void mca_spml_ucx_rmkey_unpack(sshmem_mkey_t *mkey, int pe)
{
spml_ucx_mkey_t *ucx_mkey;
ucs_status_t err;
ucx_mkey = (spml_ucx_mkey_t *)malloc(sizeof(*ucx_mkey));
if (!ucx_mkey) {
SPML_ERROR("not enough memory to allocate mkey");
goto error_fatal;
}
err = ucp_ep_rkey_unpack(mca_spml_ucx.ucp_peers[pe].ucp_conn,
mkey->u.data,
&ucx_mkey->rkey);
if (UCS_OK != err) {
SPML_ERROR("failed to unpack rkey");
goto error_fatal;
}
mkey->spml_context = ucx_mkey;
return;
error_fatal:
oshmem_shmem_abort(-1);
return;
}
static inline mca_bml_base_btl_t *get_next_btl(int dst, int *btl_id)
{
mca_bml_base_endpoint_t* endpoint;
mca_bml_base_btl_t* bml_btl;
oshmem_proc_t *proc;
mca_bml_base_btl_array_t *btl_array = 0;
int size = 0;
int shmem_index = 0;
/* get endpoint and btl */
proc = oshmem_proc_group_all(dst);
if (!proc) {
SPML_ERROR("Can not find destination proc for pe=%d", dst);
return NULL ;
}
endpoint = (mca_bml_base_endpoint_t*) proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML];
if (!endpoint) {
SPML_ERROR("pe=%d proc has no endpoint", dst);
return NULL ;
}
/* At the moment always return first transport */
size = mca_bml_base_btl_array_get_size(btl_array = &endpoint->btl_rdma);
if (0 >= size) {
/* Possibly this is SM BTL with KNEM disabled? Then we should use send based get/put */
/*
This hack is necessary for the case when KNEM is not available.
In this case we still want to use send/recv of SM BTL for put and get
but SM BTL is not in the rdma list anymore
*/
size = mca_bml_base_btl_array_get_size(btl_array =
&endpoint->btl_eager);
if (0 < size) {
/*Chose SHMEM capable btl from eager array. Not filter now: take the first
(but could appear on demand).*/
for (shmem_index = 0; shmem_index < size; shmem_index++) {
bml_btl = mca_bml_base_btl_array_get_index(btl_array, shmem_index);
_find_btl_id(bml_btl);
size = 1;
break;
}
}
}
bml_btl = mca_bml_base_btl_array_get_index(btl_array, shmem_index);
*btl_id = proc->transport_ids[0];
#if SPML_YODA_DEBUG == 1
assert(*btl_id >= 0 && *btl_id < YODA_BTL_MAX);
SPML_VERBOSE(100, "pe=%d reachable via btl %s %d", dst,
bml_btl->btl->btl_component->btl_version.mca_component_name, *btl_id);
#endif
return bml_btl;
}
static int mca_spml_ikrit_mxm_fence(int dst)
{
mca_spml_ikrit_get_request_t *fence_req;
fence_req = alloc_get_req();
if (NULL == fence_req) {
SPML_ERROR("out of get requests - aborting");
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
fence_req->mxm_req.base.mq = mca_spml_ikrit.mxm_mq;
fence_req->mxm_req.base.conn = mca_spml_ikrit.mxm_peers[dst]->mxm_conn;
#if MXM_API < MXM_VERSION(2,0)
fence_req->mxm_req.opcode = MXM_REQ_OP_FENCE;
fence_req->mxm_req.base.flags = MXM_REQ_FLAG_SEND_SYNC;
#else
fence_req->mxm_req.opcode = MXM_REQ_OP_PUT_SYNC;
fence_req->mxm_req.flags = MXM_REQ_SEND_FLAG_FENCE;
fence_req->mxm_req.op.mem.remote_vaddr = 0;
fence_req->mxm_req.op.mem.remote_mkey = &mxm_empty_mem_key;
fence_req->mxm_req.base.data_type = MXM_REQ_DATA_BUFFER;
fence_req->mxm_req.base.data.buffer.ptr = 0;
fence_req->mxm_req.base.data.buffer.length = 0;
#endif
fence_req->mxm_req.base.state = MXM_REQ_NEW;
fence_req->mxm_req.base.completed_cb = fence_completion_cb;
fence_req->mxm_req.base.context = fence_req;
OPAL_THREAD_ADD32(&mca_spml_ikrit.n_mxm_fences, 1);
SPML_IKRIT_MXM_POST_SEND(fence_req->mxm_req);
return OSHMEM_SUCCESS;
}
static
mxm_mem_key_t *mca_spml_ikrit_get_mkey_slow(int pe, void *va, int ptl_id, void **rva)
{
sshmem_mkey_t *mkey;
retry:
mkey = mca_memheap_base_get_cached_mkey(pe, va, ptl_id, rva);
if (NULL == mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable", pe, va);
oshmem_shmem_abort(-1);
return NULL;
}
if (MXM_PTL_SHM == ptl_id) {
if (mca_memheap_base_can_local_copy(mkey, va)) {
return NULL;
}
/* if dst addr is on memheap and local copy is not allowed
* disable direct shm transport
*/
if (memheap_is_va_in_segment(va, HEAP_SEG_INDEX)) {
mca_spml_ikrit.mxm_peers[pe].ptl_id = MXM_PTL_RDMA;
}
/* going via mxm must always work */
ptl_id = MXM_PTL_RDMA;
goto retry;
}
return to_mxm_mkey(mkey);
}
int mca_spml_ikrit_get(shmem_ctx_t ctx, void *src_addr, size_t size, void *dst_addr, int src)
{
mxm_send_req_t sreq;
if (0 >= size) {
return OSHMEM_SUCCESS;
}
if (OSHMEM_SUCCESS == mca_spml_ikrit_get_shm(src_addr, size, dst_addr, src))
return OSHMEM_SUCCESS;
if (OSHMEM_SUCCESS
!= mca_spml_ikrit_get_helper(&sreq,
src_addr,
size,
dst_addr,
src)) {
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
sreq.base.completed_cb = NULL;
sreq.flags = 0;
SPML_IKRIT_MXM_POST_SEND(sreq);
mca_spml_irkit_req_wait(&sreq.base);
if (MXM_OK != sreq.base.error) {
SPML_ERROR("get request failed: %s - aborting",
mxm_error_string(sreq.base.error));
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
return OSHMEM_SUCCESS;
}
static inline void put_completion_cb(void *ctx)
{
mca_spml_ikrit_put_request_t *put_req = (mca_spml_ikrit_put_request_t *) ctx;
mxm_peer_t *peer;
OPAL_THREAD_ADD_FETCH32(&mca_spml_ikrit.n_active_puts, -1);
/* TODO: keep pointer to peer in the request */
peer = &mca_spml_ikrit.mxm_peers[put_req->pe];
/* this was last put in progress. Remove peer from the list so that we do not need explicit fence */
#if SPML_IKRIT_PUT_DEBUG == 1
if (peer) {
if (peer->n_active_puts <= 0) {
/* actually this can happen because fence forces ref count to 0 while puts still may be in flight */
SPML_VERBOSE(1, "pe %d n_active_puts %d", put_req->pe, peer->n_active_puts);
}
}
if (put_req->mxm_req.base.state != MXM_REQ_COMPLETED)
SPML_ERROR("oops: pe %d uncompleted request state %d", put_req->pe, put_req->mxm_req.base.state);
#endif
if (0 < peer->n_active_puts) {
peer->n_active_puts--;
if (0 == peer->n_active_puts &&
(put_req->mxm_req.opcode == MXM_REQ_OP_PUT_SYNC)) {
opal_list_remove_item(&mca_spml_ikrit.active_peers, &peer->link);
peer->need_fence = 0;
}
}
free_put_req(put_req);
}
int mca_spml_ikrit_put(shmem_ctx_t ctx, void* dst_addr, size_t size, void* src_addr, int dst)
{
int err;
mca_spml_ikrit_put_request_t *put_req;
mxm_wait_t wait;
put_req = 0;
err = mca_spml_ikrit_put_internal(dst_addr,
size,
src_addr,
dst,
(void **) &put_req,
0);
if (OSHMEM_SUCCESS != err) {
SPML_ERROR("put failed - aborting");
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
if (!put_req)
return OSHMEM_SUCCESS;
wait.req = &put_req->mxm_req.base;
wait.state = (mxm_req_state_t)(MXM_REQ_SENT | MXM_REQ_COMPLETED);
wait.progress_cb = NULL;
wait.progress_arg = NULL;
mxm_wait(&wait);
return OSHMEM_SUCCESS;
}
static inline int mca_spml_ikrit_get_shm(void *src_addr,
size_t size,
void *dst_addr,
int src)
{
int ptl_id;
void *rva;
sshmem_mkey_t *r_mkey;
ptl_id = get_ptl_id(src);
/**
* Get the address to the remote rkey.
