/**
* @brief
* main - the initialization and main loop of pbs_daemon
*/
int
main(int argc, char *argv[])
{
char jobfile[MAXPATHLEN+1];
char jobfile_full[MAXPATHLEN+1];
pbs_net_t hostaddr = 0;
int port = -1;
int move_type = -1;
pbs_list_head attrl;
enum conn_type cntype = ToServerDIS;
int con = -1;
char *destin;
int encode_type;
int i;
job *jobp;
char job_id[PBS_MAXSVRJOBID+1];
attribute *pattr;
struct attropl *pqjatr; /* list (single) of attropl for quejob */
char script_name[MAXPATHLEN+1];
int in_server = -1;
char *param_name, *param_val;
char buf[4096];
struct hostent *hp;
struct in_addr addr;
char *credbuf = NULL;
size_t credlen = 0;
int prot = PROT_TCP;
/*the real deal or output version and exit?*/
execution_mode(argc, argv);
/* If we are not run with real and effective uid of 0, forget it */
pbs_loadconf(0);
if (!isAdminPrivilege(getlogin())) {
fprintf(stderr, "%s: Must be run by root\n", argv[0]);
exit(SEND_JOB_FATAL);
}
/* initialize the pointers in the resource_def array */
for (i = 0; i < (svr_resc_size - 1); ++i)
svr_resc_def[i].rs_next = &svr_resc_def[i+1];
/* last entry is left with null pointer */
/* set single threaded mode */
pbs_client_thread_set_single_threaded_mode();
/* disable attribute verification */
set_no_attribute_verification();
/* initialize the thread context */
if (pbs_client_thread_init_thread_context() != 0) {
fprintf(stderr, "%s: Unable to initialize thread context\n",
argv[0]);
exit(SEND_JOB_FATAL);
}
if(set_msgdaemonname("PBS_send_job")) {
fprintf(stderr, "Out of memory\n");
return 1;
}
winsock_init();
connection_init();
while (fgets(buf, sizeof(buf), stdin) != NULL) {
buf[strlen(buf)-1] = '\0'; /* gets rid of newline */
param_name = buf;
param_val = strchr(buf, '=');
if (param_val) {
*param_val = '\0';
param_val++;
} else { /* bad param_val -- skipping */
break;
}
if (strcmp(param_name, "jobfile") == 0) {
jobfile[0] = '\0';
strncpy(jobfile, param_val, MAXPATHLEN);
} else if (strcmp(param_name, "destaddr") == 0) {
hostaddr = atol(param_val);
} else if (strcmp(param_name, "destport") == 0) {
port = atoi(param_val);
} else if (strcmp(param_name, "move_type") == 0) {
move_type = atoi(param_val);
} else if (strcmp(param_name, "in_server") == 0) {
in_server = atoi(param_val);
} else if (strcmp(param_name, "server_name") == 0) {
server_name[0] = '\0';
strncpy(server_name, param_val, PBS_MAXSERVERNAME);
} else if (strcmp(param_name, "server_host") == 0) {
server_host[0] = '\0';
strncpy(server_host, param_val, (sizeof(server_host) - 1));
} else if (strcmp(param_name, "server_addr") == 0) {
pbs_server_addr = atol(param_val);
} else if (strcmp(param_name, "server_port") == 0) {
pbs_server_port_dis = atoi(param_val);
} else if (strcmp(param_name, "log_file") == 0) {
log_file = strdup(param_val);
} else if (strcmp(param_name, "path_log") == 0) {
path_log[0] = '\0';
strncpy(path_log, param_val, MAXPATHLEN);
} else if (strcmp(param_name, "path_jobs") == 0) {
path_jobs = strdup(param_val);
} else if (strcmp(param_name, "path_spool") == 0) {
path_spool = strdup(param_val);
} else if (strcmp(param_name, "path_rescdef") == 0) {
path_rescdef = strdup(param_val);
} else if (strcmp(param_name, "path_users") == 0) {
path_users = strdup(param_val);
} else if (strcmp(param_name, "path_hooks_workdir") == 0) {
path_hooks_workdir = strdup(param_val);
if (path_hooks_workdir == NULL)
exit(SEND_JOB_FATAL);
} else if (strcmp(param_name, "svr_history_enable") == 0) {
svr_history_enable = atol(param_val);
} else if (strcmp(param_name, "svr_history_duration") == 0) {
svr_history_duration = atol(param_val);
} else if (strcmp(param_name, "single_signon_password_enable") == 0) {
if (decode_b(&server.sv_attr[(int)SRV_ATR_ssignon_enable],
NULL, NULL, param_val) != 0) {
fprintf(stderr, "%s: failed to set ssignon_password_enable\n", argv[0]);
exit(SEND_JOB_FATAL);
}
} else if (strcmp(param_name, "script_name") == 0) {
strncpy(script_name, param_val, MAXPATHLEN + 1);
} else
break;
}
time(&time_now);
(void)log_open_main(log_file, path_log, 1); /* silent open */
