static void rxd_handle_connect_ack(struct rxd_ep *ep, struct ofi_ctrl_hdr *ctrl,
struct rxd_rx_buf *rx_buf)
{
struct rxd_peer *peer;
struct dlist_entry *match;
struct rxd_tx_entry *tx_entry;
FI_INFO(&rxd_prov, FI_LOG_EP_CTRL,
"connect ack- msg_id: %" PRIu64 ", segno: %d\n",
ctrl->msg_id, ctrl->seg_no);
match = dlist_find_first_match(&ep->tx_entry_list,
rxd_conn_msg_match, ctrl);
if (!match) {
FI_INFO(&rxd_prov, FI_LOG_EP_CTRL, "no matching connect\n");
goto out;
}
tx_entry = container_of(match, struct rxd_tx_entry, entry);
peer = rxd_ep_getpeer_info(ep, tx_entry->peer);
peer->state = CMAP_CONNECTED;
peer->conn_data = ctrl->conn_id;
dlist_remove(match);
rxd_tx_entry_done(ep, tx_entry);
out:
rxd_ep_repost_buff(rx_buf);
}
int rxd_handle_ack(struct rxd_ep *ep, struct ofi_ctrl_hdr *ctrl,
struct rxd_rx_buf *rx_buf)
{
int ret = 0;
uint64_t idx;
struct rxd_tx_entry *tx_entry;
struct dlist_entry *item;
struct rxd_pkt_meta *pkt;
rxd_ep_lock_if_required(ep);
FI_DBG(&rxd_prov, FI_LOG_EP_CTRL, "got ack: msg: %p - %d\n",
ctrl->msg_id, ctrl->seg_no);
idx = ctrl->msg_id & RXD_TX_IDX_BITS;
tx_entry = &ep->tx_entry_fs->buf[idx];
if (tx_entry->msg_id != ctrl->msg_id)
goto out;
item = dlist_find_first_match(&tx_entry->pkt_list, rxd_tx_pkt_match, ctrl);
if (!item)
goto out;
pkt = container_of(item, struct rxd_pkt_meta, entry);
switch (pkt->type) {
case RXD_PKT_STRT:
case RXD_PKT_DATA:
ret = rxd_tx_entry_progress(ep, tx_entry, ctrl);
break;
case RXD_PKT_LAST:
rxd_ep_free_acked_pkts(ep, tx_entry, ctrl->seg_no);
FI_DBG(&rxd_prov, FI_LOG_EP_CTRL, "reporting TX completion : %p\n", tx_entry);
if (tx_entry->op_type != RXD_TX_READ_REQ) {
rxd_cq_report_tx_comp(ep->tx_cq, tx_entry);
rxd_tx_entry_done(ep, tx_entry);
}
break;
default:
FI_WARN(&rxd_prov, FI_LOG_EP_CTRL, "invalid pkt type\n");
break;
}
out:
rxd_ep_repost_buff(rx_buf);
rxd_ep_unlock_if_required(ep);
return ret;
}
/*
* Discarded transfers were discarded by the receiving side, so we abort
* transferring the rest of the data. However, the completion is still
* reported to the sender as successful. This ensures that short and long
* messages are treated the same, since short messages would be entirely
* buffered at the receiver, with no notification that the application later
* discarded the message.
