/* ////////////////////////////////////////////////////////////////////////// */
static mca_common_sm_module_t *
attach_and_init(opal_shmem_ds_t *shmem_bufp,
size_t size,
size_t size_ctl_structure,
size_t data_seg_alignment,
bool first_call)
{
mca_common_sm_module_t *map = NULL;
mca_common_sm_seg_header_t *seg = NULL;
unsigned char *addr = NULL;
/* attach to the specified segment. note that at this point, the contents of
* *shmem_bufp have already been initialized via opal_shmem_segment_create.
*/
if (NULL == (seg = (mca_common_sm_seg_header_t *)
opal_shmem_segment_attach(shmem_bufp))) {
return NULL;
}
opal_atomic_rmb();
if (NULL == (map = OBJ_NEW(mca_common_sm_module_t))) {
OPAL_ERROR_LOG(OPAL_ERR_OUT_OF_RESOURCE);
(void)opal_shmem_segment_detach(shmem_bufp);
return NULL;
}
/* copy meta information into common sm module
* from ====> to */
if (OPAL_SUCCESS != opal_shmem_ds_copy(shmem_bufp, &map->shmem_ds)) {
(void)opal_shmem_segment_detach(shmem_bufp);
free(map);
return NULL;
}
/* the first entry in the file is the control structure. the first
* entry in the control structure is an mca_common_sm_seg_header_t
* element.
*/
map->module_seg = seg;
addr = ((unsigned char *)seg) + size_ctl_structure;
/* if we have a data segment (i.e., if 0 != data_seg_alignment),
* then make it the first aligned address after the control
* structure. IF THIS HAPPENS, THIS IS A PROGRAMMING ERROR IN
* OPEN MPI!
*/
if (0 != data_seg_alignment) {
addr = OPAL_ALIGN_PTR(addr, data_seg_alignment, unsigned char *);
/* is addr past end of the shared memory segment? */
if ((unsigned char *)seg + shmem_bufp->seg_size < addr) {
opal_show_help("help-mpi-common-sm.txt", "mmap too small", 1,
opal_proc_local_get()->proc_hostname,
(unsigned long)shmem_bufp->seg_size,
(unsigned long)size_ctl_structure,
(unsigned long)data_seg_alignment);
(void)opal_shmem_segment_detach(shmem_bufp);
free(map);
return NULL;
}
}
/**
* this routine assumes that sorted_procs is in the following state:
* o all the local procs at the beginning.
* o sorted_procs[0] is the lowest named process.
*/
int
mca_common_sm_rml_info_bcast(opal_shmem_ds_t *ds_buf,
ompi_proc_t **procs,
size_t num_procs,
int tag,
bool bcast_root,
char *msg_id_str,
opal_list_t *pending_rml_msgs)
{
int rc = OMPI_SUCCESS;
struct iovec iov[MCA_COMMON_SM_RML_MSG_LEN];
int iovrc;
size_t p;
char msg_id_str_to_tx[OPAL_PATH_MAX];
strncpy(msg_id_str_to_tx, msg_id_str, sizeof(msg_id_str_to_tx) - 1);
/* let the first item be the queueing id name */
iov[0].iov_base = (ompi_iov_base_ptr_t)msg_id_str_to_tx;
iov[0].iov_len = sizeof(msg_id_str_to_tx);
iov[1].iov_base = (ompi_iov_base_ptr_t)ds_buf;
iov[1].iov_len = sizeof(opal_shmem_ds_t);
/* figure out if i am the root proc in the group.
* if i am, bcast the message the rest of the local procs.
*/
if (bcast_root) {
opal_progress_event_users_increment();
/* first num_procs items should be local procs */
for (p = 1; p < num_procs; ++p) {
iovrc = orte_rml.send(&(procs[p]->proc_name), iov,
MCA_COMMON_SM_RML_MSG_LEN, tag, 0);
if ((ssize_t)(iov[0].iov_len + iov[1].iov_len) > iovrc) {
ORTE_ERROR_LOG(OMPI_ERR_COMM_FAILURE);
opal_progress_event_users_decrement();
rc = OMPI_ERROR;
goto out;
}
}
opal_progress_event_users_decrement();
}
else { /* i am NOT the root ("lowest") proc */
opal_list_item_t *item;
mca_common_sm_rml_pending_rml_msg_types_t *rml_msg;
/* because a component query can be performed simultaneously in multiple
* threads, the RML messages may arrive in any order. so first check to
* see if we previously received a message for me.
*/
for (item = opal_list_get_first(pending_rml_msgs);
opal_list_get_end(pending_rml_msgs) != item;
item = opal_list_get_next(item)) {
rml_msg = (mca_common_sm_rml_pending_rml_msg_types_t *)item;
/* was the message for me? */
if (0 == strcmp(rml_msg->msg_id_str, msg_id_str)) {
opal_list_remove_item(pending_rml_msgs, item);
/* from ==============> to */
opal_shmem_ds_copy(&rml_msg->shmem_ds, ds_buf);
OBJ_RELEASE(item);
break;
}
}
/* if we didn't find a message already waiting, block on receiving from
* the RML.
*/
if (opal_list_get_end(pending_rml_msgs) == item) {
do {
/* bump up the libevent polling frequency while we're in this
* RML recv, just to ensure we're checking libevent frequently.
*/
opal_progress_event_users_increment();
iovrc = orte_rml.recv(&(procs[0]->proc_name), iov,
MCA_COMMON_SM_RML_MSG_LEN, tag, 0);
opal_progress_event_users_decrement();
if (iovrc < 0) {
ORTE_ERROR_LOG(OMPI_ERR_RECV_LESS_THAN_POSTED);
rc = OMPI_ERROR;
goto out;
}
/* was the message for me? if so, we're done */
if (0 == strcmp(msg_id_str_to_tx, msg_id_str)) {
break;
}
/* if not, put it on the pending list and try again */
if (NULL == (rml_msg =
OBJ_NEW(mca_common_sm_rml_pending_rml_msg_types_t)))
{
ORTE_ERROR_LOG(OMPI_ERR_OUT_OF_RESOURCE);
rc = OMPI_ERROR;
goto out;
}
/* not for me, so place on list */
/* from ========> to */
opal_shmem_ds_copy(ds_buf, &rml_msg->shmem_ds);
memcpy(rml_msg->msg_id_str, msg_id_str_to_tx, OPAL_PATH_MAX);
opal_list_append(pending_rml_msgs, &(rml_msg->super));
} while(1);
}
}
out:
return rc;
}
