int MPIDI_nem_ib_packetized_recv_req(MPIDI_VC_t * vc, MPID_Request * rreq)
{
int mpi_errno = MPI_SUCCESS;
if (NULL == VC_FIELD(vc, connection)->packetized_recv) {
VC_FIELD(vc, connection)->packetized_recv = (void *) rreq;
} else {
mpi_errno =
MPIR_Err_create_code(mpi_errno, MPIR_ERR_FATAL,
FCNAME, __LINE__,
MPI_ERR_OTHER,
"**fail", 0);
}
DEBUG_PRINT("Add rreq %p to packetized recv\n", rreq);
return mpi_errno;
}
int
MPID_nem_gm_lmt_start_recv (MPIDI_VC_t *src_vc, struct iovec s_cookie, struct iovec r_cookie, int *completion_ctr)
{
int ret;
struct iovec *s_iov;
struct iovec *r_iov;
int s_n_iov;
int r_n_iov;
int s_offset;
int r_offset;
s_iov = s_cookie.iov_base;
s_n_iov = s_cookie.iov_len / sizeof (struct iovec);
r_iov = r_cookie.iov_base;
r_n_iov = r_cookie.iov_len / sizeof (struct iovec);
r_offset = 0;
s_offset = 0;
ret = MPID_nem_gm_lmt_do_get (VC_FIELD(src_vc, gm_node_id), VC_FIELD(src_vc, gm_port_id), &r_iov, &r_n_iov, &r_offset, &s_iov, &s_n_iov, &s_offset,
completion_ctr);
if (ret == LMT_AGAIN)
{
MPID_nem_gm_lmt_queue_t *e = MPID_nem_gm_queue_alloc (lmt);
if (!e)
{
printf ("error: malloc failed\n");
return -1;
}
e->node_id = VC_FIELD(src_vc, gm_node_id);
e->port_id = VC_FIELD(src_vc, gm_port_id);
e->r_iov = r_iov;
e->r_n_iov = r_n_iov;
e->r_offset = r_offset;
e->s_iov = s_iov;
e->s_n_iov = s_n_iov;
e->s_offset = s_offset;
e->compl_ctr = completion_ctr;
MPID_nem_gm_queue_enqueue (lmt, e);
}
else if (ret == LMT_FAILURE)
{
printf ("error: MPID_nem_gm_lmt_do_get() failed \n");
return -1;
}
return 0;
}
int
MPID_nem_newmad_vc_init (MPIDI_VC_t *vc)
{
MPIDI_CH3I_VC *vc_ch = VC_CH(vc);
char *business_card;
int mpi_errno = MPI_SUCCESS;
int val_max_sz;
int ret;
#ifdef USE_PMI2_API
val_max_sz = PMI2_MAX_VALLEN;
#else
mpi_errno = PMI_KVS_Get_value_length_max(&val_max_sz);
#endif
business_card = (char *)MPIU_Malloc(val_max_sz);
mpi_errno = vc->pg->getConnInfo(vc->pg_rank, business_card,val_max_sz,vc->pg);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* Very important */
memset(VC_FIELD(vc, url),0,MPID_NEM_NMAD_MAX_SIZE);
mpi_errno = MPID_nem_newmad_get_from_bc (business_card, VC_FIELD(vc, url));
if (mpi_errno) MPIU_ERR_POP (mpi_errno);
MPIU_Free(business_card);
ret = nm_session_connect(mpid_nem_newmad_session, &(VC_FIELD(vc,p_gate)), VC_FIELD(vc, url));
if (ret != NM_ESUCCESS) fprintf(stdout,"nm_session_connect returned ret = %d\n", ret);
nm_gate_ref_set(VC_FIELD(vc, p_gate),(void*)vc);
MPIDI_CHANGE_VC_STATE(vc, ACTIVE);
vc->eager_max_msg_sz = 32768;
vc->rndvSend_fn = NULL;
vc->sendNoncontig_fn = MPID_nem_newmad_SendNoncontig;
vc->comm_ops = &comm_ops;
vc_ch->iStartContigMsg = MPID_nem_newmad_iStartContigMsg;
vc_ch->iSendContig = MPID_nem_newmad_iSendContig;
