static int _serv_read(eio_obj_t *obj, List objs)
{
PMIXP_DEBUG("fd = %d", obj->fd);
pmixp_io_engine_t *me = (pmixp_io_engine_t *)obj->arg;
bool proceed = true;
pmixp_debug_hang(0);
/* Read and process all received messages */
while (proceed) {
switch( _process_message(me) ){
case 2:
obj->shutdown = true;
PMIXP_DEBUG("Connection finalized fd = %d", obj->fd);
/* cleanup after this connection */
eio_remove_obj(obj, objs);
xfree(me);
case 0:
proceed = 0;
case 1:
break;
}
}
return 0;
}
static int _serv_read(eio_obj_t *obj, List objs)
{
/* sanity check */
xassert(NULL != obj );
if (obj->shutdown) {
/* corresponding connection will be
* cleaned up during plugin finalize
*/
return 0;
}
PMIXP_DEBUG("fd = %d", obj->fd);
pmixp_conn_t *conn = (pmixp_conn_t *)obj->arg;
bool proceed = true;
/* debug stub */
pmixp_debug_hang(0);
/* Read and process all received messages */
while (proceed) {
if (!pmixp_conn_progress_rcv(conn)) {
proceed = 0;
}
if (!pmixp_conn_is_alive(conn)) {
obj->shutdown = true;
PMIXP_DEBUG("Connection closed fd = %d", obj->fd);
/* cleanup after this connection */
eio_remove_obj(obj, objs);
pmixp_conn_return(conn);
proceed = 0;
}
}
return 0;
}
static int _serv_write(eio_obj_t *obj, List objs)
{
/* sanity check */
xassert(NULL != obj );
if (obj->shutdown) {
/* corresponding connection will be
* cleaned up during plugin finalize
*/
return 0;
}
PMIXP_DEBUG("fd = %d", obj->fd);
pmixp_conn_t *conn = (pmixp_conn_t *)obj->arg;
/* debug stub */
pmixp_debug_hang(0);
/* progress sends */
pmixp_conn_progress_snd(conn);
/* if we are done with this connection - remove it */
if (!pmixp_conn_is_alive(conn)) {
obj->shutdown = true;
PMIXP_DEBUG("Connection finalized fd = %d", obj->fd);
/* cleanup after this connection */
eio_remove_obj(obj, objs);
pmixp_conn_return(conn);
}
return 0;
}
static void _dfwd_sent_cb(int rc, pmixp_p2p_ctx_t ctx, void *_vcbdata)
{
pmixp_coll_cbdata_t *cbdata = (pmixp_coll_cbdata_t*)_vcbdata;
pmixp_coll_t *coll = cbdata->coll;
if( PMIXP_P2P_REGULAR == ctx ){
/* lock the collective */
slurm_mutex_lock(&coll->lock);
}
if (cbdata->seq != coll->seq) {
/* it seems like this collective was reset since the time
* we initiated this send.
* Just exit to avoid data corruption.
*/
PMIXP_DEBUG("Collective was reset!");
goto exit;
}
xassert(PMIXP_COLL_DOWNFWD == coll->state);
/* Change the status */
if( SLURM_SUCCESS == rc ){
coll->dfwd_cb_cnt++;
} else {
coll->dfwd_status = PMIXP_COLL_SND_FAILED;
}
#ifdef PMIXP_COLL_DEBUG
PMIXP_DEBUG("%p: state: %s, snd_status=%s, compl_cnt=%d/%d",
coll, pmixp_coll_state2str(coll->state),
pmixp_coll_sndstatus2str(coll->dfwd_status),
coll->dfwd_cb_cnt, coll->dfwd_cb_wait);
#endif
exit:
xassert(0 < cbdata->refcntr);
cbdata->refcntr--;
if (!cbdata->refcntr) {
xfree(cbdata);
}
if( PMIXP_P2P_REGULAR == ctx ){
/* progress, in the inline case progress
* will be invoked by the caller */
_progress_coll(coll);
/* unlock the collective */
slurm_mutex_unlock(&coll->lock);
}
}
static pmix_status_t lookup_fn(const pmix_proc_t *proc, char **keys,
const pmix_info_t info[], size_t ninfo,
pmix_lookup_cbfunc_t cbfunc, void *cbdata)
{
PMIXP_DEBUG("called");
