/* initialize short copy blocks */
void
ompi_mtl_portals_short_setup()
{
int ret;
int i;
if ((ompi_mtl_portals.ptl_num_copy_blocks > 0) && (ompi_mtl_portals.ptl_copy_block_len > 0)) {
ompi_mtl_portals.ptl_short_md.length = ompi_mtl_portals.ptl_num_copy_blocks *
ompi_mtl_portals.ptl_copy_block_len;
ompi_mtl_portals.ptl_short_md.start = malloc(ompi_mtl_portals.ptl_short_md.length);
if (NULL == ompi_mtl_portals.ptl_short_md.start ) {
ompi_mtl_portals.ptl_num_copy_blocks = 0;
return;
}
ompi_mtl_portals.ptl_short_md.threshold = PTL_MD_THRESH_INF;
ompi_mtl_portals.ptl_short_md.max_size = 0;
ompi_mtl_portals.ptl_short_md.options = PTL_MD_EVENT_START_DISABLE;
ompi_mtl_portals.ptl_short_md.user_ptr = &ptl_short_request;
ompi_mtl_portals.ptl_short_md.eq_handle = ompi_mtl_portals.ptl_eq_h;
ret = PtlMDBind(ompi_mtl_portals.ptl_ni_h,
ompi_mtl_portals.ptl_short_md,
PTL_RETAIN,
&ompi_mtl_portals.ptl_short_md_h);
if (PTL_OK != ret) {
free(ompi_mtl_portals.ptl_short_md.start);
ompi_mtl_portals.ptl_num_copy_blocks = 0;
return;
}
ptl_short_request.event_callback = ompi_mtl_portals_short_callback;
ompi_mtl_portals.ptl_copy_block_free_list = malloc(ompi_mtl_portals.ptl_num_copy_blocks * sizeof(int));
if (NULL == ompi_mtl_portals.ptl_copy_block_free_list) {
free(ompi_mtl_portals.ptl_short_md.start);
ompi_mtl_portals.ptl_num_copy_blocks = 0;
return;
}
for (i=0; i<ompi_mtl_portals.ptl_num_copy_blocks; i++) {
ompi_mtl_portals.ptl_copy_block_free_list[i] = i;
}
ompi_mtl_portals.ptl_copy_block_first_free = 0;
}
}
static int
read_msg(void *start, ptl_size_t length, ptl_process_t target,
ptl_match_bits_t match_bits, ptl_size_t remote_offset,
ompi_mtl_portals4_recv_request_t *request)
{
ptl_md_t md;
int ret;
/* FIX ME: This needs to be on the send eq... */
md.start = start;
md.length = length;
md.options = 0;
md.eq_handle = ompi_mtl_portals4.send_eq_h;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(ompi_mtl_portals4.ni_h,
&md,
&request->md_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMDBind failed: %d",
__FILE__, __LINE__, ret);
return OMPI_ERR_OUT_OF_RESOURCE;
}
#if OMPI_MTL_PORTALS4_FLOW_CONTROL
while (OPAL_UNLIKELY(OPAL_THREAD_ADD32(&ompi_mtl_portals4.flowctl.send_slots, -1) < 0)) {
OPAL_THREAD_ADD32(&ompi_mtl_portals4.flowctl.send_slots, 1);
ompi_mtl_portals4_progress();
}
#endif
ret = PtlGet(request->md_h,
0,
md.length,
target,
ompi_mtl_portals4.read_idx,
match_bits,
remote_offset,
request);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlGet failed: %d",
__FILE__, __LINE__, ret);
PtlMDRelease(request->md_h);
return OMPI_ERR_OUT_OF_RESOURCE;
}
return OMPI_SUCCESS;
}
int MPID_nem_ptl_rptl_ptinit(ptl_handle_ni_t ni_handle, ptl_handle_eq_t eq_handle, ptl_pt_index_t data_pt,
ptl_pt_index_t control_pt)
{
int ret = PTL_OK;
struct rptl *rptl;
int mpi_errno = MPI_SUCCESS;
int i;
ptl_md_t md;
MPIU_CHKPMEM_DECL(2);
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_PTL_RPTL_PTINIT);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_PTL_RPTL_PTINIT);
/* setup the parts of rptls that can be done before world size or
* target information */
MPIU_CHKPMEM_MALLOC(rptl, struct rptl *, sizeof(struct rptl), mpi_errno, "rptl");
MPL_DL_APPEND(rptl_info.rptl_list, rptl);
rptl->local_state = RPTL_LOCAL_STATE_ACTIVE;
rptl->pause_ack_counter = 0;
rptl->data.ob_max_count = 0;
rptl->data.ob_curr_count = 0;
rptl->data.pt = data_pt;
rptl->control.pt = control_pt;
rptl->ni = ni_handle;
rptl->eq = eq_handle;
md.start = 0;
md.length = (ptl_size_t) (-1);
md.options = 0x0;
md.eq_handle = rptl->eq;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(rptl->ni, &md, &rptl->md);
RPTLU_ERR_POP(ret, "Error binding new global MD\n");
/* post world_size number of empty buffers on the control portal */
if (rptl->control.pt != PTL_PT_ANY) {
MPIU_CHKPMEM_MALLOC(rptl->control.me, ptl_handle_me_t *,
2 * rptl_info.world_size * sizeof(ptl_handle_me_t), mpi_errno,
"rptl target info");
for (i = 0; i < 2 * rptl_info.world_size; i++) {
ret = rptli_post_control_buffer(rptl->ni, rptl->control.pt, &rptl->control.me[i]);
RPTLU_ERR_POP(ret, "Error in rptli_post_control_buffer\n");
}
rptl->control.me_idx = 0;
}
static int handler_recv_dequeue_large(const ptl_event_t *e)
{
int mpi_errno = MPI_SUCCESS;
MPID_Request *const rreq = e->user_ptr;
MPIDI_VC_t *vc;
MPID_nem_ptl_vc_area *vc_ptl;
int ret;
int dt_contig;
MPIDI_msg_sz_t data_sz;
MPID_Datatype *dt_ptr;
MPI_Aint dt_true_lb;
MPI_Aint last;
MPIU_CHKPMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_HANDLER_RECV_DEQUEUE_LARGE);
MPIDI_FUNC_ENTER(MPID_STATE_HANDLER_RECV_DEQUEUE_LARGE);
MPIU_Assert(e->type == PTL_EVENT_PUT || e->type == PTL_EVENT_PUT_OVERFLOW);
MPIDI_Comm_get_vc(rreq->comm, NPTL_MATCH_GET_RANK(e->match_bits), &vc);
vc_ptl = VC_PTL(vc);
dequeue_req(e);
MPIDI_Datatype_get_info(rreq->dev.user_count, rreq->dev.datatype, dt_contig, data_sz, dt_ptr, dt_true_lb);
/* unpack data from unexpected buffer first */
if (e->type == PTL_EVENT_PUT_OVERFLOW) {
if (dt_contig) {
MPIU_Memcpy((char *)rreq->dev.user_buf + dt_true_lb, e->start, e->mlength);
} else {
last = e->mlength;
MPID_Segment_unpack(rreq->dev.segment_ptr, 0, &last, e->start);
MPIU_Assert(last == e->mlength);
rreq->dev.segment_first = e->mlength;
}
}
if (!(e->hdr_data & NPTL_LARGE)) {
/* all data has already been received; we're done */
mpi_errno = handler_recv_complete(e);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
goto fn_exit;
}
MPIU_Assert (e->mlength == PTL_LARGE_THRESHOLD);
/* we need to GET the rest of the data from the sender's buffer */
if (dt_contig) {
big_get((char *)rreq->dev.user_buf + dt_true_lb + PTL_LARGE_THRESHOLD, data_sz - PTL_LARGE_THRESHOLD,
vc, e->match_bits, rreq);
goto fn_exit;
}
/* noncontig recv buffer */
last = rreq->dev.segment_size;
rreq->dev.iov_count = MPL_IOV_LIMIT;
MPID_Segment_pack_vector(rreq->dev.segment_ptr, rreq->dev.segment_first, &last, rreq->dev.iov, &rreq->dev.iov_count);
if (last == rreq->dev.segment_size && rreq->dev.segment_size <= MPIDI_nem_ptl_ni_limits.max_msg_size + PTL_LARGE_THRESHOLD) {
/* Rest of message fits in one IOV */
ptl_md_t md;
md.start = rreq->dev.iov;
md.length = rreq->dev.iov_count;
md.options = PTL_IOVEC;
md.eq_handle = MPIDI_nem_ptl_origin_eq;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(MPIDI_nem_ptl_ni, &md, &REQ_PTL(rreq)->md);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmdbind", "**ptlmdbind %s", MPID_nem_ptl_strerror(ret));
REQ_PTL(rreq)->event_handler = handler_recv_complete;
ret = MPID_nem_ptl_rptl_get(REQ_PTL(rreq)->md, 0, rreq->dev.segment_size - rreq->dev.segment_first, vc_ptl->id, vc_ptl->ptg,
e->match_bits, 0, rreq);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlget", "**ptlget %s", MPID_nem_ptl_strerror(ret));
goto fn_exit;
}
/* message won't fit in a single IOV, allocate buffer and unpack when received */
/* FIXME: For now, allocate a single large buffer to hold entire message */
MPIU_CHKPMEM_MALLOC(REQ_PTL(rreq)->chunk_buffer[0], void *, data_sz - PTL_LARGE_THRESHOLD,
mpi_errno, "chunk_buffer");
big_get(REQ_PTL(rreq)->chunk_buffer[0], data_sz - PTL_LARGE_THRESHOLD, vc, e->match_bits, rreq);
fn_exit:
MPIU_CHKPMEM_COMMIT();
fn_exit2:
MPIDI_FUNC_EXIT(MPID_STATE_HANDLER_RECV_DEQUEUE_LARGE);
return mpi_errno;
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit2;
}
static int
btl_portals4_init_interface(void)
{
mca_btl_portals4_module_t *portals4_btl;
unsigned int ret, interface;
ptl_md_t md;
ptl_me_t me;
// The initialisation of EQ, PT and ME must be done after the SetMap !
for (interface=0; interface<mca_btl_portals4_component.num_btls; interface++) {
portals4_btl = mca_btl_portals4_component.btls[interface];
/* create event queue */
ret = PtlEQAlloc(portals4_btl->portals_ni_h,
mca_btl_portals4_component.recv_queue_size,
&portals4_btl->recv_eq_h);
if (PTL_OK != ret) {
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"%s:%d: PtlEQAlloc failed for NI %d: %d",
__FILE__, __LINE__, interface, ret);
goto error;
}
mca_btl_portals4_component.eqs_h[interface] = portals4_btl->recv_eq_h;
OPAL_OUTPUT_VERBOSE((90, opal_btl_base_framework.framework_output,
"PtlEQAlloc (recv_eq=%d) OK for NI %d\n", portals4_btl->recv_eq_h, interface));
/* Create recv_idx portal table entry */
ret = PtlPTAlloc(portals4_btl->portals_ni_h,
PTL_PT_ONLY_TRUNCATE,
portals4_btl->recv_eq_h,
REQ_BTL_TABLE_ID,
&portals4_btl->recv_idx);
if (PTL_OK != ret) {
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"%s:%d: PtlPTAlloc failed for NI %d: %d",
__FILE__, __LINE__, interface, ret);
goto error;
}
OPAL_OUTPUT_VERBOSE((90, opal_btl_base_framework.framework_output,
"PtlPTAlloc (recv_idx) OK for NI %d recv_idx=%d", interface, portals4_btl->recv_idx));
if (portals4_btl->recv_idx != REQ_BTL_TABLE_ID) {
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"%s:%d: PtlPTAlloc did not allocate the requested PT: %d",
__FILE__, __LINE__, portals4_btl->recv_idx);
goto error;
}
/* bind zero-length md for sending acks */
md.start = NULL;
md.length = 0;
md.options = 0;
md.eq_handle = PTL_EQ_NONE;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(portals4_btl->portals_ni_h,
&md,
&portals4_btl->zero_md_h);
if (PTL_OK != ret) {
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"%s:%d: PtlMDBind failed for NI %d: %d",
__FILE__, __LINE__, interface, ret);
goto error;
}
OPAL_OUTPUT_VERBOSE((90, opal_btl_base_framework.framework_output,
"PtlMDBind (zero-length md=%d) OK for NI %d", portals4_btl->zero_md_h, interface));
/* Bind MD across all memory */
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = 0;
md.eq_handle = portals4_btl->recv_eq_h;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(portals4_btl->portals_ni_h,
&md,
&portals4_btl->send_md_h);
if (PTL_OK != ret) {
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"%s:%d: PtlMDBind failed for NI %d: %d\n",
__FILE__, __LINE__, interface, ret);
goto error;
}
/* Handle long overflows */
me.start = NULL;
me.length = 0;
me.ct_handle = PTL_CT_NONE;
me.min_free = 0;
me.uid = PTL_UID_ANY;
me.options = PTL_ME_OP_PUT |
PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_COMM_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE;
if (mca_btl_portals4_component.use_logical) {
me.match_id.rank = PTL_RANK_ANY;
} else {
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
}
me.match_bits = BTL_PORTALS4_LONG_MSG;
me.ignore_bits = BTL_PORTALS4_CONTEXT_MASK |
BTL_PORTALS4_SOURCE_MASK |
BTL_PORTALS4_TAG_MASK;
ret = PtlMEAppend(portals4_btl->portals_ni_h,
portals4_btl->recv_idx,
&me,
PTL_OVERFLOW_LIST,
NULL,
&portals4_btl->long_overflow_me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"%s:%d: PtlMEAppend failed for NI %d: %d",
__FILE__, __LINE__, interface, ret);
goto error;
}
OPAL_OUTPUT_VERBOSE((90, opal_btl_base_framework.framework_output, "PtlMEAppend (overflow list) OK for NI %d", interface));
}
ret = mca_btl_portals4_recv_enable(portals4_btl);
if (PTL_OK != ret) {
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"%s:%d: Initialization of recv buffer failed: %d",
__FILE__, __LINE__, ret);
goto error;
}
return OPAL_SUCCESS;
error:
opal_output_verbose(1, opal_btl_base_framework.framework_output, "Error in btl_portals4_init_interface");
for (interface=0; interface<mca_btl_portals4_component.num_btls; interface++) {
portals4_btl = mca_btl_portals4_component.btls[interface];
if (NULL != portals4_btl) mca_btl_portals4_free_module(portals4_btl);
}
mca_btl_portals4_component.num_btls = 0;
if (NULL != mca_btl_portals4_component.btls) free(mca_btl_portals4_component.btls);
if (NULL != mca_btl_portals4_component.eqs_h) free(mca_btl_portals4_component.eqs_h);
mca_btl_portals4_component.btls = NULL;
mca_btl_portals4_component.eqs_h = NULL;
return OPAL_ERROR;
}
static int
ompi_mtl_portals_get_data(ompi_mtl_portals_event_t *recv_event,
struct ompi_convertor_t *convertor,
ompi_mtl_portals_request_t *ptl_request)
{
int ret;
ptl_md_t md;
ptl_handle_md_t md_h;
size_t buflen;
if (PTL_IS_SHORT_MSG(recv_event->ev.match_bits)) {
/* the buffer is sitting in the short message queue */
struct iovec iov;
uint32_t iov_count = 1;
size_t max_data;
ompi_mtl_portals_recv_short_block_t *block =
recv_event->ev.md.user_ptr;
iov.iov_base = (((char*) recv_event->ev.md.start) + recv_event->ev.offset);
iov.iov_len = recv_event->ev.mlength;
max_data = iov.iov_len;
/* see if this message filled the receive block */
if (recv_event->ev.md.length - (recv_event->ev.offset +
recv_event->ev.mlength) <
recv_event->ev.md.max_size) {
block->full = true;
}
/* pull out the data */
if (iov.iov_len > 0) {
ompi_convertor_unpack(convertor, &iov, &iov_count,
&max_data );
}
/* if synchronous, return an ack */
if (PTL_IS_SYNC_MSG(recv_event->ev)) {
md.length = 0;
md.start = (((char*) recv_event->ev.md.start) + recv_event->ev.offset);
md.threshold = 1; /* send */
md.options = PTL_MD_EVENT_START_DISABLE;
md.user_ptr = NULL;
md.eq_handle = ompi_mtl_portals.ptl_eq_h;
ret = PtlMDBind(ompi_mtl_portals.ptl_ni_h, md,
PTL_UNLINK, &md_h);
if (PTL_OK != ret) {
opal_output(fileno(stderr)," Error returned from PtlMDBind. Error code - %d \n",ret);
abort();
}
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_output,
"acking recv: 0x%016llx\n",
recv_event->ev.match_bits));
ret = PtlPut(md_h,
PTL_NO_ACK_REQ,
recv_event->ev.initiator,
OMPI_MTL_PORTALS_ACK_TABLE_ID,
0,
recv_event->ev.hdr_data,
0,
0);
if (PTL_OK != ret) {
opal_output(fileno(stderr)," Error returned from PtlPut. Error code - %d \n",ret);
abort();
}
}
/* finished with our buffer space */
ompi_mtl_portals_return_block_part(&ompi_mtl_portals, block);
ompi_convertor_get_packed_size(convertor, &buflen);
ptl_request->super.ompi_req->req_status.MPI_SOURCE =
PTL_GET_SOURCE(recv_event->ev.match_bits);
ptl_request->super.ompi_req->req_status.MPI_TAG =
PTL_GET_TAG(recv_event->ev.match_bits);
ptl_request->super.ompi_req->req_status.MPI_ERROR =
(recv_event->ev.rlength > buflen) ?
