int mca_btl_mx_proc_connect( mca_btl_mx_endpoint_t* module_endpoint )
{
int num_retry = 0;
mx_return_t mx_status;
mx_endpoint_addr_t mx_remote_addr;
module_endpoint->status = MCA_BTL_MX_CONNECTION_PENDING;
retry_connect:
mx_status = mx_connect( module_endpoint->endpoint_btl->mx_endpoint,
module_endpoint->mx_peer->nic_id, module_endpoint->mx_peer->endpoint_id,
mca_btl_mx_component.mx_filter, mca_btl_mx_component.mx_timeout, &mx_remote_addr );
if( MX_SUCCESS != mx_status ) {
if( MX_TIMEOUT == mx_status )
if( num_retry++ < mca_btl_mx_component.mx_connection_retries )
goto retry_connect;
{
char peer_name[MX_MAX_HOSTNAME_LEN];
if( MX_SUCCESS != mx_nic_id_to_hostname( module_endpoint->mx_peer->nic_id, peer_name ) )
sprintf( peer_name, "unknown %lx nic_id", (long)module_endpoint->mx_peer->nic_id );
opal_output( 0, "mx_connect fail for %s with key %x (error %s)\n\tUnique ID (local %x remote %x)\n",
peer_name, mca_btl_mx_component.mx_filter, mx_strerror(mx_status),
module_endpoint->endpoint_btl->mx_unique_network_id,
module_endpoint->mx_peer->unique_network_id );
}
module_endpoint->status = MCA_BTL_MX_NOT_REACHEABLE;
return OMPI_ERROR;
}
module_endpoint->mx_peer_addr = mx_remote_addr;
module_endpoint->status = MCA_BTL_MX_CONNECTED;
return OMPI_SUCCESS;
}
int
ompi_common_mx_initialize(void)
{
mx_return_t mx_return;
struct mca_mpool_base_resources_t mpool_resources;
int index, value;
ompi_common_mx_initialize_ref_cnt++;
if(ompi_common_mx_initialize_ref_cnt == 1) {
/* set the MX error handle to always return. This function is the
* only MX function allowed to be called before mx_init in order
* to make sure that if the MX is not up and running the MX
* library does not exit the application.
*/
mx_set_error_handler(MX_ERRORS_RETURN);
/* If we have a memory manager available, and
mpi_leave_pinned == -1, then set mpi_leave_pinned to 1.
We have a memory manager if:
- we have both FREE and MUNMAP support
- we have MUNMAP support and the linux mallopt */
value = opal_mem_hooks_support_level();
if ((value & (OPAL_MEMORY_FREE_SUPPORT | OPAL_MEMORY_MUNMAP_SUPPORT))
== (OPAL_MEMORY_FREE_SUPPORT | OPAL_MEMORY_MUNMAP_SUPPORT)) {
index = mca_base_param_find("mpi", NULL, "leave_pinned");
if (index >= 0)
if ((mca_base_param_lookup_int(index, &value) == OPAL_SUCCESS)
&& (value == -1)) {
ompi_mpi_leave_pinned = 1;
setenv("MX_RCACHE", "2", 1);
mpool_resources.regcache_clean = mx__regcache_clean;
ompi_common_mx_fake_mpool =
mca_mpool_base_module_create("fake", NULL, &mpool_resources);
if (!ompi_common_mx_fake_mpool) {
ompi_mpi_leave_pinned = 0;
setenv("MX_RCACHE", "0", 1);
opal_output(0, "Error creating fake mpool (error %s)\n",
strerror(errno));
}
}
}
/* initialize the mx library */
mx_return = mx_init();
if(MX_SUCCESS != mx_return) {
opal_output(0,
"Error in mx_init (error %s)\n",
mx_strerror(mx_return));
return OMPI_ERR_NOT_AVAILABLE;
}
}
return OMPI_SUCCESS;
}
mca_btl_base_descriptor_t* mca_btl_mx_prepare_dst( struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* endpoint,
struct mca_mpool_base_registration_t* registration,
struct ompi_convertor_t* convertor,
uint8_t order,
size_t reserve,
size_t* size,
uint32_t flags)
{
mca_btl_mx_module_t* mx_btl = (mca_btl_mx_module_t*)btl;
mca_btl_mx_frag_t* frag;
mx_return_t mx_return;
mx_segment_t mx_segment;
int rc;
MCA_BTL_MX_FRAG_ALLOC_USER(btl, frag, rc);
if( OPAL_UNLIKELY(NULL == frag) ) {
return NULL;
}
frag->segment[0].seg_len = *size;
ompi_convertor_get_current_pointer( convertor, (void**)&(frag->segment[0].seg_addr.pval) );
frag->segment[0].seg_key.key64 = (uint64_t)(intptr_t)frag;
mx_segment.segment_ptr = frag->segment[0].seg_addr.pval;
mx_segment.segment_length = frag->segment[0].seg_len;
mx_return = mx_irecv( mx_btl->mx_endpoint, &mx_segment, 1,
frag->segment[0].seg_key.key64,
BTL_MX_PUT_MASK, NULL, &(frag->mx_request) );
if( OPAL_UNLIKELY(MX_SUCCESS != mx_return) ) {
opal_output( 0, "Fail to re-register a fragment with the MX NIC ...\n" );
MCA_BTL_MX_FRAG_RETURN( btl, frag );
return NULL;
}
#ifdef HAVE_MX_FORGET
{
mx_return = mx_forget( mx_btl->mx_endpoint, &(frag->mx_request) );
if( OPAL_UNLIKELY(MX_SUCCESS != mx_return) ) {
opal_output( 0, "mx_forget failed in mca_btl_mx_prepare_dst with error %d (%s)\n",
mx_return, mx_strerror(mx_return) );
return NULL;
}
}
#endif
/* Allow the fragment to be recycled using the mca_btl_mx_free function */
