int main(int argc, char **argv)
{
struct sockaddr_in addr;
struct rdma_cm_event *event = NULL;
struct rdma_cm_id *listener = NULL;
struct rdma_event_channel *ec = NULL;
uint16_t port = 0;
memset(&addr, 0, sizeof(addr));
addr.sin_family = AF_INET;
TEST_Z(ec = rdma_create_event_channel());
TEST_NZ(rdma_create_id(ec, &listener, NULL, RDMA_PS_TCP));
TEST_NZ(rdma_bind_addr(listener, (struct sockaddr *)&addr));
TEST_NZ(rdma_listen(listener, 10)); /* backlog=10 is arbitrary */
port = ntohs(rdma_get_src_port(listener));
printf("listening on port %d.\n", port);
while (rdma_get_cm_event(ec, &event) == 0) {
struct rdma_cm_event event_copy;
memcpy(&event_copy, event, sizeof(*event));
rdma_ack_cm_event(event);
if (on_event(&event_copy))
break;
}
rdma_destroy_id(listener);
rdma_destroy_event_channel(ec);
return 0;
}
int ibrdma_transfer(struct transfer_info *tfi, int num_tfi) {
struct addrinfo *addr;
struct rdma_cm_id *cmid= NULL;
struct rdma_event_channel *ec = NULL;
struct rdma_conn_param cm_params;
TEST_NZ(getaddrinfo(host, port, NULL, &addr));
TEST_Z(ec = rdma_create_event_channel());
TEST_NZ(rdma_create_id(ec, &cmid, NULL, RDMA_PS_TCP));
TEST_NZ(rdma_resolve_addr(cmid, NULL, addr->ai_addr, TIMEOUT_IN_MS));
TEST_NZ(wait_for_event(ec, RDMA_CM_EVENT_ADDR_RESOLVED));
freeaddrinfo(addr);
build_connection(cmid);
TEST_NZ(rdma_resolve_route(cmid, TIMEOUT_IN_MS));
TEST_NZ(wait_for_event(ec, RDMA_CM_EVENT_ROUTE_RESOLVED));
build_params(&cm_params);
TEST_NZ(rdma_connect(cmid, &cm_params));
TEST_NZ(wait_for_event(ec, RDMA_CM_EVENT_ESTABLISHED));
on_connect(cmid->context);
TEST_NZ(wait_for_event(ec, RDMA_CM_EVENT_DISCONNECTED));
rdma_destroy_id(&cmid);
rdma_destroy_event_channel(&ec);
return 0;
}
int xfer_rdma_finalize(struct xfer_data *data)
{
struct rdma_cm_event *event;
int rc;
if (data->servername) {
rc = rdma_disconnect(data->cm_id);
if (rc) {
perror("rdma_disconnect");
fprintf(stderr, "%d:%s: rdma disconnect error\n", pid,
__func__);
return -1;
}
}
rdma_get_cm_event(data->cm_channel, &event);
if (event->event != RDMA_CM_EVENT_DISCONNECTED)
fprintf(stderr, "%d:%s: unexpected event during disconnect %d\n",
pid, __func__, event->event);
rdma_ack_cm_event(event);
rdma_destroy_id(data->cm_id);
rdma_destroy_event_channel(data->cm_channel);
return 0;
}
void client_disconnect(void){
struct rdma_event_channel *ec = s_ctx->ec;
struct rdma_cm_id *id = s_ctx->id;
struct rdma_cm_event *event = NULL;
struct timeval start, end, dt;
gettimeofday(&start, NULL);
printf("ready to disconnect\n");
rdma_disconnect(id);
printf("send disconnect\n");
gettimeofday(&end, NULL);
timersub(&end, &start, &dt);
long usec = dt.tv_usec + 1000000 * dt.tv_sec;
printf("[rdma_disconnect] takes %ld micro_secs.\n", usec);
while(1){
if(rdma_get_cm_event(ec, &event) == 0) {
struct rdma_cm_event event_copy;
memcpy(&event_copy, event, sizeof(*event));
rdma_ack_cm_event(event);
if (event_copy.event == RDMA_CM_EVENT_DISCONNECTED){
on_disconnect(event_copy.id);
rdma_destroy_event_channel(ec);
break;
}
}
}
return;
}
static int fi_ibv_eq_close(fid_t fid)