MPI_ERR_TRUNCATE : MPI_SUCCESS;
ptl_request->super.ompi_req->req_status._count =
recv_event->ev.mlength;
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_output,
"recv complete: 0x%016llx\n",
recv_event->ev.match_bits));
ptl_request->super.completion_callback(&ptl_request->super);
} else {
ret = ompi_mtl_datatype_recv_buf(convertor, &md.start, &buflen,
&ptl_request->free_after);
if (OMPI_SUCCESS != ret) {
opal_output(fileno(stderr)," Error returned from ompi_mtl_datatype_recv_buf. Error code - %d \n",ret);
abort();
}
md.length = (recv_event->ev.rlength > buflen) ? buflen : recv_event->ev.rlength;
md.threshold = 2; /* send and get */
md.options = PTL_MD_EVENT_START_DISABLE;
md.user_ptr = ptl_request;
md.eq_handle = ompi_mtl_portals.ptl_eq_h;
/* retain because it's unclear how many events we'll get here.
Some implementations give just the REPLY, others give SEND
and REPLY */
ret = PtlMDBind(ompi_mtl_portals.ptl_ni_h, md,
PTL_RETAIN, &md_h);
if (PTL_OK != ret) {
opal_output(fileno(stderr)," Error returned from ompi_mtl_datatype_recv_buf. Error code - %d \n",ret);
abort();
}
ptl_request->event_callback = ompi_mtl_portals_recv_progress;
ret = PtlGet(md_h,
recv_event->ev.initiator,
OMPI_MTL_PORTALS_READ_TABLE_ID,
0,
recv_event->ev.hdr_data,
0);
if (PTL_OK != ret) {
opal_output(fileno(stderr)," Error returned from PtlGet. Error code - %d \n",ret);
abort();
}
ptl_request->super.ompi_req->req_status.MPI_SOURCE =
PTL_GET_SOURCE(recv_event->ev.match_bits);
ptl_request->super.ompi_req->req_status.MPI_TAG =
PTL_GET_TAG(recv_event->ev.match_bits);
ptl_request->super.ompi_req->req_status.MPI_ERROR =
(recv_event->ev.rlength > buflen) ?
MPI_ERR_TRUNCATE : MPI_SUCCESS;
}
return OMPI_SUCCESS;
}
int
mca_btl_portals4_get(struct mca_btl_base_module_t* btl_base,
struct mca_btl_base_endpoint_t* btl_peer,
void *local_address,
uint64_t remote_address,
struct mca_btl_base_registration_handle_t *local_handle,
struct mca_btl_base_registration_handle_t *remote_handle,
size_t size,
int flags,
int order,
mca_btl_base_rdma_completion_fn_t cbfunc,
void *cbcontext,
void *cbdata)
{
mca_btl_portals4_module_t *portals4_btl = (mca_btl_portals4_module_t *) btl_base;
mca_btl_portals4_frag_t *frag = NULL;
ptl_md_t md;
int ret;
/* reserve space in the event queue for rdma operations immediately */
while (OPAL_THREAD_ADD32(&portals4_btl->portals_outstanding_ops, 1) >
portals4_btl->portals_max_outstanding_ops) {
OPAL_THREAD_ADD32(&portals4_btl->portals_outstanding_ops, -1);
OPAL_OUTPUT_VERBOSE((90, opal_btl_base_framework.framework_output, "Call to mca_btl_portals4_component_progress (1)\n"));
mca_btl_portals4_component_progress();
}
OPAL_BTL_PORTALS4_FRAG_ALLOC_USER(portals4_btl, frag);
if (NULL == frag){
OPAL_THREAD_ADD32(&portals4_btl->portals_outstanding_ops, -1);
return OPAL_ERROR;
}
OPAL_OUTPUT_VERBOSE((90, opal_btl_base_framework.framework_output,
"mca_btl_portals4_get: Incrementing portals_outstanding_ops=%d frag=%p",
portals4_btl->portals_outstanding_ops, (void *)frag));
frag->rdma_cb.func = cbfunc;
frag->rdma_cb.context = cbcontext;
frag->rdma_cb.data = cbdata;
frag->rdma_cb.local_handle = local_handle;
frag->endpoint = btl_peer;
frag->hdr.tag = MCA_BTL_TAG_MAX;
/* Bind the memory */
md.start = (void *)local_address;
md.length = size;
md.options = 0;
md.eq_handle = portals4_btl->recv_eq_h;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(portals4_btl->portals_ni_h,
&md,
&frag->md_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"%s:%d: PtlMDBind failed: %d",
__FILE__, __LINE__, ret);
return OPAL_ERROR;
}
frag->match_bits = remote_handle->key;
frag->length = md.length;
frag->peer_proc = btl_peer->ptl_proc;
OPAL_OUTPUT_VERBOSE((90, opal_btl_base_framework.framework_output, "PtlGet start=%p length=%ld nid=%x pid=%x match_bits=%lx\n",
md.start, md.length, btl_peer->ptl_proc.phys.nid, btl_peer->ptl_proc.phys.pid, frag->match_bits));
ret = PtlGet(frag->md_h,
0,
md.length,
btl_peer->ptl_proc,
portals4_btl->recv_idx,
frag->match_bits, /* match bits */
0,
frag);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"%s:%d: PtlGet failed: %d",
__FILE__, __LINE__, ret);
PtlMDRelease(frag->md_h);
frag->md_h = PTL_INVALID_HANDLE;
return OPAL_ERROR;
}
OPAL_OUTPUT_VERBOSE((90, opal_btl_base_framework.framework_output, "SUCCESS: PtlGet start=%p length=%ld nid=%x pid=%x match_bits=%lx\n",
md.start, md.length, btl_peer->ptl_proc.phys.nid, btl_peer->ptl_proc.phys.pid, frag->match_bits));
return OPAL_SUCCESS;
}
int main(int argc,
char *argv[])
{
ptl_handle_ni_t ni_logical;
ptl_pt_index_t logical_pt_index;
ptl_process_t myself;
struct timeval start, stop;
int potato = 0;
ENTRY_T potato_catcher;
HANDLE_T potato_catcher_handle;
ptl_md_t potato_launcher;
ptl_handle_md_t potato_launcher_handle;
int num_procs;
CHECK_RETURNVAL(PtlInit());
CHECK_RETURNVAL(libtest_init());
num_procs = libtest_get_size();
if (NULL != getenv("MAKELEVEL") && num_procs > 2) {
return 77;
}
CHECK_RETURNVAL(PtlNIInit
(PTL_IFACE_DEFAULT, NI_TYPE | PTL_NI_LOGICAL, PTL_PID_ANY,
NULL, NULL, &ni_logical));
CHECK_RETURNVAL(PtlSetMap(ni_logical, num_procs,
libtest_get_mapping(ni_logical)));
CHECK_RETURNVAL(PtlGetId(ni_logical, &myself));
CHECK_RETURNVAL(PtlPTAlloc
(ni_logical, 0, PTL_EQ_NONE, PTL_PT_ANY,
&logical_pt_index));
assert(logical_pt_index == 0);
/* Now do the initial setup on ni_logical */
potato_catcher.start = &potato;
potato_catcher.length = sizeof(potato);
potato_catcher.uid = PTL_UID_ANY;
potato_catcher.options = OPTIONS;
#if INTERFACE == 1
potato_catcher.match_id.rank = PTL_RANK_ANY;
potato_catcher.match_bits = 1;
potato_catcher.ignore_bits = ~potato_catcher.match_bits;
#endif
CHECK_RETURNVAL(PtlCTAlloc(ni_logical, &potato_catcher.ct_handle));
CHECK_RETURNVAL(APPEND
(ni_logical, logical_pt_index, &potato_catcher,
PTL_PRIORITY_LIST, NULL, &potato_catcher_handle));
/* Now do a barrier (on ni_physical) to make sure that everyone has their
* logical interface set up */
libtest_barrier();
/* now I can communicate between ranks with ni_logical */
/* set up the potato launcher */
potato_launcher.start = &potato;
potato_launcher.length = sizeof(potato);
potato_launcher.options =
PTL_MD_EVENT_CT_ACK | PTL_MD_EVENT_CT_SEND;
potato_launcher.eq_handle = PTL_EQ_NONE; // i.e. don't queue send events
CHECK_RETURNVAL(PtlCTAlloc(ni_logical, &potato_launcher.ct_handle));
CHECK_RETURNVAL(PtlMDBind
(ni_logical, &potato_launcher, &potato_launcher_handle));
/* rank 0 starts the potato going */
if (myself.rank == 0) {
ptl_process_t nextrank;
nextrank.rank = myself.rank + 1;
nextrank.rank *= (nextrank.rank <= num_procs - 1);
gettimeofday(&start, NULL);
CHECK_RETURNVAL(PtlPut(potato_launcher_handle, 0, potato_launcher.length,
(LOOPS == 1) ? PTL_OC_ACK_REQ : PTL_NO_ACK_REQ,
nextrank, logical_pt_index, 1, 0,
NULL, 1));
}
{ /* the potato-passing loop */
size_t waitfor;
ptl_ct_event_t ctc;
ptl_process_t nextrank;
nextrank.rank = myself.rank + 1;
nextrank.rank *= (nextrank.rank <= num_procs - 1);
for (waitfor = 1; waitfor <= LOOPS; ++waitfor) {
CHECK_RETURNVAL(PtlCTWait(potato_catcher.ct_handle, waitfor, &ctc)); // wait for potato
assert(ctc.failure == 0);
assert(ctc.success == waitfor);
/* I have the potato! */
++potato;
if (potato < LOOPS * (num_procs)) { // otherwise, the recipient may have exited
/* Bomb's away! */
if (myself.rank == 0) {
CHECK_RETURNVAL(PtlPut(potato_launcher_handle, 0,
potato_launcher.length,
(waitfor == (LOOPS - 1)) ? PTL_OC_ACK_REQ : PTL_NO_ACK_REQ,
nextrank, logical_pt_index, 3, 0, NULL, 2));
} else {
CHECK_RETURNVAL(PtlPut(potato_launcher_handle, 0,
potato_launcher.length,
(waitfor == LOOPS) ? PTL_OC_ACK_REQ : PTL_NO_ACK_REQ,
nextrank, logical_pt_index, 3, 0, NULL, 2));
}
}
}
// make sure that last send completed before exiting
CHECK_RETURNVAL(PtlCTWait(potato_launcher.ct_handle, LOOPS+1, &ctc));
assert(ctc.failure == 0);
}
if (myself.rank == 0) {
double accumulate = 0.0;
gettimeofday(&stop, NULL);
accumulate =
(stop.tv_sec + stop.tv_usec * 1e-6) - (start.tv_sec +
start.tv_usec * 1e-6);
/* calculate the average time waiting */
printf("Total time: %g secs\n", accumulate);
accumulate /= LOOPS;
printf("Average time around the loop: %g microseconds\n",
accumulate * 1e6);
accumulate /= num_procs;
printf("Average catch-to-toss latency: %g microseconds\n",
accumulate * 1e6);
}
/* cleanup */
CHECK_RETURNVAL(PtlMDRelease(potato_launcher_handle));
CHECK_RETURNVAL(PtlCTFree(potato_launcher.ct_handle));
CHECK_RETURNVAL(UNLINK(potato_catcher_handle));
CHECK_RETURNVAL(PtlCTFree(potato_catcher.ct_handle));
/* major cleanup */
CHECK_RETURNVAL(PtlPTFree(ni_logical, logical_pt_index));
CHECK_RETURNVAL(PtlNIFini(ni_logical));
CHECK_RETURNVAL(libtest_fini());
PtlFini();
return 0;
}
int
kptllnd_active_rdma(kptl_rx_t *rx, lnet_msg_t *lntmsg, int type,
unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
unsigned int offset, int nob)
{
kptl_tx_t *tx;
ptl_err_t ptlrc;
kptl_msg_t *rxmsg = rx->rx_msg;
kptl_peer_t *peer = rx->rx_peer;
unsigned long flags;
ptl_handle_md_t mdh;
LASSERT (type == TX_TYPE_PUT_RESPONSE ||
type == TX_TYPE_GET_RESPONSE);
tx = kptllnd_get_idle_tx(type);
if (tx == NULL) {
CERROR ("Can't do %s rdma to %s: can't allocate descriptor\n",
type == TX_TYPE_PUT_RESPONSE ? "GET" : "PUT",
libcfs_id2str(peer->peer_id));
return -ENOMEM;
}
kptllnd_set_tx_peer(tx, peer);
kptllnd_init_rdma_md(tx, niov, iov, kiov, offset, nob);
ptlrc = PtlMDBind(kptllnd_data.kptl_nih, tx->tx_rdma_md,
PTL_UNLINK, &mdh);
if (ptlrc != PTL_OK) {
CERROR("PtlMDBind(%s) failed: %s(%d)\n",
libcfs_id2str(peer->peer_id),
kptllnd_errtype2str(ptlrc), ptlrc);
tx->tx_status = -EIO;
kptllnd_tx_decref(tx);
return -EIO;
}
cfs_spin_lock_irqsave(&peer->peer_lock, flags);
tx->tx_lnet_msg = lntmsg;
/* lnet_finalize() will be called when tx is torn down, so I must
* return success from here on... */
tx->tx_deadline = jiffies + (*kptllnd_tunables.kptl_timeout * CFS_HZ);
tx->tx_rdma_mdh = mdh;
tx->tx_active = 1;
cfs_list_add_tail(&tx->tx_list, &peer->peer_activeq);
/* peer has now got my ref on 'tx' */
cfs_spin_unlock_irqrestore(&peer->peer_lock, flags);
tx->tx_tposted = jiffies;
if (type == TX_TYPE_GET_RESPONSE)
ptlrc = PtlPut(mdh,
tx->tx_acked ? PTL_ACK_REQ : PTL_NOACK_REQ,
rx->rx_initiator,
*kptllnd_tunables.kptl_portal,
0, /* acl cookie */
rxmsg->ptlm_u.rdma.kptlrm_matchbits,
0, /* offset */
(lntmsg != NULL) ? /* header data */
PTLLND_RDMA_OK :
PTLLND_RDMA_FAIL);
else
ptlrc = PtlGet(mdh,
rx->rx_initiator,
*kptllnd_tunables.kptl_portal,
0, /* acl cookie */
rxmsg->ptlm_u.rdma.kptlrm_matchbits,
0); /* offset */
if (ptlrc != PTL_OK) {
CERROR("Ptl%s failed: %s(%d)\n",
(type == TX_TYPE_GET_RESPONSE) ? "Put" : "Get",
kptllnd_errtype2str(ptlrc), ptlrc);
kptllnd_peer_close(peer, -EIO);
/* Everything (including this RDMA) queued on the peer will
* be completed with failure */
kptllnd_schedule_ptltrace_dump();
}
return 0;
}
static int send_msg(ptl_hdr_data_t ssend_flag, struct MPIDI_VC *vc, const void *buf, MPI_Aint count, MPI_Datatype datatype, int dest,
int tag, MPID_Comm *comm, int context_offset, struct MPID_Request **request)
{
int mpi_errno = MPI_SUCCESS;
MPID_nem_ptl_vc_area *const vc_ptl = VC_PTL(vc);
int ret;
MPIDI_msg_sz_t data_sz;
int dt_contig;
MPI_Aint dt_true_lb;
MPID_Datatype *dt_ptr;
MPID_Request *sreq = NULL;
ptl_me_t me;
int initial_iov_count, remaining_iov_count;
ptl_md_t md;
MPI_Aint last;
MPIU_CHKPMEM_DECL(2);
MPIDI_STATE_DECL(MPID_STATE_SEND_MSG);
MPIDI_FUNC_ENTER(MPID_STATE_SEND_MSG);
MPID_nem_ptl_request_create_sreq(sreq, mpi_errno, comm);
sreq->dev.match.parts.rank = dest;
sreq->dev.match.parts.tag = tag;
sreq->dev.match.parts.context_id = comm->context_id + context_offset;
sreq->ch.vc = vc;
if (!vc_ptl->id_initialized) {
mpi_errno = MPID_nem_ptl_init_id(vc);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
MPIDI_Datatype_get_info(count, datatype, dt_contig, data_sz, dt_ptr, dt_true_lb);
MPIU_DBG_MSG_FMT(CH3_CHANNEL, VERBOSE, (MPIU_DBG_FDEST, "count="MPI_AINT_FMT_DEC_SPEC" datatype=%#x contig=%d data_sz=%lu", count, datatype, dt_contig, data_sz));