frag->type = MCA_BTL_MX_SEND;
frag->base.des_dst = frag->segment;
frag->base.des_dst_cnt = 1;
frag->base.des_flags = flags;
frag->base.order = MCA_BTL_NO_ORDER;
return &frag->base;
}
int ompi_mtl_mx_module_init(){
mx_param_t mx_param;
mx_return_t mx_return;
/* setup params */
mx_param.key = MX_PARAM_UNEXP_QUEUE_MAX;
mx_param.val.unexp_queue_max = ompi_mtl_mx.mx_unexp_queue_max;
/* get a local endpoint */
mx_return = mx_open_endpoint(MX_ANY_NIC,
MX_ANY_ENDPOINT,
ompi_mtl_mx.mx_filter,
NULL,
0,
&ompi_mtl_mx.mx_endpoint);
if(mx_return != MX_SUCCESS) {
opal_output(ompi_mtl_base_output, "Error in mx_open_endpoint (error %s)\n", mx_strerror(mx_return));
return OMPI_ERROR;
}
/* get the endpoint address */
mx_return = mx_get_endpoint_addr( ompi_mtl_mx.mx_endpoint,
&ompi_mtl_mx.mx_endpoint_addr);
if(mx_return != MX_SUCCESS) {
opal_output(ompi_mtl_base_output, "Error in mx_get_endpoint_addr (error %s)\n", mx_strerror(mx_return));
return OMPI_ERROR;
}
mx_return = mx_decompose_endpoint_addr( ompi_mtl_mx.mx_endpoint_addr, &(ompi_mtl_mx.mx_addr.nic_id),
&(ompi_mtl_mx.mx_addr.endpoint_id) );
if(mx_return != MX_SUCCESS) {
opal_output(ompi_mtl_base_output, "Error in mx_decompose_endpoint_addr (error %s)\n", mx_strerror(mx_return));
return OMPI_ERROR;
}
ompi_modex_send( &mca_mtl_mx_component.super.mtl_version,
&ompi_mtl_mx.mx_addr,
sizeof(mca_mtl_mx_addr_t));
/* register the mtl mx progress function */
opal_progress_register(ompi_mtl_mx_progress);
return OMPI_SUCCESS;
}
int
ompi_mtl_mx_finalize(struct mca_mtl_base_module_t* mtl) {
mx_return_t mx_return;
opal_progress_unregister(ompi_mtl_mx_progress);
/* free resources */
mx_return = mx_close_endpoint(ompi_mtl_mx.mx_endpoint);
if(mx_return != MX_SUCCESS){
opal_output(ompi_mtl_base_output, "Error in mx_close_endpoint (error %s)\n", mx_strerror(mx_return));
return OMPI_ERROR;
}
return ompi_common_mx_finalize();
}
mca_mtl_mx_endpoint_t* mca_mtl_mx_endpoint_create(ompi_proc_t* ompi_proc) {
mca_mtl_mx_endpoint_t* mtl_mx_endpoint = NULL;
int rc;
mca_mtl_mx_addr_t *mx_peer;
size_t size;
mx_return_t mx_return;
int num_retry = 0;
/* get the remote proc's address (only one) */
rc = ompi_modex_recv(&mca_mtl_mx_component.super.mtl_version,
ompi_proc, (void**)&mx_peer, &size);
if( rc != OMPI_SUCCESS || size != sizeof(mca_mtl_mx_addr_t)) {
return NULL;
}
mtl_mx_endpoint = (mca_mtl_mx_endpoint_t*) OBJ_NEW(mca_mtl_mx_endpoint_t);
mtl_mx_endpoint->mx_peer = mx_peer;
retry_connect:
mx_return = mx_connect(ompi_mtl_mx.mx_endpoint,
mx_peer->nic_id,
mx_peer->endpoint_id,
ompi_mtl_mx.mx_filter,
ompi_mtl_mx.mx_timeout,
&mtl_mx_endpoint->mx_peer_addr);
if(MX_SUCCESS != mx_return) {
char peer_name[MX_MAX_HOSTNAME_LEN];
if(MX_TIMEOUT == mx_return) {
if( num_retry++ < ompi_mtl_mx.mx_retries ) {
goto retry_connect;
}
}
if(MX_SUCCESS != mx_nic_id_to_hostname( mx_peer->nic_id, peer_name)) {
sprintf( peer_name, "unknown %lx nic_id", (long)mx_peer->nic_id );
}
opal_output(ompi_mtl_base_output,
"mx_connect fail for %s with key %x (error %s)\n",
peer_name, ompi_mtl_mx.mx_filter, mx_strerror(mx_return) );
return NULL;
}
return mtl_mx_endpoint;
}
int
ompi_common_mx_finalize(void)
{
mx_return_t mx_return;
ompi_common_mx_initialize_ref_cnt--;
if( 0 == ompi_common_mx_initialize_ref_cnt ) {
if (ompi_common_mx_fake_mpool)
mca_mpool_base_module_destroy(ompi_common_mx_fake_mpool);
mx_return = mx_finalize();
if(mx_return != MX_SUCCESS){
opal_output(0, "Error in mx_finalize (error %s)\n", mx_strerror(mx_return));
return OMPI_ERROR;
}
}
return OMPI_SUCCESS;
}
int
ompi_mtl_mx_iprobe(struct mca_mtl_base_module_t* mtl,
struct ompi_communicator_t *comm,
int src,
int tag,
int *flag,
struct ompi_status_public_t *status)
{
uint32_t result;
mx_return_t ret;
mx_status_t mx_status;
uint64_t match_bits;
uint64_t mask_bits;
MX_SET_RECV_BITS(match_bits,
mask_bits,
comm->c_contextid,
src,
tag);
ret = mx_iprobe(ompi_mtl_mx.mx_endpoint,
match_bits,
mask_bits,
&mx_status,
&result);
if (MX_SUCCESS != ret) {
opal_output(ompi_mtl_base_output, "Error in mx_iprobe (error %s)\n", mx_strerror(ret));
return OMPI_ERROR;
}
if (result) {
if(MPI_STATUS_IGNORE != status) {
MX_GET_SRC(mx_status.match_info, status->MPI_SOURCE);
MX_GET_TAG(mx_status.match_info, status->MPI_TAG);
status->_ucount = mx_status.msg_length;
}
*flag = 1;
} else {
*flag = 0;
}
return OMPI_SUCCESS;
}
/**
* Initiate an asynchronous put.