{
struct fi_ibv_eq *eq;
struct fi_ibv_eq_entry *entry;
eq = container_of(fid, struct fi_ibv_eq, eq_fid.fid);
/* TODO: use util code, if possible, and add ref counting */
if (eq->channel)
rdma_destroy_event_channel(eq->channel);
close(eq->epfd);
while (!dlistfd_empty(&eq->list_head)) {
entry = container_of(eq->list_head.list.next,
struct fi_ibv_eq_entry, item);
dlistfd_remove(eq->list_head.list.next, &eq->list_head);
free(entry);
}
dlistfd_head_free(&eq->list_head);
fastlock_destroy(&eq->lock);
free(eq);
return 0;
}
//static int run(int argc, char **argv)
int ibrdma_send(char* host, char* port, void* data, uint64_t size)
{
struct addrinfo *addr;
struct rdma_cm_id *cmid= NULL;
struct rdma_event_channel *ec = NULL;
struct rdma_conn_param cm_params;
TEST_NZ(getaddrinfo(host, port, NULL, &addr));
TEST_Z(ec = rdma_create_event_channel());
TEST_NZ(rdma_create_id(ec, &cmid, NULL, RDMA_PS_TCP));
TEST_NZ(rdma_resolve_addr(cmid, NULL, addr->ai_addr, TIMEOUT_IN_MS));
TEST_NZ(wait_for_event(ec, RDMA_CM_EVENT_ADDR_RESOLVED));
freeaddrinfo(addr);
build_connection(cmid);
TEST_NZ(rdma_resolve_route(cmid, TIMEOUT_IN_MS));
TEST_NZ(wait_for_event(ec, RDMA_CM_EVENT_ROUTE_RESOLVED));
build_params(&cm_params);
TEST_NZ(rdma_connect(cmid, &cm_params));
TEST_NZ(wait_for_event(ec, RDMA_CM_EVENT_ESTABLISHED));
on_connect(cmid->context);
/* Init MSG send to start RDMA*/
init_tfile(data, size);
send_init(cmid->context);
/*----------------------------*/
TEST_NZ(wait_for_event(ec, RDMA_CM_EVENT_DISCONNECTED));
rdma_destroy_id(cmid);
rdma_destroy_event_channel(ec);
return 0;
}
static int migrate_channel(struct rdma_cm_id *listen_id)
{
struct rdma_event_channel *channel;
int i, ret;
printf("migrating to new event channel\n");
channel = rdma_create_event_channel();
if (!channel) {
perror("cmatose: failed to create event channel");
return -1;
}
ret = 0;
if (listen_id)
ret = rdma_migrate_id(listen_id, channel);
for (i = 0; i < connections && !ret; i++)
ret = rdma_migrate_id(test.nodes[i].cma_id, channel);
if (!ret) {
rdma_destroy_event_channel(test.channel);
test.channel = channel;
} else
perror("cmatose: failure migrating to channel");
return ret;
}
int main(int argc, char **argv)
{
struct addrinfo *addr;
struct rdma_cm_event *event = NULL;
struct rdma_cm_id *conn= NULL;
struct rdma_event_channel *ec = NULL;
if (argc != 3)
die("usage: client <server-address> <server-port>");
TEST_NZ(getaddrinfo(argv[1], argv[2], NULL, &addr));
TEST_Z(ec = rdma_create_event_channel());
TEST_NZ(rdma_create_id(ec, &conn, NULL, RDMA_PS_TCP));
TEST_NZ(rdma_resolve_addr(conn, NULL, addr->ai_addr, TIMEOUT_IN_MS));
freeaddrinfo(addr);
while (rdma_get_cm_event(ec, &event) == 0) {
struct rdma_cm_event event_copy;
memcpy(&event_copy, event, sizeof(*event));
rdma_ack_cm_event(event);
if (on_event(&event_copy))
break;
}
rdma_destroy_event_channel(ec);
return 0;
}
static int fi_ibv_eq_close(fid_t fid)
{
struct fi_ibv_eq *eq;
struct fi_ibv_eq_entry *entry;
eq = container_of(fid, struct fi_ibv_eq, eq_fid.fid);
if (eq->channel)