if (data_sz <= PTL_LARGE_THRESHOLD) {
/* Small message. Send all data eagerly */
if (dt_contig) {
void *start = (char *)buf + dt_true_lb;
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, "Small contig message");
REQ_PTL(sreq)->event_handler = handler_send;
MPIU_DBG_MSG_P(CH3_CHANNEL, VERBOSE, "&REQ_PTL(sreq)->event_handler = %p", &(REQ_PTL(sreq)->event_handler));
if (start == NULL)
ret = MPID_nem_ptl_rptl_put(MPIDI_nem_ptl_global_md, (ptl_size_t)&dummy, data_sz, PTL_NO_ACK_REQ, vc_ptl->id, vc_ptl->pt,
NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), 0, sreq,
NPTL_HEADER(ssend_flag, data_sz));
else
ret = MPID_nem_ptl_rptl_put(MPIDI_nem_ptl_global_md, (ptl_size_t)start, data_sz, PTL_NO_ACK_REQ, vc_ptl->id, vc_ptl->pt,
NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), 0, sreq,
NPTL_HEADER(ssend_flag, data_sz));
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlput", "**ptlput %s", MPID_nem_ptl_strerror(ret));
DBG_MSG_PUT("global", data_sz, vc->pg_rank, NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), NPTL_HEADER(ssend_flag, data_sz));
MPIU_DBG_MSG_D(CH3_CHANNEL, VERBOSE, "id.nid = %#x", vc_ptl->id.phys.nid);
MPIU_DBG_MSG_D(CH3_CHANNEL, VERBOSE, "id.pid = %#x", vc_ptl->id.phys.pid);
MPIU_DBG_MSG_P(CH3_CHANNEL, VERBOSE, "sreq = %p", sreq);
MPIU_DBG_MSG_D(CH3_CHANNEL, VERBOSE, "vc_ptl->pt = %d", vc_ptl->pt);
MPIU_DBG_MSG_P(CH3_CHANNEL, VERBOSE, "REQ_PTL(sreq)->event_handler = %p", REQ_PTL(sreq)->event_handler);
goto fn_exit;
}
/* noncontig data */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, "Small noncontig message");
sreq->dev.segment_ptr = MPID_Segment_alloc();
MPIR_ERR_CHKANDJUMP1(sreq->dev.segment_ptr == NULL, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment_alloc");
MPID_Segment_init(buf, count, datatype, sreq->dev.segment_ptr, 0);
sreq->dev.segment_first = 0;
sreq->dev.segment_size = data_sz;
last = sreq->dev.segment_size;
sreq->dev.iov_count = MPL_IOV_LIMIT;
MPID_Segment_pack_vector(sreq->dev.segment_ptr, sreq->dev.segment_first, &last, sreq->dev.iov, &sreq->dev.iov_count);
if (last == sreq->dev.segment_size) {
/* IOV is able to describe entire message */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, " entire message fits in IOV");
md.start = sreq->dev.iov;
md.length = sreq->dev.iov_count;
md.options = PTL_IOVEC;
md.eq_handle = MPIDI_nem_ptl_origin_eq;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(MPIDI_nem_ptl_ni, &md, &REQ_PTL(sreq)->md);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmdbind", "**ptlmdbind %s", MPID_nem_ptl_strerror(ret));
REQ_PTL(sreq)->event_handler = handler_send;
ret = MPID_nem_ptl_rptl_put(REQ_PTL(sreq)->md, 0, data_sz, PTL_NO_ACK_REQ, vc_ptl->id, vc_ptl->pt,
NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), 0, sreq,
NPTL_HEADER(ssend_flag, data_sz));
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlput", "**ptlput %s", MPID_nem_ptl_strerror(ret));
DBG_MSG_PUT("sreq", data_sz, vc->pg_rank, NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), NPTL_HEADER(ssend_flag, data_sz));
goto fn_exit;
}
/* IOV is not long enough to describe entire message */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, " IOV too long: using bounce buffer");
MPIU_CHKPMEM_MALLOC(REQ_PTL(sreq)->chunk_buffer[0], void *, data_sz, mpi_errno, "chunk_buffer");
MPID_Segment_init(buf, count, datatype, sreq->dev.segment_ptr, 0);
sreq->dev.segment_first = 0;
last = data_sz;
MPID_Segment_pack(sreq->dev.segment_ptr, sreq->dev.segment_first, &last, REQ_PTL(sreq)->chunk_buffer[0]);
MPIU_Assert(last == sreq->dev.segment_size);
REQ_PTL(sreq)->event_handler = handler_send;
ret = MPID_nem_ptl_rptl_put(MPIDI_nem_ptl_global_md, (ptl_size_t)REQ_PTL(sreq)->chunk_buffer[0], data_sz, PTL_NO_ACK_REQ,
vc_ptl->id, vc_ptl->pt, NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), 0, sreq,
NPTL_HEADER(ssend_flag, data_sz));
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlput", "**ptlput %s", MPID_nem_ptl_strerror(ret));
DBG_MSG_PUT("global", data_sz, vc->pg_rank, NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), NPTL_HEADER(ssend_flag, data_sz));
goto fn_exit;
}
/* Large message. Send first chunk of data and let receiver get the rest */
if (dt_contig) {
/* create ME for buffer so receiver can issue a GET for the data */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, "Large contig message");
big_meappend((char *)buf + dt_true_lb + PTL_LARGE_THRESHOLD, data_sz - PTL_LARGE_THRESHOLD, vc,
NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), sreq);
REQ_PTL(sreq)->event_handler = handler_send;
ret = MPID_nem_ptl_rptl_put(MPIDI_nem_ptl_global_md, (ptl_size_t)((char *)buf + dt_true_lb), PTL_LARGE_THRESHOLD, PTL_NO_ACK_REQ, vc_ptl->id, vc_ptl->pt,
NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), 0, sreq,
NPTL_HEADER(ssend_flag | NPTL_LARGE, data_sz));
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlput", "**ptlput %s", MPID_nem_ptl_strerror(ret));
DBG_MSG_PUT("global", PTL_LARGE_THRESHOLD, vc->pg_rank, NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), NPTL_HEADER(ssend_flag | NPTL_LARGE, data_sz));
goto fn_exit;
}
/* Large noncontig data */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, "Large noncontig message");
sreq->dev.segment_ptr = MPID_Segment_alloc();
MPIR_ERR_CHKANDJUMP1(sreq->dev.segment_ptr == NULL, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment_alloc");
MPID_Segment_init(buf, count, datatype, sreq->dev.segment_ptr, 0);
sreq->dev.segment_first = 0;
sreq->dev.segment_size = data_sz;
last = PTL_LARGE_THRESHOLD;
sreq->dev.iov_count = MPL_IOV_LIMIT;
MPID_Segment_pack_vector(sreq->dev.segment_ptr, sreq->dev.segment_first, &last, sreq->dev.iov, &sreq->dev.iov_count);
initial_iov_count = sreq->dev.iov_count;
sreq->dev.segment_first = last;
if (last == PTL_LARGE_THRESHOLD) {
/* first chunk of message fits into IOV */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, " first chunk fits in IOV");
if (initial_iov_count < MPL_IOV_LIMIT) {
/* There may be space for the rest of the message in this IOV */
sreq->dev.iov_count = MPL_IOV_LIMIT - sreq->dev.iov_count;
last = sreq->dev.segment_size;
MPID_Segment_pack_vector(sreq->dev.segment_ptr, sreq->dev.segment_first, &last,
&sreq->dev.iov[initial_iov_count], &sreq->dev.iov_count);
remaining_iov_count = sreq->dev.iov_count;
if (last == sreq->dev.segment_size && last <= MPIDI_nem_ptl_ni_limits.max_msg_size + PTL_LARGE_THRESHOLD) {
/* Entire message fit in one IOV */
int was_incomplete;
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, " rest of message fits in one IOV");
/* Create ME for remaining data */
me.start = &sreq->dev.iov[initial_iov_count];
me.length = remaining_iov_count;
me.ct_handle = PTL_CT_NONE;
me.uid = PTL_UID_ANY;
me.options = ( PTL_ME_OP_PUT | PTL_ME_OP_GET | PTL_ME_USE_ONCE | PTL_ME_IS_ACCESSIBLE | PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE | PTL_IOVEC );
me.match_id = vc_ptl->id;
me.match_bits = NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank);
me.ignore_bits = 0;
me.min_free = 0;
MPIU_CHKPMEM_MALLOC(REQ_PTL(sreq)->get_me_p, ptl_handle_me_t *, sizeof(ptl_handle_me_t), mpi_errno, "get_me_p");
ret = PtlMEAppend(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_get_pt, &me, PTL_PRIORITY_LIST, sreq,
&REQ_PTL(sreq)->get_me_p[0]);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmeappend", "**ptlmeappend %s", MPID_nem_ptl_strerror(ret));
DBG_MSG_MEAPPEND("CTL", vc->pg_rank, me, sreq);
/* increment the cc for the get operation */
MPIDI_CH3U_Request_increment_cc(sreq, &was_incomplete);
MPIU_Assert(was_incomplete);
/* Create MD for first chunk */
md.start = sreq->dev.iov;
md.length = initial_iov_count;
md.options = PTL_IOVEC;
md.eq_handle = MPIDI_nem_ptl_origin_eq;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(MPIDI_nem_ptl_ni, &md, &REQ_PTL(sreq)->md);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmdbind", "**ptlmdbind %s", MPID_nem_ptl_strerror(ret));
REQ_PTL(sreq)->event_handler = handler_send;
ret = MPID_nem_ptl_rptl_put(REQ_PTL(sreq)->md, 0, PTL_LARGE_THRESHOLD, PTL_NO_ACK_REQ, vc_ptl->id, vc_ptl->pt,
NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), 0, sreq,
NPTL_HEADER(ssend_flag | NPTL_LARGE, data_sz));
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlput", "**ptlput %s", MPID_nem_ptl_strerror(ret));
DBG_MSG_PUT("req", PTL_LARGE_THRESHOLD, vc->pg_rank, NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), NPTL_HEADER(ssend_flag | NPTL_LARGE, data_sz));
goto fn_exit;
}
int
shmem_transport_startup(void)
{
int ret, i;
ptl_process_t *desired = NULL;
ptl_md_t md;
ptl_le_t le;
ptl_uid_t uid = PTL_UID_ANY;
ptl_process_t my_id;
#ifdef USE_ON_NODE_COMMS
int num_on_node = 0;
#endif
#ifdef ENABLE_REMOTE_VIRTUAL_ADDRESSING
/* Make sure the heap and data bases are actually symmetric */
{
int peer;
uint64_t bases[2];
peer = (shmem_internal_my_pe + 1) % shmem_internal_num_pes;
ret = shmem_runtime_get(peer, "portals4-bases", bases, sizeof(uint64_t) * 2);
if (0 != ret) {
fprintf(stderr, "[%03d] ERROR: runtime_put failed: %d\n",
shmem_internal_my_pe, ret);
return ret;
}
if ((uintptr_t) shmem_internal_heap_base != bases[0]) {
fprintf(stderr, "[%03d] ERROR: heap base address does not match with rank %03d and virtual addressing is enabled\n",
shmem_internal_my_pe, peer);
return -1;
}
if ((uintptr_t) shmem_internal_data_base != bases[1]) {
fprintf(stderr, "[%03d] ERROR: data base address does not match with rank %03d and virtual addressing is enabled\n",
shmem_internal_my_pe, peer);
return -1;
}
}
#endif
desired = malloc(sizeof(ptl_process_t) * shmem_internal_num_pes);
if (NULL == desired) {
ret = 1;
goto cleanup;
}
ret = PtlGetPhysId(shmem_transport_portals4_ni_h, &my_id);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlGetPhysId failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
for (i = 0 ; i < shmem_internal_num_pes; ++i) {
ret = shmem_runtime_get(i, "portals4-procid",
&desired[i], sizeof(ptl_process_t));
if (0 != ret) {
fprintf(stderr, "[%03d] ERROR: runtime_get failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
#ifdef USE_ON_NODE_COMMS
/* update the connectivity map... */
if (desired[i].phys.nid == my_id.phys.nid) {
SHMEM_SET_RANK_SAME_NODE(i, num_on_node++);
if (num_on_node > 255) {
fprintf(stderr, "[%03d] ERROR: Too many local ranks.\n",
shmem_internal_my_pe);
goto cleanup;
}
}
#endif
}
ret = PtlSetMap(shmem_transport_portals4_ni_h,
shmem_internal_num_pes,
desired);
if (PTL_OK != ret && PTL_IGNORED != ret) {
fprintf(stderr, "[%03d] ERROR: PtlSetMap failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
ret = PtlGetUid(shmem_transport_portals4_ni_h, &uid);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlGetUid failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
shmem_transport_portals4_max_volatile_size = ni_limits.max_volatile_size;
shmem_transport_portals4_max_atomic_size = ni_limits.max_atomic_size;
shmem_transport_portals4_max_fetch_atomic_size = ni_limits.max_fetch_atomic_size;
shmem_transport_portals4_max_msg_size = ni_limits.max_msg_size;
if (shmem_transport_portals4_max_volatile_size < sizeof(long double complex)) {
fprintf(stderr, "[%03d] ERROR: Max volatile size found to be %lu, too small to continue\n",
shmem_internal_my_pe, (unsigned long) shmem_transport_portals4_max_volatile_size);
goto cleanup;
}
if (shmem_transport_portals4_max_atomic_size < sizeof(long double complex)) {
fprintf(stderr, "[%03d] ERROR: Max atomic size found to be %lu, too small to continue\n",
shmem_internal_my_pe, (unsigned long) shmem_transport_portals4_max_atomic_size);
goto cleanup;
}
if (shmem_transport_portals4_max_fetch_atomic_size < sizeof(long double complex)) {
fprintf(stderr, "[%03d] ERROR: Max fetch atomic size found to be %lu, too small to continue\n",
shmem_internal_my_pe, (unsigned long) shmem_transport_portals4_max_fetch_atomic_size);
goto cleanup;
}
/* create portal table entries */
ret = PtlEQAlloc(shmem_transport_portals4_ni_h,
shmem_transport_portals4_event_slots,
