int
main()
{
int ib_port = 1;
int gid_idx = 1;
int rc;
int rank, nprocs;
struct ibv_sge sge_list;
struct ibv_wc wc;
struct ibv_send_wr *sr;
unsigned long long start, end;
float time;
mypmiInit(&rank, &nprocs);
fprintf(stderr, "[%d] nprocs(%d)\n", rank, nprocs);
rc = resource_create(&res, ib_port, rank);
gid_idx = rank;
rc = connect_qp(&res, ib_port, gid_idx, rank);
create_sge(&res, buf, SIZE, &sge_list);
memset(&wc, 0, sizeof(struct ibv_wc));
sr = malloc(sizeof(*sr));
memset(sr, 0, sizeof(*sr));
mypmiBarrier();
fprintf(stderr, "[%d] START\n", rank);
memset(buf, 0, SIZE);
mypmiBarrier();
if (rank == 0) {
struct ibv_mr *mr;
for (int size = RDMA_MIN_SIZE; size < RDMA_MAX_SIZE; size += STEP) {
char *received = calloc(size, sizeof(char));
mr = ibv_reg_mr(res.pd, received, size, IBV_ACCESS_REMOTE_WRITE | IBV_ACCESS_LOCAL_WRITE);
INT_TO_BE(buf, mr->rkey);
INT_TO_BE(buf + 4, (((intptr_t)mr->addr) >> 32));
INT_TO_BE(buf + 8, (((intptr_t)mr->addr) & 0xffffffff));
if (post_ibsend(&res, IBV_WR_SEND, &sge_list, sr, 1)) {
fprintf(stderr, "[%d] failed to post SR\n", rank);
goto end;
}
while ((rc = poll_cq(&res, &wc, 1, SCQ_FLG)) == 0) {
}
/* printf("[%d] memory region is sent. key(%x) addr(%lx) rc(%d)\n", rank, mr->rkey, (intptr_t)mr->addr, rc); */
/* wait for done */
post_ibreceive(&res, &sge_list, 1);
while (poll_cq(&res, &wc, 1, RCQ_FLG) == 0) {
}
/* printf("[%d] %d byte has received (opcode=%d)\n", rank, wc.byte_len, wc.opcode); */
/* printf("[%d] Received message: %s\n", rank, buf); */
/* display_received(received, size); */
ibv_dereg_mr(mr);
free(received);
}
} else {
void client_write_once(uint32_t len){
struct timeval start, end, dt;
gettimeofday(&start, NULL);
struct connection *conn = (struct connection *)s_ctx->id->context;
post_receive(conn);
uint32_t size = len|(1UL<<31);
write_remote(conn,size);
poll_cq(NULL);//wait for write completion
poll_cq(NULL);//wait for recv completion
gettimeofday(&end, NULL);
timersub(&end, &start, &dt);
long usec = dt.tv_usec + 1000000 * dt.tv_sec;
printf("[Wriet] takes %ld micro_secs.\n", usec);
}
void openib_waitall()
{
// Poll the CQ for all outstanding data transfer
int rc = poll_cq(l_state.num_outstanding, 0);
assert(!rc);
l_state.num_outstanding = 0;
}
void client_test(char *ip, char *port) {
struct addrinfo *addr;
struct rdma_cm_event *event = NULL;
struct rdma_cm_id *conn= NULL;
struct rdma_event_channel *ec = NULL;
TEST_NZ(getaddrinfo(ip, port, 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)) {
s_ctx->ec = ec;
s_ctx->id = conn;
on_connection(event_copy.id);//send our memory information to server using post_send
//printf("wait for msg_send_completion\n");
poll_cq(NULL);//wait for send_completion
//printf("wait for msg_recv_completion\n");
poll_cq(NULL);//wait for recv_completion
break;
}
}
return;
};
/*
* Get one cq entry from c4iw and map it to openib.