**/
if (ptl_id != MXM_PTL_SHM)
return OSHMEM_ERROR;
r_mkey = mca_memheap_base_get_cached_mkey(src, src_addr, ptl_id, &rva);
if (!r_mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable",
src, src_addr);
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
if (!mca_memheap_base_can_local_copy(r_mkey, src_addr))
return OSHMEM_ERROR;
SPML_VERBOSE(100,
"shm get: pe:%d src=%p -> dst: %p sz=%d. src_rva=%p, %s",
src, src_addr, dst_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey));
memcpy(dst_addr, (void *) (unsigned long) rva, size);
opal_progress();
return OSHMEM_SUCCESS;
}
/* make global btl list&map */
static int create_btl_list(void)
{
int btl_type;
char *btl_name;
int size;
opal_list_item_t *item;
mca_btl_base_selected_module_t *btl_sm;
int i;
size = opal_list_get_size(&mca_btl_base_modules_initialized);
if (0 >= size) {
SPML_ERROR("no btl(s) available");
return OSHMEM_ERROR;
}
SPML_VERBOSE(50, "found %d capable btls", size);
mca_spml_yoda.btl_type_map =
(struct yoda_btl *) calloc(size, sizeof(struct yoda_btl));
if (!mca_spml_yoda.btl_type_map)
return OSHMEM_ERROR;
mca_spml_yoda.n_btls = 0;
for (i = 0, item = opal_list_get_first(&mca_btl_base_modules_initialized);
item != opal_list_get_end(&mca_btl_base_modules_initialized);
item = opal_list_get_next(item), i++) {
btl_sm = (mca_btl_base_selected_module_t *) item;
btl_name = btl_sm->btl_component->btl_version.mca_component_name;
btl_type = btl_name_to_id(btl_name);
SPML_VERBOSE(50, "found btl (%s) btl_type=%s", btl_name, btl_type2str(btl_type));
/* Note: we setup bml_btl in create_btl_idx() */
mca_spml_yoda.btl_type_map[mca_spml_yoda.n_btls].bml_btl = NULL;
mca_spml_yoda.btl_type_map[mca_spml_yoda.n_btls].btl =
btl_sm->btl_module;
mca_spml_yoda.btl_type_map[mca_spml_yoda.n_btls].btl_type = btl_type;
mca_spml_yoda.n_btls++;
}
if (0 == mca_spml_yoda.n_btls) {
SPML_ERROR("can not find any suitable btl");
return OSHMEM_ERROR;
}
return OSHMEM_SUCCESS;
}
/* extension. used 4 fence implementation b4 fence was added to mxm */
int mca_spml_ikrit_get_async(void *src_addr,
size_t size,
void *dst_addr,
int src)
{
mca_spml_ikrit_get_request_t *get_req;
if (OSHMEM_SUCCESS == mca_spml_ikrit_get_shm(src_addr, size, dst_addr, src))
return OSHMEM_SUCCESS;
get_req = alloc_get_req();
if (NULL == get_req) {
SPML_ERROR("out of get requests - aborting");
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
if (OSHMEM_SUCCESS
!= mca_spml_ikrit_get_helper(&get_req->mxm_req,
src_addr,
size,
dst_addr,
src)) {
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
#if MXM_API < MXM_VERSION(2,0)
get_req->mxm_req.base.flags = 0;
#else
get_req->mxm_req.flags = 0;
#endif
get_req->mxm_req.base.completed_cb = get_completion_cb;
get_req->mxm_req.base.context = get_req;
OPAL_THREAD_ADD32(&mca_spml_ikrit.n_active_gets, 1);
mxm_req_send(&get_req->mxm_req);
if (MXM_OK != get_req->mxm_req.base.error) {
SPML_ERROR("get request failed: %s - aborting",
mxm_error_string(get_req->mxm_req.base.error));
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
return OSHMEM_SUCCESS;
}
int mca_spml_ucx_del_procs(ompi_proc_t** procs, size_t nprocs)
{
int my_rank = oshmem_my_proc_id();
size_t num_reqs, max_reqs;
void *dreq, **dreqs;
ucp_ep_h ep;
size_t i, n;
oshmem_shmem_barrier();
if (!mca_spml_ucx.ucp_peers) {
return OSHMEM_SUCCESS;
}
max_reqs = mca_spml_ucx.num_disconnect;
if (max_reqs > nprocs) {
max_reqs = nprocs;
}
dreqs = malloc(sizeof(*dreqs) * max_reqs);
if (dreqs == NULL) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
num_reqs = 0;
for (i = 0; i < nprocs; ++i) {
n = (i + my_rank) % nprocs;
ep = mca_spml_ucx.ucp_peers[n].ucp_conn;
if (ep == NULL) {
continue;
}
SPML_VERBOSE(10, "disconnecting from peer %d", n);
dreq = ucp_disconnect_nb(ep);
if (dreq != NULL) {
if (UCS_PTR_IS_ERR(dreq)) {
SPML_ERROR("ucp_disconnect_nb(%d) failed: %s", n,
ucs_status_string(UCS_PTR_STATUS(dreq)));
} else {
dreqs[num_reqs++] = dreq;
}
}
mca_spml_ucx.ucp_peers[n].ucp_conn = NULL;
if ((int)num_reqs >= mca_spml_ucx.num_disconnect) {
mca_spml_ucx_waitall(dreqs, &num_reqs);
}
}
mca_spml_ucx_waitall(dreqs, &num_reqs);
free(dreqs);
opal_pmix.fence(NULL, 0);
free(mca_spml_ucx.ucp_peers);
return OSHMEM_SUCCESS;
}
int mca_spml_ucx_quiet(void)
{
ucs_status_t err;
err = ucp_worker_flush(mca_spml_ucx.ucp_worker);
if (UCS_OK != err) {
SPML_ERROR("fence failed");
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
return OSHMEM_SUCCESS;
}
static int mca_spml_ikrit_get_helper(mxm_send_req_t *sreq,
void *src_addr,
size_t size,
void *dst_addr,
int src)
{
/* shmem spec states that get() operations are blocking. So it is enough
to have single mxm request. Also we count on mxm doing copy */
void *rva;
sshmem_mkey_t *r_mkey;
int ptl_id;
ptl_id = get_ptl_id(src);
/* already tried to send via shm and failed. go via rdma */
if (ptl_id == MXM_PTL_SHM)
ptl_id = MXM_PTL_RDMA;
/**
* Get the address to the remote rkey.