*
* @param btl (IN) BTL module
* @param endpoint (IN) BTL addressing information
* @param descriptor (IN) Description of the data to be transferred
*/
static int mca_btl_mx_put( struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* endpoint,
struct mca_btl_base_descriptor_t* descriptor )
{
mca_btl_mx_module_t* mx_btl = (mca_btl_mx_module_t*)btl;
mca_btl_mx_frag_t* frag = (mca_btl_mx_frag_t*)descriptor;
mx_segment_t mx_segment[2];
mx_return_t mx_return;
uint32_t i = 0;
if( OPAL_UNLIKELY(MCA_BTL_MX_CONNECTED != ((mca_btl_mx_endpoint_t*)endpoint)->status) ) {
if( MCA_BTL_MX_NOT_REACHEABLE == ((mca_btl_mx_endpoint_t*)endpoint)->status )
return OMPI_ERROR;
if( MCA_BTL_MX_CONNECTION_PENDING == ((mca_btl_mx_endpoint_t*)endpoint)->status )
return OMPI_ERR_OUT_OF_RESOURCE;
if( OMPI_SUCCESS != mca_btl_mx_proc_connect( (mca_btl_mx_endpoint_t*)endpoint ) )
return OMPI_ERROR;
}
frag->endpoint = endpoint;
frag->type = MCA_BTL_MX_SEND;
descriptor->des_flags |= MCA_BTL_DES_SEND_ALWAYS_CALLBACK;
do {
mx_segment[i].segment_ptr = descriptor->des_src[i].seg_addr.pval;
mx_segment[i].segment_length = descriptor->des_src[i].seg_len;
} while (++i < descriptor->des_src_cnt);
mx_return = mx_isend( mx_btl->mx_endpoint, mx_segment, descriptor->des_src_cnt,
endpoint->mx_peer_addr,
descriptor->des_dst[0].seg_key.key64, frag,
&frag->mx_request );
if( OPAL_UNLIKELY(MX_SUCCESS != mx_return) ) {
opal_output( 0, "mx_isend fails with error %s\n", mx_strerror(mx_return) );
return OMPI_ERROR;
}
return OMPI_SUCCESS;
}
int
main(int argc, char **argv)
{
mx_endpoint_t ep;
uint64_t nic_id;
uint16_t my_eid;
uint64_t his_nic_id;
uint32_t board_id;
uint32_t filter;
uint16_t his_eid;
mx_endpoint_addr_t his_addr;
char *rem_host;
int len;
int iter;
int c;
int do_wait;
int do_bothways;
extern char *optarg;
mx_return_t ret;
#if DEBUG
extern int mx_debug_mask;
mx_debug_mask = 0xFFF;
#endif
mx_init();
MX_MUTEX_INIT(&stream_mutex);
/* set up defaults */
rem_host = NULL;
filter = FILTER;
my_eid = DFLT_EID;
his_eid = DFLT_EID;
board_id = MX_ANY_NIC;
len = DFLT_LEN;
iter = DFLT_ITER;
do_wait = 0;
do_bothways = 0;
num_threads = 1;
while ((c = getopt(argc, argv, "hd:e:f:n:b:r:l:N:Vvwx")) != EOF) switch(c) {
case 'd':
rem_host = optarg;
break;
case 'e':
my_eid = atoi(optarg);
break;
case 'f':
filter = atoi(optarg);
break;
case 'n':
sscanf(optarg, "%"SCNx64, &nic_id);
mx_nic_id_to_board_number(nic_id, &board_id);
break;
case 'b':
board_id = atoi(optarg);
break;
case 'r':
his_eid = atoi(optarg);
break;
case 'l':
len = atoi(optarg);
if (len > MAX_LEN) {
fprintf(stderr, "len too large, max is %d\n", MAX_LEN);
exit(1);
}
break;
case 'N':
iter = atoi(optarg);
break;
case 'V':
Verify = 1;
break;
case 'v':
do_verbose = 1;
break;
case 'w':
do_wait = 1;
break;
case 'x':
#if MX_THREAD_SAFE
do_bothways = 1;
#else
fprintf(stderr, "bi-directional mode only supported with threadsafe mx lib\n");
exit(1);
#endif
break;
case 'h':
default:
usage();
exit(1);
}
if (rem_host != NULL)
num_threads += do_bothways;
ret = mx_open_endpoint(board_id, my_eid, filter, NULL, 0, &ep);
if (ret != MX_SUCCESS) {
fprintf(stderr, "Failed to open endpoint %s\n", mx_strerror(ret));
exit(1);
}
/* If no host, we are receiver */
if (rem_host == NULL) {
if (do_verbose)
printf("Starting streaming receiver\n");
if (Verify) {
fprintf(stderr, "-V ignored. Verify must be set by sender\n");
Verify = 0;
}
if (do_wait)
receiver_blocking(ep, MATCH_VAL_MAIN, filter);
else
receiver_polling(ep, MATCH_VAL_MAIN, filter);
} else {
/* get address of destination */
mx_hostname_to_nic_id(rem_host, &his_nic_id);
mx_connect(ep, his_nic_id, his_eid, filter,
MX_INFINITE, &his_addr);
if (do_verbose)
printf("Starting streaming send to host %s\n",
rem_host);
if (Verify) printf("Verifying results\n");
/* start up the sender */
if (do_wait)
sender_blocking(ep, his_addr, iter, len,
do_bothways,MATCH_VAL_MAIN);
else
sender_polling(ep, his_addr, iter, len,
do_bothways, MATCH_VAL_MAIN);
}
mx_close_endpoint(ep);
mx_finalize();
exit(0);
}
int
ompi_mtl_mx_isend(struct mca_mtl_base_module_t* mtl,
struct ompi_communicator_t* comm,
int dest,
int tag,
struct opal_convertor_t *convertor,
mca_pml_base_send_mode_t mode,
bool blocking,
mca_mtl_request_t * mtl_request)
{
mx_return_t mx_return;
uint64_t match_bits;
mca_mtl_mx_request_t * mtl_mx_request = (mca_mtl_mx_request_t*) mtl_request;
size_t length;
ompi_proc_t* ompi_proc = ompi_comm_peer_lookup( comm, dest );
