1
1

Some checkpoint/restart cleanup.

* Remove the opal_only option. This was suffering from bit rot, and no one uses it. It can be added back fairly easily if wanted.
 * Cleanup metadata interactions at the local level.
 * Touch up some of the INC funcitonality (fix typos and a minor ordering issue)

This commit was SVN r18416.
Этот коммит содержится в:
Josh Hursey 2008-05-08 18:47:47 +00:00
родитель 8739edc580
Коммит da2f1c58e2
12 изменённых файлов: 577 добавлений и 715 удалений

Просмотреть файл

@ -35,6 +35,7 @@
#include "ompi/mca/pml/pml.h"
#include "ompi/mca/pml/base/base.h"
#include "orte/mca/rml/rml.h"
#include "orte/mca/grpcomm/grpcomm.h"
#include "ompi/proc/proc.h"
#include "bml_r2.h"
@ -93,18 +94,6 @@ int mca_bml_r2_ft_event(int state)
* preparation for being shut down.
*/
for(btl_idx = 0; btl_idx < mca_bml_r2.num_btl_modules; btl_idx++) {
/*
* Notify BTL
*/
if( NULL != (mca_bml_r2.btl_modules[btl_idx])->btl_ft_event) {
opal_output_verbose(10, ompi_cr_output,
"bml:r2: ft_event: Notify the %s BTL.\n",
(mca_bml_r2.btl_modules[btl_idx])->btl_component->btl_version.mca_component_name);
if(OMPI_SUCCESS != (ret = (mca_bml_r2.btl_modules[btl_idx])->btl_ft_event(loc_state) ) ) {
continue;
}
}
/*
* Notify Mpool
*/
@ -117,6 +106,18 @@ int mca_bml_r2_ft_event(int state)
continue;
}
}
/*
* Notify BTL
*/
if( NULL != (mca_bml_r2.btl_modules[btl_idx])->btl_ft_event) {
opal_output_verbose(10, ompi_cr_output,
"bml:r2: ft_event: Notify the %s BTL.\n",
(mca_bml_r2.btl_modules[btl_idx])->btl_component->btl_version.mca_component_name);
if(OMPI_SUCCESS != (ret = (mca_bml_r2.btl_modules[btl_idx])->btl_ft_event(loc_state) ) ) {
continue;
}
}
}
}
@ -140,7 +141,7 @@ int mca_bml_r2_ft_event(int state)
}
opal_output_verbose(10, ompi_cr_output,
"bml:r2: ft_event(reboot): Reselect BTLs\n");
"bml:r2: ft_event(Restart): Reselect BTLs\n");
/*
* Close the BTLs
@ -157,6 +158,20 @@ int mca_bml_r2_ft_event(int state)
}
else if(OPAL_CRS_RESTART == state ) {
/*
* Barrier to make all processes have been successfully restarted before
* we try to remove some restart only files.
*/
if (OMPI_SUCCESS != (ret = orte_grpcomm.barrier())) {
opal_output(0, "bml:r2: ft_event(Restart): Failed in orte_grpcomm.barrier (%d)", ret);
return ret;
}
opal_output_verbose(10, ompi_cr_output,
"bml:r2: ft_event(Restart): Cleanup restart files\n");
opal_crs_base_cleanup_flush();
/*
* Re-open the BTL framework to get the full list of components.
*/

Просмотреть файл

@ -542,13 +542,13 @@ int mca_pml_ob1_ft_event( int state )
* Add the new procs (BTLs redo modex recv's)
*/
if( OMPI_SUCCESS != (ret = mca_pml_ob1_add_procs(procs, num_procs) ) ) {
opal_output(0, "pml:ob1: fr_event(Restart): Failed in add_procs (%d)", ret);
opal_output(0, "pml:ob1: ft_event(Restart): Failed in add_procs (%d)", ret);
return ret;
}
/* Is this barrier necessary ? JJH */
if (OMPI_SUCCESS != (ret = orte_grpcomm.barrier())) {
opal_output(0, "pml:ob1: fr_event(Restart): Failed in orte_grpcomm.barrier (%d)", ret);
opal_output(0, "pml:ob1: ft_event(Restart): Failed in orte_grpcomm.barrier (%d)", ret);
return ret;
}

Просмотреть файл

@ -281,6 +281,9 @@ int ompi_proc_refresh(void) {
/* Does not change: proc->proc_name.vpid */
proc->proc_name.jobid = ORTE_PROC_MY_NAME->jobid;
/* Make sure to clear the local flag before we set it below */
proc->proc_flags = 0;
if (i == ORTE_PROC_MY_NAME->vpid) {
ompi_proc_local_proc = proc;
proc->proc_flags |= OMPI_PROC_FLAG_LOCAL;

Просмотреть файл

@ -32,6 +32,14 @@
extern "C" {
#endif
/* Some local strings to use genericly with the local metadata file */
#define CRS_METADATA_BASE ("# ")
#define CRS_METADATA_COMP ("# Component: ")
#define CRS_METADATA_PID ("# PID: ")
#define CRS_METADATA_CONTEXT ("# CONTEXT: ")
#define CRS_METADATA_MKDIR ("# MKDIR: ")
#define CRS_METADATA_TOUCH ("# TOUCH: ")
/**
* Initialize the CRS MCA framework
*
@ -106,24 +114,32 @@ extern "C" {
OPAL_DECLSPEC char * opal_crs_base_state_str(opal_crs_state_type_t state);
OPAL_DECLSPEC char * opal_crs_base_unique_snapshot_name(pid_t pid);
OPAL_DECLSPEC char * opal_crs_base_extract_expected_component(char *snapshot_loc, int *prev_pid);
OPAL_DECLSPEC int opal_crs_base_extract_expected_component(char *snapshot_loc, char ** component_name, int *prev_pid);
OPAL_DECLSPEC int opal_crs_base_init_snapshot_directory(opal_crs_base_snapshot_t *snapshot);
OPAL_DECLSPEC char * opal_crs_base_get_snapshot_directory(char *uniq_snapshot_name);
/* Opens the metadata file and places all the base information in the file.
* Options:
* 'w' = Open for writing
* 'a' = Open for writing and appending information
/*
* Read a token to the metadata file
* NULL can be passed for snapshot_loc if nit_snapshot_directory has been called.
*/
OPAL_DECLSPEC FILE *opal_crs_base_open_metadata(opal_crs_base_snapshot_t *snapshot, char mode );
OPAL_DECLSPEC int opal_crs_base_metadata_read_token(char *snapshot_loc, char * token, char ***value);
/* Open the metadata file, read off the base information and
* return the component and previous pid to the caller.
* Note: component is allocated inside this function, it is the
* callers responsibility to free this memory.
/*
* Write a token to the metadata file
* NULL can be passed for snapshot_loc if nit_snapshot_directory has been called.
*/
OPAL_DECLSPEC FILE * opal_crs_base_open_read_metadata(char *location, char **component, int *prev_pid);
OPAL_DECLSPEC int opal_crs_base_metadata_write_token(char *snapshot_loc, char * token, char *value);
/*
* Register a file for cleanup.
* Useful in C/R when files only need to temporarily exist for restart
*/
OPAL_DECLSPEC int opal_crs_base_cleanup_append(char* filename, bool is_dir);
/*
* Flush the cleanup of all registered files.
*/
OPAL_DECLSPEC int opal_crs_base_cleanup_flush(void);
#if defined(c_plusplus) || defined(__cplusplus)
}