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int
MPID_nem_mx_vc_init (MPIDI_VC_t *vc)
{
uint32_t threshold;
MPIDI_CH3I_VC *vc_ch = VC_CH(vc);
int mpi_errno = MPI_SUCCESS;
/* first make sure that our private fields in the vc fit into the area provided */
MPIU_Assert(sizeof(MPID_nem_mx_vc_area) <= MPID_NEM_VC_NETMOD_AREA_LEN);
#ifdef ONDEMAND
VC_FIELD(vc, local_connected) = 0;
VC_FIELD(vc, remote_connected) = 0;
#else
{
char *business_card;
int val_max_sz;
int ret;
#ifdef USE_PMI2_API
val_max_sz = PMI2_MAX_VALLEN;
#else
mpi_errno = PMI_KVS_Get_value_length_max(&val_max_sz);
#endif
business_card = (char *)MPIU_Malloc(val_max_sz);
mpi_errno = vc->pg->getConnInfo(vc->pg_rank, business_card,val_max_sz, vc->pg);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPID_nem_mx_get_from_bc (business_card, &VC_FIELD(vc, remote_endpoint_id), &VC_FIELD(vc, remote_nic_id));
if (mpi_errno) MPIU_ERR_POP (mpi_errno);
MPIU_Free(business_card);
ret = mx_connect(MPID_nem_mx_local_endpoint,VC_FIELD(vc, remote_nic_id),VC_FIELD(vc, remote_endpoint_id),
MPID_NEM_MX_FILTER,MX_INFINITE,&(VC_FIELD(vc, remote_endpoint_addr)));
MPIU_ERR_CHKANDJUMP1 (ret != MX_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**mx_connect", "**mx_connect %s", mx_strerror (ret));
mx_set_endpoint_addr_context(VC_FIELD(vc, remote_endpoint_addr),(void *)vc);
MPIDI_CHANGE_VC_STATE(vc, ACTIVE);
}
#endif
mx_get_info(MPID_nem_mx_local_endpoint, MX_COPY_SEND_MAX, NULL, 0, &threshold, sizeof(uint32_t));
vc->eager_max_msg_sz = threshold;
vc->rndvSend_fn = NULL;
vc->sendNoncontig_fn = MPID_nem_mx_SendNoncontig;
vc->comm_ops = &comm_ops;
vc_ch->iStartContigMsg = MPID_nem_mx_iStartContigMsg;
vc_ch->iSendContig = MPID_nem_mx_iSendContig;
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIDI_nem_ib_recv_addr(MPIDI_VC_t * vc, void *vstart)
{
MPIDI_nem_ib_pkt_address *pkt = vstart;
int i;
int ret;
#ifdef _ENABLE_XRC_
if (USE_XRC && (0 == MPIDI_CH3I_RDMA_Process.xrc_rdmafp ||
VC_XST_ISSET (vc, XF_CONN_CLOSING)))
return 1;
#endif
DEBUG_PRINT("set rdma address, dma address %p\n",
(void *)pkt->rdma_address);
/* check if it has accepted max allowing connections */
if (rdma_fp_sendconn_accepted >= rdma_polling_set_limit)
{
vbuf_address_reply_send(vc, RDMA_FP_MAX_SEND_CONN_REACHED);
goto fn_exit;
}
if (pkt->rdma_address != 0) {
/* Allocating the send vbufs for the eager RDMA flow */
ret = vbuf_fast_rdma_alloc(vc, 0);
if (ret == MPI_SUCCESS) {
for (i = 0; i < ib_hca_num_hcas; i ++) {
VC_FIELD(vc, connection)->rfp.RDMA_remote_buf_rkey[i] = pkt->rdma_hndl[i];
}
VC_FIELD(vc, connection)->rfp.remote_RDMA_buf = (void *)pkt->rdma_address;
vbuf_address_reply_send(vc, RDMA_FP_SUCCESS);
rdma_fp_sendconn_accepted++;
