ssize_t udpx_send(struct fid_ep *ep_fid, const void *buf, size_t len, void *desc,
fi_addr_t dest_addr, void *context)
{
struct udpx_ep *ep;
ssize_t ret;
ep = container_of(ep_fid, struct udpx_ep, util_ep.ep_fid.fid);
fastlock_acquire(&ep->util_ep.tx_cq->cq_lock);
if (ofi_cirque_isfull(ep->util_ep.tx_cq->cirq)) {
ret = -FI_EAGAIN;
goto out;
}
ret = sendto(ep->sock, buf, len, 0,
ip_av_get_addr(ep->util_ep.av, dest_addr),
ep->util_ep.av->addrlen);
if (ret == len) {
ep->tx_comp(ep, context);
ret = 0;
} else {
ret = -errno;
}
out:
fastlock_release(&ep->util_ep.tx_cq->cq_lock);
return ret;
}
ssize_t udpx_sendmsg(struct fid_ep *ep_fid, const struct fi_msg *msg,
uint64_t flags)
{
struct udpx_ep *ep;
struct msghdr hdr;
ssize_t ret;
ep = container_of(ep_fid, struct udpx_ep, util_ep.ep_fid.fid);
hdr.msg_name = ip_av_get_addr(ep->util_ep.av, msg->addr);
hdr.msg_namelen = ep->util_ep.av->addrlen;
hdr.msg_iov = (struct iovec *) msg->msg_iov;
hdr.msg_iovlen = msg->iov_count;
hdr.msg_control = NULL;
hdr.msg_controllen = 0;
hdr.msg_flags = 0;
fastlock_acquire(&ep->util_ep.tx_cq->cq_lock);
if (ofi_cirque_isfull(ep->util_ep.tx_cq->cirq)) {
ret = -FI_EAGAIN;
goto out;
}
ret = sendmsg(ep->sock, &hdr, 0);
if (ret >= 0) {
ep->tx_comp(ep, msg->context);
ret = 0;
} else {
ret = -errno;
}
out:
fastlock_release(&ep->util_ep.tx_cq->cq_lock);
return ret;
}
ssize_t udpx_recvmsg(struct fid_ep *ep_fid, const struct fi_msg *msg,
uint64_t flags)
{
struct udpx_ep *ep;
struct udpx_ep_entry *entry;
ssize_t ret;
ep = container_of(ep_fid, struct udpx_ep, util_ep.ep_fid.fid);
fastlock_acquire(&ep->util_ep.rx_cq->cq_lock);
if (ofi_cirque_isfull(ep->rxq)) {
ret = -FI_EAGAIN;
goto out;
}
entry = ofi_cirque_tail(ep->rxq);
entry->context = msg->context;
for (entry->iov_count = 0; entry->iov_count < msg->iov_count;
entry->iov_count++) {
entry->iov[entry->iov_count] = msg->msg_iov[entry->iov_count];
}
entry->flags = 0;
ofi_cirque_commit(ep->rxq);
ret = 0;
out:
fastlock_release(&ep->util_ep.rx_cq->cq_lock);
return ret;
}
ssize_t udpx_recv(struct fid_ep *ep_fid, void *buf, size_t len, void *desc,
fi_addr_t src_addr, void *context)
{
struct udpx_ep *ep;
struct udpx_ep_entry *entry;
ssize_t ret;
ep = container_of(ep_fid, struct udpx_ep, util_ep.ep_fid.fid);
fastlock_acquire(&ep->util_ep.rx_cq->cq_lock);
if (ofi_cirque_isfull(ep->rxq)) {
ret = -FI_EAGAIN;
goto out;
}
entry = ofi_cirque_tail(ep->rxq);
entry->context = context;
entry->iov_count = 1;
entry->iov[0].iov_base = buf;
entry->iov[0].iov_len = len;
entry->flags = 0;
ofi_cirque_commit(ep->rxq);
ret = 0;
out:
fastlock_release(&ep->util_ep.rx_cq->cq_lock);
return ret;
}
static int rxm_check_unexp_msg_list(struct util_cq *util_cq, struct rxm_recv_queue *recv_queue,
struct rxm_recv_entry *recv_entry, dlist_func_t *match)
{
struct dlist_entry *entry;
struct rxm_unexp_msg *unexp_msg;
struct rxm_recv_match_attr match_attr;
struct rxm_rx_buf *rx_buf;
int ret = 0;
fastlock_acquire(&util_cq->cq_lock);