**/
r_mkey = mca_memheap.memheap_get_cached_mkey(src,
src_addr,
ptl_id,
&rva);
if (!r_mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable",
src, src_addr);
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
SPML_VERBOSE(100,
"get: pe:%d ptl=%d src=%p -> dst: %p sz=%d. src_rva=%p, %s",
src, ptl_id, src_addr, dst_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey));
/* mxm does not really cares for get lkey */
sreq->base.mq = mca_spml_ikrit.mxm_mq;
sreq->base.conn = mca_spml_ikrit.mxm_peers[src]->mxm_conn;
sreq->base.data_type = MXM_REQ_DATA_BUFFER;
sreq->base.data.buffer.ptr = dst_addr;
sreq->base.data.buffer.length = size;
#if MXM_API < MXM_VERSION(2,0)
sreq->base.data.buffer.memh = NULL;
sreq->op.mem.remote_memh = NULL;
#else
sreq->op.mem.remote_mkey = to_mxm_mkey(r_mkey);
#endif
sreq->opcode = MXM_REQ_OP_GET;
sreq->op.mem.remote_vaddr = (intptr_t) rva;
sreq->base.state = MXM_REQ_NEW;
return OSHMEM_SUCCESS;
}
static inline int get_ptl_id(int dst)
{
ompi_proc_t *proc;
/* get endpoint and btl */
proc = oshmem_proc_group_all(dst);
if (!proc) {
SPML_ERROR("Can not find destination proc for pe=%d", dst);
oshmem_shmem_abort(-1);
return -1;
}
return OSHMEM_PROC_DATA(proc)->transport_ids[0];
}
spml_ucx_mkey_t * mca_spml_ucx_get_mkey_slow(int pe, void *va, void **rva)
{
sshmem_mkey_t *r_mkey;
r_mkey = mca_memheap_base_get_cached_mkey(pe, va, 0, rva);
if (OPAL_UNLIKELY(!r_mkey)) {
SPML_ERROR("pe=%d: %p is not address of symmetric variable",
pe, va);
oshmem_shmem_abort(-1);
return NULL;
}
return (spml_ucx_mkey_t *)(r_mkey->spml_context);
}
int mca_spml_ikrit_put_nb(void* dst_addr,
size_t size,
void* src_addr,
int dst,
void **handle)
{
int err;
err = mca_spml_ikrit_put_internal(dst_addr, size, src_addr, dst, handle, 1);
if (OSHMEM_SUCCESS != err) {
SPML_ERROR("put failed - aborting");
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
return OSHMEM_SUCCESS;
}
int mca_spml_yoda_add_procs(oshmem_proc_t** procs, size_t nprocs)
{
opal_bitmap_t reachable;
int rc;
size_t i;
if (0 == nprocs) {
return OSHMEM_SUCCESS;
}
OBJ_CONSTRUCT(&reachable, opal_bitmap_t);
rc = opal_bitmap_init(&reachable, (int) nprocs);
if (OSHMEM_SUCCESS != rc) {
return rc;
}
rc = mca_bml.bml_add_procs(nprocs, (ompi_proc_t**) procs, &reachable);
if (OSHMEM_SUCCESS != rc) {
SPML_ERROR("SPML YODA: shmem error\n");
goto cleanup_and_return;
}
rc = mca_bml.bml_register_error(mca_spml_yoda_error_handler);
if (OMPI_SUCCESS != rc) {
goto cleanup_and_return;
}
/* create btl index and map */
rc = create_btl_list();
if (OSHMEM_SUCCESS != rc) {
goto cleanup_and_return;
}
for (i = 0; i < nprocs; i++) {
rc = create_btl_idx(i);
if (OSHMEM_SUCCESS != rc) {
goto cleanup_and_return;
}
}
cleanup_and_return:
OBJ_DESTRUCT(&reachable);
return rc;
}
static inline void put_completion_cb(void *ctx)
{
mca_spml_ikrit_put_request_t *put_req = (mca_spml_ikrit_put_request_t *) ctx;
mxm_peer_t *peer;
OPAL_THREAD_ADD32(&mca_spml_ikrit.n_active_puts, -1);
peer = mca_spml_ikrit.mxm_peers[put_req->pe];
/* this was last put in progress. Remove peer from the list so that we do not need explicit fence */
#if SPML_IKRIT_PUT_DEBUG == 1
if (peer) {
if (peer->n_active_puts <= 0) {
/* actually this can happen because fence forces ref count to 0 while puts still may be in flight */
SPML_VERBOSE(1, "pe %d n_active_puts %d", put_req->pe, peer->n_active_puts);
}
}
if (put_req->mxm_req.base.state != MXM_REQ_COMPLETED)
SPML_ERROR("oops: pe %d uncompleted request state %d", put_req->pe, put_req->mxm_req.base.state);
#endif
if (0 < peer->n_active_puts) {
peer->n_active_puts--;
#if MXM_API < MXM_VERSION(2,0)
if (0 == peer->n_active_puts &&
(put_req->mxm_req.base.flags & MXM_REQ_FLAG_SEND_SYNC)) {
opal_list_remove_item(&mca_spml_ikrit.active_peers, &peer->super);
peer->need_fence = 0;
}
#else
if (0 == peer->n_active_puts &&
(put_req->mxm_req.opcode == MXM_REQ_OP_PUT_SYNC)) {
opal_list_remove_item(&mca_spml_ikrit.active_peers, &peer->super);
peer->need_fence = 0;
}
#endif
}
put_req->req_put.req_base.req_spml_complete = true;
put_req->req_put.req_base.req_oshmem.req_status.SHMEM_ERROR =
OSHMEM_SUCCESS;
oshmem_request_complete(&put_req->req_put.req_base.req_oshmem, 1);
oshmem_request_free((oshmem_request_t**) &put_req);
}
static void mca_spml_ucx_waitall(void **reqs, size_t *count_p)
{
ucs_status_t status;
size_t i;
SPML_VERBOSE(10, "waiting for %d disconnect requests", *count_p);
for (i = 0; i < *count_p; ++i) {
do {
opal_progress();
status = ucp_request_test(reqs[i], NULL);
} while (status == UCS_INPROGRESS);
if (status != UCS_OK) {
SPML_ERROR("disconnect request failed: %s",
ucs_status_string(status));
}
ucp_request_release(reqs[i]);
reqs[i] = NULL;
}
*count_p = 0;
}
int mca_spml_ikrit_get(void *src_addr, size_t size, void *dst_addr, int src)
{
mxm_send_req_t sreq;
if (0 >= size) {
return OSHMEM_SUCCESS;
}
if (OSHMEM_SUCCESS == mca_spml_ikrit_get_shm(src_addr, size, dst_addr, src))
return OSHMEM_SUCCESS;
if (OSHMEM_SUCCESS
!= mca_spml_ikrit_get_helper(&sreq,
src_addr,
size,
dst_addr,
src)) {
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
#if MXM_API < MXM_VERSION(2,0)
sreq.base.flags = MXM_REQ_FLAG_BLOCKING;
#else
sreq.flags = MXM_REQ_SEND_FLAG_BLOCKING;
#endif
sreq.base.completed_cb = NULL;
mxm_req_send(&sreq);
opal_progress();
mca_spml_irkit_req_wait(&sreq.base);
if (MXM_OK != sreq.base.error) {
SPML_ERROR("get request failed: %s - aborting",
mxm_error_string(sreq.base.error));
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
return OSHMEM_SUCCESS;
}
void mca_spml_ucx_rmkey_unpack(sshmem_mkey_t *mkey, uint32_t segno, int pe, int tr_id)
{
spml_ucx_mkey_t *ucx_mkey;
ucs_status_t err;
ucx_mkey = &mca_spml_ucx.ucp_peers[pe].mkeys[segno].key;
err = ucp_ep_rkey_unpack(mca_spml_ucx.ucp_peers[pe].ucp_conn,
mkey->u.data,
&ucx_mkey->rkey);
if (UCS_OK != err) {
SPML_ERROR("failed to unpack rkey");
goto error_fatal;
}
mkey->spml_context = ucx_mkey;
mca_spml_ucx_cache_mkey(mkey, segno, pe);
return;
error_fatal:
oshmem_shmem_abort(-1);
return;
}
void mca_spml_yoda_put_completion(mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* des,
int status)
{
mca_spml_yoda_rdma_frag_t* frag =
(mca_spml_yoda_rdma_frag_t*) des->des_cbdata;
mca_spml_yoda_put_request_t* putreq =
(mca_spml_yoda_put_request_t*) frag->rdma_req;
mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) des->des_context;
OPAL_THREAD_ADD32(&mca_spml_yoda.n_active_puts, -1);
/* check completion status */
if (OPAL_UNLIKELY(OSHMEM_SUCCESS != status)) {
