HYD_status
HYDU_sock_create_and_listen_portstr(char *iface, char *hostname, char *port_range,
char **port_str,
HYD_status(*callback) (int fd, HYD_event_t events,
void *userp), void *userp)
{
int listenfd;
char *sport, *real_port_range, *ip = NULL;
uint16_t port;
HYD_status status = HYD_SUCCESS;
/* Listen on a port in the port range */
port = 0;
real_port_range = port_range ? MPL_strdup(port_range) : NULL;
status = HYDU_sock_listen(&listenfd, real_port_range, &port);
HYDU_ERR_POP(status, "unable to listen on port\n");
/* Register the listening socket with the demux engine */
status = HYDT_dmx_register_fd(1, &listenfd, HYD_POLLIN, userp, callback);
HYDU_ERR_POP(status, "unable to register fd\n");
/* Create a port string for MPI processes to use to connect to */
if (iface) {
status = HYDU_sock_get_iface_ip(iface, &ip);
HYDU_ERR_POP(status, "unable to get network interface IP\n");
}
else if (hostname) {
ip = MPL_strdup(hostname);
}
else {
char localhost[MAX_HOSTNAME_LEN] = { 0 };
if (gethostname(localhost, MAX_HOSTNAME_LEN) < 0)
HYDU_ERR_SETANDJUMP(status, HYD_SOCK_ERROR, "unable to get local hostname\n");
ip = MPL_strdup(localhost);
}
sport = HYDU_int_to_str(port);
HYDU_MALLOC_OR_JUMP(*port_str, char *, strlen(ip) + 1 + strlen(sport) + 1, status);
MPL_snprintf(*port_str, strlen(ip) + 1 + strlen(sport) + 1, "%s:%s", ip, sport);
MPL_free(sport);
fn_exit:
if (ip)
MPL_free(ip);
return status;
fn_fail:
goto fn_exit;
}
HYD_status HYDT_bscd_ll_launch_procs(char **args, struct HYD_proxy *proxy_list,
int *control_fd)
{
int idx, i, total_procs, node_count;
int *pid, *fd_list, exec_idx;
char *targs[HYD_NUM_TMP_STRINGS], *node_list_str = NULL;
char *path = NULL, *extra_arg_list = NULL, *extra_arg, quoted_exec_string[HYD_TMP_STRLEN];
struct HYD_proxy *proxy;
struct HYDT_topo_cpuset_t cpuset;
HYD_status status = HYD_SUCCESS;
HYDU_FUNC_ENTER();
/* We use the following priority order for the executable path:
* (1) user-specified; (2) search in path; (3) Hard-coded
* location */
if (HYDT_bsci_info.launcher_exec)
path = HYDU_strdup(HYDT_bsci_info.launcher_exec);
if (!path)
path = HYDU_find_full_path("poe");
if (!path)
path = HYDU_strdup("/usr/bin/poe");
idx = 0;
targs[idx++] = HYDU_strdup(path);
if (!strcmp(HYDT_bsci_info.rmk, "ll")) {
HYDU_ERR_SETANDJUMP(status, HYD_INTERNAL_ERROR,
"ll does not support user-defined host lists\n");
}
/* Check how many nodes are being passed for the launch */
status = HYDTI_bscd_ll_query_node_count(&total_procs);
HYDU_ERR_POP(status, "unable to query for the node count\n");
node_count = 0;
for (proxy = proxy_list; proxy; proxy = proxy->next)
node_count++;
if (total_procs != node_count)
HYDU_ERR_SETANDJUMP(status, HYD_INTERNAL_ERROR,
"processes to be launched have to cover all nodes\n");
MPL_env2str("HYDRA_LAUNCHER_EXTRA_ARGS", (const char **) &extra_arg_list);
if (extra_arg_list) {
extra_arg = strtok(extra_arg_list, " ");
while (extra_arg) {
targs[idx++] = HYDU_strdup(extra_arg);
extra_arg = strtok(NULL, " ");
}
}
/* Fill in the remaining arguments */
exec_idx = idx;
for (i = 0; args[i]; i++)
targs[idx++] = HYDU_strdup(args[i]);
/* Create a quoted version of the exec string, which is only used
