void Acceptor::listen(std::string const& hostname, std::string const& port) {
auto res = create_addr_info(hostname, port);
ibv_qp_init_attr init_attr;
memset(&init_attr, 0, sizeof(init_attr));
init_attr.cap.max_send_wr = m_credits;
init_attr.cap.max_recv_wr = m_credits;
init_attr.cap.max_send_sge = 1;
init_attr.cap.max_recv_sge = 1;
init_attr.cap.max_inline_data = 0;
init_attr.sq_sig_all = 1;
init_attr.qp_type = IBV_QPT_RC;
int ret = rdma_create_ep(&m_cm_id, res, NULL, &init_attr);
destroy_addr_info(res);
if (ret) {
rdma_destroy_ep(m_cm_id);
throw exception::acceptor::generic_error(
"Error on rdma_create_ep: " + std::string(strerror(errno)));
}
if (rdma_listen(m_cm_id, 128)) {
rdma_destroy_ep(m_cm_id);
throw exception::acceptor::generic_error(
"Error on rdma_listen: " + std::string(strerror(errno)));
}
}
std::unique_ptr<Socket> Acceptor::accept() {
rdma_cm_id* new_cm_id;
if (rdma_get_request(m_cm_id, &new_cm_id)) {
throw exception::acceptor::generic_error(
"Error on rdma_get_request: " + std::string(strerror(errno)));
}
rdma_conn_param conn_param;
memset(&conn_param, 0, sizeof(rdma_conn_param));
conn_param.rnr_retry_count = m_rnr_retry_count;
if (rdma_accept(new_cm_id, &conn_param)) {
rdma_destroy_ep(new_cm_id);
throw exception::acceptor::generic_error(
"Error on rdma_accept: " + std::string(strerror(errno)));
}
ibv_qp_attr attr;
memset(&attr, 0, sizeof(ibv_qp_attr));
attr.min_rnr_timer = m_min_rtr_timer;
int flags = IBV_QP_MIN_RNR_TIMER;
if (ibv_modify_qp(new_cm_id->qp, &attr, flags)) {
rdma_destroy_ep(new_cm_id);
throw exception::acceptor::generic_error(
"Error on ibv_modify_qp: " + std::string(strerror(errno)));
}
std::unique_ptr<Socket> socket_ptr(new Socket(new_cm_id, m_credits));
return socket_ptr;
}
int fi_ibv_getinfo(uint32_t version, const char *node, const char *service,
uint64_t flags, struct fi_info *hints, struct fi_info **info)
{
struct rdma_cm_id *id;
struct rdma_addrinfo *rai;
int ret;
ret = fi_ibv_init_info();
if (ret)
goto out;
ret = fi_ibv_create_ep(node, service, flags, hints, &rai, &id);
if (ret)
goto out;
if (id->verbs) {
ret = fi_ibv_get_matching_info(ibv_get_device_name(id->verbs->device),
hints, rai, info);
} else {
ret = fi_ibv_get_matching_info(NULL, hints, rai, info);
}
rdma_destroy_ep(id);
rdma_freeaddrinfo(rai);
out:
if (!ret || ret == -FI_ENOMEM)
return ret;
else
return -FI_ENODATA;
}
static void fi_ibv_free_msg_ep(struct fi_ibv_msg_ep *ep)
{
if (ep->id)
rdma_destroy_ep(ep->id);
fi_freeinfo(ep->info);
free(ep);
}
RDMACMSocket::~RDMACMSocket() {
rdma_disconnect(this->client_id);
rdma_dereg_mr(this->verbs_mr);
rdma_destroy_ep(this->client_id);
this->verbs_buf.free();
}
void RDMACMSocket::post_recv(const Buffer& buf) {
if (rdma_post_recv(this->client_id, buf.addr, buf.addr, buf.size, this->verbs_mr) < 0) {
rdma_dereg_mr(this->verbs_mr);
this->verbs_buf.free();
rdma_destroy_ep(this->client_id);
perror("rdma_post_recv");
exit(1);
}
}
static int __fi_pep_close(fid_t fid)
{
struct __fid_pep *pep;
pep = container_of(fid, struct __fid_pep, pep_fid.fid);
if (pep->id)
rdma_destroy_ep(pep->id);
free(pep);
return 0;
}
void
kiro_server_stop (KiroServer *self)
{
g_return_if_fail (self != NULL);
KiroServerPrivate *priv = KIRO_SERVER_GET_PRIVATE (self);
if (!priv->base)
return;
//Shut down event listening
priv->close_signal = TRUE;
g_debug ("Event handling stopped");
g_list_foreach (priv->clients, disconnect_client, NULL);
g_list_free (priv->clients);
// Stop the main loop and clear its memory
g_main_loop_quit (priv->main_loop);
g_main_loop_unref (priv->main_loop);
priv->main_loop = NULL;
// Ask the main thread to join (It probably already has, but we do it
// anyways. Just in case!)
g_thread_join (priv->main_thread);
priv->main_thread = NULL;
// We don't need the connection management IO channel container any more.
