int
mca_pml_cm_enable(bool enable)
{
/* BWB - FIX ME - need to have this actually do something,
maybe? */
ompi_free_list_init_new(&mca_pml_base_send_requests,
sizeof(mca_pml_cm_hvy_send_request_t) + ompi_mtl->mtl_request_size,
opal_cache_line_size,
OBJ_CLASS(mca_pml_cm_hvy_send_request_t),
0,opal_cache_line_size,
ompi_pml_cm.free_list_num,
ompi_pml_cm.free_list_max,
ompi_pml_cm.free_list_inc,
NULL);
ompi_free_list_init_new(&mca_pml_base_recv_requests,
sizeof(mca_pml_cm_hvy_recv_request_t) + ompi_mtl->mtl_request_size,
opal_cache_line_size,
OBJ_CLASS(mca_pml_cm_hvy_recv_request_t),
0,opal_cache_line_size,
ompi_pml_cm.free_list_num,
ompi_pml_cm.free_list_max,
ompi_pml_cm.free_list_inc,
NULL);
return OMPI_SUCCESS;
}
int mca_spml_ikrit_enable(bool enable)
{
SPML_VERBOSE(50, "*** ikrit ENABLED ****");
if (false == enable) {
return OSHMEM_SUCCESS;
}
opal_free_list_init (&mca_spml_base_put_requests,
sizeof(mca_spml_ikrit_put_request_t),
opal_cache_line_size,
OBJ_CLASS(opal_free_list_item_t),
0,
opal_cache_line_size,
mca_spml_ikrit.free_list_num,
mca_spml_ikrit.free_list_max,
mca_spml_ikrit.free_list_inc,
NULL, 0, NULL, NULL, NULL);
opal_free_list_init (&mca_spml_base_get_requests,
sizeof(mca_spml_ikrit_get_request_t),
opal_cache_line_size,
OBJ_CLASS(opal_free_list_item_t),
0,
opal_cache_line_size,
mca_spml_ikrit.free_list_num,
mca_spml_ikrit.free_list_max,
mca_spml_ikrit.free_list_inc,
NULL, 0, NULL, NULL, NULL);
mca_spml_ikrit.enabled = true;
return OSHMEM_SUCCESS;
}
int mca_pml_dr_enable(bool enable)
{
if( false == enable ) return OMPI_SUCCESS;
/* requests */
ompi_free_list_init_new( &mca_pml_base_send_requests,
sizeof(mca_pml_dr_send_request_t),
opal_cache_line_size,
OBJ_CLASS(mca_pml_dr_send_request_t),
0,opal_cache_line_size,
mca_pml_dr.free_list_num,
mca_pml_dr.free_list_max,
mca_pml_dr.free_list_inc,
NULL );
ompi_free_list_init_new( &mca_pml_base_recv_requests,
sizeof(mca_pml_dr_recv_request_t),
opal_cache_line_size,
OBJ_CLASS(mca_pml_dr_recv_request_t),
0,opal_cache_line_size,
mca_pml_dr.free_list_num,
mca_pml_dr.free_list_max,
mca_pml_dr.free_list_inc,
NULL );
/* fragments */
OBJ_CONSTRUCT(&mca_pml_dr.recv_frags, ompi_free_list_t);
ompi_free_list_init_new( &mca_pml_dr.recv_frags,
sizeof(mca_pml_dr_recv_frag_t),
opal_cache_line_size,
OBJ_CLASS(mca_pml_dr_recv_frag_t),
0,opal_cache_line_size,
mca_pml_dr.free_list_num,
mca_pml_dr.free_list_max,
mca_pml_dr.free_list_inc,
NULL );
OBJ_CONSTRUCT(&mca_pml_dr.vfrags, ompi_free_list_t);
ompi_free_list_init_new( &mca_pml_dr.vfrags,
sizeof(mca_pml_dr_vfrag_t),
opal_cache_line_size,
OBJ_CLASS(mca_pml_dr_vfrag_t),
0,opal_cache_line_size,
mca_pml_dr.free_list_num,
mca_pml_dr.free_list_max,
mca_pml_dr.free_list_inc,
NULL );
OBJ_CONSTRUCT(&mca_pml_dr.send_pending, opal_list_t);
OBJ_CONSTRUCT(&mca_pml_dr.send_active, opal_list_t);
OBJ_CONSTRUCT(&mca_pml_dr.acks_pending, opal_list_t);
OBJ_CONSTRUCT(&mca_pml_dr.buffers, ompi_free_list_t);
OBJ_CONSTRUCT(&mca_pml_dr.endpoints, opal_pointer_array_t);
OBJ_CONSTRUCT(&mca_pml_dr.lock, opal_mutex_t);
mca_pml_dr.enabled = true;
return OMPI_SUCCESS;
}
int mca_spml_yoda_enable(bool enable)
{
SPML_VERBOSE(50, "*** yoda ENABLED ****");
if (false == enable) {
return OSHMEM_SUCCESS;
}
OBJ_CONSTRUCT(&mca_spml_yoda.lock, opal_mutex_t);
/**
*If we get here this is the SPML who get selected for the run. We
* should get ownership for the put and get requests list, and
* initialize them with the size of our own requests.
