static int blcr_cold_start(opal_crs_blcr_snapshot_t *snapshot) {
int ret, exit_status = OPAL_SUCCESS;
char **tmp_argv = NULL;
char * component_name = NULL;
int prev_pid;
opal_output_verbose(10, mca_crs_blcr_component.super.output_handle,
"crs:blcr: cold_start(%s)", snapshot->super.reference_name);
/*
* Find the snapshot directory, read the metadata file
*/
if( OPAL_SUCCESS != (ret = opal_crs_base_extract_expected_component(snapshot->super.local_location,
&component_name, &prev_pid) ) ) {
opal_output(mca_crs_blcr_component.super.output_handle,
"crs:blcr: blcr_cold_start: Error: Failed to extract the metadata from the local snapshot (%s). Returned %d.",
snapshot->super.local_location, ret);
exit_status = ret;
goto cleanup;
}
snapshot->super.component_name = strdup(component_name);
/* Compare the component strings to make sure this is our snapshot before going further */
if ( 0 != strncmp(mca_crs_blcr_component.super.base_version.mca_component_name,
component_name, strlen(component_name)) ) {
exit_status = OPAL_ERROR;
opal_output(mca_crs_blcr_component.super.output_handle,
"crs:blcr: blcr_cold_start: Error: This snapshot (%s) is not intended for us (%s)\n",
component_name, mca_crs_blcr_component.super.base_version.mca_component_name);
goto cleanup;
}
/*
* Context Filename
*/
opal_crs_base_metadata_read_token(snapshot->super.local_location, CRS_METADATA_CONTEXT, &tmp_argv);
if( NULL == tmp_argv ) {
opal_output(mca_crs_blcr_component.super.output_handle,
"crs:blcr: blcr_cold_start: Error: Failed to read the %s token from the local checkpoint in %s",
CRS_METADATA_CONTEXT, snapshot->super.local_location);
exit_status = OPAL_ERROR;
goto cleanup;
}
asprintf(&snapshot->context_filename, "%s/%s", snapshot->super.local_location, tmp_argv[0]);
/*
* Reset the cold_start flag
*/
snapshot->super.cold_start = false;
cleanup:
if(NULL != tmp_argv) {
opal_argv_free(tmp_argv);
tmp_argv = NULL;
}
return exit_status;
}
int opal_crs_base_extract_expected_component(FILE *metadata, char ** component_name, int *prev_pid)
{
int exit_status = OPAL_SUCCESS;
char **pid_argv = NULL;
char **name_argv = NULL;
/* Dummy check */
if( NULL == metadata ) {
exit_status = OPAL_ERROR;
goto cleanup;
}
opal_crs_base_metadata_read_token(metadata, CRS_METADATA_PID, &pid_argv);
if( NULL != pid_argv && NULL != pid_argv[0] ) {
*prev_pid = atoi(pid_argv[0]);
} else {
opal_output(0, "Error: expected_component: PID information unavailable!");
exit_status = OPAL_ERROR;
goto cleanup;
}
opal_crs_base_metadata_read_token(metadata, CRS_METADATA_COMP, &name_argv);
if( NULL != name_argv && NULL != name_argv[0] ) {
*component_name = strdup(name_argv[0]);
} else {
opal_output(0, "Error: expected_component: Component Name information unavailable!");
exit_status = OPAL_ERROR;
goto cleanup;
}
cleanup:
if( NULL != pid_argv ) {
opal_argv_free(pid_argv);
pid_argv = NULL;
}
if( NULL != name_argv ) {
opal_argv_free(name_argv);
name_argv = NULL;
}
return exit_status;
}
static int self_cold_start(opal_crs_self_snapshot_t *snapshot) {
int ret, exit_status = OPAL_SUCCESS;
char **tmp_argv = NULL;
char * component_name = NULL;
int prev_pid;
opal_output_verbose(10, mca_crs_self_component.super.output_handle,