Просмотреть файл

@ -24,16 +24,36 @@
#if HAVE_UNISTD_H
#include <unistd.h>
#endif
#ifdef HAVE_FCNTL_H
#include <fcntl.h>
#endif /* HAVE_FCNTL_H */
#ifdef HAVE_SYS_STAT_H
#include <sys/stat.h>
#endif
#include "opal/mca/mca.h"
#include "opal/mca/base/base.h"
#include "opal/include/opal/constants.h"
#include "opal/util/os_dirpath.h"
#include "opal/util/output.h"
#include "opal/util/argv.h"
#include "opal/mca/crs/crs.h"
#include "opal/mca/crs/base/base.h"
/******************
* Local Functions
******************/
static int metadata_extract_next_token(FILE *file, char **token, char **value);
static int opal_crs_base_metadata_open(FILE ** meta_data, char * location, char * mode);
static char *last_metadata_file = NULL;
static char **cleanup_file_argv = NULL;
static char **cleanup_dir_argv = NULL;
/******************
* Object stuff
******************/
static void opal_crs_base_construct(opal_crs_base_snapshot_t *snapshot)
{
snapshot->component_name = NULL;
@ -198,78 +218,112 @@ char * opal_crs_base_unique_snapshot_name(pid_t pid)
return loc_str;
}
FILE * opal_crs_base_open_read_metadata(char * location, char **component, int *prev_pid)
{
char * dir_name = NULL;
char * content = NULL;
char * tmp_str = NULL;
int len = 0;
int opal_crs_base_metadata_read_token(char *snapshot_loc, char * token, char ***value) {
int ret, exit_status = OPAL_SUCCESS;
FILE * meta_data = NULL;
char * loc_token = NULL;
char * loc_value = NULL;
int argc = 0;
*component = NULL;
*prev_pid = -1;
/* Dummy check */
if( NULL == token ) {
goto cleanup;
}
/*
* Find the snapshot directory, read the metadata file
* Open the metadata file
*/
asprintf(&dir_name, "%s/%s", location, opal_crs_base_metadata_filename);
if (NULL == (meta_data = fopen(dir_name, "r")) ) {
if( OPAL_SUCCESS != (ret = opal_crs_base_metadata_open(&meta_data, snapshot_loc, "r")) ) {
opal_output(opal_crs_base_output,
"opal:crs:base: opal_crs_base_metadata_read_token: Error: Unable to open the metadata file\n");
exit_status = ret;
goto cleanup;
}
/*
* Component Name
*/
len = 32; /* Max size for a CRS component name */
content = (char *) malloc(sizeof(char) * len);
if (NULL == fgets(content, len, meta_data) ) {
free(content);
content = NULL;
goto cleanup;
}
/* Strip of newline */
len = strlen(content);
content[len - 1] = '\0';
*component = strdup(content);
/*
* Get the PID
* Extract each token and make the records
*/
len = 128;
tmp_str = (char *) malloc(sizeof(char) * len);
if (NULL == fgets(tmp_str, len, meta_data) ) {
goto cleanup;
}
/* Strip of newline */
len = strlen(tmp_str);
if(tmp_str[len - 1] == '\n')
tmp_str[len - 1] = '\0';
*prev_pid = atoi(tmp_str);
do {
/* Get next token */
if( OPAL_SUCCESS != metadata_extract_next_token(meta_data, &loc_token, &loc_value) ) {
break;
}
/* Check token to see if it matches */
if(0 == strncmp(token, loc_token, strlen(loc_token)) ) {
opal_argv_append(&argc, value, loc_value);
}
} while(0 == feof(meta_data) );
cleanup:
return meta_data;
if(NULL != meta_data) {
fclose(meta_data);
meta_data = NULL;
}
return exit_status;
}
char * opal_crs_base_extract_expected_component(char *snapshot_loc, int *prev_pid)
{
int opal_crs_base_metadata_write_token(char *snapshot_loc, char * token, char *value) {
int ret, exit_status = OPAL_SUCCESS;
FILE * meta_data = NULL;
char * component_name = NULL;
*prev_pid = -1;
if( NULL == (meta_data = opal_crs_base_open_read_metadata(snapshot_loc, &component_name, prev_pid)) ) {
opal_output(opal_crs_base_output,
"opal:crs:base: extract_expected_component: Error: Unable to open the file (%s)\n",
snapshot_loc);
/* Dummy check */
if( NULL == token || NULL == value) {
goto cleanup;
}
cleanup:
if(NULL != meta_data)
fclose(meta_data);
/*
* Open the metadata file
*/
if( OPAL_SUCCESS != (ret = opal_crs_base_metadata_open(&meta_data, snapshot_loc, "a")) ) {
opal_output(opal_crs_base_output,
"opal:crs:base: opal_crs_base_metadata_write_token: Error: Unable to open the metadata file\n");
exit_status = ret;
goto cleanup;
}
return component_name;
fprintf(meta_data, "%s%s\n", token, value);
cleanup:
if(NULL != meta_data) {
fclose(meta_data);
meta_data = NULL;
}
return exit_status;
}
int opal_crs_base_extract_expected_component(char *snapshot_loc, char ** component_name, int *prev_pid)
{
char **pid_argv = NULL;
char **name_argv = NULL;
opal_crs_base_metadata_read_token(snapshot_loc, 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!");
}
opal_crs_base_metadata_read_token(snapshot_loc, 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!");
}
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 OPAL_SUCCESS;
}
char * opal_crs_base_get_snapshot_directory(char *uniq_snapshot_name)
@ -283,77 +337,113 @@ char * opal_crs_base_get_snapshot_directory(char *uniq_snapshot_name)
int opal_crs_base_init_snapshot_directory(opal_crs_base_snapshot_t *snapshot)
{
mode_t my_mode = S_IRWXU;
int ret, exit_status = OPAL_SUCCESS;
FILE * meta_data = NULL;
mode_t my_mode = S_IRWXU;
char * pid_str = NULL;
/*
* Make the snapshot directory from the uniq_snapshot_name
*/
if(OPAL_SUCCESS != (ret = opal_os_dirpath_create(snapshot->local_location, my_mode)) ) {
opal_output(opal_crs_base_output,
"opal:crs:base: init_snapshot_directory: Error: Unable to create directory (%s)\n",
snapshot->local_location);
exit_status = ret;
goto cleanup;
}
/*
* Initialize the metadata file at the top of that directory.
* Add 'BASE' and 'PID'
*/
if (NULL == (meta_data = opal_crs_base_open_metadata(snapshot, 'w') ) ) {
if( NULL != last_metadata_file ) {
free(last_metadata_file);
last_metadata_file = NULL;
}
last_metadata_file = strdup(snapshot->local_location);
if( OPAL_SUCCESS != (ret = opal_crs_base_metadata_write_token(NULL, CRS_METADATA_BASE, "") ) ) {
opal_output(opal_crs_base_output,
"opal:crs:base: init_snapshot_directory: Error: Unable to open the file (%s/%s)\n",
