static UCS_CLASS_INIT_FUNC(uct_self_iface_t, uct_md_h md, uct_worker_h worker,
const uct_iface_params_t *params,
const uct_iface_config_t *tl_config)
{
ucs_status_t status;
uct_self_iface_config_t *self_config = 0;
ucs_trace_func("Creating a loop-back transport self=%p rxh=%lu",
self, params->rx_headroom);
if (strcmp(params->dev_name, UCT_SELF_NAME) != 0) {
ucs_error("No device was found: %s", params->dev_name);
return UCS_ERR_NO_DEVICE;
}
UCS_CLASS_CALL_SUPER_INIT(uct_base_iface_t, &uct_self_iface_ops, md, worker,
tl_config UCS_STATS_ARG(params->stats_root)
UCS_STATS_ARG(UCT_SELF_NAME));
self_config = ucs_derived_of(tl_config, uct_self_iface_config_t);
self->id = ucs_generate_uuid((uintptr_t)self);
self->rx_headroom = params->rx_headroom;
self->data_length = self_config->super.max_bcopy;
self->release_desc.cb = uct_self_iface_release_desc;
/* create a memory pool for data transferred */
status = uct_iface_mpool_init(&self->super,
&self->msg_desc_mp,
sizeof(uct_recv_desc_t) + self->rx_headroom +
self->data_length,
sizeof(uct_recv_desc_t) + self->rx_headroom,
UCS_SYS_CACHE_LINE_SIZE,
&self_config->mp,
256,
ucs_empty_function,
"self_msg_desc");
if (UCS_OK != status) {
ucs_error("Failed to create a memory pool for the loop-back transport");
goto err;
}
/* set the message descriptor for the loop-back */
self->msg_cur_desc = ucs_mpool_get(&self->msg_desc_mp);
VALGRIND_MAKE_MEM_DEFINED(self->msg_cur_desc, sizeof(*(self->msg_cur_desc)));
if (NULL == self->msg_cur_desc) {
ucs_error("Failed to get the first descriptor in loop-back MP storage");
status = UCS_ERR_NO_RESOURCE;
goto destroy_mpool;
}
ucs_debug("Created a loop-back iface. id=0x%lx, desc=%p, len=%u, tx_hdr=%lu",
self->id, self->msg_cur_desc, self->data_length, self->rx_headroom);
return UCS_OK;
destroy_mpool:
ucs_mpool_cleanup(&self->msg_desc_mp, 1);
err:
return status;
}
ucs_status_t uct_rc_verbs_iface_common_init(uct_rc_verbs_iface_common_t *iface,
uct_rc_iface_t *rc_iface,
uct_rc_verbs_iface_common_config_t *config,
uct_rc_iface_config_t *rc_config)
{
ucs_status_t status;
size_t am_hdr_size;
memset(iface->inl_sge, 0, sizeof(iface->inl_sge));
/* Configuration */
am_hdr_size = ucs_max(config->max_am_hdr, sizeof(uct_rc_hdr_t));
iface->config.short_desc_size = ucs_max(UCT_RC_MAX_ATOMIC_SIZE, am_hdr_size);
/* Create AM headers and Atomic mempool */
status = uct_iface_mpool_init(&rc_iface->super.super,
&iface->short_desc_mp,
sizeof(uct_rc_iface_send_desc_t) +
iface->config.short_desc_size,
sizeof(uct_rc_iface_send_desc_t),
UCS_SYS_CACHE_LINE_SIZE,
&rc_config->super.tx.mp,
rc_iface->config.tx_qp_len,
uct_rc_iface_send_desc_init,
"rc_verbs_short_desc");
if (status != UCS_OK) {
return status;
}
return UCS_OK;
}
ucs_status_t uct_ib_iface_recv_mpool_init(uct_ib_iface_t *iface,
const uct_ib_iface_config_t *config,
const char *name, ucs_mpool_t *mp)
{
unsigned grow;
if (config->rx.queue_len < 1024) {
grow = 1024;
} else {
/* We want to have some free (+10%) elements to avoid mem pool expansion */
grow = ucs_min( (int)(1.1 * config->rx.queue_len + 0.5),
config->rx.mp.max_bufs);
}
return uct_iface_mpool_init(&iface->super, mp,
iface->config.rx_payload_offset + iface->config.seg_size,
iface->config.rx_hdr_offset,
UCS_SYS_CACHE_LINE_SIZE,
&config->rx.mp, grow,
uct_ib_iface_recv_desc_init,
name);
}
ucs_status_t uct_rc_verbs_iface_common_init(uct_rc_verbs_iface_common_t *iface,
uct_rc_iface_t *rc_iface,
uct_rc_verbs_iface_common_config_t *config,
uct_rc_iface_config_t *rc_config,
size_t max_hdr_size)
{
memset(iface->inl_sge, 0, sizeof(iface->inl_sge));
ucs_status_t status;
/* Configuration */
iface->config.short_desc_size = ucs_max(UCT_RC_MAX_ATOMIC_SIZE, max_hdr_size);
/* Create AM headers and Atomic mempool */
status = uct_iface_mpool_init(&rc_iface->super.super,
&iface->short_desc_mp,
sizeof(uct_rc_iface_send_desc_t) +
iface->config.short_desc_size,
sizeof(uct_rc_iface_send_desc_t),
UCS_SYS_CACHE_LINE_SIZE,
&rc_config->super.tx.mp,
rc_iface->config.tx_qp_len,
uct_rc_iface_send_desc_init,