*/
opal_free_list_init (&mca_spml_base_put_requests,
sizeof(mca_spml_yoda_put_request_t),
opal_cache_line_size,
OBJ_CLASS(mca_spml_yoda_put_request_t),
0,
opal_cache_line_size,
mca_spml_yoda.free_list_num,
mca_spml_yoda.free_list_max,
mca_spml_yoda.free_list_inc,
NULL, 0, NULL, NULL, NULL);
opal_free_list_init (&mca_spml_base_get_requests,
sizeof(mca_spml_yoda_get_request_t),
opal_cache_line_size,
OBJ_CLASS(mca_spml_yoda_get_request_t),
0,
opal_cache_line_size,
mca_spml_yoda.free_list_num,
mca_spml_yoda.free_list_max,
mca_spml_yoda.free_list_inc,
NULL, 0, NULL, NULL, NULL);
mca_spml_yoda.enabled = true;
/* The following line resolves the issue with BTL tcp and SPML yoda. In this case the
* atomic_basic_lock(root_rank) function may behave as DoS attack on root_rank, since
* all the procceses will do shmem_int_get from root_rank. These calls would go through
* bml active messaging and will trigger replays in libevent on root rank. If the flag
* OPAL_ENVLOOP_ONCE is not set then libevent will continously progress constantly
* incoming events thus causing root_rank to stuck in libevent loop.
*/
opal_progress_set_event_flag(OPAL_EVLOOP_NONBLOCK | OPAL_EVLOOP_ONCE);
#if OSHMEM_WAIT_COMPLETION_DEBUG == 1
condition_dbg_init();
#endif
return OSHMEM_SUCCESS;
}
static int
mca_btl_scif_setup_mpools (mca_btl_scif_module_t *scif_module)
{
struct mca_mpool_base_resources_t mpool_resources;
int rc;
/* initialize the grdma mpool */
mpool_resources.pool_name = "scif";
mpool_resources.reg_data = (void *) scif_module;
mpool_resources.sizeof_reg = sizeof (mca_btl_scif_reg_t);
mpool_resources.register_mem = scif_reg_mem;
mpool_resources.deregister_mem = scif_dereg_mem;
scif_module->super.btl_mpool =
mca_mpool_base_module_create("grdma", scif_module, &mpool_resources);
if (NULL == scif_module->super.btl_mpool) {
BTL_ERROR(("error creating grdma mpool"));
return OMPI_ERROR;
}
/* setup free lists for fragments. dma fragments will be used for
* rma operations and in-place sends. eager frags will be used for
* buffered sends. */
rc = ompi_free_list_init_new (&scif_module->dma_frags,
sizeof (mca_btl_scif_dma_frag_t), 64,
OBJ_CLASS(mca_btl_scif_dma_frag_t),
128, getpagesize (),
mca_btl_scif_component.scif_free_list_num,
mca_btl_scif_component.scif_free_list_max,
mca_btl_scif_component.scif_free_list_inc,
NULL);
if (OPAL_UNLIKELY(OMPI_SUCCESS != rc)) {
return rc;
}
rc = ompi_free_list_init_new (&scif_module->eager_frags,
sizeof (mca_btl_scif_eager_frag_t), 8,
OBJ_CLASS(mca_btl_scif_eager_frag_t),
128 + scif_module->super.btl_eager_limit, 64,
mca_btl_scif_component.scif_free_list_num,
mca_btl_scif_component.scif_free_list_max,
mca_btl_scif_component.scif_free_list_inc,
NULL);
if (OPAL_UNLIKELY(OMPI_SUCCESS != rc)) {
BTL_ERROR(("error creating eager receive fragment free list"));
return rc;
}
return OMPI_SUCCESS;
}
/*
* Initializes the mpool module.
*/
void mca_mpool_grdma_module_init(mca_mpool_grdma_module_t* mpool, mca_mpool_grdma_pool_t *pool)
{
OBJ_RETAIN(pool);
mpool->pool = pool;
mpool->super.mpool_component = &mca_mpool_grdma_component.super;
mpool->super.mpool_base = NULL; /* no base .. */
mpool->super.mpool_alloc = mca_mpool_grdma_alloc;
mpool->super.mpool_realloc = mca_mpool_grdma_realloc;
mpool->super.mpool_free = mca_mpool_grdma_free;
mpool->super.mpool_register = mca_mpool_grdma_register;
mpool->super.mpool_find = mca_mpool_grdma_find;
mpool->super.mpool_deregister = mca_mpool_grdma_deregister;
mpool->super.mpool_release_memory = mca_mpool_grdma_release_memory;
mpool->super.mpool_finalize = mca_mpool_grdma_finalize;
mpool->super.mpool_ft_event = mca_mpool_grdma_ft_event;
mpool->super.flags = MCA_MPOOL_FLAGS_MPI_ALLOC_MEM;
mpool->super.rcache = pool->rcache;
mpool->stat_cache_hit = mpool->stat_cache_miss = mpool->stat_evicted = 0;
mpool->stat_cache_found = mpool->stat_cache_notfound = 0;
OBJ_CONSTRUCT(&mpool->reg_list, ompi_free_list_t);
ompi_free_list_init_new(&mpool->reg_list, mpool->resources.sizeof_reg,
opal_cache_line_size,
OBJ_CLASS(mca_mpool_base_registration_t),
0, opal_cache_line_size, 0, -1, 32, NULL);
}
/*
* Initializes the mpool module.
