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;
}
int mca_spml_ucx_fence(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;
}
int mca_spml_ucx_quiet(shmem_ctx_t ctx)
{
int ret;
mca_spml_ucx_ctx_t *ucx_ctx = (mca_spml_ucx_ctx_t *)ctx;
ret = opal_common_ucx_worker_flush(ucx_ctx->ucp_worker);
if (OMPI_SUCCESS != ret) {
oshmem_shmem_abort(-1);
return ret;
}
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);
SPML_IKRIT_MXM_POST_SEND(get_req->mxm_req);
return OSHMEM_SUCCESS;
}
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;
}
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_ucx_fence(shmem_ctx_t ctx)
{
ucs_status_t err;
mca_spml_ucx_ctx_t *ucx_ctx = (mca_spml_ucx_ctx_t *)ctx;
err = ucp_worker_fence(ucx_ctx->ucp_worker);
if (UCS_OK != err) {
SPML_UCX_ERROR("fence failed: %s", ucs_status_string(err));
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
return OSHMEM_SUCCESS;
}
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;
}
static void unpack_remote_mkeys(opal_buffer_t *msg, int remote_pe)
{
int32_t cnt;
int32_t n;
int32_t tr_id;
int i;
oshmem_proc_t *proc;
proc = oshmem_proc_group_find(oshmem_group_all, remote_pe);
cnt = 1;
opal_dss.unpack(msg, &n, &cnt, OPAL_UINT32);
for (i = 0; i < n; i++) {
cnt = 1;
opal_dss.unpack(msg, &tr_id, &cnt, OPAL_UINT32);
cnt = 1;
opal_dss.unpack(msg,
&memheap_oob.mkeys[tr_id].va_base,
&cnt,
OPAL_UINT64);
if (0 == memheap_oob.mkeys[tr_id].va_base) {
cnt = 1;
opal_dss.unpack(msg, &memheap_oob.mkeys[tr_id].u.key, &cnt, OPAL_UINT64);
if (OPAL_PROC_ON_LOCAL_NODE(proc->super.proc_flags)) {
memheap_attach_segment(&memheap_oob.mkeys[tr_id], tr_id);
}
} else {
cnt = 1;
opal_dss.unpack(msg, &memheap_oob.mkeys[tr_id].len, &cnt, OPAL_UINT16);
if (0 < memheap_oob.mkeys[tr_id].len) {
memheap_oob.mkeys[tr_id].u.data = malloc(memheap_oob.mkeys[tr_id].len);
if (NULL == memheap_oob.mkeys[tr_id].u.data) {
MEMHEAP_ERROR("Failed allocate %d bytes", memheap_oob.mkeys[tr_id].len);
oshmem_shmem_abort(-1);
}
cnt = memheap_oob.mkeys[tr_id].len;
opal_dss.unpack(msg, memheap_oob.mkeys[tr_id].u.data, &cnt, OPAL_BYTE);
MCA_SPML_CALL(rmkey_unpack(&memheap_oob.mkeys[tr_id], remote_pe));
} else {
memheap_oob.mkeys[tr_id].u.key = MAP_SEGMENT_SHM_INVALID;
}
}
MEMHEAP_VERBOSE(5,
"tr_id: %d %s",
tr_id, mca_spml_base_mkey2str(&memheap_oob.mkeys[tr_id]));
}
}
/*
* Initialize module on the communicator
*/
static int mca_scoll_mpi_module_enable(mca_scoll_base_module_t *module,
oshmem_group_t *osh_group)
{
if (OSHMEM_SUCCESS != mca_scoll_mpi_save_coll_handlers(module, osh_group)){
MPI_COLL_ERROR("MPI module enable failed - aborting to prevent inconsistent application state");
/* There's no modules available */
opal_show_help("help-oshmem-scoll-mpi.txt",
"module_enable:fatal", true,
"MPI module enable failed - aborting to prevent inconsistent application state");
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
return OSHMEM_SUCCESS;
}
/*
* * Initialize module on the communicator
* */
static int mca_scoll_fca_module_enable(mca_scoll_base_module_t *module,
struct oshmem_group_t *comm)
{
mca_scoll_fca_module_t *fca_module = (mca_scoll_fca_module_t*) module;
int rc;
fca_module->comm = comm;
fca_module->rank = comm->my_pe;
rc = mca_scoll_fca_get_fca_lib(comm);
if (rc != OSHMEM_SUCCESS)
goto exit_fatal;
rc = _save_coll_handlers(fca_module);
if (rc != OSHMEM_SUCCESS)
goto exit_fatal;
rc = _get_local_ranks(fca_module);
if (rc != OSHMEM_SUCCESS)
goto exit_fatal;
rc = _create_fca_comm(fca_module);
if (rc != OSHMEM_SUCCESS)
goto exit_fatal;
FCA_MODULE_VERBOSE(fca_module, 1, "FCA Module initialized");
return OMPI_SUCCESS;
exit_fatal:
/* it is possible that other pe(s) succesfully enabled fca.