* when the executable is not launched directly, but through an
* actual launcher */
HYDU_snprintf(quoted_exec_string, HYD_TMP_STRLEN, "\"%s\"", targs[exec_idx]);
HYDU_FREE(targs[exec_idx]);
targs[exec_idx] = quoted_exec_string;
/* Increase pid list to accommodate the new pid */
HYDU_MALLOC(pid, int *, (HYD_bscu_pid_count + 1) * sizeof(int), status);
for (i = 0; i < HYD_bscu_pid_count; i++)
pid[i] = HYD_bscu_pid_list[i];
HYDU_FREE(HYD_bscu_pid_list);
HYD_bscu_pid_list = pid;
/* Increase fd list to accommodate these new fds */
HYDU_MALLOC(fd_list, int *, (HYD_bscu_fd_count + 3) * sizeof(int), status);
for (i = 0; i < HYD_bscu_fd_count; i++)
fd_list[i] = HYD_bscu_fd_list[i];
HYDU_FREE(HYD_bscu_fd_list);
HYD_bscu_fd_list = fd_list;
/* append proxy ID as -1 */
targs[idx++] = HYDU_int_to_str(-1);
targs[idx++] = NULL;
HYDT_topo_cpuset_zero(&cpuset);
status = HYDU_create_process(targs, NULL, NULL, &fd_stdout, &fd_stderr,
&HYD_bscu_pid_list[HYD_bscu_pid_count++], cpuset);
HYDU_ERR_POP(status, "create process returned error\n");
HYD_bscu_fd_list[HYD_bscu_fd_count++] = fd_stdout;
HYD_bscu_fd_list[HYD_bscu_fd_count++] = fd_stderr;
status = HYDT_dmx_register_fd(1, &fd_stdout, HYD_POLLIN,
(void *) (size_t) STDOUT_FILENO, HYDT_bscu_stdio_cb);
HYDU_ERR_POP(status, "demux returned error registering fd\n");
status = HYDT_dmx_register_fd(1, &fd_stderr, HYD_POLLIN,
(void *) (size_t) STDERR_FILENO, HYDT_bscu_stdio_cb);
HYDU_ERR_POP(status, "demux returned error registering fd\n");
fn_exit:
if (node_list_str)
HYDU_FREE(node_list_str);
HYDU_free_strlist(targs);
if (path)
HYDU_FREE(path);
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
HYD_status HYDT_bscd_slurm_launch_procs(char **args, struct HYD_proxy *proxy_list,
int *control_fd)
{
int num_hosts, idx, i;
int *pid, *fd_list;
char *targs[HYD_NUM_TMP_STRINGS], *node_list_str = NULL;
char *path = NULL, *extra_arg_list = NULL, *extra_arg;
struct HYD_proxy *proxy;
struct HYDT_topo_cpuset_t cpuset;
HYD_status status = HYD_SUCCESS;
HYDU_FUNC_ENTER();
/* We use the following priority order for the executable path:
* (1) user-specified; (2) search in path; (3) Hard-coded
* location */
if (HYDT_bsci_info.launcher_exec)
path = HYDU_strdup(HYDT_bsci_info.launcher_exec);
if (!path)
path = HYDU_find_full_path("srun");
if (!path)
path = HYDU_strdup("/usr/bin/srun");
idx = 0;
targs[idx++] = HYDU_strdup(path);
if (strcmp(HYDT_bsci_info.rmk, "slurm")) {
targs[idx++] = HYDU_strdup("--nodelist");
status = proxy_list_to_node_str(proxy_list, &node_list_str);
HYDU_ERR_POP(status, "unable to build a node list string\n");
targs[idx++] = HYDU_strdup(node_list_str);
}
num_hosts = 0;
for (proxy = proxy_list; proxy; proxy = proxy->next)
num_hosts++;
targs[idx++] = HYDU_strdup("-N");
targs[idx++] = HYDU_int_to_str(num_hosts);
targs[idx++] = HYDU_strdup("-n");
targs[idx++] = HYDU_int_to_str(num_hosts);
MPL_env2str("HYDRA_LAUNCHER_EXTRA_ARGS", (const char **) &extra_arg_list);
if (extra_arg_list) {
extra_arg = strtok(extra_arg_list, " ");
while (extra_arg) {
targs[idx++] = HYDU_strdup(extra_arg);
extra_arg = strtok(NULL, " ");
}
}
/* Fill in the remaining arguments */