if (setup_resc(1) == -1) {
/* log_buffer set in setup_resc */
log_err(-1, "pbsd_send_job(setup_resc)", log_buffer);
return (-1);
}
if( strlen(jobfile) == 0 || hostaddr == 0 || port == 0 || move_type == -1 || \
in_server == -1 || strlen(server_name) == 0 || strlen(server_host) == 0 || \
pbs_server_addr == 0 || pbs_server_port_dis == 0 || \
strlen(path_log) == 0 || path_jobs == NULL || \
path_spool == NULL || path_users == NULL ) {
log_err(-1, "pbs_send_job", "error on one of the parameters");
log_close(0); /* silent close */
exit(SEND_JOB_FATAL);
}
CLEAR_HEAD(task_list_immed);
CLEAR_HEAD(task_list_timed);
CLEAR_HEAD(task_list_event);
CLEAR_HEAD(svr_queues);
CLEAR_HEAD(svr_alljobs);
CLEAR_HEAD(svr_newjobs);
CLEAR_HEAD(svr_allresvs);
CLEAR_HEAD(svr_newresvs);
CLEAR_HEAD(svr_deferred_req);
CLEAR_HEAD(svr_unlicensedjobs);
strcpy(jobfile_full, path_jobs);
strcat(jobfile_full, jobfile);
if (chk_save_file(jobfile_full) != 0) {
sprintf(log_buffer, "Error opening jobfile=%s", jobfile);
log_err(-1, __func__, log_buffer);
goto fatal_exit;
}
if ((jobp=job_recov_fs(jobfile, RECOV_SUBJOB)) == NULL) {
sprintf(log_buffer, "Failed to recreate job in jobfile=%s", jobfile);
log_err(-1, __func__, log_buffer);
goto fatal_exit;
}
/* now delete the temp job file that was created by job_save_fs in server code
* jobs are in database now, no need to keep in filesystem
*/
unlink(jobfile_full);
if (in_server)
append_link(&svr_alljobs, &jobp->ji_alljobs, jobp);
/* select attributes/resources to send based on move type */
if (move_type == MOVE_TYPE_Exec) {
resc_access_perm = ATR_DFLAG_MOM;
encode_type = ATR_ENCODE_MOM;
cntype = ToServerDIS;
} else {
resc_access_perm = ATR_DFLAG_USWR | ATR_DFLAG_OPWR |
ATR_DFLAG_MGWR | ATR_DFLAG_SvRD;
encode_type = ATR_ENCODE_SVR;
svr_dequejob(jobp);
}
CLEAR_HEAD(attrl);
pattr = jobp->ji_wattr;
for (i=0; i < (int)JOB_ATR_LAST; i++) {
if ((job_attr_def+i)->at_flags & resc_access_perm) {
(void)(job_attr_def+i)->at_encode(pattr+i, &attrl,
(job_attr_def+i)->at_name, NULL,
encode_type, NULL);
}
}
attrl_fixlink(&attrl);
/* script name is passed from parent */
/* get host name */
pbs_loadconf(0);
addr.s_addr = htonl(hostaddr);
hp = gethostbyaddr((void *)&addr, sizeof(struct in_addr), AF_INET);
if (hp == NULL) {
sprintf(log_buffer, "%s: h_errno=%d",
inet_ntoa(addr), h_errno);
log_err(-1, __func__, log_buffer);
}
else {
/* read any credential file */
(void)get_credential(hp->h_name, jobp, PBS_GC_BATREQ,
&credbuf, &credlen);
}
/* save the job id for when after we purge the job */
(void)strcpy(job_id, jobp->ji_qs.ji_jobid);
con = -1;
DIS_tcparray_init();
for (i=0; i<RETRY; i++) {
pbs_errno = 0;
/* connect to receiving server with retries */
if (i > 0) { /* recycle after an error */
if (con >= 0)
svr_disconnect(con);
if (should_retry_route(pbs_errno) == -1) {
goto fatal_exit; /* fatal error, don't retry */
}
sleep(1<<i);
}
if ((con = svr_connect(hostaddr, port, 0, cntype, prot)) ==
PBS_NET_RC_FATAL) {
(void)sprintf(log_buffer, "send_job failed to %lx port %d",
hostaddr, port);
log_err(pbs_errno, __func__, log_buffer);
goto fatal_exit;
} else if (con == PBS_NET_RC_RETRY) {
pbs_errno = WSAECONNREFUSED; /* should retry */
continue;
}
/*
* if the job is substate JOB_SUBSTATE_TRNOUTCM which means
* we are recovering after being down or a late failure, we
* just want to send the "read-to-commit/commit"
*/
if (jobp->ji_qs.ji_substate != JOB_SUBSTATE_TRNOUTCM) {
if (jobp->ji_qs.ji_substate != JOB_SUBSTATE_TRNOUT) {
jobp->ji_qs.ji_substate = JOB_SUBSTATE_TRNOUT;
}
pqjatr = &((svrattrl *)GET_NEXT(attrl))->al_atopl;
destin = jobp->ji_qs.ji_destin;
if (PBSD_queuejob(con, jobp->ji_qs.ji_jobid, destin, pqjatr, NULL, prot, NULL)== 0) {
if (pbs_errno == PBSE_JOBEXIST && move_type == MOVE_TYPE_Exec) {
/* already running, mark it so */
log_event(PBSEVENT_ERROR,
PBS_EVENTCLASS_JOB, LOG_INFO,
jobp->ji_qs.ji_jobid, "Mom reports job already running");
goto ok_exit;
} else if ((pbs_errno == PBSE_HOOKERROR) ||
(pbs_errno == PBSE_HOOK_REJECT) ||
(pbs_errno == PBSE_HOOK_REJECT_RERUNJOB) ||
