static void postME(ptl_handle_ni_t ni,
ptl_pt_index_t index,
void *start,
ptl_size_t length,
ptl_process_t src,
int tag,
ptl_handle_me_t *mh)
{
int rc;
ptl_me_t me;
unsigned int options = PTL_ME_OP_PUT | PTL_ME_ACK_DISABLE |
PTL_ME_USE_ONCE | PTL_ME_EVENT_UNLINK_DISABLE |
PTL_ME_EVENT_LINK_DISABLE;
me.start = (char *)start;
me.length = length;
me.ct_handle = PTL_CT_NONE;
me.min_free = 0;
me.uid = PTL_UID_ANY;
me.options = options;
me.match_id = src;
me.match_bits = tag;
me.ignore_bits = 0;
rc = PtlMEAppend(ni, index, &me, PTL_PRIORITY_LIST, NULL, mh);
LIBTEST_CHECK(rc, "Error in libtest_CreateME(): PtlMEAppend");
}
static inline int
ompi_mtl_portals4_activate_block(ompi_mtl_portals4_recv_short_block_t *block)
{
ptl_match_bits_t match_bits = MTL_PORTALS4_SHORT_MSG;
ptl_match_bits_t ignore_bits;
ptl_me_t me;
int ret;
ignore_bits = MTL_PORTALS4_CONTEXT_MASK | MTL_PORTALS4_SOURCE_MASK | MTL_PORTALS4_TAG_MASK;
me.start = block->start;
me.length = ompi_mtl_portals4.recv_short_size;
me.ct_handle = PTL_CT_NONE;
me.min_free = ompi_mtl_portals4.short_limit;
me.uid = ompi_mtl_portals4.uid;
me.options =
PTL_ME_OP_PUT |
PTL_ME_EVENT_COMM_DISABLE |
PTL_ME_MANAGE_LOCAL |
PTL_ME_MAY_ALIGN;
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 = match_bits;
me.ignore_bits = ignore_bits;
OPAL_THREAD_LOCK(&ompi_mtl_portals4.short_block_mutex);
block->status = BLOCK_STATUS_WAITING_LINK;
OPAL_THREAD_UNLOCK(&ompi_mtl_portals4.short_block_mutex);
ret = PtlMEAppend(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.recv_idx,
&me,
PTL_OVERFLOW_LIST,
&block->request,
&block->me_h);
if (OPAL_LIKELY(ret == PTL_OK)) {
ret = OMPI_SUCCESS;
} else {
ret = ompi_mtl_portals4_get_error(ret);
}
return ret;
}
static void big_meappend(void *buf, ptl_size_t left_to_send, MPIDI_VC_t *vc, ptl_match_bits_t match_bits, MPID_Request *sreq)
{
int i, ret, was_incomplete;
MPID_nem_ptl_vc_area *vc_ptl;
ptl_me_t me;
vc_ptl = VC_PTL(vc);
me.start = buf;
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 );
me.match_id = vc_ptl->id;
me.match_bits = match_bits;
me.ignore_bits = 0;
me.min_free = 0;
/* allocate enough handles to cover all get operations */
REQ_PTL(sreq)->get_me_p = MPIU_Malloc(sizeof(ptl_handle_me_t) *
((left_to_send / MPIDI_nem_ptl_ni_limits.max_msg_size) + 1));
/* queue up as many entries as necessary to describe the entire message */
for (i = 0; left_to_send > 0; i++) {
/* send up to the maximum allowed by the portals interface */
if (left_to_send > MPIDI_nem_ptl_ni_limits.max_msg_size)
me.length = MPIDI_nem_ptl_ni_limits.max_msg_size;
else
me.length = left_to_send;
ret = PtlMEAppend(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_get_pt, &me, PTL_PRIORITY_LIST, sreq,
&REQ_PTL(sreq)->get_me_p[i]);
DBG_MSG_MEAPPEND("CTL", vc->pg_rank, me, sreq);
MPIU_Assert(ret == 0);
/* increment the cc for each get operation */
MPIDI_CH3U_Request_increment_cc(sreq, &was_incomplete);
MPIU_Assert(was_incomplete);