"crs:self: cold_start()");
/*
* Find the snapshot directory, read the metadata file
*/
if( NULL == snapshot->super.metadata ) {
if (NULL == (snapshot->super.metadata = fopen(snapshot->super.metadata_filename, "a")) ) {
opal_output(mca_crs_self_component.super.output_handle,
"crs:self: checkpoint(): Error: Unable to open the file (%s)",
snapshot->super.metadata_filename);
exit_status = OPAL_ERROR;
goto cleanup;
}
}
if( OPAL_SUCCESS != (ret = opal_crs_base_extract_expected_component(snapshot->super.metadata,
&component_name, &prev_pid) ) ) {
opal_output(mca_crs_self_component.super.output_handle,
"crs:self: self_cold_start: Error: Failed to extract the metadata from the local snapshot (%s). Returned %d.",
snapshot->super.metadata_filename, ret);
exit_status = ret;
goto cleanup;
}
snapshot->super.component_name = strdup(component_name);
/* Compare the strings to make sure this is our snapshot before going further */
if ( 0 != strncmp(mca_crs_self_component.super.base_version.mca_component_name,
component_name, strlen(component_name)) ) {
exit_status = OPAL_ERROR;
opal_output(mca_crs_self_component.super.output_handle,
"crs:self: self_cold_start: Error: This snapshot (%s) is not intended for us (%s)\n",
component_name, mca_crs_self_component.super.base_version.mca_component_name);
goto cleanup;
}
/*
* Restart command
* JJH: Command lines limited to 256 chars.
*/
opal_crs_base_metadata_read_token(snapshot->super.metadata, CRS_METADATA_CONTEXT, &tmp_argv);
if( NULL == tmp_argv ) {
opal_output(mca_crs_self_component.super.output_handle,
"crs:self: self_cold_start: Error: Failed to read the %s token from the local checkpoint in %s",
CRS_METADATA_CONTEXT, snapshot->super.snapshot_directory);
exit_status = OPAL_ERROR;
goto cleanup;
}
asprintf(&snapshot->cmd_line, "%s", tmp_argv[0]);
/*
* Reset the cold_start flag
*/
snapshot->super.cold_start = false;
cleanup:
if(NULL != tmp_argv) {
opal_argv_free(tmp_argv);
tmp_argv = NULL;
}
return exit_status;
}
int opal_crs_none_restart(opal_crs_base_snapshot_t *base_snapshot, bool spawn_child, pid_t *child_pid)
{
int exit_status = OPAL_SUCCESS;
char **tmp_argv = NULL;
char **cr_argv = NULL;
int status;
*child_pid = getpid();
if( NULL == base_snapshot->metadata ) {
if (NULL == (base_snapshot->metadata = fopen(base_snapshot->metadata_filename, "a")) ) {
opal_output(0,
"crs:none: checkpoint(): Error: Unable to open the file (%s)",
base_snapshot->metadata_filename);
return OPAL_ERROR;
}
}
opal_crs_base_metadata_read_token(base_snapshot->metadata, CRS_METADATA_CONTEXT, &tmp_argv);
if( NULL == tmp_argv ) {
opal_output(opal_crs_base_framework.framework_output,
"crs:none: none_restart: Error: Failed to read the %s token from the local checkpoint in %s",
CRS_METADATA_CONTEXT, base_snapshot->metadata_filename);
exit_status = OPAL_ERROR;
goto cleanup;
}
if( opal_argv_count(tmp_argv) <= 0 ) {
opal_output_verbose(10, opal_crs_base_framework.framework_output,
"crs:none: none_restart: No command line to exec, so just returning");
exit_status = OPAL_SUCCESS;
goto cleanup;
}
if ( NULL == (cr_argv = opal_argv_split(tmp_argv[0], ' ')) ) {
exit_status = OPAL_ERROR;
goto cleanup;
}
if( !spawn_child ) {