} else {
vbuf_address_reply_send(vc, RDMA_FP_SENDBUFF_ALLOC_FAILED);
return -1;
}
}
fn_exit:
return MPI_SUCCESS;
}
int MPID_nem_newmad_cancel_send(MPIDI_VC_t *vc, MPID_Request *sreq)
{
nm_sr_request_t *nmad_req = NULL;
int mpi_errno = MPI_SUCCESS;
int ret;
nmad_req = &(REQ_FIELD(sreq,newmad_req));
ret = nm_sr_scancel(mpid_nem_newmad_session,nmad_req);
if (ret == NM_ESUCCESS)
{
sreq->status.cancelled = TRUE;
(VC_FIELD(vc,pending_sends)) -= 1;
}
else
{
sreq->status.cancelled = FALSE;
}
fn_exit:
return mpi_errno;
fn_fail: ATTRIBUTE((unused))
goto fn_exit;
}
int MPID_nem_newmad_directRecv(MPIDI_VC_t *vc, MPID_Request *rreq)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_NEWMAD_DIRECTRECV);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_NEWMAD_DIRECTRECV);
if (!VC_CH(vc)->is_local)
{
nm_tag_t match_info = 0;
nm_tag_t match_mask = NEM_NMAD_MATCH_FULL_MASK;
MPIR_Rank_t source = rreq->dev.match.parts.rank;
MPIR_Context_id_t context = rreq->dev.match.parts.context_id;
Nmad_Nem_tag_t tag = rreq->dev.match.parts.tag;
int ret;
MPIDI_msg_sz_t data_sz;
int dt_contig;
MPI_Aint dt_true_lb;
MPID_Datatype *dt_ptr;
NEM_NMAD_DIRECT_MATCH(match_info,0,source,context);
if (tag != MPI_ANY_TAG)
{
NEM_NMAD_SET_TAG(match_info,tag);
}
else
{
NEM_NMAD_SET_ANYTAG(match_info);
NEM_NMAD_SET_ANYTAG(match_mask);
}
#ifdef DEBUG
fprintf(stdout,"========> Posting Recv req %p (match is %lx) \n",rreq,match_info);
#endif
MPIDI_Datatype_get_info(rreq->dev.user_count,rreq->dev.datatype, dt_contig, data_sz, dt_ptr,dt_true_lb);
rreq->dev.OnDataAvail = NULL;
if (dt_contig)
{
ret = nm_sr_irecv_with_ref_tagged(mpid_nem_newmad_session,VC_FIELD(vc,p_gate),match_info,match_mask,
(char*)(rreq->dev.user_buf) + dt_true_lb,data_sz,
&(REQ_FIELD(rreq,newmad_req)),(void*)rreq);
REQ_FIELD(rreq,iov) = NULL;
}
else
{
int num_seg = 0;
struct iovec *newmad_iov = (struct iovec *)MPIU_Malloc(NMAD_IOV_MAX_DEPTH*sizeof(struct iovec));
struct iovec *newmad_iov_ptr = &(newmad_iov[0]);
MPID_nem_newmad_process_rdtype(&rreq,dt_ptr,data_sz,&newmad_iov_ptr,&num_seg);
MPIU_Assert(num_seg <= NMAD_IOV_MAX_DEPTH);
#ifdef DEBUG
{
int index;
for(index = 0; index < num_seg ; index++)
{
fprintf(stdout,"======================\n");
fprintf(stdout,"RECV nmad_iov[%i]: [base %p][len %i]\n",index,
newmad_iov[index].iov_base,newmad_iov[index].iov_len);
}
}
#endif
ret = nm_sr_irecv_iov_with_ref_tagged(mpid_nem_newmad_session,VC_FIELD(vc,p_gate),match_info,match_mask,
newmad_iov,num_seg,&(REQ_FIELD(rreq,newmad_req)),(void*)rreq);
REQ_FIELD(rreq,iov) = newmad_iov;
}
}
else
{
/* Fixme : this might not work in the case of multiple netmods */
memset((&(REQ_FIELD(rreq,newmad_req))),0,sizeof(nm_sr_request_t));
}
