static ucs_status_t uct_perf_create_pd(ucx_perf_context_t *perf)
{
uct_pd_resource_desc_t *pd_resources;
uct_tl_resource_desc_t *tl_resources;
unsigned i, num_pd_resources;
unsigned j, num_tl_resources;
ucs_status_t status;
uct_pd_h pd;
uct_pd_config_t *pd_config;
status = uct_query_pd_resources(&pd_resources, &num_pd_resources);
if (status != UCS_OK) {
goto out;
}
for (i = 0; i < num_pd_resources; ++i) {
status = uct_pd_config_read(pd_resources[i].pd_name, NULL, NULL, &pd_config);
if (status != UCS_OK) {
goto out_release_pd_resources;
}
status = uct_pd_open(pd_resources[i].pd_name, pd_config, &pd);
uct_config_release(pd_config);
if (status != UCS_OK) {
goto out_release_pd_resources;
}
status = uct_pd_query_tl_resources(pd, &tl_resources, &num_tl_resources);
if (status != UCS_OK) {
uct_pd_close(pd);
goto out_release_pd_resources;
}
for (j = 0; j < num_tl_resources; ++j) {
if (!strcmp(perf->params.uct.tl_name, tl_resources[j].tl_name) &&
!strcmp(perf->params.uct.dev_name, tl_resources[j].dev_name))
{
uct_release_tl_resource_list(tl_resources);
perf->uct.pd = pd;
status = UCS_OK;
goto out_release_pd_resources;
}
}
uct_pd_close(pd);
uct_release_tl_resource_list(tl_resources);
}
ucs_error("Cannot use transport %s on device %s", perf->params.uct.tl_name,
perf->params.uct.dev_name);
status = UCS_ERR_NO_DEVICE;
out_release_pd_resources:
uct_release_pd_resource_list(pd_resources);
out:
return status;
}
/* Device and transport to be used are determined by minimum latency */
static ucs_status_t dev_tl_lookup()
{
int i;
int j;
uint64_t min_latency = UINT64_MAX;
int pd_index = -1;
int tl_index = -1;
ucs_status_t status;
uct_pd_resource_desc_t *pd_resources; /* Protection domain resource descriptor */
uct_tl_resource_desc_t *tl_resources; /*Communication resource descriptor */
unsigned num_pd_resources; /* Number of protected domain */
unsigned num_tl_resources; /* Number of transport resources resource objects created */
status = uct_query_pd_resources(&pd_resources, &num_pd_resources);
if (UCS_OK != status) {
fprintf(stderr, "Failed to query for protected domain resources.\n");
goto out1;
}
/* Iterate through protected domain resources */
for (i = 0; i < num_pd_resources; ++i) {
status = uct_pd_open(pd_resources[i].pd_name, &pd);
if (UCS_OK != status) {
fprintf(stderr, "Failed to open protected domain.\n"); fflush(stderr);
goto release1;
}
status = uct_pd_query_tl_resources(pd, &tl_resources, &num_tl_resources);
if (UCS_OK != status) {
fprintf(stderr, "Failed to query transport resources.\n"); fflush(stderr);
uct_pd_close(pd);
goto release1;
}
/* Go through each available transport resource for a particular protected domain
* and keep track of the fastest latency */
for (j = 0; j < num_tl_resources; ++j) {
status = resource_supported(tl_resources[j].dev_name, tl_resources[j].tl_name, 1);
if (UCS_OK == status) {
if (tl_resources[j].latency < min_latency) {
min_latency = tl_resources[j].latency;
pd_index = i;
tl_index = j;
}
}
}
uct_release_tl_resource_list(tl_resources);
uct_pd_close(pd);
}
/* Check if any valid device/transport found */
if ((-1 == pd_index) || (-1 == tl_index)) {
uct_release_pd_resource_list(pd_resources);
return UCS_ERR_UNSUPPORTED;
}
/* IMPORTANT: Certain functions that operate on an interface rely on a pointer to the protection domain that created it */
/* Reopen new protection domain and */
status = uct_pd_open(pd_resources[pd_index].pd_name, &pd);
if (UCS_OK != status) {
fprintf(stderr, "Failed to open final protected domain.\n"); fflush(stderr);
goto release1;
}
/* Open new tranport resources */
status = uct_pd_query_tl_resources(pd, &tl_resources, &num_tl_resources);
if (UCS_OK != status) {
fprintf(stderr, "Failed to query final transport resources.\n"); fflush(stderr);
uct_pd_close(pd);
goto release1;
}
/* Call resource_supported() again to set the interface */
status = resource_supported(tl_resources[tl_index].dev_name, tl_resources[tl_index].tl_name, 0);
if (UCS_OK != status) {
fprintf(stderr, "Failed to initialize final interface.\n"); fflush(stderr);
uct_pd_close(pd);
return status;
}
printf("Using %s with %s.\n", tl_resources[tl_index].dev_name, tl_resources[tl_index].tl_name);fflush(stdout);
uct_release_tl_resource_list(tl_resources);
release1:
uct_release_pd_resource_list(pd_resources);
out1:
return status;
}
static ucs_status_t ucp_fill_resources(ucp_context_h context,
const ucp_config_t *config)
{
unsigned num_tl_resources;
unsigned num_pd_resources;
uct_pd_resource_desc_t *pd_rscs;
ucs_status_t status;
ucp_rsc_index_t i;
unsigned pd_index;
uct_pd_h pd;
/* if we got here then num_resources > 0.
