Test(vc_management_auto, vc_lookup_by_id)
{
int ret;
struct gnix_vc *vc[2], *vc_chk;
struct gnix_fid_ep *ep_priv;
ep_priv = container_of(ep[0], struct gnix_fid_ep, ep_fid);
ret = _gnix_vc_alloc(ep_priv, gnix_addr[0], &vc[0]);
cr_assert_eq(ret, FI_SUCCESS);
ret = _gnix_vc_alloc(ep_priv, gnix_addr[1], &vc[1]);
cr_assert_eq(ret, FI_SUCCESS);
vc_chk = __gnix_nic_elem_by_rem_id(ep_priv->nic, vc[0]->vc_id);
cr_assert_eq(vc_chk, vc[0]);
vc_chk = __gnix_nic_elem_by_rem_id(ep_priv->nic, vc[1]->vc_id);
cr_assert_eq(vc_chk, vc[1]);
ret = _gnix_vc_destroy(vc[0]);
cr_assert_eq(ret, FI_SUCCESS);
ret = _gnix_vc_destroy(vc[1]);
cr_assert_eq(ret, FI_SUCCESS);
}
Test(vc_management_auto, vc_alloc_simple)
{
int ret;
struct gnix_vc *vc[2];
struct gnix_fid_ep *ep_priv;
ep_priv = container_of(ep[0], struct gnix_fid_ep, ep_fid);
ret = _gnix_vc_alloc(ep_priv, gnix_addr[0], &vc[0]);
cr_assert_eq(ret, FI_SUCCESS);
ret = _gnix_vc_alloc(ep_priv, gnix_addr[1], &vc[1]);
cr_assert_eq(ret, FI_SUCCESS);
/*
* vc_id's have to be different since the
* vc's were allocated using the same ep.
*/
cr_assert_neq(vc[0]->vc_id, vc[1]->vc_id);
ret = _gnix_vc_destroy(vc[0]);
cr_assert_eq(ret, FI_SUCCESS);
ret = _gnix_vc_destroy(vc[1]);
cr_assert_eq(ret, FI_SUCCESS);
}
static int gnix_ep_control(fid_t fid, int command, void *arg)
{
int i, ret = FI_SUCCESS;
struct gnix_fid_ep *ep;
struct gnix_fid_domain *dom;
struct gnix_vc *vc;
ep = container_of(fid, struct gnix_fid_ep, ep_fid);
switch (command) {
/*
* for FI_EP_RDM, post wc datagrams now
*/
case FI_ENABLE:
if (ep->type == FI_EP_RDM) {
dom = ep->domain;
for (i = 0; i < dom->fabric->n_wc_dgrams; i++) {
assert(ep->recv_cq != NULL);
ret = _gnix_vc_alloc(ep, FI_ADDR_UNSPEC, &vc);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_vc_alloc call returned %d\n", ret);
goto err;
}
ret = _gnix_vc_accept(vc);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_vc_accept returned %d\n",
ret);
_gnix_vc_destroy(vc);
goto err;
} else {
fastlock_acquire(&ep->vc_list_lock);
dlist_insert_tail(&vc->entry,
&ep->wc_vc_list);
fastlock_release(&ep->vc_list_lock);
}
}
}
break;
case FI_GETFIDFLAG:
case FI_SETFIDFLAG:
case FI_ALIAS:
default:
return -FI_ENOSYS;
}
err:
return ret;
}
Test(gnix_cancel, cancel_ep_send)
{
int ret;
struct gnix_fid_ep *gnix_ep;
struct gnix_fab_req *req;
struct fi_cq_err_entry buf;
struct gnix_vc *vc;
void *foobar_ptr = NULL;
gnix_ht_key_t *key;
/* simulate a posted request */
gnix_ep = container_of(ep[0], struct gnix_fid_ep, ep_fid);
req = _gnix_fr_alloc(gnix_ep);
req->msg.send_info[0].send_addr = 0xdeadbeef;
req->msg.cum_send_len = req->msg.send_info[0].send_len = 128;
req->user_context = foobar_ptr;
req->type = GNIX_FAB_RQ_SEND;
/* allocate, store vc */
ret = _gnix_vc_alloc(gnix_ep, NULL, &vc);
cr_assert(ret == FI_SUCCESS, "_gnix_vc_alloc failed");
key = (gnix_ht_key_t *)&gnix_ep->my_name.gnix_addr;
ret = _gnix_ht_insert(gnix_ep->vc_ht, *key, vc);
cr_assert(!ret);
/* make a dummy request */