/* account for what has been sent */
me.start = (char *)me.start + me.length;
left_to_send -= me.length;
}
}
static inline int
ompi_mtl_portals4_activate_block(ompi_mtl_portals4_recv_short_block_t *block)
{
ptl_match_bits_t match_bits = MTL_PORTALS4_SHORT_MSG;
ptl_match_bits_t ignore_bits;
ptl_me_t me;
int ret;
opal_list_remove_item(&ompi_mtl_portals4.waiting_recv_short_blocks, &block->base);
ignore_bits = MTL_PORTALS4_CONTEXT_MASK | MTL_PORTALS4_SOURCE_MASK | MTL_PORTALS4_TAG_MASK;
me.start = block->start;
me.length = ompi_mtl_portals4.recv_short_size;
me.ct_handle = PTL_CT_NONE;
me.min_free = ompi_mtl_portals4.eager_limit;
me.uid = ompi_mtl_portals4.uid;
me.options =
PTL_ME_OP_PUT |
PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_MANAGE_LOCAL |
PTL_ME_MAY_ALIGN;
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
me.match_bits = match_bits;
me.ignore_bits = ignore_bits;
ret = PtlMEAppend(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.recv_idx,
&me,
PTL_OVERFLOW_LIST,
&block->request,
&block->me_h);
if (OPAL_LIKELY(ret == PTL_OK)) {
ret = OMPI_SUCCESS;
opal_list_append(&ompi_mtl_portals4.active_recv_short_blocks,
&block->base);
} else {
ret = ompi_mtl_portals4_get_error(ret);
}
return ret;
}
int rptli_post_control_buffer(ptl_handle_ni_t ni_handle, ptl_pt_index_t pt,
ptl_handle_me_t * me_handle)
{
int ret;
ptl_me_t me;
ptl_process_t id;
MPIDI_STATE_DECL(MPID_STATE_RPTLI_POST_CONTROL_BUFFER);
MPIDI_FUNC_ENTER(MPID_STATE_RPTLI_POST_CONTROL_BUFFER);
id.phys.nid = PTL_NID_ANY;
id.phys.pid = PTL_PID_ANY;
me.start = NULL;
me.length = 0;
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);
me.match_id = id;
me.match_bits = 0;
me.ignore_bits = 0;
me.min_free = 0;
while (1) {
ret = PtlMEAppend(ni_handle, pt, &me, PTL_PRIORITY_LIST, NULL, me_handle);
if (ret != PTL_NO_SPACE)
break;
}
RPTLU_ERR_POP(ret, "Error appending empty buffer to priority list\n");
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_RPTLI_POST_CONTROL_BUFFER);
return ret;
fn_fail:
goto fn_exit;
}
static int append_overflow(int i)
{
int mpi_errno = MPI_SUCCESS;
int ret;
ptl_me_t me;
ptl_process_t id_any;
MPIDI_STATE_DECL(MPID_STATE_APPEND_OVERFLOW);
MPIDI_FUNC_ENTER(MPID_STATE_APPEND_OVERFLOW);
MPIU_Assert(i >= 0 && i < NUM_OVERFLOW_ME);
id_any.phys.pid = PTL_PID_ANY;
id_any.phys.nid = PTL_NID_ANY;
me.start = overflow_buf[i];
me.length = OVERFLOW_LENGTH;
me.ct_handle = PTL_CT_NONE;
me.uid = PTL_UID_ANY;
me.options = ( PTL_ME_OP_PUT | PTL_ME_MANAGE_LOCAL | PTL_ME_NO_TRUNCATE | PTL_ME_MAY_ALIGN |
PTL_ME_IS_ACCESSIBLE | PTL_ME_EVENT_LINK_DISABLE );
me.match_id = id_any;
me.match_bits = 0;
me.ignore_bits = ~((ptl_match_bits_t)0);
me.min_free = PTL_LARGE_THRESHOLD;
/* if there is no space to append the entry, process outstanding events and try again */
ret = PtlMEAppend(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_pt, &me, PTL_OVERFLOW_LIST, (void *)(size_t)i,