"opal:crs:base: init_snapshot_directory: Error: Unable to write BASE to the file (%s/%s)\n",
snapshot->local_location, opal_crs_base_metadata_filename);
exit_status = OPAL_ERROR;
exit_status = ret;
goto cleanup;
}
asprintf(&pid_str, "%d", getpid());
if( OPAL_SUCCESS != (ret = opal_crs_base_metadata_write_token(NULL, CRS_METADATA_PID, pid_str) ) ) {
opal_output(opal_crs_base_output,
"opal:crs:base: init_snapshot_directory: Error: Unable to write PID (%s) to the file (%s/%s)\n",
pid_str, snapshot->local_location, opal_crs_base_metadata_filename);
exit_status = ret;
goto cleanup;
}
cleanup:
if(NULL != meta_data)
fclose(meta_data);
if( NULL != pid_str) {
free(pid_str);
pid_str = NULL;
}
return OPAL_SUCCESS;
}
FILE *opal_crs_base_open_metadata(opal_crs_base_snapshot_t *snapshot, char mode )
int opal_crs_base_cleanup_append(char* filename, bool is_dir)
{
char *meta_data_fname = NULL;
FILE * meta_data = NULL;
if( NULL == filename ) {
return OPAL_SUCCESS;
}
if( is_dir ) {
opal_output_verbose(15, opal_crs_base_output,
"opal:crs: cleanup_append: Append Dir <%s>\n",
filename);
opal_argv_append_nosize(&cleanup_dir_argv, filename);
} else {
opal_output_verbose(15, opal_crs_base_output,
"opal:crs: cleanup_append: Append File <%s>\n",
filename);
opal_argv_append_nosize(&cleanup_file_argv, filename);
}
return OPAL_SUCCESS;
}
int opal_crs_base_cleanup_flush(void)
{
int argc, i;
/*
* Construct path
* Cleanup files first
*/
asprintf(&meta_data_fname, "%s/%s", snapshot->local_location, opal_crs_base_metadata_filename);
if( NULL != cleanup_file_argv ) {
argc = opal_argv_count(cleanup_file_argv);
for( i = 0; i < argc; ++i) {
opal_output_verbose(15, opal_crs_base_output,
"opal:crs: cleanup_flush: Remove File <%s>\n", cleanup_dir_argv[i]);
unlink(cleanup_file_argv[i]);
}
opal_argv_free(cleanup_file_argv);
cleanup_file_argv = NULL;
}
/*
* Open the metadata file
* Try to cleanup directories next
*/
if( mode == 'w' ) {
meta_data = fopen(meta_data_fname, "w");
}
else if( mode == 'a' ) {
meta_data = fopen(meta_data_fname, "a");
if( NULL != cleanup_dir_argv ) {
argc = opal_argv_count(cleanup_dir_argv);
for( i = 0; i < argc; ++i) {
opal_output_verbose(15, opal_crs_base_output,
"opal:crs: cleanup_flush: Remove Dir <%s>\n", cleanup_dir_argv[i]);
opal_os_dirpath_destroy(cleanup_dir_argv[i], true, NULL);
}
opal_argv_free(cleanup_dir_argv);
cleanup_dir_argv = NULL;
}
if (NULL == meta_data ) {
opal_output(opal_crs_base_output,
"opal:crs:base: open_metadata (%c): Error: Unable to open the file (%s)\n",
mode, meta_data_fname);
goto cleanup;
}
if( mode == 'w' ) {
/*
* The first line is the component name,
* everything else here is defined by the component
*/
fprintf(meta_data, "%s\n", snapshot->component_name);
fprintf(meta_data, "%d\n", getpid());
}
cleanup:
if(NULL != meta_data_fname)
free(meta_data_fname);
return meta_data;
return OPAL_SUCCESS;
}
char * opal_crs_base_state_str(opal_crs_state_type_t state)
@ -386,3 +476,154 @@ char * opal_crs_base_state_str(opal_crs_state_type_t state)
return str;
}
/******************
* Local Functions
******************/
static int opal_crs_base_metadata_open(FILE **meta_data, char * location, char * mode)
{
int exit_status = OPAL_SUCCESS;
char * dir_name = NULL;
if( NULL == location ) {
if( NULL == last_metadata_file ) {
opal_output(0, "Error: No metadata filename specified!");
exit_status = OPAL_ERROR;
goto cleanup;
} else {
location = last_metadata_file;
}
}
/*
* Find the snapshot directory, read the metadata file
*/
asprintf(&dir_name, "%s/%s", location, opal_crs_base_metadata_filename);
if (NULL == (*meta_data = fopen(dir_name, mode)) ) {
exit_status = OPAL_ERROR;
goto cleanup;
}
cleanup:
if( NULL != dir_name ) {
free(dir_name);
dir_name = NULL;
}
return exit_status;
}
static int metadata_extract_next_token(FILE *file, char **token, char **value)
{
int exit_status = OPAL_SUCCESS;
int max_len = 256;
char * line = NULL;
int line_len = 0;
int c = 0, s = 0, v = 0;
char *local_token = NULL;
char *local_value = NULL;
bool end_of_line = false;
line = (char *) malloc(sizeof(char) * max_len);
try_again:
/*
* If we are at the end of the file, then just return
*/
if(0 != feof(file) ) {
exit_status = OPAL_ERROR;
goto cleanup;
}
/*
* Other wise grab the next token/value pair
*/
if (NULL == fgets(line, max_len, file) ) {
exit_status = OPAL_ERROR;
goto cleanup;
}
line_len = strlen(line);
/* Strip off the new line if it it there */
if('\n' == line[line_len-1]) {
line[line_len-1] = '\0';
line_len--;
end_of_line = true;
}
else {
end_of_line = false;
}
/* Ignore lines with just '#' too */
if(2 >= line_len)
goto try_again;
/*
* Extract the token from the set
*/
for(c = 0;
line[c] != ':' &&
c < line_len;
++c) {
;
}
c += 2; /* For the ' ' and the '\0' */
local_token = (char *)malloc(sizeof(char) * (c + 1));
for(s = 0; s < c; ++s) {
local_token[s] = line[s];
}
local_token[s] = '\0';
*token = strdup(local_token);
if( NULL != local_token) {
free(local_token);
local_token = NULL;
}
/*
* Extract the value from the set
*/
local_value = (char *)malloc(sizeof(char) * (line_len - c + 1));
for(v = 0, s = c;
s < line_len;
++s, ++v) {
local_value[v] = line[s];
}
while(!end_of_line) {
if (NULL == fgets(line, max_len, file) ) {
exit_status = OPAL_ERROR;
goto cleanup;
}
line_len = strlen(line);
/* Strip off the new line if it it there */
if('\n' == line[line_len-1]) {
line[line_len-1] = '\0';
line_len--;
end_of_line = true;
}
else {
end_of_line = false;
}
local_value = (char *)realloc(local_value, sizeof(char) * line_len);
for(s = 0;
s < line_len;
++s, ++v) {
local_value[v] = line[s];
}
}
local_value[v] = '\0';
*value = strdup(local_value);
cleanup:
if( NULL != local_token)
free(local_token);
if( NULL != local_value)
free(local_value);
if( NULL != line)
free(line);
return exit_status;
}