if (ofi_cirque_isfull(util_cq->cirq)) {
ret = -FI_EAGAIN;
goto out;
}
match_attr.addr = recv_entry->addr;
match_attr.tag = recv_entry->tag;
match_attr.ignore = recv_entry->ignore;
entry = dlist_remove_first_match(&recv_queue->unexp_msg_list, match, &match_attr);
if (!entry)
goto out;
FI_DBG(&rxm_prov, FI_LOG_EP_DATA, "Match for posted recv found in unexp msg list\n");
unexp_msg = container_of(entry, struct rxm_unexp_msg, entry);
rx_buf = container_of(unexp_msg, struct rxm_rx_buf, unexp_msg);
rx_buf->recv_entry = recv_entry;
ret = rxm_cq_handle_data(rx_buf);
free(unexp_msg);
out:
fastlock_release(&util_cq->cq_lock);
return ret;
}
static int rxd_cq_write_ctx(struct rxd_cq *cq,
struct fi_cq_tagged_entry *cq_entry)
{
struct fi_cq_tagged_entry *comp;
if (ofi_cirque_isfull(cq->util_cq.cirq))
return -FI_ENOSPC;
comp = ofi_cirque_tail(cq->util_cq.cirq);
comp->op_context = cq_entry->op_context;
ofi_cirque_commit(cq->util_cq.cirq);
return 0;
}
static int rxd_cq_write_tagged(struct rxd_cq *cq,
struct fi_cq_tagged_entry *cq_entry)
{
struct fi_cq_tagged_entry *comp;
if (ofi_cirque_isfull(cq->util_cq.cirq))
return -FI_ENOSPC;
FI_DBG(&rxd_prov, FI_LOG_EP_CTRL,
"report completion: %p\n", cq_entry->tag);
comp = ofi_cirque_tail(cq->util_cq.cirq);
*comp = *cq_entry;
ofi_cirque_commit(cq->util_cq.cirq);
return 0;
}
static int rxd_cq_write_data(struct rxd_cq *cq,
struct fi_cq_tagged_entry *cq_entry)
{
struct fi_cq_tagged_entry *comp;
if (ofi_cirque_isfull(cq->util_cq.cirq))
return -FI_ENOSPC;
comp = ofi_cirque_tail(cq->util_cq.cirq);
comp->op_context = cq_entry->op_context;
comp->flags = cq_entry->flags;
comp->len = cq_entry->len;
comp->buf = cq_entry->buf;
comp->data = cq_entry->data;
ofi_cirque_commit(cq->util_cq.cirq);
return 0;
}
static ssize_t rxd_generic_write_inject(struct rxd_ep *rxd_ep,
const struct iovec *iov, size_t iov_count,
const struct fi_rma_iov *rma_iov, size_t rma_count,
fi_addr_t addr, void *context, uint32_t op, uint64_t data,
uint32_t rxd_flags)
{
struct rxd_x_entry *tx_entry;
fi_addr_t rxd_addr;
ssize_t ret = -FI_EAGAIN;
assert(iov_count <= RXD_IOV_LIMIT && rma_count <= RXD_IOV_LIMIT);
assert(ofi_total_iov_len(iov, iov_count) <= rxd_ep_domain(rxd_ep)->max_inline_rma);
fastlock_acquire(&rxd_ep->util_ep.lock);
fastlock_acquire(&rxd_ep->util_ep.tx_cq->cq_lock);
if (ofi_cirque_isfull(rxd_ep->util_ep.tx_cq->cirq))
goto out;
rxd_addr = rxd_ep_av(rxd_ep)->fi_addr_table[addr];
ret = rxd_send_rts_if_needed(rxd_ep, rxd_addr);
if (ret)
goto out;
tx_entry = rxd_tx_entry_init(rxd_ep, iov, iov_count, NULL, 0, rma_count, data,
0, context, rxd_addr, op, rxd_flags);
if (!tx_entry)
goto out;
ret = rxd_ep_send_op(rxd_ep, tx_entry, rma_iov, rma_count, NULL, 0, 0, 0);
if (ret) {
rxd_tx_entry_free(rxd_ep, tx_entry);
goto out;
}
if (tx_entry->op == RXD_READ_REQ)
goto out;
ret = 0;
out:
fastlock_release(&rxd_ep->util_ep.tx_cq->cq_lock);
fastlock_release(&rxd_ep->util_ep.lock);
return ret;
}