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_NEWMAD_DIRECTRECV);
return mpi_errno;
fn_fail: ATTRIBUTE((unused))
goto fn_exit;
}
/**
* FIXME: Ideally the header size should be determined by high level macros,
* instead of hacking the message header at the device layer
*/
int MPIDI_CH3I_nem_ib_parse_header(MPIDI_VC_t * vc,
vbuf * v, void **pkt, int *header_size)
{
void *vstart;
MPIDI_nem_ib_pkt_comm_header *header;
#ifdef CRC_CHECK
unsigned long crc;
#endif
int mpi_errno = MPI_SUCCESS;
int ret;
DEBUG_PRINT("[parse header] vbuf address %p\n", v);
vstart = v->pheader;
header = (MPIDI_nem_ib_pkt_comm_header *)v->iheader;
DEBUG_PRINT("[parse header] header type %d\n", header->type);
/* set it to the header size by default */
*header_size = sizeof(MPIDI_CH3_Pkt_t);
#ifdef CRC_CHECK
crc = update_crc(1, (void *)((uintptr_t)header+sizeof *header),
v->content_size - sizeof *header);
if (crc != header->mrail.crc) {
int rank; PMI_Get_rank(&rank);
MPIU_Error_printf(stderr, "CRC mismatch, get %lx, should be %lx "
"type %d, ocntent size %d\n",
crc, header->mrail.crc, header->type, v->content_size);
exit( EXIT_FAILURE );
}
#endif
switch (header->type) {
/*header caching codes */
#ifndef MV2_DISABLE_HEADER_CACHING
case (MPIDI_CH3_PKT_FAST_EAGER_SEND):
case (MPIDI_CH3_PKT_FAST_EAGER_SEND_WITH_REQ):
{
/* since header caching do not have regular iheader,
* revert back pre-adjust
*/
v->content_size += IB_PKT_HEADER_LENGTH;
vstart -= IB_PKT_HEADER_LENGTH;
v->pheader -= IB_PKT_HEADER_LENGTH;
MPIDI_nem_ib_pkt_fast_eager *fast_header = vstart;
MPIDI_CH3_Pkt_eager_send_t *eager_header =
(MPIDI_CH3_Pkt_eager_send_t *) VC_FIELD(vc, connection)->rfp.
cached_incoming;
MPIDI_nem_ib_pkt_comm_header *eager_iheader =
(MPIDI_nem_ib_pkt_comm_header *) VC_FIELD(vc, connection)->rfp.
cached_incoming_iheader;
if (MPIDI_CH3_PKT_FAST_EAGER_SEND == header->type) {
*header_size = sizeof(MPIDI_nem_ib_pkt_fast_eager);
} else {
*header_size =
sizeof(MPIDI_nem_ib_pkt_fast_eager_with_req);
eager_header->sender_req_id =
((MPIDI_nem_ib_pkt_fast_eager_with_req *)
vstart)->sender_req_id;
}
header = eager_iheader;
DEBUG_PRINT("[receiver side] cached credit %d\n",
eager_iheader->rdma_credit);
eager_header->data_sz = fast_header->bytes_in_pkt;
*pkt = (void *) eager_header;
DEBUG_PRINT
("[recv: parse header] faster headersize returned %d\n",
*header_size);
}
break;
#endif
case (MPIDI_CH3_PKT_EAGER_SEND):
{
DEBUG_PRINT("[recv: parse header] pkt eager send\n");
/* header caching codes */
#ifndef MV2_DISABLE_HEADER_CACHING
if (v->padding != NORMAL_VBUF_FLAG &&
(v->content_size - sizeof(MPIDI_CH3_Pkt_t) <= MAX_SIZE_WITH_HEADER_CACHING) ) {
/* Only cache header if the packet is from RdMA path
* XXXX: what is R3_FLAG?