// Unref and thus free it.
g_io_channel_unref (priv->conn_ec);
priv->conn_ec = NULL;
priv->close_signal = FALSE;
// kiro_destroy_connection would try to call rdma_disconnect on the given
// connection. But the server never 'connects' to anywhere, so this would
// cause a crash. We need to destroy the enpoint manually without disconnect
struct kiro_connection_context *ctx = (struct kiro_connection_context *) (priv->base->context);
kiro_destroy_connection_context (&ctx);
rdma_destroy_ep (priv->base);
priv->base = NULL;
rdma_destroy_event_channel (priv->ec);
priv->ec = NULL;
g_message ("Server stopped successfully");
}
int fi_ibv_create_ep(const char *node, const char *service,
uint64_t flags, const struct fi_info *hints,
struct rdma_addrinfo **rai, struct rdma_cm_id **id)
{
struct rdma_addrinfo *_rai;
struct sockaddr *local_addr;
int ret;
ret = fi_ibv_get_rdma_rai(node, service, flags, hints, &_rai);
if (ret) {
return ret;
}
ret = rdma_create_ep(id, _rai, NULL, NULL);
if (ret) {
VERBS_INFO_ERRNO(FI_LOG_FABRIC, "rdma_create_ep", errno);
ret = -errno;
goto err1;
}
if (rai && !_rai->ai_src_addr) {
local_addr = rdma_get_local_addr(*id);
_rai->ai_src_len = fi_ibv_sockaddr_len(local_addr);
if (!(_rai->ai_src_addr = malloc(_rai->ai_src_len))) {
ret = -FI_ENOMEM;
goto err2;
}
memcpy(_rai->ai_src_addr, local_addr, _rai->ai_src_len);
}
if (rai) {
*rai = _rai;
} else {
rdma_freeaddrinfo(_rai);
}
return ret;
err2:
rdma_destroy_ep(*id);
err1:
rdma_freeaddrinfo(_rai);
return ret;
}
RDMACMSocket* RDMACMSocket::connect(const HostAndPort& hp) {
struct rdma_cm_id* client_id = NULL;
struct rdma_addrinfo hints, *res;
memset(&hints, 0, sizeof(hints));
hints.ai_port_space = RDMA_PS_TCP;
res = NULL;
char* hostname = const_cast<char*>(hp.hostname);
char* port_str = const_cast<char*>(hp.port_str);
if (rdma_getaddrinfo(hostname, port_str, &hints, &res) < 0) {
perror("rdma_getaddrinfo");
exit(1);
}
struct ibv_qp_init_attr attr;
memset(&attr, 0, sizeof(attr));
attr.cap.max_send_wr = PACKET_WINDOW_SIZE;
attr.cap.max_recv_wr = PACKET_WINDOW_SIZE;
attr.cap.max_send_sge = 1;
attr.cap.max_recv_sge = 1;
attr.cap.max_inline_data = 0;
attr.sq_sig_all = 1;
if (rdma_create_ep(&client_id, res, NULL, &attr) < 0) {
rdma_freeaddrinfo(res);
perror("rdma_create_ep");
exit(1);
}
rdma_freeaddrinfo(res);
if (rdma_connect(client_id, NULL) < 0) {
rdma_destroy_ep(client_id);
perror("rdma_connect");
exit(1);
}
return new RDMACMSocket(client_id);
}
static int fi_ibv_msg_ep_setname(fid_t ep_fid, void *addr, size_t addrlen)
{
struct fi_ibv_msg_ep *ep;
void *save_addr;
struct rdma_cm_id *id;
int ret;
ep = container_of(ep_fid, struct fi_ibv_msg_ep, ep_fid);
if (addrlen != ep->info->src_addrlen) {
FI_INFO(&fi_ibv_prov, FI_LOG_EP_CTRL,"addrlen expected: %d, got: %d.\n",
ep->info->src_addrlen, addrlen);
return -FI_EINVAL;
}
save_addr = ep->info->src_addr;
ep->info->src_addr = malloc(ep->info->src_addrlen);
if (!ep->info->src_addr) {
ret = -FI_ENOMEM;
goto err1;
}
memcpy(ep->info->src_addr, addr, ep->info->src_addrlen);
ret = fi_ibv_create_ep(NULL, NULL, 0, ep->info, NULL, &id);
if (ret)
goto err2;
if (ep->id)
rdma_destroy_ep(ep->id);
ep->id = id;
free(save_addr);
return 0;
err2:
free(ep->info->src_addr);
err1:
ep->info->src_addr = save_addr;
return ret;
}
void RDMACMSocket::setup_verbs_buf() {
this->verbs_buf = Buffer::allocate(PACKET_SIZE * PACKET_WINDOW_SIZE * 2);
this->verbs_mr = rdma_reg_msgs(this->client_id, this->verbs_buf.addr, this->verbs_buf.size);
if (this->verbs_mr == NULL) {