return PMIX_ERR_NOT_IMPLEMENTED;
}
static pmix_status_t _disconnect_fn(const pmix_proc_t procs[], size_t nprocs,
const pmix_info_t info[], size_t ninfo,
pmix_op_cbfunc_t cbfunc, void *cbdata)
{
PMIXP_DEBUG("called");
return PMIX_ERR_NOT_SUPPORTED;
}
void pmixp_server_direct_conn(int fd)
{
eio_obj_t *obj;
pmixp_conn_t *conn;
PMIXP_DEBUG("Request from fd = %d", fd);
/* Set nonblocking */
fd_set_nonblocking(fd);
fd_set_close_on_exec(fd);
pmixp_fd_set_nodelay(fd);
conn = pmixp_conn_new_temp(PMIXP_PROTO_DIRECT, fd,
_direct_conn_establish);
/* try to process right here */
pmixp_conn_progress_rcv(conn);
if (!pmixp_conn_is_alive(conn)) {
/* success, don't need this connection anymore */
pmixp_conn_return(conn);
return;
}
/* If it is a blocking operation: create AIO object to
* handle it */
obj = eio_obj_create(fd, &direct_peer_ops, (void *)conn);
eio_new_obj(pmixp_info_io(), obj);
/* wakeup this connection to get processed */
eio_signal_wakeup(pmixp_info_io());
}
/*
* TODO: we need to keep track of the "me"
* structures created here, because we need to
* free them in "pmixp_stepd_finalize"
*/
void pmixp_server_slurm_conn(int fd)
{
eio_obj_t *obj;
pmixp_conn_t *conn = NULL;
PMIXP_DEBUG("Request from fd = %d", fd);
pmixp_debug_hang(0);
/* Set nonblocking */
fd_set_nonblocking(fd);
fd_set_close_on_exec(fd);
conn = pmixp_conn_new_temp(PMIXP_PROTO_SLURM, fd, _slurm_new_msg);
/* try to process right here */
pmixp_conn_progress_rcv(conn);
if (!pmixp_conn_is_alive(conn)) {
/* success, don't need this connection anymore */
pmixp_conn_return(conn);
return;
}
/* If it is a blocking operation: create AIO object to
* handle it */
obj = eio_obj_create(fd, &slurm_peer_ops, (void *)conn);
eio_new_obj(pmixp_info_io(), obj);
}
int pmixp_coll_ring_init(pmixp_coll_t *coll, hostlist_t *hl)
{
#ifdef PMIXP_COLL_DEBUG
PMIXP_DEBUG("called");
#endif
int i;
pmixp_coll_ring_ctx_t *coll_ctx = NULL;
pmixp_coll_ring_t *ring = &coll->state.ring;
char *p;
int rel_id = hostlist_find(*hl, pmixp_info_hostname());
/* compute the next absolute id of the neighbor */
p = hostlist_nth(*hl, (rel_id + 1) % coll->peers_cnt);
ring->next_peerid = pmixp_info_job_hostid(p);
free(p);
ring->fwrd_buf_pool = list_create(pmixp_free_buf);
ring->ring_buf_pool = list_create(pmixp_free_buf);
for (i = 0; i < PMIXP_COLL_RING_CTX_NUM; i++) {
coll_ctx = &ring->ctx_array[i];
coll_ctx->coll = coll;
coll_ctx->in_use = false;
coll_ctx->seq = coll->seq;
coll_ctx->contrib_local = false;
coll_ctx->contrib_prev = 0;
coll_ctx->state = PMIXP_COLL_RING_SYNC;
// TODO bit vector
coll_ctx->contrib_map = xmalloc(sizeof(bool) * coll->peers_cnt);
}
return SLURM_SUCCESS;
}
int pmixp_stepd_finalize(void)
{
char *path;
if (!_was_initialized) {
/* nothing to do */
return 0;
}
pmixp_libpmix_finalize();
pmixp_dmdx_finalize();
pmixp_conn_fini();
pmixp_dconn_fini();
pmixp_state_finalize();
pmixp_nspaces_finalize();
/* cleanup the UNIX socket */
PMIXP_DEBUG("Remove PMIx plugin usock");
close(pmixp_info_srv_usock_fd());
path = pmixp_info_nspace_usock(pmixp_info_namespace());
unlink(path);
xfree(path);
/* free the information */
pmixp_info_free();
return SLURM_SUCCESS;
}