(pbs_errno == PBSE_HOOK_REJECT_DELETEJOB)) {
char name_buf[MAXPATHLEN+1];
int rfd;
int len;
char *reject_msg;
int err;
err = pbs_errno;
reject_msg = pbs_geterrmsg(con);
(void)snprintf(log_buffer, sizeof(log_buffer),
"send of job to %s failed error = %d reject_msg=%s",
destin, err,
reject_msg?reject_msg:"");
log_event(PBSEVENT_JOB, PBS_EVENTCLASS_JOB,
LOG_INFO, jobp->ji_qs.ji_jobid,
log_buffer);
(void)strcpy(name_buf, path_hooks_workdir);
(void)strcat(name_buf, jobp->ji_qs.ji_jobid);
(void)strcat(name_buf, HOOK_REJECT_SUFFIX);
if ((reject_msg != NULL) &&
(reject_msg[0] != '\0')) {
if ((rfd = open(name_buf,
O_RDWR|O_CREAT|O_TRUNC, 0600)) == -1) {
snprintf(log_buffer,
sizeof(log_buffer),
"open of reject file %s failed: errno %d",
name_buf, errno);
log_event(PBSEVENT_JOB, PBS_EVENTCLASS_JOB, LOG_INFO, jobp->ji_qs.ji_jobid, log_buffer);
} else {
secure_file(name_buf, "Administrators",
READS_MASK|WRITES_MASK|STANDARD_RIGHTS_REQUIRED);
setmode(rfd, O_BINARY);
len = strlen(reject_msg)+1;
/* write also trailing null char */
if (write(rfd, reject_msg, len) != len) {
snprintf(log_buffer,
sizeof(log_buffer),
"write to file %s incomplete: errno %d", name_buf, errno);
log_event(PBSEVENT_JOB, PBS_EVENTCLASS_JOB, LOG_INFO, jobp->ji_qs.ji_jobid, log_buffer);
}
close(rfd);
}
}
if (err == PBSE_HOOKERROR)
exit(SEND_JOB_HOOKERR);
if (err == PBSE_HOOK_REJECT)
exit(SEND_JOB_HOOK_REJECT);
if (err == PBSE_HOOK_REJECT_RERUNJOB)
exit(SEND_JOB_HOOK_REJECT_RERUNJOB);
if (err == PBSE_HOOK_REJECT_DELETEJOB)
exit(SEND_JOB_HOOK_REJECT_DELETEJOB);
} else {
(void)sprintf(log_buffer, "send of job to %s failed error = %d", destin, pbs_errno);
log_event(PBSEVENT_JOB, PBS_EVENTCLASS_JOB, LOG_INFO, jobp->ji_qs.ji_jobid, log_buffer);
continue;
}
}
if (jobp->ji_qs.ji_svrflags & JOB_SVFLG_SCRIPT) {
if (PBSD_jscript(con, script_name, prot, NULL) != 0)
continue;
}
if (credlen > 0) {
int ret;
ret = PBSD_jcred(con,
jobp->ji_extended.ji_ext.ji_credtype,
credbuf, credlen, prot, NULL);
if ((ret == 0) || (i == (RETRY - 1)))
free(credbuf); /* free credbuf if credbuf is sent successfully OR */
/* at the end of all retry attempts */
if (ret != 0)
continue;
}
if ((move_type == MOVE_TYPE_Exec) &&
(jobp->ji_qs.ji_svrflags & JOB_SVFLG_HASRUN) &&
(hostaddr != pbs_server_addr)) {
/* send files created on prior run */
if ((move_job_file(con, jobp, StdOut, prot) != 0) ||
(move_job_file(con, jobp, StdErr, prot) != 0) ||
(move_job_file(con, jobp, Chkpt, prot) != 0))
continue;
}
jobp->ji_qs.ji_substate = JOB_SUBSTATE_TRNOUTCM;
}
if (PBSD_rdytocmt(con, job_id, prot, NULL) != 0)
continue;
if (PBSD_commit(con, job_id, prot, NULL) != 0)
goto fatal_exit;
goto ok_exit; /* This child process is all done */
}
if (con >= 0)
svr_disconnect(con);
/*
* If connection is actively refused by the execution node(or mother superior) OR
* the execution node(or mother superior) is rejecting request with error
* PBSE_BADHOST(failing to authorize server host), the node should be marked down.
*/
if ((move_type == MOVE_TYPE_Exec) && (pbs_errno == WSAECONNREFUSED || pbs_errno == PBSE_BADHOST)) {
i = SEND_JOB_NODEDW;
} else if (should_retry_route(pbs_errno) == -1) {
i = SEND_JOB_FATAL;
} else {
i = SEND_JOB_RETRY;
}
(void)sprintf(log_buffer, "send_job failed with error %d", pbs_errno);
log_event(PBSEVENT_DEBUG, PBS_EVENTCLASS_JOB, LOG_NOTICE,
jobp->ji_qs.ji_jobid, log_buffer);
log_close(0);
net_close(-1);
unlink(script_name);
exit(i);
fatal_exit:
if (con >= 0)
svr_disconnect(con);
log_close(0);
net_close(-1);
unlink(script_name);
exit(SEND_JOB_FATAL);
ok_exit:
if (con >= 0)
svr_disconnect(con);
log_close(0);
net_close(-1);
unlink(script_name);
exit(SEND_JOB_OK);
}
/**
*
* @brief
* Send a job over the network to some other server or MOM.
* @par
* Under Linux/Unix, this starts a child process to do the work.
* Connect to the destination host and port,
* and go through the protocol to transfer the job.
* Signals are blocked.
*
* @param[in] jobp - pointer to the job being sent.
* @param[in] hostaddr - the address of host to send job to, host byte order.
