psm2_error_t
ips_proto_am_init(struct ips_proto *proto,
int num_send_slots,
uint32_t imm_size,
struct ips_proto_am *proto_am)
{
psm2_error_t err = PSM2_OK;
int send_buf_size = proto->ep->context.ctrl->__hfi_piosize;
int num_rep_slots = calc_optimal_num_reply_slots(num_send_slots);
int num_req_slots = num_send_slots - num_rep_slots;
proto_am->proto = proto;
/* In a node pair, the number of reply send buffers on at least one of
* the nodes must be at least double the number (optimal: double + 1) of
* send descriptors on the other node. While this constraint applies
* only to the reply send buffers, allowing the caller to tune only the
* number of request send buffers would be awkward, as they have no
* knowledge of the subdivision of the memory into separate mempools for
* requests and replies. It's an internal concern at this point. */
if ((err = ips_scbctrl_init(&proto->ep->context,
num_req_slots,
num_req_slots,
imm_size,
send_buf_size,
NULL,
NULL,
&proto_am->scbc_request)))
goto fail;
if ((err = ips_scbctrl_init(&proto->ep->context,
num_rep_slots,
num_rep_slots,
imm_size,
send_buf_size,
NULL,
NULL,
&proto_am->scbc_reply)))
goto fail;
if (ips_am_msg_pool == NULL) {
union psmi_envvar_val max_msgs;
ips_am_outoforder_q.head.next = NULL;
ips_am_outoforder_q.tail = &ips_am_outoforder_q.head;
psmi_getenv("PSM2_AM_MAX_OOO_MSGS",
"Maximum number of OOO Active Messages to queue before dropping.",
PSMI_ENVVAR_LEVEL_HIDDEN, PSMI_ENVVAR_TYPE_UINT,
(union psmi_envvar_val)1024, &max_msgs);
ips_am_msg_pool = psmi_mpool_create(
sizeof(struct ips_am_message),
32, max_msgs.e_uint, 0, UNDEFINED, NULL, NULL);
}
fail:
return err;
}
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;
}
psm_error_t __psm_ep_close(psm_ep_t ep, int mode, int64_t timeout_in)
{
psm_error_t err = PSM_OK;
uint64_t t_start = get_cycles();
union psmi_envvar_val timeout_intval;
psm_ep_t tmp, mep;
PSMI_ERR_UNLESS_INITIALIZED(ep);
psmi_assert_always(ep->mctxt_master == ep);
PSMI_PLOCK();
if (psmi_opened_endpoint == NULL) {
err = psmi_handle_error(NULL, PSM_EP_WAS_CLOSED,
"PSM Endpoint is closed or does not exist");
return err;
}
tmp = psmi_opened_endpoint;
while (tmp && tmp != ep) {
tmp = tmp->user_ep_next;
}
if (!tmp) {
err = psmi_handle_error(NULL, PSM_EP_WAS_CLOSED,
"PSM Endpoint is closed or does not exist");
return err;
}
psmi_getenv("PSM_CLOSE_TIMEOUT",
"End-point close timeout over-ride.",
PSMI_ENVVAR_LEVEL_USER, PSMI_ENVVAR_TYPE_UINT,
(union psmi_envvar_val)0, &timeout_intval);
if (getenv("PSM_CLOSE_TIMEOUT")) {
timeout_in = timeout_intval.e_uint * SEC_ULL;
} else if (timeout_in > 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_in = max(timeout_in, (ep->connections * SEC_ULL) / 100);
}
if (timeout_in > 0 && timeout_in < PSMI_MIN_EP_CLOSE_TIMEOUT)
timeout_in = PSMI_MIN_EP_CLOSE_TIMEOUT;
/* Infinite and excessive close time-out are limited here to a max.