mca_mtl_mx_endpoint_t* mx_endpoint = (mca_mtl_mx_endpoint_t*) ompi_proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_MTL];
char* where;
assert(mtl == &ompi_mtl_mx.super);
MX_SET_SEND_BITS(match_bits, comm->c_contextid, comm->c_my_rank, tag);
ompi_mtl_datatype_pack(convertor,
&mtl_mx_request->mx_segment[0].segment_ptr,
&length,
&mtl_mx_request->free_after);
mtl_mx_request->mx_segment[0].segment_length = length;
mtl_mx_request->convertor = convertor;
mtl_mx_request->type = OMPI_MTL_MX_ISEND;
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_framework.framework_output,
"issend bits: 0x%016" PRIu64 "\n", match_bits));
if(mode == MCA_PML_BASE_SEND_SYNCHRONOUS) {
mx_return = mx_issend( ompi_mtl_mx.mx_endpoint,
mtl_mx_request->mx_segment,
1,
mx_endpoint->mx_peer_addr,
match_bits,
mtl_mx_request,
&mtl_mx_request->mx_request
);
where = "mx_issend";
} else {
mx_return = mx_isend( ompi_mtl_mx.mx_endpoint,
mtl_mx_request->mx_segment,
1,
mx_endpoint->mx_peer_addr,
match_bits,
mtl_mx_request,
&mtl_mx_request->mx_request
);
where = "mx_isend";
}
if( OPAL_UNLIKELY(mx_return != MX_SUCCESS) ) {
char peer_name[MX_MAX_HOSTNAME_LEN];
if(MX_SUCCESS != mx_nic_id_to_hostname( mx_endpoint->mx_peer->nic_id, peer_name)) {
sprintf( peer_name, "unknown %lx nic_id", (long)mx_endpoint->mx_peer->nic_id );
}
opal_output(ompi_mtl_base_framework.framework_output, "Error in %s (error %s) sending to %s\n",
where, mx_strerror(mx_return), peer_name);
return OMPI_ERROR;
}
return OMPI_SUCCESS;
}
int mca_btl_mx_send( struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* endpoint,
struct mca_btl_base_descriptor_t* descriptor,
mca_btl_base_tag_t tag )
{
mca_btl_mx_module_t* mx_btl = (mca_btl_mx_module_t*)btl;
mca_btl_mx_frag_t* frag = (mca_btl_mx_frag_t*)descriptor;
mx_segment_t mx_segment[2];
mx_return_t mx_return;
uint64_t total_length = 0, tag64;
uint32_t i = 0;
int btl_ownership = (descriptor->des_flags & MCA_BTL_DES_FLAGS_BTL_OWNERSHIP);
if( OPAL_UNLIKELY(MCA_BTL_MX_CONNECTED != ((mca_btl_mx_endpoint_t*)endpoint)->status) ) {
if( MCA_BTL_MX_NOT_REACHEABLE == ((mca_btl_mx_endpoint_t*)endpoint)->status )
return OMPI_ERROR;
if( MCA_BTL_MX_CONNECTION_PENDING == ((mca_btl_mx_endpoint_t*)endpoint)->status )
return OMPI_ERR_OUT_OF_RESOURCE;
if( OMPI_SUCCESS != mca_btl_mx_proc_connect( (mca_btl_mx_endpoint_t*)endpoint ) )
return OMPI_ERROR;
}
frag->endpoint = endpoint;
frag->type = MCA_BTL_MX_SEND;
do {
mx_segment[i].segment_ptr = descriptor->des_src[i].seg_addr.pval;
mx_segment[i].segment_length = descriptor->des_src[i].seg_len;
total_length += descriptor->des_src[i].seg_len;
} while (++i < descriptor->des_src_cnt);
tag64 = 0x01ULL | (((uint64_t)tag) << 8);
mx_return = mx_isend( mx_btl->mx_endpoint, mx_segment, descriptor->des_src_cnt,
endpoint->mx_peer_addr,
tag64, frag, &frag->mx_request );
if( OPAL_UNLIKELY(MX_SUCCESS != mx_return) ) {
opal_output( 0, "mx_isend fails with error %s\n", mx_strerror(mx_return) );
return OMPI_ERROR;
}
#ifdef HAVE_MX_FORGET
{
uint32_t mx_result;
mx_return = mx_ibuffered( mx_btl->mx_endpoint, &(frag->mx_request), &mx_result );
if( OPAL_UNLIKELY(MX_SUCCESS != mx_return) ) {
opal_output( 0, "mx_ibuffered failed with error %d (%s)\n",
mx_return, mx_strerror(mx_return) );
frag->base.des_flags |= MCA_BTL_DES_SEND_ALWAYS_CALLBACK;
return OMPI_ERROR;
}
if( mx_result ) {
mx_return = mx_forget( mx_btl->mx_endpoint, &(frag->mx_request) );
if( OPAL_UNLIKELY(MX_SUCCESS != mx_return) ) {
opal_output( 0, "mx_forget failed with error %d (%s)\n",
mx_return, mx_strerror(mx_return) );
frag->base.des_flags |= MCA_BTL_DES_SEND_ALWAYS_CALLBACK;
return OMPI_SUCCESS;
}
if( MCA_BTL_DES_SEND_ALWAYS_CALLBACK & frag->base.des_flags ) {
frag->base.des_cbfunc( &(mx_btl->super), frag->endpoint,
&(frag->base), OMPI_SUCCESS);
}
if( btl_ownership ) {
MCA_BTL_MX_FRAG_RETURN( mx_btl, frag );
}
return 1;
}
}
#endif
if( 2048 > total_length ) {
mx_status_t mx_status;
uint32_t mx_result;
/* let's check for completness */
mx_return = mx_test( mx_btl->mx_endpoint, &(frag->mx_request),
&mx_status, &mx_result );
if( OPAL_LIKELY(MX_SUCCESS == mx_return) ) {
if( mx_result ) {
if( MCA_BTL_DES_SEND_ALWAYS_CALLBACK & frag->base.des_flags ) {
frag->base.des_cbfunc( &(mx_btl->super), frag->endpoint,
&(frag->base), OMPI_SUCCESS);
}
if( btl_ownership ) {
MCA_BTL_MX_FRAG_RETURN( mx_btl, frag );
}
return 1;
}
}
}
frag->base.des_flags |= MCA_BTL_DES_SEND_ALWAYS_CALLBACK;
return OMPI_SUCCESS;
}
/**
* Initiate an inline send to the peer. If failure then return a descriptor.