"rc_verbs_short_desc");
if (status != UCS_OK) {
return status;
}
iface->config.notag_hdr_size = 0;
iface->am_inl_hdr = ucs_mpool_get(&iface->short_desc_mp);
if (iface->am_inl_hdr == NULL) {
ucs_error("Failed to allocate AM short header");
return UCS_ERR_NO_MEMORY;
}
return UCS_OK;
}
static UCS_CLASS_INIT_FUNC(uct_ugni_rdma_iface_t, uct_pd_h pd, uct_worker_h worker,
const char *dev_name, size_t rx_headroom,
const uct_iface_config_t *tl_config)
{
uct_ugni_rdma_iface_config_t *config = ucs_derived_of(tl_config, uct_ugni_rdma_iface_config_t);
ucs_status_t status;
pthread_mutex_lock(&uct_ugni_global_lock);
UCS_CLASS_CALL_SUPER_INIT(uct_ugni_iface_t, pd, worker, dev_name, &uct_ugni_rdma_iface_ops,
&config->super UCS_STATS_ARG(NULL));
/* Setting initial configuration */
self->config.fma_seg_size = UCT_UGNI_MAX_FMA;
self->config.rdma_max_size = UCT_UGNI_MAX_RDMA;
status = ucs_mpool_init(&self->free_desc,
0,
sizeof(uct_ugni_base_desc_t),
0, /* alignment offset */
UCS_SYS_CACHE_LINE_SIZE, /* alignment */
128, /* grow */
config->mpool.max_bufs, /* max buffers */
&uct_ugni_rdma_desc_mpool_ops,
"UGNI-DESC-ONLY");
if (UCS_OK != status) {
ucs_error("Mpool creation failed");
goto exit;
}
status = ucs_mpool_init(&self->free_desc_get,
0,
sizeof(uct_ugni_rdma_fetch_desc_t),
0, /* alignment offset */
UCS_SYS_CACHE_LINE_SIZE, /* alignment */
128 , /* grow */
config->mpool.max_bufs, /* max buffers */
&uct_ugni_rdma_desc_mpool_ops,
"UGNI-GET-DESC-ONLY");
if (UCS_OK != status) {
ucs_error("Mpool creation failed");
goto clean_desc;
}
status = ucs_mpool_init(&self->free_desc_buffer,
0,
sizeof(uct_ugni_base_desc_t) + self->config.fma_seg_size,
sizeof(uct_ugni_base_desc_t), /* alignment offset */
UCS_SYS_CACHE_LINE_SIZE, /* alignment */
128 , /* grow */
config->mpool.max_bufs, /* max buffers */
&uct_ugni_rdma_desc_mpool_ops,
"UGNI-DESC-BUFFER");
if (UCS_OK != status) {
ucs_error("Mpool creation failed");
goto clean_desc_get;
}
status = uct_iface_mpool_init(&self->super.super,
&self->free_desc_famo,
sizeof(uct_ugni_rdma_fetch_desc_t) + 8,
sizeof(uct_ugni_rdma_fetch_desc_t),/* alignment offset */
UCS_SYS_CACHE_LINE_SIZE, /* alignment */
&config->mpool, /* mpool config */
128 , /* grow */
uct_ugni_base_desc_key_init, /* memory/key init */
"UGNI-DESC-FAMO");
if (UCS_OK != status) {
ucs_error("Mpool creation failed");
goto clean_buffer;
}
status = uct_iface_mpool_init(&self->super.super,
&self->free_desc_get_buffer,
sizeof(uct_ugni_rdma_fetch_desc_t) +
self->config.fma_seg_size,
sizeof(uct_ugni_rdma_fetch_desc_t), /* alignment offset */
UCS_SYS_CACHE_LINE_SIZE, /* alignment */
&config->mpool, /* mpool config */
128 , /* grow */
uct_ugni_base_desc_key_init, /* memory/key init */
"UGNI-DESC-GET");
if (UCS_OK != status) {
ucs_error("Mpool creation failed");
goto clean_famo;
}
status = ugni_activate_iface(&self->super);
if (UCS_OK != status) {
ucs_error("Failed to activate the interface");
goto clean_get_buffer;
}
if(GNI_DEVICE_ARIES == self->super.dev->type) {
uct_ugni_rdma_iface_ops.ep_atomic_swap64 = uct_ugni_ep_atomic_swap64;
uct_ugni_rdma_iface_ops.ep_atomic_add32 = uct_ugni_ep_atomic_add32;
uct_ugni_rdma_iface_ops.ep_atomic_fadd32 = uct_ugni_ep_atomic_fadd32;
uct_ugni_rdma_iface_ops.ep_atomic_cswap32 = uct_ugni_ep_atomic_cswap32;
uct_ugni_rdma_iface_ops.ep_atomic_swap32 = uct_ugni_ep_atomic_swap32;
}
/* TBD: eventually the uct_ugni_progress has to be moved to
* rdma layer so each ugni layer will have own progress */
uct_worker_progress_register(worker, uct_ugni_progress, self);
pthread_mutex_unlock(&uct_ugni_global_lock);
return UCS_OK;
clean_get_buffer:
ucs_mpool_cleanup(&self->free_desc_get_buffer, 1);
clean_famo:
ucs_mpool_cleanup(&self->free_desc_famo, 1);
clean_buffer:
ucs_mpool_cleanup(&self->free_desc_buffer, 1);
clean_desc_get:
ucs_mpool_cleanup(&self->free_desc_get, 1);
clean_desc:
ucs_mpool_cleanup(&self->free_desc, 1);
exit:
ucs_error("Failed to activate interface");
pthread_mutex_unlock(&uct_ugni_global_lock);
return status;
}