int pmixp_coll_contrib_local(pmixp_coll_t *coll, pmixp_coll_type_t type,
char *data, size_t ndata,
void *cbfunc, void *cbdata) {
int ret = SLURM_SUCCESS;
#ifdef PMIXP_COLL_DEBUG
PMIXP_DEBUG("%p: %s seq=%d, size=%lu", coll, pmixp_coll_type2str(type),
coll->seq, ndata);
#endif
switch (type) {
case PMIXP_COLL_TYPE_FENCE_TREE:
ret = pmixp_coll_tree_local(coll, data, ndata,
cbfunc, cbdata);
break;
case PMIXP_COLL_TYPE_FENCE_RING:
ret = pmixp_coll_ring_local(coll, data, ndata,
cbfunc, cbdata);
break;
default:
ret = SLURM_ERROR;
break;
}
return ret;
}
/*
* TODO: we need to keep track of the "me"
* structures created here, because we need to
* free them in "pmixp_stepd_finalize"
*/
void pmix_server_new_conn(int fd)
{
eio_obj_t *obj;
PMIXP_DEBUG("Request from fd = %d", fd);
/* Set nonblocking */
fd_set_nonblocking(fd);
fd_set_close_on_exec(fd);
pmixp_io_engine_t *me = xmalloc(sizeof(pmixp_io_engine_t));
pmix_io_init(me, fd, srv_rcvd_header);
/* We use slurm_forward_data to send message to stepd's
* SLURM will put user ID there. We need to skip it.
*/
pmix_io_rcvd_padding(me, sizeof(uint32_t));
if( 2 == _process_message(me) ){
/* connection was fully processed here */
xfree(me);
return;
}
/* If it is a blocking operation: create AIO object to
* handle it */
obj = eio_obj_create(fd, &peer_ops, (void *)me);
eio_new_obj(pmixp_info_io(), obj);
}
int pmixp_p2p_send(const char *nodename, const char *address, const char *data,
uint32_t len, unsigned int start_delay,
unsigned int retry_cnt, int silent)
{
int retry = 0, rc;
unsigned int delay = start_delay; /* in milliseconds */
pmixp_debug_hang(0);
while (1) {
if (!silent && retry >= 1) {
PMIXP_DEBUG("send failed, rc=%d, try #%d", rc, retry);
}
rc = _pmix_p2p_send_core(nodename, address, data, len);
if (rc == SLURM_SUCCESS)
break;
retry++;
if (retry >= retry_cnt) {
PMIXP_ERROR("send failed, rc=%d, exceeded the retry limit", rc);
break;
}
/* wait with constantly increasing delay */
struct timespec ts =
{(delay / 1000), ((delay % 1000) * 1000000)};
nanosleep(&ts, NULL);
delay *= 2;
}
return rc;
}
static pmix_status_t _unpublish_fn(const pmix_proc_t *proc, char **keys,
const pmix_info_t info[], size_t ninfo,
pmix_op_cbfunc_t cbfunc, void *cbdata)
{
PMIXP_DEBUG("called");
return PMIX_ERR_NOT_SUPPORTED;
}
static pmix_status_t _fencenb_fn(const pmix_proc_t procs_v2[], size_t nprocs,
const pmix_info_t info[], size_t ninfo,
char *data, size_t ndata,
pmix_modex_cbfunc_t cbfunc, void *cbdata)
{
PMIXP_DEBUG("called");
pmixp_coll_t *coll;
pmixp_coll_type_t type = PMIXP_COLL_TYPE_FENCE;
pmix_status_t status = PMIX_SUCCESS;
int ret;
size_t i;
pmixp_proc_t *procs = xmalloc(sizeof(*procs) * nprocs);
for (i = 0; i < nprocs; i++) {
procs[i].rank = procs_v2[i].rank;
strncpy(procs[i].nspace, procs_v2[i].nspace, PMIXP_MAX_NSLEN);
}
coll = pmixp_state_coll_get(type, procs, nprocs);
ret = pmixp_coll_contrib_local(coll, data, ndata, cbfunc, cbdata);
xfree(procs);
if (SLURM_SUCCESS != ret) {
status = PMIX_ERROR;
goto error;
}
return PMIX_SUCCESS;
error:
cbfunc(status, NULL, 0, cbdata, NULL, NULL);
return status;
}
static int _pmix_p2p_send_core(const char *nodename, const char *address,