rdma_destroy_event_channel(eq->channel);
close(eq->epfd);
fastlock_acquire(&eq->lock);
while(!dlistfd_empty(&eq->list_head)) {
entry = container_of(eq->list_head.list.next, struct fi_ibv_eq_entry, item);
dlistfd_remove(eq->list_head.list.next, &eq->list_head);
free(entry);
}
dlistfd_head_free(&eq->list_head);
fastlock_destroy(&eq->lock);
free(eq);
return 0;
}
void client_disconnect(void){
struct rdma_event_channel *ec = s_ctx->ec;
struct rdma_cm_id *id = s_ctx->id;
struct rdma_cm_event *event = NULL;
printf("ready to disconnect\n");
rdma_disconnect(id);
printf("send disconnect\n");
while(1){
if(rdma_get_cm_event(ec, &event) == 0) {
struct rdma_cm_event event_copy;
memcpy(&event_copy, event, sizeof(*event));
rdma_ack_cm_event(event);
if (event_copy.event == RDMA_CM_EVENT_DISCONNECTED){
on_disconnect(event_copy.id);
rdma_destroy_event_channel(ec);
break;
}
}
}
return;
}
/// The IBConnectionGroup default destructor.
virtual ~IBConnectionGroup()
{
for (auto& c : conn_)
c = nullptr;
if (listen_id_) {
int err = rdma_destroy_id(listen_id_);
if (err) {
L_(error) << "rdma_destroy_id() failed";
}
listen_id_ = nullptr;
}
if (cq_) {
int err = ibv_destroy_cq(cq_);
if (err) {
L_(error) << "ibv_destroy_cq() failed";
}
cq_ = nullptr;
}
if (pd_) {
int err = ibv_dealloc_pd(pd_);
if (err) {
L_(error) << "ibv_dealloc_pd() failed";
}
pd_ = nullptr;
}
rdma_destroy_event_channel(ec_);
}
neigh_table_mgr::~neigh_table_mgr()
{
stop_garbage_collector();
if (m_neigh_cma_event_channel) {
rdma_destroy_event_channel(m_neigh_cma_event_channel);
}
}
int RDMA_Active_Finalize(struct RDMA_communicator *comm)
{
TEST_NZ(wait_for_event(comm->ec, RDMA_CM_EVENT_DISCONNECTED));
rdma_destroy_id(comm->cm_id);
rdma_destroy_event_channel(comm->ec);
return 0;
}
void network_release() {
ibv_dereg_mr(mr_data);
rdma_destroy_qp(cm_id);
ibv_destroy_cq(cq);
ibv_destroy_comp_channel(comp_chan);
rdma_destroy_id(cm_id);
rdma_destroy_event_channel(cm_channel);
}
static int __fi_eq_cm_close(fid_t fid)
{
struct __fid_eq_cm *eq;
eq = container_of(fid, struct __fid_eq_cm, eq_fid.fid);
if (eq->channel)
rdma_destroy_event_channel(eq->channel);
free(eq);
return 0;
}
void cfio_rdma_client_final()
{
cfio_rdma_client_wait(NULL);
// rdma_debug("final ...");
rdma_disconnect(rdma_conn->id);
wait_rdma_event();
rdma_destroy_event_channel(ec);
// rdma_debug("final successfully");
}
static int
__fi_eq_open(struct fid_fabric *fabric, const struct fi_eq_attr *attr,
struct fid_eq **eq, void *context)
{
struct __fid_eq_cm *_eq;
long flags = 0;
int ret;
if (attr->format != FI_EQ_FORMAT_CM)
return -FI_ENOSYS;
_eq = calloc(1, sizeof *_eq);
if (!_eq)
return -FI_ENOMEM;
_eq->fab = container_of(fabric, struct __fid_fabric, fabric_fid);
switch (attr->wait_obj) {
case FI_WAIT_FD:
_eq->channel = rdma_create_event_channel();
if (!_eq->channel) {
ret = -errno;
goto err1;
}
fcntl(_eq->channel->fd, F_GETFL, &flags);
ret = fcntl(_eq->channel->fd, F_SETFL, flags | O_NONBLOCK);
if (ret) {
ret = -errno;
goto err2;
}
break;
case FI_WAIT_NONE:
break;
default:
