static HYD_status create_env_file(const struct HYD_env *envlist, int num_ranks, int *ranks)
{
HYD_status status = HYD_SUCCESS;
char filename[256];
FILE *f;
const struct HYD_env *e;
int ret;
int r;
HYDU_FUNC_ENTER();
for (r = 0; r < num_ranks; ++r) {
MPL_snprintf(filename, sizeof(filename), "/tmp/hydra-env-file-%d:%d", (int) getpid(),
ranks[r]);
f = fopen(filename, "w");
HYDU_ERR_CHKANDJUMP(status, f == NULL, HYD_INTERNAL_ERROR, "fopen failed: %s\n",
strerror(errno));
for (e = envlist; e; e = e->next) {
fprintf(f, "%s=%s\n", e->env_name, e->env_value);
}
ret = fclose(f);
HYDU_ERR_CHKANDJUMP(status, ret, HYD_INTERNAL_ERROR, "fclose failed: %s\n",
strerror(errno));
}
fn_exit:
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
HYD_status HYDT_ckpoint_blcr_restart(const char *prefix, int pgid, int id, int ckpt_num,
struct HYD_env *envlist, int num_ranks, int ranks[],
int *in, int *out, int *err, int *pid)
{
HYD_status status = HYD_SUCCESS;
int ret;
int context_fd;
cr_restart_handle_t cr_handle;
cr_restart_args_t args;
char filename[256];
char port_str[64];
int port;
HYDU_FUNC_ENTER();
/* create listener socket for stdin/out/err */
status = create_stdinouterr_sock(&port);
HYDU_ERR_POP(status, "failed to create stdin/out/err socket\n");
MPL_snprintf(port_str, sizeof(port_str), "%d", port);
status = HYDU_append_env_to_list(STDINOUTERR_PORT_NAME, port_str, &envlist);
HYDU_ERR_POP(status, "failed to add to env list\n");
status = create_env_file(envlist, num_ranks, ranks);
if (status)
HYDU_ERR_POP(status, "blcr restart\n");
/* open the checkpoint file */
MPL_snprintf(filename, sizeof(filename), "%s/context-num%d-%d-%d", prefix, ckpt_num, pgid,
id);
context_fd = open(filename, O_RDONLY /* | O_LARGEFILE */);
HYDU_ERR_CHKANDJUMP(status, context_fd < 0, HYD_INTERNAL_ERROR, "open failed, %s\n",
strerror(errno));
/* ... initialize the request structure */
cr_initialize_restart_args_t(&args);
args.cr_fd = context_fd;
args.cr_flags = CR_RSTRT_RESTORE_PID;
/* ... issue the request */
ret = cr_request_restart(&args, &cr_handle);
HYDU_ERR_CHKANDJUMP(status, ret, HYD_INTERNAL_ERROR, "cr_request_restart failed, %s\n",
strerror(errno));
ret = close(context_fd);
HYDU_ERR_CHKANDJUMP(status, ret, HYD_INTERNAL_ERROR, "close failed, %s\n",
strerror(errno));
/* get fds for stdin/out/err sockets, and get pids of restarted processes */
status = wait_for_stdinouterr_sockets(num_ranks, ranks, in, out, err, pid);
if (status)
HYDU_ERR_POP(status, "blcr restart\n");
fn_exit:
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
static HYD_status fn_info_putnodeattr(int fd, char *args[])
{
struct HYD_string_stash stash;
char *key, *val, *thrid, *cmd;
struct HYD_pmcd_token *tokens = NULL;
int token_count, ret;
struct HYD_pmcd_pmi_v2_reqs *req;
HYD_status status = HYD_SUCCESS;
HYDU_FUNC_ENTER();
status = HYD_pmcd_pmi_args_to_tokens(args, &tokens, &token_count);
HYDU_ERR_POP(status, "unable to convert args to tokens\n");
key = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "key");
HYDU_ERR_CHKANDJUMP(status, key == NULL, HYD_INTERNAL_ERROR, "unable to find key token\n");
val = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "value");
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR, "unable to find value token\n");
thrid = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "thrid");
status = HYD_pmcd_pmi_add_kvs(key, val, HYD_pmcd_pmip.local.kvs, &ret);
HYDU_ERR_POP(status, "unable to put data into kvs\n");
HYD_STRING_STASH_INIT(stash);
HYD_STRING_STASH(stash, MPL_strdup("cmd=info-putnodeattr-response;"), status);
if (thrid) {
HYD_STRING_STASH(stash, MPL_strdup("thrid="), status);
HYD_STRING_STASH(stash, MPL_strdup(thrid), status);
HYD_STRING_STASH(stash, MPL_strdup(";"), status);
}
HYD_STRING_STASH(stash, MPL_strdup("rc="), status);
HYD_STRING_STASH(stash, HYDU_int_to_str(ret), status);
HYD_STRING_STASH(stash, MPL_strdup(";"), status);
HYD_STRING_SPIT(stash, cmd, status);
send_cmd_downstream(fd, cmd);
MPL_free(cmd);
for (req = pending_reqs; req; req = req->next) {
if (!strcmp(req->key, key)) {
/* Poke the progress engine before exiting */
status = poke_progress(key);
HYDU_ERR_POP(status, "poke progress error\n");
break;
}
}
fn_exit:
if (tokens)
HYD_pmcd_pmi_free_tokens(tokens, token_count);
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
HYD_status HYDT_bscd_pbs_launcher_finalize(void)
{
int err;
HYD_status status = HYD_SUCCESS;
HYDU_FUNC_ENTER();
#if defined(HAVE_TM_H)
err = tm_finalize();
HYDU_ERR_CHKANDJUMP(status, err != TM_SUCCESS, HYD_INTERNAL_ERROR,
"error calling tm_finalize\n");
#endif /* HAVE_TM_H */
if (HYDT_bscd_pbs_sys) {
if (HYDT_bscd_pbs_sys->task_id)
HYDU_FREE(HYDT_bscd_pbs_sys->task_id);
if (HYDT_bscd_pbs_sys->spawn_events)
HYDU_FREE(HYDT_bscd_pbs_sys->spawn_events);
HYDU_FREE(HYDT_bscd_pbs_sys);
}
fn_exit:
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
static HYD_status create_stdinouterr_sock(int *port)
{
HYD_status status = HYD_SUCCESS;
int ret;
struct sockaddr_in sin;
socklen_t len;
HYDU_FUNC_ENTER();
listen_fd = socket(AF_INET, SOCK_STREAM, IPPROTO_TCP);
HYDU_ERR_CHKANDJUMP(status, listen_fd < 0, HYD_INTERNAL_ERROR, "socket() failed, %s\n",
strerror(errno));
memset((void *) &sin, 0, sizeof(sin));
sin.sin_family = AF_INET;
sin.sin_addr.s_addr = htonl(INADDR_LOOPBACK);
sin.sin_port = htons(0);
ret = bind(listen_fd, (struct sockaddr *) &sin, sizeof(sin));
HYDU_ERR_CHKANDJUMP(status, ret, HYD_INTERNAL_ERROR, "bind() failed, %s\n",
strerror(errno));
ret = listen(listen_fd, SOMAXCONN);
HYDU_ERR_CHKANDJUMP(status, ret, HYD_INTERNAL_ERROR, "listen() failed, %s\n",
strerror(errno));
len = sizeof(sin);
ret = getsockname(listen_fd, (struct sockaddr *) &sin, &len);
HYDU_ERR_CHKANDJUMP(status, ret, HYD_INTERNAL_ERROR, "getsockname() failed, %s\n",
strerror(errno));
*port = ntohs(sin.sin_port);
fn_exit:
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
HYD_status HYDU_set_int(char *arg, int *var, int val)
{
HYD_status status = HYD_SUCCESS;
HYDU_ERR_CHKANDJUMP(status, *var != -1, HYD_INTERNAL_ERROR, "duplicate setting: %s\n", arg);
*var = val;
fn_exit:
return status;
fn_fail:
goto fn_exit;
}
static HYD_status control_port_fn(char *arg, char ***argv)
{
char *port = NULL, *name;
HYD_status status = HYD_SUCCESS;
HYDU_ERR_CHKANDJUMP(status, HYD_pmcd_pmip.upstream.server_name, HYD_INTERNAL_ERROR,
"duplicate control port setting\n");
port = MPL_strdup(**argv);
HYDU_ERR_CHKANDJUMP(status, NULL == port, HYD_INTERNAL_ERROR, "port not provided\n");
name = strtok(port, ":");
HYD_pmcd_pmip.upstream.server_name = name ? MPL_strdup(name) : NULL;
HYD_pmcd_pmip.upstream.server_port = strtol(strtok(NULL, ":"), NULL, 10);
(*argv)++;
fn_exit:
MPL_free(port);
return status;