/* no way to propagete errors. die */
SPML_ERROR("FATAL put completion error");
oshmem_shmem_abort(-1);
}
putreq->req_put.req_base.req_spml_complete = true;
oshmem_request_complete(&putreq->req_put.req_base.req_oshmem, 1);
oshmem_request_free((oshmem_request_t**) &putreq);
mca_bml_base_free(bml_btl, des);
}
static inline int mca_spml_yoda_put_internal(void *dst_addr,
size_t size,
void *src_addr,
int dst,
int is_nb)
{
int rc = OSHMEM_SUCCESS;
mca_spml_yoda_put_request_t *putreq = NULL;
mca_bml_base_btl_t* bml_btl;
mca_btl_base_descriptor_t* des = NULL;
mca_btl_base_segment_t* segment;
mca_spml_yoda_rdma_frag_t* frag;
int nfrags;
int i;
unsigned ncopied = 0;
unsigned int frag_size = 0;
char *p_src, *p_dst;
void* rva;
sshmem_mkey_t *r_mkey;
int btl_id = 0;
struct yoda_btl *ybtl;
int put_via_send;
mca_btl_base_registration_handle_t *local_handle = NULL, *remote_handle = NULL;
/* If nothing to put its OK.*/
if (0 >= size) {
return OSHMEM_SUCCESS;
}
/* Find bml_btl and its global btl_id */
bml_btl = get_next_btl(dst, &btl_id);
if (!bml_btl) {
SPML_ERROR("cannot reach %d pe: no appropriate btl found", oshmem_my_proc_id());
rc = OSHMEM_ERR_FATAL;
goto exit_fatal;
}
/* Check if btl has PUT method. If it doesn't - use SEND*/
put_via_send = !(bml_btl->btl->btl_flags & MCA_BTL_FLAGS_PUT);
/* Get rkey of remote PE (dst proc) which must be on memheap*/
r_mkey = mca_memheap_base_get_cached_mkey(dst, dst_addr, btl_id, &rva);
if (!r_mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable",
dst, dst_addr);
rc = OSHMEM_ERR_FATAL;
goto exit_fatal;
}
#if SPML_YODA_DEBUG == 1
SPML_VERBOSE(100, "put: pe:%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s",
dst, dst_addr, src_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey));
#endif
ybtl = &mca_spml_yoda.btl_type_map[btl_id];
if (ybtl->btl->btl_register_mem) {
assert (r_mkey->len == ybtl->btl->btl_registration_handle_size);
remote_handle = (mca_btl_base_registration_handle_t *) r_mkey->u.data;
}
/* check if we doing put into shm attached segment and if so
* just do memcpy
*/
if ((YODA_BTL_SM == ybtl->btl_type || YODA_BTL_VADER == ybtl->btl_type)
&& mca_memheap_base_can_local_copy(r_mkey, dst_addr)) {
memcpy((void *) (unsigned long) rva, src_addr, size);
return OSHMEM_SUCCESS;
}
/* We support only blocking PUT now => we always need copy for src buffer*/
calc_nfrags_put (bml_btl, size, &frag_size, &nfrags, put_via_send);
p_src = (char*) src_addr;
p_dst = (char*) (unsigned long) rva;
for (i = 0; i < nfrags; i++) {
/* Allocating send request from free list */
putreq = mca_spml_yoda_putreq_alloc(dst);
frag = &putreq->put_frag;
ncopied = i < nfrags - 1 ? frag_size :(unsigned) ((char *) src_addr + size - p_src);
/* Preparing source buffer */
/* allocate buffer */
mca_spml_yoda_bml_alloc(bml_btl,
&des,
MCA_BTL_NO_ORDER,
ncopied,
MCA_BTL_DES_SEND_ALWAYS_CALLBACK,
put_via_send);
if (OPAL_UNLIKELY(!des || !des->des_segments )) {
SPML_ERROR("src=%p nfrags = %d frag_size=%d",
src_addr, nfrags, frag_size);
SPML_ERROR("shmem OOM error need %d bytes", ncopied);
opal_show_help("help-oshmem-spml-yoda.txt",
"internal_oom_error",
true,
"Put", ncopied, mca_spml_yoda.bml_alloc_threshold);
rc = OSHMEM_ERR_FATAL;
goto exit_fatal;
}
/* copy data to allocated buffer*/
segment = des->des_segments;
spml_yoda_prepare_for_put((void*)segment->seg_addr.pval, ncopied,
(void*)p_src, (void*)p_dst, put_via_send);
if (!put_via_send && ybtl->btl->btl_register_mem) {
local_handle = ybtl->btl->btl_register_mem (ybtl->btl, bml_btl->btl_endpoint,
segment->seg_addr.pval, ncopied, 0);
if (NULL == local_handle) {
/* No free resources, Block on completion here */
SPML_ERROR("shmem error: OSHMEM_ERR_OUT_OF_RESOURCE");
oshmem_request_wait_completion(&putreq->req_put.req_base.req_oshmem);
}
}
frag->rdma_segs[0].base_seg.seg_addr.lval = (uintptr_t) p_dst;
frag->rdma_segs[0].base_seg.seg_len = (put_via_send ?
ncopied + SPML_YODA_SEND_CONTEXT_SIZE :
ncopied);
frag->rdma_req = putreq;
/* initialize callback data for put*/
des->des_cbdata = frag;
des->des_cbfunc = mca_spml_yoda_put_completion;
OPAL_THREAD_ADD32(&mca_spml_yoda.n_active_puts, 1);
/* put the data to remote side */
if (!put_via_send) {
rc = mca_bml_base_put (bml_btl, segment->seg_addr.pval, (uint64_t) (intptr_t) p_dst,
local_handle, remote_handle, ncopied, 0, 0, mca_spml_yoda_put_completion_rdma,
des);
} else {
rc = mca_bml_base_send(bml_btl, des, MCA_SPML_YODA_PUT);
if (1 == rc)
rc = OSHMEM_SUCCESS;
}
if (OPAL_UNLIKELY(OSHMEM_SUCCESS != rc)) {
if (OSHMEM_ERR_OUT_OF_RESOURCE == rc) {
/* No free resources, Block on completion here */
SPML_ERROR("shmem error: OSHMEM_ERR_OUT_OF_RESOURCE");
oshmem_request_wait_completion(&putreq->req_put.req_base.req_oshmem);
} else {
SPML_ERROR("shmem error");
}
/* exit with errro */
SPML_ERROR("shmem error: ret = %i, send_pe = %i, dest_pe = %i",
rc, oshmem_my_proc_id(), dst);
rc = OSHMEM_ERR_FATAL;
goto exit_fatal;
}
p_src += ncopied;
p_dst += ncopied;
}
return rc;
exit_fatal:
if (OSHMEM_SUCCESS != rc) {
oshmem_shmem_abort(rc);
}
return rc;
}
/* for each proc create transport ids which are indexes into global
* btl list&map
*/
static int create_btl_idx(int dst_pe)
{
oshmem_proc_t *proc;
int btl_id;
mca_bml_base_endpoint_t* endpoint;
mca_bml_base_btl_t* bml_btl = 0;
int i, size;
mca_bml_base_btl_array_t *btl_array;
int shmem_index = -1;
proc = oshmem_proc_group_find(oshmem_group_all, dst_pe);
endpoint = (mca_bml_base_endpoint_t*) proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML];
assert(endpoint);
size = mca_bml_base_btl_array_get_size(btl_array = &endpoint->btl_rdma);
if (0 >= size) {
/* Possibly this is SM BTL with KNEM disabled? Then we should use send based get/put */
/*
This hack is necessary for the case when KNEM is not available.
In this case we still want to use send/recv of SM BTL for put and get
but SM BTL is not in the rdma list anymore
*/
size = mca_bml_base_btl_array_get_size(btl_array =
&endpoint->btl_eager);
if (0 < size) {
/*Chose SHMEM capable btl from eager array. Not filter now: take the first
(but could appear on demand).*/
shmem_index = 0;
size = 1;
}
else {
SPML_ERROR("no SHMEM capable transport for dest pe=%d", dst_pe);
return OSHMEM_ERROR;
}
}
proc->transport_ids = (char *) malloc(size * sizeof(char));
if (!proc->transport_ids)
return OSHMEM_ERROR;
proc->num_transports = size;
for (i = 0; i < size; i++) {
bml_btl = mca_bml_base_btl_array_get_index(btl_array,
(shmem_index >= 0) ?