*/
static void rxd_handle_discard(struct rxd_ep *ep, struct ofi_ctrl_hdr *ctrl,
struct rxd_rx_buf *rx_buf)
{
struct rxd_tx_entry *tx_entry;
uint64_t idx;
FI_DBG(&rxd_prov, FI_LOG_EP_CTRL,
"discard- msg_id: %" PRIu64 ", segno: %d\n",
ctrl->msg_id, ctrl->seg_no);
idx = ctrl->msg_id & RXD_TX_IDX_BITS;
tx_entry = &ep->tx_entry_fs->buf[idx];
if (tx_entry->msg_id == ctrl->msg_id) {
rxd_cq_report_tx_comp(rxd_ep_tx_cq(ep), tx_entry);
rxd_cntr_report_tx_comp(ep, tx_entry);
rxd_tx_entry_done(ep, tx_entry);
}
rxd_ep_repost_buff(rx_buf);
}
void rxd_handle_connect_ack(struct rxd_ep *ep, struct ofi_ctrl_hdr *ctrl,
struct rxd_rx_buf *rx_buf)
{
struct rxd_peer *peer;
struct dlist_entry *match;
struct rxd_tx_entry *tx_entry;
rxd_ep_lock_if_required(ep);
match = dlist_find_first_match(&ep->tx_entry_list, rxd_conn_msg_match, ctrl);
if (!match)
goto out;
tx_entry = container_of(match, struct rxd_tx_entry, entry);
peer = rxd_ep_getpeer_info(ep, tx_entry->peer);
peer->addr_published = 1;
peer->conn_data = ctrl->conn_id;
dlist_remove(match);
rxd_tx_entry_done(ep, tx_entry);
out:
rxd_ep_repost_buff(rx_buf);
rxd_ep_unlock_if_required(ep);
}
void rxd_ep_handle_data_msg(struct rxd_ep *ep, struct rxd_peer *peer,
struct rxd_rx_entry *rx_entry,
struct iovec *iov, size_t iov_count,
struct ofi_ctrl_hdr *ctrl, void *data,
struct rxd_rx_buf *rx_buf)
{
uint64_t done;
ep->credits++;
done = ofi_copy_to_iov(iov, iov_count, rx_entry->done, data, ctrl->seg_size);
rx_entry->done += done;
rx_entry->window--;
rx_entry->exp_seg_no++;
if (done != ctrl->seg_size) {
/* todo: generate truncation error */
/* inform peer */
FI_WARN(&rxd_prov, FI_LOG_EP_CTRL, "TODO: message truncated\n");
}
if (rx_entry->window == 0) {
rx_entry->window = rxd_get_window_sz(ep, rx_entry->op_hdr.size - rx_entry->done);
rx_entry->last_win_seg += rx_entry->window;
ep->credits -= rx_entry->window;
FI_DBG(&rxd_prov, FI_LOG_EP_CTRL, "replying ack [%p] - %d\n",
ctrl->msg_id, ctrl->seg_no);
rxd_ep_reply_ack(ep, ctrl, ofi_ctrl_ack, rx_entry->window,
rx_entry->key, peer->conn_data, ctrl->conn_id);
}
if (rx_entry->op_hdr.size != rx_entry->done) {
if (rx_entry->window == 0) {
dlist_init(&rx_entry->wait_entry);
dlist_insert_tail(&rx_entry->wait_entry, &ep->wait_rx_list);
FI_WARN(&rxd_prov, FI_LOG_EP_CTRL, "rx-entry %p - %d enqueued\n",
ctrl->msg_id, ctrl->seg_no);
} else {
FI_DBG(&rxd_prov, FI_LOG_EP_CTRL,
"rx_entry->op_hdr.size: %d, rx_entry->done: %d\n",
rx_entry->op_hdr.size, rx_entry->done);
}
return;
}
FI_DBG(&rxd_prov, FI_LOG_EP_CTRL, "reporting RX completion event\n");
rxd_report_rx_comp(ep->rx_cq, rx_entry);
switch(rx_entry->op_hdr.op) {
case ofi_op_msg:
freestack_push(ep->recv_fs, rx_entry->recv);
break;
case ofi_op_tagged:
freestack_push(ep->trecv_fs, rx_entry->trecv);
break;
case ofi_op_read_rsp:
rxd_cq_report_tx_comp(ep->tx_cq, rx_entry->read_rsp.tx_entry);
rxd_tx_entry_done(ep, rx_entry->read_rsp.tx_entry);
break;
default:
break;
}
rxd_rx_entry_release(ep, rx_entry);
}
void rxd_ep_handle_data_msg(struct rxd_ep *ep, struct rxd_peer *peer,
struct rxd_rx_entry *rx_entry,
struct iovec *iov, size_t iov_count,
