_buf, .length = len, .lkey = local_mr->lkey };

wr.wr_id = 3;
wr.opcode = IBV_WR_RDMA_WRITE;
wr.send_flags = IBV_SEND_SIGNALED;
wr.sg_list = &sge;
wr.num_sge = 1;
wr.wr.rdma.remote_addr = remote_addr;  // from peer exchange
wr.wr.rdma.rkey = remote_rkey;

ibv_post_send(qp, &wr, &bad);

# RoCEv2 GID
cat /sys/class/infiniband/mlx5_0/ports/1/gids/3

# Server
ib_send_bw -d mlx5_0 -x 3

# Client
ib_send_bw -d mlx5_0 -x 3 <server_ip>

# Latency
ib_send_lat -d mlx5_0 <server_ip>

# RDMA write bandwidth
ib_write_bw -d mlx5_0 <server_ip>

# Cargo.toml
rdma-sys = "0.1"