(shmem_index) : (i));
btl_id = _find_btl_id(bml_btl);
SPML_VERBOSE(50,
"dst_pe(%d) use btl (%s) btl_id=%d",
dst_pe, bml_btl->btl->btl_component->btl_version.mca_component_name, btl_id);
if (0 > btl_id) {
SPML_ERROR("unknown btl: dst_pe(%d) use btl (%s) btl_id=%d",
dst_pe, bml_btl->btl->btl_component->btl_version.mca_component_name, btl_id);
return OSHMEM_ERROR;
}
proc->transport_ids[i] = btl_id;
mca_spml_yoda.btl_type_map[btl_id].bml_btl = bml_btl;
mca_spml_yoda.btl_type_map[btl_id].use_cnt++;
}
return OSHMEM_SUCCESS;
}
sshmem_mkey_t *mca_spml_yoda_register(void* addr,
size_t size,
uint64_t shmid,
int *count)
{
int i;
sshmem_mkey_t *mkeys;
struct yoda_btl *ybtl;
mca_spml_yoda_context_t* yoda_context;
SPML_VERBOSE(10, "address %p len %llu", addr, (unsigned long long)size);
*count = 0;
/* make sure everything is initialized to 0 */
mkeys = (sshmem_mkey_t *) calloc(1,
mca_spml_yoda.n_btls * sizeof(*mkeys));
if (!mkeys) {
return NULL ;
}
mca_bml.bml_register( MCA_SPML_YODA_PUT,
mca_yoda_put_callback,
NULL );
mca_bml.bml_register( MCA_SPML_YODA_GET,
mca_yoda_get_callback,
NULL );
mca_bml.bml_register( MCA_SPML_YODA_GET_RESPONSE,
mca_yoda_get_response_callback,
NULL );
/* Register proc memory in every rdma BTL. */
for (i = 0; i < mca_spml_yoda.n_btls; i++) {
ybtl = &mca_spml_yoda.btl_type_map[i];
mkeys[i].va_base = addr;
mkeys[i].u.key = MAP_SEGMENT_SHM_INVALID;
if (!ybtl->use_cnt) {
SPML_VERBOSE(10,
"%s: present but not in use. SKIP registration",
btl_type2str(ybtl->btl_type));
continue;
}
/* If we have shared memory just save its id */
if ((YODA_BTL_SM == ybtl->btl_type || YODA_BTL_VADER == ybtl->btl_type)
&& MAP_SEGMENT_SHM_INVALID != (int)shmid) {
mkeys[i].u.key = shmid;
mkeys[i].va_base = 0;
continue;
}
yoda_context = calloc(1, sizeof(*yoda_context));
mkeys[i].spml_context = yoda_context;
yoda_context->registration = NULL;
if (ybtl->btl->btl_flags & MCA_BTL_FLAGS_RDMA) {
if (NULL != ybtl->btl->btl_register_mem) {
yoda_context->registration = ybtl->btl->btl_register_mem (ybtl->btl, MCA_BTL_ENDPOINT_ANY,
addr, size, MCA_BTL_REG_FLAG_ACCESS_ANY);
if (NULL == yoda_context->registration) {
SPML_ERROR("%s: failed to register source memory: addr: %p, size: %u",
btl_type2str(ybtl->btl_type), addr, size);
/* FIXME some cleanup might be needed here
* yoda_context->btl_src_descriptor = NULL;
* *count = ???;
* free(spml_context);
*/
free(mkeys);
return NULL;
}
}
yoda_context->btl_src_descriptor = NULL;
mkeys[i].u.data = yoda_context->registration;
mkeys[i].len = yoda_context->registration ? ybtl->btl->btl_registration_handle_size : 0;
}
SPML_VERBOSE(5,
"rank %d btl %s va_base: 0x%p len: %d key %llx size %llu",
oshmem_proc_pe(oshmem_proc_local()), btl_type2str(ybtl->btl_type),
mkeys[i].va_base, mkeys[i].len, (unsigned long long)mkeys[i].u.key, (unsigned long long)size);
}
*count = mca_spml_yoda.n_btls;
return mkeys;
}
static void mca_yoda_get_callback(mca_btl_base_module_t* btl,
mca_btl_base_tag_t tag,
mca_btl_base_descriptor_t* des,
void* cbdata )
{
void** p, ** p_src, **p_dst;
size_t* size;
int* dst;
void** p_getreq;
mca_btl_base_descriptor_t* des_loc;
int rc;
mca_bml_base_btl_t* bml_btl;
mca_spml_yoda_rdma_frag_t* frag;
int btl_id;
mca_spml_yoda_put_request_t *putreq;
rc = OSHMEM_SUCCESS;
btl_id = 0;
putreq = NULL;
/* Unpack data */
p = (void **)des->des_segments->seg_addr.pval;
p_src = (void*) p;
size = (size_t*)((char*)p_src + sizeof(*p_src) );
dst = (int*)( (char*)size + sizeof(*size));
p_dst = (void*) ((char*)dst + sizeof(*dst));
p_getreq =(void**) ( (char*)p_dst + sizeof(*p_dst));
/* Prepare put via send*/
bml_btl = get_next_btl(*dst, &btl_id);
putreq = mca_spml_yoda_putreq_alloc(*dst);
frag = &putreq->put_frag;
mca_spml_yoda_bml_alloc(bml_btl,
&des_loc,
MCA_BTL_NO_ORDER,
*size,
MCA_BTL_DES_SEND_ALWAYS_CALLBACK,
1);
if (OPAL_UNLIKELY(!des_loc || !des_loc->des_segments)) {
SPML_ERROR("shmem OOM error need %d bytes", (int)*size);
oshmem_shmem_abort(-1);
}
spml_yoda_prepare_for_get_response((void*)des_loc->des_segments->seg_addr.pval, *size, (void*)*p_src, (void*) *p_dst,(void*)*p_getreq,1);
frag->rdma_req = putreq;
/* Initialize callback data for put*/
des_loc->des_cbdata = frag;
des_loc->des_cbfunc = mca_spml_yoda_put_completion;
des_loc->des_segment_count = 1;
OPAL_THREAD_ADD32(&mca_spml_yoda.n_active_puts, 1);
/* Put via send*/
rc = mca_bml_base_send(bml_btl, des_loc, MCA_SPML_YODA_GET_RESPONSE);
if (1 == rc) {
rc = OSHMEM_SUCCESS;
}
if (OPAL_UNLIKELY(OSHMEM_SUCCESS != rc)) {
if (OSHMEM_ERR_OUT_OF_RESOURCE == rc) {
/* No free resources, Block on completion here */
SPML_ERROR("shmem error: OSHMEM_ERR_OUT_OF_RESOURCE");
oshmem_request_wait_completion(&putreq->req_put.req_base.req_oshmem);
} else {
SPML_ERROR("shmem error");
}
/* exit with errro */
SPML_ERROR("shmem error: ret = %i, send_pe = %i, dest_pe = %i",
rc, oshmem_my_proc_id(), *dst);
oshmem_shmem_abort(-1);
rc = OSHMEM_ERROR;
}
}
/**
* shmem_get reads data from a remote address
* in the symmetric heap via RDMA READ.
* Get operation:
* 1. Get the rkey to the remote address.
* 2. Allocate a get request.
* 3. Allocated a temporary pre-registered buffer
* to copy the data to.
* 4. Init the request descriptor with remote side
* data and local side data.
* 5. Read the remote buffer to a pre-registered
* buffer on the local PE using RDMA READ.
* 6. Copy the received data to dst_addr if an
* intermediate pre-register buffer was used.
* 7. Clear the request and return.
*
* src_addr - address on remote pe.
* size - the amount on bytes to be read.
