int __recvpath
psmi_mq_handle_rts_outoforder(psm_mq_t mq, uint64_t tag,
uintptr_t send_buf, uint32_t send_msglen,
psm_epaddr_t peer, uint16_t msg_seqnum,
mq_rts_callback_fn_t cb,
psm_mq_req_t *req_o)
{
psm_mq_req_t req;
PSMI_PLOCK_ASSERT();
req = psmi_mq_req_alloc(mq, MQE_TYPE_RECV);
psmi_assert(req != NULL);
/* We don't know recv_msglen yet but we set it here for
* mq_iprobe */
req->send_msglen = req->recv_msglen = send_msglen;
req->state = MQ_STATE_UNEXP_RV;
req->tag = tag;
req->rts_callback = cb;
req->recv_msgoff = 0;
req->send_msgoff = 0;
req->rts_peer = peer;
req->rts_sbuf = send_buf;
req->msg_seqnum = msg_seqnum;
mq_sq_append(&peer->mctxt_master->outoforder_q, req);
peer->mctxt_master->outoforder_c++;
*req_o = req; /* no match, will callback */
_IPATH_VDBG("from=%s match=%s (req=%p) mqtag=%" PRIx64" recvlen=%d "
"sendlen=%d errcode=%d\n", psmi_epaddr_get_name(peer->epid),
"NO", req, req->tag,
req->recv_msglen, req->send_msglen, req->error_code);
return MQ_RET_UNEXP_OK;
}
int __recvpath
psmi_mq_handle_data(psm_mq_req_t req, psm_epaddr_t epaddr,
uint32_t egrid, uint32_t offset,
const void *buf, uint32_t nbytes)
{
psm_mq_t mq;
int rc;
if (req == NULL) goto no_req;
mq = req->mq;
if (req->state == MQ_STATE_MATCHED)
rc = MQ_RET_MATCH_OK;
else {
psmi_assert(req->state == MQ_STATE_UNEXP);
rc = MQ_RET_UNEXP_OK;
}
psmi_assert(req->egrid.egr_data == egrid);
psmi_mq_req_copy(req, epaddr, offset, buf, nbytes);
if (req->send_msgoff == req->send_msglen) {
if (req->type & MQE_TYPE_EGRLONG) {
STAILQ_REMOVE(&epaddr->mctxt_master->egrlong,
req, psm_mq_req, nextq);
}
if (req->state == MQ_STATE_MATCHED) {
req->state = MQ_STATE_COMPLETE;
mq_qq_append(&mq->completed_q, req);
}
else { /* MQ_STATE_UNEXP */
req->state = MQ_STATE_COMPLETE;
}
_IPATH_VDBG("epaddr=%s completed %d byte send, state=%d\n",
psmi_epaddr_get_name(epaddr->epid),
(int)req->send_msglen, req->state);
}
return rc;
no_req:
mq = epaddr->ep->mq;
req = psmi_mq_req_alloc(mq, MQE_TYPE_RECV);
psmi_assert(req != NULL);
req->egrid.egr_data = egrid;
req->recv_msgoff = offset;
req->recv_msglen = nbytes;
req->buf = psmi_mq_sysbuf_alloc(mq, nbytes);
psmi_mq_mtucpy(req->buf, buf, nbytes);
STAILQ_INSERT_TAIL(&epaddr->mctxt_master->egrdata, req, nextq);
return MQ_RET_UNEXP_OK;
}
/*
* Add ipsaddr with epid to the epstate table, return new index to caller in
* 'connidx'.