struct ofi_ctrl_hdr *ctrl, void *data,
struct rxd_rx_buf *rx_buf)
{
struct fi_cq_tagged_entry cq_entry = {0};
struct util_cntr *cntr = NULL;
uint64_t done;
struct rxd_cq *rxd_rx_cq = rxd_ep_rx_cq(ep);
ep->credits++;
done = ofi_copy_to_iov(iov, iov_count, rx_entry->done, data,
ctrl->seg_size);
rx_entry->done += done;
rx_entry->credits--;
rx_entry->exp_seg_no++;
if (done != ctrl->seg_size) {
/* todo: generate truncation error */
/* inform peer */
FI_WARN(&rxd_prov, FI_LOG_EP_CTRL, "TODO: message truncated\n");
}
if (rx_entry->credits == 0) {
rxd_set_rx_credits(ep, rx_entry);
FI_DBG(&rxd_prov, FI_LOG_EP_CTRL, "replying ack [%p] - %d\n",
ctrl->msg_id, ctrl->seg_no);
rxd_ep_reply_ack(ep, ctrl, ofi_ctrl_ack, rx_entry->credits,
rx_entry->key, peer->conn_data, ctrl->conn_id);
}
if (rx_entry->op_hdr.size != rx_entry->done) {
if (rx_entry->credits == 0) {
dlist_init(&rx_entry->wait_entry);
dlist_insert_tail(&rx_entry->wait_entry, &ep->wait_rx_list);
FI_WARN(&rxd_prov, FI_LOG_EP_CTRL, "rx-entry %p - %d enqueued\n",
ctrl->msg_id, ctrl->seg_no);
} else {
FI_DBG(&rxd_prov, FI_LOG_EP_CTRL,
"rx_entry->op_hdr.size: %d, rx_entry->done: %d\n",
rx_entry->op_hdr.size, rx_entry->done);
}
return;
}
/* todo: handle FI_COMPLETION for RX CQ comp */
switch(rx_entry->op_hdr.op) {
case ofi_op_msg:
freestack_push(ep->recv_fs, rx_entry->recv);
/* Handle cntr */
cntr = ep->util_ep.rx_cntr;
/* Handle CQ comp */
cq_entry.flags |= FI_RECV;
cq_entry.op_context = rx_entry->recv->msg.context;
cq_entry.len = rx_entry->done;
cq_entry.buf = rx_entry->recv->iov[0].iov_base;
cq_entry.data = rx_entry->op_hdr.data;
rxd_rx_cq->write_fn(rxd_rx_cq, &cq_entry);
break;
case ofi_op_tagged:
freestack_push(ep->trecv_fs, rx_entry->trecv);
/* Handle cntr */
cntr = ep->util_ep.rx_cntr;
/* Handle CQ comp */
cq_entry.flags |= (FI_RECV | FI_TAGGED);
cq_entry.op_context = rx_entry->trecv->msg.context;
cq_entry.len = rx_entry->done;
cq_entry.buf = rx_entry->trecv->iov[0].iov_base;
cq_entry.data = rx_entry->op_hdr.data;
cq_entry.tag = rx_entry->trecv->msg.tag;\
rxd_rx_cq->write_fn(rxd_rx_cq, &cq_entry);
break;
case ofi_op_atomic:
/* Handle cntr */
cntr = ep->util_ep.rem_wr_cntr;
/* Handle CQ comp */
cq_entry.flags |= FI_ATOMIC;
rxd_rx_cq->write_fn(rxd_rx_cq, &cq_entry);
break;
case ofi_op_write:
/* Handle cntr */
cntr = ep->util_ep.rem_wr_cntr;
/* Handle CQ comp */
if (rx_entry->op_hdr.flags & OFI_REMOTE_CQ_DATA) {
cq_entry.flags |= (FI_RMA | FI_REMOTE_WRITE);
cq_entry.op_context = rx_entry->trecv->msg.context;
cq_entry.len = rx_entry->done;
cq_entry.buf = rx_entry->write.iov[0].iov_base;
cq_entry.data = rx_entry->op_hdr.data;
rxd_rx_cq->write_fn(rxd_rx_cq, &cq_entry);
}
break;
case ofi_op_read_rsp:
rxd_cq_report_tx_comp(rxd_ep_tx_cq(ep), rx_entry->read_rsp.tx_entry);
rxd_cntr_report_tx_comp(ep, rx_entry->read_rsp.tx_entry);
rxd_tx_entry_done(ep, rx_entry->read_rsp.tx_entry);
break;
default:
FI_WARN(&rxd_prov, FI_LOG_EP_CTRL, "invalid op type: %d\n",
rx_entry->op_hdr.op);
break;
}
if (cntr)
cntr->cntr_fid.ops->add(&cntr->cntr_fid, 1);
rxd_rx_entry_free(ep, rx_entry);
}