this->verbs_buf.free();
rdma_destroy_ep(this->client_id);
perror("rdma_reg_msgs");
exit(1);
}
char* send_buf_begin = this->verbs_buf.addr + PACKET_SIZE * PACKET_WINDOW_SIZE;
for (int i = 0; i < PACKET_WINDOW_SIZE; ++i) {
Buffer recv_buf(this->verbs_buf.addr + i * PACKET_SIZE, PACKET_SIZE);
post_recv(recv_buf);
Buffer send_buf(send_buf_begin + i * PACKET_SIZE, PACKET_SIZE);
send_bufs.push_back(send_buf);
}
}
static int fi_ibv_copy_ifaddr(const char *name, const char *service, uint64_t flags,
struct fi_info *info)
{
struct rdma_addrinfo *rai;
struct fi_info *fi;
struct rdma_cm_id *id;
int ret;
ret = fi_ibv_get_rdma_rai(name, service, flags, NULL, &rai);
if (ret) {
FI_WARN(&fi_ibv_prov, FI_LOG_FABRIC,
"rdma_getaddrinfo failed for name:%s\n", name);
return ret;
}
ret = rdma_create_ep(&id, rai, NULL, NULL);
if (!ret) {
for (fi = info; fi; fi = fi->next)
if (!strncmp(id->verbs->device->name, fi->domain_attr->name,
strlen(id->verbs->device->name)))
break;
if (!fi) {
FI_WARN(&fi_ibv_prov, FI_LOG_FABRIC,
"No matching fi_info for device: "
"%s with address: %s\n",
id->verbs->device->name, name);
} else {
if (fi->src_addr) {
free(fi->src_addr);
fi->src_addr = NULL;
}
fi_ibv_rai_to_fi(rai, fi);
}
rdma_destroy_ep(id);
}
rdma_freeaddrinfo(rai);
return 0;
}
void fi_ibv_destroy_ep(struct rdma_addrinfo *rai, struct rdma_cm_id **id)
{
rdma_freeaddrinfo(rai);
rdma_destroy_ep(*id);
}
/* Builds a list of interfaces that correspond to active verbs devices */
static int fi_ibv_getifaddrs(struct dlist_entry *verbs_devs)
{
struct ifaddrs *ifaddr, *ifa;
char name[INET6_ADDRSTRLEN];
struct rdma_addrinfo *rai;
struct rdma_cm_id *id;
const char *ret_ptr;
int ret, num_verbs_ifs = 0;
char *iface = NULL;
size_t iface_len = 0;
int exact_match = 0;
ret = getifaddrs(&ifaddr);
if (ret) {
VERBS_WARN(FI_LOG_FABRIC,
"Unable to get interface addresses\n");
return ret;
}
/* select best iface name based on user's input */
if (fi_param_get_str(&fi_ibv_prov, "iface", &iface) == FI_SUCCESS) {
iface_len = strlen(iface);
if (iface_len > IFNAMSIZ) {
VERBS_INFO(FI_LOG_EP_CTRL,
"Too long iface name: %s, max: %d\n",
iface, IFNAMSIZ);
return -FI_EINVAL;
}
for (ifa = ifaddr; ifa && !exact_match; ifa = ifa->ifa_next)
exact_match = !strcmp(ifa->ifa_name, iface);
}
for (ifa = ifaddr; ifa; ifa = ifa->ifa_next) {
if (!ifa->ifa_addr || !(ifa->ifa_flags & IFF_UP) ||
!strcmp(ifa->ifa_name, "lo"))
continue;
if(iface) {
if(exact_match) {
if(strcmp(ifa->ifa_name, iface))
continue;
} else {
if(strncmp(ifa->ifa_name, iface, iface_len))
continue;
}
}
switch (ifa->ifa_addr->sa_family) {
case AF_INET:
ret_ptr = inet_ntop(AF_INET, &ofi_sin_addr(ifa->ifa_addr),
name, INET6_ADDRSTRLEN);
break;
case AF_INET6:
ret_ptr = inet_ntop(AF_INET6, &ofi_sin6_addr(ifa->ifa_addr),
name, INET6_ADDRSTRLEN);
break;
default:
continue;
}
if (!ret_ptr) {
VERBS_WARN(FI_LOG_FABRIC,
"inet_ntop failed: %s(%d)\n",
strerror(errno), errno);
goto err1;
}
ret = fi_ibv_create_ep(name, NULL, FI_NUMERICHOST | FI_SOURCE,
NULL, &rai, &id);
if (ret)
continue;
ret = fi_ibv_add_rai(verbs_devs, id, rai);
if (ret)
goto err2;
VERBS_DBG(FI_LOG_FABRIC, "Found active interface for verbs device: "
"%s with address: %s\n",
ibv_get_device_name(id->verbs->device), name);
rdma_destroy_ep(id);
num_verbs_ifs++;
}
freeifaddrs(ifaddr);
return num_verbs_ifs ? 0 : -FI_ENODATA;
err2:
rdma_destroy_ep(id);
err1:
fi_ibv_verbs_devs_free(verbs_devs);
freeifaddrs(ifaddr);
return ret;
}
static int run(void)
{
struct rdma_addrinfo hints, *res;