const char *data, uint32_t len)
{
int rc, timeout;
slurm_msg_t msg;
forward_data_msg_t req;
List ret_list;
ret_data_info_t *ret_data_info = NULL;
pmixp_debug_hang(0);
slurm_msg_t_init(&msg);
PMIXP_DEBUG("nodelist=%s, address=%s, len=%u", nodename, address, len);
req.address = (char *)address;
req.len = len;
/* there is not much we can do - just cast) */
req.data = (char*)data;
msg.msg_type = REQUEST_FORWARD_DATA;
msg.data = &req;
if (slurm_conf_get_addr(nodename, &msg.address) == SLURM_ERROR) {
PMIXP_ERROR("Can't find address for host "
"%s, check slurm.conf", nodename);
return SLURM_ERROR;
}
timeout = slurm_get_msg_timeout() * 1000;
msg.forward.timeout = timeout;
msg.forward.cnt = 0;
msg.forward.nodelist = NULL;
ret_list = slurm_send_addr_recv_msgs(&msg, (char*)nodename, timeout);
if (!ret_list) {
/* This should never happen (when this was
* written slurm_send_addr_recv_msgs always
* returned a list */
PMIXP_ERROR("No return list given from "
"slurm_send_addr_recv_msgs spawned for %s",
nodename);
return SLURM_ERROR;
} else if ((errno != SLURM_COMMUNICATIONS_CONNECTION_ERROR) &&
!list_count(ret_list)) {
PMIXP_ERROR("failed to send to %s, errno=%d", nodename, errno);
return SLURM_ERROR;
}
rc = SLURM_SUCCESS;
while ((ret_data_info = list_pop(ret_list))) {
int temp_rc = slurm_get_return_code(ret_data_info->type,
ret_data_info->data);
if (temp_rc != SLURM_SUCCESS)
rc = temp_rc;
destroy_data_info(ret_data_info);
}
FREE_NULL_LIST(ret_list);
return rc;
}
static pmix_status_t _spawn_fn(const pmix_proc_t *proc,
const pmix_info_t job_info[], size_t ninfo,
const pmix_app_t apps[], size_t napps,
pmix_spawn_cbfunc_t cbfunc, void *cbdata)
{
PMIXP_DEBUG("called");
return PMIX_ERR_NOT_SUPPORTED;
}
static pmix_status_t _dmodex_fn(const pmix_proc_t *proc,
const pmix_info_t info[], size_t ninfo,
pmix_modex_cbfunc_t cbfunc, void *cbdata)
{
int rc;
PMIXP_DEBUG("called");
rc = pmixp_dmdx_get(proc->nspace, proc->rank, cbfunc, cbdata);
return (SLURM_SUCCESS == rc) ? PMIX_SUCCESS : PMIX_ERROR;
}
static pmix_status_t abort_fn(const pmix_proc_t *proc, void *server_object,
int status, const char msg[], pmix_proc_t procs[],
size_t nprocs, pmix_op_cbfunc_t cbfunc, void *cbdata)
{
/* Just kill this stepid for now. Think what we can do for FT here? */
PMIXP_DEBUG("called: status = %d, msg = %s", status, msg);
slurm_kill_job_step(pmixp_info_jobid(), pmixp_info_stepid(), SIGKILL);
if (NULL != cbfunc) {
cbfunc(PMIX_SUCCESS, cbdata);
}
return PMIX_SUCCESS;
}
static int _ring_forward_data(pmixp_coll_ring_ctx_t *coll_ctx, uint32_t contrib_id,
uint32_t hop_seq, void *data, size_t size)
{
pmixp_coll_ring_msg_hdr_t hdr;
pmixp_coll_t *coll = _ctx_get_coll(coll_ctx);
pmixp_coll_ring_t *ring = &coll->state.ring;
hdr.nodeid = coll->my_peerid;
hdr.msgsize = size;
hdr.seq = coll_ctx->seq;
hdr.hop_seq = hop_seq;
hdr.contrib_id = contrib_id;
pmixp_ep_t *ep = (pmixp_ep_t*)xmalloc(sizeof(*ep));
pmixp_coll_ring_cbdata_t *cbdata = NULL;
uint32_t offset = 0;
Buf buf = _get_fwd_buf(coll_ctx);
int rc = SLURM_SUCCESS;
pmixp_coll_ring_ctx_sanity_check(coll_ctx);
#ifdef PMIXP_COLL_DEBUG