* So differnt frameworks will be used for collective ops
*/
FCA_ERROR("FCA module enable failed - aborting to prevent inconsistent application state");
/* There's no modules available */
opal_show_help("help-oshmem-scoll-fca.txt",
"module_enable:fatal", true,
"FCA module enable failed - aborting to prevent inconsistent application state");
oshmem_shmem_abort(-1);
return OMPI_ERROR;
}
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;
}
static inline void _shmem_init(int required, int *provided)
{
int err = OSHMEM_SUCCESS;
if (oshmem_shmem_initialized) {
/*
* SPEC: If start_pes() is called multiple times, subsequent calls have no effect.
*/
return;
}
err = oshmem_shmem_init(0, NULL, required, provided);
if (OSHMEM_SUCCESS != err) {
/* since spec does not propagete error to user we can only abort */
SHMEM_API_ERROR("SHMEM failed to initialize - aborting");
oshmem_shmem_abort(-1);
}
OPAL_CR_INIT_LIBRARY();
#if HAVE_ON_EXIT
on_exit(shmem_onexit, NULL);
#endif
}
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.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;
}
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;
}
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);
}
int cache_group(oshmem_group_t *group,
int PE_start,
int logPE_stride,
int PE_size)
{
oshmem_group_cache_t *cached_group = NULL;
cached_group = OBJ_NEW(oshmem_group_cache_t);
#if OPAL_ENABLE_DEBUG
cached_group->item.opal_list_item_belong_to = NULL;
cached_group->item.opal_list_item_refcount = 0;
#endif
cached_group->group = group;
cached_group->cache_id[0] = PE_start;
cached_group->cache_id[1] = logPE_stride;
cached_group->cache_id[2] = PE_size;
if (opal_list_get_size(&oshmem_group_cache_list)
< oshmem_group_cache_size) {
opal_list_append(&oshmem_group_cache_list,
(opal_list_item_t *)cached_group);
} else {
#if ABORT_ON_CACHE_OVERFLOW
opal_output(0,
"error: group cache overflow on rank %i: cache_size = %u: try encreasing oshmem_group_cache_size mca parameter",
group->my_pe,
oshmem_group_cache_size);
oshmem_shmem_abort(-1);
#else
/*This part of code makes FIFO group cache management. Define ABORT_ON_CACHE_OVERFLOW as 0 to enable this.*/
oshmem_group_cache_t *cached_group_to_remove = (oshmem_group_cache_t *)opal_list_remove_first(&oshmem_group_cache_list);
oshmem_proc_group_destroy(cached_group_to_remove->group);
OBJ_RELEASE(cached_group_to_remove);
opal_list_append(&oshmem_group_cache_list,(opal_list_item_t *)cached_group);
#endif
}
return OSHMEM_SUCCESS;
}
static void memheap_attach_segment(sshmem_mkey_t *mkey, int tr_id)
{
/* process special case when va was got using sshmem
* this case is notable for:
* - key is set as (seg_id);
* - va_base is set as 0;
* - len is set as 0;
*/
assert(mkey->va_base == 0);
assert(mkey->len == 0);
MEMHEAP_VERBOSE(5,
"shared memory usage tr_id: %d va_base: 0x%p len: %d key %llx",
tr_id,
mkey->va_base, mkey->len, (unsigned long long)mkey->u.key);
mca_sshmem_segment_attach(&(memheap_map->mem_segs[HEAP_SEG_INDEX]), mkey);
if ((void *) -1 == (void *) mkey->va_base) {
MEMHEAP_ERROR("tr_id: %d key %llx attach failed: errno = %d",
tr_id, (unsigned long long)mkey->u.key, errno);
oshmem_shmem_abort(-1);
}
}
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;
}
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;
}
void mca_memheap_modex_recv_all(void)
{
int i;
int j;
int nprocs, my_pe;
oshmem_proc_t *proc;
mca_spml_mkey_t *mkey;
void* dummy_rva;
if (!mca_memheap_base_key_exchange)
return;
/* init rkey cache */
nprocs = oshmem_num_procs();
my_pe = oshmem_my_proc_id();
/* Note:
* Doing exchange via rml till we figure out problem with grpcomm.modex and barrier
*/
for (i = 0; i < nprocs; i++) {
if (i == my_pe)
continue;
proc = oshmem_proc_group_find(oshmem_group_all, i);
for (j = 0; j < memheap_map->n_segments; j++) {
mkey =
mca_memheap_base_get_cached_mkey(i,
memheap_map->mem_segs[j].start,
proc->transport_ids[0],
&dummy_rva);
if (!mkey) {
MEMHEAP_ERROR("Failed to receive mkeys");
oshmem_shmem_abort(-1);
}
}
}
/*
* There is an issue with orte_grpcomm.barrier usage as
* ess/pmi directs to use grpcomm/pmi in case slurm srun() call grpcomm/pmi calls PMI_Barrier()
* that is a function of external library.
* There is no opal_progress() in such way. As a result slow PEs send a request (MEMHEAP_RKEY_REQ) to
* fast PEs waiting on barrier and do not get a respond (MEMHEAP_RKEY_RESP).