fn_fail:
goto fn_exit;
}
static HYD_status fn_info_getjobattr(int fd, char *args[])
{
struct HYD_string_stash stash;
char *cmd, *key, *thrid;
struct HYD_pmcd_token *tokens;
int token_count;
HYD_status status = HYD_SUCCESS;
HYDU_FUNC_ENTER();
status = HYD_pmcd_pmi_args_to_tokens(args, &tokens, &token_count);
HYDU_ERR_POP(status, "unable to convert args to tokens\n");
key = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "key");
HYDU_ERR_CHKANDJUMP(status, key == NULL, HYD_INTERNAL_ERROR, "unable to find token: key\n");
thrid = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "thrid");
if (!strcmp(key, "PMI_process_mapping")) {
HYD_STRING_STASH_INIT(stash);
HYD_STRING_STASH(stash, MPL_strdup("cmd=info-getjobattr-response;"), status);
if (thrid) {
HYD_STRING_STASH(stash, MPL_strdup("thrid="), status);
HYD_STRING_STASH(stash, MPL_strdup(thrid), status);
HYD_STRING_STASH(stash, MPL_strdup(";"), status);
}
HYD_STRING_STASH(stash, MPL_strdup("found=TRUE;value="), status);
HYD_STRING_STASH(stash, MPL_strdup(HYD_pmcd_pmip.system_global.pmi_process_mapping),
status);
HYD_STRING_STASH(stash, MPL_strdup(";rc=0;"), status);
HYD_STRING_SPIT(stash, cmd, status);
send_cmd_downstream(fd, cmd);
MPL_free(cmd);
} else {
status = send_cmd_upstream("cmd=info-getjobattr;", fd, args);
HYDU_ERR_POP(status, "error sending command upstream\n");
}
fn_exit:
HYD_pmcd_pmi_free_tokens(tokens, token_count);
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
HYD_status HYDU_set_str(char *arg, char **var, const char *val)
{
HYD_status status = HYD_SUCCESS;
HYDU_ERR_CHKANDJUMP(status, *var, HYD_INTERNAL_ERROR, "duplicate setting: %s\n", arg);
if (val == NULL)
HYDU_ERR_SETANDJUMP(status, HYD_INTERNAL_ERROR, "cannot assign NULL object\n");
*var = MPL_strdup(val);
fn_exit:
return status;
fn_fail:
goto fn_exit;
}
static HYD_status control_port_fn(char *arg, char ***argv)
{
char *port = NULL;
HYD_status status = HYD_SUCCESS;
HYDU_ERR_CHKANDJUMP(status, HYD_pmcd_pmip.upstream.server_name, HYD_INTERNAL_ERROR,
"duplicate control port setting\n");
port = MPL_strdup(**argv);
HYD_pmcd_pmip.upstream.server_name = MPL_strdup(strtok(port, ":"));
HYD_pmcd_pmip.upstream.server_port = atoi(strtok(NULL, ":"));
(*argv)++;
fn_exit:
if (port)
MPL_free(port);
return status;
fn_fail:
goto fn_exit;
}
static HYD_status fn_kvs_get(int fd, int pid, int pgid, char *args[])
{
int i, idx, found;
struct HYD_pmcd_pmi_pg_scratch *pg_scratch;
struct HYD_pg *pg;
struct HYD_proxy *proxy;
struct HYD_pmcd_pmi_kvs_pair *run;
char *key, *thrid, *cmd;
struct HYD_string_stash stash;
struct HYD_pmcd_token *tokens;
int token_count;
HYD_status status = HYD_SUCCESS;
HYDU_FUNC_ENTER();
status = HYD_pmcd_pmi_args_to_tokens(args, &tokens, &token_count);
HYDU_ERR_POP(status, "unable to convert args to tokens\n");
key = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "key");
HYDU_ERR_CHKANDJUMP(status, key == NULL, HYD_INTERNAL_ERROR, "unable to find key token\n");
thrid = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "thrid");
proxy = HYD_pmcd_pmi_find_proxy(fd);
HYDU_ASSERT(proxy, status);
pg_scratch = (struct HYD_pmcd_pmi_pg_scratch *) proxy->pg->pg_scratch;
found = 0;
for (run = pg_scratch->kvs->key_pair; run; run = run->next) {
if (!strcmp(run->key, key)) {
found = 1;
break;
}
}
if (!found) {
pg = proxy->pg;
pg_scratch = (struct HYD_pmcd_pmi_pg_scratch *) pg->pg_scratch;
idx = -1;
for (i = 0; i < pg->pg_process_count; i++)
if (pg_scratch->ecount[i].fd == fd && pg_scratch->ecount[i].pid == pid) {
idx = i;
break;
}
HYDU_ASSERT(idx != -1, status);
for (i = 0; i < pg->pg_process_count; i++) {
if (pg_scratch->ecount[i].epoch < pg_scratch->ecount[idx].epoch) {
/* We haven't reached a barrier yet; queue up request */
status = HYD_pmcd_pmi_v2_queue_req(fd, pid, pgid, args, key, &pending_reqs);
HYDU_ERR_POP(status, "unable to queue request\n");
/* We are done */
goto fn_exit;
}
}
}
HYD_STRING_STASH_INIT(stash);
HYD_STRING_STASH(stash, HYDU_strdup("cmd=kvs-get-response;"), status);
if (thrid) {
HYD_STRING_STASH(stash, HYDU_strdup("thrid="), status);
HYD_STRING_STASH(stash, HYDU_strdup(thrid), status);
HYD_STRING_STASH(stash, HYDU_strdup(";"), status);
}
if (found) {
HYD_STRING_STASH(stash, HYDU_strdup("found=TRUE;value="), status);
HYD_STRING_STASH(stash, HYDU_strdup(run->val), status);
HYD_STRING_STASH(stash, HYDU_strdup(";"), status);
}
else {
HYD_STRING_STASH(stash, HYDU_strdup("found=FALSE;"), status);
}
HYD_STRING_STASH(stash, HYDU_strdup("rc=0;"), status);
HYD_STRING_SPIT(stash, cmd, status);
status = cmd_response(fd, pid, cmd);
HYDU_ERR_POP(status, "send command failed\n");
HYDU_FREE(cmd);
fn_exit:
HYD_pmcd_pmi_free_tokens(tokens, token_count);
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
static HYD_status fn_info_getjobattr(int fd, int pid, int pgid, char *args[])
{
struct HYD_proxy *proxy;
struct HYD_pmcd_pmi_pg_scratch *pg_scratch;
struct HYD_pmcd_pmi_kvs_pair *run;
const char *key;
char *thrid, *val, *cmd;
struct HYD_string_stash stash;
struct HYD_pmcd_token *tokens;
int token_count;
HYD_status status = HYD_SUCCESS;
HYDU_FUNC_ENTER();
status = HYD_pmcd_pmi_args_to_tokens(args, &tokens, &token_count);
HYDU_ERR_POP(status, "unable to convert args to tokens\n");
key = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "key");
HYDU_ERR_CHKANDJUMP(status, key == NULL, HYD_INTERNAL_ERROR, "unable to find key token\n");
thrid = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "thrid");
proxy = HYD_pmcd_pmi_find_proxy(fd);
HYDU_ASSERT(proxy, status);
pg_scratch = (struct HYD_pmcd_pmi_pg_scratch *) proxy->pg->pg_scratch;
val = NULL;
if (!strcmp(key, "PMI_dead_processes"))
val = pg_scratch->dead_processes;
/* Try to find the key */
for (run = pg_scratch->kvs->key_pair; run; run = run->next) {
if (!strcmp(run->key, key)) {
val = run->val;
break;
}
}
HYD_STRING_STASH_INIT(stash);
HYD_STRING_STASH(stash, HYDU_strdup("cmd=info-getjobattr-response;"), status);
if (thrid) {
HYD_STRING_STASH(stash, HYDU_strdup("thrid="), status);
HYD_STRING_STASH(stash, HYDU_strdup(thrid), status);
HYD_STRING_STASH(stash, HYDU_strdup(";"), status);
}
HYD_STRING_STASH(stash, HYDU_strdup("found="), status);
if (val) {
HYD_STRING_STASH(stash, HYDU_strdup("TRUE;value="), status);
HYD_STRING_STASH(stash, HYDU_strdup(val), status);
HYD_STRING_STASH(stash, HYDU_strdup(";rc=0;"), status);
}
else {
HYD_STRING_STASH(stash, HYDU_strdup("FALSE;rc=0;"), status);
}
HYD_STRING_SPIT(stash, cmd, status);
status = cmd_response(fd, pid, cmd);
HYDU_ERR_POP(status, "send command failed\n");
HYDU_FREE(cmd);
fn_exit:
HYD_pmcd_pmi_free_tokens(tokens, token_count);
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
static HYD_status fn_info_getnodeattr(int fd, char *args[])