PMIXP_DEBUG("%p: transit data to nodeid=%d, seq=%d, hop=%d, size=%lu, contrib=%d",
coll_ctx, _ring_next_id(coll), hdr.seq,
hdr.hop_seq, hdr.msgsize, hdr.contrib_id);
#endif
if (!buf) {
rc = SLURM_ERROR;
goto exit;
}
ep->type = PMIXP_EP_NOIDEID;
ep->ep.nodeid = ring->next_peerid;
/* pack ring info */
_pack_coll_ring_info(coll, &hdr, buf);
/* insert payload to buf */
offset = get_buf_offset(buf);
pmixp_server_buf_reserve(buf, size);
memcpy(get_buf_data(buf) + offset, data, size);
set_buf_offset(buf, offset + size);
cbdata = xmalloc(sizeof(pmixp_coll_ring_cbdata_t));
cbdata->buf = buf;
cbdata->coll = coll;
cbdata->coll_ctx = coll_ctx;
cbdata->seq = coll_ctx->seq;
rc = pmixp_server_send_nb(ep, PMIXP_MSG_RING, coll_ctx->seq, buf,
_ring_sent_cb, cbdata);
exit:
return rc;
}
int pmixp_coll_contrib_local(pmixp_coll_t *coll, char *data, size_t size)
{
PMIXP_DEBUG("%s:%d: get local contribution", pmixp_info_namespace(),
pmixp_info_nodeid());
/* sanity check */
pmixp_coll_sanity_check(coll);
/* lock the structure */
slurm_mutex_lock(&coll->lock);
/* change the collective state if need */
if (PMIXP_COLL_SYNC == coll->state) {
PMIXP_DEBUG(
"%s:%d: get local contribution: switch to PMIXP_COLL_FAN_IN",
pmixp_info_namespace(), pmixp_info_nodeid());
coll->state = PMIXP_COLL_FAN_IN;
coll->ts = time(NULL);
}
xassert(PMIXP_COLL_FAN_IN == coll->state);
/* save & mark local contribution */
coll->contrib_local = true;
grow_buf(coll->buf, size);
memcpy(get_buf_data(coll->buf) + get_buf_offset(coll->buf), data, size);
set_buf_offset(coll->buf, get_buf_offset(coll->buf) + size);
/* unlock the structure */
slurm_mutex_unlock(&coll->lock);
/* check if the collective is ready to progress */
_progress_fan_in(coll);
PMIXP_DEBUG("%s:%d: get local contribution: finish",
pmixp_info_namespace(), pmixp_info_nodeid());
return SLURM_SUCCESS;
}
void _progres_fan_out(pmixp_coll_t *coll, Buf buf)
{
PMIXP_DEBUG("%s:%d: start", pmixp_info_namespace(), pmixp_info_nodeid());
pmixp_coll_sanity_check(coll);
xassert(PMIXP_COLL_FAN_OUT == coll->state || PMIXP_COLL_FAN_OUT_IN == coll->state);
/* update the database */
if (NULL != coll->cbfunc) {
void *data = get_buf_data(buf) + get_buf_offset(buf);
size_t size = remaining_buf(buf);
PMIXP_DEBUG("%s:%d: use the callback", pmixp_info_namespace(),
pmixp_info_nodeid());
coll->cbfunc(PMIX_SUCCESS, data, size, coll->cbdata,
pmixp_free_Buf, (void *)buf);
}
/* Prepare for the next collective operation */
_fan_out_finished(coll);
PMIXP_DEBUG("%s:%d: collective is prepared for the next use",
pmixp_info_namespace(), pmixp_info_nodeid());
}
static void _ring_sent_cb(int rc, pmixp_p2p_ctx_t ctx, void *_cbdata)
{
pmixp_coll_ring_cbdata_t *cbdata = (pmixp_coll_ring_cbdata_t*)_cbdata;
pmixp_coll_ring_ctx_t *coll_ctx = cbdata->coll_ctx;
pmixp_coll_t *coll = cbdata->coll;
Buf buf = cbdata->buf;
pmixp_coll_sanity_check(coll);
if (PMIXP_P2P_REGULAR == ctx) {
/* lock the collective */
slurm_mutex_lock(&coll->lock);
}
#ifdef PMIXP_COLL_DEBUG
PMIXP_DEBUG("%p: called %d", coll_ctx, coll_ctx->seq);
#endif
if (cbdata->seq != coll_ctx->seq) {
/* it seems like this collective was reset since the time
* we initiated this send.
