/*
* Look for an existing usnic proc based on a hashed RTE process
* name.
*/
ompi_btl_usnic_endpoint_t *
ompi_btl_usnic_proc_lookup_endpoint(ompi_btl_usnic_module_t *receiver,
uint64_t sender_hashed_rte_name)
{
ompi_btl_usnic_proc_t *proc;
ompi_btl_usnic_endpoint_t *endpoint;
opal_list_item_t *item;
MSGDEBUG1_OUT("lookup_endpoint: recvmodule=%p sendhash=0x%" PRIx64,
(void *)receiver, sender_hashed_rte_name);
for (item = opal_list_get_first(&receiver->all_endpoints);
item != opal_list_get_end(&receiver->all_endpoints);
item = opal_list_get_next(item)) {
endpoint = container_of(item, ompi_btl_usnic_endpoint_t,
endpoint_endpoint_li);
proc = endpoint->endpoint_proc;
if (ompi_rte_hash_name(&proc->proc_ompi->proc_name) ==
sender_hashed_rte_name) {
MSGDEBUG1_OUT("lookup_endpoint: matched endpoint=%p",
(void *)endpoint);
return endpoint;
}
}
/* Didn't find it */
return NULL;
}
/*
* Look for an existing usnic proc based on a hashed RTE process
* name.
*/
opal_btl_usnic_endpoint_t *
opal_btl_usnic_proc_lookup_endpoint(opal_btl_usnic_module_t *receiver,
uint64_t sender_proc_name)
{
opal_btl_usnic_proc_t *proc;
opal_btl_usnic_endpoint_t *endpoint;
opal_list_item_t *item;
MSGDEBUG1_OUT("lookup_endpoint: recvmodule=%p sendhash=0x%" PRIx64,
(void *)receiver, sender_hashed_rte_name);
opal_mutex_lock(&receiver->all_endpoints_lock);
for (item = opal_list_get_first(&receiver->all_endpoints);
item != opal_list_get_end(&receiver->all_endpoints);
item = opal_list_get_next(item)) {
endpoint = container_of(item, opal_btl_usnic_endpoint_t,
endpoint_endpoint_li);
proc = endpoint->endpoint_proc;
/* Note that this works today because opal_proc_t->proc_name
is unique across the universe. George is potentially
working to give handles instead of proc names, and then
have a function pointer to perform comparisons. This would
be bad here in the critical path, though... */
if (proc->proc_opal->proc_name == sender_proc_name) {
MSGDEBUG1_OUT("lookup_endpoint: matched endpoint=%p",
(void *)endpoint);
opal_mutex_unlock(&receiver->all_endpoints_lock);
return endpoint;
}
}
opal_mutex_unlock(&receiver->all_endpoints_lock);
/* Didn't find it */
return NULL;
}
/*
* Create an endpoint and claim the matched modex slot
*/
int
opal_btl_usnic_create_endpoint(opal_btl_usnic_module_t *module,
opal_btl_usnic_proc_t *proc,
opal_btl_usnic_endpoint_t **endpoint_o)
{
int err;
int modex_index;
opal_btl_usnic_endpoint_t *endpoint;
/* look for matching modex info */
err = match_modex(module, proc, &modex_index);
if (OPAL_SUCCESS != err) {
opal_output_verbose(5, USNIC_OUT,
"btl:usnic:create_endpoint: did not match usnic modex info for peer %s",
OPAL_NAME_PRINT(proc->proc_opal->proc_name));
return err;
}
endpoint = OBJ_NEW(opal_btl_usnic_endpoint_t);
if (NULL == endpoint) {
return OPAL_ERR_OUT_OF_RESOURCE;
}
/* Initalize the endpoint */
endpoint->endpoint_module = module;
assert(modex_index >= 0 && modex_index < (int)proc->proc_modex_count);
endpoint->endpoint_remote_addr = proc->proc_modex[modex_index];
/* Initialize endpoint sequence number info */
endpoint->endpoint_next_seq_to_send = module->local_addr.isn;
endpoint->endpoint_ack_seq_rcvd = endpoint->endpoint_next_seq_to_send - 1;
endpoint->endpoint_next_contig_seq_to_recv =
endpoint->endpoint_remote_addr.isn;
endpoint->endpoint_highest_seq_rcvd =
endpoint->endpoint_next_contig_seq_to_recv - 1;
endpoint->endpoint_rfstart = WINDOW_SIZE_MOD(endpoint->endpoint_next_contig_seq_to_recv);
/* Defer creating the ibv_ah. Since calling ibv_create_ah() may
trigger ARP resolution, it's better to batch all the endpoints'
calls to ibv_create_ah() together to get some parallelism. */
endpoint->endpoint_remote_ah = NULL;
/* Now claim that modex slot */
proc->proc_modex_claimed[modex_index] = true;
MSGDEBUG1_OUT("create_endpoint: module=%p claimed endpoint=%p on proc=%p (hash=0x%" PRIx64 ")\n",
(void *)module, (void *)endpoint, (void *)proc,
proc->proc_opal->proc_name);
/* Save the endpoint on this proc's array of endpoints */
proc->proc_endpoints[proc->proc_endpoint_count] = endpoint;
endpoint->endpoint_proc_index = proc->proc_endpoint_count;
endpoint->endpoint_proc = proc;
++proc->proc_endpoint_count;
OBJ_RETAIN(proc);
/* also add endpoint to module's list of endpoints (done here and
not in the endpoint constructor because we aren't able to pass
the module as a constructor argument -- doh!). */
opal_mutex_lock(&module->all_endpoints_lock);
opal_list_append(&(module->all_endpoints),
&(endpoint->endpoint_endpoint_li));
endpoint->endpoint_on_all_endpoints = true;
opal_mutex_unlock(&module->all_endpoints_lock);
*endpoint_o = endpoint;
return OPAL_SUCCESS;
}
/* Responsible for handling "large" frags (reserve + *size > max_frag_payload)
* in the same manner as btl_prepare_src. Must return a smaller amount than
* requested if the given convertor cannot process the entire (*size).