*
* there are following ways to solve one:
* 1. calculate requests from remote PEs and do ORTE_PROGRESSED_WAIT waiting for expected value;
* 2. use shmem_barrier_all();
* 3. rework pmi/barrier to use opal_progress();
* 4. use orte_grpcomm.barrier carefully;
*
* It seems there is no need to use orte_grpcomm.barrier here
*/
if (memheap_map->mem_segs[HEAP_SEG_INDEX].shmid != MEMHEAP_SHM_INVALID) {
/* unfortunately we must do barrier here to assure that everyone are attached to our segment
* good thing that this code path only invoked on older linuxes (-mca shmalloc_use_hugepages 3|4)
* try to minimize damage here by waiting 5 seconds and doing progress
*/
shmem_barrier_all();
/* keys exchanged, segments attached, now we can safely cleanup */
if (memheap_map->mem_segs[HEAP_SEG_INDEX].type
== MAP_SEGMENT_ALLOC_SHM) {
shmctl(memheap_map->mem_segs[HEAP_SEG_INDEX].shmid,
IPC_RMID,
NULL );
}
}
}
void mca_memheap_modex_recv_all(void)
{
int i;
int j;
int nprocs, my_pe;
opal_buffer_t *msg = NULL;
void *send_buffer = NULL;
char *rcv_buffer = NULL;
int size;
int *rcv_size = NULL;
int *rcv_n_transports = NULL;
int *rcv_offsets = NULL;
int rc = OSHMEM_SUCCESS;
size_t buffer_size;
if (!mca_memheap_base_key_exchange) {
oshmem_shmem_barrier();
return;
}
nprocs = oshmem_num_procs();
my_pe = oshmem_my_proc_id();
/* buffer allocation for num_transports
* message sizes and offsets */
rcv_size = (int *)malloc(nprocs * sizeof(int));
if (NULL == rcv_size) {
MEMHEAP_ERROR("failed to get rcv_size buffer");
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto exit_fatal;
}
rcv_offsets = (int *)malloc(nprocs * sizeof(int));
if (NULL == rcv_offsets) {
MEMHEAP_ERROR("failed to get rcv_offsets buffer");
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto exit_fatal;
}
rcv_n_transports = (int *)malloc(nprocs * sizeof(int));
if (NULL == rcv_offsets) {
MEMHEAP_ERROR("failed to get rcv_offsets buffer");
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto exit_fatal;
}
/* serialize our own mkeys */
msg = OBJ_NEW(opal_buffer_t);
if (NULL == msg) {
MEMHEAP_ERROR("failed to get msg buffer");
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto exit_fatal;
}
for (j = 0; j < memheap_map->n_segments; j++) {
pack_local_mkeys(msg, 0, j, 1);
}
/* we assume here that int32_t returned by opal_dss.unload
* is equal to size of int we use for MPI_Allgather, MPI_Allgatherv */
assert(sizeof(int32_t) == sizeof(int));
/* Do allgather */
opal_dss.unload(msg, &send_buffer, &size);
MEMHEAP_VERBOSE(1, "local keys packed into %d bytes, %d segments", size, memheap_map->n_segments);
/* we need to send num_transports and message sizes separately
* since message sizes depend on types of btl used */
rc = oshmem_shmem_allgather(&memheap_map->num_transports, rcv_n_transports, sizeof(int));
if (MPI_SUCCESS != rc) {
MEMHEAP_ERROR("allgather failed");
goto exit_fatal;
}
rc = oshmem_shmem_allgather(&size, rcv_size, sizeof(int));
if (MPI_SUCCESS != rc) {
MEMHEAP_ERROR("allgather failed");
goto exit_fatal;
}
/* calculating offsets (displacements) for allgatherv */
rcv_offsets[0] = 0;
for (i = 1; i < nprocs; i++) {
rcv_offsets[i] = rcv_offsets[i - 1] + rcv_size[i - 1];
}
buffer_size = rcv_offsets[nprocs - 1] + rcv_size[nprocs - 1];
rcv_buffer = malloc (buffer_size);
if (NULL == rcv_buffer) {
MEMHEAP_ERROR("failed to allocate recieve buffer");
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto exit_fatal;
}
rc = oshmem_shmem_allgatherv(send_buffer, rcv_buffer, size, rcv_size, rcv_offsets);
if (MPI_SUCCESS != rc) {
free (rcv_buffer);
MEMHEAP_ERROR("allgatherv failed");
goto exit_fatal;
}
opal_dss.load(msg, rcv_buffer, buffer_size);
/* deserialize mkeys */
OPAL_THREAD_LOCK(&memheap_oob.lck);
for (i = 0; i < nprocs; i++) {
if (i == my_pe) {
continue;
}
msg->unpack_ptr = (void *)((intptr_t) msg->base_ptr + rcv_offsets[i]);
for (j = 0; j < memheap_map->n_segments; j++) {
map_segment_t *s;
s = &memheap_map->mem_segs[j];
if (NULL != s->mkeys_cache[i]) {
MEMHEAP_VERBOSE(10, "PE%d: segment%d already exists, mkey will be replaced", i, j);
} else {
s->mkeys_cache[i] = (sshmem_mkey_t *) calloc(rcv_n_transports[i],
sizeof(sshmem_mkey_t));
if (NULL == s->mkeys_cache[i]) {