* if the user's device list is empty, there is no match */
if (0 == config->devices.count) {
ucs_error("The device list is empty. Please specify the devices you would like to use "
"or omit the UCX_DEVICES so that the default will be used.");
status = UCS_ERR_NO_ELEM;
goto err;
}
/* if we got here then num_resources > 0.
* if the user's tls list is empty, there is no match */
if (0 == config->tls.count) {
ucs_error("The TLs list is empty. Please specify the transports you would like to use "
"or omit the UCX_TLS so that the default will be used.");
status = UCS_ERR_NO_ELEM;
goto err;
}
/* List protection domain resources */
status = uct_query_pd_resources(&pd_rscs, &num_pd_resources);
if (status != UCS_OK) {
goto err;
}
/* Error check: Make sure there is at least one PD */
if (num_pd_resources == 0) {
ucs_error("No pd resources found");
status = UCS_ERR_NO_DEVICE;
goto err_release_pd_resources;
}
if (num_pd_resources >= UCP_MAX_PDS) {
ucs_error("Only up to %ld PDs are supported", UCP_MAX_PDS);
status = UCS_ERR_EXCEEDS_LIMIT;
goto err_release_pd_resources;
}
context->num_pds = 0;
context->pd_rscs = NULL;
context->pds = NULL;
context->pd_attrs = NULL;
context->num_tls = 0;
context->tl_rscs = NULL;
/* Allocate array of PD resources we would actually use */
context->pd_rscs = ucs_calloc(num_pd_resources, sizeof(*context->pd_rscs),
"ucp_pd_resources");
if (context->pd_rscs == NULL) {
status = UCS_ERR_NO_MEMORY;
goto err_free_context_resources;
}
/* Allocate array of protection domains */
context->pds = ucs_calloc(num_pd_resources, sizeof(*context->pds), "ucp_pds");
if (context->pds == NULL) {
status = UCS_ERR_NO_MEMORY;
goto err_free_context_resources;
}
/* Allocate array of protection domains attributes */
context->pd_attrs = ucs_calloc(num_pd_resources, sizeof(*context->pd_attrs),
"ucp_pd_attrs");
if (context->pd_attrs == NULL) {
status = UCS_ERR_NO_MEMORY;
goto err_free_context_resources;
}
/* Open all protection domains, keep only those which have at least one TL
* resources selected on them.
*/
pd_index = 0;
for (i = 0; i < num_pd_resources; ++i) {
status = uct_pd_open(pd_rscs[i].pd_name, &pd);
if (status != UCS_OK) {
goto err_free_context_resources;
}
context->pd_rscs[pd_index] = pd_rscs[i];
context->pds[pd_index] = pd;
/* Save PD attributes */
status = uct_pd_query(pd, &context->pd_attrs[pd_index]);
if (status != UCS_OK) {
goto err_free_context_resources;
}
/* Add communication resources of each PD */
status = ucp_add_tl_resources(context, pd, pd_index, config,
&num_tl_resources);
if (status != UCS_OK) {
goto err_free_context_resources;
}
/* If the PD does not have transport resources, don't use it */
if (num_tl_resources > 0) {
++pd_index;
++context->num_pds;
} else {
ucs_debug("closing pd %s because it has no selected transport resources",
pd_rscs[i].pd_name);
uct_pd_close(pd);
}
}
/* Error check: Make sure there is at least one transport */
if (0 == context->num_tls) {
ucs_error("There are no available resources matching the configured criteria");
status = UCS_ERR_NO_DEVICE;
goto err_free_context_resources;
}
/* Error check: Make sure there are no too many transports */
if (context->num_tls >= UCP_MAX_TLS) {
ucs_error("Exceeded resources limit (%u requested, up to %d are supported)",
context->num_tls, UCP_MAX_TLS);
status = UCS_ERR_EXCEEDS_LIMIT;
goto err_free_context_resources;
}
uct_release_pd_resource_list(pd_rscs);
return UCS_OK;
err_free_context_resources:
ucp_free_resources(context);
err_release_pd_resources:
uct_release_pd_resource_list(pd_rscs);
err:
return status;
}