*/
void mca_mpool_rgpusm_module_init(mca_mpool_rgpusm_module_t* mpool)
{
mpool->super.mpool_component = &mca_mpool_rgpusm_component.super;
mpool->super.mpool_base = NULL; /* no base .. */
mpool->super.mpool_alloc = NULL;
mpool->super.mpool_realloc = NULL;
mpool->super.mpool_free = mca_mpool_rgpusm_free;
mpool->super.mpool_register = mca_mpool_rgpusm_register;
mpool->super.mpool_find = mca_mpool_rgpusm_find;
mpool->super.mpool_deregister = mca_mpool_rgpusm_deregister;
mpool->super.mpool_release_memory = NULL;
mpool->super.mpool_finalize = mca_mpool_rgpusm_finalize;
mpool->super.mpool_ft_event = mca_mpool_rgpusm_ft_event;
mpool->super.rcache =
mca_rcache_base_module_create(mca_mpool_rgpusm_component.rcache_name);
mpool->super.flags = 0;
mpool->resources.reg_data = NULL;
mpool->resources.sizeof_reg = sizeof(struct mca_mpool_common_cuda_reg_t);
mpool->resources.register_mem = cuda_openmemhandle;
mpool->resources.deregister_mem = cuda_closememhandle;
OBJ_CONSTRUCT(&mpool->reg_list, ompi_free_list_t);
ompi_free_list_init_new(&mpool->reg_list, mpool->resources.sizeof_reg,
opal_cache_line_size,
OBJ_CLASS(mca_mpool_base_registration_t),
0,opal_cache_line_size,
0, -1, 32, NULL);
OBJ_CONSTRUCT(&mpool->lru_list, opal_list_t);
mpool->stat_cache_hit = mpool->stat_cache_miss = mpool->stat_evicted = 0;
mpool->stat_cache_found = mpool->stat_cache_notfound = 0;
mpool->stat_cache_valid = mpool->stat_cache_invalid = 0;
}
int32_t opal_datatype_init( void )
{
const opal_datatype_t* datatype;
int32_t i;
opal_arch_compute_local_id( &opal_local_arch );
/**
* Force he initialization of the opal_datatype_t class. This will allow us to
* call OBJ_DESTRUCT without going too deep in the initialization process.
*/
opal_class_initialize(OBJ_CLASS(opal_datatype_t));
for( i = OPAL_DATATYPE_FIRST_TYPE; i < OPAL_DATATYPE_MAX_PREDEFINED; i++ ) {
datatype = opal_datatype_basicDatatypes[i];
/* All of the predefined OPAL types don't have any GAPS! */
datatype->desc.desc[0].elem.common.flags = OPAL_DATATYPE_FLAG_PREDEFINED |
OPAL_DATATYPE_FLAG_DATA |
OPAL_DATATYPE_FLAG_CONTIGUOUS |
OPAL_DATATYPE_FLAG_NO_GAPS;
datatype->desc.desc[0].elem.common.type = i;
/* datatype->desc.desc[0].elem.blocklen XXX not set at the moment, it will be needed later */
datatype->desc.desc[0].elem.count = 1;
datatype->desc.desc[0].elem.disp = 0;
datatype->desc.desc[0].elem.extent = datatype->size;
datatype->desc.desc[1].end_loop.common.flags = 0;
datatype->desc.desc[1].end_loop.common.type = OPAL_DATATYPE_END_LOOP;
datatype->desc.desc[1].end_loop.items = 1;
datatype->desc.desc[1].end_loop.first_elem_disp = datatype->desc.desc[0].elem.disp;
datatype->desc.desc[1].end_loop.size = datatype->size;
}
return OPAL_SUCCESS;
}
static int mca_bcol_iboffload_alloc_reg_qp_resource(int qp_index, mca_bcol_iboffload_device_t *device)
{
int length;
mca_bcol_iboffload_component_t *cm = &mca_bcol_iboffload_component;
ompi_free_list_t *frags_free = &device->frags_free[qp_index];
OBJ_CONSTRUCT(frags_free, ompi_free_list_t);
length = cm->qp_infos[qp_index].size;
IBOFFLOAD_VERBOSE(10, ("free list len %d\n", length));
if (OMPI_SUCCESS != ompi_free_list_init_ex_new(frags_free,
sizeof(mca_bcol_iboffload_frag_t), MCA_IBOFFLOAD_CACHE_LINE_SIZE,
OBJ_CLASS(mca_bcol_iboffload_frag_t),
length, cm->buffer_alignment,
cm->free_list_num,
cm->free_list_max,
cm->free_list_inc,
device->mpool,
mca_bcol_iboffload_frag_init,
(void *) &cm->qp_infos[qp_index].qp_index)) {
IBOFFLOAD_ERROR(("Failed to allocate frags_free"));
return OMPI_ERROR;
}
return OMPI_SUCCESS;
}
static int
portals4_open(void)
{
int ret;
mca_coll_portals4_component.ni_h = PTL_INVALID_HANDLE;
mca_coll_portals4_component.uid = PTL_UID_ANY;
mca_coll_portals4_component.pt_idx = -1;
mca_coll_portals4_component.finish_pt_idx = -1;
mca_coll_portals4_component.eq_h = PTL_INVALID_HANDLE;
mca_coll_portals4_component.unex_me_h = PTL_INVALID_HANDLE;
mca_coll_portals4_component.finish_me_h = PTL_INVALID_HANDLE;
mca_coll_portals4_component.zero_md_h = PTL_INVALID_HANDLE;
mca_coll_portals4_component.data_md_h = PTL_INVALID_HANDLE;