{
int found;
struct HYD_pmcd_pmi_kvs_pair *run;
char *key, *waitval, *thrid;
struct HYD_string_stash stash;
char *cmd;
struct HYD_pmcd_token *tokens = NULL;
int token_count;
HYD_status status = HYD_SUCCESS;
HYDU_FUNC_ENTER();
status = HYD_pmcd_pmi_args_to_tokens(args, &tokens, &token_count);
HYDU_ERR_POP(status, "unable to convert args to tokens\n");
key = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "key");
HYDU_ERR_CHKANDJUMP(status, key == NULL, HYD_INTERNAL_ERROR, "unable to find key token\n");
waitval = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "wait");
thrid = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "thrid");
/* if a predefined value is not found, we let the code fall back
* to regular search and return an error to the client */
found = 0;
for (run = HYD_pmcd_pmip.local.kvs->key_pair; run; run = run->next) {
if (!strcmp(run->key, key)) {
found = 1;
break;
}
}
if (found) { /* We found the attribute */
HYD_STRING_STASH_INIT(stash);
HYD_STRING_STASH(stash, MPL_strdup("cmd=info-getnodeattr-response;"), status);
if (thrid) {
HYD_STRING_STASH(stash, MPL_strdup("thrid="), status);
HYD_STRING_STASH(stash, MPL_strdup(thrid), status);
HYD_STRING_STASH(stash, MPL_strdup(";"), status);
}
HYD_STRING_STASH(stash, MPL_strdup("found=TRUE;value="), status);
HYD_STRING_STASH(stash, MPL_strdup(run->val), status);
HYD_STRING_STASH(stash, MPL_strdup(";rc=0;"), status);
HYD_STRING_SPIT(stash, cmd, status);
send_cmd_downstream(fd, cmd);
MPL_free(cmd);
} else if (waitval && !strcmp(waitval, "TRUE")) {
/* The client wants to wait for a response; queue up the request */
status = HYD_pmcd_pmi_v2_queue_req(fd, -1, -1, args, key, &pending_reqs);
HYDU_ERR_POP(status, "unable to queue request\n");
goto fn_exit;
} else {
/* Tell the client that we can't find the attribute */
HYD_STRING_STASH_INIT(stash);
HYD_STRING_STASH(stash, MPL_strdup("cmd=info-getnodeattr-response;"), status);
if (thrid) {
HYD_STRING_STASH(stash, MPL_strdup("thrid="), status);
HYD_STRING_STASH(stash, MPL_strdup(thrid), status);
HYD_STRING_STASH(stash, MPL_strdup(";"), status);
}
HYD_STRING_STASH(stash, MPL_strdup("found=FALSE;rc=0;"), status);
HYD_STRING_SPIT(stash, cmd, status);
send_cmd_downstream(fd, cmd);
MPL_free(cmd);
}
fn_exit:
if (tokens)
HYD_pmcd_pmi_free_tokens(tokens, token_count);
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
static HYD_status fn_kvs_put(int fd, int pid, int pgid, char *args[])
{
struct HYD_string_stash stash;
char *key, *val, *thrid, *cmd;
int ret;
struct HYD_proxy *proxy;
struct HYD_pmcd_pmi_pg_scratch *pg_scratch;
struct HYD_pmcd_token *tokens;
int token_count;
struct HYD_pmcd_pmi_v2_reqs *req;
HYD_status status = HYD_SUCCESS;
HYDU_FUNC_ENTER();
status = HYD_pmcd_pmi_args_to_tokens(args, &tokens, &token_count);
HYDU_ERR_POP(status, "unable to convert args to tokens\n");
key = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "key");
HYDU_ERR_CHKANDJUMP(status, key == NULL, HYD_INTERNAL_ERROR, "unable to find key token\n");
val = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "value");
if (val == NULL) {
/* the user sent an empty string */
val = HYDU_strdup("");
}
thrid = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "thrid");
proxy = HYD_pmcd_pmi_find_proxy(fd);
HYDU_ASSERT(proxy, status);
pg_scratch = (struct HYD_pmcd_pmi_pg_scratch *) proxy->pg->pg_scratch;
status = HYD_pmcd_pmi_add_kvs(key, val, pg_scratch->kvs, &ret);
HYDU_ERR_POP(status, "unable to put data into kvs\n");
HYD_STRING_STASH_INIT(stash);
HYD_STRING_STASH(stash, HYDU_strdup("cmd=kvs-put-response;"), status);
if (thrid) {
HYD_STRING_STASH(stash, HYDU_strdup("thrid="), status);
HYD_STRING_STASH(stash, HYDU_strdup(thrid), status);
HYD_STRING_STASH(stash, HYDU_strdup(";"), status);
}
HYD_STRING_STASH(stash, HYDU_strdup("rc="), status);
HYD_STRING_STASH(stash, HYDU_int_to_str(ret), status);
HYD_STRING_STASH(stash, HYDU_strdup(";"), status);
HYD_STRING_SPIT(stash, cmd, status);
status = cmd_response(fd, pid, cmd);
HYDU_ERR_POP(status, "send command failed\n");
HYDU_FREE(cmd);
for (req = pending_reqs; req; req = req->next) {
if (!strcmp(req->key, key)) {
/* Poke the progress engine before exiting */
status = poke_progress(key);
HYDU_ERR_POP(status, "poke progress error\n");
break;
}
}
fn_exit:
HYD_pmcd_pmi_free_tokens(tokens, token_count);
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
HYD_status HYDT_ckpoint_blcr_checkpoint(const char *prefix, int pgid, int id, int ckpt_num)
{
HYD_status status = HYD_SUCCESS;
int ret;
int fd;
cr_checkpoint_args_t my_args;
cr_checkpoint_handle_t my_handle;
char filename[256];
HYDU_FUNC_ENTER();
/* build the checkpoint filename */
MPL_snprintf(filename, sizeof(filename), "%s/context-num%d-%d-%d", prefix, ckpt_num, pgid,
id);
/* remove existing checkpoint file, if any */
(void) unlink(filename);
/* open the checkpoint file */
fd = open(filename, O_WRONLY | O_CREAT | O_TRUNC /* | O_LARGEFILE */ , 0600);
HYDU_ERR_CHKANDJUMP(status, fd < 0, HYD_INTERNAL_ERROR, "open failed: %s\n",
strerror(errno));
cr_initialize_checkpoint_args_t(&my_args);
my_args.cr_fd = fd;
my_args.cr_scope = CR_SCOPE_TREE;
/* issue the request */
ret = cr_request_checkpoint(&my_args, &my_handle);
if (ret < 0) {
HYDU_ERR_CHKANDJUMP(status, errno == CR_ENOSUPPORT, HYD_INTERNAL_ERROR,
"Checkpointing failed. Make sure BLCR kernel module is loaded. %s\n",
strerror(errno));
HYDU_ERR_SETANDJUMP(status, HYD_INTERNAL_ERROR, "cr_request_checkpoint failed, %s\n",
strerror(errno));
}
/* wait for the request to complete */
while (1) {
ret = cr_poll_checkpoint(&my_handle, NULL);
if (ret < 0) {
if ((ret == CR_POLL_CHKPT_ERR_POST) && (errno == CR_ERESTARTED)) {
HYDU_ERR_SETANDJUMP(status, HYD_INTERNAL_ERROR,
"trying to restart in a checkpoint\n");
}
else if (errno == EINTR) {
/* poll was interrupted by a signal -- retry */
}
else {
HYDU_ERR_SETANDJUMP(status, HYD_INTERNAL_ERROR,
"cr_poll_checkpoint failed: %s\n", strerror(errno));
}
}
else if (ret == 0) {
HYDU_ERR_SETANDJUMP(status, HYD_INTERNAL_ERROR,
"cr_poll_checkpoint returned 0 unexpectedly\n");
}
else {
break;
}
}
ret = close(my_args.cr_fd);
HYDU_ERR_CHKANDJUMP(status, ret, HYD_INTERNAL_ERROR, "close failed, %s\n",
strerror(errno));
fn_exit:
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
/* This waits for the restarted processes to reconnect their
stdin/out/err sockets, then sets the appropriate entries in the in
out and err arrays. This also gets the pids of the restarted
processes. */
static HYD_status wait_for_stdinouterr_sockets(int num_ranks, int *ranks, int *in, int *out,
int *err, int *pid)
{
HYD_status status = HYD_SUCCESS;
int ret;
int fd;
int i, c;
sock_ident_t id;