* @param[in] port - the destination port, host byte order
* @param[in] move_type - the type of move (e.g. MOVE_TYPE_exec)
* @param[in] post_func - the function to execute once the child process
* sending job completes (Linux/Unix only)
* @param[in] data - input data to 'post_func'
*
* @return int
* @retval 2 parent : success (child forked)
* @retval -1 parent : on failure (pbs_errno set to error number)
* @retval SEND_JOB_OK child : 0 success, job sent
* @retval SEND_JOB_FATAL child : 1 permenent failure or rejection,
* @retval SEND_JOB_RETRY child : 2 failed but try again
* @retval SEND_JOB_NODEDW child : 3 execution node down, retry different node
*/
int
send_job(job *jobp, pbs_net_t hostaddr, int port, int move_type,
void (*post_func)(struct work_task *), struct batch_request *preq)
{
#ifdef WIN32
char cmdline[80];
pio_handles pio;
char buf[4096];
struct work_task *ptask;
int newstate;
int newsub;
long tempval;
char script_name[MAXPATHLEN+1];
int gridproxy_cred = 0;
#ifdef PBS_CRED_GRIDPROXY
if (jobp->ji_extended.ji_ext.ji_credtype == PBS_CREDTYPE_GRIDPROXY)
gridproxy_cred = 1;
#endif
if (pbs_conf.pbs_use_tcp == 1 && move_type == MOVE_TYPE_Exec && gridproxy_cred == 0) {
return (send_job_exec(jobp, hostaddr, port, preq));
}
sprintf(cmdline, "%s/sbin/pbs_send_job", pbs_conf.pbs_exec_path);
if (win_popen(cmdline, "w", &pio, NULL) == 0) {
errno = GetLastError();
pbs_errno = errno;
(void)sprintf(log_buffer, "executing %s for job %s failed errno=%d", cmdline, jobp->ji_qs.ji_jobid, errno);
log_event(PBSEVENT_DEBUG, PBS_EVENTCLASS_JOB, LOG_ERR,
jobp->ji_qs.ji_jobid, log_buffer);
/* force re-eval of job state out of Transit */
svr_evaljobstate(jobp, &newstate, &newsub, 1);
svr_setjobstate(jobp, newstate, newsub);
win_pclose(&pio);
return (-1);
}
ptask = set_task(WORK_Deferred_Child, (long)pio.pi.hProcess, post_func, preq);
if (!ptask) {
log_err(errno, __func__, msg_err_malloc);
errno = ENOMEM;
pbs_errno = errno;
win_pclose(&pio);
/* force re-eval of job state out of Transit */
svr_evaljobstate(jobp, &newstate, &newsub, 1);
svr_setjobstate(jobp, newstate, newsub);
return (-1);
} else {
ptask->wt_parm2 = jobp;
append_link(&((job *)jobp)->ji_svrtask, &ptask->wt_linkobj, ptask);
}
script_name[0] = '\0';
/* if job has a script read it from database */
if (jobp->ji_qs.ji_svrflags & JOB_SVFLG_SCRIPT) {
/*
* copy the job script from database to a temp file
* PBSD_jscript works with a file
* delete it at the end of the send
*/
if (svr_create_tmp_jobscript(jobp, &script_name) != 0) {
pbs_errno = PBSE_SYSTEM;
snprintf(log_buffer, sizeof(log_buffer),
"Failed to create temporary job script for job %s",
jobp->ji_qs.ji_jobid);
log_err(pbs_errno, "send_job", log_buffer);
win_pclose2(&pio);
return (-1);
}
}
addpid(pio.pi.hProcess);
/* our job is to calc eligible time accurately and save it */
/* on new server, accrue type should be calc afresh */
/* Note: if job is being sent for execution on mom, then don't calc eligible time */
if ((jobp->ji_wattr[(int)JOB_ATR_accrue_type].at_val.at_long == JOB_ELIGIBLE) &&
(server.sv_attr[(int)SRV_ATR_EligibleTimeEnable].at_val.at_long == 1) &&
(move_type != MOVE_TYPE_Exec)) {
tempval = ((long)time_now - jobp->ji_wattr[(int)JOB_ATR_sample_starttime].at_val.at_long);
jobp->ji_wattr[(int)JOB_ATR_eligible_time].at_val.at_long += tempval;
jobp->ji_wattr[(int)JOB_ATR_eligible_time].at_flags |= ATR_VFLAG_MODCACHE;
}
/* in windows code, a child process "w32_send_job" handles the send
* This needs the job information, so we save using the filesystem
* This avoids the child process from having to "connect" to the database again
* The file is deleted by the send_job child process when it has done recovering the job
*/
job_save_fs(jobp, SAVEJOB_FULLFORCE); /* so the spawned process can get a fresh copy of job */
if (*jobp->ji_qs.ji_fileprefix != '\0')
sprintf(buf, "jobfile=%s%s\n", jobp->ji_qs.ji_fileprefix, JOB_FILE_SUFFIX);
else
sprintf(buf, "jobfile=%s%s\n", jobp->ji_qs.ji_jobid, JOB_FILE_SUFFIX);
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "destaddr=%ld\n", hostaddr);
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "destport=%d\n", port);
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "move_type=%d\n", move_type);
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "in_server=%d\n", is_linked(&svr_alljobs, &jobp->ji_alljobs));
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "server_name=%s\n", (server_name?server_name:""));
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "server_host=%s\n", (server_host?server_host:""));