*
* @param btl (IN) BTL module
* @param peer (IN) BTL peer addressing
*/
static int mca_btl_mx_sendi( struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* endpoint,
struct ompi_convertor_t* convertor,
void* header,
size_t header_size,
size_t payload_size,
uint8_t order,
uint32_t flags,
mca_btl_base_tag_t tag,
mca_btl_base_descriptor_t** descriptor )
{
mca_btl_mx_module_t* mx_btl = (mca_btl_mx_module_t*) btl;
size_t max_data;
if( OPAL_UNLIKELY(MCA_BTL_MX_CONNECTED != ((mca_btl_mx_endpoint_t*)endpoint)->status) ) {
if( MCA_BTL_MX_NOT_REACHEABLE == ((mca_btl_mx_endpoint_t*)endpoint)->status )
return OMPI_ERROR;
if( MCA_BTL_MX_CONNECTION_PENDING == ((mca_btl_mx_endpoint_t*)endpoint)->status )
return OMPI_ERR_OUT_OF_RESOURCE;
if( OMPI_SUCCESS != mca_btl_mx_proc_connect( (mca_btl_mx_endpoint_t*)endpoint ) )
return OMPI_ERROR;
}
if( !ompi_convertor_need_buffers(convertor) ) {
uint32_t mx_segment_count = 0;
uint64_t tag64 = 0x01ULL | (((uint64_t)tag) << 8);
mx_return_t mx_return;
mx_request_t mx_request;
mx_segment_t mx_segments[2], *mx_segment = mx_segments;
if( 0 != header_size ) {
mx_segment->segment_ptr = header;
mx_segment->segment_length = header_size;
mx_segment++;
mx_segment_count++;
}
if( 0 != payload_size ) {
struct iovec iov;
uint32_t iov_count = 1;
iov.iov_base = NULL;
iov.iov_len = payload_size;
(void)ompi_convertor_pack( convertor, &iov, &iov_count, &max_data );
assert( max_data == payload_size );
mx_segment->segment_ptr = iov.iov_base;
mx_segment->segment_length = max_data;
mx_segment_count++;
}
mx_return = mx_isend( mx_btl->mx_endpoint, mx_segments, mx_segment_count,
endpoint->mx_peer_addr, tag64, NULL, &mx_request );
if( OPAL_UNLIKELY(MX_SUCCESS != mx_return) ) {
opal_output( 0, "mx_isend fails with error %s\n", mx_strerror(mx_return) );
return OMPI_ERROR;
}
#ifdef HAVE_MX_FORGET
{
uint32_t mx_result;
mx_return = mx_ibuffered( mx_btl->mx_endpoint, &mx_request, &mx_result );
if( OPAL_UNLIKELY(MX_SUCCESS != mx_return) ) {
opal_output( 0, "mx_ibuffered failed with error %d (%s)\n",
mx_return, mx_strerror(mx_return) );
return OMPI_SUCCESS;
}
if( mx_result ) {
mx_return = mx_forget( mx_btl->mx_endpoint, &mx_request );
if( OPAL_UNLIKELY(MX_SUCCESS != mx_return) ) {
opal_output( 0, "mx_forget failed with error %d (%s)\n",
mx_return, mx_strerror(mx_return) );
}
}
return OMPI_SUCCESS;
}
#endif
}
/* No optimization on this path. Just allocate a descriptor and return it
* to the user.
*/
*descriptor = mca_btl_mx_alloc( btl, endpoint, order,
header_size + payload_size, flags );
return OMPI_ERR_RESOURCE_BUSY;
}
int
ompi_mtl_mx_send(struct mca_mtl_base_module_t* mtl,
struct ompi_communicator_t* comm,
int dest,
int tag,
struct opal_convertor_t *convertor,
mca_pml_base_send_mode_t mode)
{
mx_return_t mx_return;
uint64_t match_bits;
mca_mtl_mx_request_t mtl_mx_request;
size_t length;
mx_status_t mx_status;
uint32_t result;
ompi_proc_t* ompi_proc = ompi_comm_peer_lookup( comm, dest );
mca_mtl_mx_endpoint_t* mx_endpoint = (mca_mtl_mx_endpoint_t*) ompi_proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_MTL];
char* where;
assert(mtl == &ompi_mtl_mx.super);
MX_SET_SEND_BITS(match_bits, comm->c_contextid, comm->c_my_rank, tag);
ompi_mtl_datatype_pack(convertor,
&mtl_mx_request.mx_segment[0].segment_ptr,
&length,
&mtl_mx_request.free_after);
mtl_mx_request.mx_segment[0].segment_length = length;
mtl_mx_request.convertor = convertor;
mtl_mx_request.type = OMPI_MTL_MX_ISEND;
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_framework.framework_output,
"issend bits: 0x%016" PRIu64 "\n",
match_bits));
if(mode == MCA_PML_BASE_SEND_SYNCHRONOUS) {
mx_return = mx_issend( ompi_mtl_mx.mx_endpoint,
mtl_mx_request.mx_segment,
1,
mx_endpoint->mx_peer_addr,
match_bits,
&mtl_mx_request,
&mtl_mx_request.mx_request
);
where = "mx_issend";
} else {
mx_return = mx_isend( ompi_mtl_mx.mx_endpoint,
mtl_mx_request.mx_segment,
1,
mx_endpoint->mx_peer_addr,
match_bits,
&mtl_mx_request,
&mtl_mx_request.mx_request
);
where = "mx_isend";
}
if( OPAL_UNLIKELY(mx_return != MX_SUCCESS) ) {
char peer_name[MX_MAX_HOSTNAME_LEN];
if(MX_SUCCESS != mx_nic_id_to_hostname( mx_endpoint->mx_peer->nic_id, peer_name)) {
sprintf( peer_name, "unknown %lx nic_id", (long)mx_endpoint->mx_peer->nic_id );
}
opal_output(ompi_mtl_base_framework.framework_output, "Error in %s (error %s) sending to %s\n",
where, mx_strerror(mx_return), peer_name);
/* Free buffer if needed */
if(mtl_mx_request.free_after) {
free(mtl_mx_request.mx_segment[0].segment_ptr);
}
return OMPI_ERROR;
}
do {
mx_return = mx_test(ompi_mtl_mx.mx_endpoint,