*/
psm2_error_t
ips_epstate_add(struct ips_epstate *eps, struct ips_epaddr *ipsaddr,
ips_epstate_idx *connidx_o)
{
int i, j;
ips_epstate_idx connidx;
if (++eps->eps_tabsizeused > eps->eps_tabsize) { /* realloc */
struct ips_epstate_entry *newtab;
eps->eps_tabsize += PTL_EPADDR_ALLOC_CHUNK;
newtab = (struct ips_epstate_entry *)
psmi_calloc(eps->context->ep, PER_PEER_ENDPOINT,
eps->eps_tabsize,
sizeof(struct ips_epstate_entry));
if (newtab == NULL)
return PSM2_NO_MEMORY;
else if (eps->eps_tab) { /* NOT first alloc */
for (i = 0;
i < eps->eps_tabsize - PTL_EPADDR_ALLOC_CHUNK; i++)
newtab[i] = eps->eps_tab[i]; /* deep copy */
psmi_free(eps->eps_tab);
}
eps->eps_tab = newtab;
}
/* Find the next free hole. We can afford to do this since connect is not
* in the critical path */
for (i = 0, j = eps->eps_tab_nextidx; i < eps->eps_tabsize; i++, j++) {
if (j == eps->eps_tabsize)
j = 0;
if (eps->eps_tab[j].ipsaddr == NULL) {
eps->eps_tab_nextidx = j + 1;
if (eps->eps_tab_nextidx == eps->eps_tabsize)
eps->eps_tab_nextidx = 0;
break;
}
}
psmi_assert_always(i != eps->eps_tabsize);
connidx = (j - eps->eps_base_idx) & (IPS_EPSTATE_CONNIDX_MAX-1);
_HFI_VDBG("node %s gets connidx=%d (table idx %d)\n",
psmi_epaddr_get_name(((psm2_epaddr_t) ipsaddr)->epid), connidx,
j);
eps->eps_tab[j].ipsaddr = ipsaddr;
if (j >= IPS_EPSTATE_CONNIDX_MAX) {
return psmi_handle_error(eps->context->ep,
PSM2_TOO_MANY_ENDPOINTS,
"Can't connect to more than %d non-local endpoints",
IPS_EPSTATE_CONNIDX_MAX);
}
*connidx_o = connidx;
return PSM2_OK;
}
int __recvpath
psmi_mq_handle_rts(psm_mq_t mq, uint64_t tag,
uintptr_t send_buf, uint32_t send_msglen,
psm_epaddr_t peer, mq_rts_callback_fn_t cb,
psm_mq_req_t *req_o)
{
psm_mq_req_t req;
uint32_t msglen;
int rc;
PSMI_PLOCK_ASSERT();
req = mq_req_match(&(mq->expected_q), tag, 1);
if (req) { /* we have a match, no need to callback */
msglen = mq_set_msglen(req, req->buf_len, send_msglen);
req->type = MQE_TYPE_RECV;
req->state = MQ_STATE_MATCHED;
req->tag = tag;
req->recv_msgoff = 0;
req->rts_peer = peer;
req->rts_sbuf = send_buf;
*req_o = req; /* yes match */
rc = MQ_RET_MATCH_OK;
}
else { /* No match, keep track of callback */
req = psmi_mq_req_alloc(mq, MQE_TYPE_RECV);
psmi_assert(req != NULL);
req->type = MQE_TYPE_RECV;
/* We don't know recv_msglen yet but we set it here for
* mq_iprobe */
req->send_msglen = req->recv_msglen = send_msglen;
req->state = MQ_STATE_UNEXP_RV;
req->tag = tag;
req->rts_callback = cb;
req->recv_msgoff = 0;
req->rts_peer = peer;
req->rts_sbuf = send_buf;
mq_sq_append(&mq->unexpected_q, req);
*req_o = req; /* no match, will callback */
rc = MQ_RET_UNEXP_OK;
}
_IPATH_VDBG("from=%s match=%s (req=%p) mqtag=%" PRIx64" recvlen=%d "
"sendlen=%d errcode=%d\n", psmi_epaddr_get_name(peer->epid),
rc == MQ_RET_MATCH_OK ? "YES" : "NO", req, req->tag,
req->recv_msglen, req->send_msglen, req->error_code);
return rc;
}
int __recvpath
psmi_mq_handle_data(psm_mq_req_t req, psm_epaddr_t epaddr,
const void *buf, uint32_t nbytes)
{
psm_mq_t mq = req->mq;
int rc;
if (req->state == MQ_STATE_MATCHED)
rc = MQ_RET_MATCH_OK;
else {
psmi_assert(req->state == MQ_STATE_UNEXP);
rc = MQ_RET_UNEXP_OK;
}
psmi_mq_req_copy(req, epaddr, buf, nbytes);
if (req->send_msgoff == req->send_msglen) {
if (req->type & MQE_TYPE_EGRLONG) {
int flowid = req->egrid.egr_flowid;
psmi_assert(STAILQ_FIRST(&epaddr->egrlong[flowid]) == req);
STAILQ_REMOVE_HEAD(&epaddr->egrlong[flowid], nextq);
}
/* Whatever is leftover in the posted message should be now marked as
* undefined.