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "server_addr=%ld\n", pbs_server_addr);
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "server_port=%d\n", pbs_server_port_dis);
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "log_file=%s\n", (log_file?log_file:""));
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "path_log=%s\n", (path_log?path_log:""));
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "path_jobs=%s\n", (path_jobs?path_jobs:""));
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "path_spool=%s\n", (path_spool?path_spool:""));
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "path_rescdef=%s\n", (path_rescdef?path_rescdef:""));
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "path_users=%s\n", (path_users?path_users:""));
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "path_hooks_workdir=%s\n",
(path_hooks_workdir?path_hooks_workdir:""));
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "svr_history_enable=%ld\n", svr_history_enable);
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "svr_history_duration=%ld\n", svr_history_duration);
win_pwrite(&pio, buf, strlen(buf));
if ( (server.sv_attr[SRV_ATR_ssignon_enable].at_flags & \
ATR_VFLAG_SET) && \
(server.sv_attr[SRV_ATR_ssignon_enable].at_val.at_long == 1) )
strcpy(buf, "single_signon_password_enable=1\n");
else
strcpy(buf, "single_signon_password_enable=0\n");
win_pwrite(&pio, buf, strlen(buf));
sprintf(buf, "script_name=%s\n", script_name);
win_pwrite(&pio, buf, strlen(buf));
strcpy(buf, "quit\n");
win_pwrite(&pio, buf, strlen(buf));
win_pclose2(&pio); /* closes all handles except the process handle */
return (2);
#else
pbs_list_head attrl;
enum conn_type cntype = ToServerDIS;
int con;
char *credbuf = NULL;
size_t credlen = 0;
char *destin = jobp->ji_qs.ji_destin;
int encode_type;
int i;
char job_id[PBS_MAXSVRJOBID+1];
attribute *pattr;
pid_t pid;
struct attropl *pqjatr; /* list (single) of attropl for quejob */
char script_name[MAXPATHLEN+1];
struct work_task *ptask;
struct hostent *hp;
struct in_addr addr;
long tempval;
int gridproxy_cred = 0;
int rpp = 0;
#ifdef PBS_CRED_GRIDPROXY
if (jobp->ji_extended.ji_ext.ji_credtype == PBS_CREDTYPE_GRIDPROXY)
gridproxy_cred = 1;
#endif
if (pbs_conf.pbs_use_tcp == 1 && move_type == MOVE_TYPE_Exec && gridproxy_cred == 0) {
return (send_job_exec(jobp, hostaddr, port, preq));
}
script_name[0] = '\0';
/* if job has a script read it from database */
if (jobp->ji_qs.ji_svrflags & JOB_SVFLG_SCRIPT) {
/*
* copy the job script from database to a temp file
* PBSD_jscript works with a file
* delete it at the end of the send
*/
if (svr_create_tmp_jobscript(jobp, script_name) != 0) {
pbs_errno = PBSE_SYSTEM;
snprintf(log_buffer, sizeof(log_buffer),
"Failed to create temporary job script for job %s",
jobp->ji_qs.ji_jobid);
log_err(pbs_errno, "send_job", log_buffer);
return -1;
}
}
pid = fork();
if (pid == -1) { /* Error on fork */
log_err(errno, __func__, "fork failed\n");
pbs_errno = PBSE_SYSTEM;
return -1;
}
if (pid != 0) { /* The parent (main server) */
ptask = set_task(WORK_Deferred_Child, pid, post_func, preq);
if (!ptask) {
log_err(errno, __func__, msg_err_malloc);
return (-1);
} else {
ptask->wt_parm2 = jobp;
append_link(&((job *)jobp)->ji_svrtask,
&ptask->wt_linkobj, ptask);
}
return 2;
}
/*
* the child process
*
* set up signal cather for error return
*/
DBPRT(("%s: child started, sending to port %d\n", __func__, port))
rpp_terminate();
/* Unprotect child from being killed by kernel */
daemon_protect(0, PBS_DAEMON_PROTECT_OFF);
#ifdef WIN32
/* get host name */
/*
* If host address is loopback address then do not resolve with dns
* Use "localhost" as the host name.
*/
if ((htonl(hostaddr) == loopback_addr->sin_addr.s_addr)) {
(void)get_credential(LOCALHOST_SHORTNAME, jobp, PBS_GC_BATREQ,
&credbuf, &credlen);
} else {
#endif
addr.s_addr = htonl(hostaddr);
hp = gethostbyaddr((void *)&addr, sizeof(struct in_addr), AF_INET);
if (hp == NULL) {
sprintf(log_buffer, "%s: h_errno=%d",
inet_ntoa(addr), h_errno);
log_err(-1, __func__, log_buffer);
} else {
/* read any credential file */
(void)get_credential(hp->h_name, jobp, PBS_GC_BATREQ,
&credbuf, &credlen);
}
#ifdef WIN32
}
#endif
/* encode job attributes to be moved */
CLEAR_HEAD(attrl);
/* select attributes/resources to send based on move type */
if (move_type == MOVE_TYPE_Exec) {
resc_access_perm = ATR_DFLAG_MOM;
encode_type = ATR_ENCODE_MOM;
cntype = ToServerDIS;