OBJ_CONSTRUCT(&mca_coll_portals4_component.requests, opal_free_list_t);
ret = opal_free_list_init(&mca_coll_portals4_component.requests,
sizeof(ompi_coll_portals4_request_t),
opal_cache_line_size,
OBJ_CLASS(ompi_coll_portals4_request_t),
0, 0, 8, 0, 8, NULL, 0, NULL, NULL, NULL);
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: ompi_free_list_init failed: %d\n",
__FILE__, __LINE__, ret);
return ret;
}
return OMPI_SUCCESS;
}
static void initFreeList(void)
{
mca_base_var_register("oshmem", "shmem", "java", "eager",
"Java buffers eager size",
MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,
OPAL_INFO_LVL_5,
MCA_BASE_VAR_SCOPE_READONLY,
&oshmem_shmem_java_eager);
OBJ_CONSTRUCT(&shmem_java_buffers, opal_free_list_t);
int r = opal_free_list_init(&shmem_java_buffers,
sizeof(shmem_java_buffer_t),
opal_cache_line_size,
OBJ_CLASS(shmem_java_buffer_t),
0, /* payload size */
0, /* payload align */
2, /* initial elements to alloc */
-1, /* max elements */
2, /* num elements per alloc */
NULL, /* mpool */
0, /* mpool reg flags */
NULL, /* unused0 */
NULL, /* item_init */
NULL /* inem_init context */);
if(r != OPAL_SUCCESS)
{
fprintf(stderr, "Unable to initialize shmem_java_buffers.\n");
exit(1);
}
}
/*
* Initializes the mpool module.
*/
void mca_mpool_gpusm_module_init(mca_mpool_gpusm_module_t* mpool)
{
mpool->super.mpool_component = &mca_mpool_gpusm_component.super;
mpool->super.mpool_base = NULL;
mpool->super.mpool_alloc = NULL;
mpool->super.mpool_realloc = NULL;
mpool->super.mpool_free = NULL;
mpool->super.mpool_register = mca_mpool_gpusm_register;
mpool->super.mpool_find = mca_mpool_gpusm_find;
mpool->super.mpool_deregister = mca_mpool_gpusm_deregister;
mpool->super.mpool_release_memory = NULL;
mpool->super.mpool_finalize = mca_mpool_gpusm_finalize;
mpool->super.mpool_ft_event = mca_mpool_gpusm_ft_event;
mpool->super.rcache = NULL;
mpool->super.flags = 0;
mpool->resources.reg_data = NULL;
mpool->resources.sizeof_reg = sizeof(struct mca_mpool_common_cuda_reg_t);
mpool->resources.register_mem = cuda_getmemhandle;
mpool->resources.deregister_mem = cuda_ungetmemhandle;
OBJ_CONSTRUCT(&mpool->reg_list, ompi_free_list_t);
/* Start with 0 entries in the free list since CUDA may not have
* been initialized when this free list is created and there is
* some CUDA specific activities that need to be done. */
ompi_free_list_init_new(&mpool->reg_list, mpool->resources.sizeof_reg,
opal_cache_line_size,
OBJ_CLASS(mca_mpool_gpusm_registration_t),
0,opal_cache_line_size,
0, -1, 64, NULL);
}
/*
* Initializes the mpool module.
*/
void mca_mpool_rdma_module_init(mca_mpool_rdma_module_t* mpool)
{
mpool->super.mpool_component = &mca_mpool_rdma_component.super;
mpool->super.mpool_base = NULL; /* no base .. */
mpool->super.mpool_alloc = mca_mpool_rdma_alloc;
mpool->super.mpool_realloc = mca_mpool_rdma_realloc;
mpool->super.mpool_free = mca_mpool_rdma_free;
mpool->super.mpool_register = mca_mpool_rdma_register;
mpool->super.mpool_find = mca_mpool_rdma_find;
mpool->super.mpool_deregister = mca_mpool_rdma_deregister;
mpool->super.mpool_release_memory = mca_mpool_rdma_release_memory;
mpool->super.mpool_finalize = mca_mpool_rdma_finalize;
mpool->super.mpool_ft_event = mca_mpool_rdma_ft_event;
mpool->super.rcache =
mca_rcache_base_module_create(mca_mpool_rdma_component.rcache_name);
mpool->super.flags = MCA_MPOOL_FLAGS_MPI_ALLOC_MEM;
OBJ_CONSTRUCT(&mpool->reg_list, ompi_free_list_t);
ompi_free_list_init_new(&mpool->reg_list, mpool->resources.sizeof_reg,
opal_cache_line_size,
OBJ_CLASS(mca_mpool_base_registration_t),
0,opal_cache_line_size,
0, -1, 32, NULL);
OBJ_CONSTRUCT(&mpool->mru_list, opal_list_t);
OBJ_CONSTRUCT(&mpool->gc_list, opal_list_t);
mpool->stat_cache_hit = mpool->stat_cache_miss = mpool->stat_evicted = 0;
mpool->stat_cache_found = mpool->stat_cache_notfound = 0;
/* Set this here (vs in component.c) because
ompi_mpi_leave_pinned* may have been set after MCA params were
read (e.g., by the openib btl) */
mca_mpool_rdma_component.leave_pinned = (int)
(1 == ompi_mpi_leave_pinned || ompi_mpi_leave_pinned_pipeline);
}
/*
* Initializes the mpool module.