struct ibv_qp_init_attr attr;
struct ibv_wc wc;
int ret;
memset(&hints, 0, sizeof hints);
hints.ai_port_space = RDMA_PS_TCP;
ret = rdma_getaddrinfo(server, port, &hints, &res);
if (ret) {
printf("rdma_getaddrinfo %d\n", errno);
return ret;
}
memset(&attr, 0, sizeof attr);
attr.cap.max_send_wr = attr.cap.max_recv_wr = 1;
attr.cap.max_send_sge = attr.cap.max_recv_sge = 1;
attr.cap.max_inline_data = 16;
attr.qp_context = id;
attr.sq_sig_all = 1;
ret = rdma_create_ep(&id, res, NULL, &attr);
rdma_freeaddrinfo(res);
if (ret) {
printf("rdma_create_ep %d\n", errno);
return ret;
}
mr = rdma_reg_msgs(id, recv_msg, 16);
if (!mr) {
printf("rdma_reg_msgs %d\n", errno);
return ret;
}
ret = rdma_post_recv(id, NULL, recv_msg, 16, mr);
if (ret) {
printf("rdma_post_recv %d\n", errno);
return ret;
}
ret = rdma_connect(id, NULL);
if (ret) {
printf("rdma_connect %d\n", errno);
return ret;
}
s = get_dtime();
ret = rdma_post_send(id, NULL, send_msg, 16, NULL, IBV_SEND_INLINE);
if (ret) {
printf("rdma_post_send %d\n", errno);
return ret;
}
e = get_dtime();
ret = rdma_get_recv_comp(id, &wc);
if (ret <= 0) {
printf("rdma_get_recv_comp %d\n", ret);
return ret;
}
printf("time %f\n", e - s);
rdma_disconnect(id);
rdma_dereg_mr(mr);
rdma_destroy_ep(id);
return 0;
}
static int xrc_test(void)
{
struct rdma_cm_id *conn_id, *lookup_id;
struct ibv_qp_init_attr attr;
struct rdma_conn_param param;
struct rdma_cm_event *event;
struct ibv_wc wc;
int ret;
conn_id = xrc_listen_recv();
if (!conn_id)
return -1;
ret = xrc_create_srq_listen(rdma_get_local_addr(conn_id),
sizeof(struct sockaddr_storage));
if (ret)
return -1;
memset(&attr, 0, sizeof attr);
attr.qp_type = IBV_QPT_XRC_RECV;
attr.ext.xrc_recv.xrcd = srq_id->srq->ext.xrc.xrcd;
ret = rdma_create_qp(conn_id, NULL, &attr);
if (ret) {
printf("Unable to create xrc recv qp %d\n", errno);
return ret;
}
ret = rdma_accept(conn_id, NULL);
if (ret) {
printf("rdma_accept failed for xrc recv qp %d\n", errno);
return ret;
}
ret = rdma_get_request(srq_id, &lookup_id);
if (ret) {
printf("rdma_get_request %d\n", errno);
return ret;
}
mr = rdma_reg_msgs(srq_id, recv_msg, sizeof recv_msg);
if (!mr) {
printf("ibv_reg_msgs %d\n", errno);
return ret;
}
ret = rdma_post_recv(srq_id, NULL, recv_msg, sizeof recv_msg, mr);
if (ret) {
printf("rdma_post_recv %d\n", errno);
return ret;
}
memset(¶m, 0, sizeof param);
param.qp_num = srq_id->srq->ext.xrc.srq_num;
ret = rdma_accept(lookup_id, ¶m);
if (ret) {
printf("rdma_accept failed for srqn lookup %d\n", errno);
return ret;
}
rdma_destroy_id(lookup_id);
ret = rdma_get_recv_comp(srq_id, &wc);
if (ret <= 0) {
printf("rdma_get_recv_comp %d\n", ret);
return ret;
}
ret = rdma_get_cm_event(conn_id->channel, &event);
if (ret || event->event != RDMA_CM_EVENT_DISCONNECTED) {
printf("Failed to get disconnect event\n");
return -1;
}
rdma_ack_cm_event(event);
rdma_disconnect(conn_id);
rdma_destroy_ep(conn_id);
rdma_dereg_mr(mr);
rdma_destroy_ep(srq_id);
rdma_destroy_ep(listen_id);
return 0;
}
static int rc_test(void)
{
struct rdma_addrinfo hints, *res;
struct ibv_qp_init_attr attr;
struct ibv_wc wc;
int ret;
memset(&hints, 0, sizeof hints);
hints.ai_flags = RAI_PASSIVE;
hints.ai_port_space = RDMA_PS_TCP;
ret = rdma_getaddrinfo(NULL, port, &hints, &res);
if (ret) {
printf("rdma_getaddrinfo %d\n", errno);
return ret;
}
memset(&attr, 0, sizeof attr);
attr.cap.max_send_wr = attr.cap.max_recv_wr = 1;
attr.cap.max_send_sge = attr.cap.max_recv_sge = 1;
attr.cap.max_inline_data = sizeof send_msg;
attr.sq_sig_all = 1;