MEMHEAP_ERROR("PE%d: segment%d: Failed to allocate mkeys cache entry", i, j);
oshmem_shmem_abort(-1);
}
}
memheap_oob.mkeys = s->mkeys_cache[i];
unpack_remote_mkeys(msg, i);
}
}
OPAL_THREAD_UNLOCK(&memheap_oob.lck);
exit_fatal:
if (rcv_size) {
free(rcv_size);
}
if (rcv_offsets) {
free(rcv_offsets);
}
if (rcv_n_transports) {
free(rcv_n_transports);
}
if (send_buffer) {
free(send_buffer);
}
if (msg) {
OBJ_RELEASE(msg);
}
/* This function requires abort in any error case */
if (OSHMEM_SUCCESS != rc) {
oshmem_shmem_abort(rc);
}
}
sshmem_mkey_t *mca_spml_ucx_register(void* addr,
size_t size,
uint64_t shmid,
int *count)
{
sshmem_mkey_t *mkeys;
ucs_status_t status;
spml_ucx_mkey_t *ucx_mkey;
size_t len;
ucp_mem_map_params_t mem_map_params;
int segno;
map_segment_t *mem_seg;
unsigned flags;
int my_pe = oshmem_my_proc_id();
*count = 0;
mkeys = (sshmem_mkey_t *) calloc(1, sizeof(*mkeys));
if (!mkeys) {
return NULL;
}
segno = memheap_find_segnum(addr);
mem_seg = memheap_find_seg(segno);
ucx_mkey = &mca_spml_ucx_ctx_default.ucp_peers[my_pe].mkeys[segno].key;
mkeys[0].spml_context = ucx_mkey;
/* if possible use mem handle already created by ucx allocator */
if (MAP_SEGMENT_ALLOC_UCX != mem_seg->type) {
flags = 0;
if (mca_spml_ucx.heap_reg_nb && memheap_is_va_in_segment(addr, HEAP_SEG_INDEX)) {
flags = UCP_MEM_MAP_NONBLOCK;
}
mem_map_params.field_mask = UCP_MEM_MAP_PARAM_FIELD_ADDRESS |
UCP_MEM_MAP_PARAM_FIELD_LENGTH |
UCP_MEM_MAP_PARAM_FIELD_FLAGS;
mem_map_params.address = addr;
mem_map_params.length = size;
mem_map_params.flags = flags;
status = ucp_mem_map(mca_spml_ucx.ucp_context, &mem_map_params, &ucx_mkey->mem_h);
if (UCS_OK != status) {
goto error_out;
}
} else {
ucx_mkey->mem_h = (ucp_mem_h)mem_seg->context;
}
status = ucp_rkey_pack(mca_spml_ucx.ucp_context, ucx_mkey->mem_h,
&mkeys[0].u.data, &len);
if (UCS_OK != status) {
goto error_unmap;
}
if (len >= 0xffff) {
SPML_UCX_ERROR("packed rkey is too long: %llu >= %d",
(unsigned long long)len,
0xffff);
oshmem_shmem_abort(-1);
}
status = ucp_ep_rkey_unpack(mca_spml_ucx_ctx_default.ucp_peers[oshmem_group_self->my_pe].ucp_conn,
mkeys[0].u.data,
&ucx_mkey->rkey);
if (UCS_OK != status) {
SPML_UCX_ERROR("failed to unpack rkey");
goto error_unmap;
}
mkeys[0].len = len;
mkeys[0].va_base = addr;
*count = 1;
mca_spml_ucx_cache_mkey(&mca_spml_ucx_ctx_default, &mkeys[0], segno, my_pe);
return mkeys;
error_unmap:
ucp_mem_unmap(mca_spml_ucx.ucp_context, ucx_mkey->mem_h);
error_out:
free(mkeys);
return NULL ;
}
/*
* * Invoked when there's a new communicator that has been created.
* * Look at the communicator and decide which set of functions and
* * priority we want to return.
* */
mca_scoll_base_module_t *
mca_scoll_fca_comm_query(struct oshmem_group_t *comm, int *priority)
{
mca_scoll_base_module_t *module;
int size = comm->proc_count;
int local_peers = 0;
mca_scoll_fca_module_t *fca_module;
*priority = 0;
module = NULL;
if (!mca_scoll_fca_component.fca_enable) {
FCA_VERBOSE(20, "FCA is disable on user request => exiting");
goto exit;
}
if (mca_memheap.memheap_component == NULL ) {
FCA_VERBOSE(20, "No memheap => exiting");
goto exit;
}
if (NULL == mca_scoll_fca_component.ret) {
MCA_MEMHEAP_CALL(private_alloc(sizeof(int),(void **)&mca_scoll_fca_component.ret));
MCA_MEMHEAP_CALL(private_alloc(oshmem_group_all->proc_count*sizeof(*mca_scoll_fca_component.rcounts), (void **)&mca_scoll_fca_component.rcounts ));
MCA_MEMHEAP_CALL(private_alloc(/*info_size*/20,&mca_scoll_fca_component.my_info_exchangeable));
MCA_MEMHEAP_CALL(private_alloc(sizeof(fca_comm_desc_t), &mca_scoll_fca_component.fca_comm_desc_exchangeable));
}
if (size < mca_scoll_fca_component.fca_np) {
FCA_VERBOSE(20,
"size(%d) < fca_np(%d)",
size, mca_scoll_fca_component.fca_np);
goto exit;
}
if (size < 2) {
FCA_VERBOSE(20, "size(%d) < 2", size);
goto exit;
}
if (!have_remote_peers(comm,
size,
&local_peers) /* || OMPI_COMM_IS_INTER(comm)*/) {
FCA_VERBOSE(1,
"all peers in group are on the same node, fca disabled\n");
goto exit;
}
fca_module = OBJ_NEW(mca_scoll_fca_module_t);
if (!fca_module) {
goto exit_fatal;
}
fca_module->super.scoll_module_enable = mca_scoll_fca_module_enable;
fca_module->super.scoll_collect =
mca_scoll_fca_component.fca_enable_allgather ?
