static mca_mtl_base_module_t *
ompi_mtl_psm2_component_init(bool enable_progress_threads,
bool enable_mpi_threads)
{
psm_error_t err;
int verno_major = PSM_VERNO_MAJOR;
int verno_minor = PSM_VERNO_MINOR;
int local_rank = -1, num_local_procs = 0;
int num_total_procs = 0;
/* Compute the total number of processes on this host and our local rank
* on that node. We need to provide PSM with these values so it can
* allocate hardware contexts appropriately across processes.
*/
if (OMPI_SUCCESS != get_num_local_procs(&num_local_procs)) {
opal_output(0, "Cannot determine number of local processes. "
"Cannot continue.\n");
return NULL;
}
if (OMPI_SUCCESS != get_local_rank(&local_rank)) {
opal_output(0, "Cannot determine local rank. Cannot continue.\n");
return NULL;
}
if (OMPI_SUCCESS != get_num_total_procs(&num_total_procs)) {
opal_output(0, "Cannot determine total number of processes. "
"Cannot continue.\n");
return NULL;
}
err = psm_error_register_handler(NULL /* no ep */,
PSM_ERRHANDLER_NOP);
if (err) {
opal_output(0, "Error in psm_error_register_handler (error %s)\n",
psm_error_get_string(err));
return NULL;
}
if (num_local_procs == num_total_procs) {
setenv("PSM_DEVICES", "self,shm", 0);
}
err = psm_init(&verno_major, &verno_minor);
if (err) {
opal_show_help("help-mtl-psm.txt",
"psm init", true,
psm_error_get_string(err));
return NULL;
}
/* Complete PSM initialization */
ompi_mtl_psm2_module_init(local_rank, num_local_procs);
ompi_mtl_psm2.super.mtl_request_size =
sizeof(mca_mtl_psm2_request_t) -
sizeof(struct mca_mtl_request_t);
return &ompi_mtl_psm2.super;
}
static int psmx_init_lib(void)
{
int major, minor;
int ret = 0, err;
if (psmx_lib_initialized)
return 0;
pthread_mutex_lock(&psmx_lib_mutex);
if (psmx_lib_initialized)
goto out;
psm_error_register_handler(NULL, PSM_ERRHANDLER_NO_HANDLER);
major = PSM_VERNO_MAJOR;
minor = PSM_VERNO_MINOR;
err = psm_init(&major, &minor);
if (err != PSM_OK) {
FI_WARN(&psmx_prov, FI_LOG_CORE,
"psm_init failed: %s\n", psm_error_get_string(err));
ret = err;
goto out;
}
FI_INFO(&psmx_prov, FI_LOG_CORE,
"PSM header version = (%d, %d)\n", PSM_VERNO_MAJOR, PSM_VERNO_MINOR);
FI_INFO(&psmx_prov, FI_LOG_CORE,
"PSM library version = (%d, %d)\n", major, minor);
if (major != PSM_VERNO_MAJOR) {
psmx_am_compat_mode = 1;
FI_INFO(&psmx_prov, FI_LOG_CORE,
"PSM AM compat mode enabled: appliation %d.%d, library %d.%d.\n",
PSM_VERNO_MAJOR, PSM_VERNO_MINOR, major, minor);
}
if (major > 1) {
psmx_compat_lib = 1;
FI_INFO(&psmx_prov, FI_LOG_CORE,
"PSM is supported via the psm2-compat library over PSM2.\n");
}
psmx_lib_initialized = 1;
out:
pthread_mutex_unlock(&psmx_lib_mutex);
return ret;
}
int initialize(void)
{
int verno_major = PSM_VERNO_MAJOR;
int verno_minor = PSM_VERNO_MINOR;
psm_ep_errhandler_t handler;
int err;
//err = psm_error_register_handler(NULL, // Global handler
// PSM_ERRHANDLER_NO_HANDLER); // return errors
if (err) {
fprintf(stderr, "t register global handler: %s\n",
psm_error_get_string(err));
return -1;
}
err = psm_init(&verno_major, &verno_minor);
if (err || verno_major > PSM_VERNO_MAJOR) {
if (err)
fprintf(stderr, "PSM initialization failure: %s\n",
psm_error_get_string(err));
else
fprintf(stderr, "PSM loaded an unexpected/unsupported "
"version (%d.%d)\n", verno_major, verno_minor);
return -1;
}
// We were able to initialize PSM but will defer all further error
// handling since most of the errors beyond this point will be fatal.