* XXX Sends not supported yet.
*/
#if 0
#ifdef PSM_VALGRIND
if (req->send_msglen < req->buf_len)
VALGRIND_MAKE_MEM_UNDEFINED(
(void *) ((uintptr_t) req->buf + req->send_msglen),
req->buf_len - req->send_msglen);
#endif
#endif
if (req->state == MQ_STATE_MATCHED) {
req->state = MQ_STATE_COMPLETE;
mq_qq_append(&mq->completed_q, req);
}
else { /* MQ_STATE_UNEXP */
req->state = MQ_STATE_COMPLETE;
}
_IPATH_VDBG("epaddr=%s completed %d byte send, state=%d\n",
psmi_epaddr_get_name(epaddr->epid),
(int)req->send_msglen, req->state);
}
return rc;
}
int __recvpath
psmi_mq_handle_envelope_outoforder(psm_mq_t mq, uint16_t mode,
psm_epaddr_t epaddr, uint16_t msg_seqnum,
uint64_t tag, psmi_egrid_t egrid, uint32_t send_msglen,
const void *payload, uint32_t paylen)
{
psm_mq_req_t req;
uint32_t msglen;
req = psmi_mq_req_alloc(mq, MQE_TYPE_RECV);
psmi_assert(req != NULL);
req->tag = tag;
req->recv_msgoff = 0;
req->recv_msglen = req->send_msglen = req->buf_len = msglen = send_msglen;
_IPATH_VDBG(
"from=%s match=NO (req=%p) mode=%x mqtag=%" PRIx64
" send_msglen=%d\n", psmi_epaddr_get_name(epaddr->epid),
req, mode, tag, send_msglen);
switch (mode) {
case MQ_MSG_TINY:
if (msglen > 0) {
req->buf = psmi_mq_sysbuf_alloc(mq, msglen);
mq_copy_tiny((uint32_t *)req->buf, (uint32_t *)payload, msglen);
}
else
req->buf = NULL;
req->state = MQ_STATE_COMPLETE;
break;
case MQ_MSG_SHORT:
req->buf = psmi_mq_sysbuf_alloc(mq, msglen);
psmi_mq_mtucpy(req->buf, payload, msglen);
req->state = MQ_STATE_COMPLETE;
break;
case MQ_MSG_LONG:
req->egrid = egrid;
req->epaddr = epaddr;
req->send_msgoff = 0;
req->buf = psmi_mq_sysbuf_alloc(mq, msglen);
req->state = MQ_STATE_UNEXP;
req->type |= MQE_TYPE_EGRLONG;
STAILQ_INSERT_TAIL(&epaddr->mctxt_master->egrlong, req, nextq);
_IPATH_VDBG("unexp MSG_LONG %d of length %d bytes pay=%d\n",
egrid.egr_msgno, msglen, paylen);
if (paylen > 0)
psmi_mq_handle_data(req, epaddr,
egrid.egr_data, 0, payload, paylen);
psmi_mq_handle_egrdata(mq, req, epaddr);
break;
default:
psmi_handle_error(PSMI_EP_NORETURN, PSM_INTERNAL_ERR,
"Internal error, unknown packet 0x%x", mode);
}
req->msg_seqnum = msg_seqnum;
mq_sq_append(&epaddr->mctxt_master->outoforder_q, req);
epaddr->mctxt_master->outoforder_c++;
mq->stats.rx_sys_bytes += msglen;
mq->stats.rx_sys_num++;
return MQ_RET_UNEXP_OK;
}
/*
