mca_spml_mkey_t *mca_spml_yoda_register(void* addr,
size_t size,
uint64_t shmid,
int *count)
{
int i;
mca_btl_base_descriptor_t* des = NULL;
const opal_datatype_t *datatype = &opal_datatype_wchar;
opal_convertor_t convertor;
mca_spml_mkey_t *mkeys;
struct yoda_btl *ybtl;
oshmem_proc_t *proc_self;
mca_spml_yoda_context_t* yoda_context;
struct iovec iov;
uint32_t iov_count = 1;
SPML_VERBOSE(10, "address %p len %llu", addr, (unsigned long long)size);
*count = 0;
/* make sure everything is initialized to 0 */
mkeys = (mca_spml_mkey_t *) calloc(1,
mca_spml_yoda.n_btls * sizeof(*mkeys));
if (!mkeys) {
return NULL ;
}
proc_self = oshmem_proc_group_find(oshmem_group_all, oshmem_my_proc_id());
/* create convertor */
OBJ_CONSTRUCT(&convertor, opal_convertor_t);
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;
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
&& MEMHEAP_SHM_INVALID != (int) MEMHEAP_SHM_GET_ID(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 (NULL != ybtl->btl->btl_prepare_src) {
/* initialize convertor for source descriptor*/
opal_convertor_copy_and_prepare_for_recv(proc_self->proc_convertor,
datatype,
size,
addr,
0,
&convertor);
if (NULL != ybtl->btl->btl_mpool && NULL != ybtl->btl->btl_mpool->mpool_register) {
iov.iov_len = size;
iov.iov_base = NULL;
opal_convertor_pack(&convertor, &iov, &iov_count, &size);
ybtl->btl->btl_mpool->mpool_register(ybtl->btl->btl_mpool,
iov.iov_base, size, 0, &yoda_context->registration);
}
/* initialize convertor for source descriptor*/
opal_convertor_copy_and_prepare_for_recv(proc_self->proc_convertor,
datatype,
size,
addr,
0,
&convertor);
/* register source memory */
des = ybtl->btl->btl_prepare_src(ybtl->btl,
0,
yoda_context->registration,
&convertor,
MCA_BTL_NO_ORDER,
0,
&size,
0);
if (NULL == des) {
SPML_ERROR("%s: failed to register source memory. ",
btl_type2str(ybtl->btl_type));
}
yoda_context->btl_src_descriptor = des;
mkeys[i].u.data = des->des_src;
mkeys[i].len = ybtl->btl->btl_seg_size;
}
SPML_VERBOSE(5,
"rank %d btl %s address 0x%p len %llu shmid 0x%X|0x%X",
oshmem_proc_local_proc->proc_name.vpid, btl_type2str(ybtl->btl_type),
mkeys[i].va_base, (unsigned long long)size, MEMHEAP_SHM_GET_TYPE(shmid), MEMHEAP_SHM_GET_ID(shmid));
}
OBJ_DESTRUCT(&convertor);
*count = mca_spml_yoda.n_btls;
return mkeys;
}
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);
}
}
}