return -FI_ENOSYS;
}
_eq->flags = attr->flags;
_eq->eq_fid.fid.fclass = FID_CLASS_EQ;
_eq->eq_fid.fid.context = context;
_eq->eq_fid.fid.ops = &__fi_eq_cm_ops;
_eq->eq_fid.ops = &__fi_eq_cm_data_ops;
*eq = &_eq->eq_fid;
return 0;
err2:
if (_eq->channel)
rdma_destroy_event_channel(_eq->channel);
err1:
free(_eq);
return ret;
}
void
diod_rdma_destroy (diod_rdma_t rdma)
{
if (rdma->listen_id) {
rdma_destroy_id(rdma->listen_id);
rdma->listen_id = NULL;
}
if (rdma->event_channel) {
rdma_destroy_event_channel (rdma->event_channel);
rdma->event_channel = NULL;
}
free (rdma);
}
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");
}
static int ibw_ctx_priv_destruct(struct ibw_ctx_priv *pctx)
{
DEBUG(DEBUG_DEBUG, ("ibw_ctx_priv_destruct(%p)\n", pctx));
/*
* tevent_fd must be removed before the fd is closed
*/
TALLOC_FREE(pctx->cm_channel_event);
/* destroy cm */
if (pctx->cm_channel) {
rdma_destroy_event_channel(pctx->cm_channel);
pctx->cm_channel = NULL;
}
if (pctx->cm_id) {
rdma_destroy_id(pctx->cm_id);
pctx->cm_id = NULL;
}
return 0;
}
int main(int argc, char **argv)
{
struct addrinfo *addr;
struct rdma_cm_event *event = NULL;
struct rdma_cm_id *conn= NULL;
struct rdma_event_channel *ec = NULL;
if (argc != 4)
usage(argv[0]);
if (strcmp(argv[1], "write") == 0)
set_mode(M_WRITE);
else if (strcmp(argv[1], "read") == 0)
set_mode(M_READ);
else
usage(argv[0]);
TEST_NZ(getaddrinfo(argv[2], argv[3], NULL, &addr));
TEST_Z(ec = rdma_create_event_channel());
TEST_NZ(rdma_create_id(ec, &conn, NULL, RDMA_PS_TCP));
TEST_NZ(rdma_resolve_addr(conn, NULL, addr->ai_addr, TIMEOUT_IN_MS));
freeaddrinfo(addr);
while (rdma_get_cm_event(ec, &event) == 0) {
struct rdma_cm_event event_copy;
memcpy(&event_copy, event, sizeof(*event));
rdma_ack_cm_event(event);
if (on_event(&event_copy))
break;
}
rdma_destroy_event_channel(ec);
return 0;
}
int ibrdma_transfer(struct transfer_info tfi, int num_tfi) {
struct addrinfo *addr;
struct rdma_cm_id *cmid= NULL;
struct rdma_event_channel *ec = NULL;
struct rdma_conn_param cm_params;
int i,j;
/*Allocation buffer space for reading from local fs to memory*/
struct transfer_file *ffile = tfi.tfiles;
int nf = tfi.tfiles;
char* host = tfi.ib_host;
char* port; sprintf(port,"%d",tfi.ib_port);
for (i = 0; i < NUM_FILE_BUF_C; i++) {
tfi.fbufs[i].fbuf = (char *)malloc(FILE_BUF_SIZE_C);
tfi.fbufs[i].size = 0;
}
TEST_NZ(getaddrinfo(host, port, NULL, &addr));
TEST_Z(ec = rdma_create_event_channel());
TEST_NZ(rdma_create_id(ec, &cmid, NULL, RDMA_PS_TCP));
TEST_NZ(rdma_resolve_addr(cmid, NULL, addr->ai_addr, TIMEOUT_IN_MS));
TEST_NZ(wait_for_event(ec, RDMA_CM_EVENT_ADDR_RESOLVED));
freeaddrinfo(addr);
build_connection(cmid);
TEST_NZ(rdma_resolve_route(cmid, TIMEOUT_IN_MS));
TEST_NZ(wait_for_event(ec, RDMA_CM_EVENT_ROUTE_RESOLVED));
build_params(&cm_params);
TEST_NZ(rdma_connect(cmid, &cm_params));
TEST_NZ(wait_for_event(ec, RDMA_CM_EVENT_ESTABLISHED));
on_connect(cmid->context);
TEST_NZ(wait_for_event(ec, RDMA_CM_EVENT_DISCONNECTED));
rdma_destroy_id(&cmid);
rdma_destroy_event_channel(&ec);