int num_expected_connections = num_ranks * 2; /* wait for connections for stdout and err */
HYDU_FUNC_ENTER();
/* if one of the processes is rank 0, we should wait for an
* additional connection for stdin */
for (i = 0; i < num_ranks; ++i)
if (ranks[i] == 0) {
++num_expected_connections;
break;
}
for (c = 0; c < num_expected_connections; ++c) {
size_t len;
char *id_p;
/* wait for a connection */
do {
struct sockaddr_in rmt_addr;
socklen_t sa_len = sizeof(rmt_addr);;
fd = accept(listen_fd, (struct sockaddr *) &rmt_addr, &sa_len);
} while (fd && errno == EINTR);
HYDU_ERR_CHKANDJUMP(status, fd == -1, HYD_INTERNAL_ERROR, "accept failed, %s\n",
strerror(errno));
/* read the socket identifier */
len = sizeof(id);
id_p = (char *) &id;
do {
do {
ret = read(fd, id_p, len);
} while (ret == 0 || (ret == -1 && errno == EINTR));
HYDU_ERR_CHKANDJUMP(status, ret == -1, HYD_INTERNAL_ERROR, "read failed, %s\n",
strerror(errno));
len -= ret;
id_p += ret;
} while (len);
/* determine the index for this process in the stdout/err
* arrays */
for (i = 0; i < num_ranks; ++i)
if (ranks[i] == id.rank)
break;
HYDU_ASSERT(i < num_ranks, status);
/* assign the fd */
switch (id.socktype) {
case IN_SOCK:
HYDU_ASSERT(id.rank == 0, status);
*in = fd;
break;
case OUT_SOCK:
out[i] = fd;
break;
case ERR_SOCK:
err[i] = fd;
break;
default:
HYDU_ASSERT(0, status);
break;
}
/* assign the pid */
pid[i] = id.pid;
}
ret = close(listen_fd);
HYDU_ERR_CHKANDJUMP(status, ret, HYD_INTERNAL_ERROR, "close of listener port failed, %s\n",
strerror(errno));
fn_exit:
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
static HYD_status fn_get(int fd, int pid, int pgid, char *args[])
{
int i;
struct HYD_proxy *proxy;
struct HYD_pmcd_pmi_pg_scratch *pg_scratch;
struct HYD_pmcd_pmi_kvs_pair *run;
char *kvsname, *key, *val;
char *tmp[HYD_NUM_TMP_STRINGS], *cmd;
struct HYD_pmcd_token *tokens;
int token_count;
HYD_status status = HYD_SUCCESS;
HYDU_FUNC_ENTER();
status = HYD_pmcd_pmi_args_to_tokens(args, &tokens, &token_count);
HYDU_ERR_POP(status, "unable to convert args to tokens\n");
kvsname = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "kvsname");
HYDU_ERR_CHKANDJUMP(status, kvsname == NULL, HYD_INTERNAL_ERROR,
"unable to find token: kvsname\n");
key = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "key");
HYDU_ERR_CHKANDJUMP(status, key == NULL, HYD_INTERNAL_ERROR,
"unable to find token: key\n");
proxy = HYD_pmcd_pmi_find_proxy(fd);
HYDU_ASSERT(proxy, status);
pg_scratch = (struct HYD_pmcd_pmi_pg_scratch *) proxy->pg->pg_scratch;
val = NULL;
if (!strcmp(key, "PMI_dead_processes")) {
val = pg_scratch->dead_processes;
goto found_val;
}
if (strcmp(pg_scratch->kvs->kvs_name, kvsname))
HYDU_ERR_SETANDJUMP(status, HYD_INTERNAL_ERROR,
"kvsname (%s) does not match this group's kvs space (%s)\n",
kvsname, pg_scratch->kvs->kvs_name);
/* Try to find the key */
for (run = pg_scratch->kvs->key_pair; run; run = run->next) {
if (!strcmp(run->key, key)) {
val = run->val;
break;
}
}
found_val:
i = 0;
tmp[i++] = HYDU_strdup("cmd=get_result rc=");
if (val) {
tmp[i++] = HYDU_strdup("0 msg=success value=");
tmp[i++] = HYDU_strdup(val);
}
else {
tmp[i++] = HYDU_strdup("-1 msg=key_");
tmp[i++] = HYDU_strdup(key);
tmp[i++] = HYDU_strdup("_not_found value=unknown");
}
tmp[i++] = HYDU_strdup("\n");
tmp[i++] = NULL;
status = HYDU_str_alloc_and_join(tmp, &cmd);
HYDU_ERR_POP(status, "unable to join strings\n");
HYDU_free_strlist(tmp);
status = cmd_response(fd, pid, cmd);
HYDU_ERR_POP(status, "error writing PMI line\n");
HYDU_FREE(cmd);
fn_exit:
HYD_pmcd_pmi_free_tokens(tokens, token_count);
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
static HYD_status fn_put(int fd, int pid, int pgid, char *args[])
{
int i, ret;
struct HYD_proxy *proxy;
struct HYD_pmcd_pmi_pg_scratch *pg_scratch;
char *kvsname, *key, *val;
char *tmp[HYD_NUM_TMP_STRINGS], *cmd;
struct HYD_pmcd_token *tokens;
int token_count;
HYD_status status = HYD_SUCCESS;
HYDU_FUNC_ENTER();
status = HYD_pmcd_pmi_args_to_tokens(args, &tokens, &token_count);
HYDU_ERR_POP(status, "unable to convert args to tokens\n");
kvsname = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "kvsname");
HYDU_ERR_CHKANDJUMP(status, kvsname == NULL, HYD_INTERNAL_ERROR,
"unable to find token: kvsname\n");
key = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "key");
HYDU_ERR_CHKANDJUMP(status, key == NULL, HYD_INTERNAL_ERROR,
"unable to find token: key\n");
val = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "value");
if (val == NULL) {
/* the user sent an empty string */
val = HYDU_strdup("");
}
proxy = HYD_pmcd_pmi_find_proxy(fd);
HYDU_ASSERT(proxy, status);
pg_scratch = (struct HYD_pmcd_pmi_pg_scratch *) proxy->pg->pg_scratch;
if (strcmp(pg_scratch->kvs->kvs_name, kvsname))
HYDU_ERR_SETANDJUMP(status, HYD_INTERNAL_ERROR,
"kvsname (%s) does not match this group's kvs space (%s)\n",
kvsname, pg_scratch->kvs->kvs_name);
status = HYD_pmcd_pmi_add_kvs(key, val, pg_scratch->kvs, &ret);
HYDU_ERR_POP(status, "unable to add keypair to kvs\n");
i = 0;
tmp[i++] = HYDU_strdup("cmd=put_result rc=");
tmp[i++] = HYDU_int_to_str(ret);
if (ret == 0) {
tmp[i++] = HYDU_strdup(" msg=success");
}
else {
tmp[i++] = HYDU_strdup(" msg=duplicate_key");
tmp[i++] = HYDU_strdup(key);
}
tmp[i++] = HYDU_strdup("\n");
tmp[i++] = NULL;
status = HYDU_str_alloc_and_join(tmp, &cmd);
HYDU_ERR_POP(status, "unable to join strings\n");
HYDU_free_strlist(tmp);
status = cmd_response(fd, pid, cmd);
HYDU_ERR_POP(status, "error writing PMI line\n");
HYDU_FREE(cmd);
fn_exit:
HYD_pmcd_pmi_free_tokens(tokens, token_count);
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
static HYD_status fn_spawn(int fd, int pid, int pgid, char *args[])
{
struct HYD_pg *pg;
struct HYD_pmcd_pmi_pg_scratch *pg_scratch;
struct HYD_proxy *proxy;
struct HYD_pmcd_token *tokens;
struct HYD_exec *exec_list = NULL, *exec;
struct HYD_env *env;
struct HYD_node *node;
char key[PMI_MAXKEYLEN], *val;
int nprocs, preput_num, info_num, ret;
char *execname, *path = NULL;
struct HYD_pmcd_token_segment *segment_list = NULL;
int token_count, i, j, k, new_pgid, total_spawns;
int argcnt, num_segments;
char *control_port, *proxy_args[HYD_NUM_TMP_STRINGS] = { NULL };
char *tmp[HYD_NUM_TMP_STRINGS];
HYD_status status = HYD_SUCCESS;
HYDU_FUNC_ENTER();
for (i = 0; args[i]; i++)
mcmd_args[mcmd_num_args++] = HYDU_strdup(args[i]);
mcmd_args[mcmd_num_args] = NULL;
status = HYD_pmcd_pmi_args_to_tokens(mcmd_args, &tokens, &token_count);
HYDU_ERR_POP(status, "unable to convert args to tokens\n");
/* Here's the order of things we do:
*
* 1. Break the token list into multiple segments, each segment
* corresponding to a command. Each command represents
* information for one executable.