&shmem_transport_portals4_eq_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlEQAlloc failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
#ifdef ENABLE_REMOTE_VIRTUAL_ADDRESSING
ret = PtlPTAlloc(shmem_transport_portals4_ni_h,
0,
shmem_transport_portals4_eq_h,
shmem_transport_portals4_pt,
&all_pt);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlPTAlloc of table entry failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
#else
ret = PtlPTAlloc(shmem_transport_portals4_ni_h,
0,
shmem_transport_portals4_eq_h,
shmem_transport_portals4_data_pt,
&data_pt);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlPTAlloc of data table failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
ret = PtlPTAlloc(shmem_transport_portals4_ni_h,
0,
shmem_transport_portals4_eq_h,
shmem_transport_portals4_heap_pt,
&heap_pt);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlPTAlloc of heap table failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
#endif
#ifndef ENABLE_HARD_POLLING
/* target ct */
ret = PtlCTAlloc(shmem_transport_portals4_ni_h, &shmem_transport_portals4_target_ct_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlCTAlloc of target ct failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
le.ct_handle = shmem_transport_portals4_target_ct_h;
#endif
le.uid = uid;
le.options = PTL_LE_OP_PUT | PTL_LE_OP_GET |
PTL_LE_EVENT_LINK_DISABLE |
PTL_LE_EVENT_SUCCESS_DISABLE;
#if !defined(ENABLE_HARD_POLLING)
le.options |= PTL_LE_EVENT_CT_COMM;
#endif
#ifdef ENABLE_REMOTE_VIRTUAL_ADDRESSING
le.start = NULL;
le.length = PTL_SIZE_MAX;
ret = PtlLEAppend(shmem_transport_portals4_ni_h,
shmem_transport_portals4_pt,
&le,
PTL_PRIORITY_LIST,
NULL,
&shmem_transport_portals4_le_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlLEAppend of all memory failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
#else
/* Open LE to heap section */
le.start = shmem_internal_heap_base;
le.length = shmem_internal_heap_length;
ret = PtlLEAppend(shmem_transport_portals4_ni_h,
shmem_transport_portals4_heap_pt,
&le,
PTL_PRIORITY_LIST,
NULL,
&shmem_transport_portals4_heap_le_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlLEAppend of heap section failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
/* Open LE to data section */
le.start = shmem_internal_data_base;
le.length = shmem_internal_data_length;
ret = PtlLEAppend(shmem_transport_portals4_ni_h,
shmem_transport_portals4_data_pt,
&le,
PTL_PRIORITY_LIST,
NULL,
&shmem_transport_portals4_data_le_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlLEAppend of data section failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
#endif
/* Open MD to all memory */
ret = PtlCTAlloc(shmem_transport_portals4_ni_h, &shmem_transport_portals4_put_ct_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlCTAlloc of put ct failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
ret = PtlCTAlloc(shmem_transport_portals4_ni_h, &shmem_transport_portals4_get_ct_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlCTAlloc of get ct failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = PTL_MD_EVENT_CT_ACK;
if (1 == PORTALS4_TOTAL_DATA_ORDERING) {
md.options |= PTL_MD_UNORDERED;
}
md.eq_handle = shmem_transport_portals4_eq_h;
md.ct_handle = shmem_transport_portals4_put_ct_h;
ret = PtlMDBind(shmem_transport_portals4_ni_h,
&md,
&shmem_transport_portals4_put_event_md_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlMDBind of put MD failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = PTL_MD_EVENT_CT_ACK |
PTL_MD_EVENT_SUCCESS_DISABLE |
PTL_MD_VOLATILE;
if (1 == PORTALS4_TOTAL_DATA_ORDERING) {
md.options |= PTL_MD_UNORDERED;
}
md.eq_handle = shmem_transport_portals4_eq_h;
md.ct_handle = shmem_transport_portals4_put_ct_h;
ret = PtlMDBind(shmem_transport_portals4_ni_h,
&md,
&shmem_transport_portals4_put_volatile_md_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlMDBind of put MD failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = PTL_MD_EVENT_CT_ACK |
PTL_MD_EVENT_SUCCESS_DISABLE;
if (1 == PORTALS4_TOTAL_DATA_ORDERING) {
md.options |= PTL_MD_UNORDERED;
}
md.eq_handle = shmem_transport_portals4_eq_h;
md.ct_handle = shmem_transport_portals4_put_ct_h;
ret = PtlMDBind(shmem_transport_portals4_ni_h,
&md,
&shmem_transport_portals4_put_cntr_md_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlMDBind of put cntr MD failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = PTL_MD_EVENT_CT_REPLY |
PTL_MD_EVENT_SUCCESS_DISABLE;
if (1 == PORTALS4_TOTAL_DATA_ORDERING) {
md.options |= PTL_MD_UNORDERED;
}
md.eq_handle = shmem_transport_portals4_eq_h;
md.ct_handle = shmem_transport_portals4_get_ct_h;
ret = PtlMDBind(shmem_transport_portals4_ni_h,
&md,
&shmem_transport_portals4_get_md_h);
if (PTL_OK != ret) {
fprintf(stderr, "[%03d] ERROR: PtlMDBind of get MD failed: %d\n",
shmem_internal_my_pe, ret);
goto cleanup;
}
ret = 0;
cleanup:
if (NULL != desired) free(desired);
return ret;
}
void test_prepostME(int cache_size,
int *cache_buf,
ptl_handle_ni_t ni,
int npeers,
int nmsgs,
int nbytes,
int niters)
{
int i, j, k;
double tmp, total = 0;
ptl_handle_md_t send_md_handle;
ptl_md_t send_md;
ptl_process_t dest;
ptl_size_t offset;
ptl_pt_index_t index;
ptl_handle_eq_t recv_eq_handle;
ptl_handle_me_t me_handles[npeers * nmsgs];
ptl_event_t event;
ptl_assert(PtlEQAlloc(ni, nmsgs * npeers + 1,
&send_md.eq_handle), PTL_OK);
send_md.start = send_buf;
send_md.length = SEND_BUF_SIZE;
send_md.options = PTL_MD_UNORDERED;
send_md.ct_handle = PTL_CT_NONE;
ptl_assert(PtlMDBind(ni, &send_md, &send_md_handle), PTL_OK);
ptl_assert(PtlEQAlloc(ni, nmsgs * npeers + 1, &recv_eq_handle), PTL_OK);
ptl_assert(PtlPTAlloc(ni, 0, recv_eq_handle, TestSameDirectionIndex,
&index), PTL_OK);
ptl_assert(TestSameDirectionIndex, index);
tmp = timer();
for (j = 0; j < npeers; ++j) {
for (k = 0; k < nmsgs; ++k) {
ptl_process_t src;
src.rank = recv_peers[j];
postME(ni, index, recv_buf + (nbytes * (k + j * nmsgs)),
nbytes, src, magic_tag, &me_handles[k + j * nmsgs]);
}
}
total += (timer() - tmp);
for (i = 0; i < niters - 1; ++i) {
cache_invalidate(cache_size, cache_buf);
libtest_Barrier();
tmp = timer();
for (j = 0; j < npeers; ++j) {
for (k = 0; k < nmsgs; ++k) {
offset = (nbytes * (k + j * nmsgs));
dest.rank = send_peers[npeers - j - 1],
ptl_assert(libtest_Put_offset(send_md_handle, offset, nbytes,
dest, index, magic_tag, offset), PTL_OK);
}
}
/* wait for sends */
for (j = 0; j < npeers * nmsgs; ++j) {
ptl_assert(PtlEQWait(send_md.eq_handle, &event), PTL_OK);
ptl_assert(event.type, PTL_EVENT_SEND);
}
/* wait for receives */
for (j = 0; j < npeers * nmsgs; j++) {
PtlEQWait(recv_eq_handle, &event);
}
for (j = 0; j < npeers; ++j) {
for (k = 0; k < nmsgs; ++k) {
ptl_process_t src;
src.rank = recv_peers[j];
postME(ni, index, recv_buf + (nbytes * (k + j * nmsgs)),
nbytes, src, magic_tag, &me_handles[k + j * nmsgs]);
}
}
total += (timer() - tmp);
}
libtest_Barrier();
tmp = timer();
for (j = 0; j < npeers; ++j) {
for (k = 0; k < nmsgs; ++k) {
offset = (nbytes * (k + j * nmsgs));
dest.rank = send_peers[npeers - j - 1],
ptl_assert(libtest_Put_offset(send_md_handle, offset, nbytes, dest,
index, magic_tag, offset), PTL_OK);
}
}
/* wait for sends */
for (j = 0; j < npeers * nmsgs; ++j) {
ptl_assert(PtlEQWait(send_md.eq_handle, &event), PTL_OK);
ptl_assert(event.type, PTL_EVENT_SEND);
}
/* wait for receives */
for (j = 0; j < npeers * nmsgs; j++) {
PtlEQWait(recv_eq_handle, &event);
}
total += (timer() - tmp);
ptl_assert(PtlEQFree(send_md.eq_handle), PTL_OK);
ptl_assert(PtlMDRelease(send_md_handle), PTL_OK);
ptl_assert(PtlEQFree(recv_eq_handle), PTL_OK);
ptl_assert(PtlPTFree(ni, index), PTL_OK);
tmp = libtest_AllreduceDouble(total, PTL_SUM);
display_result("pre-post", (niters * npeers * nmsgs * 2) / (tmp / world_size));
}
int main(int argc,
char *argv[])
{
ptl_handle_ni_t ni_h;
ptl_pt_index_t pt_index;
uint64_t *buf;
ENTRY_T entry;
HANDLE_T entry_h;
ptl_md_t md;
ptl_handle_md_t md_h;
int rank;
int num_procs;
int ret;
ptl_process_t *procs;
ptl_handle_eq_t eq_h;
ptl_event_t ev;
ptl_hdr_data_t rcvd = 0;
ptl_hdr_data_t goal = 0;
ptl_hdr_data_t hdr_data = 1;
ptl_size_t offset = sizeof(uint64_t);
uint32_t distance;
int sends = 0;
CHECK_RETURNVAL(PtlInit());
CHECK_RETURNVAL(libtest_init());
rank = libtest_get_rank();
num_procs = libtest_get_size();
/* This test only succeeds if we have more than one rank */
if (num_procs < 2) return 77;
CHECK_RETURNVAL(PtlNIInit(PTL_IFACE_DEFAULT, NI_TYPE | PTL_NI_LOGICAL,
PTL_PID_ANY, NULL, NULL, &ni_h));
procs = libtest_get_mapping(ni_h);
CHECK_RETURNVAL(PtlSetMap(ni_h, num_procs, procs));
CHECK_RETURNVAL(PtlEQAlloc(ni_h, 1024, &eq_h));
CHECK_RETURNVAL(PtlPTAlloc(ni_h, 0, eq_h, 0, &pt_index));
assert(pt_index == 0);
buf = malloc(sizeof(uint64_t) * num_procs);
assert(NULL != buf);
md.start = buf;
md.length = sizeof(uint64_t) * num_procs;
md.options = PTL_MD_UNORDERED;
md.eq_handle = eq_h;
md.ct_handle = PTL_CT_NONE;
CHECK_RETURNVAL(PtlMDBind(ni_h, &md, &md_h));
entry.start = buf;
entry.length = sizeof(uint64_t) * num_procs;
entry.ct_handle = PTL_CT_NONE;
entry.uid = PTL_UID_ANY;
entry.options = OPTIONS;
#if MATCHING == 1
entry.match_id.rank = PTL_RANK_ANY;
entry.match_bits = 0;
entry.ignore_bits = 0;
entry.min_free = 0;
#endif
CHECK_RETURNVAL(APPEND(ni_h, pt_index, &entry,
PTL_PRIORITY_LIST, NULL, &entry_h));
/* ensure ME is linked before the barrier */
CHECK_RETURNVAL(PtlEQWait(eq_h, &ev));
assert( ev.type == PTL_EVENT_LINK );
libtest_barrier();
/* Bruck's Concatenation Algorithm */
memcpy(buf, &rank, sizeof(uint64_t));
for (distance = 1; distance < num_procs; distance *= 2) {
ptl_size_t to_xfer;
int peer;
ptl_process_t proc;
if (rank >= distance) {
peer = rank - distance;
} else {
peer = rank + (num_procs - distance);
}
to_xfer = sizeof(uint64_t) * MIN(distance, num_procs - distance);
proc.rank = peer;
CHECK_RETURNVAL(PtlPut(md_h,
0,
to_xfer,
PTL_NO_ACK_REQ,
proc,
0,
0,
offset,
NULL,
hdr_data));
sends += 1;
/* wait for completion of the proper receive, and keep count
of uncompleted sends. "rcvd" is an accumulator to deal
with out-of-order receives, which are IDed by the
hdr_data */
goal |= hdr_data;
while ((rcvd & goal) != goal) {
ret = PtlEQWait(eq_h, &ev);
switch (ret) {
case PTL_OK:
if (ev.type == PTL_EVENT_SEND) {
sends -= 1;
} else {
rcvd |= ev.hdr_data;
assert(ev.type == PTL_EVENT_PUT);
assert(ev.rlength == ev.mlength);
assert((ev.rlength == to_xfer) || (ev.hdr_data != hdr_data));
}
break;
default:
fprintf(stderr, "PtlEQWait failure: %d\n", ret);
abort();
}
}
hdr_data <<= 1;
offset += to_xfer;
}
/* wait for any SEND_END events not yet seen */
while (sends) {
ret = PtlEQWait(eq_h, &ev);
switch (ret) {
case PTL_OK:
assert( ev.type == PTL_EVENT_SEND );
sends -= 1;
break;
default:
fprintf(stderr, "PtlEQWait failure: %d\n", ret);
abort();
}
}
CHECK_RETURNVAL(UNLINK(entry_h));
CHECK_RETURNVAL(PtlMDRelease(md_h));
free(buf);
libtest_barrier();
/* cleanup */
CHECK_RETURNVAL(PtlPTFree(ni_h, pt_index));
CHECK_RETURNVAL(PtlEQFree(eq_h));
CHECK_RETURNVAL(PtlNIFini(ni_h));
CHECK_RETURNVAL(libtest_fini());
PtlFini();
return 0;
}
int
ompi_mtl_portals_isend(struct mca_mtl_base_module_t* mtl,
struct ompi_communicator_t* comm,
int dest,
int tag,
struct ompi_convertor_t *convertor,
mca_pml_base_send_mode_t mode,
bool blocking,
mca_mtl_request_t *mtl_request)
{
int ret;
ptl_match_bits_t match_bits;
ptl_md_t md;
ptl_handle_md_t md_h;
ptl_handle_me_t me_h;
ompi_proc_t* ompi_proc = ompi_comm_peer_lookup( comm, dest );
mca_mtl_base_endpoint_t *endpoint = (mca_mtl_base_endpoint_t*) ompi_proc->proc_pml;
ompi_mtl_portals_request_t *ptl_request =
(ompi_mtl_portals_request_t*) mtl_request;
size_t buflen;
assert(mtl == &ompi_mtl_portals.base);
ret = ompi_mtl_datatype_pack(convertor, &md.start, &buflen,
&(ptl_request->free_after));
if (OMPI_SUCCESS != ret) return ret;
md.length = buflen;
ptl_request->event_callback = ompi_mtl_portals_send_progress;
if ((MCA_PML_BASE_SEND_READY == mode)) {
/* ready send (length doesn't matter) or short non-sync send.