Просмотреть файл

@ -252,7 +252,7 @@ int opal_crs_blcr_module_finalize(void)
int opal_crs_blcr_checkpoint(pid_t pid, opal_crs_base_snapshot_t *base_snapshot, opal_crs_state_type_t *state)
{
int ret;
int ret, exit_status = OPAL_SUCCESS;
opal_crs_blcr_snapshot_t *snapshot = OBJ_NEW(opal_crs_blcr_snapshot_t);
char * tmp_str = NULL;
@ -272,15 +272,15 @@ int opal_crs_blcr_checkpoint(pid_t pid, opal_crs_base_snapshot_t *base_snapshot,
snapshot->super.remote_location = strdup(base_snapshot->remote_location);
/*
* Create the snapshot directory
* Update the snapshot metadata
*/
snapshot->super.component_name = strdup(mca_crs_blcr_component.super.base_version.mca_component_name);
if( OPAL_SUCCESS != (ret = opal_crs_base_init_snapshot_directory(&snapshot->super) )) {
*state = OPAL_CRS_ERROR;
if( OPAL_SUCCESS != (ret = opal_crs_base_metadata_write_token(NULL, CRS_METADATA_COMP, snapshot->super.component_name) ) ) {
opal_output(mca_crs_blcr_component.super.output_handle,
"crs:blcr: checkpoint(): Error: Unable to initialize the directory for (%s).",
"crs:blcr: checkpoint(): Error: Unable to write component name to the directory for (%s).",
snapshot->super.reference_name);
return ret;
exit_status = ret;
goto cleanup;
}
/*
@ -328,7 +328,8 @@ int opal_crs_blcr_checkpoint(pid_t pid, opal_crs_base_snapshot_t *base_snapshot,
opal_output(mca_crs_blcr_component.super.output_handle,
"crs:blcr: checkpoint(): Error: Unable to checkpoint pid (%d)",
pid);
return ret;
exit_status = ret;
goto cleanup;
}
*state = blcr_current_state;
@ -347,7 +348,9 @@ int opal_crs_blcr_checkpoint(pid_t pid, opal_crs_base_snapshot_t *base_snapshot,
snapshot->context_filename, tmp_str, ret);
perror("crs:blcr: checkpoint");
free(tmp_str);
return ret;
exit_status = ret;
goto cleanup;
}
/*
@ -358,7 +361,8 @@ int opal_crs_blcr_checkpoint(pid_t pid, opal_crs_base_snapshot_t *base_snapshot,
opal_output(mca_crs_blcr_component.super.output_handle,
"crs:blcr: checkpoint(): Error: Unable to update metadata for snapshot (%s).",
snapshot->super.reference_name);
return ret;
exit_status = ret;
goto cleanup;
}
}
@ -367,10 +371,13 @@ int opal_crs_blcr_checkpoint(pid_t pid, opal_crs_base_snapshot_t *base_snapshot,
*/
base_snapshot = &(snapshot->super);
if(NULL != tmp_str)
cleanup:
if(NULL != tmp_str) {
free(tmp_str);
tmp_str = NULL;
}
return OPAL_SUCCESS;
return exit_status;
}
int opal_crs_blcr_restart(opal_crs_base_snapshot_t *base_snapshot, bool spawn_child, pid_t *child_pid)
@ -760,8 +767,6 @@ static int blcr_get_checkpoint_filename(char **fname, pid_t pid)
}
static int blcr_update_snapshot_metadata(opal_crs_blcr_snapshot_t *snapshot) {
char * dir_name = NULL;
FILE * meta_data = NULL;
int exit_status = OPAL_SUCCESS;
opal_output_verbose(10, mca_crs_blcr_component.super.output_handle,
@ -779,34 +784,19 @@ static int blcr_update_snapshot_metadata(opal_crs_blcr_snapshot_t *snapshot) {
}
/*
* Append to the metadata file:
* the relative path of the context filename
* Append to the metadata file the context filename
*/
if( NULL == (meta_data = opal_crs_base_open_metadata(&snapshot->super, 'w') ) ) {
exit_status = OPAL_ERROR;
goto cleanup;
}
/* Context Filename -- Relative path */
fprintf(meta_data, "%s\n", snapshot->context_filename);
opal_crs_base_metadata_write_token(snapshot->super.local_location, CRS_METADATA_CONTEXT, snapshot->context_filename);
cleanup:
if(NULL != meta_data)
fclose(meta_data);
if(NULL != dir_name)
free(dir_name);
return exit_status;
}
static int blcr_cold_start(opal_crs_blcr_snapshot_t *snapshot) {
char * content = NULL;
int ret, exit_status = OPAL_SUCCESS;
char **tmp_argv = NULL;
char * component_name = NULL;
int prev_pid;
int len = 0;
FILE * meta_data = NULL;
int exit_status = OPAL_SUCCESS;
opal_output_verbose(10, mca_crs_blcr_component.super.output_handle,
"crs:blcr: cold_start(%s)", snapshot->super.reference_name);
@ -814,9 +804,9 @@ static int blcr_cold_start(opal_crs_blcr_snapshot_t *snapshot) {
/*
* Find the snapshot directory, read the metadata file
*/
if( NULL == (meta_data = opal_crs_base_open_read_metadata(snapshot->super.local_location,
&component_name, &prev_pid) ) ) {
exit_status = OPAL_ERROR;
if( OPAL_SUCCESS != (ret = opal_crs_base_extract_expected_component(snapshot->super.local_location,
&component_name, &prev_pid) ) ) {
exit_status = ret;
goto cleanup;
}
@ -835,19 +825,8 @@ static int blcr_cold_start(opal_crs_blcr_snapshot_t *snapshot) {
/*
* Context Filename
*/
len = 256; /* Max size for a BLCR filename */
content = (char *) malloc(sizeof(char) * len);
if (NULL == fgets(content, len, meta_data) ) {
free(content);
content = NULL;
goto cleanup;
}
/* Strip of newline */
len = strlen(content);
content[len - 1] = '\0';
/* save the filename in the structure */
asprintf(&snapshot->context_filename, "%s/%s", snapshot->super.local_location, content);
opal_crs_base_metadata_read_token(snapshot->super.local_location, CRS_METADATA_CONTEXT, &tmp_argv);
asprintf(&snapshot->context_filename, "%s/%s", snapshot->super.local_location, tmp_argv[0]);
/*
* Reset the cold_start flag
@ -855,10 +834,10 @@ static int blcr_cold_start(opal_crs_blcr_snapshot_t *snapshot) {
snapshot->super.cold_start = false;
cleanup:
if(NULL != meta_data)
fclose(meta_data);
if(NULL != content)
free(content);
if(NULL != tmp_argv) {
opal_argv_free(tmp_argv);
tmp_argv = NULL;
}
return exit_status;
}

Просмотреть файл

@ -291,13 +291,12 @@ int opal_crs_self_checkpoint(pid_t pid, opal_crs_base_snapshot_t *base_snapshot,
}
/*
* Create the snapshot directory
* Update the snapshot metadata
*/
snapshot->super.component_name = strdup(mca_crs_self_component.super.base_version.mca_component_name);
if( OPAL_SUCCESS != (ret = opal_crs_base_init_snapshot_directory(&snapshot->super) )) {
*state = OPAL_CRS_ERROR;
if( OPAL_SUCCESS != (ret = opal_crs_base_metadata_write_token(NULL, CRS_METADATA_COMP, snapshot->super.component_name) ) ) {
opal_output(mca_crs_self_component.super.output_handle,
"crs:self: checkpoint(): Error: Unable to initialize the directory for (%s).",
"crs:self: checkpoint(): Error: Unable to write component name to the directory for (%s).",
snapshot->super.reference_name);
exit_status = ret;
goto cleanup;
@ -592,12 +591,10 @@ static int opal_crs_self_restart_cmd(opal_crs_self_snapshot_t *snapshot, char **
}
static int self_cold_start(opal_crs_self_snapshot_t *snapshot) {
char * content = NULL;
int ret, exit_status = OPAL_SUCCESS;
char **tmp_argv = NULL;
char * component_name = NULL;
int prev_pid;
int len = 0;
FILE * meta_data = NULL;
int exit_status = OPAL_SUCCESS;
opal_output_verbose(10, mca_crs_self_component.super.output_handle,
"crs:self: cold_start(%s)", snapshot->super.reference_name);
@ -605,9 +602,9 @@ static int self_cold_start(opal_crs_self_snapshot_t *snapshot) {
/*
* Find the snapshot directory, read the metadata file
*/
if( NULL == (meta_data = opal_crs_base_open_read_metadata(snapshot->super.local_location,
&component_name, &prev_pid) ) ) {
exit_status = OPAL_ERROR;
if( OPAL_SUCCESS != (ret = opal_crs_base_extract_expected_component(snapshot->super.local_location,
&component_name, &prev_pid) ) ) {
exit_status = ret;
goto cleanup;
}
@ -627,19 +624,8 @@ static int self_cold_start(opal_crs_self_snapshot_t *snapshot) {
* Restart command
* JJH: Command lines limited to 256 chars.
*/
len = 256; /* Max size for a SELF filename */
content = (char *) malloc(sizeof(char) * len);
if (NULL == fgets(content, len, meta_data) ) {
free(content);
content = NULL;
goto cleanup;
}
/* Strip of newline */
len = strlen(content);
content[len - 1] = '\0';
/* save the command line in the structure */
asprintf(&snapshot->cmd_line, "%s", content);
opal_crs_base_metadata_read_token(snapshot->super.local_location, CRS_METADATA_CONTEXT, &tmp_argv);
asprintf(&snapshot->cmd_line, "%s", tmp_argv[0]);
/*
* Reset the cold_start flag
@ -647,51 +633,35 @@ static int self_cold_start(opal_crs_self_snapshot_t *snapshot) {
snapshot->super.cold_start = false;
cleanup:
if(NULL != meta_data)
fclose(meta_data);
if(NULL != content)
free(content);
if(NULL != tmp_argv) {
opal_argv_free(tmp_argv);
tmp_argv = NULL;
}
return exit_status;
}
static int self_update_snapshot_metadata(opal_crs_self_snapshot_t *snapshot) {
char * dir_name = NULL;
FILE *meta_data = NULL;
int exit_status = OPAL_SUCCESS;
if(NULL == snapshot->cmd_line) {
opal_show_help("help-opal-crs-self.txt", "self:no-restart-cmd",
true);
exit_status = OPAL_ERROR;
goto cleanup;
}
opal_output_verbose(10, mca_crs_self_component.super.output_handle,
"crs:self: update_snapshot_metadata(%s)",
snapshot->super.reference_name);
/*
* Append to the metadata file:
* the relative path of the context filename
* Append to the metadata file the command line to restart with
* - How user wants us to restart
*/
if( NULL == (meta_data = opal_crs_base_open_metadata(&snapshot->super, 'w') ) ) {
exit_status = OPAL_ERROR;
goto cleanup;
}
/* How user wants us to restart */
if(NULL != snapshot->cmd_line) {
fprintf(meta_data, "%s\n", snapshot->cmd_line);
}
else {
opal_show_help("help-opal-crs-self.txt", "self:no-restart-cmd",
true);
exit_status = OPAL_ERROR;
}
opal_crs_base_metadata_write_token(snapshot->super.local_location, CRS_METADATA_CONTEXT, snapshot->cmd_line);
cleanup:
if(NULL != meta_data) {
fclose(meta_data);
}
if(NULL != dir_name) {
free(dir_name);
dir_name = NULL;
}
return exit_status;
}

