static int
_handle_completion(int fd, slurmd_job_t *job, uid_t uid)
{
int rc = SLURM_SUCCESS;
int errnum = 0;
int first;
int last;
jobacctinfo_t *jobacct = NULL;
int step_rc;
debug("_handle_completion for job %u.%u",
job->jobid, job->stepid);
debug3(" uid = %d", uid);
if (!_slurm_authorized_user(uid)) {
debug("step completion message from uid %ld for job %u.%u ",
(long)uid, job->jobid, job->stepid);
rc = -1;
errnum = EPERM;
/* Send the return code and errno */
safe_write(fd, &rc, sizeof(int));
safe_write(fd, &errnum, sizeof(int));
return SLURM_SUCCESS;
}
safe_read(fd, &first, sizeof(int));
safe_read(fd, &last, sizeof(int));
safe_read(fd, &step_rc, sizeof(int));
jobacct = jobacct_gather_g_create(NULL);
jobacct_gather_g_getinfo(jobacct, JOBACCT_DATA_PIPE, &fd);
/*
* Record the completed nodes
*/
pthread_mutex_lock(&step_complete.lock);
if (! step_complete.wait_children) {
rc = -1;
errnum = ETIMEDOUT; /* not used anyway */
goto timeout;
}
/* SlurmUser or root can craft a launch without a valid credential
* ("srun --no-alloc ...") and no tree information can be built
* without the hostlist from the credential. */
if (step_complete.rank >= 0) {
#if 0
char bits_string[128];
debug2("Setting range %d (bit %d) through %d(bit %d)",
first, first-(step_complete.rank+1),
last, last-(step_complete.rank+1));
bit_fmt(bits_string, sizeof(bits_string), step_complete.bits);
debug2(" before bits: %s", bits_string);
#endif
bit_nset(step_complete.bits,
first - (step_complete.rank+1),
last - (step_complete.rank+1));
#if 0
bit_fmt(bits_string, sizeof(bits_string), step_complete.bits);
debug2(" after bits: %s", bits_string);
#endif
}
step_complete.step_rc = MAX(step_complete.step_rc, step_rc);
/************* acct stuff ********************/
jobacct_gather_g_aggregate(step_complete.jobacct, jobacct);
timeout:
jobacct_gather_g_destroy(jobacct);
/*********************************************/
/* Send the return code and errno, we do this within the locked
* region to ensure that the stepd doesn't exit before we can
* perform this send. */
safe_write(fd, &rc, sizeof(int));
safe_write(fd, &errnum, sizeof(int));
pthread_cond_signal(&step_complete.cond);
pthread_mutex_unlock(&step_complete.lock);
return SLURM_SUCCESS;
rwfail:
return SLURM_FAILURE;
}
static int
_handle_completion(int fd, slurmd_job_t *job, uid_t uid, int protocol)
{
int rc = SLURM_SUCCESS;
int errnum = 0;
int first;
int last;
jobacctinfo_t *jobacct = NULL;
int step_rc;
char* buf;
int len;
Buf buffer;
int version; /* For future use */
bool lock_set = false;
debug("_handle_completion for job %u.%u",
job->jobid, job->stepid);
debug3(" uid = %d", uid);
if (!_slurm_authorized_user(uid)) {
debug("step completion message from uid %ld for job %u.%u ",
(long)uid, job->jobid, job->stepid);
rc = -1;
errnum = EPERM;
/* Send the return code and errno */
safe_write(fd, &rc, sizeof(int));
safe_write(fd, &errnum, sizeof(int));
return SLURM_SUCCESS;
}
if (protocol >= 2)
safe_read(fd, &version, sizeof(int));
safe_read(fd, &first, sizeof(int));
safe_read(fd, &last, sizeof(int));
safe_read(fd, &step_rc, sizeof(int));
if (protocol >= 2) {
/*
* We must not use getinfo over a pipe with slurmd here
* Indeed, slurmstepd does a large use of setinfo over a pipe
* with slurmd and doing the reverse can result in a deadlock
* scenario with slurmd :
* slurmd(lockforread,write)/slurmstepd(write,lockforread)
* Do pack/unpack instead to be sure of independances of
* slurmd and slurmstepd
*/
safe_read(fd, &len, sizeof(int));
buf = xmalloc(len);
safe_read(fd, buf, len);
buffer = create_buf(buf, len);
jobacct_gather_g_unpack(&jobacct, SLURM_PROTOCOL_VERSION,
buffer);
free_buf(buffer);
} else {
jobacct = jobacct_gather_g_create(NULL);
jobacct_gather_g_getinfo(jobacct, JOBACCT_DATA_PIPE, &fd);
}
/*
* Record the completed nodes
*/
pthread_mutex_lock(&step_complete.lock);
lock_set = true;
if (! step_complete.wait_children) {
rc = -1;
errnum = ETIMEDOUT; /* not used anyway */
goto timeout;
}
/* SlurmUser or root can craft a launch without a valid credential
* ("srun --no-alloc ...") and no tree information can be built
* without the hostlist from the credential. */
if (step_complete.rank >= 0) {
#if 0
char bits_string[128];
debug2("Setting range %d (bit %d) through %d(bit %d)",
first, first-(step_complete.rank+1),
last, last-(step_complete.rank+1));
bit_fmt(bits_string, sizeof(bits_string), step_complete.bits);
debug2(" before bits: %s", bits_string);
#endif
bit_nset(step_complete.bits,
first - (step_complete.rank+1),
last - (step_complete.rank+1));
#if 0
bit_fmt(bits_string, sizeof(bits_string), step_complete.bits);
debug2(" after bits: %s", bits_string);
#endif
}
step_complete.step_rc = MAX(step_complete.step_rc, step_rc);
/************* acct stuff ********************/
jobacct_gather_g_aggregate(step_complete.jobacct, jobacct);
timeout:
jobacct_gather_g_destroy(jobacct);
/*********************************************/
/* Send the return code and errno, we do this within the locked
* region to ensure that the stepd doesn't exit before we can
* perform this send. */
safe_write(fd, &rc, sizeof(int));
safe_write(fd, &errnum, sizeof(int));
pthread_cond_signal(&step_complete.cond);
pthread_mutex_unlock(&step_complete.lock);
return SLURM_SUCCESS;
rwfail: if (lock_set) {
pthread_cond_signal(&step_complete.cond);
pthread_mutex_unlock(&step_complete.lock);
}
return SLURM_FAILURE;
}