*
* Returns:
* 0 cqe returned
* -ENODATA EMPTY;
* -EAGAIN caller must try again
* any other -errno fatal error
*/
static int c4iw_poll_cq_one(struct c4iw_cq *chp, struct ib_wc *wc)
{
struct c4iw_qp *qhp = NULL;
struct t4_cqe cqe = {0, 0}, *rd_cqe;
struct t4_wq *wq;
u32 credit = 0;
u8 cqe_flushed;
u64 cookie = 0;
int ret;
ret = t4_next_cqe(&chp->cq, &rd_cqe);
if (ret)
return ret;
qhp = get_qhp(chp->rhp, CQE_QPID(rd_cqe));
if (!qhp)
wq = NULL;
else {
spin_lock(&qhp->lock);
wq = &(qhp->wq);
}
ret = poll_cq(wq, &(chp->cq), &cqe, &cqe_flushed, &cookie, &credit);
if (ret)
goto out;
wc->wr_id = cookie;
wc->qp = &qhp->ibqp;
wc->vendor_err = CQE_STATUS(&cqe);
wc->wc_flags = 0;
PDBG("%s qpid 0x%x type %d opcode %d status 0x%x len %u wrid hi 0x%x "
"lo 0x%x cookie 0x%llx\n", __func__, CQE_QPID(&cqe),
CQE_TYPE(&cqe), CQE_OPCODE(&cqe), CQE_STATUS(&cqe), CQE_LEN(&cqe),
CQE_WRID_HI(&cqe), CQE_WRID_LOW(&cqe), (unsigned long long)cookie);
if (CQE_TYPE(&cqe) == 0) {
if (!CQE_STATUS(&cqe))
wc->byte_len = CQE_LEN(&cqe);
else
wc->byte_len = 0;
wc->opcode = IB_WC_RECV;
if (CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_INV ||
CQE_OPCODE(&cqe) == FW_RI_SEND_WITH_SE_INV) {
wc->ex.invalidate_rkey = CQE_WRID_STAG(&cqe);
wc->wc_flags |= IB_WC_WITH_INVALIDATE;
}
} else {
switch (CQE_OPCODE(&cqe)) {
case FW_RI_RDMA_WRITE:
wc->opcode = IB_WC_RDMA_WRITE;
break;
case FW_RI_READ_REQ:
wc->opcode = IB_WC_RDMA_READ;
wc->byte_len = CQE_LEN(&cqe);
break;
case FW_RI_SEND_WITH_INV:
case FW_RI_SEND_WITH_SE_INV:
wc->opcode = IB_WC_SEND;
wc->wc_flags |= IB_WC_WITH_INVALIDATE;
break;
case FW_RI_SEND:
case FW_RI_SEND_WITH_SE:
wc->opcode = IB_WC_SEND;
break;
case FW_RI_BIND_MW:
wc->opcode = IB_WC_BIND_MW;
break;
case FW_RI_LOCAL_INV:
wc->opcode = IB_WC_LOCAL_INV;
break;
case FW_RI_FAST_REGISTER:
wc->opcode = IB_WC_FAST_REG_MR;
break;
default:
printk(KERN_ERR MOD "Unexpected opcode %d "
"in the CQE received for QPID=0x%0x\n",
CQE_OPCODE(&cqe), CQE_QPID(&cqe));
ret = -EINVAL;
goto out;
}
}
if (cqe_flushed)
wc->status = IB_WC_WR_FLUSH_ERR;
else {
switch (CQE_STATUS(&cqe)) {
case T4_ERR_SUCCESS:
wc->status = IB_WC_SUCCESS;
break;
case T4_ERR_STAG:
wc->status = IB_WC_LOC_ACCESS_ERR;
break;
case T4_ERR_PDID:
wc->status = IB_WC_LOC_PROT_ERR;
break;
case T4_ERR_QPID:
case T4_ERR_ACCESS:
wc->status = IB_WC_LOC_ACCESS_ERR;
break;
case T4_ERR_WRAP:
wc->status = IB_WC_GENERAL_ERR;
break;
case T4_ERR_BOUND:
wc->status = IB_WC_LOC_LEN_ERR;
break;
case T4_ERR_INVALIDATE_SHARED_MR:
case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
wc->status = IB_WC_MW_BIND_ERR;
break;
case T4_ERR_CRC:
case T4_ERR_MARKER:
case T4_ERR_PDU_LEN_ERR:
case T4_ERR_OUT_OF_RQE:
case T4_ERR_DDP_VERSION:
case T4_ERR_RDMA_VERSION:
case T4_ERR_DDP_QUEUE_NUM:
case T4_ERR_MSN:
case T4_ERR_TBIT:
case T4_ERR_MO:
case T4_ERR_MSN_RANGE:
case T4_ERR_IRD_OVERFLOW:
case T4_ERR_OPCODE:
case T4_ERR_INTERNAL_ERR:
wc->status = IB_WC_FATAL_ERR;
break;
case T4_ERR_SWFLUSH:
wc->status = IB_WC_WR_FLUSH_ERR;
break;
default:
printk(KERN_ERR MOD
"Unexpected cqe_status 0x%x for QPID=0x%0x\n",
CQE_STATUS(&cqe), CQE_QPID(&cqe));
ret = -EINVAL;
}
}
out:
if (wq)
spin_unlock(&qhp->lock);
return ret;
}
static int __c4iw_poll_cq_one(struct c4iw_cq *chp, struct c4iw_qp *qhp,
struct ib_wc *wc, struct c4iw_srq *srq)
{
struct t4_cqe uninitialized_var(cqe);
struct t4_wq *wq = qhp ? &qhp->wq : NULL;
u32 credit = 0;
u8 cqe_flushed;
u64 cookie = 0;
int ret;
ret = poll_cq(wq, &(chp->cq), &cqe, &cqe_flushed, &cookie, &credit,
srq ? &srq->wq : NULL);
if (ret)
goto out;
wc->wr_id = cookie;
wc->qp = qhp ? &qhp->ibqp : NULL;
wc->vendor_err = CQE_STATUS(&cqe);
wc->wc_flags = 0;
/*
* Simulate a SRQ_LIMIT_REACHED HW notification if required.