static void smr_post_fetch_resp(struct smr_ep *ep, struct smr_cmd *cmd,
const struct iovec *result_iov, size_t count)
{
struct smr_cmd *pend;
struct smr_resp *resp;
assert(!ofi_cirque_isfull(smr_resp_queue(ep->region)));
resp = ofi_cirque_tail(smr_resp_queue(ep->region));
cmd->msg.hdr.data = (uint64_t) ((char **) resp -
(char **) ep->region);
pend = freestack_pop(ep->pend_fs);
smr_post_pend_resp(cmd, pend, resp);
memcpy(pend->msg.data.iov, result_iov,
sizeof(*result_iov) * count);
pend->msg.data.iov_count = count;
ofi_cirque_commit(smr_resp_queue(ep->region));
}
ssize_t rxd_generic_rma(struct rxd_ep *rxd_ep, const struct iovec *iov,
size_t iov_count, const struct fi_rma_iov *rma_iov, size_t rma_count,
void **desc, fi_addr_t addr, void *context, uint32_t op, uint64_t data,
uint32_t rxd_flags)
{
struct rxd_x_entry *tx_entry;
fi_addr_t rxd_addr;
ssize_t ret = -FI_EAGAIN;
if (rxd_flags & RXD_INJECT)
return rxd_generic_write_inject(rxd_ep, iov, iov_count, rma_iov,
rma_count, addr, context, op,
data, rxd_flags);
assert(iov_count <= RXD_IOV_LIMIT && rma_count <= RXD_IOV_LIMIT);
fastlock_acquire(&rxd_ep->util_ep.lock);
fastlock_acquire(&rxd_ep->util_ep.tx_cq->cq_lock);
if (ofi_cirque_isfull(rxd_ep->util_ep.tx_cq->cirq))
goto out;
rxd_addr = rxd_ep_av(rxd_ep)->fi_addr_table[addr];
ret = rxd_send_rts_if_needed(rxd_ep, rxd_addr);
if (ret)
goto out;
tx_entry = rxd_tx_entry_init(rxd_ep, iov, iov_count, NULL, 0, rma_count,
data, 0, context, rxd_addr, op, rxd_flags);
if (!tx_entry)
goto out;
ret = rxd_ep_send_op(rxd_ep, tx_entry, rma_iov, rma_count, NULL, 0, 0, 0);
if (ret)
rxd_tx_entry_free(rxd_ep, tx_entry);
out:
fastlock_release(&rxd_ep->util_ep.tx_cq->cq_lock);
fastlock_release(&rxd_ep->util_ep.lock);
return ret;
}
struct tcpx_xfer_entry *tcpx_xfer_entry_alloc(struct tcpx_cq *tcpx_cq,
enum tcpx_xfer_op_codes type)
{
struct tcpx_xfer_entry *xfer_entry;
tcpx_cq->util_cq.cq_fastlock_acquire(&tcpx_cq->util_cq.cq_lock);
/* optimization: don't allocate queue_entry when cq is full */
if (ofi_cirque_isfull(tcpx_cq->util_cq.cirq)) {
tcpx_cq->util_cq.cq_fastlock_release(&tcpx_cq->util_cq.cq_lock);
return NULL;
}
xfer_entry = ofi_buf_alloc(tcpx_cq->buf_pools[type].pool);
if (!xfer_entry) {
tcpx_cq->util_cq.cq_fastlock_release(&tcpx_cq->util_cq.cq_lock);
FI_INFO(&tcpx_prov, FI_LOG_DOMAIN,"failed to get buffer\n");
return NULL;
}
tcpx_cq->util_cq.cq_fastlock_release(&tcpx_cq->util_cq.cq_lock);
return xfer_entry;
}
static ssize_t smr_generic_atomic(struct fid_ep *ep_fid,
const struct fi_ioc *ioc, void **desc, size_t count,
const struct fi_ioc *compare_ioc, void **compare_desc,
size_t compare_count, struct fi_ioc *result_ioc,
void **result_desc, size_t result_count,
fi_addr_t addr, const struct fi_rma_ioc *rma_ioc,
size_t rma_count, enum fi_datatype datatype,
enum fi_op atomic_op, void *context, uint32_t op)
{
struct smr_ep *ep;
struct smr_domain *domain;
struct smr_region *peer_smr;
struct smr_inject_buf *tx_buf;
struct smr_cmd *cmd;
struct iovec iov[SMR_IOV_LIMIT];