* Just exit to avoid data corruption.
*/
PMIXP_DEBUG("%p: collective was reset!", coll_ctx);
goto exit;
}
coll_ctx->forward_cnt++;
_progress_coll_ring(coll_ctx);
exit:
pmixp_server_buf_reset(buf);
list_push(coll->state.ring.fwrd_buf_pool, buf);
if (PMIXP_P2P_REGULAR == ctx) {
/* unlock the collective */
slurm_mutex_unlock(&coll->lock);
}
xfree(cbdata);
}
static void _reset_coll_ring(pmixp_coll_ring_ctx_t *coll_ctx)
{
pmixp_coll_t *coll = _ctx_get_coll(coll_ctx);
#ifdef PMIXP_COLL_DEBUG
PMIXP_DEBUG("%p: called", coll_ctx);
#endif
pmixp_coll_ring_ctx_sanity_check(coll_ctx);
coll_ctx->in_use = false;
coll_ctx->state = PMIXP_COLL_RING_SYNC;
coll_ctx->contrib_local = false;
coll_ctx->contrib_prev = 0;
coll_ctx->forward_cnt = 0;
coll->ts = time(NULL);
memset(coll_ctx->contrib_map, 0, sizeof(bool) * coll->peers_cnt);
coll_ctx->ring_buf = NULL;
}
void pmixp_coll_bcast(pmixp_coll_t *coll, Buf buf)
{
PMIXP_DEBUG("%s:%d: start", pmixp_info_namespace(), pmixp_info_nodeid());
/* lock the structure */
slurm_mutex_lock(&coll->lock);
_progres_fan_out(coll, buf);
/* unlock the structure */
slurm_mutex_unlock(&coll->lock);
/* We may already start next collective. Try to progress!
* its OK if we in SYNC - there will be no-op */
_progress_fan_in(coll);
}
int pmixp_coll_ring_local(pmixp_coll_t *coll, char *data, size_t size,
void *cbfunc, void *cbdata)
{
int ret = SLURM_SUCCESS;
pmixp_coll_ring_ctx_t *coll_ctx = NULL;
/* lock the structure */
slurm_mutex_lock(&coll->lock);
/* sanity check */
pmixp_coll_sanity_check(coll);
/* setup callback info */
coll->cbfunc = cbfunc;
coll->cbdata = cbdata;
coll_ctx = pmixp_coll_ring_ctx_new(coll);
if (!coll_ctx) {
PMIXP_ERROR("Can not get new ring collective context, seq=%u",
coll->seq);
ret = SLURM_ERROR;
goto exit;
}
#ifdef PMIXP_COLL_DEBUG
PMIXP_DEBUG("%p: contrib/loc: seqnum=%u, state=%d, size=%lu",
coll_ctx, coll_ctx->seq, coll_ctx->state, size);
#endif
if (_pmixp_coll_contrib(coll_ctx, coll->my_peerid, 0, data, size)) {
goto exit;
}
/* mark local contribution */
coll_ctx->contrib_local = true;
_progress_coll_ring(coll_ctx);
exit:
/* unlock the structure */
slurm_mutex_unlock(&coll->lock);
return ret;
}
static void _libpmix_cb(void *_vcbdata)
{
pmixp_coll_cbdata_t *cbdata = (pmixp_coll_cbdata_t*)_vcbdata;
pmixp_coll_t *coll = cbdata->coll;
/* lock the collective */
slurm_mutex_lock(&coll->lock);
if (cbdata->seq != coll->seq) {
/* it seems like this collective was reset since the time
* we initiated this send.