Eagerly send data and don't wait for completion */
PTL_SET_SEND_BITS(match_bits, comm->c_contextid,
comm->c_my_rank,
tag, PTL_READY_MSG);
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_output,
"ready send bits: 0x%016llx\n",
match_bits));
md.threshold = 1;
md.options = PTL_MD_EVENT_START_DISABLE;
md.user_ptr = ptl_request;
md.eq_handle = ompi_mtl_portals.ptl_eq_h;
ret = PtlMDBind(ompi_mtl_portals.ptl_ni_h,
md,
PTL_UNLINK,
&(md_h));
if (OMPI_SUCCESS != ret) {
if (ptl_request->free_after) free(md.start);
return ompi_common_portals_error_ptl_to_ompi(ret);
}
ptl_request->event_callback = ompi_mtl_portals_send_progress_no_ack;
ret = PtlPut(md_h,
PTL_NO_ACK_REQ,
endpoint->ptl_proc,
OMPI_MTL_PORTALS_SEND_TABLE_ID,
0,
match_bits,
0,
0);
if (OMPI_SUCCESS != ret) {
PtlMDUnlink(md_h);
if (ptl_request->free_after) free(md.start);
return ompi_common_portals_error_ptl_to_ompi(ret);
}
} else if (md.length > ompi_mtl_portals.eager_limit) {
/* it's a long message - same protocol for all send modes
other than ready */
PTL_SET_SEND_BITS(match_bits, comm->c_contextid,
comm->c_my_rank,
tag, PTL_LONG_MSG);
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_output,
"long send bits: 0x%016llx (%d)\n",
match_bits, dest));
md.threshold = 2; /* send, {ack, get} */
md.options = PTL_MD_OP_GET | PTL_MD_EVENT_START_DISABLE;
md.user_ptr = ptl_request;
md.eq_handle = ompi_mtl_portals.ptl_eq_h;
ret = PtlMEAttach(ompi_mtl_portals.ptl_ni_h,
OMPI_MTL_PORTALS_READ_TABLE_ID,
endpoint->ptl_proc,
(ptl_match_bits_t)(uintptr_t) ptl_request,
0,
PTL_UNLINK,
PTL_INS_AFTER,
&me_h);
if (OMPI_SUCCESS != ret) {
if (ptl_request->free_after) free(md.start);
return ompi_common_portals_error_ptl_to_ompi(ret);
}
ret = PtlMDAttach(me_h,
md,
PTL_UNLINK,
&(md_h));
if (OMPI_SUCCESS != ret) {
PtlMEUnlink(me_h);
if (ptl_request->free_after) free(md.start);
return ompi_common_portals_error_ptl_to_ompi(ret);
}
ret = PtlPut(md_h,
PTL_ACK_REQ,
endpoint->ptl_proc,
OMPI_MTL_PORTALS_SEND_TABLE_ID,
0,
match_bits,
0,
(ptl_hdr_data_t)(uintptr_t) ptl_request);
if (OMPI_SUCCESS != ret) {
PtlMDUnlink(md_h);
if (ptl_request->free_after) free(md.start);
return ompi_common_portals_error_ptl_to_ompi(ret);
}
} else if (MCA_PML_BASE_SEND_SYNCHRONOUS == mode) {
/* short synchronous message */
PTL_SET_SEND_BITS(match_bits, comm->c_contextid,
comm->c_my_rank,
tag, PTL_SHORT_MSG);
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_output,
"short ssend bits: 0x%016llx (%d)\n",
match_bits, dest));
md.threshold = 2; /* send, {ack, put} */
md.options = PTL_MD_OP_PUT | PTL_MD_EVENT_START_DISABLE;
md.user_ptr = ptl_request;
md.eq_handle = ompi_mtl_portals.ptl_eq_h;
ret = PtlMEAttach(ompi_mtl_portals.ptl_ni_h,
OMPI_MTL_PORTALS_ACK_TABLE_ID,
endpoint->ptl_proc,
(ptl_match_bits_t)(uintptr_t) ptl_request,
0,
PTL_UNLINK,
PTL_INS_AFTER,
&me_h);
if (OMPI_SUCCESS != ret) {
if (ptl_request->free_after) free(md.start);
return ompi_common_portals_error_ptl_to_ompi(ret);
}
ret = PtlMDAttach(me_h,
md,
PTL_UNLINK,
&(md_h));
if (OMPI_SUCCESS != ret) {
PtlMEUnlink(me_h);
if (ptl_request->free_after) free(md.start);
return ompi_common_portals_error_ptl_to_ompi(ret);
}
ret = PtlPut(md_h,
PTL_ACK_REQ,
endpoint->ptl_proc,
OMPI_MTL_PORTALS_SEND_TABLE_ID,
0,
match_bits,
0,
(ptl_hdr_data_t)(uintptr_t) ptl_request);
if (OMPI_SUCCESS != ret) {
PtlMDUnlink(md_h);
if (ptl_request->free_after) free(md.start);
return ompi_common_portals_error_ptl_to_ompi(ret);
}
} else {
/* short send message */
PTL_SET_SEND_BITS(match_bits, comm->c_contextid,
comm->c_my_rank,
tag, PTL_SHORT_MSG);
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_output,
"short send bits: 0x%016llx\n",
match_bits));
md.threshold = 1;
md.options = PTL_MD_EVENT_START_DISABLE;
md.user_ptr = ptl_request;
md.eq_handle = ompi_mtl_portals.ptl_eq_h;
ret = PtlMDBind(ompi_mtl_portals.ptl_ni_h,
md,
PTL_UNLINK,
&(md_h));
if (OMPI_SUCCESS != ret) {
if (ptl_request->free_after) free(md.start);
return ompi_common_portals_error_ptl_to_ompi(ret);
}
ptl_request->event_callback = ompi_mtl_portals_send_progress_no_ack;
ret = PtlPut(md_h,
PTL_NO_ACK_REQ,
endpoint->ptl_proc,
OMPI_MTL_PORTALS_SEND_TABLE_ID,
0,
match_bits,
0,
0);
if (OMPI_SUCCESS != ret) {
PtlMDUnlink(md_h);
if (ptl_request->free_after) free(md.start);
return ompi_common_portals_error_ptl_to_ompi(ret);
}
}
return OMPI_SUCCESS;
}
static int
portals4_init_interface(void)
{
unsigned int ret;
ptl_md_t md;
ptl_me_t me;
/* create event queues */
ret = PtlEQAlloc(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.send_queue_size,
&ompi_mtl_portals4.send_eq_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlEQAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
ret = PtlEQAlloc(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.recv_queue_size,
&ompi_mtl_portals4.recv_eq_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlEQAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
/* Create send and long message (read) portal table entries */
ret = PtlPTAlloc(ompi_mtl_portals4.ni_h,
PTL_PT_ONLY_USE_ONCE |
PTL_PT_ONLY_TRUNCATE |
PTL_PT_FLOWCTRL,
ompi_mtl_portals4.recv_eq_h,
REQ_RECV_TABLE_ID,
&ompi_mtl_portals4.recv_idx);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlPTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
if (ompi_mtl_portals4.recv_idx != REQ_RECV_TABLE_ID) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlPTAlloc did not allocate the requested PT: %d\n",
__FILE__, __LINE__, ompi_mtl_portals4.recv_idx);
goto error;
}
ret = PtlPTAlloc(ompi_mtl_portals4.ni_h,
PTL_PT_ONLY_USE_ONCE |
PTL_PT_ONLY_TRUNCATE,
ompi_mtl_portals4.send_eq_h,
REQ_READ_TABLE_ID,
&ompi_mtl_portals4.read_idx);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlPTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
if (ompi_mtl_portals4.read_idx != REQ_READ_TABLE_ID) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlPTAlloc did not allocate the requested PT: %d\n",
__FILE__, __LINE__, ompi_mtl_portals4.read_idx);
goto error;
}
/* bind zero-length md for sending acks */
md.start = NULL;
md.length = 0;
md.options = 0;
md.eq_handle = PTL_EQ_NONE;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(ompi_mtl_portals4.ni_h,
&md,
&ompi_mtl_portals4.zero_md_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMDBind failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
/* Bind MD across all memory */
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = 0;
md.eq_handle = ompi_mtl_portals4.send_eq_h;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(ompi_mtl_portals4.ni_h,
&md,
&ompi_mtl_portals4.send_md_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMDBind failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
/* Handle long overflows */
me.start = NULL;
me.length = 0;
me.ct_handle = PTL_CT_NONE;
me.min_free = 0;
me.uid = ompi_mtl_portals4.uid;
me.options = PTL_ME_OP_PUT |
PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_COMM_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE;
if (ompi_mtl_portals4.use_logical) {
me.match_id.rank = PTL_RANK_ANY;
} else {
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
}
me.match_bits = MTL_PORTALS4_LONG_MSG;
me.ignore_bits = MTL_PORTALS4_CONTEXT_MASK |
MTL_PORTALS4_SOURCE_MASK |
MTL_PORTALS4_TAG_MASK;
ret = PtlMEAppend(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.recv_idx,
&me,
PTL_OVERFLOW_LIST,
NULL,
&ompi_mtl_portals4.long_overflow_me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
/* attach short unex recv blocks */
ret = ompi_mtl_portals4_recv_short_init();
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: short receive block initialization failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
ompi_mtl_portals4.opcount = 0;
#if OPAL_ENABLE_DEBUG
ompi_mtl_portals4.recv_opcount = 0;
#endif
#if OMPI_MTL_PORTALS4_FLOW_CONTROL
ret = ompi_mtl_portals4_flowctl_init();
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: ompi_mtl_portals4_flowctl_init failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
#endif
return OMPI_SUCCESS;
error:
if (!PtlHandleIsEqual(ompi_mtl_portals4.long_overflow_me_h, PTL_INVALID_HANDLE)) {
PtlMEUnlink(ompi_mtl_portals4.long_overflow_me_h);
}
if (!PtlHandleIsEqual(ompi_mtl_portals4.zero_md_h, PTL_INVALID_HANDLE)) {
PtlMDRelease(ompi_mtl_portals4.zero_md_h);
}
if (!PtlHandleIsEqual(ompi_mtl_portals4.send_md_h, PTL_INVALID_HANDLE)) {
PtlMDRelease(ompi_mtl_portals4.send_md_h);
}
if (ompi_mtl_portals4.read_idx != (ptl_pt_index_t) ~0UL) {
PtlPTFree(ompi_mtl_portals4.ni_h, ompi_mtl_portals4.read_idx);
}
if (ompi_mtl_portals4.recv_idx != (ptl_pt_index_t) ~0UL) {
PtlPTFree(ompi_mtl_portals4.ni_h, ompi_mtl_portals4.recv_idx);
}
if (!PtlHandleIsEqual(ompi_mtl_portals4.send_eq_h, PTL_INVALID_HANDLE)) {
PtlEQFree(ompi_mtl_portals4.send_eq_h);
}
if (!PtlHandleIsEqual(ompi_mtl_portals4.recv_eq_h, PTL_INVALID_HANDLE)) {
PtlEQFree(ompi_mtl_portals4.recv_eq_h);
}
return OMPI_ERROR;
}
int main(int argc,
char *argv[])
{
ptl_handle_ni_t ni_handle;
ptl_process_t *procs;
int rank;
ptl_pt_index_t pt_index, signal_pt_index;
HANDLE_T signal_e_handle;
HANDLE_T signal_e2_handle;
int num_procs;
ptl_handle_eq_t eq_handle;
ptl_handle_ct_t ct_handle;
ptl_handle_md_t md_handle;
ptl_ni_limits_t limits_reqd, limits_actual;
ENTRY_T value_e;
limits_reqd.max_entries = 1024;
limits_reqd.max_unexpected_headers = ITERS*2;
limits_reqd.max_mds = 1024;
limits_reqd.max_eqs = 1024;
limits_reqd.max_cts = 1024;
limits_reqd.max_pt_index = 64;
limits_reqd.max_iovecs = 1024;
limits_reqd.max_list_size = 1024;
limits_reqd.max_triggered_ops = 1024;
limits_reqd.max_msg_size = 1048576;
limits_reqd.max_atomic_size = 1048576;
limits_reqd.max_fetch_atomic_size = 1048576;
limits_reqd.max_waw_ordered_size = 1048576;
limits_reqd.max_war_ordered_size = 1048576;
limits_reqd.max_volatile_size = 1048576;
limits_reqd.features = 0;
CHECK_RETURNVAL(PtlInit());
CHECK_RETURNVAL(libtest_init());
rank = libtest_get_rank();
num_procs = libtest_get_size();
if (num_procs < 2) {
fprintf(stderr, "test_flowctl_noeq requires at least two processes\n");
return 77;
}
int iters;
if (num_procs < ITERS)
iters = ITERS*2+1;
else
iters = ITERS;
CHECK_RETURNVAL(PtlNIInit(PTL_IFACE_DEFAULT, NI_TYPE | PTL_NI_LOGICAL,
PTL_PID_ANY, &limits_reqd, &limits_actual, &ni_handle));
procs = libtest_get_mapping(ni_handle);
CHECK_RETURNVAL(PtlSetMap(ni_handle, num_procs, procs));
if (0 == rank) {
/* create data PT space */
CHECK_RETURNVAL(PtlEQAlloc(ni_handle, (num_procs - 1) * iters + 64, &eq_handle));
CHECK_RETURNVAL(PtlPTAlloc(ni_handle, PTL_PT_FLOWCTRL, eq_handle, 5,
&pt_index));
/* create signal ME */
CHECK_RETURNVAL(PtlCTAlloc(ni_handle, &ct_handle));
CHECK_RETURNVAL(PtlPTAlloc(ni_handle, 1, eq_handle, 6,
&signal_pt_index));
value_e.start = NULL;
value_e.length = 0;
value_e.ct_handle = ct_handle;
value_e.uid = PTL_UID_ANY;
value_e.options = OPTIONS | PTL_LE_EVENT_CT_COMM;
#if INTERFACE == 1
value_e.match_id.rank = PTL_RANK_ANY;
value_e.match_bits = 0;
value_e.ignore_bits = 0;
#endif
CHECK_RETURNVAL(APPEND(ni_handle, 5, &value_e, PTL_OVERFLOW_LIST, NULL, &signal_e_handle));
} else {
ptl_md_t md;
/* 16 extra just in case... */
CHECK_RETURNVAL(PtlEQAlloc(ni_handle, iters*2 + 16, &eq_handle));
md.start = NULL;
md.length = 0;
md.options = 0;
md.eq_handle = eq_handle;
md.ct_handle = PTL_CT_NONE;
CHECK_RETURNVAL(PtlMDBind(ni_handle, &md, &md_handle));
}
fprintf(stderr,"at barrier \n");
libtest_barrier();
if (0 == rank) {
ptl_ct_event_t ct;
ptl_event_t ev;
int ret, count = 0, saw_flowctl = 0;
fprintf(stderr,"begin ctwait \n");
/* wait for signal counts */
CHECK_RETURNVAL(PtlCTWait(ct_handle, iters / 2 , &ct));
if (ct.success != iters / 2 || ct.failure != 0) {
return 1;
}
fprintf(stderr,"done CT wait \n");
/* wait for event entries */
while (1) {
ret = PtlEQGet(eq_handle, &ev);
if (PTL_OK == ret) {
count++;
fprintf(stderr, "found EQ value \n");
} else if (ret == PTL_EQ_EMPTY) {
continue;
} else {