*
* 2. Allocate a process group for the new set of spawned
* processes
*
* 3. Get all the common keys and deal with them
*
* 4. Create an executable list based on the segments.
*
* 5. Create a proxy list using the created executable list and
* spawn it.
*/
/* Break the token list into multiple segments and create an
* executable list based on the segments. */
val = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "totspawns");
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: totspawns\n");
total_spawns = atoi(val);
HYDU_MALLOC(segment_list, struct HYD_pmcd_token_segment *,
total_spawns * sizeof(struct HYD_pmcd_token_segment), status);
segment_tokens(tokens, token_count, segment_list, &num_segments);
if (num_segments != total_spawns) {
/* We didn't read the entire PMI string; wait for the rest to
* arrive */
goto fn_exit;
}
else {
/* Got the entire PMI string; free the arguments and reset */
HYDU_free_strlist(mcmd_args);
mcmd_num_args = 0;
}
/* Allocate a new process group */
for (pg = &HYD_server_info.pg_list; pg->next; pg = pg->next);
new_pgid = pg->pgid + 1;
status = HYDU_alloc_pg(&pg->next, new_pgid);
HYDU_ERR_POP(status, "unable to allocate process group\n");
pg = pg->next;
proxy = HYD_pmcd_pmi_find_proxy(fd);
HYDU_ASSERT(proxy, status);
pg->spawner_pg = proxy->pg;
for (j = 0; j < total_spawns; j++) {
/* For each segment, we create an exec structure */
val = HYD_pmcd_pmi_find_token_keyval(&tokens[segment_list[j].start_idx],
segment_list[j].token_count, "nprocs");
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: nprocs\n");
nprocs = atoi(val);
pg->pg_process_count += nprocs;
val = HYD_pmcd_pmi_find_token_keyval(&tokens[segment_list[j].start_idx],
segment_list[j].token_count, "argcnt");
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: argcnt\n");
argcnt = atoi(val);
val = HYD_pmcd_pmi_find_token_keyval(&tokens[segment_list[j].start_idx],
segment_list[j].token_count, "info_num");
if (val)
info_num = atoi(val);
else
info_num = 0;
if (exec_list == NULL) {
status = HYDU_alloc_exec(&exec_list);
HYDU_ERR_POP(status, "unable to allocate exec\n");
exec_list->appnum = 0;
exec = exec_list;
}
else {
for (exec = exec_list; exec->next; exec = exec->next);
status = HYDU_alloc_exec(&exec->next);
HYDU_ERR_POP(status, "unable to allocate exec\n");
exec->next->appnum = exec->appnum + 1;
exec = exec->next;
}
/* Info keys */
for (i = 0; i < info_num; i++) {
char *info_key, *info_val;
HYDU_snprintf(key, PMI_MAXKEYLEN, "info_key_%d", i);
val = HYD_pmcd_pmi_find_token_keyval(&tokens[segment_list[j].start_idx],
segment_list[j].token_count, key);
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: %s\n", key);
info_key = val;
HYDU_snprintf(key, PMI_MAXKEYLEN, "info_val_%d", i);
val = HYD_pmcd_pmi_find_token_keyval(&tokens[segment_list[j].start_idx],
segment_list[j].token_count, key);
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: %s\n", key);
info_val = val;
if (!strcmp(info_key, "path")) {
path = HYDU_strdup(info_val);
}
else if (!strcmp(info_key, "wdir")) {
exec->wdir = HYDU_strdup(info_val);
}
else if (!strcmp(info_key, "host")) {
status = HYDU_process_mfile_token(info_val, 1, &pg->user_node_list);
HYDU_ERR_POP(status, "error create node list\n");
}
else if (!strcmp(info_key, "hostfile")) {
status =
HYDU_parse_hostfile(info_val, &pg->user_node_list, HYDU_process_mfile_token);
HYDU_ERR_POP(status, "error parsing hostfile\n");
}
else {
/* Unrecognized info key; ignore */
}
}
status = HYDU_correct_wdir(&exec->wdir);
HYDU_ERR_POP(status, "unable to correct wdir\n");
val = HYD_pmcd_pmi_find_token_keyval(&tokens[segment_list[j].start_idx],
segment_list[j].token_count, "execname");
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: execname\n");
if (path == NULL)
execname = HYDU_strdup(val);
else {
i = 0;
tmp[i++] = HYDU_strdup(path);
tmp[i++] = HYDU_strdup("/");
tmp[i++] = HYDU_strdup(val);
tmp[i++] = NULL;
status = HYDU_str_alloc_and_join(tmp, &execname);
HYDU_ERR_POP(status, "error while joining strings\n");
HYDU_free_strlist(tmp);
}
i = 0;
exec->exec[i++] = execname;
for (k = 0; k < argcnt; k++) {
HYDU_snprintf(key, PMI_MAXKEYLEN, "arg%d", k + 1);
val = HYD_pmcd_pmi_find_token_keyval(&tokens[segment_list[j].start_idx],
segment_list[j].token_count, key);
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: %s\n", key);
exec->exec[i++] = HYDU_strdup(val);
}
exec->exec[i++] = NULL;
exec->proc_count = nprocs;
/* It is not clear what kind of environment needs to get
* passed to the spawned process. Don't set anything here, and
* let the proxy do whatever it does by default. */
exec->env_prop = NULL;
status = HYDU_env_create(&env, "PMI_SPAWNED", "1");
HYDU_ERR_POP(status, "unable to create PMI_SPAWNED environment\n");
exec->user_env = env;
}
status = HYD_pmcd_pmi_alloc_pg_scratch(pg);
HYDU_ERR_POP(status, "unable to allocate pg scratch space\n");
if (pg->user_node_list) {
pg->pg_core_count = 0;
for (i = 0, node = pg->user_node_list; node; node = node->next, i++) {
pg->pg_core_count += node->core_count;
node->node_id = i;
}
}
else {
pg->pg_core_count = HYD_server_info.pg_list.pg_core_count;
}
pg->pg_process_count = 0;
for (exec = exec_list; exec; exec = exec->next)
pg->pg_process_count += exec->proc_count;
pg_scratch = (struct HYD_pmcd_pmi_pg_scratch *) pg->pg_scratch;
/* Get the common keys and deal with them */
val = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "preput_num");
if (val)
preput_num = atoi(val);
else
preput_num = 0;
for (i = 0; i < preput_num; i++) {
char *preput_key, *preput_val;
HYDU_snprintf(key, PMI_MAXKEYLEN, "preput_key_%d", i);
val = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, key);
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: %s\n", key);
preput_key = val;
HYDU_snprintf(key, PMI_MAXKEYLEN, "preput_val_%d", i);
val = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, key);
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: %s\n", key);
preput_val = val;
status = HYD_pmcd_pmi_add_kvs(preput_key, preput_val, pg_scratch->kvs, &ret);
HYDU_ERR_POP(status, "unable to add keypair to kvs\n");
}
/* Create the proxy list */
if (pg->user_node_list) {
status = HYDU_create_proxy_list(exec_list, pg->user_node_list, pg);
HYDU_ERR_POP(status, "error creating proxy list\n");
}
else {
status = HYDU_create_proxy_list(exec_list, HYD_server_info.node_list, pg);
HYDU_ERR_POP(status, "error creating proxy list\n");
}
HYDU_free_exec_list(exec_list);
status = HYDU_sock_create_and_listen_portstr(HYD_server_info.user_global.iface,
HYD_server_info.local_hostname,
HYD_server_info.port_range, &control_port,
HYD_pmcd_pmiserv_control_listen_cb,
(void *) (size_t) new_pgid);
HYDU_ERR_POP(status, "unable to create PMI port\n");
if (HYD_server_info.user_global.debug)
HYDU_dump(stdout, "Got a control port string of %s\n", control_port);
/* Go to the last PG */
for (pg = &HYD_server_info.pg_list; pg->next; pg = pg->next);