ret = rdma_create_ep(&listen_id, res, NULL, &attr);
rdma_freeaddrinfo(res);
if (ret) {
printf("rdma_create_ep %d\n", errno);
return ret;
}
ret = rdma_listen(listen_id, 0);
if (ret) {
printf("rdma_listen %d\n", errno);
return ret;
}
ret = rdma_get_request(listen_id, &id);
if (ret) {
printf("rdma_get_request %d\n", errno);
return ret;
}
mr = rdma_reg_msgs(id, recv_msg, sizeof recv_msg);
if (!mr) {
printf("rdma_reg_msgs %d\n", errno);
return ret;
}
ret = rdma_post_recv(id, NULL, recv_msg, sizeof recv_msg, mr);
if (ret) {
printf("rdma_post_recv %d\n", errno);
return ret;
}
ret = rdma_accept(id, NULL);
if (ret) {
printf("rdma_accept %d\n", errno);
return ret;
}
ret = rdma_get_recv_comp(id, &wc);
if (ret <= 0) {
printf("rdma_get_recv_comp %d\n", ret);
return ret;
}
ret = rdma_post_send(id, NULL, send_msg, sizeof send_msg, NULL, IBV_SEND_INLINE);
if (ret) {
printf("rdma_post_send %d\n", errno);
return ret;
}
ret = rdma_get_send_comp(id, &wc);
if (ret <= 0) {
printf("rdma_get_send_comp %d\n", ret);
return ret;
}
rdma_disconnect(id);
rdma_dereg_mr(mr);
rdma_destroy_ep(id);
rdma_destroy_ep(listen_id);
return 0;
}
static gboolean
process_rdma_event (GIOChannel *source, GIOCondition condition, gpointer data)
{
// Right now, we don't need 'source'
// Tell the compiler to ignore it by (void)-ing it
(void) source;
if (!G_TRYLOCK (rdma_handling)) {
g_debug ("RDMA handling will wait for the next dispatch.");
return TRUE;
}
g_debug ("Got message on condition: %i", condition);
void *payload = ((GList *)data)->data;
struct kiro_client_connection *cc = (struct kiro_client_connection *)payload;
struct ibv_wc wc;
gint num_comp = ibv_poll_cq (cc->conn->recv_cq, 1, &wc);
if (!num_comp) {
g_critical ("RDMA event handling was triggered, but there is no completion on the queue");
goto end_rmda_eh;
}
if (num_comp < 0) {
g_critical ("Failure getting receive completion event from the queue: %s", strerror (errno));
goto end_rmda_eh;
}
g_debug ("Got %i receive events from the queue", num_comp);
void *cq_ctx;
struct ibv_cq *cq;
int err = ibv_get_cq_event (cc->conn->recv_cq_channel, &cq, &cq_ctx);
if (!err)
ibv_ack_cq_events (cq, 1);
struct kiro_connection_context *ctx = (struct kiro_connection_context *)cc->conn->context;
guint type = ((struct kiro_ctrl_msg *)ctx->cf_mr_recv->mem)->msg_type;
g_debug ("Received a message from Client %u of type %u", cc->id, type);
switch (type) {
case KIRO_PING:
{
struct kiro_ctrl_msg *msg = (struct kiro_ctrl_msg *) (ctx->cf_mr_send->mem);
msg->msg_type = KIRO_PONG;
if (!send_msg (cc->conn, ctx->cf_mr_send)) {
g_warning ("Failure while trying to post PONG send: %s", strerror (errno));
goto done;
}
break;
}
case KIRO_ACK_RDMA:
{
g_debug ("ACK received");
if (G_TRYLOCK (realloc_timeout)) {
g_debug ("Client %i has ACKed the reallocation request", cc->id);
GList *client = g_list_find (realloc_list, (gpointer)cc);
if (client) {
realloc_list = g_list_remove_link (realloc_list, client);
if (cc->backup_mri->mr)
ibv_dereg_mr (cc->backup_mri->mr);
g_free (cc->backup_mri);
cc->backup_mri = NULL;
g_debug ("Client %i removed from realloc_list", cc->id);
}
G_UNLOCK (realloc_timeout);
}
break;
}
default:
g_debug ("Message Type is unknow. Ignoring...");
}
done:
//Post a generic receive in order to stay responsive to any messages from
//the client
if (rdma_post_recv (cc->conn, cc->conn, ctx->cf_mr_recv->mem, ctx->cf_mr_recv->size, ctx->cf_mr_recv->mr)) {
//TODO: Connection teardown in an event handler routine? Not a good
//idea...