Просмотреть файл

@ -73,7 +73,6 @@
/******************
* Global Var Decls
******************/
bool opal_cr_allow_opal_only = false;
bool opal_cr_stall_check = false;
bool opal_cr_currently_stalled = false;
int opal_cr_output;
@ -81,18 +80,13 @@ int opal_cr_output;
/******************
* Local Functions & Var Decls
******************/
static int cr_notify_response(opal_cr_ckpt_cmd_state_t resp);
static int extract_env_vars(int prev_pid);
static int cr_entry_point_notify_reopen_files(int *prog_read_fd, int *prog_write_fd);
static void opal_cr_entry_point_signal_handler (int signo);
static void opal_cr_sigpipe_debug_signal_handler (int signo);
static opal_cr_coord_callback_fn_t cur_coord_callback = NULL;
static opal_cr_notify_callback_fn_t cur_notify_callback = NULL;
static char *prog_named_pipe_r = NULL;
static char *prog_named_pipe_w = NULL;
/******************
* Interface Functions & Vars
******************/
@ -100,10 +94,12 @@ char * opal_cr_pipe_dir = NULL;
int opal_cr_entry_point_signal = 0;
bool opal_cr_is_enabled = true;
bool opal_cr_is_tool = false;
/* Current checkpoint state */
int opal_cr_checkpointing = OPAL_CR_STATUS_NONE;
int opal_cr_checkpointing_state = OPAL_CR_STATUS_NONE;
/* Current checkpoint request channel state */
int opal_cr_checkpoint_request = OPAL_CR_STATUS_NONE;
int opal_cr_checkpoint_request = OPAL_CR_STATUS_NONE;
static bool opal_cr_debug_sigpipe = false;
@ -241,22 +237,6 @@ int opal_cr_init(void )
opal_cr_thread_sleep_check, opal_cr_thread_sleep_wait);
#endif
/*
* Whether or not to allow OPAL only checkpointing.
* By default we rely on ORTE to provide this functionality for us, but
* if the application is OPAL only then we need to fallback to the signal
* method which is activated by setting this MCA parameter to 'true'.
*/
mca_base_param_reg_int_name("opal_cr", "allow_opal_only",
"Enable OPAL Only checkpointing [Default: Disabled]",
true, false,
0, &val);
opal_cr_allow_opal_only = OPAL_INT_TO_BOOL(val);
opal_output_verbose(10, opal_cr_output,
"opal_cr: init: OPAL CR Allow OPAL Only: %d",
val);
mca_base_param_reg_int_name("opal_cr", "is_tool",
"Is this a tool program, meaning does it require a fully operational OPAL or just enough to exec.",
false, false,
@ -315,13 +295,6 @@ int opal_cr_init(void )
opal_cr_stall_check = false;
opal_cr_currently_stalled = false;
/*
* Register the entry point
*/
if( OPAL_SUCCESS != (ret = opal_cr_entry_point_init()) ) {
exit_status = OPAL_ERROR;
goto cleanup;
}
} /* End opal_cr_is_tool = true */
/*
@ -381,7 +354,7 @@ int opal_cr_init(void )
int opal_cr_finalize(void)
{
int ret, exit_status = OPAL_SUCCESS;
int exit_status = OPAL_SUCCESS;
if( --opal_cr_initalized != 0 ) {
if( opal_cr_initalized < 0 ) {
@ -407,13 +380,9 @@ int opal_cr_finalize(void)
}
#endif /* OPAL_ENABLE_FT_THREAD == 1 */
if( OPAL_SUCCESS != (ret = opal_cr_entry_point_finalize()) ) {
exit_status = ret;
}
/* Nothing to do for just process notifications */
opal_cr_checkpointing = OPAL_CR_STATUS_TERM;
opal_cr_checkpoint_request = OPAL_CR_STATUS_TERM;
opal_cr_checkpointing_state = OPAL_CR_STATUS_TERM;
opal_cr_checkpoint_request = OPAL_CR_STATUS_TERM;
}
#if OPAL_ENABLE_FT == 1
@ -452,7 +421,7 @@ void opal_cr_test_if_checkpoint_ready(void)
* - If a request is pending then cancel it
* - o.w., skip it.
*/
if(OPAL_CR_STATUS_RUNNING == opal_cr_checkpointing ) {
if(OPAL_CR_STATUS_RUNNING == opal_cr_checkpointing_state ) {
if( OPAL_SUCCESS != (ret = cur_notify_callback(OPAL_CHECKPOINT_CMD_IN_PROGRESS) ) ) {
opal_output(opal_cr_output,
"Error: opal_cr: test_if_checkpoint_ready: Respond [In Progress] Failed. (%d)",
@ -478,8 +447,8 @@ void opal_cr_test_if_checkpoint_ready(void)
/*
* Start the checkpoint
*/
opal_cr_checkpointing = OPAL_CR_STATUS_RUNNING;
opal_cr_checkpoint_request = OPAL_CR_STATUS_NONE;
opal_cr_checkpointing_state = OPAL_CR_STATUS_RUNNING;
opal_cr_checkpoint_request = OPAL_CR_STATUS_NONE;
STAGE_1:
if( OPAL_SUCCESS != (ret = cur_notify_callback(OPAL_CHECKPOINT_CMD_START) ) ) {
@ -525,8 +494,12 @@ int opal_cr_inc_core(pid_t pid, opal_crs_base_snapshot_t *snapshot, bool term, i
}
if(*state == OPAL_CRS_CONTINUE) {
if(term)
if(term) {
*state = OPAL_CRS_TERM;
opal_cr_checkpointing_state = OPAL_CR_STATUS_TERM;
} else {
opal_cr_checkpointing_state = OPAL_CR_STATUS_CONTINUE;
}
}
else {
term = false;
@ -537,6 +510,7 @@ int opal_cr_inc_core(pid_t pid, opal_crs_base_snapshot_t *snapshot, bool term, i
*/
if(*state == OPAL_CRS_RESTART) {
extract_env_vars(prev_pid);
opal_cr_checkpointing_state = OPAL_CR_STATUS_RESTART_PRE;
}
/*
@ -594,11 +568,10 @@ int opal_cr_coord(int state)
/*
* Here we are returning to either:
* - opal_notify()
* If we have an OPAL only opplication.
* - [orte | ompi]_notify()
* If we have an ORTE or OPAL application.
*/
opal_cr_checkpointing_state = OPAL_CR_STATUS_RESTART_POST;
return OPAL_SUCCESS;
}
@ -720,83 +693,6 @@ static int extract_env_vars(int prev_pid)
/*****************************************
* OPAL CR Entry Point Functionality
*****************************************/
int opal_cr_entry_point_init(void)
{
int exit_status = OPAL_SUCCESS;
char *tmp_pid = NULL;
opal_cr_notify_callback_fn_t prev_notify_func;
if( !opal_cr_allow_opal_only ) {
return OPAL_SUCCESS;
}
opal_cr_reg_notify_callback(cr_notify_response, &prev_notify_func);
/* String representation of the PID */
asprintf(&tmp_pid, "%d", getpid());
asprintf(&prog_named_pipe_r, "%s/%s.%s", opal_cr_pipe_dir, OPAL_CR_NAMED_PROG_R, tmp_pid);
asprintf(&prog_named_pipe_w, "%s/%s.%s", opal_cr_pipe_dir, OPAL_CR_NAMED_PROG_W, tmp_pid);
opal_output_verbose(15, opal_cr_output,