*/
void mca_mpool_rdma_module_init(mca_mpool_rdma_module_t* mpool)
{
mpool->super.mpool_component = &mca_mpool_rdma_component.super;
mpool->super.mpool_base = NULL; /* no base .. */
mpool->super.mpool_alloc = mca_mpool_rdma_alloc;
mpool->super.mpool_realloc = mca_mpool_rdma_realloc;
mpool->super.mpool_free = mca_mpool_rdma_free;
mpool->super.mpool_register = mca_mpool_rdma_register;
mpool->super.mpool_find = mca_mpool_rdma_find;
mpool->super.mpool_deregister = mca_mpool_rdma_deregister;
mpool->super.mpool_release_memory = mca_mpool_rdma_release_memory;
if(mca_mpool_rdma_component.print_stats == true)
mpool->super.mpool_finalize = mca_mpool_rdma_finalize;
else
mpool->super.mpool_finalize = NULL;
mpool->super.rcache =
mca_rcache_base_module_create(mca_mpool_rdma_component.rcache_name);
mpool->super.flags = MCA_MPOOL_FLAGS_MPI_ALLOC_MEM;
OBJ_CONSTRUCT(&mpool->reg_list, ompi_free_list_t);
ompi_free_list_init(&mpool->reg_list, mpool->resources.sizeof_reg,
OBJ_CLASS(mca_mpool_base_registration_t), 0, -1, 32,
NULL);
OBJ_CONSTRUCT(&mpool->mru_list, opal_list_t);
mpool->stat_cache_hit = mpool->stat_cache_miss = mpool->stat_evicted = 0;
mpool->stat_cache_found = mpool->stat_cache_notfound = 0;
}
static int vader_btl_first_time_init(mca_btl_vader_t *vader_btl, int n)
{
mca_btl_vader_component_t *component = &mca_btl_vader_component;
int rc;
/* generate the endpoints */
component->endpoints = (struct mca_btl_base_endpoint_t *) calloc (n, sizeof (struct mca_btl_base_endpoint_t));
component->segment_offset = (n - 1) * MCA_BTL_VADER_FBOX_PEER_SIZE + 4096;
/* initialize fragment descriptor free lists */
/* initialize free list for put/get/single copy/inline fragments */
rc = ompi_free_list_init_ex_new(&component->vader_frags_user,
sizeof(mca_btl_vader_frag_t),
opal_cache_line_size, OBJ_CLASS(mca_btl_vader_frag_t),
0, opal_cache_line_size,
component->vader_free_list_num,
component->vader_free_list_max,
component->vader_free_list_inc,
NULL, mca_btl_vader_frag_init,
(void *) (sizeof(mca_btl_vader_hdr_t) +
mca_btl_vader_component.max_inline_send));
if (OMPI_SUCCESS != rc) {
return rc;
}
/* initialize free list for buffered send fragments */
rc = ompi_free_list_init_ex_new(&component->vader_frags_eager,
sizeof (mca_btl_vader_frag_t),
opal_cache_line_size, OBJ_CLASS(mca_btl_vader_frag_t),
0, opal_cache_line_size,
component->vader_free_list_num,
component->vader_free_list_max,
component->vader_free_list_inc,
NULL, mca_btl_vader_frag_init,
(void *) (sizeof (mca_btl_vader_hdr_t) +
mca_btl_vader.super.btl_eager_limit));
if (OMPI_SUCCESS != rc) {
return rc;
}
/* set flag indicating btl has been inited */
vader_btl->btl_inited = true;
return OMPI_SUCCESS;
}
int mca_spml_yoda_enable(bool enable)
{
SPML_VERBOSE(50, "*** yoda ENABLED ****");
if (false == enable) {
return OSHMEM_SUCCESS;
}
OBJ_CONSTRUCT(&mca_spml_yoda.lock, opal_mutex_t);
/**
*If we get here this is the SPML who get selected for the run. We
* should get ownership for the put and get requests list, and
* initialize them with the size of our own requests.