* The "rationale" is that there is no point waiting around forever for
* graceful termination. Normal (or forced) process termination should clean
* up the context state correctly even if termination is not graceful. */
if (timeout_in <= 0 || timeout_in < PSMI_MAX_EP_CLOSE_TIMEOUT)
timeout_in = PSMI_MAX_EP_CLOSE_TIMEOUT;
_HFI_PRDBG("Closing endpoint %p with force=%s and to=%.2f seconds and "
"%d connections\n",
ep, mode == PSM_EP_CLOSE_FORCE ? "YES" : "NO",
(double)timeout_in / 1e9, (int)ep->connections);
/* XXX We currently cheat in the sense that we leave each PTL the allowed
* timeout. There's no good way to do this until we change the PTL
* interface to allow asynchronous finalization
*/
mep = ep;
tmp = ep->mctxt_prev;
do {
ep = tmp;
tmp = ep->mctxt_prev;
PSM_MCTXT_REMOVE(ep);
if (psmi_ep_device_is_enabled(ep, PTL_DEVID_AMSH))
err =
psmi_ptl_amsh.fini(ep->ptl_amsh.ptl, mode,
timeout_in);
if ((err == PSM_OK || err == PSM_TIMEOUT) &&
psmi_ep_device_is_enabled(ep, PTL_DEVID_IPS))
err =
psmi_ptl_ips.fini(ep->ptl_ips.ptl, mode,
timeout_in);
/* If there's timeouts in the disconnect requests,
* still make sure that we still get to close the
*endpoint and mark it closed */
if (psmi_ep_device_is_enabled(ep, PTL_DEVID_IPS))
psmi_context_close(&ep->context);
psmi_free(ep->epaddr);
psmi_free(ep->context_mylabel);
/*
* Before freeing the master ep itself,
* remove it from the global linklist.
* We do it here to let atexit handler in ptl_am directory
* to search the global linklist and free the shared memory file.
*/
if (ep == mep) {
if (psmi_opened_endpoint == ep) {
psmi_opened_endpoint = ep->user_ep_next;
} else {
tmp = psmi_opened_endpoint;
while (tmp->user_ep_next != ep) {
tmp = tmp->user_ep_next;
}
tmp->user_ep_next = ep->user_ep_next;
}
psmi_opened_endpoint_count--;
}
psmi_free(ep);
} while ((err == PSM_OK || err == PSM_TIMEOUT) && tmp != ep);
PSMI_PUNLOCK();
_HFI_PRDBG("Closed endpoint in %.3f secs\n",
(double)cycles_to_nanosecs(get_cycles() -
t_start) / SEC_ULL);
return err;
}
psm_error_t
__psm_ep_open(psm_uuid_t const unique_job_key,
struct psm_ep_open_opts const *opts_i, psm_ep_t *epo,
psm_epid_t *epido)
{
psm_error_t err;
psm_mq_t mq;
psm_epid_t epid;
psm_ep_t ep, tmp;
uint32_t units[HFI_MAX_RAILS];
uint16_t ports[HFI_MAX_RAILS];
int i, num_rails = 0;
char *uname = "HFI_UNIT";
char *pname = "HFI_PORT";
char uvalue[4], pvalue[4];
int devid_enabled[PTL_MAX_INIT];
union psmi_envvar_val devs;
PSMI_ERR_UNLESS_INITIALIZED(NULL);
/* Currently only one endpoint is supported. */
if (psmi_opened_endpoint_count > 0)
return PSM_TOO_MANY_ENDPOINTS;
PSMI_PLOCK();
/* Matched Queue initialization. We do this early because we have to
* make sure ep->mq exists and is valid before calling ips_do_work.