* dst_addr - address on the local pe.
* src - the pe of remote process.
*/
int mca_spml_yoda_get(void* src_addr, size_t size, void* dst_addr, int src)
{
int rc = OSHMEM_SUCCESS;
sshmem_mkey_t *r_mkey, *l_mkey;
void* rva;
unsigned ncopied = 0;
unsigned int frag_size = 0;
char *p_src, *p_dst;
int i;
int nfrags;
mca_bml_base_btl_t* bml_btl = NULL;
mca_btl_base_segment_t* segment;
mca_btl_base_descriptor_t* des = NULL;
mca_spml_yoda_rdma_frag_t* frag = NULL;
struct mca_spml_yoda_getreq_parent get_holder;
struct yoda_btl *ybtl;
int btl_id = 0;
int get_via_send;
mca_btl_base_registration_handle_t *local_handle, *remote_handle = NULL;
mca_spml_yoda_get_request_t* getreq = NULL;
/*If nothing to get its OK.*/
if (0 >= size) {
return rc;
}
/* Find bml_btl and its global btl_id */
bml_btl = get_next_btl(src, &btl_id);
if (!bml_btl) {
SPML_ERROR("cannot reach %d pe: no appropriate btl found", oshmem_my_proc_id());
rc = OSHMEM_ERR_FATAL;
goto exit_fatal;
}
/* Check if btl has GET method. If it doesn't - use SEND*/
get_via_send = ! ( (bml_btl->btl->btl_flags & (MCA_BTL_FLAGS_GET)) &&
(bml_btl->btl->btl_flags & (MCA_BTL_FLAGS_PUT)) );
/* Get rkey of remote PE (src proc) which must be on memheap*/
r_mkey = mca_memheap_base_get_cached_mkey(src, src_addr, btl_id, &rva);
if (!r_mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable",
src, src_addr);
rc = OSHMEM_ERR_FATAL;
goto exit_fatal;
}
#if SPML_YODA_DEBUG == 1
SPML_VERBOSE(100, "get: pe:%d src=%p -> dst: %p sz=%d. src_rva=%p, %s",
src, src_addr, dst_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey));
#endif
ybtl = &mca_spml_yoda.btl_type_map[btl_id];
if (ybtl->btl->btl_register_mem) {
assert(ybtl->btl->btl_registration_handle_size == r_mkey->len);
remote_handle = (mca_btl_base_registration_handle_t *) r_mkey->u.data;
}
nfrags = 1;
/* check if we doing get into shm attached segment and if so
* just do memcpy
*/
if ((YODA_BTL_SM == ybtl->btl_type || YODA_BTL_VADER == ybtl->btl_type)
&& mca_memheap_base_can_local_copy(r_mkey, src_addr)) {
memcpy(dst_addr, (void *) rva, size);
/* must call progress here to avoid deadlock. Scenarion:
* pe1 pols pe2 via shm get. pe2 tries to get static variable from node one, which goes to sm btl
* In this case pe2 is stuck forever because pe1 never calls opal_progress.
* May be we do not need to call progress on every get() here but rather once in a while.
*/
opal_progress();
return OSHMEM_SUCCESS;
}
l_mkey = mca_memheap.memheap_get_local_mkey(dst_addr,
btl_id);
/*
* Need a copy if local memory has not been registered or
* we make GET via SEND
*/
frag_size = ncopied;
if ((NULL == l_mkey) || get_via_send) {
calc_nfrags_get (bml_btl, size, &frag_size, &nfrags, get_via_send);
}
p_src = (char*) (unsigned long) rva;
p_dst = (char*) dst_addr;
get_holder.active_count = 0;
for (i = 0; i < nfrags; i++) {
/**
* Allocating a get request from a pre-allocated
* and pre-registered free list.
*/
getreq = mca_spml_yoda_getreq_alloc(src);
assert(getreq);
getreq->p_dst = NULL;
frag = &getreq->get_frag;
getreq->parent = &get_holder;
ncopied = i < nfrags - 1 ? frag_size :(unsigned) ((char *) dst_addr + size - p_dst);
frag->allocated = 0;
/* Prepare destination descriptor*/
memcpy(&frag->rdma_segs[0].base_seg,
r_mkey->u.data,
r_mkey->len);
frag->rdma_segs[0].base_seg.seg_len = (get_via_send ? ncopied + SPML_YODA_SEND_CONTEXT_SIZE : ncopied);
if (get_via_send) {
frag->use_send = 1;
frag->allocated = 1;
/**
* Allocate a temporary buffer on the local PE.
* The local buffer will store the data read
* from the remote address.
*/
mca_spml_yoda_bml_alloc(bml_btl,
&des,
MCA_BTL_NO_ORDER,
(int)frag_size,
MCA_BTL_DES_SEND_ALWAYS_CALLBACK,
get_via_send);
if (OPAL_UNLIKELY(!des || !des->des_segments)) {
SPML_ERROR("shmem OOM error need %d bytes", ncopied);
SPML_ERROR("src=%p nfrags = %d frag_size=%d",
src_addr, nfrags, frag_size);
rc = OSHMEM_ERR_FATAL;
goto exit_fatal;
}
segment = des->des_segments;
spml_yoda_prepare_for_get((void*)segment->seg_addr.pval, ncopied, (void*)p_src, oshmem_my_proc_id(), (void*)p_dst, (void*) getreq);
des->des_cbfunc = mca_spml_yoda_get_response_completion;
des->des_cbdata = frag;
OPAL_THREAD_ADD32(&mca_spml_yoda.n_active_gets, 1);
}
else {
/*
* Register src memory if do GET via GET
*/
if (NULL == l_mkey && ybtl->btl->btl_register_mem) {
local_handle = ybtl->btl->btl_register_mem (ybtl->btl, bml_btl->btl_endpoint, p_dst, ncopied,
MCA_BTL_REG_FLAG_LOCAL_WRITE);
if (NULL == local_handle) {
SPML_ERROR("%s: failed to register destination memory %p.",
btl_type2str(ybtl->btl_type), p_dst);
}
frag->local_handle = local_handle;
} else {
local_handle = ((mca_spml_yoda_context_t*)l_mkey->spml_context)->registration;
frag->local_handle = NULL;
}
frag->rdma_segs[0].base_seg.seg_addr.lval = (uintptr_t) p_src;
getreq->p_dst = (uint64_t*) p_dst;
frag->size = ncopied;
OPAL_THREAD_ADD32(&mca_spml_yoda.n_active_gets, 1);
}
/**
* Initialize the remote data fragment
* with remote address data required for
* executing RDMA READ from a remote buffer.