*/
ompi_free_list_init_new(&mca_spml_base_put_requests,
sizeof(mca_spml_yoda_put_request_t),
opal_cache_line_size,
OBJ_CLASS(mca_spml_yoda_put_request_t),
0,
opal_cache_line_size,
mca_spml_yoda.free_list_num,
mca_spml_yoda.free_list_max,
mca_spml_yoda.free_list_inc,
NULL );
ompi_free_list_init_new(&mca_spml_base_get_requests,
sizeof(mca_spml_yoda_get_request_t),
opal_cache_line_size,
OBJ_CLASS(mca_spml_yoda_get_request_t),
0,
opal_cache_line_size,
mca_spml_yoda.free_list_num,
mca_spml_yoda.free_list_max,
mca_spml_yoda.free_list_inc,
NULL );
mca_spml_yoda.enabled = true;
#if OSHMEM_WAIT_COMPLETION_DEBUG == 1
condition_dbg_init();
#endif
return OSHMEM_SUCCESS;
}
/**
* the constructor function. creates the free list to get the nodes from
*
* @param object the tree that is to be used
*
* @retval NONE
*/
static void opal_rb_tree_construct(opal_object_t * object)
{
opal_rb_tree_t * tree = (opal_rb_tree_t *) object;
tree->root_ptr = NULL;
OBJ_CONSTRUCT(&(tree->free_list), opal_free_list_t);
opal_free_list_init (&(tree->free_list), sizeof(opal_rb_tree_node_t),
opal_cache_line_size, OBJ_CLASS(opal_rb_tree_node_t),
0,opal_cache_line_size, 0, -1 , 128, NULL, 0, NULL, NULL, NULL);
}
static void initFreeList(void)
{
OBJ_CONSTRUCT(&ompi_java_buffers, opal_free_list_t);
int r = opal_free_list_init(&ompi_java_buffers, sizeof(ompi_java_buffer_t),
OBJ_CLASS(ompi_java_buffer_t), 2, -1, 2);
if(r != OPAL_SUCCESS)
{
fprintf(stderr, "Unable to initialize ompi_java_buffers.\n");
exit(1);
}
}
/*
* Setup the freelist of IO requests. This does not need to be
* protected with a lock because it's called during MPI_INIT.
*/
int mca_io_base_request_create_freelist(void)
{
opal_list_item_t *p;
const mca_base_component_t *component;
const mca_io_base_component_2_0_0_t *v200;
size_t size = 0;
int i, init, incr;
/* Find the maximum additional number of bytes required by all io
components for requests and make that the request size */
for (p = opal_list_get_first(&mca_io_base_components_available);
p != opal_list_get_end(&mca_io_base_components_available);
p = opal_list_get_next(p)) {
component = ((mca_base_component_priority_list_item_t *)
p)->super.cli_component;
/* Only know how to handle v2.0.0 components for now */
if (component->mca_type_major_version == 2 &&
component->mca_type_minor_version == 0 &&
component->mca_type_release_version == 0) {
v200 = (mca_io_base_component_2_0_0_t *) component;
if (v200->io_request_bytes > size) {
size = v200->io_request_bytes;
}
}
}
/* Construct and initialized the freelist of IO requests. */
OBJ_CONSTRUCT(&mca_io_base_requests, ompi_free_list_t);
mca_io_base_requests_valid = true;
i = mca_base_param_find("io", NULL, "base_freelist_initial_size");
mca_base_param_lookup_int(i, &init);
i = mca_base_param_find("io", NULL, "base_freelist_increment");
mca_base_param_lookup_int(i, &incr);
ompi_free_list_init_new(&mca_io_base_requests,
sizeof(mca_io_base_request_t) + size,
CACHE_LINE_SIZE,
OBJ_CLASS(mca_io_base_request_t),
0,CACHE_LINE_SIZE,
init, -1, incr,
NULL);
/* All done */
return OMPI_SUCCESS;
}
int
ompi_osc_pt2pt_component_init(bool enable_progress_threads,
bool enable_mpi_threads)
{
size_t aligned_size;
OBJ_CONSTRUCT(&mca_osc_pt2pt_component.p2p_c_sendreqs, opal_free_list_t);
opal_free_list_init(&mca_osc_pt2pt_component.p2p_c_sendreqs,
sizeof(ompi_osc_pt2pt_sendreq_t),
OBJ_CLASS(ompi_osc_pt2pt_sendreq_t),
1, -1, 1);
OBJ_CONSTRUCT(&mca_osc_pt2pt_component.p2p_c_replyreqs, opal_free_list_t);
opal_free_list_init(&mca_osc_pt2pt_component.p2p_c_replyreqs,
sizeof(ompi_osc_pt2pt_replyreq_t),
OBJ_CLASS(ompi_osc_pt2pt_replyreq_t),
1, -1, 1);
OBJ_CONSTRUCT(&mca_osc_pt2pt_component.p2p_c_longreqs, opal_free_list_t);
opal_free_list_init(&mca_osc_pt2pt_component.p2p_c_longreqs,
sizeof(ompi_osc_pt2pt_longreq_t),
OBJ_CLASS(ompi_osc_pt2pt_longreq_t),
1, -1, 1);
/* adjust size to be multiple of ompi_ptr_t to avoid alignment issues*/
aligned_size = sizeof(ompi_osc_pt2pt_buffer_t) +
(sizeof(ompi_osc_pt2pt_buffer_t) % sizeof(ompi_ptr_t)) +