mca_scoll_fca_collect : NULL;
fca_module->super.scoll_reduce =
mca_scoll_fca_component.fca_enable_allreduce ?
mca_scoll_fca_reduce : NULL;
fca_module->super.scoll_barrier =
mca_scoll_fca_component.fca_enable_barrier ? mca_scoll_fca_barrier :
NULL;
fca_module->super.scoll_broadcast =
mca_scoll_fca_component.fca_enable_bcast ? mca_scoll_fca_broadcast :
NULL;
*priority = mca_scoll_fca_component.fca_priority;
module = &fca_module->super;
exit:
FCA_VERBOSE(4,
"Query FCA module for comm %p size %d rank %d local_peers=%d: priority=%d %s",
(void *)comm, size, comm->my_pe, local_peers, *priority, module ? "enabled" : "disabled");
return module;
exit_fatal:
/* it is possible that other pe(s) succesfully initialized fca.
* So differnt frameworks will be used for collective ops
*/
FCA_ERROR("FCA module query failed - aborting");
oshmem_shmem_abort(-1);
return NULL ;
}
int mca_atomic_mxm_fadd(void *target,
void *prev,
const void *value,
size_t nlong,
int pe,
struct oshmem_op_t *op)
{
unsigned my_pe;
uint8_t nlong_order;
void *remote_addr;
int ptl_id;
mxm_send_req_t sreq;
mxm_error_t mxm_err;
sshmem_mkey_t *r_mkey;
static char dummy_buf[8];
my_pe = oshmem_my_proc_id();
ptl_id = -1;
mxm_err = MXM_OK;
if (!target || !value) {
ATOMIC_ERROR("[#%d] target or value are not defined", my_pe);
oshmem_shmem_abort(-1);
return OSHMEM_ERR_BAD_PARAM;
}
if ((pe < 0) || (pe >= oshmem_num_procs())) {
ATOMIC_ERROR("[#%d] PE=%d not valid", my_pe, pe);
oshmem_shmem_abort(-1);
return OSHMEM_ERR_BAD_PARAM;
}
switch (nlong) {
case 1:
nlong_order = 0;
break;
case 2:
nlong_order = 1;
break;
case 4:
nlong_order = 2;
break;
case 8:
nlong_order = 3;
break;
default:
ATOMIC_ERROR("[#%d] Type size must be 1/2/4 or 8 bytes.", my_pe);
oshmem_shmem_abort(-1);
return OSHMEM_ERR_BAD_PARAM;
}
ptl_id = oshmem_proc_group_all(pe)->transport_ids[0];
if (MXM_PTL_SHM == ptl_id) {
ptl_id = MXM_PTL_RDMA;
}
r_mkey = mca_memheap.memheap_get_cached_mkey(pe,
target,
ptl_id,
&remote_addr);
if (!r_mkey) {
ATOMIC_ERROR("[#%d] %p is not address of symmetric variable",
my_pe, target);
oshmem_shmem_abort(-1);
return OSHMEM_ERR_BAD_PARAM;
}
/* mxm request init */
sreq.base.state = MXM_REQ_NEW;
sreq.base.mq = mca_spml_self->mxm_mq;
sreq.base.conn = mca_spml_self->mxm_peers[pe]->mxm_hw_rdma_conn;
sreq.base.completed_cb = NULL;
sreq.base.data_type = MXM_REQ_DATA_BUFFER;
sreq.op.atomic.remote_vaddr = (uintptr_t) remote_addr;
#if MXM_API < MXM_VERSION(2,0)
sreq.op.atomic.remote_memh = MXM_INVALID_MEM_HANDLE;
memcpy(&sreq.op.atomic.value8, value, nlong);
#else
sreq.op.atomic.remote_mkey = to_mxm_mkey(r_mkey);
memcpy(&sreq.op.atomic.value, value, nlong);
#endif
sreq.op.atomic.order = nlong_order;
/* Do we need atomic 'add' or atomic 'fetch and add'? */
if (NULL == prev) {
sreq.base.data.buffer.ptr = dummy_buf;
sreq.base.data.buffer.length = nlong;
sreq.base.data.buffer.memh = MXM_INVALID_MEM_HANDLE;
#if MXM_API < MXM_VERSION(2,0)
sreq.base.flags = MXM_REQ_FLAG_SEND_SYNC;
sreq.opcode = MXM_REQ_OP_ATOMIC_ADD;
#else
sreq.flags = 0;
sreq.opcode = MXM_REQ_OP_ATOMIC_FADD;
#endif
} else {
sreq.base.data.buffer.ptr = prev;
sreq.base.data.buffer.length = nlong;
sreq.base.data.buffer.memh = MXM_INVALID_MEM_HANDLE;
#if MXM_API < MXM_VERSION(2,0)
sreq.base.flags = 0;
#else
sreq.flags = 0;
#endif
sreq.opcode = MXM_REQ_OP_ATOMIC_FADD;
}
if (MXM_OK != (mxm_err = mxm_req_send(&sreq))) {
ATOMIC_ERROR("[#%d] mxm_req_send failed, mxm_error = %d",
my_pe, mxm_err);
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
mxm_req_wait(&sreq.base);
if (MXM_OK != sreq.base.error) {