fprintf(stderr, "0: Unexpected return code from EQGet: %d\n", ret);
return 1;
}
if (ev.type == PTL_EVENT_PT_DISABLED) {
saw_flowctl++;
break;
}
}
fprintf(stderr, "0: Saw %d flowctl\n", saw_flowctl);
if (saw_flowctl == 0) {
return 1;
}
/* Now clear out all of the unexpected messages so we can clean up everything */
CHECK_RETURNVAL(APPEND(ni_handle, 5, &value_e, PTL_PRIORITY_LIST, NULL, &signal_e2_handle));
ret = PTL_OK;
while (ret != PTL_EQ_EMPTY)
ret = PtlEQGet(eq_handle, &ev);
} else {
ptl_process_t target;
ptl_event_t ev;
int ret, count = 0, fails = 0;
int i;
target.rank = 0;
printf("beginning puts \n");
for (i = 0 ; i < iters ; ++i) {
CHECK_RETURNVAL(PtlPut(md_handle,
0,
0,
PTL_ACK_REQ,
target,
5,
0,
0,
NULL,
0));
usleep(100);
}
while (count < iters) {
ret = PtlEQGet(eq_handle, &ev);
if (PTL_EQ_EMPTY == ret) {
continue;
} else if (PTL_OK != ret) {
fprintf(stderr, "%d: PtlEQGet returned %d\n", rank, ret);
return 1;
}
if (ev.ni_fail_type == PTL_NI_OK) {
if (ev.type == PTL_EVENT_SEND) {
continue;
} else if (ev.type == PTL_EVENT_ACK) {
count++;
} else {
fprintf(stderr, "%d: Unexpected event type %d\n", rank, ev.type);
}
} else if (ev.ni_fail_type == PTL_NI_PT_DISABLED) {
count++;
fails++;
} else if (ev.ni_fail_type == PTL_EQ_EMPTY) {
continue;
} else if (ev.ni_fail_type == PTL_EQ_DROPPED) {
continue;
} else {
fprintf(stderr, "%d: Unexpected fail type: %d\n", rank, ev.ni_fail_type);
return 1;
}
}
fprintf(stderr, "%d: Saw %d of %d ACKs as fails\n", rank, fails, count);
}
fprintf(stderr,"at final barrier \n");
libtest_barrier();
if (0 == rank) {
CHECK_RETURNVAL(UNLINK(signal_e_handle));
CHECK_RETURNVAL(UNLINK(signal_e2_handle));
CHECK_RETURNVAL(PtlPTFree(ni_handle, signal_pt_index));
CHECK_RETURNVAL(PtlCTFree(ct_handle));
CHECK_RETURNVAL(PtlPTFree(ni_handle, pt_index));
CHECK_RETURNVAL(PtlEQFree(eq_handle));
} else {
CHECK_RETURNVAL(PtlMDRelease(md_handle));
CHECK_RETURNVAL(PtlEQFree(eq_handle));
}
fprintf(stderr,"final cleanup \n");
CHECK_RETURNVAL(PtlNIFini(ni_handle));
CHECK_RETURNVAL(libtest_fini());
PtlFini();
return 0;
}
int main(int argc,
char *argv[])
{
ptl_handle_ni_t ni_handle;
ptl_process_t *procs;
int rank;
ptl_pt_index_t pt_index, signal_pt_index;
HANDLE_T value_e_handle, signal_e_handle;
int num_procs;
ptl_handle_eq_t eq_handle;
ptl_handle_ct_t ct_handle;
ptl_handle_md_t md_handle;
CHECK_RETURNVAL(PtlInit());
CHECK_RETURNVAL(libtest_init());
rank = libtest_get_rank();
num_procs = libtest_get_size();
if (num_procs < 2) {
fprintf(stderr, "test_flowctl_noeq requires at least two processes\n");
return 77;
}
CHECK_RETURNVAL(PtlNIInit(PTL_IFACE_DEFAULT, NI_TYPE | PTL_NI_LOGICAL,
PTL_PID_ANY, NULL, NULL, &ni_handle));
procs = libtest_get_mapping(ni_handle);
CHECK_RETURNVAL(PtlSetMap(ni_handle, num_procs, procs));
if (0 == rank) {
ENTRY_T value_e;
/* create data ME */
CHECK_RETURNVAL(PtlEQAlloc(ni_handle, (num_procs - 1) * ITERS / 2, &eq_handle));
CHECK_RETURNVAL(PtlPTAlloc(ni_handle, PTL_PT_FLOWCTRL, eq_handle, 5,
&pt_index));
value_e.start = NULL;
value_e.length = 0;
value_e.ct_handle = PTL_CT_NONE;
value_e.uid = PTL_UID_ANY;
value_e.options = OPTIONS;
#if INTERFACE == 1
value_e.match_id.rank = PTL_RANK_ANY;
value_e.match_bits = 0;
value_e.ignore_bits = 0;
#endif
CHECK_RETURNVAL(APPEND(ni_handle, 5, &value_e, PTL_PRIORITY_LIST, NULL, &value_e_handle));
/* create signal ME */
CHECK_RETURNVAL(PtlCTAlloc(ni_handle, &ct_handle));
CHECK_RETURNVAL(PtlPTAlloc(ni_handle, 0, PTL_EQ_NONE, 6,
&signal_pt_index));
value_e.start = NULL;
value_e.length = 0;
value_e.ct_handle = ct_handle;
value_e.uid = PTL_UID_ANY;
value_e.options = OPTIONS | PTL_LE_EVENT_SUCCESS_DISABLE | PTL_LE_EVENT_CT_COMM;
#if INTERFACE == 1
value_e.match_id.rank = PTL_RANK_ANY;
value_e.match_bits = 0;
value_e.ignore_bits = 0;
#endif
CHECK_RETURNVAL(APPEND(ni_handle, 6, &value_e, PTL_PRIORITY_LIST, NULL, &signal_e_handle));
} else {
ptl_md_t md;
/* 16 extra just in case... */
CHECK_RETURNVAL(PtlEQAlloc(ni_handle, ITERS * 2 + 16, &eq_handle));
md.start = NULL;
md.length = 0;
md.options = 0;
md.eq_handle = eq_handle;
md.ct_handle = PTL_CT_NONE;
CHECK_RETURNVAL(PtlMDBind(ni_handle, &md, &md_handle));
}
libtest_barrier();
if (0 == rank) {
ptl_ct_event_t ct;
ptl_event_t ev;
int ret, count = 0, saw_dropped = 0, saw_flowctl = 0;
/* wait for signal counts */
CHECK_RETURNVAL(PtlCTWait(ct_handle, num_procs - 1, &ct));
if (ct.success != num_procs - 1 || ct.failure != 0) {
return 1;
}
/* wait for event entries */
while (count < ITERS * (num_procs - 1)) {
ret = PtlEQWait(eq_handle, &ev);
if (PTL_OK == ret) {
;
} else if (PTL_EQ_DROPPED == ret) {
saw_dropped++;
if (ev.type == PTL_EVENT_PT_DISABLED){
saw_flowctl++;
CHECK_RETURNVAL(PtlPTEnable(ni_handle, pt_index));
}
break;
} else {
fprintf(stderr, "0: Unexpected return code from EQWait: %d\n", ret);
return 1;
}
if (ev.type == PTL_EVENT_PT_DISABLED) {
CHECK_RETURNVAL(PtlPTEnable(ni_handle, pt_index));
saw_flowctl++;
} else {
count++;
}
}
fprintf(stderr, "0: Saw %d dropped, %d flowctl\n", saw_dropped, saw_flowctl);
if (saw_flowctl == 0) {
return 1;
}
} else {
ptl_process_t target;
ptl_event_t ev;
int ret, count = 0, fails = 0;
int i;
int *fail_seen;
fail_seen = malloc(sizeof(int) * ITERS);
if (NULL == fail_seen) {
fprintf(stderr, "%d: malloc failed\n", rank);
return 1;
}
memset(fail_seen, 0, sizeof(int) * ITERS);
target.rank = 0;
for (i = 0 ; i < ITERS ; ++i) {
CHECK_RETURNVAL(PtlPut(md_handle,
0,
0,
PTL_ACK_REQ,
target,
5,
0,
0,
(void*)(size_t)i,
0));
usleep(100);
}
while (count < ITERS) {
ret = PtlEQGet(eq_handle, &ev);
if (PTL_EQ_EMPTY == ret) {
continue;
} else if (PTL_OK != ret) {
fprintf(stderr, "%d: PtlEQGet returned %d\n", rank, ret);
return 1;
}
if (ev.ni_fail_type == PTL_NI_OK) {
if (ev.type == PTL_EVENT_SEND) {
continue;
} else if (ev.type == PTL_EVENT_ACK) {
count++;
} else {
fprintf(stderr, "%d: Unexpected event type %d\n", rank, ev.type);
}
} else if (ev.ni_fail_type == PTL_NI_PT_DISABLED) {
int iter = (size_t) ev.user_ptr;
if (fail_seen[iter]++ > 0) {
fprintf(stderr, "%d: Double report of PT_DISABLED for "
"iteration %d\n", rank, iter);
return 1;
}
count++;
fails++;
} else {
fprintf(stderr, "%d: Unexpected fail type: %d\n", rank, ev.ni_fail_type);
return 1;
}
}
fprintf(stderr, "%d: Saw %d of %d events as fails\n", rank, fails, count);
CHECK_RETURNVAL(PtlPut(md_handle,
0,
0,
PTL_NO_ACK_REQ,
target,
6,
0,
0,
NULL,
0));
/* wait for the send event on the last put */
CHECK_RETURNVAL(PtlEQWait(eq_handle, &ev));
while (fails > 0) {
CHECK_RETURNVAL(PtlPut(md_handle,
0,
0,
PTL_ACK_REQ,
target,
5,
0,
0,
NULL,
0));
while (1) {
ret = PtlEQWait(eq_handle, &ev);
if (PTL_OK != ret) {
fprintf(stderr, "%d: PtlEQWait returned %d\n", rank, ret);
return 1;
}
if (ev.ni_fail_type == PTL_NI_OK) {
if (ev.type == PTL_EVENT_SEND) {
continue;
} else if (ev.type == PTL_EVENT_ACK) {
fails--;
break;
} else {
fprintf(stderr, "%d: Unexpected event type %d\n", rank, ev.type);
}
} else if (ev.ni_fail_type == PTL_NI_PT_DISABLED) {
break;
} else {
fprintf(stderr, "%d: Unexpected fail type: %d\n", rank, ev.ni_fail_type);
return 1;
}
}
}
}
libtest_barrier();
if (0 == rank) {
CHECK_RETURNVAL(UNLINK(signal_e_handle));
CHECK_RETURNVAL(PtlPTFree(ni_handle, signal_pt_index));
CHECK_RETURNVAL(PtlCTFree(ct_handle));
CHECK_RETURNVAL(UNLINK(value_e_handle));
CHECK_RETURNVAL(PtlPTFree(ni_handle, pt_index));
CHECK_RETURNVAL(PtlEQFree(eq_handle));
} else {
CHECK_RETURNVAL(PtlMDRelease(md_handle));
CHECK_RETURNVAL(PtlEQFree(eq_handle));
}
CHECK_RETURNVAL(PtlNIFini(ni_handle));
CHECK_RETURNVAL(libtest_fini());
PtlFini();
return 0;
}
/*
/!\ Called for each processes /!\
*/
static int
portals4_init_query(bool enable_progress_threads,
bool enable_mpi_threads)
{
int ret;
ptl_md_t md;
ptl_me_t me;
/* Initialize Portals and create a physical, matching interface */
ret = PtlInit();
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlInit failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
ret = PtlNIInit(PTL_IFACE_DEFAULT,
PTL_NI_PHYSICAL | PTL_NI_MATCHING,
PTL_PID_ANY,
NULL,
&mca_coll_portals4_component.ni_limits,
&mca_coll_portals4_component.ni_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlNIInit failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
ret = PtlGetId(mca_coll_portals4_component.ni_h, &mca_coll_portals4_component.id);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlGetid failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* FIX ME: Need to make sure our ID matches with the MTL... */
ret = PtlGetUid(mca_coll_portals4_component.ni_h, &mca_coll_portals4_component.uid);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlGetUid failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
ret = PtlEQAlloc(mca_coll_portals4_component.ni_h,
MCA_COLL_PORTALS4_EQ_SIZE,
&mca_coll_portals4_component.eq_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlEQAlloc failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
ret = PtlPTAlloc(mca_coll_portals4_component.ni_h,
0,
mca_coll_portals4_component.eq_h,
REQ_COLL_TABLE_ID,
&mca_coll_portals4_component.pt_idx);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlPTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
if (mca_coll_portals4_component.pt_idx != REQ_COLL_TABLE_ID) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlPTAlloc return wrong pt_idx: %d\n",
__FILE__, __LINE__,
mca_coll_portals4_component.finish_pt_idx);
return OMPI_ERROR;
}
ret = PtlPTAlloc(mca_coll_portals4_component.ni_h,
0,
mca_coll_portals4_component.eq_h,
REQ_COLL_FINISH_TABLE_ID,
&mca_coll_portals4_component.finish_pt_idx);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlPTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
if (mca_coll_portals4_component.finish_pt_idx != REQ_COLL_FINISH_TABLE_ID) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlPTAlloc return wrong pt_idx: %d\n",
__FILE__, __LINE__,
mca_coll_portals4_component.finish_pt_idx);
return OMPI_ERROR;
}
/* Bind MD/MDs across all memory. We prefer (for obvious reasons)
to have a single MD across all of memory */
memset(&md, 0, sizeof(ptl_md_t));
md.start = 0;
md.length = 0;
md.options = 0;
md.eq_handle = PTL_EQ_NONE;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(mca_coll_portals4_component.ni_h,
&md,
&mca_coll_portals4_component.zero_md_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlMDBind failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = 0;
md.eq_handle = PTL_EQ_NONE;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(mca_coll_portals4_component.ni_h,
&md,
&mca_coll_portals4_component.data_md_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlMDBind failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
OPAL_OUTPUT_VERBOSE((90, ompi_coll_base_framework.framework_output, "PtlMDBind start=%p length=%x\n", md.start, md.length));
/* setup finish ack ME */
me.start = NULL;
me.length = 0;
me.ct_handle = PTL_CT_NONE;
me.min_free = 0;
me.uid = mca_coll_portals4_component.uid;
me.options = PTL_ME_OP_PUT |
PTL_ME_EVENT_LINK_DISABLE | PTL_ME_EVENT_UNLINK_DISABLE;
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
me.match_bits = 0;
me.ignore_bits = 0;
ret = PtlMEAppend(mca_coll_portals4_component.ni_h,
mca_coll_portals4_component.finish_pt_idx,
&me,
PTL_PRIORITY_LIST,
NULL,
&mca_coll_portals4_component.finish_me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlMEAppend of barrier unexpected failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* This ME is used for RTR exchange only */
me.start = NULL;
me.length = 0;
me.ct_handle = PTL_CT_NONE;
me.min_free = 0;
me.uid = mca_coll_portals4_component.uid;
me.options = PTL_ME_OP_PUT |
PTL_ME_EVENT_SUCCESS_DISABLE | PTL_ME_EVENT_OVER_DISABLE |
PTL_ME_EVENT_LINK_DISABLE | PTL_ME_EVENT_UNLINK_DISABLE;
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
/* Note : the RTR bit must be set to match this ME,
* this allows to discriminate the RTR from data flow
* (especially for the Barrier operations)