mca_osc_pt2pt_component.p2p_c_eager_size;
OBJ_CONSTRUCT(&mca_osc_pt2pt_component.p2p_c_buffers, opal_free_list_t);
opal_free_list_init(&mca_osc_pt2pt_component.p2p_c_buffers,
aligned_size,
OBJ_CLASS(ompi_osc_pt2pt_buffer_t),
1, -1, 1);
return OMPI_SUCCESS;
}
static void ompi_free_list_construct(ompi_free_list_t* fl)
{
OBJ_CONSTRUCT(&fl->fl_lock, opal_mutex_t);
OBJ_CONSTRUCT(&fl->fl_condition, opal_condition_t);
fl->fl_max_to_alloc = 0;
fl->fl_num_allocated = 0;
fl->fl_num_per_alloc = 0;
fl->fl_num_waiting = 0;
fl->fl_elem_size = sizeof(ompi_free_list_item_t);
fl->fl_elem_class = OBJ_CLASS(ompi_free_list_item_t);
fl->fl_header_space = 0;
fl->fl_alignment = 0;
fl->fl_mpool = 0;
OBJ_CONSTRUCT(&(fl->fl_allocations), opal_list_t);
}
/*
* initialize the rb tree
*/
int mca_mpool_base_tree_init(void) {
int rc;
OBJ_CONSTRUCT(&mca_mpool_base_tree, opal_rb_tree_t);
OBJ_CONSTRUCT(&mca_mpool_base_tree_item_free_list, opal_free_list_t);
OBJ_CONSTRUCT(&tree_lock, opal_mutex_t);
rc = opal_free_list_init (&mca_mpool_base_tree_item_free_list,
sizeof(mca_mpool_base_tree_item_t),
opal_cache_line_size,
OBJ_CLASS(mca_mpool_base_tree_item_t),
0,opal_cache_line_size,
0, -1 , 4, NULL, 0, NULL, NULL, NULL);
if(OPAL_SUCCESS == rc) {
rc = opal_rb_tree_init(&mca_mpool_base_tree, mca_mpool_base_tree_node_compare);
}
return rc;
}
/*
* initialize the rb tree
*/
int mca_mpool_base_tree_init(void) {
int rc;
OBJ_CONSTRUCT(&mca_mpool_base_tree, ompi_rb_tree_t);
OBJ_CONSTRUCT(&mca_mpool_base_tree_item_free_list, ompi_free_list_t);
OBJ_CONSTRUCT(&tree_lock, opal_mutex_t);
rc = ompi_free_list_init_new(&mca_mpool_base_tree_item_free_list,
sizeof(mca_mpool_base_tree_item_t),
CACHE_LINE_SIZE,
OBJ_CLASS(mca_mpool_base_tree_item_t),
0,CACHE_LINE_SIZE,
0, -1 , 4, NULL);
if(OMPI_SUCCESS == rc) {
rc = ompi_rb_tree_init(&mca_mpool_base_tree, mca_mpool_base_tree_node_compare);
}
return rc;
}
mca_pml_base_module_t* mca_pml_crcpw_component_init(int* priority,
bool enable_progress_threads,
bool enable_mpi_threads)
{
/* We use the PML_SELECT_WRAPPER_PRIORITY to indicate when this
* component should wrap around what is already selected
* If it is not set to this seminal value, then we are doing a
* normal selection operation
*/
if(*priority == PML_SELECT_WRAPPER_PRIORITY ) {
opal_output_verbose( 20, mca_pml_crcpw_component.output_handle,
"pml:crcpw: component_init: Wrap the selected component %s",
mca_pml_base_selected_component.pmlm_version.mca_component_name);
mca_pml_crcpw_module.wrapped_pml_component = mca_pml_base_selected_component;
mca_pml_crcpw_module.wrapped_pml_module = mca_pml;
mca_pml_crcpw_component.pml_crcp_wrapped = true;
opal_output_verbose( 20, mca_pml_crcpw_component.output_handle,
"pml:crcpw: component_init: Initalize Wrapper");
OBJ_CONSTRUCT(&pml_state_list, ompi_free_list_t);
ompi_free_list_init_new( &pml_state_list,
sizeof(ompi_crcp_base_pml_state_t),
opal_cache_line_size,
OBJ_CLASS(ompi_crcp_base_pml_state_t),
0,opal_cache_line_size,
5, /* Initial number */
-1, /* Max = Unlimited */
64, /* Increment by */
NULL);
}
else {
opal_output_verbose( 20, mca_pml_crcpw_component.output_handle,
"pml:crcpw: component_init: Priority %d",
mca_pml_crcpw_component.priority);
}
*priority = mca_pml_crcpw_component.priority;
pml_crcpw_is_finalized = false;
return &mca_pml_crcpw_module.super;
}
static void opal_free_list_construct(opal_free_list_t* fl)
{
OBJ_CONSTRUCT(&fl->fl_lock, opal_mutex_t);
OBJ_CONSTRUCT(&fl->fl_condition, opal_condition_t);
fl->fl_max_to_alloc = 0;
fl->fl_num_allocated = 0;
fl->fl_num_per_alloc = 0;
fl->fl_num_waiting = 0;
fl->fl_frag_size = sizeof(opal_free_list_item_t);
fl->fl_frag_alignment = 0;
fl->fl_payload_buffer_size = 0;
fl->fl_payload_buffer_alignment = 0;
fl->fl_frag_class = OBJ_CLASS(opal_free_list_item_t);
fl->fl_mpool = NULL;
/* default flags */
fl->fl_mpool_reg_flags = MCA_MPOOL_FLAGS_CACHE_BYPASS |
MCA_MPOOL_FLAGS_CUDA_REGISTER_MEM;
fl->ctx = NULL;
OBJ_CONSTRUCT(&(fl->fl_allocations), opal_list_t);
}
/*
* Initializes the rcache module.