*/
frag->rdma_req = getreq;
/**
* Do GET operation
*/
if (get_via_send) {
rc = mca_bml_base_send(bml_btl, des, MCA_SPML_YODA_GET);
if (1 == rc)
rc = OSHMEM_SUCCESS;
} else {
rc = mca_bml_base_get(bml_btl, p_dst, (uint64_t) (intptr_t) p_src, local_handle,
remote_handle, ncopied, 0, 0, mca_spml_yoda_get_completion, frag);
}
if (OPAL_UNLIKELY(OSHMEM_SUCCESS != rc)) {
if (OSHMEM_ERR_OUT_OF_RESOURCE == rc) {
/* No free resources, Block on completion here */
oshmem_request_wait_completion(&getreq->req_get.req_base.req_oshmem);
return OSHMEM_SUCCESS;
} else {
SPML_ERROR("oshmem_get: error %d", rc);
goto exit_fatal;
}
}
p_dst += ncopied;
p_src += ncopied;
OPAL_THREAD_ADD32(&get_holder.active_count, 1);
}
/* revisit if we really need this for self and sm */
/* if (YODA_BTL_SELF == ybtl->btl_type) */
opal_progress();
/* Wait for completion on request */
while (get_holder.active_count > 0)
oshmem_request_wait_completion(&getreq->req_get.req_base.req_oshmem);
return rc;
exit_fatal:
if (OSHMEM_SUCCESS != rc) {
oshmem_shmem_abort(rc);
}
return rc;
}
sshmem_mkey_t *mca_spml_ikrit_register(void* addr,
size_t size,
uint64_t shmid,
int *count)
{
int i;
sshmem_mkey_t *mkeys;
mxm_error_t err;
mxm_mem_key_t *m_key;
int my_rank = oshmem_my_proc_id();
*count = 0;
mkeys = (sshmem_mkey_t *) calloc(1, MXM_PTL_LAST * sizeof(*mkeys));
if (!mkeys) {
return NULL ;
}
for (i = 0; i < MXM_PTL_LAST; i++) {
mkeys[i].u.key = MAP_SEGMENT_SHM_INVALID;
switch (i) {
case MXM_PTL_SHM:
if ((int)shmid != MAP_SEGMENT_SHM_INVALID) {
mkeys[i].u.key = shmid;
mkeys[i].va_base = 0;
} else {
mkeys[i].len = 0;
mkeys[i].va_base = addr;
}
mkeys[i].spml_context = 0;
break;
case MXM_PTL_RDMA:
mkeys[i].va_base = addr;
mkeys[i].spml_context = 0;
if (mca_spml_ikrit.ud_only) {
mkeys[i].len = 0;
break;
}
err = mxm_mem_map(mca_spml_ikrit.mxm_context, &addr, &size, 0, 0, 0);
if (MXM_OK != err) {
SPML_ERROR("Failed to register memory: %s", mxm_error_string(err));
goto error_out;
}
mkeys[i].spml_context = (void *)(unsigned long)size;
m_key = malloc(sizeof(*m_key));
if (NULL == m_key) {
SPML_ERROR("Failed to allocate m_key memory");
goto error_out;
}
mkeys[i].len = sizeof(*m_key);
mkeys[i].u.data = m_key;
err = mxm_mem_get_key(mca_spml_ikrit.mxm_context, addr, m_key);
if (MXM_OK != err) {
SPML_ERROR("Failed to get memory key: %s", mxm_error_string(err));
goto error_out;
}
break;
default:
SPML_ERROR("unsupported PTL: %d", i);
goto error_out;
}
SPML_VERBOSE(5,
"rank %d ptl %d addr %p size %llu %s",
my_rank, i, addr, (unsigned long long)size,
mca_spml_base_mkey2str(&mkeys[i]));
mca_spml_ikrit_cache_mkeys(&mkeys[i], memheap_find_segnum(addr), my_rank, i);
}
*count = MXM_PTL_LAST;
return mkeys;
error_out:
mca_spml_ikrit_deregister(mkeys);
return NULL;
}
int mca_spml_ikrit_add_procs(ompi_proc_t** procs, size_t nprocs)
{
spml_ikrit_mxm_ep_conn_info_t *ep_info = NULL;
spml_ikrit_mxm_ep_conn_info_t *ep_hw_rdma_info = NULL;
spml_ikrit_mxm_ep_conn_info_t my_ep_info;
size_t mxm_addr_len = MXM_MAX_ADDR_LEN;
mxm_error_t err;
size_t i, n;
int rc = OSHMEM_ERROR;
ompi_proc_t *proc_self;
int my_rank = oshmem_my_proc_id();
OBJ_CONSTRUCT(&mca_spml_ikrit.active_peers, opal_list_t);
/* Allocate connection requests */
ep_info = calloc(sizeof(spml_ikrit_mxm_ep_conn_info_t), nprocs);
if (NULL == ep_info) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
if (mca_spml_ikrit.hw_rdma_channel) {
ep_hw_rdma_info = calloc(sizeof(spml_ikrit_mxm_ep_conn_info_t), nprocs);
if (NULL == ep_hw_rdma_info) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
}
mca_spml_ikrit.mxm_peers = (mxm_peer_t *) calloc(nprocs , sizeof(mxm_peer_t));
if (NULL == mca_spml_ikrit.mxm_peers) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
memset(&my_ep_info, 0, sizeof(my_ep_info));
if (mca_spml_ikrit.hw_rdma_channel) {
err = mxm_ep_get_address(mca_spml_ikrit.mxm_hw_rdma_ep, &my_ep_info.addr.ep_addr, &mxm_addr_len);
if (MXM_OK != err) {
orte_show_help("help-oshmem-spml-ikrit.txt", "unable to get endpoint address", true,
mxm_error_string(err));
rc = OSHMEM_ERROR;
goto bail;
}
oshmem_shmem_allgather(&my_ep_info, ep_hw_rdma_info,
sizeof(spml_ikrit_mxm_ep_conn_info_t));
}
err = mxm_ep_get_address(mca_spml_ikrit.mxm_ep, &my_ep_info.addr.ep_addr, &mxm_addr_len);
if (MXM_OK != err) {
orte_show_help("help-oshmem-spml-ikrit.txt", "unable to get endpoint address", true,
mxm_error_string(err));
rc = OSHMEM_ERROR;
goto bail;
}
oshmem_shmem_allgather(&my_ep_info, ep_info,
sizeof(spml_ikrit_mxm_ep_conn_info_t));
opal_progress_register(spml_ikrit_progress);
/* Get the EP connection requests for all the processes from modex */
for (n = 0; n < nprocs; ++n) {
/* mxm 2.0 keeps its connections on a list. Make sure
* that list have different order on every rank */
i = (my_rank + n) % nprocs;
mxm_peer_construct(&mca_spml_ikrit.mxm_peers[i]);
err = mxm_ep_connect(mca_spml_ikrit.mxm_ep, ep_info[i].addr.ep_addr, &mca_spml_ikrit.mxm_peers[i].mxm_conn);
if (MXM_OK != err) {
SPML_ERROR("MXM returned connect error: %s\n", mxm_error_string(err));
goto bail;
}
mxm_conn_ctx_set(mca_spml_ikrit.mxm_peers[i].mxm_conn, &mca_spml_ikrit.mxm_peers[i]);
if (mca_spml_ikrit.hw_rdma_channel) {
err = mxm_ep_connect(mca_spml_ikrit.mxm_hw_rdma_ep, ep_hw_rdma_info[i].addr.ep_addr, &mca_spml_ikrit.mxm_peers[i].mxm_hw_rdma_conn);
if (MXM_OK != err) {
SPML_ERROR("MXM returned connect error: %s\n", mxm_error_string(err));
goto bail;
}
} else {
mca_spml_ikrit.mxm_peers[i].mxm_hw_rdma_conn = mca_spml_ikrit.mxm_peers[i].mxm_conn;
}
}
if (ep_info)
free(ep_info);
if (ep_hw_rdma_info)
free(ep_hw_rdma_info);
if (mca_spml_ikrit.bulk_connect) {
/* Need a barrier to ensure remote peers already created connection */
oshmem_shmem_barrier();
mxm_ep_wireup(mca_spml_ikrit.mxm_ep);
}
proc_self = oshmem_proc_group_find(oshmem_group_all, my_rank);
/* identify local processes and change transport to SHM */
for (i = 0; i < nprocs; i++) {
if (procs[i]->super.proc_name.jobid != proc_self->super.proc_name.jobid ||
!OPAL_PROC_ON_LOCAL_NODE(procs[i]->super.proc_flags)) {
continue;
}
if (procs[i] == proc_self)
continue;