&overflow_me_handle[i]);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmeappend", "**ptlmeappend %s", MPID_nem_ptl_strerror(ret));
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_APPEND_OVERFLOW);
return mpi_errno;
fn_fail:
goto fn_exit;
}
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;
}
mca_btl_base_registration_handle_t *
mca_btl_portals4_register_mem(mca_btl_base_module_t *btl_base,
mca_btl_base_endpoint_t *endpoint,
void *base,
size_t size,
uint32_t flags)
{
struct mca_btl_portals4_module_t *portals4_btl = (struct mca_btl_portals4_module_t*) btl_base;
mca_btl_base_registration_handle_t *handle = NULL;
ptl_me_t me;
int ret;
handle = (mca_btl_base_registration_handle_t *)malloc(sizeof(mca_btl_base_registration_handle_t));
if (!handle) {
return NULL;
}
handle->key = OPAL_THREAD_ADD_FETCH64(&(portals4_btl->portals_rdma_key), 1);
handle->remote_offset = 0;
OPAL_OUTPUT_VERBOSE((90, opal_btl_base_framework.framework_output,
"mca_btl_portals4_register_mem NI=%d base=%p size=%ld handle=%p key=%ld flags=%d",
portals4_btl->interface_num, base, size, (void *)handle, handle->key, flags));
/* create a match entry */
me.start = base;
me.length = size;
me.ct_handle = PTL_CT_NONE;
me.min_free = 0;
me.uid = PTL_UID_ANY;
me.options = PTL_ME_OP_GET |
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 = endpoint->ptl_proc.rank;
} else {
me.match_id.phys.nid = endpoint->ptl_proc.phys.nid;
me.match_id.phys.pid = endpoint->ptl_proc.phys.pid;
}
me.match_bits = handle->key;
me.ignore_bits = BTL_PORTALS4_PROTOCOL_MASK |
BTL_PORTALS4_CONTEXT_MASK |
BTL_PORTALS4_SOURCE_MASK;
me.ignore_bits = 0;
ret = PtlMEAppend(portals4_btl->portals_ni_h,
portals4_btl->recv_idx,
&me,
PTL_PRIORITY_LIST,
handle,
&(handle->me_h));
if (PTL_OK != ret) {
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
OPAL_THREAD_ADD_FETCH32(&portals4_btl->portals_outstanding_ops, -1);
return NULL;
}
OPAL_OUTPUT_VERBOSE((90, opal_btl_base_framework.framework_output,
"PtlMEAppend (mca_btl_portals4_register_mem) handle=%p, me_h=%d start=%p length=%ld rank=%x nid=%x pid=%x match_bits=%lx\n",
(void *)handle, handle->me_h, me.start, me.length,
me.match_id.rank, me.match_id.phys.nid, me.match_id.phys.pid, me.match_bits));
return handle;
}
int
ompi_mtl_portals4_irecv(struct mca_mtl_base_module_t* mtl,
struct ompi_communicator_t *comm,
int src,
int tag,
struct opal_convertor_t *convertor,
mca_mtl_request_t *mtl_request)
{
ptl_match_bits_t match_bits, ignore_bits;
int ret = OMPI_SUCCESS;
ptl_process_t remote_proc;
mca_mtl_base_endpoint_t *endpoint = NULL;
ompi_mtl_portals4_recv_request_t *ptl_request =
(ompi_mtl_portals4_recv_request_t*) mtl_request;
void *start;
size_t length;
bool free_after;
ptl_me_t me;
if (MPI_ANY_SOURCE == src) {
remote_proc.phys.nid = PTL_NID_ANY;
remote_proc.phys.pid = PTL_PID_ANY;
} else {
ompi_proc_t* ompi_proc = ompi_comm_peer_lookup( comm, src );
endpoint = (mca_mtl_base_endpoint_t*) ompi_proc->proc_pml;
remote_proc = endpoint->ptl_proc;
}