"opal_cr: init: Named Pipes (%s) (%s)",
prog_named_pipe_r, prog_named_pipe_w);
/*
* Setup a signal handler to catch and start the proper thread
* to handle the checkpoint
*/
if( SIG_ERR == signal(opal_cr_entry_point_signal, opal_cr_entry_point_signal_handler) ) {
exit_status = OPAL_ERROR;
goto cleanup;
}
cleanup:
if( NULL != tmp_pid) {
free(tmp_pid);
tmp_pid = NULL;
}
return exit_status;
}
int opal_cr_entry_point_finalize(void)
{
if( !opal_cr_allow_opal_only ) {
return OPAL_SUCCESS;
}
if( NULL != prog_named_pipe_r) {
free(prog_named_pipe_r);
prog_named_pipe_r = NULL;
}
if( NULL != prog_named_pipe_w) {
free(prog_named_pipe_w);
prog_named_pipe_w = NULL;
}
return OPAL_SUCCESS;
}
/*
* C/R Signal Handler.
* Once a signal is received then the notification thread is notified
* so it can communicate with the checkpoint command to take the approprate
* action.
*/
static void opal_cr_entry_point_signal_handler (int signo)
{
if( opal_cr_entry_point_signal != signo ) {
/* Not our signal */
return;
}
/*
* Signal thread to start checkpoint handshake
*/
opal_cr_checkpoint_request = OPAL_CR_STATUS_REQUESTED;
}
/*
* Used only for debugging SIGPIPE problems
*/
@ -818,346 +714,6 @@ static void opal_cr_sigpipe_debug_signal_handler (int signo)
}
}
/*
* Respond to an asynchronous checkpoint request
*/
int cr_notify_response(opal_cr_ckpt_cmd_state_t resp)
{
static int app_term = 0, app_pid = 0;
static opal_crs_base_snapshot_t *snapshot = NULL;
static int prog_named_read_pipe_fd, prog_named_write_pipe_fd;
static int len = 0;
static int cr_state;
int ret, exit_status = OPAL_SUCCESS;
int tmp_resp;
char *tmp_str = NULL;
ssize_t tmp_size = 0;
/* Commands from the command line tool */
unsigned char app_cmd;
if( opal_cr_currently_stalled ) {
goto STAGE_1;
}
/*
* Open a named pipe for our application
*/
if (OPAL_SUCCESS != (ret = cr_entry_point_notify_reopen_files(&prog_named_read_pipe_fd, &prog_named_write_pipe_fd))) {
goto ckpt_cleanup;
}
/*
* Get the initial handshake command
*/
if( sizeof(int) != (ret = read(prog_named_read_pipe_fd, &len, sizeof(int))) ) {
opal_output(opal_cr_output,
"opal_cr: cr_notify_response: Error: Unable to read the first handshake from named pipe (%s). %d\n",
prog_named_pipe_r, ret);
goto ckpt_cleanup;
}
tmp_resp = (int)resp;
if( sizeof(int) != (ret = write(prog_named_write_pipe_fd, &tmp_resp, sizeof(int)) ) ) {
opal_output(opal_cr_output,
"opal_cr: cr_notify_response: %d: Error: Unable to write to pipe (%s) ret = %d [Line %d]\n",
tmp_resp, prog_named_pipe_w, ret, __LINE__);
goto ckpt_cleanup;
}
/*
* Respond that the checkpoint is currently in progress
*/
if( OPAL_CHECKPOINT_CMD_IN_PROGRESS == resp ) {
opal_output_verbose(10, opal_cr_output,
"opal_cr: cr_notify_response: Checkpoint in progress, cannot start (%d)",
getpid());
goto ckpt_cleanup;
}
/*
* Respond that the application is unable to be checkpointed
*/
else if( OPAL_CHECKPOINT_CMD_NULL == resp ) {
opal_output_verbose(10, opal_cr_output,
"opal_cr: cr_notify_response: Non-checkpointable application, cannot start (%d)",
getpid());
goto ckpt_cleanup;
}
/*
* Respond that some error has occurred such that the application is
* not able to be checkpointed
*/
else if( OPAL_CHECKPOINT_CMD_ERROR == resp ) {
opal_output_verbose(10, opal_cr_output,
"opal_cr: cr_notify_response: Error generated, cannot start (%d)",
getpid());
goto ckpt_cleanup;
}
/*
* Respond signalng that we wish to respond to this request
*/
opal_output_verbose(10, opal_cr_output,
"opal_cr: cr_notify_response: Starting checkpoint request (%d)",
getpid());
/*
* Wait for a notify command from command line tool
*/
if( sizeof(app_cmd) != (ret = read(prog_named_read_pipe_fd, &app_cmd, sizeof(app_cmd))) ) {
opal_output(opal_cr_output,
"opal_cr: cr_notify_response: Error: Unable to read the requested command from named pipe (%s). %d\n",
prog_named_pipe_r, ret);
goto ckpt_cleanup;
}
/* get PID argument */
if( sizeof(int) != (ret = read(prog_named_read_pipe_fd, &app_pid, sizeof(int))) ) {
opal_output(opal_cr_output,
"opal_cr: cr_notify_response: Error: Unable to read the pid from named pipe (%s). %d\n",
prog_named_pipe_r, ret);
goto ckpt_cleanup;
}
/* get term argument */
if( sizeof(int) != (ret = read(prog_named_read_pipe_fd, &app_term, sizeof(int))) ) {
opal_output(opal_cr_output,
"opal_cr: cr_notify_response: Error: Unable to read the term from named pipe (%s). %d\n",
prog_named_pipe_r, ret);
goto ckpt_cleanup;
}
/* get Snapshot Handle argument */
if( sizeof(int) != (ret = read(prog_named_read_pipe_fd, &len, sizeof(int))) ) {
opal_output(opal_cr_output,
"opal_cr: cr_notify_response: Error: Unable to read the snapshot_handle len from named pipe (%s). %d\n",
prog_named_pipe_r, ret);
goto ckpt_cleanup;
}
tmp_size = sizeof(char) * len;
tmp_str = (char *) malloc(sizeof(char) * len);
if( tmp_size != (ret = read(prog_named_read_pipe_fd, tmp_str, (sizeof(char) * len))) ) {
opal_output(opal_cr_output,
"opal_cr: cr_notify_response: Error: Unable to read the snapshot_handle from named pipe (%s). %d\n",
prog_named_pipe_r, ret);
goto ckpt_cleanup;
}
/*
* If they didn't send anything of meaning then use the defaults
*/
snapshot = OBJ_NEW(opal_crs_base_snapshot_t);
if( 1 < strlen(tmp_str) ) {
if( NULL != snapshot->reference_name)
free( snapshot->reference_name );
snapshot->reference_name = strdup(tmp_str);
if( NULL != snapshot->local_location )
free( snapshot->local_location );
snapshot->local_location = opal_crs_base_get_snapshot_directory(snapshot->reference_name);
if( NULL != snapshot->remote_location )
free( snapshot->remote_location );