return 0;
}
void close(boost::system::error_code &ec)
{
if(!event_channel_)
{
HPX_IBVERBS_RESET_EC(ec);
return;
}
if(event_channel_)
{
rdma_destroy_event_channel(event_channel_);
event_channel_ = 0;
}
else {
HPX_IBVERBS_THROWS_IF(ec, boost::asio::error::not_connected);
}
if(listener_)
{
rdma_destroy_id(listener_);
listener_ = 0;
}
HPX_IBVERBS_RESET_EC(ec);
}
int main(int argc, char *argv[])
{
struct rping_cb *cb;
int op;
int ret = 0;
int persistent_server = 0;
cb = malloc(sizeof(*cb));
if (!cb)
return -ENOMEM;
memset(cb, 0, sizeof(*cb));
cb->server = -1;
cb->state = IDLE;
cb->size = 64;
cb->sin.ss_family = PF_INET;
cb->port = htons(7174);
sem_init(&cb->sem, 0, 0);
opterr = 0;
while ((op=getopt(argc, argv, "a:Pp:C:S:t:scvVd")) != -1) {
switch (op) {
case 'a':
ret = get_addr(optarg, (struct sockaddr *) &cb->sin);
break;
case 'P':
persistent_server = 1;
break;
case 'p':
cb->port = htons(atoi(optarg));
DEBUG_LOG("port %d\n", (int) atoi(optarg));
break;
case 's':
cb->server = 1;
DEBUG_LOG("server\n");
break;
case 'c':
cb->server = 0;
DEBUG_LOG("client\n");
break;
case 'S':
cb->size = atoi(optarg);
if ((cb->size < RPING_MIN_BUFSIZE) ||
(cb->size > (RPING_BUFSIZE - 1))) {
fprintf(stderr, "Invalid size %d "
"(valid range is %Zd to %d)\n",
cb->size, RPING_MIN_BUFSIZE, RPING_BUFSIZE);
ret = EINVAL;
} else
DEBUG_LOG("size %d\n", (int) atoi(optarg));
break;
case 'C':
cb->count = atoi(optarg);
if (cb->count < 0) {
fprintf(stderr, "Invalid count %d\n",
cb->count);
ret = EINVAL;
} else
DEBUG_LOG("count %d\n", (int) cb->count);
break;
case 'v':
cb->verbose++;
DEBUG_LOG("verbose\n");
break;
case 'V':
cb->validate++;
DEBUG_LOG("validate data\n");
break;
case 'd':
debug++;
break;
default:
usage("rping");
ret = EINVAL;
goto out;
}
}
if (ret)
goto out;
if (cb->server == -1) {
usage("rping");
ret = EINVAL;
goto out;
}
cb->cm_channel = rdma_create_event_channel();
if (!cb->cm_channel) {
perror("rdma_create_event_channel");
ret = errno;
goto out;
}
ret = rdma_create_id(cb->cm_channel, &cb->cm_id, cb, RDMA_PS_TCP);
if (ret) {
perror("rdma_create_id");
goto out2;
}
DEBUG_LOG("created cm_id %p\n", cb->cm_id);
ret = pthread_create(&cb->cmthread, NULL, cm_thread, cb);
if (ret) {
perror("pthread_create");
goto out2;
}
if (cb->server) {
if (persistent_server)
ret = rping_run_persistent_server(cb);
else
ret = rping_run_server(cb);
} else {
ret = rping_run_client(cb);
}
DEBUG_LOG("destroy cm_id %p\n", cb->cm_id);
rdma_destroy_id(cb->cm_id);
out2:
rdma_destroy_event_channel(cb->cm_channel);
out:
free(cb);
return ret;
}
int main(int argc, char **argv)
{
int op, ret;
while ((op = getopt(argc, argv, "s:b:c:C:S:t:p:mT")) != -1) {
switch (op) {
case 's':
dst_addr = optarg;
break;
case 'b':
src_addr = optarg;
break;
case 'c':
connections = atoi(optarg);
break;
case 'C':
message_count = atoi(optarg);
break;
case 'S':
message_size = atoi(optarg);
break;
case 't':
set_tos = 1;
tos = (uint8_t) strtoul(optarg, NULL, 0);
break;
case 'p':
port = optarg;
break;
case 'm':
migrate = 1;
break;
case 'T':
set_ts = 1;
break;