g_critical ("Posting generic receive for event handling failed: %s", strerror (errno));
kiro_destroy_connection_context (&ctx);
rdma_destroy_ep (cc->conn);
goto end_rmda_eh;
}
ibv_req_notify_cq (cc->conn->recv_cq, 0); // Make the respective Queue push events onto the channel
g_debug ("Finished RDMA event handling");
end_rmda_eh:
G_UNLOCK (rdma_handling);
return TRUE;
}
int
kiro_server_start (KiroServer *self, const char *address, const char *port, void *mem, size_t mem_size)
{
g_return_val_if_fail (self != NULL, -1);
KiroServerPrivate *priv = KIRO_SERVER_GET_PRIVATE (self);
if (priv->base) {
g_debug ("Server already started.");
return -1;
}
if (!mem || mem_size == 0) {
g_warning ("Invalid memory given to provide.");
return -1;
}
struct rdma_addrinfo hints, *res_addrinfo;
memset (&hints, 0, sizeof (hints));
hints.ai_port_space = RDMA_PS_IB;
hints.ai_flags = RAI_PASSIVE;
char *addr_c = g_strdup (address);
char *port_c = g_strdup (port);
int rtn = rdma_getaddrinfo (addr_c, port_c, &hints, &res_addrinfo);
g_free (addr_c);
g_free (port_c);
if (rtn) {
g_critical ("Failed to create address information: %s", strerror (errno));
return -1;
}
struct ibv_qp_init_attr qp_attr;
memset (&qp_attr, 0, sizeof (qp_attr));
qp_attr.cap.max_send_wr = 10;
qp_attr.cap.max_recv_wr = 10;
qp_attr.cap.max_send_sge = 1;
qp_attr.cap.max_recv_sge = 1;
qp_attr.qp_context = priv->base;
qp_attr.sq_sig_all = 1;
if (rdma_create_ep (& (priv->base), res_addrinfo, NULL, &qp_attr)) {
g_critical ("Endpoint creation failed: %s", strerror (errno));
g_free (res_addrinfo);
return -1;
}
g_free (res_addrinfo); // No longer needed
g_debug ("Endpoint created");
char *addr_local = NULL;
struct sockaddr *src_addr = rdma_get_local_addr (priv->base);
if (!src_addr) {
addr_local = "NONE";
}
else {
addr_local = inet_ntoa (((struct sockaddr_in *)src_addr)->sin_addr);
/*
if(src_addr->sa_family == AF_INET)
addr_local = &(((struct sockaddr_in*)src_addr)->sin_addr);
else
addr_local = &(((struct sockaddr_in6*)src_addr)->sin6_addr);
*/
}
g_message ("Server bound to address %s:%s", addr_local, port);
if (rdma_listen (priv->base, 0)) {
g_critical ("Failed to put server into listening state: %s", strerror (errno));
rdma_destroy_ep (priv->base);
return -1;
}
priv->mem = mem;
priv->mem_size = mem_size;
priv->ec = rdma_create_event_channel();
if (rdma_migrate_id (priv->base, priv->ec)) {
g_critical ("Was unable to migrate connection to new Event Channel: %s", strerror (errno));
rdma_destroy_ep (priv->base);
return -1;
}
priv->main_loop = g_main_loop_new (NULL, FALSE);
priv->conn_ec = g_io_channel_unix_new (priv->ec->fd);
g_io_add_watch (priv->conn_ec, G_IO_IN | G_IO_PRI, process_cm_event, (gpointer)priv);
priv->main_thread = g_thread_new ("KIRO Server main loop", start_server_main_loop, priv->main_loop);
// We gave control to the main_loop (with add_watch) and don't need our ref
// any longer
g_io_channel_unref (priv->conn_ec);
g_message ("Enpoint listening");
return 0;
}
Acceptor::~Acceptor() {
rdma_destroy_ep(m_cm_id);
}
static int
fi_ibv_mr_reg(struct fid *fid, const void *buf, size_t len,
uint64_t access, uint64_t offset, uint64_t requested_key,
uint64_t flags, struct fid_mr **mr, void *context)
{
struct fi_ibv_mem_desc *md;
int fi_ibv_access = 0;
struct fid_domain *domain;
if (flags)
return -FI_EBADFLAGS;
if (fid->fclass != FI_CLASS_DOMAIN) {
return -FI_EINVAL;
}
domain = container_of(fid, struct fid_domain, fid);
md = calloc(1, sizeof *md);
if (!md)
return -FI_ENOMEM;
md->domain = container_of(domain, struct fi_ibv_domain, domain_fid);
md->mr_fid.fid.fclass = FI_CLASS_MR;
md->mr_fid.fid.context = context;
md->mr_fid.fid.ops = &fi_ibv_mr_ops;
/* Enable local write access by default for FI_EP_RDM which hides local
* registration requirements. This allows to avoid buffering or double
* registration */
if (!(md->domain->info->caps & FI_LOCAL_MR))
fi_ibv_access |= IBV_ACCESS_LOCAL_WRITE;
/* Local read access to an MR is enabled by default in verbs */
if (access & FI_RECV)