&mtl_mx_request.mx_request,
&mx_status,
&result);
if( OPAL_UNLIKELY(mx_return != MX_SUCCESS) ) {
opal_output(ompi_mtl_base_framework.framework_output, "Error in mx_wait (error %s)\n", mx_strerror(mx_return));
abort();
}
if( OPAL_UNLIKELY(result && mx_status.code != MX_STATUS_SUCCESS) ) {
opal_output(ompi_mtl_base_framework.framework_output,
"Error in ompi_mtl_mx_send, mx_wait returned something other than MX_STATUS_SUCCESS: mx_status.code = %d.\n",
mx_status.code);
abort();
}
} while(!result);
/* Free buffer if needed */
if(mtl_mx_request.free_after) {
free(mtl_mx_request.mx_segment[0].segment_ptr);
}
return OMPI_SUCCESS;
}
int MPID_nem_mx_cancel_recv(MPIDI_VC_t *vc, MPID_Request *rreq)
{
mx_request_t *mx_request = NULL;
mx_return_t ret;
uint32_t result;
int mpi_errno = MPI_SUCCESS;
int handled = FALSE;
mx_request = &(REQ_FIELD(rreq,mx_request));
/* FIXME this test is probably not correct with multiple netmods */
/* We need to know to which netmod a recv request actually "belongs" to */
if(mx_request != NULL)
{
ret = mx_cancel(MPID_nem_mx_local_endpoint,mx_request,&result);
MPIU_ERR_CHKANDJUMP1(ret != MX_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**mx_cancel", "**mx_cancel %s", mx_strerror(ret));
if (result)
{
int found;
rreq->status.cancelled = TRUE;
found = MPIDI_CH3U_Recvq_DP(rreq);
MPIU_Assert(found);
rreq->status.count = 0;
MPID_REQUEST_SET_COMPLETED(rreq);
MPID_Request_release(rreq);
}
else
{
rreq->status.cancelled = FALSE;
MPIU_DBG_MSG_P(CH3_OTHER,VERBOSE,
"request 0x%08x already matched, unable to cancel", rreq->handle);
}
handled = TRUE;
}
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int ompi_mtl_mx_progress( void ) {
mx_return_t mx_return;
mx_request_t mx_request;
mx_status_t mx_status;
uint32_t result;
mca_mtl_mx_request_t* mtl_mx_request;
int completed = 0;
while(1){
mx_return = mx_ipeek(ompi_mtl_mx.mx_endpoint,
&mx_request,
&result);
if( OPAL_UNLIKELY(mx_return != MX_SUCCESS) ) {
opal_output(ompi_mtl_base_framework.framework_output, "Error in mx_ipeek (error %s)\n", mx_strerror(mx_return));
}
if(result) {
completed++;
mx_return = mx_test(ompi_mtl_mx.mx_endpoint,
&mx_request,
&mx_status,
&result);
if( OPAL_UNLIKELY(mx_return != MX_SUCCESS) ) {
opal_output(ompi_mtl_base_framework.framework_output, "Error in mx_test (error %s)\n", mx_strerror(mx_return));
abort();
}
if( OPAL_UNLIKELY(0 == result) ) {
opal_output(ompi_mtl_base_framework.framework_output, "Error in ompi_mtl_mx_progress, mx_ipeek returned a request, mx_test on the request resulted failure.\n");
abort();
}
mtl_mx_request = (mca_mtl_mx_request_t*) mx_status.context;
if(OMPI_MTL_MX_ISEND == mtl_mx_request->type) {
if(mtl_mx_request->free_after) {
free(mtl_mx_request->mx_segment[0].segment_ptr);
}
} else {
assert( OMPI_MTL_MX_IRECV == mtl_mx_request->type );
ompi_mtl_datatype_unpack(mtl_mx_request->convertor,
mtl_mx_request->mx_segment[0].segment_ptr,
mx_status.xfer_length);
/* set the status */
MX_GET_SRC(mx_status.match_info,
mtl_mx_request->super.ompi_req->req_status.MPI_SOURCE);
MX_GET_TAG(mx_status.match_info,
mtl_mx_request->super.ompi_req->req_status.MPI_TAG);
mtl_mx_request->super.ompi_req->req_status._ucount =
mx_status.xfer_length;
}
/* suppose everything went just fine ... */
mtl_mx_request->super.ompi_req->req_status.MPI_ERROR = OMPI_SUCCESS;
if( OPAL_UNLIKELY(MX_STATUS_SUCCESS != mx_status.code) ) {
if( MX_STATUS_TRUNCATED == mx_status.code ) {
mtl_mx_request->super.ompi_req->req_status.MPI_ERROR = MPI_ERR_TRUNCATE;
} else {
mtl_mx_request->super.ompi_req->req_status.MPI_ERROR = MPI_ERR_INTERN;
}
return completed;
}
mtl_mx_request->super.completion_callback(&mtl_mx_request->super);
return completed;
} else {
return completed;
}
}
}
int
MPID_nem_mx_vc_init (MPIDI_VC_t *vc)
{
uint32_t threshold;
MPIDI_CH3I_VC *vc_ch = VC_CH(vc);
int mpi_errno = MPI_SUCCESS;
/* first make sure that our private fields in the vc fit into the area provided */
MPIU_Assert(sizeof(MPID_nem_mx_vc_area) <= MPID_NEM_VC_NETMOD_AREA_LEN);
#ifdef ONDEMAND
VC_FIELD(vc, local_connected) = 0;
VC_FIELD(vc, remote_connected) = 0;
#else
{
char *business_card;
int val_max_sz;
int ret;
#ifdef USE_PMI2_API
val_max_sz = PMI2_MAX_VALLEN;
#else
mpi_errno = PMI_KVS_Get_value_length_max(&val_max_sz);
#endif
business_card = (char *)MPIU_Malloc(val_max_sz);
mpi_errno = vc->pg->getConnInfo(vc->pg_rank, business_card,val_max_sz, vc->pg);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPID_nem_mx_get_from_bc (business_card, &VC_FIELD(vc, remote_endpoint_id), &VC_FIELD(vc, remote_nic_id));
if (mpi_errno) MPIU_ERR_POP (mpi_errno);