*/
err = psmi_mq_malloc(&mq);
if (err != PSM_OK)
goto fail;
/* See which ptl devices we want to use for this ep to be opened */
psmi_getenv("PSM_DEVICES",
"Ordered list of PSM-level devices",
PSMI_ENVVAR_LEVEL_USER,
PSMI_ENVVAR_TYPE_STR,
(union psmi_envvar_val)PSMI_DEVICES_DEFAULT, &devs);
if ((err = psmi_parse_devices(devid_enabled, devs.e_str)))
goto fail;
if (psmi_device_is_enabled(devid_enabled, PTL_DEVID_IPS)) {
err = psmi_ep_multirail(&num_rails, units, ports);
if (err != PSM_OK)
goto fail;
/* If multi-rail is used, set the first ep unit/port */
if (num_rails > 0) {
snprintf(uvalue, 4, "%1d", units[0]);
snprintf(pvalue, 4, "%1d", ports[0]);
setenv(uname, uvalue, 1);
setenv(pname, pvalue, 1);
}
}
err = __psm_ep_open_internal(unique_job_key,
devid_enabled, opts_i, mq, &ep, &epid);
if (err != PSM_OK)
goto fail;
if (psmi_opened_endpoint == NULL) {
psmi_opened_endpoint = ep;
} else {
tmp = psmi_opened_endpoint;
while (tmp->user_ep_next)
tmp = tmp->user_ep_next;
tmp->user_ep_next = ep;
}
psmi_opened_endpoint_count++;
ep->mctxt_prev = ep->mctxt_next = ep;
ep->mctxt_master = ep;
mq->ep = ep;
/* Active Message initialization */
err = psmi_am_init_internal(ep);
if (err != PSM_OK)
goto fail;
*epo = ep;
*epido = epid;
if (psmi_device_is_enabled(devid_enabled, PTL_DEVID_IPS)) {
for (i = 1; i < num_rails; i++) {
snprintf(uvalue, 4, "%1d", units[i]);
snprintf(pvalue, 4, "%1d", ports[i]);
setenv(uname, uvalue, 1);
setenv(pname, pvalue, 1);
/* Create slave EP */
err = __psm_ep_open_internal(unique_job_key,
devid_enabled, opts_i, mq,
&tmp, &epid);
if (err)
goto fail;
/* Point back to shared resources on the master EP */
tmp->am_htable = ep->am_htable;
/* Link slave EP after master EP. */
PSM_MCTXT_APPEND(ep, tmp);
}
}
/* Once we've initialized all devices, we can update the MQ with its
* default values */
if (err == PSM_OK)
err = psmi_mq_initialize_defaults(mq);
_HFI_VDBG("psm_ep_open() OK....\n");
fail:
PSMI_PUNLOCK();
return err;
}
psm_error_t
__psm_ep_open_internal(psm_uuid_t const unique_job_key, int *devid_enabled,
struct psm_ep_open_opts const *opts_i, psm_mq_t mq,
psm_ep_t *epo, psm_epid_t *epido)
{
psm_ep_t ep = NULL;
uint32_t num_units;
size_t len;
psm_error_t err;
psm_epaddr_t epaddr = NULL;
char buf[128], *p, *e;
union psmi_envvar_val envvar_val;
size_t ptl_sizes;
struct psm_ep_open_opts opts;
ptl_t *amsh_ptl, *ips_ptl, *self_ptl;
int i;
/* First get the set of default options, we overwrite with the user's
* desired values afterwards */
if ((err = psm_ep_open_opts_get_defaults(&opts)))
goto fail;
if (opts_i != NULL) {
if (opts_i->timeout != -1)
opts.timeout = opts_i->timeout;
if (opts_i->unit != -1)
opts.unit = opts_i->unit;
if (opts_i->affinity != -1)
opts.affinity = opts_i->affinity;
if (opts_i->sendbufs_num != -1)
opts.sendbufs_num = opts_i->sendbufs_num;
if (opts_i->network_pkey != HFI_DEFAULT_P_KEY)
opts.network_pkey = opts_i->network_pkey;
if (opts_i->port != 0)
opts.port = opts_i->port;
if (opts_i->outsl != -1)
opts.outsl = opts_i->outsl;
if (opts_i->service_id)
opts.service_id = (uint64_t) opts_i->service_id;
if (opts_i->path_res_type != PSM_PATH_RES_NONE)
opts.path_res_type = opts_i->path_res_type;
if (opts_i->senddesc_num)
opts.senddesc_num = opts_i->senddesc_num;
if (opts_i->imm_size)
opts.imm_size = opts_i->imm_size;
}
/* Get Service ID from environment */
if (!psmi_getenv("PSM_IB_SERVICE_ID",
"HFI Service ID for path resolution",
PSMI_ENVVAR_LEVEL_USER,
PSMI_ENVVAR_TYPE_ULONG_ULONG,
(union psmi_envvar_val)HFI_DEFAULT_SERVICE_ID,
&envvar_val)) {
opts.service_id = (uint64_t) envvar_val.e_ulonglong;
}
/* Get Path resolution type from environment Possible choices are:
*
* NONE : Default same as previous instances. Utilizes static data.