snapshot->remote_location = strdup(snapshot->local_location);
free(tmp_str);
tmp_str = NULL;
}
/* get Snapshot location argument */
if( sizeof(int) != (ret = read(prog_named_read_pipe_fd, &len, sizeof(int))) ) {
opal_output(opal_cr_output,
"opal_cr: cr_notify_response: Error: Unable to read the snapshot_location len from named pipe (%s). %d\n",
prog_named_pipe_r, ret);
goto ckpt_cleanup;
}
tmp_str = (char *) malloc(sizeof(char) * len);
tmp_size = sizeof(char) * len;
if( tmp_size != (ret = read(prog_named_read_pipe_fd, tmp_str, (sizeof(char) * len))) ) {
opal_output(opal_cr_output,
"opal_cr: cr_notify_response: Error: Unable to read the snapshot_location from named pipe (%s). %d\n",
prog_named_pipe_r, ret);
goto ckpt_cleanup;
}
/*
* If they didn't send anything of meaning then use the defaults
*/
if( 1 < strlen(tmp_str) ) {
if( NULL != snapshot->local_location)
free( snapshot->local_location );
asprintf(&(snapshot->local_location), "%s/%s", tmp_str, snapshot->reference_name);
if( NULL != snapshot->remote_location)
free( snapshot->remote_location );
snapshot->remote_location = strdup(snapshot->local_location);
free(tmp_str);
tmp_str = NULL;
}
/*
* Raise the notification flag.
* This will trigger the coordination, and checkpoint of the
* application if it is possible
*/
STAGE_1:
opal_cr_currently_stalled = false;
ret = opal_cr_inc_core(app_pid, snapshot, app_term, &cr_state);
if( OPAL_EXISTS == ret ) {
opal_output_verbose(5, opal_cr_output,
"opal_cr: cr_notify_response: Stalling the checkpoint progress until state is stable again (PID = %d)\n",
getpid());
opal_cr_currently_stalled = true;
return exit_status;
}
else if(OPAL_SUCCESS != ret) {
opal_output(opal_cr_output,
"opal_cr: cr_notify_response: Error: checkpoint notification failed. %d\n", ret);
goto ckpt_cleanup;
}
/* Don't stall any longer */
opal_cr_stall_check = false;
if(OPAL_CRS_RESTART == cr_state) {
opal_output_verbose(10, opal_cr_output,
"opal_cr: cr_notify_response: Restarting...(%d)\n",
getpid());
app_term = false;
/* Do not respond to the non-existent command line tool */
goto ckpt_cleanup;
}
else if(cr_state == OPAL_CRS_CONTINUE) {
; /* Don't need to do anything here */
}
else if(cr_state == OPAL_CRS_TERM ) {
; /* Don't need to do anything here */
}
else {
opal_output_verbose(5, opal_cr_output,
"opal_cr: cr_notify_response: Unknown cr_state(%d) [%d]",
cr_state, getpid());
}
/*
* Return the expected variables to the command line tool
*/
len = strlen(snapshot->reference_name);
len++; /* To account for the Null character */
if( sizeof(int) != (ret = write(prog_named_write_pipe_fd, &len, sizeof(int))) ) {
opal_output(opal_cr_output,
"opal_cr: cr_notify_response: Error: Unable to write fname length to named pipe (%s). %d.\n",
prog_named_pipe_w, ret);
goto ckpt_cleanup;
}
if(len > 0) {
if( (ssize_t)(sizeof(char) * len) !=
(ret = write(prog_named_write_pipe_fd, snapshot->reference_name, (sizeof(char) * len))) ) {
opal_output(opal_cr_output,
"opal_cr: cr_notify_response: Error: Unable to write snapshot->reference_name to named pipe (%s). %d\n",
prog_named_pipe_w, ret);
goto ckpt_cleanup;
}
}
if( sizeof(int) != (ret = write(prog_named_write_pipe_fd, &cr_state, sizeof(int))) ) {
opal_output(opal_cr_output,
"opal_cr: cr_notify_response: Error: Unable to write cr_state to named pipe (%s). %d\n",
prog_named_pipe_w, ret);
goto ckpt_cleanup;
}
ckpt_cleanup:
close(prog_named_write_pipe_fd);
close(prog_named_read_pipe_fd);
remove(prog_named_pipe_r);
remove(prog_named_pipe_w);
if(app_term) {
opal_output_verbose(10, opal_cr_output,
"opal_cr: cr_notify_response: User has asked to terminate the application");
exit(OPAL_SUCCESS);
}
/* Prepare to wait for another checkpoint action */
opal_cr_checkpointing = OPAL_CR_STATUS_NONE;
opal_cr_currently_stalled = false;
return exit_status;
}
static int cr_entry_point_notify_reopen_files(int *prog_read_fd, int *prog_write_fd)
{
int ret = OPAL_ERR_NOT_IMPLEMENTED;
#ifndef HAVE_MKFIFO
return ret;
#else
#ifdef __WINDOWS__
return ret;
#else
/*
* Open up the read pipe
*/
if( (ret = mkfifo(prog_named_pipe_r, 0660)) < 0) {
if(EEXIST == ret || -1 == ret ) {
opal_output_verbose(10, opal_cr_output,
"opal_cr: notify_reopen_files: mkfifo failed because file (%s) already exists, attempting to use this pipe. (%d)",
prog_named_pipe_r, ret);
}
else {
opal_output(opal_cr_output,
"opal_cr: notify_reopen_files: Error: mkfifo failed to make named pipe (%s). (%d)\n",
prog_named_pipe_r, ret);
return OPAL_ERROR;
}
}
*prog_read_fd = open(prog_named_pipe_r, O_RDWR);
if(*prog_read_fd < 0) {
opal_output(opal_cr_output,
"opal_cr: init: Error: open failed to open the named pipe (%s). %d\n",
prog_named_pipe_r, *prog_read_fd);
return OPAL_ERROR;
}
/*
* Open up the write pipe
*/
if( (ret = mkfifo(prog_named_pipe_w, 0660)) < 0) {
if(EEXIST == ret || -1 == ret ) {
opal_output_verbose(10, opal_cr_output,
"opal_cr: notify_reopen_files: mkfifo failed because file (%s) already exists, attempting to use this pipe. (%d)",
prog_named_pipe_w, ret);
}
else {
opal_output(opal_cr_output,
"opal_cr: notify_reopen_files: Error: mkfifo failed to make named pipe (%s). (%d)\n",
prog_named_pipe_w, ret);
return OPAL_ERROR;
}
}
*prog_write_fd = open(prog_named_pipe_w, O_WRONLY);
if(*prog_write_fd < 0) {
opal_output(opal_cr_output,
"opal_cr: notify_reopen_files: Error: open failed to open the named pipe (%s). (%d)\n",
prog_named_pipe_w, *prog_write_fd);
return OPAL_ERROR;
}
return OPAL_SUCCESS;
#endif /* __WINDOWS__ */
#endif /* HAVE_MKFIFO */
}
#if OPAL_ENABLE_FT_THREAD == 1
static void* opal_cr_thread_fn(opal_object_t *obj)
{