*/
static opal_btl_usnic_send_frag_t *
prepare_src_large(
struct opal_btl_usnic_module_t* module,
struct mca_btl_base_endpoint_t* endpoint,
struct opal_convertor_t* convertor,
uint8_t order,
size_t reserve,
size_t* size,
uint32_t flags)
{
opal_btl_usnic_send_frag_t *frag;
opal_btl_usnic_large_send_frag_t *lfrag;
int rc;
/* Get holder for the msg */
lfrag = opal_btl_usnic_large_send_frag_alloc(module);
if (OPAL_UNLIKELY(NULL == lfrag)) {
return NULL;
}
frag = &lfrag->lsf_base;
/* The header location goes in SG[0], payload in SG[1]. If we are using a
* convertor then SG[1].seg_len is accurate but seg_addr is NULL. */
frag->sf_base.uf_base.USNIC_SEND_LOCAL_COUNT = 2;
/* stash header location, PML will write here */
frag->sf_base.uf_local_seg[0].seg_addr.pval = &lfrag->lsf_ompi_header;
frag->sf_base.uf_local_seg[0].seg_len = reserve;
/* make sure upper header small enough */
assert(reserve <= sizeof(lfrag->lsf_ompi_header));
if (OPAL_UNLIKELY(opal_convertor_need_buffers(convertor))) {
/* threshold == -1 means always pack eagerly */
if (mca_btl_usnic_component.pack_lazy_threshold >= 0 &&
*size >= (size_t)mca_btl_usnic_component.pack_lazy_threshold) {
MSGDEBUG1_OUT("packing frag %p on the fly", (void *)frag);
lfrag->lsf_pack_on_the_fly = true;
/* tell the PML we will absorb as much as possible while still
* respecting indivisible element boundaries in the convertor */
*size = opal_btl_usnic_convertor_pack_peek(convertor, *size);
/* Clone the convertor b/c we (the BTL) don't own it and the PML
* might mutate it after we return from this function. */
rc = opal_convertor_clone(convertor, &frag->sf_convertor,
/*copy_stack=*/true);
if (OPAL_UNLIKELY(OPAL_SUCCESS != rc)) {
BTL_ERROR(("unexpected convertor clone error"));
abort(); /* XXX */
}
}
else {
/* pack everything in the convertor into a chain of segments now,
* leaving space for the PML header in the first segment */
lfrag->lsf_base.sf_base.uf_local_seg[0].seg_addr.pval =
pack_chunk_seg_chain_with_reserve(module, lfrag, reserve,
convertor, *size, size);
}
/* We set SG[1] to {NULL,bytes_packed} so that various calculations
* by both PML and this BTL will be correct. For example, the PML adds
* up the bytes in the descriptor segments to determine if an MPI-level
* request is complete or not. */
frag->sf_base.uf_local_seg[1].seg_addr.pval = NULL;
frag->sf_base.uf_local_seg[1].seg_len = *size;
} else {
/* convertor not needed, just save the payload pointer in SG[1] */
lfrag->lsf_pack_on_the_fly = true;
opal_convertor_get_current_pointer(convertor,
&frag->sf_base.uf_local_seg[1].seg_addr.pval);
frag->sf_base.uf_local_seg[1].seg_len = *size;
}
frag->sf_base.uf_base.des_flags = flags;
frag->sf_endpoint = endpoint;
return frag;
}