MPIU_Free(business_card);
ret = mx_connect(MPID_nem_mx_local_endpoint,VC_FIELD(vc, remote_nic_id),VC_FIELD(vc, remote_endpoint_id),
MPID_NEM_MX_FILTER,MX_INFINITE,&(VC_FIELD(vc, remote_endpoint_addr)));
MPIU_ERR_CHKANDJUMP1 (ret != MX_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**mx_connect", "**mx_connect %s", mx_strerror (ret));
mx_set_endpoint_addr_context(VC_FIELD(vc, remote_endpoint_addr),(void *)vc);
MPIDI_CHANGE_VC_STATE(vc, ACTIVE);
}
#endif
mx_get_info(MPID_nem_mx_local_endpoint, MX_COPY_SEND_MAX, NULL, 0, &threshold, sizeof(uint32_t));
vc->eager_max_msg_sz = threshold;
vc->rndvSend_fn = NULL;
vc->sendNoncontig_fn = MPID_nem_mx_SendNoncontig;
vc->comm_ops = &comm_ops;
vc_ch->iStartContigMsg = MPID_nem_mx_iStartContigMsg;
vc_ch->iSendContig = MPID_nem_mx_iSendContig;
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int init_mx( MPIDI_PG_t *pg_p )
{
mx_endpoint_addr_t local_endpoint_addr;
mx_return_t ret;
mx_param_t param;
int mpi_errno = MPI_SUCCESS;
int r;
r = MPL_putenv("MX_DISABLE_SHARED=1");
MPIU_ERR_CHKANDJUMP(r, mpi_errno, MPI_ERR_OTHER, "**putenv");
r = MPL_putenv("MX_DISABLE_SELF=1");
MPIU_ERR_CHKANDJUMP(r, mpi_errno, MPI_ERR_OTHER, "**putenv");
ret = mx_init();
MPIU_ERR_CHKANDJUMP1 (ret != MX_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**mx_init", "**mx_init %s", mx_strerror (ret));
mx_set_error_handler(MX_ERRORS_RETURN);
/*
ret = mx_get_info(NULL, MX_NIC_COUNT, NULL, 0, &nic_count, sizeof(int));
MPIU_ERR_CHKANDJUMP1 (ret != MX_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**mx_get_info", "**mx_get_info %s", mx_strerror (ret));
count = ++nic_count;
mx_nics = (uint64_t *)MPIU_Malloc(count*sizeof(uint64_t));
ret = mx_get_info(NULL, MX_NIC_IDS, NULL, 0, mx_nics, count*sizeof(uint64_t));
MPIU_ERR_CHKANDJUMP1 (ret != MX_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**mx_get_info", "**mx_get_info %s", mx_strerror (ret));
do{
ret = mx_nic_id_to_board_number(mx_nics[index],&mx_board_num);
index++;
}while(ret != MX_SUCCESS);
*/
#ifndef USE_CTXT_AS_MARK
param.key = MX_PARAM_CONTEXT_ID;
param.val.context_id.bits = NEM_MX_MATCHING_BITS - SHIFT_TYPE;
param.val.context_id.shift = SHIFT_TYPE;
ret = mx_open_endpoint(MX_ANY_NIC,MX_ANY_ENDPOINT,MPID_NEM_MX_FILTER,¶m,1,&MPID_nem_mx_local_endpoint);
#else
ret = mx_open_endpoint(MX_ANY_NIC,MX_ANY_ENDPOINT,MPID_NEM_MX_FILTER,NULL,0,&MPID_nem_mx_local_endpoint);
#endif
MPIU_ERR_CHKANDJUMP1 (ret != MX_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**mx_open_endpoint", "**mx_open_endpoint %s", mx_strerror (ret));
ret = mx_get_endpoint_addr(MPID_nem_mx_local_endpoint,&local_endpoint_addr);
MPIU_ERR_CHKANDJUMP1 (ret != MX_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**mx_get_endpoint_addr", "**mx_get_endpoint_addr %s", mx_strerror (ret));
ret = mx_decompose_endpoint_addr(local_endpoint_addr,&MPID_nem_mx_local_nic_id,&MPID_nem_mx_local_endpoint_id);
MPIU_ERR_CHKANDJUMP1 (ret != MX_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**mx_decompose_endpoint_addr", "**mx_decompose_endpoint_addr %s", mx_strerror (ret));
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int mca_btl_mx_register( struct mca_btl_base_module_t* btl,
mca_btl_base_tag_t tag,
mca_btl_base_module_recv_cb_fn_t cbfunc,
void* cbdata )
{
mca_btl_mx_module_t* mx_btl = (mca_btl_mx_module_t*) btl;
#if 0
if( (NULL != cbfunc) && ( 0 == mca_btl_mx_component.mx_use_unexpected) ) {
#endif
if( NULL != cbfunc ) {
mca_btl_mx_frag_t* frag;
mx_return_t mx_return;
mx_segment_t mx_segment;
int i, rc;
/* Post the receives if there is no unexpected handler */
for( i = 0; i < mca_btl_mx_component.mx_max_posted_recv; i++ ) {
MCA_BTL_MX_FRAG_ALLOC_EAGER( mx_btl, frag, rc );
if( NULL == frag ) {
opal_output( 0, "mca_btl_mx_register: unable to allocate more eager fragments\n" );
if( 0 == i ) {
return OMPI_ERROR;
}
break; /* some fragments are already registered. Try to continue... */
}
frag->base.des_dst = frag->segment;
frag->base.des_dst_cnt = 1;
frag->base.des_src = NULL;
frag->base.des_src_cnt = 0;
frag->mx_frag_list = NULL;
frag->type = MCA_BTL_MX_RECV;
mx_segment.segment_ptr = (void*)(frag+1);
mx_segment.segment_length = mx_btl->super.btl_eager_limit;
mx_return = mx_irecv( mx_btl->mx_endpoint, &mx_segment, 1,
0x01ULL, BTL_MX_RECV_MASK,
frag, &(frag->mx_request) );
if( MX_SUCCESS != mx_return ) {
opal_output( 0, "mca_btl_mx_register: mx_irecv failed with status %d (%s)\n",
mx_return, mx_strerror(mx_return) );
MCA_BTL_MX_FRAG_RETURN( mx_btl, frag );
return OMPI_ERROR;
}
}
}
return OMPI_SUCCESS;
}
/**
* Allocate a segment.