handler = error_handler;
err = psm_error_register_handler(NULL, // Global handler
// PSM_ERRHANDLER_PSM_HANDLER);
handler);
if (err) {
fprintf(stderr, "t register global errhandler: %s\n",
psm_error_get_string(err));
return -1;
}
return 0;
}
static
int pspsm_init(void)
{
static pspsm_init_state_t init_state = PSPSM_INIT_START;
int verno_minor = PSM_VERNO_MINOR;
int verno_major = PSM_VERNO_MAJOR;
psm_error_t ret;
if (init_state == PSPSM_INIT_START) {
/* Check for an available /dev/ipath */
ret = pspsm_check_dev_ipath();
if (ret != 0) {
goto err_dev_ipath;
}
ret = psm_init(&verno_major, &verno_minor);
if (ret != PSM_OK) {
goto err_init;
}
/*
* All processes wanting to communicate need to use
* the same UUID.
*
* It is unclear whether there are drawbacks from
* simply using the same UUID for groups of processes
* that will never communicate.
*
* On top of a constant fill pattern, we use:
*
* - PSP_PSM_UNIQ_ID if set and not zero, or
* - PMI_ID, if set and not zero - that's not entirely
* clean, but a practical solution for MPI apps (as
* long as we do not implement communication between
* two sets of MPI processes not sharing a
* communicator).
*/
memset(pspsm_uuid.as_uuid, DEFAULT_UUID_PATTERN,
sizeof(pspsm_uuid.as_uuid));
if (pscom.env.psm_uniq_id) {
pspsm_dprint(2, "seeding PSM UUID with %u", pscom.env.psm_uniq_id);
pspsm_uuid.as_uint = pscom.env.psm_uniq_id;
}
/* Open the endpoint here in init with the hope that
every mpi rank call indirect psm_ep_open() before
transmitting any data from or to this endpoint.
This is to avoid a race condition in
libpsm_infinipath. Downside: We consume PSM
Contexts even in the case of only local
communication. You could use PSP_PSM=0 in this
case.
*/
if (pspsm_open_endpoint()) goto err_ep;
if (pspsm_init_mq()) goto err_mq;
pspsm_dprint(2, "pspsm_init: OK");
init_state = PSPSM_INIT_DONE;
}
return init_state; /* 0 = success, -1 = error */
err_dev_ipath:
pspsm_dprint(2, "pspsm_init: No \"/dev/ipath\" found. Arch psm is disabled.");
goto err_exit;
err_init:
pspsm_err(psm_error_get_string(ret));
pspsm_dprint(1, "pspsm_init: %s", pspsm_err_str);
// Fall through
err_ep:
err_mq:
err_exit:
init_state = PSPSM_INIT_FAILED;
return init_state; /* 0 = success, -1 = error */
}
static mca_mtl_base_module_t *
ompi_mtl_psm_component_init(bool enable_progress_threads,
bool enable_mpi_threads)
{
psm_error_t err;
int verno_major = PSM_VERNO_MAJOR;
int verno_minor = PSM_VERNO_MINOR;
int local_rank = -1, num_local_procs = 0;
int num_total_procs = 0;
/* Compute the total number of processes on this host and our local rank
* on that node. We need to provide PSM with these values so it can
* allocate hardware contexts appropriately across processes.
*/
if (OMPI_SUCCESS != get_num_local_procs(&num_local_procs)) {
opal_output(0, "Cannot determine number of local processes. "
"Cannot continue.\n");
return NULL;
}
if (OMPI_SUCCESS != get_local_rank(&local_rank)) {
opal_output(0, "Cannot determine local rank. Cannot continue.\n");
return NULL;
}
if (OMPI_SUCCESS != get_num_total_procs(&num_total_procs)) {
opal_output(0, "Cannot determine total number of processes. "
"Cannot continue.\n");
return NULL;
}
#if PSM_VERNO >= 0x010c
/* Set infinipath debug level */
err = psm_setopt(PSM_COMPONENT_CORE, 0, PSM_CORE_OPT_DEBUG,
(const void*) &ompi_mtl_psm.debug_level,
sizeof(unsigned));
if (err) {
/* Non fatal error. Can continue */
opal_show_help("help-mtl-psm.txt",
"psm init", false,
psm_error_get_string(err));
}
#endif
if (getenv("PSM_DEVICES") == NULL) {
/* Only allow for shm and ipath devices in 2.0 and earlier releases
* (unless the user overrides the setting).
