int MPID_nem_ib_lmt_start_recv_core(struct MPID_Request *req, void *raddr, uint32_t rkey,
void *write_to_buf)
{
int mpi_errno = MPI_SUCCESS;
int ibcom_errno;
struct MPIDI_VC *vc = req->ch.vc;
MPID_nem_ib_vc_area *vc_ib = VC_IB(vc);
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_IB_LMT_START_RECV_CORE);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_IB_LMT_START_RECV_CORE);
ibcom_errno =
MPID_nem_ib_com_lrecv(vc_ib->sc->fd, (uint64_t) req, raddr, req->ch.lmt_data_sz, rkey,
write_to_buf);
MPID_nem_ib_ncqe += 1;
//dprintf("start_recv,ncqe=%d\n", MPID_nem_ib_ncqe);
MPIU_ERR_CHKANDJUMP(ibcom_errno, mpi_errno, MPI_ERR_OTHER, "**MPID_nem_ib_com_lrecv");
dprintf("lmt_start_recv_core,MPID_nem_ib_ncqe=%d\n", MPID_nem_ib_ncqe);
dprintf
("lmt_start_recv_core,req=%p,sz=%ld,write_to_buf=%p,lmt_pack_buf=%p,user_buf=%p,raddr=%p,rkey=%08x,tail=%p=%02x\n",
req, req->ch.lmt_data_sz, write_to_buf, REQ_FIELD(req, lmt_pack_buf), req->dev.user_buf,
raddr, rkey, write_to_buf + req->ch.lmt_data_sz - sizeof(uint8_t),
*((uint8_t *) (write_to_buf + req->ch.lmt_data_sz - sizeof(uint8_t))));
#ifdef MPID_NEM_IB_LMT_GET_CQE
MPID_nem_ib_ncqe_to_drain += 1; /* use CQE instead of polling */
#else
/* drain_scq and ib_poll is not ordered, so both can decrement ref_count */
MPIR_Request_add_ref(req);
/* register to poll list in ib_poll() */
/* don't use req->dev.next because it causes unknown problem */
MPID_nem_ib_lmtq_enqueue(&MPID_nem_ib_lmtq, req);
dprintf("lmt_start_recv_core,lmtq enqueue\n");
//volatile uint8_t* tailmagic = (uint8_t*)((void*)req->dev.user_buf + req->ch.lmt_data_sz - sizeof(uint8_t));
//dprintf("start_recv_core,cur_tail=%02x,lmt_receiver_tail=%02x\n", *tailmagic, REQ_FIELD(req, lmt_receiver_tail));
#endif
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_IB_LMT_START_RECV_CORE);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_nem_ib_lmt_start_recv_core(struct MPID_Request *req, void *raddr, uint32_t rkey, long len,
void *write_to_buf, uint32_t max_msg_sz, int end)
{
int mpi_errno = MPI_SUCCESS;
int ibcom_errno;
struct MPIDI_VC *vc = req->ch.vc;
MPID_nem_ib_vc_area *vc_ib = VC_IB(vc);
int i;
int divide;
int posted_num;
int last;
uint32_t r_max_msg_sz; /* responder's max_msg_sz */
void *write_pos;
void *addr;
long data_sz;
MPIDI_msg_sz_t rest_data_sz;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_IB_LMT_START_RECV_CORE);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_IB_LMT_START_RECV_CORE);
MPID_nem_ib_com_get_info_conn(vc_ib->sc->fd, MPID_NEM_IB_COM_INFOKEY_PATTR_MAX_MSG_SZ,
&r_max_msg_sz, sizeof(uint32_t));
divide = (max_msg_sz + r_max_msg_sz - 1) / r_max_msg_sz;
write_pos = write_to_buf;
posted_num = 0;
last = MPID_NEM_IB_LMT_PART_OF_SEGMENT;
rest_data_sz = len;
addr = raddr;
for (i = 0; i < divide; i++) {
if (i == divide - 1)
data_sz = max_msg_sz - i * r_max_msg_sz;
else
data_sz = r_max_msg_sz;
if (i == divide - 1) {
if (end)
last = MPID_NEM_IB_LMT_LAST_PKT; /* last part of last segment packet */
else
last = MPID_NEM_IB_LMT_SEGMENT_LAST; /* last part of this segment */
/* last data may be smaller than initiator's max_msg_sz */
if (rest_data_sz < max_msg_sz)
data_sz = rest_data_sz;
}
ibcom_errno =
MPID_nem_ib_com_lrecv(vc_ib->sc->fd, (uint64_t) req, addr, data_sz, rkey,
write_pos, last);
MPIU_ERR_CHKANDJUMP(ibcom_errno, mpi_errno, MPI_ERR_OTHER, "**MPID_nem_ib_com_lrecv");
/* update position */
write_pos = (void *) ((char *) write_pos + data_sz);
addr = (void *) ((char *) addr + data_sz);
/* update rest data size */
rest_data_sz -= data_sz;
/* count request number */
posted_num++;
}
MPIU_Assert(rest_data_sz == 0);
MPID_nem_ib_ncqe += posted_num;
//dprintf("start_recv,ncqe=%d\n", MPID_nem_ib_ncqe);
dprintf("lmt_start_recv_core,MPID_nem_ib_ncqe=%d\n", MPID_nem_ib_ncqe);
dprintf
("lmt_start_recv_core,req=%p,sz=%ld,write_to_buf=%p,lmt_pack_buf=%p,user_buf=%p,raddr=%p,rkey=%08x,tail=%p=%02x\n",
req, req->ch.lmt_data_sz, write_to_buf, REQ_FIELD(req, lmt_pack_buf), req->dev.user_buf,
raddr, rkey, write_to_buf + req->ch.lmt_data_sz - sizeof(uint8_t),
*((uint8_t *) (write_to_buf + req->ch.lmt_data_sz - sizeof(uint8_t))));
//fflush(stdout);
#ifdef MPID_NEM_IB_LMT_GET_CQE
MPID_nem_ib_ncqe_to_drain += posted_num; /* use CQE instead of polling */
#else
/* drain_scq and ib_poll is not ordered, so both can decrement ref_count */
MPIR_Request_add_ref(req);
/* register to poll list in ib_poll() */
/* don't use req->dev.next because it causes unknown problem */
MPID_nem_ib_lmtq_enqueue(&MPID_nem_ib_lmtq, req);
dprintf("lmt_start_recv_core,lmtq enqueue\n");
//volatile uint8_t* tailmagic = (uint8_t*)((void*)req->dev.user_buf + req->ch.lmt_data_sz - sizeof(uint8_t));
//dprintf("start_recv_core,cur_tail=%02x,lmt_receiver_tail=%02x\n", *tailmagic, REQ_FIELD(req, lmt_receiver_tail));
#endif
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_IB_LMT_START_RECV_CORE);
return mpi_errno;
fn_fail:
goto fn_exit;
}