ATOMIC_ERROR("[#%d] mxm_req_wait got non MXM_OK error: %d",
my_pe, sreq.base.error);
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
return OSHMEM_SUCCESS;
}
sshmem_mkey_t *mca_spml_ucx_register(void* addr,
size_t size,
uint64_t shmid,
int *count)
{
sshmem_mkey_t *mkeys;
ucs_status_t err;
spml_ucx_mkey_t *ucx_mkey;
size_t len;
*count = 0;
mkeys = (sshmem_mkey_t *) calloc(1, sizeof(*mkeys));
if (!mkeys) {
return NULL ;
}
ucx_mkey = (spml_ucx_mkey_t *)malloc(sizeof(*ucx_mkey));
if (!ucx_mkey) {
goto error_out;
}
mkeys[0].spml_context = ucx_mkey;
err = ucp_mem_map(mca_spml_ucx.ucp_context,
&addr, size, 0, &ucx_mkey->mem_h);
if (UCS_OK != err) {
goto error_out1;
}
err = ucp_rkey_pack(mca_spml_ucx.ucp_context, ucx_mkey->mem_h,
&mkeys[0].u.data, &len);
if (UCS_OK != err) {
goto error_unmap;
}
if (len >= 0xffff) {
SPML_ERROR("packed rkey is too long: %llu >= %d",
(unsigned long long)len,
0xffff);
oshmem_shmem_abort(-1);
}
err = ucp_ep_rkey_unpack(mca_spml_ucx.ucp_peers[oshmem_group_self->my_pe].ucp_conn,
mkeys[0].u.data,
&ucx_mkey->rkey);
if (UCS_OK != err) {
SPML_ERROR("failed to unpack rkey");
goto error_unmap;
}
mkeys[0].len = len;
mkeys[0].va_base = addr;
*count = 1;
return mkeys;
error_unmap:
ucp_mem_unmap(mca_spml_ucx.ucp_context, ucx_mkey->mem_h);
error_out1:
free(ucx_mkey);
error_out:
free(mkeys);
return NULL ;
}
static void memheap_attach_segment(mca_spml_mkey_t *mkey, int tr_id)
{
/* process special case when va was got using shmget(IPC_PRIVATE)
* this case is notable for:
* - key is set as (type|shmid);
* - va_base is set as 0;
*/
if (!mkey->va_base
&& ((int) MEMHEAP_SHM_GET_ID(mkey->key) != MEMHEAP_SHM_INVALID)) {
MEMHEAP_VERBOSE(5,
"shared memory usage tr_id: %d key %llx base_va %p shmid 0x%X|0x%X",
tr_id,
(unsigned long long)mkey->key,
mkey->va_base,
MEMHEAP_SHM_GET_TYPE(mkey->key),
MEMHEAP_SHM_GET_ID(mkey->key));
if (MEMHEAP_SHM_GET_TYPE(mkey->key) == MAP_SEGMENT_ALLOC_SHM) {
mkey->va_base = shmat(MEMHEAP_SHM_GET_ID(mkey->key),
0,
0);
} else if (MEMHEAP_SHM_GET_TYPE(mkey->key) == MAP_SEGMENT_ALLOC_IBV) {
#if defined(MPAGE_ENABLE) && (MPAGE_ENABLE > 0)
openib_device_t *device = NULL;
struct ibv_mr *ib_mr;
void *addr;
static int mr_count;
int access_flag = IBV_ACCESS_LOCAL_WRITE |
IBV_ACCESS_REMOTE_WRITE |
IBV_ACCESS_REMOTE_READ |
IBV_ACCESS_NO_RDMA;
device = (openib_device_t *)memheap_map->mem_segs[HEAP_SEG_INDEX].context;
assert(device);
/* workaround mtt problem - request aligned addresses */
++mr_count;
addr = (void *)(mca_memheap_base_start_address + mca_memheap_base_mr_interleave_factor*1024ULL*1024ULL*1024ULL*mr_count);
ib_mr = ibv_reg_shared_mr(MEMHEAP_SHM_GET_ID(mkey->key),
device->ib_pd, addr, access_flag);
if (NULL == ib_mr)
{
mkey->va_base = (void*)-1;
MEMHEAP_ERROR("error to ibv_reg_shared_mr() errno says %d: %s",
errno, strerror(errno));
}
else
{
if (ib_mr->addr != addr) {
MEMHEAP_WARN("Failed to map shared region to address %p got addr %p. Try to increase 'memheap_mr_interleave_factor' from %d", addr, ib_mr->addr, mca_memheap_base_mr_interleave_factor);
}
opal_value_array_append_item(&device->ib_mr_array, &ib_mr);
mkey->va_base = ib_mr->addr;
}
#endif /* MPAGE_ENABLE */
} else {
MEMHEAP_ERROR("tr_id: %d key %llx attach failed: incorrect shmid 0x%X|0x%X",
tr_id,
(unsigned long long)mkey->key,
MEMHEAP_SHM_GET_TYPE(mkey->key),
MEMHEAP_SHM_GET_ID(mkey->key));
oshmem_shmem_abort(-1);
}
if ((void *) -1 == (void *) mkey->va_base) {
MEMHEAP_ERROR("tr_id: %d key %llx attach failed: errno = %d",
tr_id, (unsigned long long)mkey->key, errno);
oshmem_shmem_abort(-1);
}
}
}
/**
* TODO: using put request as handle is not good.