struct iovec compare_iov[SMR_IOV_LIMIT];
struct iovec result_iov[SMR_IOV_LIMIT];
int peer_id, err = 0;
uint16_t flags = 0;
ssize_t ret = 0;
size_t msg_len, total_len;
assert(count <= SMR_IOV_LIMIT);
assert(result_count <= SMR_IOV_LIMIT);
assert(compare_count <= SMR_IOV_LIMIT);
assert(rma_count <= SMR_IOV_LIMIT);
ep = container_of(ep_fid, struct smr_ep, util_ep.ep_fid.fid);
domain = container_of(ep->util_ep.domain, struct smr_domain, util_domain);
peer_id = (int) addr;
ret = smr_verify_peer(ep, peer_id);
if(ret)
return ret;
peer_smr = smr_peer_region(ep->region, peer_id);
fastlock_acquire(&peer_smr->lock);
if (peer_smr->cmd_cnt < 2) {
ret = -FI_EAGAIN;
goto unlock_region;
}
fastlock_acquire(&ep->util_ep.tx_cq->cq_lock);
if (ofi_cirque_isfull(ep->util_ep.tx_cq->cirq)) {
ret = -FI_EAGAIN;
goto unlock_cq;
}
cmd = ofi_cirque_tail(smr_cmd_queue(peer_smr));
msg_len = total_len = ofi_datatype_size(datatype) *
ofi_total_ioc_cnt(ioc, count);
switch (op) {
case ofi_op_atomic_compare:
assert(compare_ioc);
ofi_ioc_to_iov(compare_ioc, compare_iov, compare_count,
ofi_datatype_size(datatype));
total_len *= 2;
/* fall through */
case ofi_op_atomic_fetch:
assert(result_ioc);
ofi_ioc_to_iov(result_ioc, result_iov, result_count,
ofi_datatype_size(datatype));
if (!domain->fast_rma)
flags |= SMR_RMA_REQ;
/* fall through */
case ofi_op_atomic:
if (atomic_op != FI_ATOMIC_READ) {
assert(ioc);
ofi_ioc_to_iov(ioc, iov, count, ofi_datatype_size(datatype));
} else {
count = 0;
}
break;
default:
break;
}
if (total_len <= SMR_MSG_DATA_LEN && !(flags & SMR_RMA_REQ)) {
smr_format_inline_atomic(cmd, smr_peer_addr(ep->region)[peer_id].addr,
iov, count, compare_iov, compare_count,
op, datatype, atomic_op);
} else if (total_len <= SMR_INJECT_SIZE) {
tx_buf = smr_freestack_pop(smr_inject_pool(peer_smr));
smr_format_inject_atomic(cmd, smr_peer_addr(ep->region)[peer_id].addr,
iov, count, result_iov, result_count,
compare_iov, compare_count, op, datatype,
atomic_op, peer_smr, tx_buf);
} else {
FI_WARN(&smr_prov, FI_LOG_EP_CTRL,
"message too large\n");
ret = -FI_EINVAL;
goto unlock_cq;
}
cmd->msg.hdr.op_flags |= flags;
ofi_cirque_commit(smr_cmd_queue(peer_smr));
peer_smr->cmd_cnt--;
if (op != ofi_op_atomic) {
if (flags & SMR_RMA_REQ) {
smr_post_fetch_resp(ep, cmd,
(const struct iovec *) result_iov,
result_count);
goto format_rma;
}
err = smr_fetch_result(ep, peer_smr, result_iov, result_count,
rma_ioc, rma_count, datatype, msg_len);
if (err)
FI_WARN(&smr_prov, FI_LOG_EP_CTRL,
"unable to fetch results");
}
ret = ep->tx_comp(ep, context, ofi_tx_cq_flags(op), err);
if (ret) {
FI_WARN(&smr_prov, FI_LOG_EP_CTRL,
"unable to process tx completion\n");
}
format_rma:
cmd = ofi_cirque_tail(smr_cmd_queue(peer_smr));
smr_format_rma_ioc(cmd, rma_ioc, rma_count);
ofi_cirque_commit(smr_cmd_queue(peer_smr));
peer_smr->cmd_cnt--;
unlock_cq:
fastlock_release(&ep->util_ep.tx_cq->cq_lock);
unlock_region:
fastlock_release(&peer_smr->lock);
return ret;
}