*/
if (srq && !(srq->flags & T4_SRQ_LIMIT_SUPPORT) && srq->armed &&
srq->wq.in_use < srq->srq_limit)
c4iw_dispatch_srq_limit_reached_event(srq);
pr_debug("qpid 0x%x type %d opcode %d status 0x%x len %u wrid hi 0x%x lo 0x%x cookie 0x%llx\n",
CQE_QPID(&cqe),
CQE_TYPE(&cqe), CQE_OPCODE(&cqe),
CQE_STATUS(&cqe), CQE_LEN(&cqe),
CQE_WRID_HI(&cqe), CQE_WRID_LOW(&cqe),
(unsigned long long)cookie);
if (CQE_TYPE(&cqe) == 0) {
if (!CQE_STATUS(&cqe))
wc->byte_len = CQE_LEN(&cqe);
else
wc->byte_len = 0;
switch (CQE_OPCODE(&cqe)) {
case FW_RI_SEND:
wc->opcode = IB_WC_RECV;
break;
case FW_RI_SEND_WITH_INV:
case FW_RI_SEND_WITH_SE_INV:
wc->opcode = IB_WC_RECV;
wc->ex.invalidate_rkey = CQE_WRID_STAG(&cqe);
wc->wc_flags |= IB_WC_WITH_INVALIDATE;
c4iw_invalidate_mr(qhp->rhp, wc->ex.invalidate_rkey);
break;
case FW_RI_WRITE_IMMEDIATE:
wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
wc->ex.imm_data = CQE_IMM_DATA(&cqe);
wc->wc_flags |= IB_WC_WITH_IMM;
break;
default:
pr_err("Unexpected opcode %d in the CQE received for QPID=0x%0x\n",
CQE_OPCODE(&cqe), CQE_QPID(&cqe));
ret = -EINVAL;
goto out;
}
} else {
switch (CQE_OPCODE(&cqe)) {
case FW_RI_WRITE_IMMEDIATE:
case FW_RI_RDMA_WRITE:
wc->opcode = IB_WC_RDMA_WRITE;
break;
case FW_RI_READ_REQ:
wc->opcode = IB_WC_RDMA_READ;
wc->byte_len = CQE_LEN(&cqe);
break;
case FW_RI_SEND_WITH_INV:
case FW_RI_SEND_WITH_SE_INV:
wc->opcode = IB_WC_SEND;
wc->wc_flags |= IB_WC_WITH_INVALIDATE;
break;
case FW_RI_SEND:
case FW_RI_SEND_WITH_SE:
wc->opcode = IB_WC_SEND;
break;
case FW_RI_LOCAL_INV:
wc->opcode = IB_WC_LOCAL_INV;
break;
case FW_RI_FAST_REGISTER:
wc->opcode = IB_WC_REG_MR;
/* Invalidate the MR if the fastreg failed */
if (CQE_STATUS(&cqe) != T4_ERR_SUCCESS)
c4iw_invalidate_mr(qhp->rhp,
CQE_WRID_FR_STAG(&cqe));
break;
default:
pr_err("Unexpected opcode %d in the CQE received for QPID=0x%0x\n",
CQE_OPCODE(&cqe), CQE_QPID(&cqe));
ret = -EINVAL;
goto out;
}
}
if (cqe_flushed)
wc->status = IB_WC_WR_FLUSH_ERR;
else {
switch (CQE_STATUS(&cqe)) {
case T4_ERR_SUCCESS:
wc->status = IB_WC_SUCCESS;
break;
case T4_ERR_STAG:
wc->status = IB_WC_LOC_ACCESS_ERR;
break;
case T4_ERR_PDID:
wc->status = IB_WC_LOC_PROT_ERR;
break;
case T4_ERR_QPID:
case T4_ERR_ACCESS:
wc->status = IB_WC_LOC_ACCESS_ERR;
break;
case T4_ERR_WRAP:
wc->status = IB_WC_GENERAL_ERR;
break;
case T4_ERR_BOUND:
wc->status = IB_WC_LOC_LEN_ERR;
break;
case T4_ERR_INVALIDATE_SHARED_MR:
case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
wc->status = IB_WC_MW_BIND_ERR;
break;
case T4_ERR_CRC:
case T4_ERR_MARKER:
case T4_ERR_PDU_LEN_ERR:
case T4_ERR_OUT_OF_RQE:
case T4_ERR_DDP_VERSION:
case T4_ERR_RDMA_VERSION:
case T4_ERR_DDP_QUEUE_NUM:
case T4_ERR_MSN:
case T4_ERR_TBIT:
case T4_ERR_MO:
case T4_ERR_MSN_RANGE:
case T4_ERR_IRD_OVERFLOW:
case T4_ERR_OPCODE:
case T4_ERR_INTERNAL_ERR:
wc->status = IB_WC_FATAL_ERR;
break;
case T4_ERR_SWFLUSH:
wc->status = IB_WC_WR_FLUSH_ERR;
break;
default:
pr_err("Unexpected cqe_status 0x%x for QPID=0x%0x\n",
CQE_STATUS(&cqe), CQE_QPID(&cqe));
wc->status = IB_WC_FATAL_ERR;
}
}
out:
return ret;
}
void client_test(char *ip, char *port) {
struct addrinfo *addr;
struct rdma_cm_event *event = NULL;
struct rdma_cm_id *conn= NULL;
struct rdma_event_channel *ec = NULL;
struct timeval t1, t2, t3, t11, t12;
struct timeval dt, dt1, dt2, dt11, dt12, dt13;
gettimeofday(&t1, NULL);
TEST_NZ(getaddrinfo(ip, port, NULL, &addr));
TEST_Z(ec = rdma_create_event_channel());
gettimeofday(&t11, NULL);
TEST_NZ(rdma_create_id(ec, &conn, NULL, RDMA_PS_TCP));
gettimeofday(&t12, NULL);
TEST_NZ(rdma_resolve_addr(conn, NULL, addr->ai_addr, TIMEOUT_IN_MS));
freeaddrinfo(addr);
gettimeofday(&t2, 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);
if (on_event(&event_copy)) {
s_ctx->ec = ec;
s_ctx->id = conn;
on_connection(event_copy.id);//send our memory information to server using post_send
poll_cq(NULL);//wait for send_completion
poll_cq(NULL);//wait for recv_completion
break;
}
}
gettimeofday(&t3, NULL);
timersub(&t3, &t1, &dt);
timersub(&t3, &t2, &dt2);
timersub(&t2, &t1, &dt1);
timersub(&t2, &t12, &dt13);
timersub(&t12, &t11, &dt12);
timersub(&t11, &t1, &dt11);
long usec = dt.tv_usec + 10000 * dt.tv_sec;
printf("[dt]:\t%ld us.\n", usec);
printf("[dt1]:\t%ld us.\n", dt1.tv_usec+1000000 *dt1.tv_sec);
printf("Including the following steps: \n");
printf("[dt11]:\t%ld us.\n", dt11.tv_usec+1000000 *dt11.tv_sec);
printf("[dt12]:\t%ld us.\n", dt12.tv_usec+1000000 *dt12.tv_sec);
printf("[dt13]:\t%ld us.\n", dt13.tv_usec+1000000 *dt13.tv_sec);
printf("[dt2] takes %ld micro_secs.\n", dt2.tv_usec+1000000*dt2.tv_sec);
printf("[dt]:total time\t[dt1]:pre_setup\t[dt2]:send/recv\t.\n");
printf("[dt11]:create_event_channel\t[dt12]:create_id\t[dt13]:resolve_address.\n");
return;
};
void iterate_interface()
{
post_one_read();
poll_cq(1);
}