* OPP : Use OFED Plus Plus library to do path record queries.
* UMAD : Use raw libibumad interface to form and process path records.
*/
if (!psmi_getenv("PSM_PATH_REC",
"Mechanism to query HFI path record (default is no path query)",
PSMI_ENVVAR_LEVEL_USER, PSMI_ENVVAR_TYPE_STR,
(union psmi_envvar_val)"none", &envvar_val)) {
if (!strcasecmp(envvar_val.e_str, "none"))
opts.path_res_type = PSM_PATH_RES_NONE;
else if (!strcasecmp(envvar_val.e_str, "opp"))
opts.path_res_type = PSM_PATH_RES_OPP;
else if (!strcasecmp(envvar_val.e_str, "umad"))
opts.path_res_type = PSM_PATH_RES_UMAD;
else {
_HFI_ERROR("Unknown path resolution type %s. "
"Disabling use of path record query.\n",
envvar_val.e_str);
opts.path_res_type = PSM_PATH_RES_NONE;
}
}
/* If a specific unit is set in the environment, use that one. */
if (!psmi_getenv("HFI_UNIT", "Device Unit number (-1 autodetects)",
PSMI_ENVVAR_LEVEL_USER, PSMI_ENVVAR_TYPE_LONG,
(union psmi_envvar_val)HFI_UNIT_ID_ANY, &envvar_val)) {
opts.unit = envvar_val.e_long;
}
/* Get user specified port number to use. */
if (!psmi_getenv("HFI_PORT", "IB Port number (0 autodetects)",
PSMI_ENVVAR_LEVEL_USER, PSMI_ENVVAR_TYPE_LONG,
(union psmi_envvar_val)HFI_PORT_NUM_ANY,
&envvar_val)) {
opts.port = envvar_val.e_long;
}
/* Get service level from environment, path-query overrides it */
if (!psmi_getenv
("HFI_SL", "HFI outging ServiceLevel number (default 0)",
PSMI_ENVVAR_LEVEL_USER, PSMI_ENVVAR_TYPE_LONG,
(union psmi_envvar_val)PSMI_SL_DEFAULT, &envvar_val)) {
opts.outsl = envvar_val.e_long;
}
/* Get network key from environment. MVAPICH and other vendor MPIs do not
* specify it on ep open and we may require it for vFabrics.
* path-query will override it.
*/
if (!psmi_getenv("PSM_PKEY",
"HFI PKey to use for endpoint",
PSMI_ENVVAR_LEVEL_USER,
PSMI_ENVVAR_TYPE_ULONG,
(union psmi_envvar_val)HFI_DEFAULT_P_KEY,
&envvar_val)) {
opts.network_pkey = (uint64_t) envvar_val.e_ulong;
}
/* BACKWARDS COMPATIBILITY: Open MPI likes to choose its own PKEY of
0x7FFF. That's no longer a valid default, so override it if the
client was compiled against PSM v1 */
if (PSMI_VERNO_GET_MAJOR(psmi_verno_client()) < 2 &&
opts.network_pkey == 0x7FFF) {
opts.network_pkey = HFI_DEFAULT_P_KEY;
}
/* Get number of default send buffers from environment */
if (!psmi_getenv("PSM_NUM_SEND_BUFFERS",
"Number of send buffers to allocate [1024]",
PSMI_ENVVAR_LEVEL_USER,
PSMI_ENVVAR_TYPE_UINT,
(union psmi_envvar_val)1024, &envvar_val)) {
opts.sendbufs_num = envvar_val.e_uint;
}
/* Get immediate data size - transfers less than immediate data size do
* not consume a send buffer and require just a send descriptor.
*/
if (!psmi_getenv("PSM_SEND_IMMEDIATE_SIZE",
"Immediate data send size not requiring a buffer [128]",
PSMI_ENVVAR_LEVEL_USER,
PSMI_ENVVAR_TYPE_UINT,
(union psmi_envvar_val)128, &envvar_val)) {
opts.imm_size = envvar_val.e_uint;
}
/* Get numner of send descriptors - by default this is 4 times the number
* of send buffers - mainly used for short/inlined messages.