*/
MPIU_Memcpy((VC_FIELD(vc, connection)->rfp.cached_incoming), vstart,
sizeof(MPIDI_CH3_Pkt_eager_send_t));
MPIU_Memcpy((VC_FIELD(vc, connection)->rfp.cached_incoming_iheader), header,
sizeof(MPIDI_nem_ib_pkt_comm_header));
}
#endif
*pkt = (MPIDI_CH3_Pkt_t *) vstart;
*header_size = sizeof(MPIDI_CH3_Pkt_t);
DEBUG_PRINT("[recv: parse header] headersize returned %d\n",
*header_size);
}
break;
case (MPIDI_CH3_PKT_RNDV_REQ_TO_SEND):
case (MPIDI_CH3_PKT_RNDV_CLR_TO_SEND):
case MPIDI_CH3_PKT_EAGER_SYNC_ACK:
case MPIDI_NEM_PKT_LMT_RTS:
case MPIDI_NEM_PKT_LMT_CTS:
case MPIDI_NEM_PKT_LMT_DONE:
case MPIDI_NEM_PKT_LMT_COOKIE:
/* CKPT codes */
#ifdef CKPT
case MPIDI_CH3_PKT_CM_SUSPEND:
case MPIDI_CH3_PKT_CM_REACTIVATION_DONE:
case MPIDI_CH3_PKT_CR_REMOTE_UPDATE:
#endif
{
*pkt = vstart;
}
break;
case MPIDI_CH3_PKT_CANCEL_SEND_REQ:
{
*pkt = vstart;
/*Fix: Need to unregister and free the rndv buffer in get protocol.*/
}
break;
case MPIDI_CH3_PKT_CANCEL_SEND_RESP:
{
MPID_Request *req;
*pkt = vstart;
MPID_Request_get_ptr(((MPIDI_CH3_Pkt_cancel_send_resp_t *)(*pkt))->sender_req_id, req);
if (req != NULL) {
/* unregister and free the rndv buffer */
MPIDI_NEM_IB_RREQ_RNDV_FINISH(req);
}
}
break;
case (MPIDI_CH3_PKT_NOOP):
{
*pkt = v->iheader;
}
break;
/* rfp codes */
case MPIDI_CH3_PKT_ADDRESS:
{
*pkt = v->iheader;
MPIDI_nem_ib_recv_addr(vc, vstart);
break;
}
case MPIDI_CH3_PKT_ADDRESS_REPLY:
{
*pkt = v->iheader;
MPIDI_nem_ib_recv_addr_reply(vc, vstart);
break;
}
case MPIDI_CH3_PKT_PACKETIZED_SEND_START:
{
*pkt = vstart;
*header_size = sizeof(MPIDI_CH3_Pkt_packetized_send_start_t);
break;
}
case MPIDI_CH3_PKT_PACKETIZED_SEND_DATA:
{
*header_size = sizeof(MPIDI_CH3_Pkt_packetized_send_data_t);
*pkt = vstart;
break;
}
case MPIDI_CH3_PKT_RNDV_R3_DATA:
{
*header_size = sizeof(MPIDI_CH3_Pkt_rndv_r3_data_t);
*pkt = vstart;
break;
}
case MPIDI_CH3_PKT_RNDV_R3_ACK:
{
*pkt = v->iheader;
MPIDI_nem_ib_lmt_r3_recv_ack(vc, vstart);
break;
}
#if defined(USE_EAGER_SHORT)
case MPIDI_CH3_PKT_EAGERSHORT_SEND:
#endif
case MPIDI_CH3_PKT_EAGER_SYNC_SEND:
case MPIDI_CH3_PKT_READY_SEND:
{
*pkt = vstart;
break;
}
case MPIDI_CH3_PKT_PUT:
{
*pkt = vstart;
break;
}
case MPIDI_CH3_PKT_GET:
{
*pkt = vstart;
break;
}
case MPIDI_CH3_PKT_GET_RESP: /*15 */
{
*pkt = vstart;
break;
}
case MPIDI_CH3_PKT_ACCUMULATE:
{
*pkt = vstart;
break;
}
case MPIDI_CH3_PKT_LOCK:
{
*pkt = vstart;
break;
}
case MPIDI_CH3_PKT_LOCK_GRANTED:
{
*pkt = vstart;
break;
}
case MPIDI_CH3_PKT_PT_RMA_DONE:
{
*pkt = vstart;
break;
}