* Just exit to avoid data corruption.
*/
PMIXP_ERROR("%p: collective was reset: myseq=%u, curseq=%u",
coll, cbdata->seq, coll->seq);
goto exit;
}
xassert(PMIXP_COLL_DOWNFWD == coll->state);
coll->dfwd_cb_cnt++;
#ifdef PMIXP_COLL_DEBUG
PMIXP_DEBUG("%p: state: %s, snd_status=%s, compl_cnt=%d/%d",
coll, pmixp_coll_state2str(coll->state),
pmixp_coll_sndstatus2str(coll->dfwd_status),
coll->dfwd_cb_cnt, coll->dfwd_cb_wait);
#endif
_progress_coll(coll);
exit:
xassert(0 < cbdata->refcntr);
cbdata->refcntr--;
if (!cbdata->refcntr) {
xfree(cbdata);
}
/* unlock the collective */
slurm_mutex_unlock(&coll->lock);
}
pmix_status_t fencenb_fn(const pmix_proc_t procs[], size_t nprocs,
const pmix_info_t info[], size_t ninfo,
char *data, size_t ndata,
pmix_modex_cbfunc_t cbfunc, void *cbdata)
{
PMIXP_DEBUG("called");
pmixp_coll_t *coll;
pmixp_coll_type_t type = PMIXP_COLL_TYPE_FENCE;
pmix_status_t status = PMIX_SUCCESS;
pmixp_debug_hang(0);
coll = pmixp_state_coll_get(type, procs, nprocs);
pmixp_coll_set_callback(coll, cbfunc, cbdata);
if (SLURM_SUCCESS != pmixp_coll_contrib_local(coll, data, ndata)) {
goto error;
}
return PMIX_SUCCESS;
error:
cbfunc(status, NULL, 0, cbdata, NULL, NULL);
return status;
}
static int _progress_ufwd_wpc(pmixp_coll_t *coll)
{
xassert(PMIXP_COLL_UPFWD_WPC == coll->state);
if (!coll->contrib_prnt) {
return false;
}
/* Need to wait only for the local completion callback if installed*/
coll->dfwd_status = PMIXP_COLL_SND_ACTIVE;
coll->dfwd_cb_wait = 0;
/* move to the next state */
coll->state = PMIXP_COLL_DOWNFWD;
/* local delivery */
if (coll->cbfunc) {
pmixp_coll_cbdata_t *cbdata;
cbdata = xmalloc(sizeof(pmixp_coll_cbdata_t));
cbdata->coll = coll;
cbdata->seq = coll->seq;
cbdata->refcntr = 1;
char *data = get_buf_data(coll->dfwd_buf) + coll->dfwd_offset;
size_t size = get_buf_offset(coll->dfwd_buf) -
coll->dfwd_offset;
coll->cbfunc(PMIX_SUCCESS, data, size, coll->cbdata,
_libpmix_cb, (void *)cbdata);
coll->dfwd_cb_wait++;
#ifdef PMIXP_COLL_DEBUG
PMIXP_DEBUG("%p: local delivery, size = %lu",
coll, (uint64_t)size);
#endif
}
/* events observed - need another iteration */
return true;
}
int pmixp_stepd_send(const char *nodelist, const char *address,
const char *data, uint32_t len,
unsigned int start_delay,
unsigned int retry_cnt, int silent)
{
int retry = 0, rc;
unsigned int delay = start_delay; /* in milliseconds */
char *copy_of_nodelist = xstrdup(nodelist);
while (1) {
if (!silent && retry >= 1) {
PMIXP_DEBUG("send failed, rc=%d, try #%d", rc, retry);
}
rc = slurm_forward_data(©_of_nodelist, (char *)address,
len, data);
if (rc == SLURM_SUCCESS)
break;
retry++;
if (retry >= retry_cnt) {
PMIXP_ERROR("send failed, rc=%d, exceeded the retry limit", rc);
break;
}
/* wait with constantly increasing delay */
struct timespec ts =
{(delay / 1000), ((delay % 1000) * 1000000)};
nanosleep(&ts, NULL);
delay *= 2;
}
xfree(copy_of_nodelist);
return rc;
}