/* We do not need to create a quoted version of the string for
* SLURM. It seems to be internally quoting it anyway. */
for (i = 0; args[i]; i++)
targs[idx++] = HYDU_strdup(args[i]);
/* Increase pid list to accommodate the new pid */
HYDU_MALLOC(pid, int *, (HYD_bscu_pid_count + 1) * sizeof(int), status);
for (i = 0; i < HYD_bscu_pid_count; i++)
pid[i] = HYD_bscu_pid_list[i];
HYDU_FREE(HYD_bscu_pid_list);
HYD_bscu_pid_list = pid;
/* Increase fd list to accommodate these new fds */
HYDU_MALLOC(fd_list, int *, (HYD_bscu_fd_count + 3) * sizeof(int), status);
for (i = 0; i < HYD_bscu_fd_count; i++)
fd_list[i] = HYD_bscu_fd_list[i];
HYDU_FREE(HYD_bscu_fd_list);
HYD_bscu_fd_list = fd_list;
/* append proxy ID as -1 */
targs[idx++] = HYDU_int_to_str(-1);
targs[idx++] = NULL;
if (HYDT_bsci_info.debug) {
HYDU_dump(stdout, "Launch arguments: ");
HYDU_print_strlist(targs);
}
HYDT_topo_cpuset_zero(&cpuset);
status = HYDU_create_process(targs, NULL, NULL, &fd_stdout, &fd_stderr,
&HYD_bscu_pid_list[HYD_bscu_pid_count++], cpuset);
HYDU_ERR_POP(status, "create process returned error\n");
HYD_bscu_fd_list[HYD_bscu_fd_count++] = fd_stdout;
HYD_bscu_fd_list[HYD_bscu_fd_count++] = fd_stderr;
status = HYDT_dmx_register_fd(1, &fd_stdout, HYD_POLLIN,
(void *) (size_t) STDOUT_FILENO, HYDT_bscu_stdio_cb);
HYDU_ERR_POP(status, "demux returned error registering fd\n");
status = HYDT_dmx_register_fd(1, &fd_stderr, HYD_POLLIN,
(void *) (size_t) STDERR_FILENO, HYDT_bscu_stdio_cb);
HYDU_ERR_POP(status, "demux returned error registering fd\n");
fn_exit:
if (node_list_str)
HYDU_FREE(node_list_str);
HYDU_free_strlist(targs);
if (path)
HYDU_FREE(path);
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
int main(int argc, char **argv)
{
int i, count, pid, ret_status, sent, closed, ret, done;
struct HYD_pmcd_hdr hdr;
HYD_status status = HYD_SUCCESS;
status = HYDU_dbg_init("proxy:unset");
HYDU_ERR_POP(status, "unable to initialization debugging\n");
status = HYDU_set_signal(SIGPIPE, signal_cb);
HYDU_ERR_POP(status, "unable to set SIGPIPE\n");
status = HYDU_set_signal(SIGTSTP, signal_cb);
HYDU_ERR_POP(status, "unable to set SIGTSTP\n");
status = HYDU_set_common_signals(signal_cb);
HYDU_ERR_POP(status, "unable to set common signals\n");
status = init_params();
HYDU_ERR_POP(status, "Error initializing proxy params\n");
status = HYD_pmcd_pmip_get_params(argv);
HYDU_ERR_POP(status, "bad parameters passed to the proxy\n");
status = HYDT_dmx_init(&HYD_pmcd_pmip.user_global.demux);
HYDU_ERR_POP(status, "unable to initialize the demux engine\n");
status = HYDT_ftb_init();
HYDU_ERR_POP(status, "unable to initialize FTB\n");
/* See if HYDI_CONTROL_FD is set before trying to connect upstream */
ret = MPL_env2int("HYDI_CONTROL_FD", &HYD_pmcd_pmip.upstream.control);
if (ret < 0) {
HYDU_ERR_POP(status, "error reading HYDI_CONTROL_FD environment\n");
}
else if (ret == 0) {
status = HYDU_sock_connect(HYD_pmcd_pmip.upstream.server_name,
HYD_pmcd_pmip.upstream.server_port,
&HYD_pmcd_pmip.upstream.control,
HYD_pmcd_pmip.local.retries, HYD_CONNECT_DELAY);
HYDU_ERR_POP(status,
"unable to connect to server %s at port %d (check for firewalls!)\n",