* This handles the regular (i.e. non-rendezvous MPI envelopes)
*/
int __recvpath
psmi_mq_handle_envelope(psm_mq_t mq, uint16_t mode, psm_epaddr_t epaddr,
uint64_t tag, psmi_egrid_t egrid, uint32_t send_msglen,
const void *payload, uint32_t paylen)
{
psm_mq_req_t req;
uint32_t msglen;
int rc;
psmi_assert(epaddr != NULL);
req = mq_req_match(&(mq->expected_q), tag, 1);
if (req) { /* we have a match */
psmi_assert(MQE_TYPE_IS_RECV(req->type));
req->tag = tag;
msglen = mq_set_msglen(req, req->buf_len, send_msglen);
_IPATH_VDBG("from=%s match=YES (req=%p) mode=%x mqtag=%"
PRIx64" msglen=%d paylen=%d\n", psmi_epaddr_get_name(epaddr->epid),
req, mode, tag, msglen, paylen);
switch(mode) {
case MQ_MSG_TINY:
PSM_VALGRIND_DEFINE_MQ_RECV(req->buf, req->buf_len, msglen);
mq_copy_tiny((uint32_t *)req->buf, (uint32_t *)payload, msglen);
req->state = MQ_STATE_COMPLETE;
mq_qq_append(&mq->completed_q, req);
break;
case MQ_MSG_SHORT: /* message fits in 1 payload */
PSM_VALGRIND_DEFINE_MQ_RECV(req->buf, req->buf_len, msglen);
psmi_mq_mtucpy(req->buf, payload, msglen);
req->state = MQ_STATE_COMPLETE;
mq_qq_append(&mq->completed_q, req);
break;
case MQ_MSG_LONG:
req->egrid = egrid;
req->state = MQ_STATE_MATCHED;
req->type |= MQE_TYPE_EGRLONG;
req->send_msgoff = req->recv_msgoff = 0;
STAILQ_INSERT_TAIL(&epaddr->mctxt_master->egrlong, req, nextq);
_IPATH_VDBG("exp MSG_LONG %d of length %d bytes pay=%d\n",
egrid.egr_msgno, msglen, paylen);
if (paylen > 0)
psmi_mq_handle_data(req, epaddr,
egrid.egr_data, 0, payload, paylen);
psmi_mq_handle_egrdata(mq, req, epaddr);
break;
default:
psmi_handle_error(PSMI_EP_NORETURN, PSM_INTERNAL_ERR,
"Internal error, unknown packet 0x%x", mode);
}
mq->stats.rx_user_bytes += msglen;
mq->stats.rx_user_num++;
rc = MQ_RET_MATCH_OK;
if (mode == MQ_MSG_LONG)
return rc;
}
else
rc = psmi_mq_handle_envelope_unexpected(mq, mode, epaddr, tag,
egrid, send_msglen, payload, paylen);
return rc;
}
int __recvpath
psmi_mq_handle_envelope_unexpected(
psm_mq_t mq, uint16_t mode, psm_epaddr_t epaddr,
uint64_t tag, psmi_egrid_t egrid, uint32_t send_msglen,
const void *payload, uint32_t paylen)
{
psm_mq_req_t req;
uint32_t msglen;
/*
* Keep a callback here in case we want to fit some other high-level
* protocols over MQ (i.e. shmem). These protocols would bypass the
* normal mesage handling and go to higher-level message handlers.