case MPIDI_CH3_PKT_LOCK_PUT_UNLOCK:
{
*pkt = vstart;
break;
}
case MPIDI_CH3_PKT_LOCK_GET_UNLOCK:
{
*pkt = vstart;
break;
}
case MPIDI_CH3_PKT_LOCK_ACCUM_UNLOCK:
case MPIDI_CH3_PKT_ACCUM_IMMED:
{
*pkt = vstart;
break;
}
case MPIDI_CH3_PKT_FLOW_CNTL_UPDATE:
{
*pkt = vstart;
break;
}
case MPIDI_CH3_PKT_CLOSE:
{
*pkt = vstart;
}
break;
default:
{
/* Header is corrupted if control has reached here in prototype */
/* */
MPIU_ERR_SETFATALANDJUMP2(mpi_errno,
MPI_ERR_OTHER,
"**fail",
"**fail %s %d",
"Control shouldn't reach here "
"in prototype, header %d\n",
header->type);
}
}
DEBUG_PRINT("Before set credit, vc: %p, v->rail: %d, "
"pkt: %p, pheader: %p\n", vc, v->rail, pkt, v->pheader);
SET_CREDIT(header, VC_FIELD(vc, connection), (v->rail));
if (VC_FIELD(vc, connection)->srp.credits[v->rail].remote_credit > 0 &&
VC_FIELD(vc, connection)->srp.credits[v->rail].backlog.len > 0) {
/* backlog send codes */
MRAILI_Backlog_send(vc, v->rail);
}
/* if any credits remain, schedule rendezvous progress */
if ((VC_FIELD(vc, connection)->srp.credits[v->rail].remote_credit > 0
/* rfp codes */
|| (VC_FIELD(vc, connection)->rfp.ptail_RDMA_send !=
VC_FIELD(vc, connection)->rfp.phead_RDMA_send)
)
&& (VC_FIELD(vc, connection)->sreq_head != NULL)) {
/* rndv codes */
#if 0
PUSH_FLOWLIST(vc);
#endif
}
/* rfp codes */
if ((VC_FIELD(vc, connection)->rfp.RDMA_recv_buf == NULL) && /*(c->initialized) && */
num_rdma_buffer && !VC_FIELD(vc, connection)->rfp.rdma_failed) {
if ((process_info.polling_group_size + rdma_pending_conn_request) <
rdma_polling_set_limit) {
VC_FIELD(vc, connection)->rfp.eager_start_cnt++;
if (rdma_polling_set_threshold <
VC_FIELD(vc, connection)->rfp.eager_start_cnt) {
{
ret = vbuf_fast_rdma_alloc(vc, 1);
if (ret == MPI_SUCCESS) {
vbuf_address_send(vc);
rdma_pending_conn_request++;
} else {
VC_FIELD(vc, connection)->rfp.rdma_failed = 1;
}
goto fn_exit;
}
}
}
}
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIDI_nem_ib_packetized_recv_data(MPIDI_VC_t * vc, vbuf *v)
{
int mpi_errno = MPI_SUCCESS;
int skipsize = sizeof(MPIDI_CH3_Pkt_packetized_send_data_t);
int nb, complete;
MPID_Request *rreq = VC_FIELD(vc, connection)->packetized_recv;
if (NULL == VC_FIELD(vc, connection)->packetized_recv) {
mpi_errno =
MPIR_Err_create_code(mpi_errno, MPIR_ERR_FATAL,
FCNAME, __LINE__,
MPI_ERR_OTHER,
"**fail", 0);
goto fn_exit;
}
DEBUG_PRINT("[pkt recv], rreq %p, offset %d, count %d\n", rreq, rreq->dev.iov_offset,
rreq->dev.iov_count);
mpi_errno = MPIDI_nem_ib_fill_request(rreq, v, skipsize, &nb);