fastlock_acquire(&vc->tx_queue_lock);
dlist_insert_head(&req->dlist, &vc->tx_queue);
fastlock_release(&vc->tx_queue_lock);
/* cancel simulated request */
ret = fi_cancel(&ep[0]->fid, foobar_ptr);
cr_assert(ret == FI_SUCCESS, "fi_cancel failed");
/* check for event */
ret = fi_cq_readerr(msg_cq[0], &buf, FI_SEND);
cr_assert(ret == 1, "did not find one error event");
cr_assert(buf.buf == (void *) 0xdeadbeef, "buffer mismatch");
cr_assert(buf.data == 0, "data mismatch");
cr_assert(buf.err == FI_ECANCELED, "error code mismatch");
cr_assert(buf.prov_errno == FI_ECANCELED, "prov error code mismatch");
cr_assert(buf.len == 128, "length mismatch");
}
Test(vc_management_auto, vc_connect)
{
int ret;
struct gnix_vc *vc_conn;
struct gnix_fid_ep *ep_priv[2];
gnix_ht_key_t key;
enum gnix_vc_conn_state state;
ep_priv[0] = container_of(ep[0], struct gnix_fid_ep, ep_fid);
ep_priv[1] = container_of(ep[1], struct gnix_fid_ep, ep_fid);
ret = _gnix_vc_alloc(ep_priv[0], gnix_addr[1], &vc_conn);
cr_assert_eq(ret, FI_SUCCESS);
memcpy(&key, &gni_addr[1],
sizeof(gnix_ht_key_t));
ret = _gnix_ht_insert(ep_priv[0]->vc_ht, key, vc_conn);
cr_assert_eq(ret, FI_SUCCESS);
vc_conn->modes |= GNIX_VC_MODE_IN_HT;
ret = _gnix_vc_connect(vc_conn);
cr_assert_eq(ret, FI_SUCCESS);
/*
* since we asked for FI_PROGRESS_AUTO for control
* we can just spin here. add a yield in case the
* test is only being run on one cpu.
*/
state = GNIX_VC_CONN_NONE;
while (state != GNIX_VC_CONNECTED) {
pthread_yield();
state = _gnix_vc_state(vc_conn);
}
ret = _gnix_vc_disconnect(vc_conn);
cr_assert_eq(ret, FI_SUCCESS);
/* VC is destroyed by the EP */
}
DIRECT_FN STATIC int gnix_accept(struct fid_ep *ep, const void *param,
size_t paramlen)
{
int ret;
struct gnix_vc *vc;
struct gnix_fid_ep *ep_priv;
struct gnix_pep_sock_conn *conn;
struct gnix_pep_sock_connresp resp;
struct fi_eq_cm_entry eq_entry, *eqe_ptr;
struct gnix_mbox *mbox = NULL;
struct gnix_av_addr_entry av_entry;
if (!ep || (paramlen && !param) || paramlen > GNIX_CM_DATA_MAX_SIZE)
return -FI_EINVAL;
ep_priv = container_of(ep, struct gnix_fid_ep, ep_fid.fid);
COND_ACQUIRE(ep_priv->requires_lock, &ep_priv->vc_lock);
/* Look up and unpack the connection request used to create this EP. */
conn = (struct gnix_pep_sock_conn *)ep_priv->info->handle;
if (!conn || conn->fid.fclass != FI_CLASS_CONNREQ) {
ret = -FI_EINVAL;
goto err_unlock;
}
/* Create new VC without CM data. */
av_entry.gnix_addr = ep_priv->dest_addr.gnix_addr;
av_entry.cm_nic_cdm_id = ep_priv->dest_addr.cm_nic_cdm_id;
ret = _gnix_vc_alloc(ep_priv, &av_entry, &vc);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL, "Failed to create VC: %d\n", ret);
goto err_unlock;
}
ep_priv->vc = vc;
ep_priv->vc->peer_caps = conn->req.peer_caps;
ep_priv->vc->peer_id = conn->req.vc_id;
ret = _gnix_mbox_alloc(vc->ep->nic->mbox_hndl, &mbox);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_DATA, "_gnix_mbox_alloc returned %s\n",
fi_strerror(-ret));
goto err_mbox_alloc;