Просмотреть файл

@ -61,28 +61,39 @@ enum opal_cr_ckpt_cmd_state_t {
OPAL_CR_STATUS_NONE, /* No checkpoint in progress */
OPAL_CR_STATUS_REQUESTED, /* Checkpoint has been requested */
OPAL_CR_STATUS_RUNNING, /* Checkpoint is currently running */
OPAL_CR_STATUS_TERM /* Checkpoint is running and will terminate process upon completion */
OPAL_CR_STATUS_TERM, /* Checkpoint is running and will terminate process upon completion */
/* State of the continue operation */
OPAL_CR_STATUS_CONTINUE,
/* State of the restart operation */
OPAL_CR_STATUS_RESTART_PRE,
OPAL_CR_STATUS_RESTART_POST
};
typedef enum opal_cr_ckpt_cmd_state_t opal_cr_ckpt_cmd_state_t;
/* An output handle to be used by the cr runtime
* functionality as an argument to opal_output() */
OPAL_DECLSPEC extern int opal_cr_output;
/* Directory containing the named pipes for communication
* with the opal-checkpoint tool */
OPAL_DECLSPEC extern char * opal_cr_pipe_dir;
/* Signal that opal-checkpoint uses to contact the
* application process */
OPAL_DECLSPEC extern int opal_cr_entry_point_signal;
/* If Checkpointing is enabled in this application */
OPAL_DECLSPEC extern bool opal_cr_is_enabled;
/* If the application running is a tool
* (e.g., opal-checkpoint, orted, ...) */
OPAL_DECLSPEC extern bool opal_cr_is_tool;
/* An output handle to be used by the cr runtime
* functionality as an argument to opal_output() */
OPAL_DECLSPEC extern int opal_cr_output;
/* If a checkpoint has been requested */
OPAL_DECLSPEC extern int opal_cr_checkpoint_request;
/* The current state of a checkpoint operation */
OPAL_DECLSPEC extern int opal_cr_checkpointing;
OPAL_DECLSPEC extern int opal_cr_checkpointing_state;
/*
* If this is an application that doesn't want to have
@ -217,16 +228,6 @@ typedef enum opal_cr_ckpt_cmd_state_t opal_cr_ckpt_cmd_state_t;
/*******************************
* Notification Routines
*******************************/
/*
* Init OPAL entry point functionality
*/
OPAL_DECLSPEC int opal_cr_entry_point_init(void);
/*
* Finalize OPAL entry point functionality
*/
OPAL_DECLSPEC int opal_cr_entry_point_finalize(void);
/**
* A function to respond to the async checkpoint request
* this is useful when figuring out who should respond

Просмотреть файл

@ -41,6 +41,9 @@
#ifdef HAVE_SYS_STAT_H
#include <sys/stat.h>
#endif
#ifdef HAVE_FCNTL_H
#include <fcntl.h>
#endif /* HAVE_FCNTL_H */
#ifdef HAVE_SYS_TYPES_H
#include <sys/types.h>
#endif
@ -75,7 +78,7 @@ static int initialize(int argc, char *argv[]);
static int finalize(void);
static int parse_args(int argc, char *argv[]);
static int check_file(char *given_filename);
static int post_env_vars(int prev_pid);
static int post_env_vars(int prev_pid, char *location);
/*****************************************
* Global Vars for Command line Arguments
@ -185,7 +188,7 @@ main(int argc, char *argv[])
char * base = NULL;
base = opal_crs_base_get_snapshot_directory(opal_restart_globals.filename);
expected_crs_comp = strdup(opal_crs_base_extract_expected_component(base, &prev_pid));
opal_crs_base_extract_expected_component(base, &expected_crs_comp, &prev_pid);
free(base);
}
@ -260,12 +263,18 @@ main(int argc, char *argv[])
/* Since some checkpoint/restart systems don't pass along env vars to the
* restarted app, we need to take care of that.
*
* Included here is the creation of any files or directories that need to be
* created before the process is restarted.
*/
if(OPAL_SUCCESS != (ret = post_env_vars(prev_pid) ) ) {
if(OPAL_SUCCESS != (ret = post_env_vars(prev_pid, snapshot->local_location) ) ) {
exit_status = ret;
goto cleanup;
}
/*
* Do the actual restart
*/
ret = opal_crs.crs_restart(snapshot,
opal_restart_globals.forked,
&child_pid);
@ -513,12 +522,15 @@ static int check_file(char *given_filename)
return exit_status;
}
static int post_env_vars(int prev_pid)
static int post_env_vars(int prev_pid, char *location)
{
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",
@ -535,17 +547,82 @@ static int post_env_vars(int prev_pid)
asprintf(&proc_file, "/tmp/%s-%d", OPAL_CR_BASE_ENV_NAME, prev_pid);
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
*/
opal_crs_base_metadata_read_token(location, CRS_METADATA_MKDIR, &loc_mkdir);
argc = opal_argv_count(loc_mkdir);
for( i = 0; i < argc; ++i ) {
if( NULL != command ) {
free(command);
command = NULL;
}
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(location, CRS_METADATA_TOUCH, &loc_touch);
argc = opal_argv_count(loc_touch);
for( i = 0; i < argc; ++i ) {
if( NULL != command ) {
free(command);
command = NULL;
}
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)
if( NULL != command) {
free(command);
if( NULL != proc_file)
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;
}
return exit_status;
}

Просмотреть файл

@ -631,7 +631,7 @@ int orte_snapc_base_add_vpid_metadata( orte_process_name_t *proc,
char *snapshot_ref,
char *snapshot_location)
{
int exit_status = ORTE_SUCCESS;
int ret, exit_status = ORTE_SUCCESS;
FILE * meta_data = NULL;
char * meta_data_fname = NULL;
char * crs_comp = NULL;
@ -659,8 +659,7 @@ int orte_snapc_base_add_vpid_metadata( orte_process_name_t *proc,
orte_util_convert_process_name_to_string(&proc_name, proc);
/* Extract the checkpointer */
crs_comp = opal_crs_base_extract_expected_component(snapshot_location, &prev_pid);
if( NULL == crs_comp ) {
if( OPAL_SUCCESS != (ret = opal_crs_base_extract_expected_component(snapshot_location, &crs_comp, &prev_pid)) ) {
exit_status = ORTE_ERROR;
goto cleanup;
}

Просмотреть файл

@ -86,7 +86,6 @@ int app_coord_init() {
* Register the INC notification callback
*/
opal_cr_reg_notify_callback(snapc_full_app_notify_response, &prev_notify_func);
opal_cr_entry_point_finalize();
/* String representation of the PID */
asprintf(&tmp_pid, "%d", getpid());
@ -198,9 +197,15 @@ int snapc_full_app_notify_response(opal_cr_ckpt_cmd_state_t resp)
/*
* Begin checkpoint
* - Init the checkpoint metadata file
*/
OPAL_OUTPUT_VERBOSE((10, mca_snapc_full_component.super.output_handle,
"App) notify_response: Start checkpoint..."));
if( OPAL_SUCCESS != (ret = opal_crs_base_init_snapshot_directory(local_snapshot) ) ) {
opal_output(0, "App) Error: Unable to initalize the snapshot directory!\n");
exit_status = ret;
goto ckpt_cleanup;
}
STAGE_1:
opal_cr_currently_stalled = false;
@ -269,8 +274,8 @@ int snapc_full_app_notify_response(opal_cr_ckpt_cmd_state_t resp)
}
/* Prepare to wait for another checkpoint action */
opal_cr_checkpointing = OPAL_CR_STATUS_NONE;
opal_cr_currently_stalled = false;
opal_cr_checkpointing_state = OPAL_CR_STATUS_NONE;
opal_cr_currently_stalled = false;
return exit_status;
}