default:
printf("usage: %s\n", argv[0]);
printf("\t[-s server_address]\n");
printf("\t[-b bind_address]\n");
printf("\t[-c connections]\n");
printf("\t[-C message_count]\n");
printf("\t[-S message_size]\n");
printf("\t[-t type_of_service]\n");
printf("\t[-p port_number]\n");
printf("\t[-m(igrate)]\n");
printf("\t[-T(imestamping)]\n");
exit(1);
}
}
test.connects_left = connections;
test.channel = rdma_create_event_channel();
if (!test.channel) {
printf("failed to create event channel\n");
exit(1);
}
if (alloc_nodes())
exit(1);
if (dst_addr)
ret = run_client();
else
ret = run_server();
printf("test complete\n");
destroy_nodes();
rdma_destroy_event_channel(test.channel);
if (test.rai)
rdma_freeaddrinfo(test.rai);
printf("return status %d\n", ret);
return ret;
}
int main(int argc, char **argv)
{
int op, ret;
while ((op = getopt(argc, argv, "m:M:sb:c:C:S:p:")) != -1) {
switch (op) {
case 'm':
dst_addr = optarg;
break;
case 'M':
unmapped_addr = 1;
dst_addr = optarg;
break;
case 's':
is_sender = 1;
break;
case 'b':
src_addr = optarg;
test.src_addr = (struct sockaddr *) &test.src_in;
break;
case 'c':
connections = atoi(optarg);
break;
case 'C':
message_count = atoi(optarg);
break;
case 'S':
message_size = atoi(optarg);
break;
case 'p':
port_space = strtol(optarg, NULL, 0);
break;
default:
printf("usage: %s\n", argv[0]);
printf("\t-m multicast_address\n");
printf("\t[-M unmapped_multicast_address]\n"
"\t replaces -m and requires -b\n");
printf("\t[-s(ender)]\n");
printf("\t[-b bind_address]\n");
printf("\t[-c connections]\n");
printf("\t[-C message_count]\n");
printf("\t[-S message_size]\n");
printf("\t[-p port_space - %#x for UDP (default), "
"%#x for IPOIB]\n", RDMA_PS_UDP, RDMA_PS_IPOIB);
exit(1);
}
}
if (unmapped_addr && !src_addr) {
printf("unmapped multicast address requires binding "
"to source address\n");
exit(1);
}
test.dst_addr = (struct sockaddr *) &test.dst_in;
test.connects_left = connections;
test.channel = rdma_create_event_channel();
if (!test.channel) {
perror("failed to create event channel");
exit(1);
}
if (alloc_nodes())
exit(1);
ret = run();
printf("test complete\n");
destroy_nodes();
rdma_destroy_event_channel(test.channel);
printf("return status %d\n", ret);
return ret;
}
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;
}
static int run(int argc, char **argv)
{
struct addrinfo *addr;
//struct rdma_cm_event *event = NULL;
struct rdma_cm_id *cmid= NULL;
struct rdma_event_channel *ec = NULL;
struct rdma_conn_param cm_params;
if (argc != 4)
usage(argv[0]);
if (strcmp(argv[1], "write") == 0)
set_mode(M_WRITE);
else if (strcmp(argv[1], "read") == 0)
set_mode(M_READ);
else
usage(argv[0]);
TEST_NZ(getaddrinfo(argv[2], argv[3], NULL, &addr));
TEST_Z(ec = rdma_create_event_channel());
/*create rdma socket*/
TEST_NZ(rdma_create_id(ec, &cmid, NULL, RDMA_PS_TCP));
/* int rdma_resolve_addr (struct rdma_cm_id *id, struct sockaddr *src_addr,
struct sockaddr dst_addr, int timeout_ms)
id RDMA identifier
src_addr Source address information. This parameter may be NULL.