}
vc->smsg_mbox = mbox;
/* Initialize the GNI connection. */
ret = _gnix_vc_smsg_init(vc, conn->req.vc_id,
&conn->req.vc_mbox_attr,
&conn->req.cq_irq_mdh);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_vc_smsg_init returned %s\n",
fi_strerror(-ret));
goto err_smsg_init;
}
vc->conn_state = GNIX_VC_CONNECTED;
/* Send ACK with VC attrs to allow peer to initialize GNI connection. */
resp.cmd = GNIX_PEP_SOCK_RESP_ACCEPT;
resp.vc_id = vc->vc_id;
resp.vc_mbox_attr.msg_type = GNI_SMSG_TYPE_MBOX_AUTO_RETRANSMIT;
resp.vc_mbox_attr.msg_buffer = mbox->base;
resp.vc_mbox_attr.buff_size = vc->ep->nic->mem_per_mbox;
resp.vc_mbox_attr.mem_hndl = *mbox->memory_handle;
resp.vc_mbox_attr.mbox_offset = (uint64_t)mbox->offset;
resp.vc_mbox_attr.mbox_maxcredit =
ep_priv->domain->params.mbox_maxcredit;
resp.vc_mbox_attr.msg_maxsize =
ep_priv->domain->params.mbox_msg_maxsize;
resp.cq_irq_mdh = ep_priv->nic->irq_mem_hndl;
resp.peer_caps = ep_priv->caps;
resp.cm_data_len = paramlen;
if (paramlen) {
eqe_ptr = (struct fi_eq_cm_entry *)resp.eqe_buf;
memcpy(eqe_ptr->data, param, paramlen);
}
ret = write(conn->sock_fd, &resp, sizeof(resp));
if (ret != sizeof(resp)) {
GNIX_WARN(FI_LOG_EP_CTRL,
"Failed to send resp, errno: %d\n",
errno);
ret = -FI_EIO;
goto err_write;
}
/* Notify user that this side is connected. */
eq_entry.fid = &ep_priv->ep_fid.fid;
ret = fi_eq_write(&ep_priv->eq->eq_fid, FI_CONNECTED, &eq_entry,
sizeof(eq_entry), 0);
if (ret != sizeof(eq_entry)) {
GNIX_WARN(FI_LOG_EP_CTRL,
"fi_eq_write failed, err: %d\n", ret);
goto err_eq_write;
}
/* Free the connection request. */
free(conn);
ep_priv->conn_state = GNIX_EP_CONNECTED;
COND_RELEASE(ep_priv->requires_lock, &ep_priv->vc_lock);
GNIX_DEBUG(FI_LOG_EP_CTRL, "Sent conn accept: %p\n", ep_priv);
return FI_SUCCESS;
err_eq_write:
err_write:
err_smsg_init:
_gnix_mbox_free(ep_priv->vc->smsg_mbox);
ep_priv->vc->smsg_mbox = NULL;
err_mbox_alloc:
_gnix_vc_destroy(ep_priv->vc);
ep_priv->vc = NULL;
err_unlock:
COND_RELEASE(ep_priv->requires_lock, &ep_priv->vc_lock);
return ret;
}
DIRECT_FN STATIC int gnix_connect(struct fid_ep *ep, const void *addr,
const void *param, size_t paramlen)
{
int ret;
struct gnix_fid_ep *ep_priv;
struct sockaddr_in saddr;
struct gnix_pep_sock_connreq req;
struct fi_eq_cm_entry *eqe_ptr;
struct gnix_vc *vc;
struct gnix_mbox *mbox = NULL;
struct gnix_av_addr_entry av_entry;
if (!ep || !addr || (paramlen && !param) ||
paramlen > GNIX_CM_DATA_MAX_SIZE)
return -FI_EINVAL;
ep_priv = container_of(ep, struct gnix_fid_ep, ep_fid.fid);
COND_ACQUIRE(ep_priv->requires_lock, &ep_priv->vc_lock);
if (ep_priv->conn_state != GNIX_EP_UNCONNECTED) {
ret = -FI_EINVAL;
goto err_unlock;
}
ret = _gnix_pe_to_ip(addr, &saddr);
if (ret != FI_SUCCESS) {
GNIX_INFO(FI_LOG_EP_CTRL,
"Failed to translate gnix_ep_name to IP\n");
goto err_unlock;
}
/* Create new VC without CM data. */
av_entry.gnix_addr = ep_priv->dest_addr.gnix_addr;
av_entry.cm_nic_cdm_id = ep_priv->dest_addr.cm_nic_cdm_id;