*/
COLL_PORTALS4_SET_BITS(me.match_bits, 0, 0, 1, 0, 0, 0);
me.ignore_bits = ~COLL_PORTALS4_RTR_MASK;
ret = PtlMEAppend(mca_coll_portals4_component.ni_h,
mca_coll_portals4_component.pt_idx,
&me,
PTL_OVERFLOW_LIST,
NULL,
&mca_coll_portals4_component.unex_me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlMEAppend of barrier unexpected failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* activate progress callback */
ret = opal_progress_register(portals4_progress);
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: opal_progress_register failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
return OMPI_SUCCESS;
}
static int ptl_init(MPIDI_PG_t *pg_p, int pg_rank, char **bc_val_p, int *val_max_sz_p)
{
int mpi_errno = MPI_SUCCESS;
int ret;
ptl_md_t md;
ptl_ni_limits_t desired;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_PTL_INIT);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_PTL_INIT);
/* first make sure that our private fields in the vc and req fit into the area provided */
MPIR_Assert(sizeof(MPID_nem_ptl_vc_area) <= MPIDI_NEM_VC_NETMOD_AREA_LEN);
MPIR_Assert(sizeof(MPID_nem_ptl_req_area) <= MPIDI_NEM_REQ_NETMOD_AREA_LEN);
/* Make sure our IOV is the same as portals4's IOV */
MPIR_Assert(sizeof(ptl_iovec_t) == sizeof(MPL_IOV));
MPIR_Assert(((void*)&(((ptl_iovec_t*)0)->iov_base)) == ((void*)&(((MPL_IOV*)0)->MPL_IOV_BUF)));
MPIR_Assert(((void*)&(((ptl_iovec_t*)0)->iov_len)) == ((void*)&(((MPL_IOV*)0)->MPL_IOV_LEN)));
MPIR_Assert(sizeof(((ptl_iovec_t*)0)->iov_len) == sizeof(((MPL_IOV*)0)->MPL_IOV_LEN));
mpi_errno = MPIDI_CH3I_Register_anysource_notification(MPID_nem_ptl_anysource_posted, MPID_nem_ptl_anysource_matched);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIDI_Anysource_improbe_fn = MPID_nem_ptl_anysource_improbe;
/* init portals */
ret = PtlInit();
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlinit", "**ptlinit %s", MPID_nem_ptl_strerror(ret));
/* do an interface pre-init to get the default limits struct */
ret = PtlNIInit(PTL_IFACE_DEFAULT, PTL_NI_MATCHING | PTL_NI_PHYSICAL,
PTL_PID_ANY, NULL, &desired, &MPIDI_nem_ptl_ni);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlniinit", "**ptlniinit %s", MPID_nem_ptl_strerror(ret));
/* finalize the interface so we can re-init with our desired maximums */
ret = PtlNIFini(MPIDI_nem_ptl_ni);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlnifini", "**ptlnifini %s", MPID_nem_ptl_strerror(ret));
/* set higher limits if they are determined to be too low */
if (desired.max_unexpected_headers < UNEXPECTED_HDR_COUNT && getenv("PTL_LIM_MAX_UNEXPECTED_HEADERS") == NULL)
desired.max_unexpected_headers = UNEXPECTED_HDR_COUNT;
if (desired.max_list_size < LIST_SIZE && getenv("PTL_LIM_MAX_LIST_SIZE") == NULL)
desired.max_list_size = LIST_SIZE;
if (desired.max_entries < ENTRY_COUNT && getenv("PTL_LIM_MAX_ENTRIES") == NULL)
desired.max_entries = ENTRY_COUNT;
/* do the real init */
ret = PtlNIInit(PTL_IFACE_DEFAULT, PTL_NI_MATCHING | PTL_NI_PHYSICAL,
PTL_PID_ANY, &desired, &MPIDI_nem_ptl_ni_limits, &MPIDI_nem_ptl_ni);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlniinit", "**ptlniinit %s", MPID_nem_ptl_strerror(ret));
/* allocate EQs for each portal */
ret = PtlEQAlloc(MPIDI_nem_ptl_ni, EVENT_COUNT, &MPIDI_nem_ptl_eq);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptleqalloc", "**ptleqalloc %s", MPID_nem_ptl_strerror(ret));
ret = PtlEQAlloc(MPIDI_nem_ptl_ni, EVENT_COUNT, &MPIDI_nem_ptl_get_eq);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptleqalloc", "**ptleqalloc %s", MPID_nem_ptl_strerror(ret));
ret = PtlEQAlloc(MPIDI_nem_ptl_ni, EVENT_COUNT, &MPIDI_nem_ptl_control_eq);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptleqalloc", "**ptleqalloc %s", MPID_nem_ptl_strerror(ret));
ret = PtlEQAlloc(MPIDI_nem_ptl_ni, EVENT_COUNT, &MPIDI_nem_ptl_rpt_eq);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptleqalloc", "**ptleqalloc %s", MPID_nem_ptl_strerror(ret));
/* allocate a separate EQ for origin events. with this, we can implement rate-limit operations
to prevent a locally triggered flow control even */
ret = PtlEQAlloc(MPIDI_nem_ptl_ni, EVENT_COUNT, &MPIDI_nem_ptl_origin_eq);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptleqalloc", "**ptleqalloc %s", MPID_nem_ptl_strerror(ret));
/* allocate portal for matching messages */
ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_eq,
PTL_PT_ANY, &MPIDI_nem_ptl_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
/* allocate portal for large messages where receiver does a get */
ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_get_eq,
PTL_PT_ANY, &MPIDI_nem_ptl_get_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
/* allocate portal for MPICH control messages */
ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_control_eq,
PTL_PT_ANY, &MPIDI_nem_ptl_control_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
/* allocate portal for MPICH control messages */
ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_rpt_eq,
PTL_PT_ANY, &MPIDI_nem_ptl_rpt_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
/* allocate portal for MPICH control messages */
ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_rpt_eq,
PTL_PT_ANY, &MPIDI_nem_ptl_get_rpt_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
/* allocate portal for MPICH control messages */
ret = PtlPTAlloc(MPIDI_nem_ptl_ni, PTL_PT_ONLY_USE_ONCE | PTL_PT_ONLY_TRUNCATE | PTL_PT_FLOWCTRL, MPIDI_nem_ptl_rpt_eq,
PTL_PT_ANY, &MPIDI_nem_ptl_control_rpt_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
/* create an MD that covers all of memory */
md.start = 0;
md.length = (ptl_size_t)-1;
md.options = 0x0;
md.eq_handle = MPIDI_nem_ptl_origin_eq;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(MPIDI_nem_ptl_ni, &md, &MPIDI_nem_ptl_global_md);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmdbind", "**ptlmdbind %s", MPID_nem_ptl_strerror(ret));
/* currently, rportlas only works with a single NI and EQ */
ret = MPID_nem_ptl_rptl_init(MPIDI_Process.my_pg->size, ORIGIN_EVENTS, get_target_info);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlniinit", "**ptlniinit %s", MPID_nem_ptl_strerror(ret));
/* allow rportal to manage the primary portal and retransmit if needed */
ret = MPID_nem_ptl_rptl_ptinit(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_origin_eq, MPIDI_nem_ptl_pt, MPIDI_nem_ptl_rpt_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
/* allow rportal to manage the get and control portals, but we
* don't expect retransmission to be needed on the get portal, so
* we pass PTL_PT_ANY as the dummy portal. unfortunately, portals
* does not have an "invalid" PT constant, which would have been
* more appropriate to pass over here. */
ret = MPID_nem_ptl_rptl_ptinit(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_origin_eq, MPIDI_nem_ptl_get_pt, MPIDI_nem_ptl_get_rpt_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
ret = MPID_nem_ptl_rptl_ptinit(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_origin_eq, MPIDI_nem_ptl_control_pt, MPIDI_nem_ptl_control_rpt_pt);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlptalloc", "**ptlptalloc %s", MPID_nem_ptl_strerror(ret));
/* create business card */
mpi_errno = get_business_card(pg_rank, bc_val_p, val_max_sz_p);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* init other modules */
mpi_errno = MPID_nem_ptl_poll_init();
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPID_nem_ptl_nm_init();
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_PTL_INIT);
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int
component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit,
struct ompi_communicator_t *comm, struct ompi_info_t *info,
int flavor, int *model)
{
ompi_osc_portals4_module_t *module = NULL;
int ret = OMPI_ERROR;
int tmp;
ptl_md_t md;
ptl_me_t me;
char *name;
if (MPI_WIN_FLAVOR_SHARED == flavor) return OMPI_ERR_NOT_SUPPORTED;
/* create module structure */
module = (ompi_osc_portals4_module_t*)
calloc(1, sizeof(ompi_osc_portals4_module_t));
if (NULL == module) return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
/* fill in the function pointer part */
memcpy(module, &ompi_osc_portals4_module_template,
sizeof(ompi_osc_base_module_t));
/* fill in our part */
if (MPI_WIN_FLAVOR_ALLOCATE == flavor) {
module->free_after = *base = malloc(size);
if (NULL == *base) goto error;
} else {
module->free_after = NULL;
}
ret = ompi_comm_dup(comm, &module->comm);
if (OMPI_SUCCESS != ret) goto error;
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"portals4 component creating window with id %d",
ompi_comm_get_cid(module->comm));
asprintf(&name, "portals4 window %d", ompi_comm_get_cid(module->comm));
ompi_win_set_name(win, name);
free(name);
/* share everyone's displacement units. Only do an allgather if
strictly necessary, since it requires O(p) state. */
tmp = disp_unit;
ret = module->comm->c_coll.coll_bcast(&tmp, 1, MPI_INT, 0,
module->comm,
module->comm->c_coll.coll_bcast_module);
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: MPI_Bcast failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
tmp = (tmp == disp_unit) ? 1 : 0;
ret = module->comm->c_coll.coll_allreduce(MPI_IN_PLACE, &tmp, 1, MPI_INT, MPI_LAND,
module->comm, module->comm->c_coll.coll_allreduce_module);
if (OMPI_SUCCESS != ret) goto error;
if (tmp == 1) {
module->disp_unit = disp_unit;
module->disp_units = NULL;
} else {
module->disp_unit = -1;
module->disp_units = malloc(sizeof(int) * ompi_comm_size(module->comm));
ret = module->comm->c_coll.coll_allgather(&disp_unit, 1, MPI_INT,
module->disp_units, 1, MPI_INT,
module->comm,
module->comm->c_coll.coll_allgather_module);
if (OMPI_SUCCESS != ret) goto error;
}
module->ni_h = mca_osc_portals4_component.matching_ni_h;
module->pt_idx = mca_osc_portals4_component.matching_pt_idx;
ret = PtlCTAlloc(module->ni_h, &(module->ct_h));
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: PtlCTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = PTL_MD_EVENT_SUCCESS_DISABLE | PTL_MD_EVENT_CT_REPLY | PTL_MD_EVENT_CT_ACK;
md.eq_handle = mca_osc_portals4_component.matching_eq_h;
md.ct_handle = module->ct_h;
ret = PtlMDBind(module->ni_h, &md, &module->md_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: PtlMDBind failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
md.start = 0;
md.length = PTL_SIZE_MAX;
md.options = PTL_MD_EVENT_CT_REPLY | PTL_MD_EVENT_CT_ACK;
md.eq_handle = mca_osc_portals4_component.matching_eq_h;
md.ct_handle = module->ct_h;
ret = PtlMDBind(module->ni_h, &md, &module->req_md_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: PtlMDBind failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
if (MPI_WIN_FLAVOR_DYNAMIC == flavor) {
me.start = 0;
me.length = PTL_SIZE_MAX;
} else {
me.start = *base;
me.length = size;
}
me.ct_handle = PTL_CT_NONE;
me.uid = PTL_UID_ANY;
me.options = PTL_ME_OP_PUT | PTL_ME_OP_GET | PTL_ME_NO_TRUNCATE | PTL_ME_EVENT_SUCCESS_DISABLE;
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
me.match_bits = module->comm->c_contextid;
me.ignore_bits = 0;
ret = PtlMEAppend(module->ni_h,
module->pt_idx,
&me,
PTL_PRIORITY_LIST,
NULL,
&module->data_me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
me.start = &module->state;
me.length = sizeof(module->state);
me.ct_handle = PTL_CT_NONE;
me.uid = PTL_UID_ANY;
me.options = PTL_ME_OP_PUT | PTL_ME_OP_GET | PTL_ME_NO_TRUNCATE | PTL_ME_EVENT_SUCCESS_DISABLE;
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
me.match_bits = module->comm->c_contextid | OSC_PORTALS4_MB_CONTROL;
me.ignore_bits = 0;
ret = PtlMEAppend(module->ni_h,
module->pt_idx,
&me,
PTL_PRIORITY_LIST,
NULL,
&module->control_me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
module->opcount = 0;
module->match_bits = module->comm->c_contextid;
module->atomic_max = (check_config_value_equal("accumulate_ordering", info, "none")) ?