} else {
resc_access_perm = ATR_DFLAG_USWR | ATR_DFLAG_OPWR |
ATR_DFLAG_MGWR | ATR_DFLAG_SvRD;
encode_type = ATR_ENCODE_SVR;
svr_dequejob(jobp); /* clears default resource settings */
}
/* our job is to calc eligible time accurately and save it */
/* on new server, accrue type should be calc afresh */
/* Note: if job is being sent for execution on mom, then don't calc eligible time */
if ((jobp->ji_wattr[(int)JOB_ATR_accrue_type].at_val.at_long == JOB_ELIGIBLE) &&
(server.sv_attr[(int)SRV_ATR_EligibleTimeEnable].at_val.at_long == 1) &&
(move_type != MOVE_TYPE_Exec)) {
tempval = ((long)time_now - jobp->ji_wattr[(int)JOB_ATR_sample_starttime].at_val.at_long);
jobp->ji_wattr[(int)JOB_ATR_eligible_time].at_val.at_long += tempval;
jobp->ji_wattr[(int)JOB_ATR_eligible_time].at_flags |= ATR_VFLAG_MODCACHE;
}
pattr = jobp->ji_wattr;
for (i=0; i < (int)JOB_ATR_LAST; i++) {
if ((job_attr_def+i)->at_flags & resc_access_perm) {
(void)(job_attr_def+i)->at_encode(pattr+i, &attrl,
(job_attr_def+i)->at_name, (char *)0,
encode_type, NULL);
}
}
attrl_fixlink(&attrl);
/* save the job id for when after we purge the job */
(void)strcpy(job_id, jobp->ji_qs.ji_jobid);
pbs_errno = 0;
con = -1;
for (i=0; i<RETRY; i++) {
/* connect to receiving server with retries */
if (i > 0) { /* recycle after an error */
if (con >= 0)
svr_disconnect(con);
if (should_retry_route(pbs_errno) == -1) {
/* delete the temp script file */
unlink(script_name);
exit(SEND_JOB_FATAL); /* fatal error, don't retry */
}
sleep(1<<i);
}
if ((con = svr_connect(hostaddr, port, 0, cntype, rpp)) ==
PBS_NET_RC_FATAL) {
(void)sprintf(log_buffer, "send_job failed to %lx port %d",
hostaddr, port);
log_err(pbs_errno, __func__, log_buffer);
/* delete the temp script file */
unlink(script_name);
if ((move_type == MOVE_TYPE_Exec) && (pbs_errno == PBSE_BADCRED))
exit(SEND_JOB_NODEDW);
exit(SEND_JOB_FATAL);
} else if (con == PBS_NET_RC_RETRY) {
pbs_errno = ECONNREFUSED; /* should retry */
continue;
}
/*
* if the job is substate JOB_SUBSTATE_TRNOUTCM which means
* we are recovering after being down or a late failure, we
* just want to send the commit"
*/
if (jobp->ji_qs.ji_substate != JOB_SUBSTATE_TRNOUTCM) {
if (jobp->ji_qs.ji_substate != JOB_SUBSTATE_TRNOUT) {
jobp->ji_qs.ji_substate = JOB_SUBSTATE_TRNOUT;
}
pqjatr = &((svrattrl *)GET_NEXT(attrl))->al_atopl;
if (PBSD_queuejob(con, jobp->ji_qs.ji_jobid, destin,
pqjatr, (char *)0, rpp, NULL) == 0) {
if (pbs_errno == PBSE_JOBEXIST &&
move_type == MOVE_TYPE_Exec) {
/* already running, mark it so */
log_event(PBSEVENT_ERROR, PBS_EVENTCLASS_JOB,
LOG_INFO, jobp->ji_qs.ji_jobid,
"Mom reports job already running");
exit(SEND_JOB_OK);
}
else if ((pbs_errno == PBSE_HOOKERROR) ||
(pbs_errno == PBSE_HOOK_REJECT) ||
(pbs_errno == PBSE_HOOK_REJECT_RERUNJOB) ||
(pbs_errno == PBSE_HOOK_REJECT_DELETEJOB)) {
char name_buf[MAXPATHLEN+1];
int rfd;
int len;
char *reject_msg;
int err;
err = pbs_errno;
reject_msg = pbs_geterrmsg(con);
(void)sprintf(log_buffer,
"send of job to %s failed error = %d reject_msg=%s",
destin, err,
reject_msg?reject_msg:"");
log_event(PBSEVENT_JOB, PBS_EVENTCLASS_JOB,
LOG_INFO, jobp->ji_qs.ji_jobid,
log_buffer);
(void)strcpy(name_buf, path_hooks_workdir);
(void)strcat(name_buf, jobp->ji_qs.ji_jobid);
(void)strcat(name_buf, HOOK_REJECT_SUFFIX);
if ((reject_msg != NULL) &&
(reject_msg[0] != '\0')) {
if ((rfd = open(name_buf,
O_RDWR|O_CREAT|O_TRUNC, 0600)) == -1) {
sprintf(log_buffer,
"open of reject file %s failed: errno %d",
name_buf, errno);
log_event(PBSEVENT_JOB,
PBS_EVENTCLASS_JOB,
LOG_INFO, jobp->ji_qs.ji_jobid,
log_buffer);
} else {
#ifdef WIN32
secure_file(name_buf, "Administrators",
READS_MASK|WRITES_MASK|STANDARD_RIGHTS_REQUIRED);
setmode(rfd, O_BINARY);
#endif
len = strlen(reject_msg)+1;
/* write also trailing null char */
if (write(rfd, reject_msg, len) != len) {
sprintf(log_buffer,
"write to file %s incomplete: errno %d", name_buf, errno);
log_event(PBSEVENT_JOB,
PBS_EVENTCLASS_JOB,
LOG_INFO, jobp->ji_qs.ji_jobid,
log_buffer);
}
close(rfd);
}
}
if (err == PBSE_HOOKERROR)
exit(SEND_JOB_HOOKERR);
if (err == PBSE_HOOK_REJECT)
exit(SEND_JOB_HOOK_REJECT);
if (err == PBSE_HOOK_REJECT_RERUNJOB)
exit(SEND_JOB_HOOK_REJECT_RERUNJOB);
if (err == PBSE_HOOK_REJECT_DELETEJOB)
exit(SEND_JOB_HOOK_REJECT_DELETEJOB);
}
else {
(void)sprintf(log_buffer,
"send of job to %s failed error = %d",
destin, pbs_errno);
log_event(PBSEVENT_JOB, PBS_EVENTCLASS_JOB,
LOG_INFO, jobp->ji_qs.ji_jobid,
log_buffer);
continue;
}
}