if (mpi_errno != MPI_SUCCESS) {
mpi_errno =
MPIR_Err_create_code(mpi_errno, MPIR_ERR_FATAL,
FCNAME, __LINE__,
MPI_ERR_OTHER, "**fail",
0);
goto fn_exit;
}
skipsize += nb;
if (MPIDI_nem_ib_request_adjust_iov(rreq, nb)) {
mpi_errno = MPIDI_CH3U_Handle_recv_req(vc, rreq, &complete);
DEBUG_PRINT("[recv: handle read] adjust req fine, complete %d\n",
complete);
if (mpi_errno != MPI_SUCCESS) {
mpi_errno =
MPIR_Err_create_code(mpi_errno,
MPIR_ERR_RECOVERABLE, FCNAME,
__LINE__, MPI_ERR_OTHER, "**fail", 0);
goto fn_exit;
}
while (complete != TRUE) {
mpi_errno = MPIDI_nem_ib_fill_request(rreq, v, skipsize, &nb);
if (mpi_errno != MPI_SUCCESS) {
mpi_errno =
MPIR_Err_create_code(mpi_errno, MPIR_ERR_FATAL,
FCNAME, __LINE__,
MPI_ERR_OTHER, "**fail",
0);
goto fn_exit;
}
if (!MPIDI_nem_ib_request_adjust_iov(rreq, nb)) {
goto fn_exit;
}
skipsize += nb;
mpi_errno = MPIDI_CH3U_Handle_recv_req(vc, rreq, &complete);
DEBUG_PRINT("[recv: handle read] adjust req fine, complete %d\n",
complete);
if (mpi_errno != MPI_SUCCESS) {
mpi_errno =
MPIR_Err_create_code(mpi_errno,
MPIR_ERR_RECOVERABLE, FCNAME,
__LINE__, MPI_ERR_OTHER, "**fail", 0);
goto fn_exit;
}
}
if (TRUE == complete) {
VC_FIELD(vc, connection)->packetized_recv = NULL;
}
}
fn_exit:
return mpi_errno;
}
int MPIDI_nem_ib_recv_addr_reply(MPIDI_VC_t * vc, void *vstart)
{
int hca_index;
int ret;
MPIDI_nem_ib_pkt_address_reply *pkt = vstart;
DEBUG_PRINT("Received addr reply packet. reply data :%d\n", pkt->reply_data);
if (pkt->reply_data == RDMA_FP_SENDBUFF_ALLOC_FAILED
|| pkt->reply_data == RDMA_FP_MAX_SEND_CONN_REACHED) {
DEBUG_PRINT("RDMA FP setup failed. clean up recv buffers\n ");
/* de-regster the recv buffers */
for (hca_index = 0; hca_index < ib_hca_num_hcas; hca_index++) {
if (VC_FIELD(vc, connection)->rfp.RDMA_recv_buf_mr[hca_index]) {
ret = deregister_memory(VC_FIELD(vc, connection)->rfp.RDMA_recv_buf_mr[hca_index]);
if (ret) {
MPIU_Error_printf("Failed to deregister mr (%d)\n", ret);
} else {
VC_FIELD(vc, connection)->rfp.RDMA_recv_buf_mr[hca_index] = NULL;
}
}
}
/* deallocate recv RDMA buffers */
if (VC_FIELD(vc, connection)->rfp.RDMA_recv_buf_DMA) {
MPIU_Free(VC_FIELD(vc, connection)->rfp.RDMA_recv_buf_DMA);
VC_FIELD(vc, connection)->rfp.RDMA_recv_buf_DMA = NULL;
}
/* deallocate vbuf struct buffers */
if (VC_FIELD(vc, connection)->rfp.RDMA_recv_buf) {
MPIU_Free(VC_FIELD(vc, connection)->rfp.RDMA_recv_buf);
VC_FIELD(vc, connection)->rfp.RDMA_recv_buf = NULL;
}
/* set flag to mark that FP setup is failed/rejected.