mca_osc_portals4_component.matching_atomic_max :
MIN(mca_osc_portals4_component.matching_atomic_max,
mca_osc_portals4_component.matching_atomic_ordered_size);
module->fetch_atomic_max = (check_config_value_equal("accumulate_ordering", info, "none")) ?
mca_osc_portals4_component.matching_fetch_atomic_max :
MIN(mca_osc_portals4_component.matching_fetch_atomic_max,
mca_osc_portals4_component.matching_atomic_ordered_size);
module->zero = 0;
module->one = 1;
module->start_group = NULL;
module->post_group = NULL;
module->state.post_count = 0;
module->state.complete_count = 0;
if (check_config_value_bool("no_locks", info)) {
module->state.lock = LOCK_ILLEGAL;
} else {
module->state.lock = LOCK_UNLOCKED;
}
OBJ_CONSTRUCT(&module->outstanding_locks, opal_list_t);
module->passive_target_access_epoch = false;
#if OPAL_ASSEMBLY_ARCH == OPAL_AMD64 || OPAL_ASSEMBLY_ARCH == OPAL_IA32
*model = MPI_WIN_UNIFIED;
#else
*model = MPI_WIN_SEPARATE;
#endif
win->w_osc_module = &module->super;
PtlAtomicSync();
/* Make sure that everyone's ready to receive. */
module->comm->c_coll.coll_barrier(module->comm,
module->comm->c_coll.coll_barrier_module);
return OMPI_SUCCESS;
error:
/* BWB: FIX ME: This is all wrong... */
if (0 != module->ct_h) PtlCTFree(module->ct_h);
if (0 != module->data_me_h) PtlMEUnlink(module->data_me_h);
if (0 != module->req_md_h) PtlMDRelease(module->req_md_h);
if (0 != module->md_h) PtlMDRelease(module->md_h);
if (NULL != module->comm) ompi_comm_free(&module->comm);
if (NULL != module) free(module);
return ret;
}
/* called when a receive should be progressed */
int
ompi_mtl_portals4_recv_progress(ptl_event_t *ev,
ompi_mtl_portals4_base_request_t* ptl_base_request)
{
int ret;
ompi_mtl_portals4_recv_request_t* ptl_request =
(ompi_mtl_portals4_recv_request_t*) ptl_base_request;
size_t msg_length = 0;
switch (ev->type) {
case PTL_EVENT_PUT:
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_output, "Recv %d (0x%lx) got put event",
ptl_request->opcount, ev->hdr_data));
if (ev->ni_fail_type != PTL_NI_OK) {
opal_output_verbose(1, ompi_mtl_base_output,
"%s:%d: PTL_EVENT_PUT with ni_fail_type: %d",
__FILE__, __LINE__, ev->ni_fail_type);
goto callback_error;
}
msg_length = MTL_PORTALS4_GET_LENGTH(ev->hdr_data);
ptl_request->super.super.ompi_req->req_status.MPI_SOURCE =
MTL_PORTALS4_GET_SOURCE(ev->match_bits);
ptl_request->super.super.ompi_req->req_status.MPI_TAG =
MTL_PORTALS4_GET_TAG(ev->match_bits);
if (msg_length > ptl_request->delivery_len) {
opal_output_verbose(1, ompi_mtl_base_output, "truncate expected: %ld %ld",
msg_length, ptl_request->delivery_len);
ptl_request->super.super.ompi_req->req_status.MPI_ERROR = MPI_ERR_TRUNCATE;
}
#if OPAL_ENABLE_DEBUG
ptl_request->hdr_data = ev->hdr_data;
#endif
if (!MTL_PORTALS4_IS_SHORT_MSG(ev->match_bits) && ompi_mtl_portals4.protocol == rndv) {
ptl_md_t md;
md.start = (char*) ptl_request->delivery_ptr + ompi_mtl_portals4.eager_limit;
md.length = ((msg_length > ptl_request->delivery_len) ?
ptl_request->delivery_len : msg_length) - ompi_mtl_portals4.eager_limit;
md.options = 0;
md.eq_handle = ompi_mtl_portals4.eq_h;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(ompi_mtl_portals4.ni_h,
&md,
&ptl_request->md_h);
if (PTL_OK != ret) {
if (NULL != ptl_request->buffer_ptr) free(ptl_request->buffer_ptr);
opal_output_verbose(1, ompi_mtl_base_output,
"%s:%d: PtlMDBind failed: %d",
__FILE__, __LINE__, ret);
goto callback_error;
}
ret = PtlGet(ptl_request->md_h,
0,
md.length,
ev->initiator,
ompi_mtl_portals4.read_idx,
ev->hdr_data,
ompi_mtl_portals4.eager_limit,
ptl_request);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_mtl_base_output,
"%s:%d: PtlGet failed: %d",
__FILE__, __LINE__, ret);
PtlMDRelease(ptl_request->md_h);
if (NULL != ptl_request->buffer_ptr) free(ptl_request->buffer_ptr);
goto callback_error;
}
} else {
/* make sure the data is in the right place */
ret = ompi_mtl_datatype_unpack(ptl_request->convertor,
ev->start,
ev->mlength);
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_mtl_base_output,
"%s:%d: ompi_mtl_datatype_unpack failed: %d",
__FILE__, __LINE__, ret);
ptl_request->super.super.ompi_req->req_status.MPI_ERROR = ret;
}
ptl_request->super.super.ompi_req->req_status._ucount = ev->mlength;
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_output, "Recv %d (0x%lx) completed, expected",
ptl_request->opcount, ptl_request->hdr_data));
ptl_request->super.super.completion_callback(&ptl_request->super.super);
}
break;
case PTL_EVENT_REPLY:
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_output, "Recv %d (0x%lx) got reply event",
ptl_request->opcount, ptl_request->hdr_data));
if (ev->ni_fail_type != PTL_NI_OK) {
opal_output_verbose(1, ompi_mtl_base_output,
"%s:%d: PTL_EVENT_REPLY with ni_fail_type: %d",
__FILE__, __LINE__, ev->ni_fail_type);
PtlMDRelease(ptl_request->md_h);
goto callback_error;
}
/* set the status - most of this filled in right after issuing
the PtlGet */
ptl_request->super.super.ompi_req->req_status._ucount = ev->mlength;
if (ompi_mtl_portals4.protocol == rndv) {
ptl_request->super.super.ompi_req->req_status._ucount +=
ompi_mtl_portals4.eager_limit;
}
/* make sure the data is in the right place. Use _ucount for
the total length because it will be set correctly for all
three protocols. mlength is only correct for eager, and
delivery_len is the length of the buffer, not the length of
the send. */
ret = ompi_mtl_datatype_unpack(ptl_request->convertor,
ptl_request->delivery_ptr,
ptl_request->super.super.ompi_req->req_status._ucount);
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_mtl_base_output,
"%s:%d: ompi_mtl_datatype_unpack failed: %d",
__FILE__, __LINE__, ret);
ptl_request->super.super.ompi_req->req_status.MPI_ERROR = ret;
}
PtlMDRelease(ptl_request->md_h);
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_output, "Recv %d (0x%lx) completed, reply",
ptl_request->opcount, ptl_request->hdr_data));
ptl_request->super.super.completion_callback(&ptl_request->super.super);
break;
case PTL_EVENT_PUT_OVERFLOW:
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_output, "Recv %d (0x%lx) got put_overflow event",
ptl_request->opcount, ev->hdr_data));
if (ev->ni_fail_type != PTL_NI_OK) {
opal_output_verbose(1, ompi_mtl_base_output,
"%s:%d: PTL_EVENT_PUT_OVERFLOW with ni_fail_type: %d",
__FILE__, __LINE__, ev->ni_fail_type);
goto callback_error;
}
msg_length = MTL_PORTALS4_GET_LENGTH(ev->hdr_data);
ptl_request->super.super.ompi_req->req_status.MPI_SOURCE =
MTL_PORTALS4_GET_SOURCE(ev->match_bits);
ptl_request->super.super.ompi_req->req_status.MPI_TAG =
MTL_PORTALS4_GET_TAG(ev->match_bits);
if (msg_length > ptl_request->delivery_len) {
opal_output_verbose(1, ompi_mtl_base_output, "truncate unexpected: %ld %ld %d",
msg_length, ptl_request->delivery_len, MTL_PORTALS4_IS_SHORT_MSG(ev->match_bits));
ptl_request->super.super.ompi_req->req_status.MPI_ERROR = MPI_ERR_TRUNCATE;
}
#if OPAL_ENABLE_DEBUG
ptl_request->hdr_data = ev->hdr_data;
#endif
/* overflow case. Short messages have the buffer stashed
somewhere. Long messages left in buffer at the source */
if (MTL_PORTALS4_IS_SHORT_MSG(ev->match_bits)) {
ptl_request->super.super.ompi_req->req_status._ucount = ev->mlength;
if (ev->mlength > 0) {
struct iovec iov;
uint32_t iov_count = 1;
size_t max_data;
iov.iov_base = (char*) ev->start;
iov.iov_len = ev->mlength;
max_data = iov.iov_len;
ret = opal_convertor_unpack(ptl_request->convertor,
&iov, &iov_count,
&max_data );
if (NULL != ptl_request->buffer_ptr) free(ptl_request->buffer_ptr);
if (ret < 0) {
opal_output_verbose(1, ompi_mtl_base_output,
"%s:%d: opal_convertor_unpack failed: %d",
__FILE__, __LINE__, ret);
goto callback_error;
}
}
/* if it's a sync, send the ack */
if (MTL_PORTALS4_IS_SYNC_MSG(ev->hdr_data)) {
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_output, "Recv %d (0x%lx) sending sync ack",
ptl_request->opcount, ptl_request->hdr_data));
ret = PtlPut(ompi_mtl_portals4.zero_md_h,
0,
0,
PTL_NO_ACK_REQ,
ev->initiator,
ompi_mtl_portals4.read_idx,
ev->hdr_data,
0,
NULL,
0);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_mtl_base_output,
"%s:%d: PtlPut failed: %d",
__FILE__, __LINE__, ret);
goto callback_error;
}
}
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_output, "Recv %d (0x%lx) completed, unexpected short (0x%lx)",
ptl_request->opcount, ptl_request->hdr_data, (long) ev->start));
ptl_request->super.super.completion_callback(&ptl_request->super.super);
} else {
ptl_md_t md;
if (ev->mlength > 0) {
/* if rndv or triggered, copy the eager part to the right place */
memcpy(ptl_request->delivery_ptr, ev->start, ev->mlength);
}
md.start = (char*) ptl_request->delivery_ptr + ev->mlength;
md.length = ((msg_length > ptl_request->delivery_len) ?
ptl_request->delivery_len : msg_length) - ev->mlength;
md.options = 0;
md.eq_handle = ompi_mtl_portals4.eq_h;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(ompi_mtl_portals4.ni_h,
&md,
&ptl_request->md_h);
if (PTL_OK != ret) {
if (NULL != ptl_request->buffer_ptr) free(ptl_request->buffer_ptr);
opal_output_verbose(1, ompi_mtl_base_output,
"%s:%d: PtlMDBind failed: %d",
__FILE__, __LINE__, ret);
goto callback_error;
}
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_output, "Recv %d (0x%lx) getting long data",
ptl_request->opcount, ptl_request->hdr_data));
ret = PtlGet(ptl_request->md_h,
0,
md.length,
ev->initiator,
ompi_mtl_portals4.read_idx,
ev->hdr_data,
ev->mlength,
ptl_request);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_mtl_base_output,
"%s:%d: PtlGet failed: %d",
__FILE__, __LINE__, ret);
if (NULL != ptl_request->buffer_ptr) free(ptl_request->buffer_ptr);
PtlMDRelease(ptl_request->md_h);
goto callback_error;
}
}
break;
default:
opal_output_verbose(1, ompi_mtl_base_output,
"Unhandled receive callback with event type %d",
ev->type);
return OMPI_ERROR;
}
return OMPI_SUCCESS;
callback_error:
ptl_request->super.super.ompi_req->req_status.MPI_ERROR =
ompi_mtl_portals4_get_error(ret);
ptl_request->super.super.completion_callback(&ptl_request->super.super);
return OMPI_SUCCESS;
}