int pmixp_libpmix_init(void)
{
int rc;
mode_t rights = (S_IRUSR | S_IWUSR | S_IXUSR) | (S_IRGRP | S_IWGRP | S_IXGRP);
pmix_info_t *kvp;
/* NOTE: we need user who owns the job to access PMIx usock
* file. According to 'man 7 unix':
* "... In the Linux implementation, sockets which are visible in the file system
* honor the permissions of the directory they are in... "
* Our case is the following: slurmstepd is usually running as root, user application will
* be "sudo'ed". To provide both of them with acces to the unix socket we do the following:
* 1. Owner ID is set to the job owner.
* 2. Group ID corresponds to slurmstepd.
* 3. Set 0770 access mode */
if (0 != mkdir(pmixp_info_tmpdir_lib(), rights) ) {
PMIXP_ERROR_STD("Cannot create directory \"%s\"", pmixp_info_tmpdir_lib());
return errno;
}
/* There might be umask that will drop essential rights. Fix it explicitly.
* TODO: is there more elegant solution? */
if (chmod(pmixp_info_tmpdir_lib(), rights) < 0) {
error("chown(%s): %m", pmixp_info_tmpdir_lib());
return errno;
}
if (chown(pmixp_info_tmpdir_lib(), (uid_t) pmixp_info_jobuid(), (gid_t) -1) < 0) {
error("chown(%s): %m", pmixp_info_tmpdir_lib());
return errno;
}
setenv(PMIXP_PMIXLIB_TMPDIR, pmixp_info_tmpdir_lib(), 1);
PMIXP_ALLOC_KEY(kvp, PMIX_USERID);
PMIX_VAL_SET(&kvp->value, uint32_t, pmixp_info_jobuid());
/* setup the server library */
if (PMIX_SUCCESS != (rc = PMIx_server_init(&_slurm_pmix_cb, kvp, 1))) {
PMIXP_ERROR_STD("PMIx_server_init failed with error %d\n", rc);
return SLURM_ERROR;
}
PMIXP_FREE_KEY(kvp);
/*
if( pmixp_fixrights(pmixp_info_tmpdir_lib(),
(uid_t) pmixp_info_jobuid(), rights) ){
}
*/
/* register the errhandler */
PMIx_Register_errhandler(NULL, 0, errhandler, errhandler_reg_callbk,
NULL);
return 0;
}
int main(int argc, char **argv)
{
int rc;
pmix_value_t value;
pmix_value_t *val = &value;
pmix_proc_t proc;
uint32_t nprocs;
/* init us */
if (PMIX_SUCCESS != (rc = PMIx_Init(&myproc))) {
pmix_output(0, "Client ns %s rank %d: PMIx_Init failed: %d", myproc.nspace, myproc.rank, rc);
exit(0);
}
pmix_output(0, "Client ns %s rank %d: Running", myproc.nspace, myproc.rank);
/* get our universe size */
if (PMIX_SUCCESS != (rc = PMIx_Get(&myproc, PMIX_UNIV_SIZE, NULL, 0, &val))) {
pmix_output(0, "Client ns %s rank %d: PMIx_Get universe size failed: %d", myproc.nspace, myproc.rank, rc);
goto done;
}
nprocs = val->data.uint32;
PMIX_VALUE_RELEASE(val);
pmix_output(0, "Client %s:%d universe size %d", myproc.nspace, myproc.rank, nprocs);
completed = false;
/* register our errhandler */
PMIx_Register_errhandler(NULL, 0, notification_fn, errhandler_reg_callbk, NULL);
/* call fence to sync */
PMIX_PROC_CONSTRUCT(&proc);
(void)strncpy(proc.nspace, myproc.nspace, PMIX_MAX_NSLEN);
proc.rank = PMIX_RANK_WILDCARD;
if (PMIX_SUCCESS != (rc = PMIx_Fence(&proc, 1, NULL, 0))) {
pmix_output(0, "Client ns %s rank %d: PMIx_Fence failed: %d", myproc.nspace, myproc.rank, rc);
goto done;
}
/* rank=0 calls abort */
if (0 == myproc.rank) {
PMIx_Abort(PMIX_ERR_OUT_OF_RESOURCE, "Eat rocks",
&proc, 1);
pmix_output(0, "Client ns %s rank %d: Abort called", myproc.nspace, myproc.rank);
}
/* everyone simply waits */
while (!completed) {
struct timespec ts;
ts.tv_sec = 0;
ts.tv_nsec = 100000;
nanosleep(&ts, NULL);
}
done:
/* finalize us */
pmix_output(0, "Client ns %s rank %d: Finalizing", myproc.nspace, myproc.rank);
PMIx_Deregister_errhandler(0, op_callbk, NULL);
if (PMIX_SUCCESS != (rc = PMIx_Finalize())) {
fprintf(stderr, "Client ns %s rank %d:PMIx_Finalize failed: %d\n", myproc.nspace, myproc.rank, rc);
} else {
fprintf(stderr, "Client ns %s rank %d:PMIx_Finalize successfully completed\n", myproc.nspace, myproc.rank);
}
fflush(stderr);
return(0);
}