*/
if (mode >= MQ_MSG_USER_FIRST && mq->unexpected_callback) {
mq->unexpected_callback(mq,mode,epaddr,tag,send_msglen,payload,paylen);
return MQ_RET_UNEXP_OK;
}
req = psmi_mq_req_alloc(mq, MQE_TYPE_RECV);
psmi_assert(req != NULL);
req->tag = tag;
req->recv_msgoff = 0;
req->recv_msglen = req->send_msglen = req->buf_len = msglen = send_msglen;
_IPATH_VDBG(
"from=%s match=NO (req=%p) mode=%x mqtag=%" PRIx64
" send_msglen=%d\n", psmi_epaddr_get_name(epaddr->epid),
req, mode, tag, send_msglen);
#if 0
if (mq->cur_sysbuf_bytes+msglen > mq->max_sysbuf_bytes) {
_IPATH_VDBG("req=%p with len=%d exceeds limit of %llu sysbuf_bytes\n",
req, msglen, (unsigned long long) mq->max_sysbuf_bytes);
return MQ_RET_UNEXP_NO_RESOURCES;
}
#endif
switch (mode) {
case MQ_MSG_TINY:
if (msglen > 0) {
req->buf = psmi_mq_sysbuf_alloc(mq, msglen);
mq_copy_tiny((uint32_t *)req->buf, (uint32_t *)payload, msglen);
}
else
req->buf = NULL;
req->state = MQ_STATE_COMPLETE;
break;
case MQ_MSG_SHORT:
req->buf = psmi_mq_sysbuf_alloc(mq, msglen);
psmi_mq_mtucpy(req->buf, payload, msglen);
req->state = MQ_STATE_COMPLETE;
break;
case MQ_MSG_LONG:
req->egrid = egrid;
req->send_msgoff = 0;
req->buf = psmi_mq_sysbuf_alloc(mq, msglen);
req->state = MQ_STATE_UNEXP;
req->type |= MQE_TYPE_EGRLONG;
STAILQ_INSERT_TAIL(&epaddr->mctxt_master->egrlong, req, nextq);
_IPATH_VDBG("unexp MSG_LONG %d of length %d bytes pay=%d\n",
egrid.egr_msgno, msglen, paylen);
if (paylen > 0)
psmi_mq_handle_data(req, epaddr,
egrid.egr_data, 0, payload, paylen);
psmi_mq_handle_egrdata(mq, req, epaddr);
break;
default:
psmi_handle_error(PSMI_EP_NORETURN, PSM_INTERNAL_ERR,
"Internal error, unknown packet 0x%x", mode);
}
mq_sq_append(&mq->unexpected_q, req);
mq->stats.rx_sys_bytes += msglen;
mq->stats.rx_sys_num++;
return MQ_RET_UNEXP_OK;
}
psm2_error_t
__psm2_ep_connect(psm2_ep_t ep, int num_of_epid, psm2_epid_t const *array_of_epid,
int const *array_of_epid_mask, /* can be NULL */
psm2_error_t *array_of_errors, psm2_epaddr_t *array_of_epaddr,
int64_t timeout)
{
psm2_error_t err = PSM2_OK;
ptl_ctl_t *ptlctl;
ptl_t *ptl;
int i, j, dup_idx;
int num_toconnect = 0;
int *epid_mask = NULL;
int *epid_mask_isdupof = NULL;
char *device;
uint64_t t_start = get_cycles();
uint64_t t_left;
union psmi_envvar_val timeout_intval;
PSM2_LOG_MSG("entering");
PSMI_ERR_UNLESS_INITIALIZED(ep);
PSMI_PLOCK();
/*
* Normally we would lock here, but instead each implemented ptl component
* does its own locking. This is mostly because the ptl components are
* ahead of the PSM interface in that they can disconnect their peers.