*/
void mca_rcache_rgpusm_module_init(mca_rcache_rgpusm_module_t* rcache)
{
rcache->super.rcache_component = &mca_rcache_rgpusm_component.super;
rcache->super.rcache_register = mca_rcache_rgpusm_register;
rcache->super.rcache_find = mca_rcache_rgpusm_find;
rcache->super.rcache_deregister = mca_rcache_rgpusm_deregister;
rcache->super.rcache_finalize = mca_rcache_rgpusm_finalize;
rcache->vma_module = mca_rcache_base_vma_module_alloc ();
OBJ_CONSTRUCT(&rcache->reg_list, opal_free_list_t);
opal_free_list_init (&rcache->reg_list, sizeof(struct mca_rcache_common_cuda_reg_t),
opal_cache_line_size,
OBJ_CLASS(mca_rcache_base_registration_t),
0,opal_cache_line_size,
0, -1, 32, NULL, 0, NULL, NULL, NULL);
OBJ_CONSTRUCT(&rcache->lru_list, opal_list_t);
rcache->stat_cache_hit = rcache->stat_cache_miss = rcache->stat_evicted = 0;
rcache->stat_cache_found = rcache->stat_cache_notfound = 0;
rcache->stat_cache_valid = rcache->stat_cache_invalid = 0;
}
void hcoll_rte_fns_setup(void)
{
init_module_fns();
OBJ_CONSTRUCT(&mca_coll_hcoll_component.requests, ompi_free_list_t);
ompi_free_list_init_ex_new(
&(mca_coll_hcoll_component.requests),
sizeof(ompi_request_t),
/* no special alignment needed */
8,
OBJ_CLASS(ompi_request_t),
/* no payload data */
0, 0,
/* NOTE: hack - need to parametrize this */
10,
50,
10,
/* No Mpool */
NULL,
NULL,
NULL
);
}
int ofi_comp_list_init(opal_free_list_t *comp_list)
{
int rc;
OBJ_CONSTRUCT(comp_list, opal_free_list_t);
rc = opal_free_list_init(comp_list,
sizeof(mca_btl_ofi_completion_t),
opal_cache_line_size,
OBJ_CLASS(mca_btl_ofi_completion_t),
0,
0,
128,
-1,
128,
NULL,
0,
NULL,
NULL,
NULL);
if (rc != OPAL_SUCCESS) {
BTL_VERBOSE(("cannot allocate completion freelist"));
}
return rc;
}
/*
* Initializes the rcache module.
*/
void mca_rcache_grdma_module_init(mca_rcache_grdma_module_t* rcache, mca_rcache_grdma_cache_t *cache)
{
OBJ_RETAIN(cache);
rcache->cache = cache;
rcache->super.rcache_component = &mca_rcache_grdma_component.super;
rcache->super.rcache_register = mca_rcache_grdma_register;
rcache->super.rcache_find = mca_rcache_grdma_find;
rcache->super.rcache_deregister = mca_rcache_grdma_deregister;
rcache->super.rcache_invalidate_range = mca_rcache_grdma_invalidate_range;
rcache->super.rcache_finalize = mca_rcache_grdma_finalize;
rcache->super.rcache_evict = mca_rcache_grdma_evict;
rcache->stat_cache_hit = rcache->stat_cache_miss = rcache->stat_evicted = 0;
rcache->stat_cache_found = rcache->stat_cache_notfound = 0;
OBJ_CONSTRUCT(&rcache->reg_list, opal_free_list_t);
opal_free_list_init (&rcache->reg_list, rcache->resources.sizeof_reg,
opal_cache_line_size,
OBJ_CLASS(mca_rcache_base_registration_t),
0, opal_cache_line_size, 0, -1, 32, NULL, 0,
NULL, NULL, NULL);
}
/** VPROTOCOL level functions (same as PML one)
*/
static mca_vprotocol_base_module_t *mca_vprotocol_pessimist_component_init( int* priority,
bool enable_progress_threads,
bool enable_mpi_threads)
{
V_OUTPUT_VERBOSE(500, "vprotocol_pessimist: component_init");
*priority = _priority;
/* sanity check */
if(enable_mpi_threads)
{
opal_output(0, "vprotocol_pessimist: component_init: threads are enabled, and not supported by vprotocol pessimist fault tolerant layer, will not load");
return NULL;
}
mca_vprotocol_pessimist.clock = 1;
mca_vprotocol_pessimist.replay = false;
OBJ_CONSTRUCT(&mca_vprotocol_pessimist.replay_events, opal_list_t);
OBJ_CONSTRUCT(&mca_vprotocol_pessimist.pending_events, opal_list_t);
OBJ_CONSTRUCT(&mca_vprotocol_pessimist.events_pool, opal_free_list_t);
opal_free_list_init (&mca_vprotocol_pessimist.events_pool,
sizeof(mca_vprotocol_pessimist_event_t),
opal_cache_line_size,
OBJ_CLASS(mca_vprotocol_pessimist_event_t),
0,opal_cache_line_size,
_free_list_num,
_free_list_max,
_free_list_inc,
NULL, 0, NULL, NULL, NULL);
mca_vprotocol_pessimist.event_buffer_max_length =
_event_buffer_size / sizeof(vprotocol_pessimist_mem_event_t);
mca_vprotocol_pessimist.event_buffer_length = 0;
mca_vprotocol_pessimist.event_buffer =
(vprotocol_pessimist_mem_event_t *) malloc(_event_buffer_size);
mca_vprotocol_pessimist.el_comm = MPI_COMM_NULL;
return &mca_vprotocol_pessimist.super;
}