ret = _gnix_vc_alloc(ep_priv, &av_entry, &vc);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"Failed to create VC:: %d\n",
ret);
goto err_unlock;
}
ep_priv->vc = vc;
ret = _gnix_mbox_alloc(vc->ep->nic->mbox_hndl, &mbox);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_mbox_alloc returned %s\n",
fi_strerror(-ret));
goto err_mbox_alloc;
}
vc->smsg_mbox = mbox;
ep_priv->conn_fd = socket(AF_INET, SOCK_STREAM, 0);
if (ep_priv->conn_fd < 0) {
GNIX_WARN(FI_LOG_EP_CTRL,
"Failed to create connect socket, errno: %d\n",
errno);
ret = -FI_ENOSPC;
goto err_socket;
}
/* Currently blocks until connected. */
ret = connect(ep_priv->conn_fd, (struct sockaddr *)&saddr,
sizeof(saddr));
if (ret) {
GNIX_WARN(FI_LOG_EP_CTRL,
"Failed to connect, errno: %d\n",
errno);
ret = -FI_EIO;
goto err_connect;
}
req.info = *ep_priv->info;
/* Note addrs are swapped. */
memcpy(&req.dest_addr, (void *)&ep_priv->src_addr,
sizeof(req.dest_addr));
memcpy(&ep_priv->dest_addr, addr, sizeof(ep_priv->dest_addr));
memcpy(&req.src_addr, addr, sizeof(req.src_addr));
if (ep_priv->info->tx_attr)
req.tx_attr = *ep_priv->info->tx_attr;
if (ep_priv->info->rx_attr)
req.rx_attr = *ep_priv->info->rx_attr;
if (ep_priv->info->ep_attr)
req.ep_attr = *ep_priv->info->ep_attr;
if (ep_priv->info->domain_attr)
req.domain_attr = *ep_priv->info->domain_attr;
if (ep_priv->info->fabric_attr)
req.fabric_attr = *ep_priv->info->fabric_attr;
req.fabric_attr.fabric = NULL;
req.domain_attr.domain = NULL;
req.vc_id = vc->vc_id;
req.vc_mbox_attr.msg_type = GNI_SMSG_TYPE_MBOX_AUTO_RETRANSMIT;
req.vc_mbox_attr.msg_buffer = mbox->base;
req.vc_mbox_attr.buff_size = vc->ep->nic->mem_per_mbox;
req.vc_mbox_attr.mem_hndl = *mbox->memory_handle;
req.vc_mbox_attr.mbox_offset = (uint64_t)mbox->offset;
req.vc_mbox_attr.mbox_maxcredit =
ep_priv->domain->params.mbox_maxcredit;
req.vc_mbox_attr.msg_maxsize = ep_priv->domain->params.mbox_msg_maxsize;
req.cq_irq_mdh = ep_priv->nic->irq_mem_hndl;
req.peer_caps = ep_priv->caps;
req.cm_data_len = paramlen;
if (paramlen) {
eqe_ptr = (struct fi_eq_cm_entry *)req.eqe_buf;
memcpy(eqe_ptr->data, param, paramlen);
}
ret = write(ep_priv->conn_fd, &req, sizeof(req));
if (ret != sizeof(req)) {
GNIX_WARN(FI_LOG_EP_CTRL,
"Failed to send req, errno: %d\n",
errno);
ret = -FI_EIO;
goto err_write;
}
/* set fd to non-blocking now since we can't block within the eq
* progress system
*/
fi_fd_nonblock(ep_priv->conn_fd);
ep_priv->conn_state = GNIX_EP_CONNECTING;
COND_RELEASE(ep_priv->requires_lock, &ep_priv->vc_lock);
GNIX_DEBUG(FI_LOG_EP_CTRL, "Sent conn req: %p, %s\n",
ep_priv, inet_ntoa(saddr.sin_addr));
return FI_SUCCESS;
err_write:
err_connect:
close(ep_priv->conn_fd);
ep_priv->conn_fd = -1;
err_socket:
_gnix_mbox_free(ep_priv->vc->smsg_mbox);
ep_priv->vc->smsg_mbox = NULL;
err_mbox_alloc:
_gnix_vc_destroy(ep_priv->vc);
ep_priv->vc = NULL;
err_unlock:
COND_RELEASE(ep_priv->requires_lock, &ep_priv->vc_lock);
return ret;
}