/* use zcopy for put/get via sysv shared memory with fallback to RDMA */
mca_spml_ikrit.mxm_peers[i].ptl_id = MXM_PTL_SHM;
}
SPML_VERBOSE(50, "*** ADDED PROCS ***");
return OSHMEM_SUCCESS;
bail:
if (ep_info)
free(ep_info);
if (ep_hw_rdma_info)
free(ep_hw_rdma_info);
SPML_ERROR("add procs FAILED rc=%d", rc);
return rc;
}
int mca_spml_ikrit_add_procs(ompi_proc_t** procs, size_t nprocs)
{
spml_ikrit_mxm_ep_conn_info_t *ep_info = NULL;
spml_ikrit_mxm_ep_conn_info_t *ep_hw_rdma_info = NULL;
spml_ikrit_mxm_ep_conn_info_t my_ep_info = {{0}};
#if MXM_API < MXM_VERSION(2,0)
mxm_conn_req_t *conn_reqs;
int timeout;
#else
size_t mxm_addr_len = MXM_MAX_ADDR_LEN;
#endif
mxm_error_t err;
size_t i, n;
int rc = OSHMEM_ERROR;
ompi_proc_t *proc_self;
int my_rank = oshmem_my_proc_id();
OBJ_CONSTRUCT(&mca_spml_ikrit.active_peers, opal_list_t);
/* Allocate connection requests */
#if MXM_API < MXM_VERSION(2,0)
conn_reqs = malloc(nprocs * sizeof(mxm_conn_req_t));
if (NULL == conn_reqs) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
memset(conn_reqs, 0x0, sizeof(mxm_conn_req_t));
#endif
ep_info = calloc(sizeof(spml_ikrit_mxm_ep_conn_info_t), nprocs);
if (NULL == ep_info) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
if (mca_spml_ikrit.hw_rdma_channel) {
ep_hw_rdma_info = calloc(sizeof(spml_ikrit_mxm_ep_conn_info_t), nprocs);
if (NULL == ep_hw_rdma_info) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
}
mca_spml_ikrit.mxm_peers = (mxm_peer_t **) malloc(nprocs
* sizeof(*(mca_spml_ikrit.mxm_peers)));
if (NULL == mca_spml_ikrit.mxm_peers) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
#if MXM_API < MXM_VERSION(2,0)
if (OSHMEM_SUCCESS
!= spml_ikrit_get_ep_address(&my_ep_info, MXM_PTL_SELF)) {
rc = OSHMEM_ERROR;
goto bail;
}
if (OSHMEM_SUCCESS
!= spml_ikrit_get_ep_address(&my_ep_info, MXM_PTL_RDMA)) {
rc = OSHMEM_ERROR;
goto bail;
}
#else
if (mca_spml_ikrit.hw_rdma_channel) {
err = mxm_ep_get_address(mca_spml_ikrit.mxm_hw_rdma_ep, &my_ep_info.addr.ep_addr, &mxm_addr_len);
if (MXM_OK != err) {
orte_show_help("help-oshmem-spml-ikrit.txt", "unable to get endpoint address", true,
mxm_error_string(err));
rc = OSHMEM_ERROR;
goto bail;
}
oshmem_shmem_allgather(&my_ep_info, ep_hw_rdma_info,
sizeof(spml_ikrit_mxm_ep_conn_info_t));
}
err = mxm_ep_get_address(mca_spml_ikrit.mxm_ep, &my_ep_info.addr.ep_addr, &mxm_addr_len);
if (MXM_OK != err) {
orte_show_help("help-oshmem-spml-ikrit.txt", "unable to get endpoint address", true,
mxm_error_string(err));
rc = OSHMEM_ERROR;
goto bail;
}
#endif
oshmem_shmem_allgather(&my_ep_info, ep_info,
sizeof(spml_ikrit_mxm_ep_conn_info_t));
opal_progress_register(spml_ikrit_progress);
/* Get the EP connection requests for all the processes from modex */
for (n = 0; n < nprocs; ++n) {
/* mxm 2.0 keeps its connections on a list. Make sure
* that list have different order on every rank */
i = (my_rank + n) % nprocs;
mca_spml_ikrit.mxm_peers[i] = OBJ_NEW(mxm_peer_t);
if (NULL == mca_spml_ikrit.mxm_peers[i]) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
mca_spml_ikrit.mxm_peers[i]->pe = i;
#if MXM_API < MXM_VERSION(2,0)
conn_reqs[i].ptl_addr[MXM_PTL_SELF] =
(struct sockaddr *) &ep_info[i].addr.ptl_addr[MXM_PTL_SELF];
conn_reqs[i].ptl_addr[MXM_PTL_SHM] = NULL;
conn_reqs[i].ptl_addr[MXM_PTL_RDMA] =
(struct sockaddr *) &ep_info[i].addr.ptl_addr[MXM_PTL_RDMA];
#else
err = mxm_ep_connect(mca_spml_ikrit.mxm_ep, ep_info[i].addr.ep_addr, &mca_spml_ikrit.mxm_peers[i]->mxm_conn);
if (MXM_OK != err) {
SPML_ERROR("MXM returned connect error: %s\n", mxm_error_string(err));
goto bail;
}
if (OSHMEM_SUCCESS != create_ptl_idx(i))
goto bail;
mxm_conn_ctx_set(mca_spml_ikrit.mxm_peers[i]->mxm_conn, mca_spml_ikrit.mxm_peers[i]);
if (mca_spml_ikrit.hw_rdma_channel) {
err = mxm_ep_connect(mca_spml_ikrit.mxm_hw_rdma_ep, ep_hw_rdma_info[i].addr.ep_addr, &mca_spml_ikrit.mxm_peers[i]->mxm_hw_rdma_conn);
if (MXM_OK != err) {
SPML_ERROR("MXM returned connect error: %s\n", mxm_error_string(err));
goto bail;
}
} else {
mca_spml_ikrit.mxm_peers[i]->mxm_hw_rdma_conn = mca_spml_ikrit.mxm_peers[i]->mxm_conn;
}
#endif
}
#if MXM_API < MXM_VERSION(2,0)
/* Connect to remote peers */
if (mxm_get_version() < MXM_VERSION(1,5)) {
timeout = 1000;
} else {
timeout = -1;
}
err = mxm_ep_connect(mca_spml_ikrit.mxm_ep, conn_reqs, nprocs, timeout);
if (MXM_OK != err) {
SPML_ERROR("MXM returned connect error: %s\n", mxm_error_string(err));
for (i = 0; i < nprocs; ++i) {
if (MXM_OK != conn_reqs[i].error) {
SPML_ERROR("MXM EP connect to %s error: %s\n",
procs[i]->proc_hostname, mxm_error_string(conn_reqs[i].error));
}
}
rc = OSHMEM_ERR_CONNECTION_FAILED;
goto bail;
}
/* Save returned connections */
for (i = 0; i < nprocs; ++i) {
mca_spml_ikrit.mxm_peers[i]->mxm_conn = conn_reqs[i].conn;
if (OSHMEM_SUCCESS != create_ptl_idx(i)) {
rc = OSHMEM_ERR_CONNECTION_FAILED;
goto bail;
}
mxm_conn_ctx_set(conn_reqs[i].conn, mca_spml_ikrit.mxm_peers[i]);
}
if (conn_reqs)
free(conn_reqs);
#endif
if (ep_info)
free(ep_info);
if (ep_hw_rdma_info)
free(ep_hw_rdma_info);
#if MXM_API >= MXM_VERSION(2,0)
if (mca_spml_ikrit.bulk_connect) {
/* Need a barrier to ensure remote peers already created connection */
oshmem_shmem_barrier();
mxm_ep_wireup(mca_spml_ikrit.mxm_ep);
}
#endif
proc_self = oshmem_proc_group_find(oshmem_group_all, my_rank);
/* identify local processes and change transport to SHM */
for (i = 0; i < nprocs; i++) {
if (procs[i]->super.proc_name.jobid != proc_self->super.proc_name.jobid ||
!OPAL_PROC_ON_LOCAL_NODE(procs[i]->super.proc_flags)) {
continue;
}
if (procs[i] == proc_self)
continue;
/* use zcopy for put/get via sysv shared memory */
OSHMEM_PROC_DATA(procs[i])->transport_ids[0] = MXM_PTL_SHM;
OSHMEM_PROC_DATA(procs[i])->transport_ids[1] = MXM_PTL_RDMA;
OSHMEM_PROC_DATA(procs[i])->num_transports = 2;
}
SPML_VERBOSE(50, "*** ADDED PROCS ***");
return OSHMEM_SUCCESS;
bail:
#if MXM_API < MXM_VERSION(2,0)
if (conn_reqs)
free(conn_reqs);
#endif
if (ep_info)
free(ep_info);
if (ep_hw_rdma_info)
free(ep_hw_rdma_info);
SPML_ERROR("add procs FAILED rc=%d", rc);
return rc;
}