opal_output_verbose(10, opal_crs_base_framework.framework_output,
"crs:none: none_restart: exec :(%s, %s):",
cr_argv[0], tmp_argv[0]);
status = execvp(cr_argv[0], cr_argv);
if(status < 0) {
opal_output(opal_crs_base_framework.framework_output,
"crs:none: none_restart: Child failed to execute :(%d):", status);
}
opal_output(opal_crs_base_framework.framework_output,
"crs:none: none_restart: execvp returned %d", status);
exit_status = status;
goto cleanup;
} else {
opal_output(opal_crs_base_framework.framework_output,
"crs:none: none_restart: Spawn not implemented");
exit_status = OPAL_ERR_NOT_IMPLEMENTED;
goto cleanup;
}
cleanup:
if (cr_argv) {
opal_argv_free (cr_argv);
}
fclose(base_snapshot->metadata);
return exit_status;
}
static int post_env_vars(int prev_pid, opal_crs_base_snapshot_t *snapshot)
{
int ret, exit_status = OPAL_SUCCESS;
char *command = NULL;
char *proc_file = NULL;
char **loc_touch = NULL;
char **loc_mkdir = NULL;
int argc, i;
if( 0 > prev_pid ) {
opal_output(opal_restart_globals.output,
"Invalid PID (%d)\n",
prev_pid);
exit_status = OPAL_ERROR;
goto cleanup;
}
/*
* This is needed so we can pass the previous environment to the restarted
* application process.
*/
opal_asprintf(&proc_file, "%s/%s-%d", opal_tmp_directory(), OPAL_CR_BASE_ENV_NAME, prev_pid);
opal_asprintf(&command, "env | grep OMPI_ > %s", proc_file);
opal_output_verbose(5, opal_restart_globals.output,
"post_env_vars: Execute: <%s>", command);
ret = system(command);
if( 0 > ret) {
exit_status = ret;
goto cleanup;
}
/*
* Any directories that need to be created
*/
if( NULL == (snapshot->metadata = fopen(snapshot->metadata_filename, "r")) ) {
opal_show_help("help-opal-restart.txt", "invalid_metadata", true,
opal_restart_globals.snapshot_metadata,
snapshot->metadata_filename);
exit_status = OPAL_ERROR;
goto cleanup;
}
opal_crs_base_metadata_read_token(snapshot->metadata, CRS_METADATA_MKDIR, &loc_mkdir);
argc = opal_argv_count(loc_mkdir);
for( i = 0; i < argc; ++i ) {
if( NULL != command ) {
free(command);
command = NULL;
}
opal_asprintf(&command, "mkdir -p %s", loc_mkdir[i]);
opal_output_verbose(5, opal_restart_globals.output,
"post_env_vars: Execute: <%s>", command);
ret = system(command);
if( 0 > ret) {
exit_status = ret;
goto cleanup;
}
}
if( 0 < argc ) {
system("sync ; sync");
}
/*
* Any files that need to exist
*/
opal_crs_base_metadata_read_token(snapshot->metadata, CRS_METADATA_TOUCH, &loc_touch);
argc = opal_argv_count(loc_touch);
for( i = 0; i < argc; ++i ) {
if( NULL != command ) {
free(command);
command = NULL;
}
opal_asprintf(&command, "touch %s", loc_touch[i]);
opal_output_verbose(5, opal_restart_globals.output,
"post_env_vars: Execute: <%s>", command);
ret = system(command);
if( 0 > ret) {
exit_status = ret;
goto cleanup;
}
}
if( 0 < argc ) {
system("sync ; sync");
}
cleanup:
if( NULL != command) {
free(command);
command = NULL;
}
if( NULL != proc_file) {
free(proc_file);
proc_file = NULL;
}
if( NULL != loc_mkdir ) {
opal_argv_free(loc_mkdir);
loc_mkdir = NULL;
}
if( NULL != loc_touch ) {
opal_argv_free(loc_touch);
loc_touch = NULL;
}
if( NULL != snapshot->metadata ) {
fclose(snapshot->metadata);
snapshot->metadata = NULL;
}
return exit_status;
}