static int __gnix_vc_hndl_conn_req(struct gnix_cm_nic *cm_nic,
char *msg_buffer,
struct gnix_address src_cm_nic_addr)
{
int ret = FI_SUCCESS;
gni_return_t __attribute__((unused)) status;
struct gnix_fid_ep *ep = NULL;
gnix_ht_key_t *key_ptr;
struct gnix_av_addr_entry entry;
struct gnix_address src_addr, target_addr;
struct gnix_vc *vc = NULL;
struct gnix_vc *vc_try = NULL;
struct gnix_work_req *work_req;
int src_vc_id;
gni_smsg_attr_t src_smsg_attr;
uint64_t src_vc_ptr;
struct wq_hndl_conn_req *data = NULL;
ssize_t __attribute__((unused)) len;
GNIX_TRACE(FI_LOG_EP_CTRL, "\n");
/*
* unpack the message
*/
__gnix_vc_unpack_conn_req(msg_buffer,
&target_addr,
&src_addr,
&src_vc_id,
&src_vc_ptr,
&src_smsg_attr);
GNIX_DEBUG(FI_LOG_EP_CTRL,
"conn req rx: (From Aries addr 0x%x Id %d to Aries 0x%x Id %d src vc 0x%lx )\n",
src_addr.device_addr,
src_addr.cdm_id,
target_addr.device_addr,
target_addr.cdm_id,
src_vc_ptr);
/*
* lookup the ep from the addr_to_ep_ht using the target_addr
* in the datagram
*/
key_ptr = (gnix_ht_key_t *)&target_addr;
ep = (struct gnix_fid_ep *)_gnix_ht_lookup(cm_nic->addr_to_ep_ht,
*key_ptr);
if (ep == NULL) {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_ht_lookup addr_to_ep failed\n");
ret = -FI_ENOENT;
goto err;
}
/*
* look to see if there is a VC already for the
* address of the connecting EP.
*/
key_ptr = (gnix_ht_key_t *)&src_addr;
fastlock_acquire(&ep->vc_ht_lock);
vc = (struct gnix_vc *)_gnix_ht_lookup(ep->vc_ht,
*key_ptr);
/*
* if there is no corresponding vc in the hash,
* or there is an entry and its not in connecting state
* go down the conn req ack route.
*/
if ((vc == NULL) ||
(vc->conn_state == GNIX_VC_CONN_NONE)) {
if (vc == NULL) {
entry.gnix_addr = src_addr;
entry.cm_nic_cdm_id = src_cm_nic_addr.cdm_id;
ret = _gnix_vc_alloc(ep,
&entry,
&vc_try);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_vc_alloc returned %s\n",
fi_strerror(-ret));
goto err;
}
vc_try->conn_state = GNIX_VC_CONNECTING;
ret = _gnix_ht_insert(ep->vc_ht,
*key_ptr,
vc_try);
if (likely(ret == FI_SUCCESS)) {
vc = vc_try;
vc->modes |= GNIX_VC_MODE_IN_HT;
} else if (ret == -FI_ENOSPC) {
_gnix_vc_destroy(vc_try);
} else {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_ht_insert returned %s\n",
fi_strerror(-ret));
goto err;
}
} else
vc->conn_state = GNIX_VC_CONNECTING;
/*
* prepare a work request to
* initiate an request response
*/
work_req = calloc(1, sizeof(*work_req));
if (work_req == NULL) {
ret = -FI_ENOMEM;
goto err;
}
data = calloc(1, sizeof(struct wq_hndl_conn_req));
if (data == NULL) {
ret = -FI_ENOMEM;
goto err;
}
memcpy(&data->src_smsg_attr,
&src_smsg_attr,
sizeof(src_smsg_attr));
data->vc = vc;
data->src_vc_id = src_vc_id;
data->src_vc_ptr = src_vc_ptr;
work_req->progress_fn = __gnix_vc_conn_ack_prog_fn;
work_req->data = data;
work_req->completer_fn = __gnix_vc_conn_ack_comp_fn;
work_req->completer_data = data;
/*
* add the work request to the tail of the
* cm_nic's work queue, progress the cm_nic.