*/
if (!psmi_getenv("PSM_NUM_SEND_DESCRIPTORS",
"Number of send descriptors to allocate [4096]",
PSMI_ENVVAR_LEVEL_USER,
PSMI_ENVVAR_TYPE_UINT,
(union psmi_envvar_val)4096, &envvar_val)) {
opts.senddesc_num = envvar_val.e_uint;
}
if (psmi_device_is_enabled(devid_enabled, PTL_DEVID_IPS)) {
if ((err = psm_ep_num_devunits(&num_units)) != PSM_OK)
goto fail;
} else
num_units = 0;
/* do some error checking */
if (opts.timeout < -1) {
err = psmi_handle_error(NULL, PSM_PARAM_ERR,
"Invalid timeout value %lld",
(long long)opts.timeout);
goto fail;
} else if (num_units && (opts.unit < -1 || opts.unit >= (int)num_units)) {
err = psmi_handle_error(NULL, PSM_PARAM_ERR,
"Invalid Device Unit ID %d (%d units found)",
opts.unit, num_units);
goto fail;
} else if (opts.port < 0 || opts.port > HFI_MAX_PORT) {
err = psmi_handle_error(NULL, PSM_PARAM_ERR,
"Invalid Device port number %d",
opts.port);
goto fail;
} else if (opts.affinity < 0
|| opts.affinity > PSM_EP_OPEN_AFFINITY_FORCE) {
err =
psmi_handle_error(NULL, PSM_PARAM_ERR,
"Invalid Affinity option: %d",
opts.affinity);
goto fail;
} else if (opts.outsl < PSMI_SL_MIN || opts.outsl > PSMI_SL_MAX) {
err = psmi_handle_error(NULL, PSM_PARAM_ERR,
"Invalid SL number: %lld",
(unsigned long long)opts.outsl);
goto fail;
}
/* Set environment variable if PSM is not allowed to set affinity */
if (opts.affinity == PSM_EP_OPEN_AFFINITY_SKIP)
setenv("HFI_NO_CPUAFFINITY", "1", 1);
/* Allocate end point structure storage */
ptl_sizes =
(psmi_device_is_enabled(devid_enabled, PTL_DEVID_SELF) ?
psmi_ptl_self.sizeof_ptl() : 0) +
(psmi_device_is_enabled(devid_enabled, PTL_DEVID_IPS) ?
psmi_ptl_ips.sizeof_ptl() : 0) +
(psmi_device_is_enabled(devid_enabled, PTL_DEVID_AMSH) ?
psmi_ptl_amsh.sizeof_ptl() : 0);
if (ptl_sizes == 0)
return PSM_EP_NO_DEVICE;
ep = (psm_ep_t) psmi_memalign(PSMI_EP_NONE, UNDEFINED, 64,
sizeof(struct psm_ep) + ptl_sizes);
epaddr = (psm_epaddr_t) psmi_calloc(PSMI_EP_NONE, PER_PEER_ENDPOINT,
1, sizeof(struct psm_epaddr));
if (ep == NULL || epaddr == NULL) {
err = psmi_handle_error(NULL, PSM_NO_MEMORY,
"Couldn't allocate memory for %s structure",
ep == NULL ? "psm_ep" : "psm_epaddr");
goto fail;
}
/* Copy PTL enabled status */
for (i = 0; i < PTL_MAX_INIT; i++)
ep->devid_enabled[i] = devid_enabled[i];
/* Matched Queue initialization. We do this early because we have to
* make sure ep->mq exists and is valid before calling ips_do_work.