*/
static inline int mca_spml_ikrit_put_internal(void* dst_addr,
size_t size,
void* src_addr,
int dst,
void **handle,
int zcopy)
{
void *rva;
mca_spml_ikrit_put_request_t *put_req;
int ptl_id;
sshmem_mkey_t *r_mkey;
static int count;
int need_progress = 0;
if (0 >= size) {
return OSHMEM_SUCCESS;
}
ptl_id = get_ptl_id(dst);
/* Get rkey of remote PE (dst proc) which must be on memheap */
r_mkey = mca_memheap_base_get_cached_mkey(dst, dst_addr, ptl_id, &rva);
if (!r_mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable",
dst, dst_addr);
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
#if SPML_IKRIT_PUT_DEBUG == 1
SPML_VERBOSE(100, "put: pe:%d ptl=%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s",
dst, ptl_id, dst_addr, src_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey));
#endif
if (ptl_id == MXM_PTL_SHM) {
if (mca_memheap_base_can_local_copy(r_mkey, dst_addr)) {
memcpy((void *) (unsigned long) rva, src_addr, size);
/* call progress as often as we would have with regular put */
if (++count % SPML_IKRIT_PACKETS_PER_SYNC == 0)
mxm_progress(mca_spml_ikrit.mxm_context);
return OSHMEM_SUCCESS;
}
/* segment not mapped - fallback to rmda */
ptl_id = MXM_PTL_RDMA;
r_mkey = mca_memheap_base_get_cached_mkey(dst, dst_addr, ptl_id, &rva);
if (!r_mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable",
dst, dst_addr);
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
}
#if SPML_IKRIT_PUT_DEBUG == 1
SPML_VERBOSE(100, "put: pe:%d ptl=%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s",
dst, ptl_id, dst_addr, src_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey));
#endif
put_req = alloc_put_req();
if (NULL == put_req) {
SPML_ERROR("out of put requests - aborting");
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
if (handle)
*handle = put_req;
/* fill out request */
put_req->mxm_req.base.mq = mca_spml_ikrit.mxm_mq;
/* request immediate responce if we are getting low on send buffers. We only get responce from remote on ack timeout.
* Also request explicit ack once in a while */
#if MXM_API < MXM_VERSION(2,0)
put_req->mxm_req.opcode = MXM_REQ_OP_PUT;
if (mca_spml_ikrit.free_list_max - mca_spml_ikrit.n_active_puts <= SPML_IKRIT_PUT_LOW_WATER ||
(mca_spml_ikrit.mxm_peers[dst]->n_active_puts + 1) % SPML_IKRIT_PACKETS_PER_SYNC == 0) {
put_req->mxm_req.base.flags = MXM_REQ_FLAG_SEND_SYNC;
need_progress = 1;
} else {
put_req->mxm_req.base.flags = MXM_REQ_FLAG_SEND_LAZY|MXM_REQ_FLAG_SEND_SYNC;
}
#else
put_req->mxm_req.flags = 0;
if (mca_spml_ikrit.free_list_max - mca_spml_ikrit.n_active_puts <= SPML_IKRIT_PUT_LOW_WATER ||
(int)opal_list_get_size(&mca_spml_ikrit.active_peers) > mca_spml_ikrit.unsync_conn_max ||
(mca_spml_ikrit.mxm_peers[dst]->n_active_puts + 1) % SPML_IKRIT_PACKETS_PER_SYNC == 0) {
need_progress = 1;
put_req->mxm_req.opcode = MXM_REQ_OP_PUT_SYNC;
} else {
put_req->mxm_req.opcode = MXM_REQ_OP_PUT;
}
if (!zcopy) {
if (size < mca_spml_ikrit.put_zcopy_threshold) {
put_req->mxm_req.flags |= MXM_REQ_SEND_FLAG_BLOCKING;
} else {
put_req->mxm_req.opcode = MXM_REQ_OP_PUT_SYNC;
}
}
#endif
put_req->mxm_req.base.conn = mca_spml_ikrit.mxm_peers[dst]->mxm_conn;
put_req->mxm_req.base.data_type = MXM_REQ_DATA_BUFFER;
put_req->mxm_req.base.data.buffer.ptr = src_addr;
put_req->mxm_req.base.data.buffer.length = size;
put_req->mxm_req.base.completed_cb = put_completion_cb;
put_req->mxm_req.base.context = put_req;
put_req->mxm_req.op.mem.remote_vaddr = (intptr_t) rva;
put_req->mxm_req.base.state = MXM_REQ_NEW;