HYD_pmcd_pmip.upstream.server_name, HYD_pmcd_pmip.upstream.server_port);
}
status = HYDU_sock_write(HYD_pmcd_pmip.upstream.control,
&HYD_pmcd_pmip.local.id, sizeof(HYD_pmcd_pmip.local.id), &sent,
&closed, HYDU_SOCK_COMM_MSGWAIT);
HYDU_ERR_POP(status, "unable to send the proxy ID to the server\n");
if (closed)
goto fn_fail;
status = HYDT_dmx_register_fd(1, &HYD_pmcd_pmip.upstream.control,
HYD_POLLIN, NULL, HYD_pmcd_pmip_control_cmd_cb);
HYDU_ERR_POP(status, "unable to register fd\n");
while (1) {
/* Wait for some event to occur */
status = HYDT_dmx_wait_for_event(-1);
HYDU_ERR_POP(status, "demux engine error waiting for event\n");
/* Check to see if there's any open read socket left; if there
* are, we will just wait for more events. */
count = 0;
for (i = 0; i < HYD_pmcd_pmip.local.proxy_process_count; i++) {
if (HYD_pmcd_pmip.downstream.out[i] != HYD_FD_CLOSED)
count++;
if (HYD_pmcd_pmip.downstream.err[i] != HYD_FD_CLOSED)
count++;
if (count)
break;
}
if (!count)
break;
}
/* Now wait for the processes to finish */
done = 0;
while (1) {
pid = waitpid(-1, &ret_status, 0);
/* Find the pid and mark it as complete. */
if (pid > 0)
for (i = 0; i < HYD_pmcd_pmip.local.proxy_process_count; i++)
if (HYD_pmcd_pmip.downstream.pid[i] == pid) {
if (HYD_pmcd_pmip.downstream.forced_cleanup) {
/* If it is a forced cleanup, the exit status
* is either already set or we have to ignore
* it */
if (HYD_pmcd_pmip.downstream.exit_status[i] == -1)
HYD_pmcd_pmip.downstream.exit_status[i] = 0;
else
HYD_pmcd_pmip.downstream.exit_status[i] = ret_status;
}
else {
HYD_pmcd_pmip.downstream.exit_status[i] = ret_status;
}
done++;
}
/* If no more processes are pending, break out */
if (done == HYD_pmcd_pmip.local.proxy_process_count)
break;
/* Check if there are any messages from the launcher */
status = HYDT_dmx_wait_for_event(0);
HYDU_IGNORE_TIMEOUT(status);
HYDU_ERR_POP(status, "demux engine error waiting for event\n");
}
/* Send the exit status upstream */
HYD_pmcd_init_header(&hdr);
hdr.cmd = EXIT_STATUS;
status =
HYDU_sock_write(HYD_pmcd_pmip.upstream.control, &hdr, sizeof(hdr), &sent, &closed,
HYDU_SOCK_COMM_MSGWAIT);
HYDU_ERR_POP(status, "unable to send EXIT_STATUS command upstream\n");
HYDU_ASSERT(!closed, status);
status = HYDU_sock_write(HYD_pmcd_pmip.upstream.control,
HYD_pmcd_pmip.downstream.exit_status,
HYD_pmcd_pmip.local.proxy_process_count * sizeof(int), &sent,
&closed, HYDU_SOCK_COMM_MSGWAIT);
HYDU_ERR_POP(status, "unable to return exit status upstream\n");
HYDU_ASSERT(!closed, status);
status = HYDT_dmx_deregister_fd(HYD_pmcd_pmip.upstream.control);
HYDU_ERR_POP(status, "unable to deregister fd\n");
close(HYD_pmcd_pmip.upstream.control);
status = HYDT_dmx_finalize();
HYDU_ERR_POP(status, "error returned from demux finalize\n");
status = HYDT_ftb_finalize();
HYDU_ERR_POP(status, "unable to finalize FTB\n");
status = HYDT_bsci_finalize();
HYDU_ERR_POP(status, "unable to finalize the bootstrap device\n");
/* cleanup the params structure */
cleanup_params();
fn_exit:
HYDU_dbg_finalize();
return status;
fn_fail:
/* kill all processes */
HYD_pmcd_pmip_send_signal(SIGKILL);
goto fn_exit;
}