MTL_PORTALS4_SET_RECV_BITS(match_bits, ignore_bits, comm->c_contextid,
src, tag);
ret = ompi_mtl_datatype_recv_buf(convertor, &start, &length, &free_after);
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_mtl_base_output,
"%s:%d: PtlMEAppend failed: %d",
__FILE__, __LINE__, ret);
return ret;
}
#if OPAL_ENABLE_DEBUG
ptl_request->opcount = ++ompi_mtl_portals4.recv_opcount;
ptl_request->hdr_data = 0;
#endif
ptl_request->super.event_callback = ompi_mtl_portals4_recv_progress;
ptl_request->buffer_ptr = (free_after) ? start : NULL;
ptl_request->convertor = convertor;
ptl_request->delivery_ptr = start;
ptl_request->delivery_len = length;
ptl_request->super.super.ompi_req->req_status.MPI_ERROR = OMPI_SUCCESS;
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_output,
"Recv %d from %x,%x of length %d (0x%lx, 0x%lx)\n",
ptl_request->opcount,
remote_proc.phys.nid, remote_proc.phys.pid,
(int)length, match_bits, ignore_bits));
me.start = start;
me.length = length;
me.ct_handle = PTL_CT_NONE;
me.min_free = 0;
me.uid = PTL_UID_ANY;
me.options = PTL_ME_OP_PUT | PTL_ME_USE_ONCE | PTL_ME_EVENT_UNLINK_DISABLE;
me.match_id = remote_proc;
me.match_bits = match_bits;
me.ignore_bits = ignore_bits;
ret = PtlMEAppend(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.send_idx,
&me,
PTL_PRIORITY_LIST,
ptl_request,
&ptl_request->me_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: PtlMEAppend failed: %d",
__FILE__, __LINE__, ret);
return ompi_mtl_portals4_get_error(ret);
}
return OMPI_SUCCESS;
}
int
ompi_mtl_portals4_irecv(struct mca_mtl_base_module_t* mtl,
struct ompi_communicator_t *comm,
int src,
int tag,
struct opal_convertor_t *convertor,
mca_mtl_request_t *mtl_request)
{
ptl_match_bits_t match_bits, ignore_bits;
int ret = OMPI_SUCCESS;
ptl_process_t remote_proc;
ompi_mtl_portals4_recv_request_t *ptl_request =
(ompi_mtl_portals4_recv_request_t*) mtl_request;
void *start;
size_t length;
bool free_after;
ptl_me_t me;
if (MPI_ANY_SOURCE == src) {
if (ompi_mtl_portals4.use_logical) {
remote_proc.rank = PTL_RANK_ANY;
} else {
remote_proc.phys.nid = PTL_NID_ANY;
remote_proc.phys.pid = PTL_PID_ANY;
}
} else if ((ompi_mtl_portals4.use_logical) && (MPI_COMM_WORLD == comm)) {
remote_proc.rank = src;
} else {
ompi_proc_t* ompi_proc = ompi_comm_peer_lookup( comm, src );
remote_proc = *((ptl_process_t*) ompi_mtl_portals4_get_endpoint (mtl, ompi_proc));
}
MTL_PORTALS4_SET_RECV_BITS(match_bits, ignore_bits, comm->c_contextid,
src, tag);
ret = ompi_mtl_datatype_recv_buf(convertor, &start, &length, &free_after);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
return ret;
}
ptl_request->super.type = portals4_req_recv;
ptl_request->super.event_callback = ompi_mtl_portals4_recv_progress;
#if OPAL_ENABLE_DEBUG
ptl_request->opcount = OPAL_THREAD_ADD64((int64_t*) &ompi_mtl_portals4.recv_opcount, 1);
ptl_request->hdr_data = 0;
#endif
ptl_request->buffer_ptr = (free_after) ? start : NULL;
ptl_request->convertor = convertor;
ptl_request->delivery_ptr = start;