status = HYD_pmcd_pmi_fill_in_proxy_args(proxy_args, control_port, new_pgid);
HYDU_ERR_POP(status, "unable to fill in proxy arguments\n");
HYDU_FREE(control_port);
status = HYD_pmcd_pmi_fill_in_exec_launch_info(pg);
HYDU_ERR_POP(status, "unable to fill in executable arguments\n");
status = HYDT_bsci_launch_procs(proxy_args, pg->proxy_list, NULL);
HYDU_ERR_POP(status, "launcher cannot launch processes\n");
{
char *cmd_str[HYD_NUM_TMP_STRINGS], *cmd;
i = 0;
cmd_str[i++] = HYDU_strdup("cmd=spawn_result rc=0");
cmd_str[i++] = HYDU_strdup("\n");
cmd_str[i++] = NULL;
status = HYDU_str_alloc_and_join(cmd_str, &cmd);
HYDU_ERR_POP(status, "unable to join strings\n");
HYDU_free_strlist(cmd_str);
status = cmd_response(fd, pid, cmd);
HYDU_ERR_POP(status, "error writing PMI line\n");
HYDU_FREE(cmd);
}
fn_exit:
HYD_pmcd_pmi_free_tokens(tokens, token_count);
HYDU_free_strlist(proxy_args);
if (segment_list)
HYDU_FREE(segment_list);
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
static HYD_status fn_spawn(int fd, int pid, int pgid, char *args[])
{
struct HYD_pg *pg;
struct HYD_pmcd_pmi_pg_scratch *pg_scratch;
struct HYD_proxy *proxy;
struct HYD_pmcd_token *tokens;
struct HYD_exec *exec_list = NULL, *exec;
struct HYD_env *env;
struct HYD_node *node;
char *thrid;
char key[PMI_MAXKEYLEN], *val;
int maxprocs, preputcount, infokeycount, ret;
int ncmds;
char *execname, *path = NULL;
struct HYD_pmcd_token_segment *segment_list = NULL;
int token_count, i, j, k, new_pgid;
int argcnt, num_segments;
struct HYD_string_stash proxy_stash;
char *control_port;
struct HYD_string_stash stash;
HYD_status status = HYD_SUCCESS;
HYDU_FUNC_ENTER();
status = HYD_pmcd_pmi_args_to_tokens(args, &tokens, &token_count);
HYDU_ERR_POP(status, "unable to convert args to tokens\n");
thrid = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "thrid");
/* Here's the order of things we do:
*
* 1. Break the token list into multiple segments, each segment
* corresponding to a command. Each command represents
* information for one executable.
*
* 2. Allocate a process group for the new set of spawned
* processes
*
* 3. Get all the common keys and deal with them
*
* 4. Create an executable list based on the segments.
*
* 5. Create a proxy list using the created executable list and
* spawn it.
*/
val = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "ncmds");
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR, "unable to find token: ncmds\n");
ncmds = atoi(val);
HYDU_MALLOC(segment_list, struct HYD_pmcd_token_segment *,
(ncmds + 1) * sizeof(struct HYD_pmcd_token_segment), status);
segment_tokens(tokens, token_count, segment_list, &num_segments);
HYDU_ASSERT((ncmds + 1) == num_segments, status);
/* Allocate a new process group */
for (pg = &HYD_server_info.pg_list; pg->next; pg = pg->next);
new_pgid = pg->pgid + 1;
status = HYDU_alloc_pg(&pg->next, new_pgid);
HYDU_ERR_POP(status, "unable to allocate process group\n");
pg = pg->next;
proxy = HYD_pmcd_pmi_find_proxy(fd);
HYDU_ASSERT(proxy, status);
pg->spawner_pg = proxy->pg;
for (j = 1; j <= ncmds; j++) {
/* For each segment, we create an exec structure */
val = HYD_pmcd_pmi_find_token_keyval(&tokens[segment_list[j].start_idx],
segment_list[j].token_count, "maxprocs");
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: maxprocs\n");
maxprocs = atoi(val);
pg->pg_process_count += maxprocs;
val = HYD_pmcd_pmi_find_token_keyval(&tokens[segment_list[j].start_idx],
segment_list[j].token_count, "argc");
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: argc\n");
argcnt = atoi(val);
val = HYD_pmcd_pmi_find_token_keyval(&tokens[segment_list[j].start_idx],
segment_list[j].token_count, "infokeycount");
if (val)
infokeycount = atoi(val);
else
infokeycount = 0;
if (exec_list == NULL) {
status = HYDU_alloc_exec(&exec_list);
HYDU_ERR_POP(status, "unable to allocate exec\n");
exec_list->appnum = 0;
exec = exec_list;
}
else {
for (exec = exec_list; exec->next; exec = exec->next);
status = HYDU_alloc_exec(&exec->next);
HYDU_ERR_POP(status, "unable to allocate exec\n");
exec->next->appnum = exec->appnum + 1;
exec = exec->next;
}
/* Info keys */
for (i = 0; i < infokeycount; i++) {
char *info_key, *info_val;
HYDU_snprintf(key, PMI_MAXKEYLEN, "infokey%d", i);
val = HYD_pmcd_pmi_find_token_keyval(&tokens[segment_list[j].start_idx],
segment_list[j].token_count, key);
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: %s\n", key);
info_key = val;
HYDU_snprintf(key, PMI_MAXKEYLEN, "infoval%d", i);
val = HYD_pmcd_pmi_find_token_keyval(&tokens[segment_list[j].start_idx],
segment_list[j].token_count, key);
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: %s\n", key);
info_val = val;
if (!strcmp(info_key, "path")) {
path = HYDU_strdup(info_val);
}
else if (!strcmp(info_key, "wdir")) {
exec->wdir = HYDU_strdup(info_val);
}
else if (!strcmp(info_key, "host") || !strcmp(info_key, "hosts")) {
char *host = strtok(info_val, ",");
while (host) {
status = HYDU_process_mfile_token(host, 1, &pg->user_node_list);
HYDU_ERR_POP(status, "error creating node list\n");
host = strtok(NULL, ",");
}
}
else if (!strcmp(info_key, "hostfile")) {
status = HYDU_parse_hostfile(info_val, &pg->user_node_list,
HYDU_process_mfile_token);
HYDU_ERR_POP(status, "error parsing hostfile\n");
}
else {
/* Unrecognized info key; ignore */
}
}
status = HYDU_correct_wdir(&exec->wdir);
HYDU_ERR_POP(status, "unable to correct wdir\n");
val = HYD_pmcd_pmi_find_token_keyval(&tokens[segment_list[j].start_idx],
segment_list[j].token_count, "subcmd");
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: subcmd\n");
if (path == NULL)
execname = HYDU_strdup(val);
else {
HYD_STRING_STASH_INIT(stash);
HYD_STRING_STASH(stash, HYDU_strdup(path), status);
HYD_STRING_STASH(stash, HYDU_strdup("/"), status);
HYD_STRING_STASH(stash, HYDU_strdup(val), status);
HYD_STRING_SPIT(stash, execname, status);
}
i = 0;
exec->exec[i++] = execname;
for (k = 0; k < argcnt; k++) {
HYDU_snprintf(key, PMI_MAXKEYLEN, "argv%d", k);
val = HYD_pmcd_pmi_find_token_keyval(&tokens[segment_list[j].start_idx],
segment_list[j].token_count, key);
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: %s\n", key);
exec->exec[i++] = HYDU_strdup(val);
}
exec->exec[i++] = NULL;
exec->proc_count = maxprocs;
/* It is not clear what kind of environment needs to get
* passed to the spawned process. Don't set anything here, and
* let the proxy do whatever it does by default. */
exec->env_prop = NULL;
status = HYDU_env_create(&env, "PMI_SPAWNED", "1");
HYDU_ERR_POP(status, "unable to create PMI_SPAWNED environment\n");
exec->user_env = env;
}
status = HYD_pmcd_pmi_alloc_pg_scratch(pg);
HYDU_ERR_POP(status, "unable to allocate pg scratch space\n");
pg->pg_process_count = 0;
for (exec = exec_list; exec; exec = exec->next)
pg->pg_process_count += exec->proc_count;
pg_scratch = (struct HYD_pmcd_pmi_pg_scratch *) pg->pg_scratch;
/* Get the common keys and deal with them */
val = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "preputcount");
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: preputcount\n");
preputcount = atoi(val);
for (i = 0; i < preputcount; i++) {
char *preput_key, *preput_val;
HYDU_snprintf(key, PMI_MAXKEYLEN, "ppkey%d", i);
val = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, key);
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: %s\n", key);
preput_key = val;
HYDU_snprintf(key, PMI_MAXKEYLEN, "ppval%d", i);
val = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, key);
HYDU_ERR_CHKANDJUMP(status, val == NULL, HYD_INTERNAL_ERROR,
"unable to find token: %s\n", key);
preput_val = val;
status = HYD_pmcd_pmi_add_kvs(preput_key, preput_val, pg_scratch->kvs, &ret);
HYDU_ERR_POP(status, "unable to add keypair to kvs\n");
}
/* Create the proxy list */
if (pg->user_node_list) {
status = HYDU_create_proxy_list(exec_list, pg->user_node_list, pg);
HYDU_ERR_POP(status, "error creating proxy list\n");
}
else {
status = HYDU_create_proxy_list(exec_list, HYD_server_info.node_list, pg);
HYDU_ERR_POP(status, "error creating proxy list\n");
}
HYDU_free_exec_list(exec_list);
if (pg->user_node_list) {
pg->pg_core_count = 0;
for (i = 0, node = pg->user_node_list; node; node = node->next, i++)
pg->pg_core_count += node->core_count;
}
else {
pg->pg_core_count = 0;
for (proxy = pg->proxy_list; proxy; proxy = proxy->next)
pg->pg_core_count += proxy->node->core_count;
}
status = HYDU_sock_create_and_listen_portstr(HYD_server_info.user_global.iface,
HYD_server_info.localhost,
HYD_server_info.port_range, &control_port,
HYD_pmcd_pmiserv_control_listen_cb,
(void *) (size_t) new_pgid);
HYDU_ERR_POP(status, "unable to create PMI port\n");
if (HYD_server_info.user_global.debug)
HYDU_dump(stdout, "Got a control port string of %s\n", control_port);
/* Go to the last PG */
for (pg = &HYD_server_info.pg_list; pg->next; pg = pg->next);
status = HYD_pmcd_pmi_fill_in_proxy_args(&proxy_stash, control_port, new_pgid);
HYDU_ERR_POP(status, "unable to fill in proxy arguments\n");
HYDU_FREE(control_port);
status = HYD_pmcd_pmi_fill_in_exec_launch_info(pg);
HYDU_ERR_POP(status, "unable to fill in executable arguments\n");
status = HYDT_bsci_launch_procs(proxy_stash.strlist, pg->proxy_list, NULL);
HYDU_ERR_POP(status, "launcher cannot launch processes\n");
{
char *cmd;
HYD_STRING_STASH_INIT(stash);
HYD_STRING_STASH(stash, HYDU_strdup("cmd=spawn-response;"), status);
if (thrid) {
HYD_STRING_STASH(stash, HYDU_strdup("thrid="), status);
HYD_STRING_STASH(stash, HYDU_strdup(thrid), status);
HYD_STRING_STASH(stash, HYDU_strdup(";"), status);
}
HYD_STRING_STASH(stash, HYDU_strdup("rc=0;"), status);
HYD_STRING_STASH(stash, HYDU_strdup("jobid="), status);
HYD_STRING_STASH(stash, HYDU_strdup(pg_scratch->kvs->kvsname), status);
HYD_STRING_STASH(stash, HYDU_strdup(";"), status);
HYD_STRING_STASH(stash, HYDU_strdup("nerrs=0;"), status);
HYD_STRING_SPIT(stash, cmd, status);
status = cmd_response(fd, pid, cmd);
HYDU_ERR_POP(status, "send command failed\n");
HYDU_FREE(cmd);
}
fn_exit:
HYD_pmcd_pmi_free_tokens(tokens, token_count);
HYD_STRING_STASH_FREE(proxy_stash);
if (segment_list)
HYDU_FREE(segment_list);
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
static HYD_status handle_bitmap_binding(const char *binding, const char *mapping)
{
int i, j, k, bind_count, map_count, cache_depth = 0, bind_depth = 0, map_depth = 0;
int total_map_objs, total_bind_objs, num_pus_in_map_domain, num_pus_in_bind_domain,
total_map_domains;
hwloc_obj_t map_obj, bind_obj, *start_pu;
hwloc_cpuset_t *map_domains;
char *bind_str, *map_str;
HYD_status status = HYD_SUCCESS;
HYDU_FUNC_ENTER();
/* split out the count fields */
status = split_count_field(binding, &bind_str, &bind_count);
HYDU_ERR_POP(status, "error splitting count field\n");
status = split_count_field(mapping, &map_str, &map_count);
HYDU_ERR_POP(status, "error splitting count field\n");
/* get the binding object */
if (!strcmp(bind_str, "board"))
bind_depth = hwloc_get_type_or_above_depth(topology, HWLOC_OBJ_MACHINE);
else if (!strcmp(bind_str, "numa"))
bind_depth = hwloc_get_type_or_above_depth(topology, HWLOC_OBJ_NODE);
else if (!strcmp(bind_str, "socket"))
bind_depth = hwloc_get_type_or_above_depth(topology, HWLOC_OBJ_SOCKET);
else if (!strcmp(bind_str, "core"))
bind_depth = hwloc_get_type_or_above_depth(topology, HWLOC_OBJ_CORE);
else if (!strcmp(bind_str, "hwthread"))
bind_depth = hwloc_get_type_or_above_depth(topology, HWLOC_OBJ_PU);
else {
/* check if it's in the l*cache format */
cache_depth = parse_cache_string(bind_str);
if (!cache_depth) {
HYDU_ERR_SETANDJUMP(status, HYD_INTERNAL_ERROR,
"unrecognized binding string \"%s\"\n", binding);
}
bind_depth = hwloc_get_cache_type_depth(topology, cache_depth, -1);
}
/* get the mapping */
if (!strcmp(map_str, "board"))
map_depth = hwloc_get_type_or_above_depth(topology, HWLOC_OBJ_MACHINE);
else if (!strcmp(map_str, "numa"))
map_depth = hwloc_get_type_or_above_depth(topology, HWLOC_OBJ_NODE);
else if (!strcmp(map_str, "socket"))
map_depth = hwloc_get_type_or_above_depth(topology, HWLOC_OBJ_SOCKET);
else if (!strcmp(map_str, "core"))
map_depth = hwloc_get_type_or_above_depth(topology, HWLOC_OBJ_CORE);
else if (!strcmp(map_str, "hwthread"))
map_depth = hwloc_get_type_or_above_depth(topology, HWLOC_OBJ_PU);
else {
cache_depth = parse_cache_string(map_str);
if (!cache_depth) {
HYDU_ERR_SETANDJUMP(status, HYD_INTERNAL_ERROR,
"unrecognized mapping string \"%s\"\n", mapping);
}
map_depth = hwloc_get_cache_type_depth(topology, cache_depth, -1);
}
/*
* Process Affinity Algorithm:
*
* The code below works in 3 stages. The end result is an array of all the possible
* binding bitmaps for a system, based on the options specified.
*
* 1. Define all possible mapping "domains" in a system. A mapping domain is a group
* of hardware elements found by traversing the topology. Each traversal skips the
* number of elements the user specified in the mapping string. The traversal ends
* when the next mapping domain == the first mapping domain. Note that if the
* mapping string defines a domain that is larger than the system size, we exit
* with an error.
*
* 2. Define the number of possible binding domains within a mapping domain. This
* process is similar to step 1, in that we traverse the mapping domain finding
* all possible bind combinations, stopping when a duplicate of the first binding
* is reached. If a binding is larger (in # of PUs) than the mapping domain,
* the number of possible bindings for that domain is 1. In this stage, we also
* locate the first PU in each mapping domain for use later during binding.
*
* 3. Create the binding bitmaps. We allocate an array of bitmaps and fill them in
* with all possible bindings. The starting PU in each mapping domain is advanced
* if and when we wrap around to the beginning of the mapping domains. This ensures
* that we do not repeat.