if (jobp->ji_qs.ji_svrflags & JOB_SVFLG_SCRIPT) {
if (PBSD_jscript(con, script_name, rpp, NULL) != 0)
continue;
}
if (credlen > 0) {
int ret;
ret = PBSD_jcred(con,
jobp->ji_extended.ji_ext.ji_credtype,
credbuf, credlen, rpp, NULL);
if ((ret == 0) || (i == (RETRY - 1)))
free(credbuf); /* free credbuf if cred info is sent successfully OR */
/* at the end of all retry attempts */
if (ret != 0)
continue;
}
if ((move_type == MOVE_TYPE_Exec) &&
(jobp->ji_qs.ji_svrflags & JOB_SVFLG_HASRUN) &&
(hostaddr != pbs_server_addr)) {
/* send files created on prior run */
if ((move_job_file(con, jobp, StdOut, rpp, NULL) != 0) ||
(move_job_file(con, jobp, StdErr, rpp, NULL) != 0) ||
(move_job_file(con, jobp, Chkpt, rpp, NULL) != 0))
continue;
}
jobp->ji_qs.ji_substate = JOB_SUBSTATE_TRNOUTCM;
}
if (PBSD_rdytocmt(con, job_id, rpp, NULL) != 0)
continue;
if (PBSD_commit(con, job_id, rpp, NULL) != 0) {
/* delete the temp script file */
unlink(script_name);
exit(SEND_JOB_FATAL);
}
svr_disconnect(con);
/* delete the temp script file */
unlink(script_name);
exit(SEND_JOB_OK); /* This child process is all done */
}
if (con >= 0)
svr_disconnect(con);
/*
* If connection is actively refused by the execution node(or mother superior) OR
* the execution node(or mother superior) is rejecting request with error
* PBSE_BADHOST(failing to authorize server host), the node should be marked down.
*/
if ((move_type == MOVE_TYPE_Exec) && (pbs_errno == ECONNREFUSED || pbs_errno == PBSE_BADHOST)) {
i = SEND_JOB_NODEDW;
} else if (should_retry_route(pbs_errno) == -1) {
i = SEND_JOB_FATAL;
} else {
i = SEND_JOB_RETRY;
}
(void)sprintf(log_buffer, "send_job failed with error %d", pbs_errno);
log_event(PBSEVENT_DEBUG, PBS_EVENTCLASS_JOB, LOG_NOTICE,
jobp->ji_qs.ji_jobid, log_buffer);
/* delete the temp script file */
unlink(script_name);
exit(i);
return -1; /* NOT REACHED */
#endif /* !WIN32 */
}
/**
* @brief
* -submit job request
*
* @param[in] c - communication handle
* @param[in] attrib - ponter to attr list
* @param[in] script - job script
* @param[in] destination - host where job submitted
* @param[in] extend - buffer to hold cred info
*
* @return string
* @retval jobid success
* @retval NULL error
*
*/
char *
__pbs_submit(int c, struct attropl *attrib, char *script, char *destination, char *extend)
{
struct attropl *pal;
char *return_jobid = NULL;
int rc;
struct pbs_client_thread_context *ptr;
struct cred_info *cred_info = NULL;
/* initialize the thread context data, if not already initialized */
if (pbs_client_thread_init_thread_context() != 0)
return return_jobid;
ptr = (struct pbs_client_thread_context *)
pbs_client_thread_get_context_data();
if (!ptr) {
pbs_errno = PBSE_INTERNAL;
return return_jobid;
}
/* first verify the attributes, if verification is enabled */
rc = pbs_verify_attributes(c, PBS_BATCH_QueueJob,
MGR_OBJ_JOB, MGR_CMD_NONE, attrib);
if (rc)
return return_jobid;
/* lock pthread mutex here for this connection */
/* blocking call, waits for mutex release */
if (pbs_client_thread_lock_connection(c) != 0)
return return_jobid;
/* first be sure that the script is readable if specified ... */
if ((script != NULL) && (*script != '\0')) {
if (access(script, R_OK) != 0) {
pbs_errno = PBSE_BADSCRIPT;
if ((connection[c].ch_errtxt = strdup("cannot access script file")) == NULL)
pbs_errno = PBSE_SYSTEM;
goto error;
}
}
/* initiate the queueing of the job */
for (pal = attrib; pal; pal = pal->next)
pal->op = SET; /* force operator to SET */
/* Queue job with null string for job id */
return_jobid = PBSD_queuejob(c, "", destination, attrib, extend, 0, NULL);
if (return_jobid == NULL)
goto error;
/* send script across */
if ((script != NULL) && (*script != '\0')) {
if ((rc = PBSD_jscript(c, script, 0, NULL)) != 0) {
if (rc == PBSE_JOBSCRIPTMAXSIZE)
pbs_errno = rc;
else
pbs_errno = PBSE_BADSCRIPT;
goto error;
}
}
/* OK, the script got across, apparently, so we are */
/* ready to commit */
cred_info = (struct cred_info *) ptr->th_cred_info;
/* opaque information */
if (cred_info && cred_info->cred_len > 0) {
if (PBSD_jcred(c, cred_info->cred_type,
cred_info->cred_buf,
cred_info->cred_len, 0, NULL) != 0) {
pbs_errno = PBSE_BADCRED;
goto error;
}
}
if (PBSD_commit(c, return_jobid, 0, NULL) != 0)
goto error;
/* unlock the thread lock and update the thread context data */
if (pbs_client_thread_unlock_connection(c) != 0)
return NULL;
return return_jobid;
error:
(void)pbs_client_thread_unlock_connection(c);
return NULL;
}
/**
* @brief
* Send execution job on connected rpp stream.