we sholdn't try further on this vc */
VC_FIELD(vc, connection)->rfp.rdma_failed = 1;
} else if (pkt->reply_data == RDMA_FP_SUCCESS) {
/* set pointers */
VC_FIELD(vc, connection)->rfp.p_RDMA_recv = 0;
VC_FIELD(vc, connection)->rfp.p_RDMA_recv_tail = num_rdma_buffer - 1;
/* Add the connection to the RDMA polling list */
MPIU_Assert(process_info.polling_group_size < rdma_polling_set_limit);
process_info.polling_set
[process_info.polling_group_size] = vc;
process_info.polling_group_size++;
VC_FIELD(vc, cmanager)->num_channels += 1;
VC_FIELD(vc, cmanager)->num_local_pollings = 1;
VC_FIELD(vc, connection)->rfp.in_polling_set = 1;
} else {
ibv_va_error_abort(GEN_EXIT_ERR,
"Invalid reply data received. reply_data: pkt->reply_data%d\n",
pkt->reply_data);
}
rdma_pending_conn_request--;
return MPI_SUCCESS;
}
int MPID_nem_newmad_iprobe(MPIDI_VC_t *vc, int source, int tag, MPID_Comm *comm, int context_offset, int *flag, MPI_Status *status)
{
nm_tag_t match_info = 0;
nm_tag_t match_mask = NEM_NMAD_MATCH_FULL_MASK;
nm_gate_t out_gate;
nm_gate_t in_gate;
nm_tag_t out_tag;
int size;
int mpi_errno = MPI_SUCCESS;
int ret;
NEM_NMAD_SET_CTXT(match_info,comm->context_id + context_offset);
if( source == MPI_ANY_SOURCE)
{
NEM_NMAD_SET_ANYSRC(match_info);
NEM_NMAD_SET_ANYSRC(match_mask);
in_gate = NM_ANY_GATE;
}
else
{
NEM_NMAD_SET_SRC(match_info,source);
in_gate = VC_FIELD(vc,p_gate);
}
if (tag != MPI_ANY_TAG)
{
NEM_NMAD_SET_TAG(match_info,tag);
}
else
{
NEM_NMAD_SET_ANYTAG(match_info);
NEM_NMAD_SET_ANYTAG(match_mask);
}
ret = nm_sr_probe(mpid_nem_newmad_session,in_gate,&out_gate,
match_info,match_mask,&out_tag,&size);
if (ret == NM_ESUCCESS)
{
if (source != MPI_ANY_SOURCE)
status->MPI_SOURCE = source;
else
{
MPIDI_VC_t *vc;
int index;
vc = (MPIDI_VC_t *)nm_gate_ref_get(out_gate);
for(index = 0 ; index < comm->local_size ; index ++)
if (vc == comm->vcr[index])
break;
status->MPI_SOURCE = index;
}
if (tag != MPI_ANY_TAG)
status->MPI_TAG = tag;
else
NEM_NMAD_MATCH_GET_TAG(out_tag,status->MPI_TAG);
status->count = size;
*flag = TRUE;
}
else
*flag = FALSE;
fn_exit:
return mpi_errno;
fn_fail: ATTRIBUTE((unused))
goto fn_exit;
}