*/
ep->mq = mq;
/* Get ready for PTL initialization */
memcpy(&ep->uuid, (void *)unique_job_key, sizeof(psm_uuid_t));
ep->epaddr = epaddr;
ep->memmode = mq->memmode;
ep->hfi_num_sendbufs = opts.sendbufs_num;
ep->service_id = opts.service_id;
ep->path_res_type = opts.path_res_type;
ep->hfi_num_descriptors = opts.senddesc_num;
ep->hfi_imm_size = opts.imm_size;
ep->errh = psmi_errhandler_global; /* by default use the global one */
ep->ptl_amsh.ep_poll = psmi_poll_noop;
ep->ptl_ips.ep_poll = psmi_poll_noop;
ep->connections = 0;
/* See how many iterations we want to spin before yielding */
psmi_getenv("PSM_YIELD_SPIN_COUNT",
"Spin poll iterations before yield",
PSMI_ENVVAR_LEVEL_HIDDEN,
PSMI_ENVVAR_TYPE_UINT,
(union psmi_envvar_val)PSMI_BLOCKUNTIL_POLLS_BEFORE_YIELD,
&envvar_val);
ep->yield_spin_cnt = envvar_val.e_uint;
ptl_sizes = 0;
amsh_ptl = ips_ptl = self_ptl = NULL;
if (psmi_ep_device_is_enabled(ep, PTL_DEVID_AMSH)) {
amsh_ptl = (ptl_t *) (ep->ptl_base_data + ptl_sizes);
ptl_sizes += psmi_ptl_amsh.sizeof_ptl();
}
if (psmi_ep_device_is_enabled(ep, PTL_DEVID_IPS)) {
ips_ptl = (ptl_t *) (ep->ptl_base_data + ptl_sizes);
ptl_sizes += psmi_ptl_ips.sizeof_ptl();
}
if (psmi_ep_device_is_enabled(ep, PTL_DEVID_SELF)) {
self_ptl = (ptl_t *) (ep->ptl_base_data + ptl_sizes);
ptl_sizes += psmi_ptl_self.sizeof_ptl();
}
if ((err = psmi_ep_open_device(ep, &opts, unique_job_key,
&(ep->context), &ep->epid)))
goto fail;
psmi_assert_always(ep->epid != 0);
ep->epaddr->epid = ep->epid;
_HFI_VDBG("psmi_ep_open_device() passed\n");
/* Set our new label as soon as we know what it is */
strncpy(buf, psmi_gethostname(), sizeof(buf) - 1);
buf[sizeof(buf) - 1] = '\0';
p = buf + strlen(buf);
/* If our rank is set, use it. If not, use context.subcontext notation */
if (((e = getenv("MPI_RANKID")) != NULL && *e) ||
((e = getenv("PSC_MPI_RANK")) != NULL && *e))
len = snprintf(p, sizeof(buf) - strlen(buf), ":%d.", atoi(e));
else
len = snprintf(p, sizeof(buf) - strlen(buf), ":%d.%d.",
(uint32_t) psm_epid_context(ep->epid),
(uint32_t) psmi_epid_subcontext(ep->epid));
*(p + len) = '\0';
ep->context_mylabel = psmi_strdup(ep, buf);
if (ep->context_mylabel == NULL) {
err = PSM_NO_MEMORY;
goto fail;
}
/* hfi_set_mylabel(ep->context_mylabel); */
if ((err = psmi_epid_set_hostname(psm_epid_nid(ep->epid), buf, 0)))
goto fail;
_HFI_VDBG("start ptl device init...\n");
if (psmi_ep_device_is_enabled(ep, PTL_DEVID_SELF)) {
if ((err = psmi_ptl_self.init(ep, self_ptl, &ep->ptl_self)))
goto fail;
}
if (psmi_ep_device_is_enabled(ep, PTL_DEVID_IPS)) {
if ((err = psmi_ptl_ips.init(ep, ips_ptl, &ep->ptl_ips)))
goto fail;
}
/* If we're shm-only, this device is enabled above */
if (psmi_ep_device_is_enabled(ep, PTL_DEVID_AMSH)) {
if ((err = psmi_ptl_amsh.init(ep, amsh_ptl, &ep->ptl_amsh)))
goto fail;
} else {
/* We may have pre-attached as part of getting our rank for enabling
* shared contexts. */
}
_HFI_VDBG("finish ptl device init...\n");
/*
* Keep only IPS since only IPS support multi-rail, other devices
* are only setup once. IPS device can come to this function again.