fi_ibv_access |= IBV_ACCESS_LOCAL_WRITE;
/* iWARP spec requires Remote Write access for an MR that is used
* as a data sink for a Remote Read */
if (access & FI_READ) {
fi_ibv_access |= IBV_ACCESS_LOCAL_WRITE;
if (md->domain->verbs->device->transport_type == IBV_TRANSPORT_IWARP)
fi_ibv_access |= IBV_ACCESS_REMOTE_WRITE;
}
if (access & FI_WRITE)
fi_ibv_access |= IBV_ACCESS_LOCAL_WRITE;
if (access & FI_REMOTE_READ)
fi_ibv_access |= IBV_ACCESS_REMOTE_READ;
/* Verbs requires Local Write access too for Remote Write access */
if (access & FI_REMOTE_WRITE)
fi_ibv_access |= IBV_ACCESS_LOCAL_WRITE |
IBV_ACCESS_REMOTE_WRITE | IBV_ACCESS_REMOTE_ATOMIC;
md->mr = ibv_reg_mr(md->domain->pd, (void *) buf, len, fi_ibv_access);
if (!md->mr)
goto err;
md->mr_fid.mem_desc = (void *) (uintptr_t) md->mr->lkey;
md->mr_fid.key = md->mr->rkey;
*mr = &md->mr_fid;
if(md->domain->eq && (md->domain->eq_flags & FI_REG_MR)) {
struct fi_eq_entry entry = {
.fid = &md->mr_fid.fid,
.context = context
};
fi_ibv_eq_write_event(md->domain->eq, FI_MR_COMPLETE,
&entry, sizeof(entry));
}
return 0;
err:
free(md);
return -errno;
}
static int fi_ibv_mr_regv(struct fid *fid, const struct iovec * iov,
size_t count, uint64_t access, uint64_t offset, uint64_t requested_key,
uint64_t flags, struct fid_mr **mr, void *context)
{
if (count > VERBS_MR_IOV_LIMIT) {
VERBS_WARN(FI_LOG_FABRIC,
"iov count > %d not supported\n",
VERBS_MR_IOV_LIMIT);
return -FI_EINVAL;
}
return fi_ibv_mr_reg(fid, iov->iov_base, iov->iov_len, access, offset,
requested_key, flags, mr, context);
}
static int fi_ibv_mr_regattr(struct fid *fid, const struct fi_mr_attr *attr,
uint64_t flags, struct fid_mr **mr)
{
return fi_ibv_mr_regv(fid, attr->mr_iov, attr->iov_count, attr->access,
0, attr->requested_key, flags, mr, attr->context);
}
static int fi_ibv_domain_bind(struct fid *fid, struct fid *bfid, uint64_t flags)
{
struct fi_ibv_domain *domain;
struct fi_ibv_eq *eq;
domain = container_of(fid, struct fi_ibv_domain, domain_fid.fid);
switch (bfid->fclass) {
case FI_CLASS_EQ:
eq = container_of(bfid, struct fi_ibv_eq, eq_fid);
domain->eq = eq;
domain->eq_flags = flags;
break;
default:
return -EINVAL;
}
return 0;
}
static int fi_ibv_domain_close(fid_t fid)
{
struct fi_ibv_domain *domain;
int ret;
domain = container_of(fid, struct fi_ibv_domain, domain_fid.fid);
if (domain->rdm) {
rdma_destroy_ep(domain->rdm_cm->listener);
free(domain->rdm_cm);
}
if (domain->pd) {
ret = ibv_dealloc_pd(domain->pd);
if (ret)
return -ret;
domain->pd = NULL;
}
fi_freeinfo(domain->info);
free(domain);
return 0;
}
static int fi_ibv_open_device_by_name(struct fi_ibv_domain *domain, const char *name)
{
struct ibv_context **dev_list;
int i, ret = -FI_ENODEV;
if (!name)
return -FI_EINVAL;
dev_list = rdma_get_devices(NULL);
if (!dev_list)
return -errno;
for (i = 0; dev_list[i] && ret; i++) {
if (domain->rdm) {
ret = strncmp(name, ibv_get_device_name(dev_list[i]->device),
strlen(name) - strlen(verbs_rdm_domain.suffix));
} else {
ret = strcmp(name, ibv_get_device_name(dev_list[i]->device));
}
if (!ret)
domain->verbs = dev_list[i];
}
rdma_free_devices(dev_list);
return ret;
}
static struct fi_ops fi_ibv_fid_ops = {
.size = sizeof(struct fi_ops),
.close = fi_ibv_domain_close,
.bind = fi_ibv_domain_bind,
.control = fi_no_control,
.ops_open = fi_no_ops_open,
};
static struct fi_ops_mr fi_ibv_domain_mr_ops = {
.size = sizeof(struct fi_ops_mr),
.reg = fi_ibv_mr_reg,
.regv = fi_ibv_mr_regv,
.regattr = fi_ibv_mr_regattr,
};
static struct fi_ops_domain fi_ibv_domain_ops = {
.size = sizeof(struct fi_ops_domain),
.av_open = fi_no_av_open,
.cq_open = fi_ibv_cq_open,
.endpoint = fi_ibv_open_ep,
.scalable_ep = fi_no_scalable_ep,
.cntr_open = fi_no_cntr_open,
.poll_open = fi_no_poll_open,
.stx_ctx = fi_no_stx_context,
.srx_ctx = fi_ibv_srq_context,
};
static struct fi_ops_domain fi_ibv_rdm_domain_ops = {
.size = sizeof(struct fi_ops_domain),
.av_open = fi_ibv_rdm_av_open,
.cq_open = fi_ibv_rdm_cq_open,
.endpoint = fi_ibv_rdm_open_ep,
.scalable_ep = fi_no_scalable_ep,