*/
fastlock_acquire(&cm_nic->wq_lock);
dlist_insert_before(&work_req->list, &cm_nic->cm_nic_wq);
fastlock_release(&cm_nic->wq_lock);
fastlock_release(&ep->vc_ht_lock);
_gnix_vc_schedule(vc);
ret = _gnix_cm_nic_progress(cm_nic);
if (ret != FI_SUCCESS)
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_cm_nic_progress returned %s\n",
fi_strerror(-ret));
} else {
/*
* we can only be in connecting state if we
* reach here. We have all the informatinon,
* and the other side will get the information
* at some point, so go ahead and build SMSG connection.
*/
if (vc->conn_state != GNIX_VC_CONNECTING) {
GNIX_WARN(FI_LOG_EP_CTRL,
"vc %p not in connecting state nor in cm wq\n",
vc, vc->conn_state);
ret = -FI_EINVAL;
goto err;
}
ret = __gnix_vc_smsg_init(vc, src_vc_id,
&src_smsg_attr);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_vc_smsg_init returned %s\n",
fi_strerror(-ret));
goto err;
}
vc->conn_state = GNIX_VC_CONNECTED;
GNIX_DEBUG(FI_LOG_EP_CTRL, "moving vc %p state to connected\n",
vc);
fastlock_release(&ep->vc_ht_lock);
ret = _gnix_vc_schedule(vc);
ret = _gnix_cm_nic_progress(cm_nic);
if (ret != FI_SUCCESS)
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_cm_nic_progress returned %s\n",
fi_strerror(-ret));
}
err:
return ret;
}
/*
* connect to self, since we use a lock here
* the only case we need to deal with is one
* vc connect request with the peer vc not yet
* being set up
*/
static int __gnix_vc_connect_to_same_cm_nic(struct gnix_vc *vc)
{
int ret = FI_SUCCESS;
struct gnix_fid_domain *dom = NULL;
struct gnix_fid_ep *ep = NULL;
struct gnix_fid_ep *ep_peer = NULL;
struct gnix_cm_nic *cm_nic = NULL;
struct gnix_mbox *mbox = NULL, *mbox_peer = NULL;
struct gnix_vc *vc_peer;
gni_smsg_attr_t smsg_mbox_attr;
gni_smsg_attr_t smsg_mbox_attr_peer;
gnix_ht_key_t *key_ptr;
struct gnix_av_addr_entry entry;
GNIX_TRACE(FI_LOG_EP_CTRL, "\n");
ep = vc->ep;
if (ep == NULL)
return -FI_EINVAL;
cm_nic = ep->cm_nic;
if (cm_nic == NULL) {
ret = -FI_EINVAL;
goto exit;
}
dom = ep->domain;
if (dom == NULL) {
ret = -FI_EINVAL;
goto exit;
}
fastlock_acquire(&ep->vc_ht_lock);
if ((vc->conn_state == GNIX_VC_CONNECTING) ||
(vc->conn_state == GNIX_VC_CONNECTED)) {
fastlock_release(&ep->vc_ht_lock);
return FI_SUCCESS;
} else
vc->conn_state = GNIX_VC_CONNECTING;
fastlock_release(&ep->vc_ht_lock);
GNIX_DEBUG(FI_LOG_EP_CTRL, "moving vc %p state to connecting\n", vc);
if (vc->smsg_mbox == NULL) {
ret = _gnix_mbox_alloc(vc->ep->nic->mbox_hndl,
&mbox);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_mbox_alloc returned %s\n",
fi_strerror(-ret));
goto exit;
}
vc->smsg_mbox = mbox;
} else
mbox = vc->smsg_mbox;
smsg_mbox_attr.msg_type = GNI_SMSG_TYPE_MBOX_AUTO_RETRANSMIT;
smsg_mbox_attr.msg_buffer = mbox->base;
smsg_mbox_attr.buff_size = vc->ep->nic->mem_per_mbox;
smsg_mbox_attr.mem_hndl = *mbox->memory_handle;
smsg_mbox_attr.mbox_offset = (uint64_t)mbox->offset;
smsg_mbox_attr.mbox_maxcredit = dom->params.mbox_maxcredit;
smsg_mbox_attr.msg_maxsize = dom->params.mbox_msg_maxsize;
