int __recvpath
psmi_mq_handle_rts(psm_mq_t mq, uint64_t tag,
uintptr_t send_buf, uint32_t send_msglen,
psm_epaddr_t peer, mq_rts_callback_fn_t cb,
psm_mq_req_t *req_o)
{
psm_mq_req_t req;
uint32_t msglen;
int rc;
PSMI_PLOCK_ASSERT();
req = mq_req_match(&(mq->expected_q), tag, 1);
if (req) { /* we have a match, no need to callback */
msglen = mq_set_msglen(req, req->buf_len, send_msglen);
req->type = MQE_TYPE_RECV;
req->state = MQ_STATE_MATCHED;
req->tag = tag;
req->recv_msgoff = 0;
req->rts_peer = peer;
req->rts_sbuf = send_buf;
*req_o = req; /* yes match */
rc = MQ_RET_MATCH_OK;
}
else { /* No match, keep track of callback */
req = psmi_mq_req_alloc(mq, MQE_TYPE_RECV);
psmi_assert(req != NULL);
req->type = MQE_TYPE_RECV;
/* We don't know recv_msglen yet but we set it here for
* mq_iprobe */
req->send_msglen = req->recv_msglen = send_msglen;
req->state = MQ_STATE_UNEXP_RV;
req->tag = tag;
req->rts_callback = cb;
req->recv_msgoff = 0;
req->rts_peer = peer;
req->rts_sbuf = send_buf;
mq_sq_append(&mq->unexpected_q, req);
*req_o = req; /* no match, will callback */
rc = MQ_RET_UNEXP_OK;
}
_IPATH_VDBG("from=%s match=%s (req=%p) mqtag=%" PRIx64" recvlen=%d "
"sendlen=%d errcode=%d\n", psmi_epaddr_get_name(peer->epid),
rc == MQ_RET_MATCH_OK ? "YES" : "NO", req, req->tag,
req->recv_msglen, req->send_msglen, req->error_code);
return rc;
}
/*
* Note, epaddr is the master.
*/
int __recvpath
psmi_mq_handle_outoforder_queue(psm_epaddr_t epaddr)
{
psm_mq_t mq = epaddr->ep->mq;
psm_mq_req_t ureq, ereq;
uint32_t msglen;
next_ooo:
ureq = mq_ooo_match(&epaddr->outoforder_q, epaddr->mctxt_recv_seqnum);
if (ureq == NULL) return 0;
epaddr->mctxt_recv_seqnum++;
epaddr->outoforder_c--;
ereq = mq_req_match(&(mq->expected_q), ureq->tag, 1);
if (ereq == NULL) {
mq_sq_append(&mq->unexpected_q, ureq);
if (epaddr->outoforder_c) goto next_ooo;
return 0;
}
psmi_assert(MQE_TYPE_IS_RECV(ereq->type));
ereq->tag = ureq->tag;
msglen = mq_set_msglen(ereq, ereq->buf_len, ureq->send_msglen);
switch (ureq->state) {
case MQ_STATE_COMPLETE:
if (ureq->buf != NULL) { /* 0-byte don't alloc a sysbuf */
psmi_mq_mtucpy(ereq->buf,
(const void *)ureq->buf, msglen);
psmi_mq_sysbuf_free(mq, ureq->buf);
}
ereq->state = MQ_STATE_COMPLETE;
mq_qq_append(&mq->completed_q, ereq);
break;
case MQ_STATE_UNEXP: /* not done yet */
ereq->type = ureq->type;
ereq->egrid = ureq->egrid;
ereq->epaddr = ureq->epaddr;
ereq->send_msgoff = ureq->send_msgoff;
ereq->recv_msgoff = min(ureq->recv_msgoff, msglen);
psmi_mq_mtucpy(ereq->buf,
(const void *)ureq->buf, ereq->recv_msgoff);
psmi_mq_sysbuf_free(mq, ureq->buf);
ereq->state = MQ_STATE_MATCHED;