*
* @param[in] jobp - pointer to the job being sent
* @param[in] hostaddr - the address of host to send job to, host byte order
* @param[in] port - the destination port, host byte order
* @param[in] request - The batch request associated with this send job call
*
* @return int
* @retval 2 : success
* @retval -1 : failure (pbs_errno set to error number)
*
*/
int
send_job_exec(job *jobp, pbs_net_t hostaddr, int port, struct batch_request *request)
{
pbs_list_head attrl;
attribute *pattr;
mominfo_t *pmom = NULL;
int stream = -1;
int encode_type;
char *destin = jobp->ji_qs.ji_destin;
int i;
size_t credlen = 0;
char *credbuf = NULL;
char job_id[PBS_MAXSVRJOBID + 1];
struct attropl *pqjatr; /* list (single) of attropl for quejob */
int rc;
int rpp = 1;
char *jobid = NULL;
char *script = NULL;
char *msgid = NULL;
char *dup_msgid = NULL;
struct work_task *ptask = NULL;
/* if job has a script read it from database */
if (jobp->ji_qs.ji_svrflags & JOB_SVFLG_SCRIPT) {
/*
* copy the job script from database to a temp file
* PBSD_jscript works with a file
* delete it at the end of the send
*/
if ((script = svr_load_jobscript(jobp)) == NULL) {
pbs_errno = PBSE_SYSTEM;
snprintf(log_buffer, sizeof(log_buffer),
"Failed to load job script for job %s",
jobp->ji_qs.ji_jobid);
log_err(pbs_errno, "send_job", log_buffer);
goto send_err;
}
}
stream = svr_connect(hostaddr, port, NULL, ToServerDIS, rpp);
if (stream < 0) {
log_event(PBSEVENT_ERROR, PBS_EVENTCLASS_REQUEST, LOG_WARNING, "", "Could not connect to Mom");
goto send_err;
}
pmom = tfind2((unsigned long) jobp->ji_qs.ji_un.ji_exect.ji_momaddr,
jobp->ji_qs.ji_un.ji_exect.ji_momport,
&ipaddrs);
if (!pmom || (((mom_svrinfo_t *)(pmom->mi_data))->msr_state & INUSE_DOWN))
goto send_err;
CLEAR_HEAD(attrl);
resc_access_perm = ATR_DFLAG_MOM;
encode_type = ATR_ENCODE_MOM;
pattr = jobp->ji_wattr;
for (i = 0; i < (int) JOB_ATR_LAST; i++) {
if ((job_attr_def + i)->at_flags & resc_access_perm) {
(void)(job_attr_def + i)->at_encode(pattr + i, &attrl,
(job_attr_def + i)->at_name, (char *) 0, encode_type,
NULL);
}
}
attrl_fixlink(&attrl);
/* save the job id for when after we purge the job */
/* read any credential file */
(void)get_credential(pmom->mi_host, jobp, PBS_GC_BATREQ, &credbuf, &credlen);
(void) strcpy(job_id, jobp->ji_qs.ji_jobid);
pbs_errno = 0;
pqjatr = &((svrattrl *) GET_NEXT(attrl))->al_atopl;
jobid = PBSD_queuejob(stream, jobp->ji_qs.ji_jobid, destin, pqjatr, (char *) 0, rpp, &msgid);
free_attrlist(&attrl);
if (jobid == NULL)
goto send_err;
rpp_add_close_func(stream, process_DreplyRPP); /* register a close handler */
/* adding msgid to deferred list, dont free msgid */
if ((ptask = add_mom_deferred_list(stream, pmom, post_sendmom, msgid, request, jobp)) == NULL)
goto send_err;
/* add to pjob->svrtask list so its automatically cleared when job is purged */
append_link(&jobp->ji_svrtask, &ptask->wt_linkobj, ptask);
/* we cannot use the same msgid, since it is not part of the preq,
* make a dup of it, and we can freely free it
*/
if ((dup_msgid = strdup(msgid)) == NULL)
goto send_err;
/*
* henceforth use the same msgid, since we mean to say all this is
* part of a single logical request to the mom
* and we will be hanging off one request to be answered to finally
*/
if (jobp->ji_qs.ji_svrflags & JOB_SVFLG_SCRIPT) {
if (PBSD_jscript_direct(stream, script, rpp, &dup_msgid) != 0)
goto send_err;
}
free(script);
script = NULL;
if (credlen > 0) {
rc = PBSD_jcred(stream, jobp->ji_extended.ji_ext.ji_credtype, credbuf, credlen, rpp, &dup_msgid);
if (credbuf)
free(credbuf);
if (rc != 0)
goto send_err;
}
if ((jobp->ji_qs.ji_svrflags & JOB_SVFLG_HASRUN)
&& (hostaddr != pbs_server_addr)) {
if ((move_job_file(stream, jobp, StdOut, rpp, &dup_msgid) != 0) ||
(move_job_file(stream, jobp, StdErr, rpp, &dup_msgid) != 0) ||
(move_job_file(stream, jobp, Chkpt, rpp, &dup_msgid) != 0))
goto send_err;
}
if (PBSD_commit(stream, job_id, rpp, &dup_msgid) != 0)
goto send_err;
free(dup_msgid); /* free this as it is not part of any work task */
return 2;
send_err:
if (dup_msgid)
free(dup_msgid);
if (script)
free(script);
if (ptask) {
if (ptask->wt_event2)
free(ptask->wt_event2);
delete_task(ptask);
}
sprintf(log_buffer, "send of job to %s failed error = %d", destin, pbs_errno);
log_event(PBSEVENT_JOB, PBS_EVENTCLASS_JOB, LOG_INFO, jobp->ji_qs.ji_jobid, log_buffer);
return (-1);
}