*/
if (ep == NULL || array_of_epaddr == NULL || array_of_epid == NULL ||
num_of_epid < 1) {
err = psmi_handle_error(ep, PSM2_PARAM_ERR,
"Invalid psm2_ep_connect parameters");
goto fail;
}
/* We need two of these masks to detect duplicates */
err = PSM2_NO_MEMORY;
epid_mask =
(int *)psmi_malloc(ep, UNDEFINED, sizeof(int) * num_of_epid);
if (epid_mask == NULL)
goto fail;
epid_mask_isdupof =
(int *)psmi_malloc(ep, UNDEFINED, sizeof(int) * num_of_epid);
if (epid_mask_isdupof == NULL)
goto fail;
err = PSM2_OK;
/* Eventually handle timeouts across all connects. */
for (j = 0; j < num_of_epid; j++) {
if (array_of_epid_mask != NULL && !array_of_epid_mask[j])
epid_mask[j] = 0;
else {
epid_mask[j] = 1;
array_of_errors[j] = PSM2_EPID_UNKNOWN;
array_of_epaddr[j] = NULL;
num_toconnect++;
}
epid_mask_isdupof[j] = -1;
}
psmi_getenv("PSM2_CONNECT_TIMEOUT",
"End-point connection timeout over-ride. 0 for no time-out.",
PSMI_ENVVAR_LEVEL_USER, PSMI_ENVVAR_TYPE_UINT,
(union psmi_envvar_val)0, &timeout_intval);
if (getenv("PSM2_CONNECT_TIMEOUT")) {
timeout = timeout_intval.e_uint * SEC_ULL;
} else if (timeout > 0) {
/* The timeout parameter provides the minimum timeout. A heuristic
* is used to scale up the timeout linearly with the number of
* endpoints, and we allow one second per 100 endpoints. */
timeout = max(timeout, (num_toconnect * SEC_ULL) / 100);
}
if (timeout > 0 && timeout < PSMI_MIN_EP_CONNECT_TIMEOUT)
timeout = PSMI_MIN_EP_CONNECT_TIMEOUT;
_HFI_PRDBG("Connect to %d endpoints with time-out of %.2f secs\n",
num_toconnect, (double)timeout / 1e9);
/* Look for duplicates in input array */
for (i = 0; i < num_of_epid; i++) {
for (j = i + 1; j < num_of_epid; j++) {
if (array_of_epid[i] == array_of_epid[j] &&
epid_mask[i] && epid_mask[j]) {
epid_mask[j] = 0; /* don't connect more than once */
epid_mask_isdupof[j] = i;
}
}
}
for (i = 0; i < PTL_MAX_INIT; i++) {
if (ep->devid_enabled[i] == -1)
continue;
/* Set up the right connect ptrs */
switch (ep->devid_enabled[i]) {
case PTL_DEVID_IPS:
ptlctl = &ep->ptl_ips;
ptl = ep->ptl_ips.ptl;
device = "ips";
break;
case PTL_DEVID_AMSH:
ptlctl = &ep->ptl_amsh;
ptl = ep->ptl_amsh.ptl;
device = "amsh";
break;
case PTL_DEVID_SELF:
ptlctl = &ep->ptl_self;
ptl = ep->ptl_self.ptl;
device = "self";
break;
default:
device = "unknown";
ptlctl = &ep->ptl_ips; /*no-unused */
ptl = ep->ptl_ips.ptl; /*no-unused */
device = "ips"; /*no-unused */
psmi_handle_error(PSMI_EP_NORETURN, PSM2_INTERNAL_ERR,
"Unknown/unhandled PTL id %d\n",
ep->devid_enabled[i]);
break;
}
t_left = psmi_cycles_left(t_start, timeout);
_HFI_VDBG("Trying to connect with device %s\n", device);
if ((err = ptlctl->ep_connect(ptl, num_of_epid, array_of_epid,
epid_mask, array_of_errors,
array_of_epaddr,
cycles_to_nanosecs(t_left)))) {
_HFI_PRDBG("Connect failure in device %s err=%d\n",
device, err);
goto connect_fail;
}
/* Now process what's been connected */
for (j = 0; j < num_of_epid; j++) {
dup_idx = epid_mask_isdupof[j];