STAILQ_INSERT_AFTER(&ureq->epaddr->mctxt_master->egrlong,
ureq, ereq, nextq);
STAILQ_REMOVE(&ureq->epaddr->mctxt_master->egrlong,
ureq, psm_mq_req, nextq);
break;
case MQ_STATE_UNEXP_RV: /* rendez-vous ... */
ereq->state = MQ_STATE_MATCHED;
ereq->rts_peer = ureq->rts_peer;
ereq->rts_sbuf = ureq->rts_sbuf;
ereq->send_msgoff = 0;
ereq->rts_callback = ureq->rts_callback;
ereq->rts_reqidx_peer = ureq->rts_reqidx_peer;
ereq->type = ureq->type;
ereq->rts_callback(ereq, 0);
break;
default:
fprintf(stderr, "Unexpected state %d in req %p\n", ureq->state, ureq);
fprintf(stderr, "type=%d, mq=%p, tag=%p\n",
ureq->type, ureq->mq, (void *)(uintptr_t)ureq->tag);
abort();
}
psmi_mq_req_free(ureq);
if (epaddr->outoforder_c) goto next_ooo;
return 0;
}
/*
* This handles the regular (i.e. non-rendezvous MPI envelopes)
*/
int __recvpath
psmi_mq_handle_envelope(psm_mq_t mq, uint16_t mode, psm_epaddr_t epaddr,
uint64_t tag, psmi_egrid_t egrid, uint32_t send_msglen,
const void *payload, uint32_t paylen)
{
psm_mq_req_t req;
uint32_t msglen;
int rc;
psmi_assert(epaddr != NULL);
req = mq_req_match(&(mq->expected_q), tag, 1);
if (req) { /* we have a match */
psmi_assert(MQE_TYPE_IS_RECV(req->type));
req->tag = tag;
msglen = mq_set_msglen(req, req->buf_len, send_msglen);
_IPATH_VDBG("from=%s match=YES (req=%p) mode=%x mqtag=%"
PRIx64" msglen=%d paylen=%d\n", psmi_epaddr_get_name(epaddr->epid),
req, mode, tag, msglen, paylen);
switch(mode) {
case MQ_MSG_TINY:
PSM_VALGRIND_DEFINE_MQ_RECV(req->buf, req->buf_len, msglen);
mq_copy_tiny((uint32_t *)req->buf, (uint32_t *)payload, msglen);
req->state = MQ_STATE_COMPLETE;
mq_qq_append(&mq->completed_q, req);
break;
case MQ_MSG_SHORT: /* message fits in 1 payload */
PSM_VALGRIND_DEFINE_MQ_RECV(req->buf, req->buf_len, msglen);
psmi_mq_mtucpy(req->buf, payload, msglen);
req->state = MQ_STATE_COMPLETE;
mq_qq_append(&mq->completed_q, req);
break;
case MQ_MSG_LONG:
req->egrid = egrid;
req->state = MQ_STATE_MATCHED;
req->type |= MQE_TYPE_EGRLONG;
req->send_msgoff = req->recv_msgoff = 0;
STAILQ_INSERT_TAIL(&epaddr->mctxt_master->egrlong, req, nextq);
_IPATH_VDBG("exp MSG_LONG %d of length %d bytes pay=%d\n",
egrid.egr_msgno, msglen, paylen);
if (paylen > 0)
psmi_mq_handle_data(req, epaddr,
egrid.egr_data, 0, payload, paylen);
psmi_mq_handle_egrdata(mq, req, epaddr);
break;
default:
psmi_handle_error(PSMI_EP_NORETURN, PSM_INTERNAL_ERR,
"Internal error, unknown packet 0x%x", mode);
}
mq->stats.rx_user_bytes += msglen;
mq->stats.rx_user_num++;
rc = MQ_RET_MATCH_OK;
if (mode == MQ_MSG_LONG)
return rc;
}
else
rc = psmi_mq_handle_envelope_unexpected(mq, mode, epaddr, tag,
egrid, send_msglen, payload, paylen);
return rc;
}