.cntr_open = fi_rbv_rdm_cntr_open,
.poll_open = fi_no_poll_open,
.stx_ctx = fi_no_stx_context,
.srx_ctx = fi_no_srx_context,
};
static int
fi_ibv_domain(struct fid_fabric *fabric, struct fi_info *info,
struct fid_domain **domain, void *context)
{
struct fi_ibv_domain *_domain;
struct fi_ibv_fabric *fab;
struct fi_info *fi;
int ret;
fi = fi_ibv_get_verbs_info(info->domain_attr->name);
if (!fi)
return -FI_EINVAL;
fab = container_of(fabric, struct fi_ibv_fabric, util_fabric.fabric_fid);
ret = ofi_check_domain_attr(&fi_ibv_prov, fabric->api_version,
fi->domain_attr, info->domain_attr);
if (ret)
return ret;
_domain = calloc(1, sizeof *_domain);
if (!_domain)
return -FI_ENOMEM;
_domain->info = fi_dupinfo(info);
if (!_domain->info)
goto err1;
_domain->rdm = FI_IBV_EP_TYPE_IS_RDM(info);
if (_domain->rdm) {
_domain->rdm_cm = calloc(1, sizeof(*_domain->rdm_cm));
if (!_domain->rdm_cm) {
ret = -FI_ENOMEM;
goto err2;
}
}
ret = fi_ibv_open_device_by_name(_domain, info->domain_attr->name);
if (ret)
goto err2;
_domain->pd = ibv_alloc_pd(_domain->verbs);
if (!_domain->pd) {
ret = -errno;
goto err2;
}
_domain->domain_fid.fid.fclass = FI_CLASS_DOMAIN;
_domain->domain_fid.fid.context = context;
_domain->domain_fid.fid.ops = &fi_ibv_fid_ops;
_domain->domain_fid.mr = &fi_ibv_domain_mr_ops;
if (_domain->rdm) {
_domain->domain_fid.ops = &fi_ibv_rdm_domain_ops;
_domain->rdm_cm->ec = rdma_create_event_channel();
if (!_domain->rdm_cm->ec) {
VERBS_INFO(FI_LOG_EP_CTRL,
"Failed to create listener event channel: %s\n",
strerror(errno));
ret = -FI_EOTHER;
goto err2;
}
if (fi_fd_nonblock(_domain->rdm_cm->ec->fd) != 0) {
VERBS_INFO_ERRNO(FI_LOG_EP_CTRL, "fcntl", errno);
ret = -FI_EOTHER;
goto err3;
}
if (rdma_create_id(_domain->rdm_cm->ec,
&_domain->rdm_cm->listener, NULL, RDMA_PS_TCP))
{
VERBS_INFO(FI_LOG_EP_CTRL, "Failed to create cm listener: %s\n",
strerror(errno));
ret = -FI_EOTHER;
goto err3;
}
_domain->rdm_cm->is_bound = 0;
} else {
_domain->domain_fid.ops = &fi_ibv_domain_ops;
}
_domain->fab = fab;
*domain = &_domain->domain_fid;
return 0;
err3:
if (_domain->rdm)
rdma_destroy_event_channel(_domain->rdm_cm->ec);
err2:
if (_domain->rdm)
free(_domain->rdm_cm);
fi_freeinfo(_domain->info);
err1:
free(_domain);
return ret;
}
static int fi_ibv_trywait(struct fid_fabric *fabric, struct fid **fids, int count)
{
struct fi_ibv_cq *cq;
int ret, i;
for (i = 0; i < count; i++) {
switch (fids[i]->fclass) {
case FI_CLASS_CQ:
cq = container_of(fids[i], struct fi_ibv_cq, cq_fid.fid);
ret = cq->trywait(fids[i]);
if (ret)
return ret;
break;
case FI_CLASS_EQ:
/* We are always ready to wait on an EQ since
* rdmacm EQ is based on an fd */
continue;
case FI_CLASS_CNTR:
case FI_CLASS_WAIT:
return -FI_ENOSYS;
default:
return -FI_EINVAL;
}
}
return FI_SUCCESS;
}
static int fi_ibv_fabric_close(fid_t fid)
{
struct fi_ibv_fabric *fab;
int ret;
fab = container_of(fid, struct fi_ibv_fabric, util_fabric.fabric_fid.fid);
ret = ofi_fabric_close(&fab->util_fabric);
if (ret)
return ret;
free(fab);
return 0;
}
static struct fi_ops fi_ibv_fi_ops = {
.size = sizeof(struct fi_ops),
.close = fi_ibv_fabric_close,
.bind = fi_no_bind,
.control = fi_no_control,
.ops_open = fi_no_ops_open,
};
static struct fi_ops_fabric fi_ibv_ops_fabric = {
.size = sizeof(struct fi_ops_fabric),
.domain = fi_ibv_domain,
.passive_ep = fi_ibv_passive_ep,
.eq_open = fi_ibv_eq_open,
.wait_open = fi_no_wait_open,
.trywait = fi_ibv_trywait
};
int fi_ibv_fabric(struct fi_fabric_attr *attr, struct fid_fabric **fabric,
void *context)
{
struct fi_ibv_fabric *fab;
struct fi_info *info;
int ret;
ret = fi_ibv_init_info();
if (ret)
return ret;
fab = calloc(1, sizeof(*fab));
if (!fab)
return -FI_ENOMEM;
for (info = verbs_info; info; info = info->next) {
ret = ofi_fabric_init(&fi_ibv_prov, info->fabric_attr, attr,
&fab->util_fabric, context);
if (ret != -FI_ENODATA)
break;
}
if (ret) {
free(fab);
return ret;
}
*fabric = &fab->util_fabric.fabric_fid;
(*fabric)->fid.ops = &fi_ibv_fi_ops;
(*fabric)->ops = &fi_ibv_ops_fabric;
return 0;
}