*/
if (PSM_VERNO >= 0x0104) {
if (num_local_procs == num_total_procs) {
setenv("PSM_DEVICES", "self,shm", 0);
} else {
setenv("PSM_DEVICES", "self,shm,ipath", 0);
}
}
else {
if (num_local_procs == num_total_procs) {
setenv("PSM_DEVICES", "shm", 0);
} else {
setenv("PSM_DEVICES", "shm,ipath", 0);
}
}
}
err = psm_init(&verno_major, &verno_minor);
if (err) {
opal_show_help("help-mtl-psm.txt",
"psm init", true,
psm_error_get_string(err));
return NULL;
}
/* Complete PSM initialization */
ompi_mtl_psm_module_init(local_rank, num_local_procs);
ompi_mtl_psm.super.mtl_request_size =
sizeof(mca_mtl_psm_request_t) -
sizeof(struct mca_mtl_request_t);
/* don't register the err handler until we know we will be active */
err = psm_error_register_handler(NULL /* no ep */,
PSM_ERRHANDLER_NOP);
if (err) {
opal_output(0, "Error in psm_error_register_handler (error %s)\n",
psm_error_get_string(err));
return NULL;
}
return &ompi_mtl_psm.super;
}
static mca_mtl_base_module_t*
ompi_mtl_psm_component_init(bool enable_progress_threads,
bool enable_mpi_threads)
{
psm_error_t err;
int rc;
int verno_major = PSM_VERNO_MAJOR;
int verno_minor = PSM_VERNO_MINOR;
ompi_proc_t *my_proc, **procs;
size_t num_total_procs, proc;
int local_rank = -1, num_local_procs = 0;
/* Compute the total number of processes on this host and our local rank
* on that node. We need to provide PSM with these values so it can
* allocate hardware contexts appropriately across processes.
*/
if ((rc = ompi_proc_refresh()) != OMPI_SUCCESS) {
return NULL;
}
my_proc = ompi_proc_local();
if (NULL == (procs = ompi_proc_world(&num_total_procs))) {
return NULL;
}
for (proc = 0; proc < num_total_procs; proc++) {
if (my_proc == procs[proc]) {
local_rank = num_local_procs++;
continue;
}
if (OPAL_PROC_ON_LOCAL_NODE(procs[proc]->proc_flags)) {
num_local_procs++;
}
}
assert(local_rank >= 0 && num_local_procs > 0);
free(procs);
err = psm_error_register_handler(NULL /* no ep */,
PSM_ERRHANDLER_NOP);
if (err) {
opal_output(0, "Error in psm_error_register_handler (error %s)\n",
psm_error_get_string(err));
return NULL;
}
#if PSM_VERNO >= 0x010c
/* Set infinipath debug level */
err = psm_setopt(PSM_COMPONENT_CORE, 0, PSM_CORE_OPT_DEBUG,
(const void*) &ompi_mtl_psm.debug_level,
sizeof(unsigned));
if (err) {
/* Non fatal error. Can continue */
orte_show_help("help-mtl-psm.txt",
"psm init", false,
psm_error_get_string(err));
}
#endif
/* Only allow for shm and ipath devices in 2.0 and earlier releases
* (unless the user overrides the setting).
*/
if (PSM_VERNO >= 0x0104) {
setenv("PSM_DEVICES", "self,shm,ipath", 0);
}
else {
setenv("PSM_DEVICES", "shm,ipath", 0);
}
err = psm_init(&verno_major, &verno_minor);
if (err) {
orte_show_help("help-mtl-psm.txt",
"psm init", true,
psm_error_get_string(err));
return NULL;
}
/* Complete PSM initialization */
ompi_mtl_psm_module_init(local_rank, num_local_procs);
ompi_mtl_psm.super.mtl_request_size =
sizeof(mca_mtl_psm_request_t) -
sizeof(struct mca_mtl_request_t);
return &ompi_mtl_psm.super;
}