put_req->pe = dst;
#if MXM_API < MXM_VERSION(2,0)
put_req->mxm_req.base.data.buffer.memh = NULL;
put_req->mxm_req.op.mem.remote_memh = NULL;
#else
put_req->mxm_req.op.mem.remote_mkey = to_mxm_mkey(r_mkey);
#endif
OPAL_THREAD_ADD32(&mca_spml_ikrit.n_active_puts, 1);
if (mca_spml_ikrit.mxm_peers[dst]->need_fence == 0) {
opal_list_append(&mca_spml_ikrit.active_peers,
&mca_spml_ikrit.mxm_peers[dst]->super);
mca_spml_ikrit.mxm_peers[dst]->need_fence = 1;
}
mca_spml_ikrit.mxm_peers[dst]->n_active_puts++;
SPML_IKRIT_MXM_POST_SEND(put_req->mxm_req);
if (need_progress)
mxm_progress(mca_spml_ikrit.mxm_context);
return OSHMEM_SUCCESS;
}
/* simple buffered put implementation. NOT IN USE
* Problems:
* - slighly worse performance than impl based on non buffered put
* - fence complexity is O(n_active_connections) instead of O(n_connections_with_outstanding_puts).
* Later is bounded by the network RTT & mxm ack timer.
*/
int mca_spml_ikrit_put_simple(void* dst_addr,
size_t size,
void* src_addr,
int dst)
{
void *rva;
mxm_send_req_t mxm_req;
mxm_wait_t wait;
int ptl_id;
sshmem_mkey_t *r_mkey;
static int count;
ptl_id = get_ptl_id(dst);
/* Get rkey of remote PE (dst proc) which must be on memheap */
r_mkey = mca_memheap_base_get_cached_mkey(dst, dst_addr, ptl_id, &rva);
if (!r_mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable",
dst, dst_addr);
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
#if SPML_IKRIT_PUT_DEBUG == 1
SPML_VERBOSE(100, "put: pe:%d ptl=%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s",
dst, ptl_id, dst_addr, src_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey));
#endif
if (ptl_id == MXM_PTL_SHM) {
if (mca_memheap_base_can_local_copy(r_mkey, dst_addr)) {
memcpy((void *) (unsigned long) rva, src_addr, size);
/* call progress as often as we would have with regular put */
if (++count % SPML_IKRIT_PACKETS_PER_SYNC == 0)
mxm_progress(mca_spml_ikrit.mxm_context);
return OSHMEM_SUCCESS;
}
/* segment not mapped - fallback to rmda */
ptl_id = MXM_PTL_RDMA;
r_mkey = mca_memheap_base_get_cached_mkey(dst,
//(unsigned long) dst_addr,
dst_addr,
ptl_id,
&rva);
if (!r_mkey) {
SPML_ERROR("pe=%d: %p is not address of shared variable",
dst, dst_addr);
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
}
#if SPML_IKRIT_PUT_DEBUG == 1
SPML_VERBOSE(100, "put: pe:%d ptl=%d dst=%p <- src: %p sz=%d. dst_rva=%p, %s",
dst, ptl_id, dst_addr, src_addr, (int)size, (void *)rva, mca_spml_base_mkey2str(r_mkey));
#endif
/* fill out request */
mxm_req.base.mq = mca_spml_ikrit.mxm_mq;
#if MXM_API < MXM_VERSION(2,0)
mxm_req.base.flags = MXM_REQ_FLAG_BLOCKING;
#else
mxm_req.flags = MXM_REQ_SEND_FLAG_BLOCKING;
#endif
mxm_req.base.conn = mca_spml_ikrit.mxm_peers[dst]->mxm_conn;
mxm_req.base.data_type = MXM_REQ_DATA_BUFFER;
mxm_req.base.data.buffer.ptr = src_addr;
mxm_req.base.data.buffer.length = size;
mxm_req.base.completed_cb = 0;
mxm_req.base.context = 0;
mxm_req.opcode = MXM_REQ_OP_PUT;
mxm_req.op.mem.remote_vaddr = (intptr_t) rva;
mxm_req.base.state = MXM_REQ_NEW;
mxm_req.base.error = MXM_OK;
#if MXM_API < MXM_VERSION(2, 0)
mxm_req.base.data.buffer.memh = NULL;
mxm_req.op.mem.remote_memh = NULL;
#else
mxm_req.op.mem.remote_mkey = to_mxm_mkey(r_mkey);
#endif
if (mca_spml_ikrit.mxm_peers[dst]->need_fence == 0) {
opal_list_append(&mca_spml_ikrit.active_peers,
&mca_spml_ikrit.mxm_peers[dst]->super);
mca_spml_ikrit.mxm_peers[dst]->need_fence = 1;
}
SPML_IKRIT_MXM_POST_SEND(mxm_req);
wait.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;
}