if (!epid_mask[j] && dup_idx == -1)
continue;
if (dup_idx != -1) { /* dup */
array_of_epaddr[j] = array_of_epaddr[dup_idx];
array_of_errors[j] = array_of_errors[dup_idx];
epid_mask_isdupof[j] = -1;
}
if (array_of_errors[j] == PSM2_OK) {
epid_mask[j] = 0; /* don't try on next ptl */
ep->connections++;
}
}
}
for (i = 0; i < num_of_epid; i++) {
ptl_ctl_t *c = NULL;
if (array_of_epid_mask != NULL && !array_of_epid_mask[i])
continue;
/* If we see unreachable here, that means some PTLs were not enabled */
if (array_of_errors[i] == PSM2_EPID_UNREACHABLE) {
err = PSM2_EPID_UNREACHABLE;
break;
}
psmi_assert_always(array_of_epaddr[i] != NULL);
c = array_of_epaddr[i]->ptlctl;
psmi_assert_always(c != NULL);
_HFI_VDBG("%-20s DEVICE %s (%p)\n",
psmi_epaddr_get_name(array_of_epid[i]),
c == &ep->ptl_ips ? "hfi" :
(c == &ep->ptl_amsh ? "amsh" : "self"),
(void *)array_of_epaddr[i]->ptlctl->ptl);
}
connect_fail:
/* If the error is a timeout (at worse) and the client is OPA MPI,
* just return timeout to let OPA MPI handle the hostnames that
* timed out */
if (err != PSM2_OK) {
char errbuf[PSM2_ERRSTRING_MAXLEN];
size_t len;
int j = 0;
if (err == PSM2_EPID_UNREACHABLE) {
char *deverr = "of an incorrect setting";
char *eperr = " ";
char *devname = NULL;
if (!psmi_ep_device_is_enabled(ep, PTL_DEVID_AMSH)) {
deverr =
"there is no shared memory PSM device (shm)";
eperr = " shared memory ";
} else
if (!psmi_ep_device_is_enabled(ep, PTL_DEVID_IPS)) {
deverr =
"there is no OPA PSM device (hfi)";
eperr = " OPA ";
}
len = snprintf(errbuf, sizeof(errbuf) - 1,
"Some%sendpoints could not be connected because %s "
"in the currently enabled PSM_DEVICES (",
eperr, deverr);
for (i = 0; i < PTL_MAX_INIT && len < sizeof(errbuf) - 1;
i++) {
switch (ep->devid_enabled[i]) {
case PTL_DEVID_IPS:
devname = "hfi";
break;
case PTL_DEVID_AMSH:
devname = "shm";
break;
case PTL_DEVID_SELF:
default:
devname = "self";
break;
}
len +=
snprintf(errbuf + len,
sizeof(errbuf) - len - 1, "%s,",
devname);
}
if (len < sizeof(errbuf) - 1 && devname != NULL)
/* parsed something, remove trailing comma */
errbuf[len - 1] = ')';
} else
len = snprintf(errbuf, sizeof(errbuf) - 1,
"%s", err == PSM2_TIMEOUT ?
"Dectected connection timeout" :
psm2_error_get_string(err));
/* first pass, look for all nodes with the error */
for (i = 0; i < num_of_epid && len < sizeof(errbuf) - 1; i++) {
if (array_of_epid_mask != NULL
&& !array_of_epid_mask[i])
continue;
if (array_of_errors[i] == PSM2_OK)
continue;
if (array_of_errors[i] == PSM2_EPID_UNREACHABLE &&
err != PSM2_EPID_UNREACHABLE)
continue;
if (err == array_of_errors[i]) {
len +=
snprintf(errbuf + len,
sizeof(errbuf) - len - 1, "%c %s",
j == 0 ? ':' : ',',
psmi_epaddr_get_hostname
(array_of_epid[i]));
j++;
}
}
errbuf[sizeof(errbuf) - 1] = '\0';
err = psmi_handle_error(ep, err, errbuf);
}
fail:
PSMI_PUNLOCK();
if (epid_mask != NULL)
psmi_free(epid_mask);
if (epid_mask_isdupof != NULL)
psmi_free(epid_mask_isdupof);
PSM2_LOG_MSG("leaving");
return err;
}