dst_addr Destination address information
timeout_ms Time to wait for resolution to complete
Description:
Resolve destination and optional source addresses from IP addresses
to an RDMA address. If suc- cessful,
the specified rdma_cm_id will be bound to a local device.
*/
TEST_NZ(rdma_resolve_addr(cmid, NULL, addr->ai_addr, TIMEOUT_IN_MS));
TEST_NZ(wait_for_event(ec, RDMA_CM_EVENT_ADDR_RESOLVED));
freeaddrinfo(addr);
build_connection(cmid);
sprintf(get_local_message_region(cmid->context), "message from active/client side with pid %d", getpid());
/*--------------------*/
/* int rdma_resolve_route (struct rdma_cm_id *id, int timeout_ms);
id RDMA identifier
timeout_ms Time to wait for resolution to complete
Description:
Resolves an RDMA route to the destination address in order
to establish a connection. The destination address must have
already been resolved by calling rdma_resolve_addr.
*/
TEST_NZ(rdma_resolve_route(cmid, TIMEOUT_IN_MS));
TEST_NZ(wait_for_event(ec, RDMA_CM_EVENT_ROUTE_RESOLVED));
/* -------------------- */
print_path_rec(cmid);
/* int rdma_connect (struct rdma_cm_id *id, struct rdma_conn_param *conn_param);
id RDMA identifier
conn_param connection parameters
Description:
For an rdma_cm_id of type RDMA_PS_TCP, this call initiates a connection
request to a remote destination. For an rdma_cm_id of type RDMA_PS_UDP,
it initiates a lookup of the remote QP providing the datagram service
*/
build_params(&cm_params);
printf("Connecting ...\n");
TEST_NZ(rdma_connect(cmid, &cm_params));
TEST_NZ(wait_for_event(ec, RDMA_CM_EVENT_ESTABLISHED));
printf("Connected !\n");
/* --------------------- */
/*TODO: do something */
on_connect(cmid->context);
send_mr(cmid->context);
/*--------------------*/
rdma_disconnect(cmid);
rdma_destroy_id(cmid);
rdma_destroy_event_channel(ec);
return 0;
/*=================*/
/*=================*/
/*
while (rdma_get_cm_event(ec, &event) == 0) {
memcpy(&event_copy, event, sizeof(*event));
rdma_ack_cm_event(event);
if (on_event(&event_copy))
break;
}
*/
}
int fi_ibv_eq_open(struct fid_fabric *fabric, struct fi_eq_attr *attr,
struct fid_eq **eq, void *context)
{
struct fi_ibv_eq *_eq;
struct epoll_event event;
int ret;
_eq = calloc(1, sizeof *_eq);
if (!_eq)
return -ENOMEM;
_eq->fab = container_of(fabric, struct fi_ibv_fabric, fabric_fid);
fastlock_init(&_eq->lock);
ret = dlistfd_head_init(&_eq->list_head);
if (ret) {
FI_INFO(&fi_ibv_prov, FI_LOG_EQ, "Unable to initialize dlistfd\n");
goto err1;
}
_eq->epfd = epoll_create1(0);
if (_eq->epfd < 0) {
ret = -errno;
goto err2;
}
memset(&event, 0, sizeof(event));
event.events = EPOLLIN;
if (epoll_ctl(_eq->epfd, EPOLL_CTL_ADD,
_eq->list_head.signal.fd[FI_READ_FD], &event)) {