*/
for (i = 0; i < PTL_MAX_INIT; i++) {
if (devid_enabled[i] != PTL_DEVID_IPS) {
devid_enabled[i] = -1;
}
}
*epido = ep->epid;
*epo = ep;
return PSM_OK;
fail:
if (ep != NULL) {
if (ep->context.fd != -1)
close(ep->context.fd);
psmi_free(ep);
}
if (epaddr != NULL)
psmi_free(epaddr);
return err;
}
static
psm_error_t
psmi_ep_open_device(const psm_ep_t ep,
const struct psm_ep_open_opts *opts,
const psm_uuid_t unique_job_key,
struct psmi_context *context, psm_epid_t *epid)
{
psm_error_t err = PSM_OK;
/* Skip affinity. No affinity if:
* 1. User explicitly sets no-affinity=YES in environment.
* 2. User doesn't set affinity in environment and PSM is opened with
* option affinity skip.
*/
if (psmi_ep_device_is_enabled(ep, PTL_DEVID_IPS)) {
uint32_t rcvthread_flags;
union psmi_envvar_val env_rcvthread;
static int norcvthread; /* only for first rail */
ep->out_sl = opts->outsl;
if ((err =
psmi_context_open(ep, opts->unit, opts->port,
unique_job_key, opts->timeout,
context)) != PSM_OK)
goto fail;
_HFI_DBG("[%d]use unit %d port %d\n", getpid(),
context->ctrl->__hfi_unit, 1);
/* At this point, we have the unit id and port number, so
* check if pkey is not 0x0/0x7fff/0xffff, and match one
* of the pkey in table.
*/
if ((err =
psmi_ep_verify_pkey(ep, (uint16_t) opts->network_pkey,
&ep->network_pkey)) != PSM_OK)
goto fail;
/* See if we want to activate support for receive thread */
psmi_getenv("PSM_RCVTHREAD",
"Recv thread flags (0 disables thread)",
PSMI_ENVVAR_LEVEL_USER, PSMI_ENVVAR_TYPE_UINT_FLAGS,
(union psmi_envvar_val)(norcvthread++ ? 0 :
PSMI_RCVTHREAD_FLAGS),
&env_rcvthread);
rcvthread_flags = env_rcvthread.e_uint;
/* If enabled, use the pollurg capability to implement a receive
* interrupt thread that can handle urg packets */
if (rcvthread_flags) {
context->runtime_flags |= PSMI_RUNTIME_RCVTHREAD;
#ifdef PSMI_PLOCK_IS_NOLOCK
psmi_handle_error(PSMI_EP_NORETURN, PSM_INTERNAL_ERR,
"#define PSMI_PLOCK_IS_NOLOCK not functional yet "
"with RCVTHREAD on");
#endif
}
context->rcvthread_flags = rcvthread_flags;
*epid = context->epid;
} else if (psmi_ep_device_is_enabled(ep, PTL_DEVID_AMSH)) {
int rank;
/* In shm-only mode, we need to derive a valid epid
* based on our rank. We try to get it from the
* environment if its available, or resort to using
* our PID as the rank.
*/
union psmi_envvar_val env_rankid;
if (psmi_getenv
("MPI_LOCALRANKID", "Shared context rankid",
PSMI_ENVVAR_LEVEL_HIDDEN, PSMI_ENVVAR_TYPE_INT,
(union psmi_envvar_val)-1, &env_rankid)) {
if (psmi_getenv
("PSC_MPI_NODE_RANK",
"Shared context rankid",
PSMI_ENVVAR_LEVEL_HIDDEN,
PSMI_ENVVAR_TYPE_INT,
(union psmi_envvar_val)-1, &env_rankid)) {
rank = getpid();
} else
rank = env_rankid.e_int;
} else
rank = env_rankid.e_int;
/*
* We use a LID of 0 for non-HFI communication.
* Since a jobkey is not available from IPS, pull the
* first 16 bits from the UUID.
*/
*epid = PSMI_EPID_PACK(((uint16_t *) unique_job_key)[0],
(rank >> 3), rank, 0,
PSMI_HFI_TYPE_DEFAULT, 0x7, rank);
} else {