*
*/
/* calculate the number of map domains */
total_map_objs = hwloc_get_nbobjs_by_depth(topology, map_depth);
num_pus_in_map_domain = (HYDT_topo_hwloc_info.total_num_pus / total_map_objs) * map_count;
HYDU_ERR_CHKANDJUMP(status, num_pus_in_map_domain > HYDT_topo_hwloc_info.total_num_pus,
HYD_INTERNAL_ERROR, "mapping option \"%s\" larger than total system size\n",
mapping);
/* The number of total_map_domains should be large enough to
* contain all contiguous map object collections of length
* map_count. For example, if the map object is "socket" and the
* map_count is 3, on a system with 4 sockets, the following map
* domains should be included: (0,1,2), (3,0,1), (2,3,0), (1,2,3).
* We do this by finding how many times we need to replicate the
* list of the map objects so that an integral number of map
* domains can map to them. In the above case, the list of map
* objects is replicated 3 times. */
for (i = 1; (i * total_map_objs) % map_count; i++);
total_map_domains = (i * total_map_objs) / map_count;
/* initialize the map domains */
HYDU_MALLOC_OR_JUMP(map_domains, hwloc_bitmap_t *, total_map_domains * sizeof(hwloc_bitmap_t),
status);
HYDU_MALLOC_OR_JUMP(start_pu, hwloc_obj_t *, total_map_domains * sizeof(hwloc_obj_t), status);
/* For each map domain, find the next map object (first map object
* for the first map domain) and add the following "map_count"
* number of contiguous map objects, wrapping to the first one if
* needed, to the map domain. Store the first PU in the first map
* object of the map domain as "start_pu". This is needed later
* for the actual binding. */
map_obj = NULL;
for (i = 0; i < total_map_domains; i++) {
map_domains[i] = hwloc_bitmap_alloc();
hwloc_bitmap_zero(map_domains[i]);
for (j = 0; j < map_count; j++) {
map_obj = hwloc_get_next_obj_by_depth(topology, map_depth, map_obj);
/* map_obj will be NULL if it reaches the end. call again to wrap around */
if (!map_obj)
map_obj = hwloc_get_next_obj_by_depth(topology, map_depth, map_obj);
if (j == 0)
start_pu[i] =
hwloc_get_obj_inside_cpuset_by_type(topology, map_obj->cpuset, HWLOC_OBJ_PU, 0);
hwloc_bitmap_or(map_domains[i], map_domains[i], map_obj->cpuset);
}
}
/* Find the possible binding domains is similar to that of map
* domains. But if a binding domain is larger (in # of PUs) than
* the mapping domain, the number of possible bindings for that
* domain is 1. */
/* calculate the number of possible bindings and allocate bitmaps for them */
total_bind_objs = hwloc_get_nbobjs_by_depth(topology, bind_depth);
num_pus_in_bind_domain = (HYDT_topo_hwloc_info.total_num_pus / total_bind_objs) * bind_count;
if (num_pus_in_bind_domain < num_pus_in_map_domain) {
for (i = 1; (i * num_pus_in_map_domain) % num_pus_in_bind_domain; i++);
HYDT_topo_hwloc_info.num_bitmaps =
(i * num_pus_in_map_domain * total_map_domains) / num_pus_in_bind_domain;
}
else {
HYDT_topo_hwloc_info.num_bitmaps = total_map_domains;
}
/* initialize bitmaps */
HYDU_MALLOC_OR_JUMP(HYDT_topo_hwloc_info.bitmap, hwloc_bitmap_t *,
HYDT_topo_hwloc_info.num_bitmaps * sizeof(hwloc_bitmap_t), status);
for (i = 0; i < HYDT_topo_hwloc_info.num_bitmaps; i++) {
HYDT_topo_hwloc_info.bitmap[i] = hwloc_bitmap_alloc();
hwloc_bitmap_zero(HYDT_topo_hwloc_info.bitmap[i]);
}
/* do bindings */
i = 0;
while (i < HYDT_topo_hwloc_info.num_bitmaps) {
for (j = 0; j < total_map_domains; j++) {
bind_obj = hwloc_get_ancestor_obj_by_depth(topology, bind_depth, start_pu[j]);
for (k = 0; k < bind_count; k++) {
hwloc_bitmap_or(HYDT_topo_hwloc_info.bitmap[i], HYDT_topo_hwloc_info.bitmap[i],
bind_obj->cpuset);
/* if the binding is smaller than the mapping domain, wrap around inside that domain */
if (num_pus_in_bind_domain < num_pus_in_map_domain) {
bind_obj =
hwloc_get_next_obj_inside_cpuset_by_depth(topology, map_domains[j],
bind_depth, bind_obj);
if (!bind_obj)
bind_obj =
hwloc_get_next_obj_inside_cpuset_by_depth(topology, map_domains[j],
bind_depth, bind_obj);
}
else {
bind_obj = hwloc_get_next_obj_by_depth(topology, bind_depth, bind_obj);
if (!bind_obj)
bind_obj = hwloc_get_next_obj_by_depth(topology, bind_depth, bind_obj);
}
}
i++;
/* advance the starting position for this map domain, if needed */
if (num_pus_in_bind_domain < num_pus_in_map_domain) {
for (k = 0; k < num_pus_in_bind_domain; k++) {
start_pu[j] = hwloc_get_next_obj_inside_cpuset_by_type(topology, map_domains[j],
HWLOC_OBJ_PU,
start_pu[j]);
if (!start_pu[j])
start_pu[j] =
hwloc_get_next_obj_inside_cpuset_by_type(topology, map_domains[j],
HWLOC_OBJ_PU, start_pu[j]);
}
}
}
}
/* free temporary memory */
MPL_free(map_domains);
MPL_free(start_pu);
fn_exit:
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
static HYD_status fn_fullinit(int fd, char *args[])
{
int id, i;
char *rank_str;
struct HYD_string_stash stash;
char *cmd;
struct HYD_pmcd_token *tokens;
int token_count;
HYD_status status = HYD_SUCCESS;
HYDU_FUNC_ENTER();
status = HYD_pmcd_pmi_args_to_tokens(args, &tokens, &token_count);
HYDU_ERR_POP(status, "unable to convert args to tokens\n");
rank_str = HYD_pmcd_pmi_find_token_keyval(tokens, token_count, "pmirank");
HYDU_ERR_CHKANDJUMP(status, rank_str == NULL, HYD_INTERNAL_ERROR,
"unable to find pmirank token\n");
id = atoi(rank_str);
/* Store the PMI_RANK to fd mapping */
for (i = 0; i < HYD_pmcd_pmip.local.proxy_process_count; i++) {
if (HYD_pmcd_pmip.downstream.pmi_rank[i] == id) {
HYD_pmcd_pmip.downstream.pmi_fd[i] = fd;
HYD_pmcd_pmip.downstream.pmi_fd_active[i] = 1;
break;
}
}
HYDU_ASSERT(i < HYD_pmcd_pmip.local.proxy_process_count, status);
HYD_STRING_STASH_INIT(stash);
HYD_STRING_STASH(stash,
MPL_strdup("cmd=fullinit-response;pmi-version=2;pmi-subversion=0;rank="),
status);
HYD_STRING_STASH(stash, HYDU_int_to_str(id), status);
HYD_STRING_STASH(stash, MPL_strdup(";size="), status);
HYD_STRING_STASH(stash, HYDU_int_to_str(HYD_pmcd_pmip.system_global.global_process_count),
status);
HYD_STRING_STASH(stash, MPL_strdup(";appnum=0"), status);
if (HYD_pmcd_pmip.local.spawner_kvsname) {
HYD_STRING_STASH(stash, MPL_strdup(";spawner-jobid="), status);
HYD_STRING_STASH(stash, MPL_strdup(HYD_pmcd_pmip.local.spawner_kvsname), status);
}
if (HYD_pmcd_pmip.user_global.debug) {
HYD_STRING_STASH(stash, MPL_strdup(";debugged=TRUE;pmiverbose=TRUE"), status);
} else {
HYD_STRING_STASH(stash, MPL_strdup(";debugged=FALSE;pmiverbose=FALSE"), status);
}
HYD_STRING_STASH(stash, MPL_strdup(";rc=0;"), status);
HYD_STRING_SPIT(stash, cmd, status);
send_cmd_downstream(fd, cmd);
MPL_free(cmd);
fn_exit:
HYD_pmcd_pmi_free_tokens(tokens, token_count);
HYDU_FUNC_EXIT();
return status;
fn_fail:
goto fn_exit;
}