ret = -errno;
goto err3;
}
switch (attr->wait_obj) {
case FI_WAIT_NONE:
case FI_WAIT_UNSPEC:
case FI_WAIT_FD:
_eq->channel = rdma_create_event_channel();
if (!_eq->channel) {
ret = -errno;
goto err3;
}
ret = fi_fd_nonblock(_eq->channel->fd);
if (ret)
goto err4;
if (epoll_ctl(_eq->epfd, EPOLL_CTL_ADD, _eq->channel->fd, &event)) {
ret = -errno;
goto err4;
}
break;
default:
ret = -FI_ENOSYS;
goto err1;
}
_eq->flags = attr->flags;
_eq->eq_fid.fid.fclass = FI_CLASS_EQ;
_eq->eq_fid.fid.context = context;
_eq->eq_fid.fid.ops = &fi_ibv_eq_fi_ops;
_eq->eq_fid.ops = &fi_ibv_eq_ops;
*eq = &_eq->eq_fid;
return 0;
err4:
if (_eq->channel)
rdma_destroy_event_channel(_eq->channel);
err3:
close(_eq->epfd);
err2:
dlistfd_head_free(&_eq->list_head);
err1:
fastlock_destroy(&_eq->lock);
free(_eq);
return ret;
}
int main(int argc, char **argv)
{
int op, ret;
hints.ai_port_space = RDMA_PS_TCP;
while ((op = getopt(argc, argv, "s:b:f:P:c:C:S:t:p:m")) != -1) {
switch (op) {
case 's':
dst_addr = optarg;
break;
case 'b':
src_addr = optarg;
break;
case 'f':
if (!strncasecmp("ip", optarg, 2)) {
hints.ai_flags = RAI_NUMERICHOST;
} else if (!strncasecmp("gid", optarg, 3)) {
hints.ai_flags = RAI_NUMERICHOST | RAI_FAMILY;
hints.ai_family = AF_IB;
} else if (strncasecmp("name", optarg, 4)) {
fprintf(stderr, "Warning: unknown address format\n");
}
break;
case 'P':
if (!strncasecmp("ib", optarg, 2)) {
hints.ai_port_space = RDMA_PS_IB;
} else if (strncasecmp("tcp", optarg, 3)) {
fprintf(stderr, "Warning: unknown port space format\n");
}
break;
case 'c':
connections = atoi(optarg);
break;
case 'C':
message_count = atoi(optarg);
break;
case 'S':
message_size = atoi(optarg);
break;
case 't':
set_tos = 1;
tos = (uint8_t) strtoul(optarg, NULL, 0);
break;
case 'p':
port = optarg;
break;
case 'm':
migrate = 1;
break;
default:
printf("usage: %s\n", argv[0]);
printf("\t[-s server_address]\n");
printf("\t[-b bind_address]\n");
printf("\t[-f address_format]\n");
printf("\t name, ip, ipv6, or gid\n");
printf("\t[-P port_space]\n");
printf("\t tcp or ib\n");
printf("\t[-c connections]\n");
printf("\t[-C message_count]\n");
printf("\t[-S message_size]\n");
printf("\t[-t type_of_service]\n");
printf("\t[-p port_number]\n");
printf("\t[-m(igrate)]\n");
exit(1);
}
}
test.connects_left = connections;
test.channel = create_first_event_channel();
if (!test.channel) {
exit(1);
}
if (alloc_nodes())
exit(1);
if (dst_addr) {
ret = run_client();
} else {
hints.ai_flags |= RAI_PASSIVE;
ret = run_server();
}
printf("test complete\n");
destroy_nodes();
rdma_destroy_event_channel(test.channel);
if (test.rai)
rdma_freeaddrinfo(test.rai);
printf("return status %d\n", ret);
return ret;
}