key_ptr = (gnix_ht_key_t *)&vc->peer_addr;
ep_peer = (struct gnix_fid_ep *)_gnix_ht_lookup(cm_nic->addr_to_ep_ht,
*key_ptr);
if (ep_peer == NULL) {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_ht_lookup addr_to_ep failed\n");
ret = -FI_ENOENT;
goto exit;
}
key_ptr = (gnix_ht_key_t *)&ep->my_name.gnix_addr;
fastlock_acquire(&ep_peer->vc_ht_lock);
vc_peer = (struct gnix_vc *)_gnix_ht_lookup(ep_peer->vc_ht,
*key_ptr);
/*
* handle the special case of connecting to self
*/
if (vc_peer == vc) {
ret = __gnix_vc_smsg_init(vc, vc->vc_id, &smsg_mbox_attr);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_vc_smsg_init returned %s\n",
fi_strerror(-ret));
goto exit_w_lock;
}
vc->conn_state = GNIX_VC_CONNECTED;
GNIX_DEBUG(FI_LOG_EP_CTRL, "moving vc %p state to connected\n",
vc);
goto exit_w_lock;
}
if ((vc_peer != NULL) &&
(vc_peer->conn_state != GNIX_VC_CONN_NONE)) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_vc_connect self, vc_peer in inconsistent state\n");
ret = -FI_ENOSPC;
goto exit_w_lock;
}
if (vc_peer == NULL) {
entry.gnix_addr = ep->my_name.gnix_addr;
entry.cm_nic_cdm_id = ep->my_name.cm_nic_cdm_id;
ret = _gnix_vc_alloc(ep_peer,
&entry,
&vc_peer);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_vc_alloc returned %s\n",
fi_strerror(-ret));
goto exit_w_lock;
}
ret = _gnix_ht_insert(ep_peer->vc_ht,
*key_ptr,
vc_peer);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_ht_insert returned %s\n",
fi_strerror(-ret));
goto exit_w_lock;
}
vc_peer->modes |= GNIX_VC_MODE_IN_HT;
}
vc_peer->conn_state = GNIX_VC_CONNECTING;
GNIX_DEBUG(FI_LOG_EP_CTRL, "moving vc %p state to connecting\n",
vc_peer);
if (vc_peer->smsg_mbox == NULL) {
ret = _gnix_mbox_alloc(vc_peer->ep->nic->mbox_hndl,
&mbox_peer);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_mbox_alloc returned %s\n",
fi_strerror(-ret));
goto exit_w_lock;
}
vc_peer->smsg_mbox = mbox_peer;
} else
mbox_peer = vc_peer->smsg_mbox;
smsg_mbox_attr_peer.msg_type = GNI_SMSG_TYPE_MBOX_AUTO_RETRANSMIT;
smsg_mbox_attr_peer.msg_buffer = mbox_peer->base;
smsg_mbox_attr_peer.buff_size = vc_peer->ep->nic->mem_per_mbox;
smsg_mbox_attr_peer.mem_hndl = *mbox_peer->memory_handle;
smsg_mbox_attr_peer.mbox_offset = (uint64_t)mbox_peer->offset;
smsg_mbox_attr_peer.mbox_maxcredit = dom->params.mbox_maxcredit;
smsg_mbox_attr_peer.msg_maxsize = dom->params.mbox_msg_maxsize;
ret = __gnix_vc_smsg_init(vc, vc_peer->vc_id, &smsg_mbox_attr_peer);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_vc_smsg_init returned %s\n",
fi_strerror(-ret));
goto exit_w_lock;
}
ret = __gnix_vc_smsg_init(vc_peer, vc->vc_id, &smsg_mbox_attr);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_vc_smsg_init returned %s\n",
fi_strerror(-ret));
goto exit_w_lock;
}
vc->conn_state = GNIX_VC_CONNECTED;
GNIX_DEBUG(FI_LOG_EP_CTRL, "moving vc %p state to connected\n",
vc);
vc_peer->conn_state = GNIX_VC_CONNECTED;
GNIX_DEBUG(FI_LOG_EP_CTRL, "moving vc %p state to connected\n",
vc_peer);
exit_w_lock:
fastlock_release(&ep_peer->vc_ht_lock);
exit:
return ret;
}