*
* @param btl (IN) BTL module
* @param size (IN) Request segment size.
*/
mca_btl_base_descriptor_t* mca_btl_mx_alloc( struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* endpoint,
uint8_t order,
size_t size,
uint32_t flags)
{
mca_btl_mx_module_t* mx_btl = (mca_btl_mx_module_t*) btl;
mca_btl_mx_frag_t* frag;
int rc;
MCA_BTL_MX_FRAG_ALLOC_EAGER(mx_btl, frag, rc);
if( OPAL_UNLIKELY(NULL == frag) ) {
return NULL;
}
frag->segment[0].seg_len =
size <= mx_btl->super.btl_eager_limit ?
size : mx_btl->super.btl_eager_limit ;
frag->segment[0].seg_addr.pval = (void*)(frag+1);
frag->base.des_src = frag->segment;
frag->base.des_src_cnt = 1;
frag->base.des_flags = flags;
frag->base.order = MCA_BTL_NO_ORDER;
return (mca_btl_base_descriptor_t*)frag;
}
int ompi_mtl_mx_module_init(){
mx_param_t mx_param;
mx_return_t mx_return;
int32_t nic, ep;
/* setup params */
mx_param.key = MX_PARAM_UNEXP_QUEUE_MAX;
mx_param.val.unexp_queue_max = ompi_mtl_mx.mx_unexp_queue_max;
/* get a local endpoint */
nic = ompi_mtl_mx.mx_board_num;
if (nic < 0) {
nic = MX_ANY_NIC;
}
ep = ompi_mtl_mx.mx_endpoint_num;
if (ep < 0) {
ep = MX_ANY_ENDPOINT;
}
mx_return = mx_open_endpoint(nic,
ep,
ompi_mtl_mx.mx_filter,
NULL,
0,
&ompi_mtl_mx.mx_endpoint);
if(mx_return != MX_SUCCESS) {
opal_output(ompi_mtl_base_framework.framework_output, "Error in mx_open_endpoint (error %s)\n", mx_strerror(mx_return));
return OMPI_ERROR;
}
/* get the endpoint address */
mx_return = mx_get_endpoint_addr( ompi_mtl_mx.mx_endpoint,
&ompi_mtl_mx.mx_endpoint_addr);
if(mx_return != MX_SUCCESS) {
opal_output(ompi_mtl_base_framework.framework_output, "Error in mx_get_endpoint_addr (error %s)\n", mx_strerror(mx_return));
return OMPI_ERROR;
}
mx_return = mx_decompose_endpoint_addr( ompi_mtl_mx.mx_endpoint_addr, &(ompi_mtl_mx.mx_addr.nic_id),
&(ompi_mtl_mx.mx_addr.endpoint_id) );
if(mx_return != MX_SUCCESS) {
opal_output(ompi_mtl_base_framework.framework_output, "Error in mx_decompose_endpoint_addr (error %s)\n", mx_strerror(mx_return));
return OMPI_ERROR;
}
opal_output_verbose(10, ompi_mtl_base_framework.framework_output,
"mtl:mx: local nic %d, endpoint %d, got nic %d, ep %d\n", nic, ep,
(int)ompi_mtl_mx.mx_addr.nic_id,
ompi_mtl_mx.mx_addr.endpoint_id);
ompi_modex_send( &mca_mtl_mx_component.super.mtl_version,
&ompi_mtl_mx.mx_addr,
sizeof(mca_mtl_mx_addr_t));
/* register the mtl mx progress function */
opal_progress_register(ompi_mtl_mx_progress);
return OMPI_SUCCESS;
}
int MPID_nem_mx_cancel_send(MPIDI_VC_t *vc, MPID_Request *sreq)
{
mx_request_t *mx_request = NULL;
mx_return_t ret;
uint32_t result;
int mpi_errno = MPI_SUCCESS;
int handled = FALSE;
if (!VC_CH(vc)->is_local)
{
mx_request = &(REQ_FIELD(sreq,mx_request));
ret = mx_cancel(MPID_nem_mx_local_endpoint,mx_request,&result);
MPIU_ERR_CHKANDJUMP1(ret != MX_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**mx_cancel", "**mx_cancel %s", mx_strerror(ret));
if (result)
{
sreq->status.cancelled = TRUE;
sreq->cc = 0;
MPIU_Object_set_ref(sreq, 1);
MPID_nem_mx_pending_send_req--;
}
else
{
sreq->status.cancelled = FALSE;
}
handled = TRUE;
}
fn_exit:
return handled;
fn_fail:
goto fn_exit;
}