ptl_request->delivery_len = length;
ptl_request->req_started = false;
ptl_request->super.super.ompi_req->req_status.MPI_ERROR = OMPI_SUCCESS;
ptl_request->pending_reply = 0;
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_framework.framework_output,
"Recv %lu from %x,%x of length %ld (0x%lx, 0x%lx, 0x%lx)\n",
ptl_request->opcount,
remote_proc.phys.nid, remote_proc.phys.pid,
(int64_t)length, match_bits, ignore_bits, (unsigned long) ptl_request));
me.start = start;
me.length = length;
me.ct_handle = PTL_CT_NONE;
me.min_free = 0;
me.uid = ompi_mtl_portals4.uid;
me.options =
PTL_ME_OP_PUT |
PTL_ME_USE_ONCE |
PTL_ME_EVENT_UNLINK_DISABLE;
if (length <= ompi_mtl_portals4.short_limit) {
me.options |= PTL_ME_EVENT_LINK_DISABLE;
}
me.match_id = remote_proc;
me.match_bits = match_bits;
me.ignore_bits = ignore_bits;
ret = PtlMEAppend(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.recv_idx,
&me,
PTL_PRIORITY_LIST,
ptl_request,
&ptl_request->me_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
if (NULL != ptl_request->buffer_ptr) free(ptl_request->buffer_ptr);
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d",
__FILE__, __LINE__, ret);
return ompi_mtl_portals4_get_error(ret);
}
/* if a long message, spin until we either have a comm event or a
link event, guaranteeing progress for long unexpected
messages. */
if (length > ompi_mtl_portals4.short_limit) {
while (true != ptl_request->req_started) {
ompi_mtl_portals4_progress();
}
}
return OMPI_SUCCESS;
}
static int
setup_sync_handles(struct ompi_communicator_t *comm,
ompi_coll_portals4_request_t *request,
mca_coll_portals4_module_t *portals4_module)
{
int ret, line;
ptl_me_t me;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
"coll:portals4:setup_sync_handles enter rank %d", request->u.scatter.my_rank));
/**********************************/
/* Setup Sync Handles */
/**********************************/
COLL_PORTALS4_SET_BITS(request->u.scatter.sync_match_bits, ompi_comm_get_cid(comm),
0, 1, COLL_PORTALS4_SCATTER, 0, request->u.scatter.coll_count);
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
"coll:portals4:setup_sync_handles rank(%d) sync_match_bits(0x%016lX)",
request->u.scatter.my_rank, request->u.scatter.sync_match_bits));
ret = PtlCTAlloc(mca_coll_portals4_component.ni_h,
&request->u.scatter.sync_cth);
if (PTL_OK != ret) {
ret = OMPI_ERR_TEMP_OUT_OF_RESOURCE;
line = __LINE__;
goto err_hdlr;
}
request->u.scatter.sync_mdh = mca_coll_portals4_component.zero_md_h;
me.start = NULL;
me.length = 0;
me.ct_handle = request->u.scatter.sync_cth;
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_LINK_DISABLE | PTL_ME_EVENT_UNLINK_DISABLE |
PTL_ME_EVENT_CT_COMM | PTL_ME_EVENT_CT_OVERFLOW;
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
me.match_bits = request->u.scatter.sync_match_bits;
me.ignore_bits = 0;
ret = PtlMEAppend(mca_coll_portals4_component.ni_h,
mca_coll_portals4_component.pt_idx,
&me,
PTL_PRIORITY_LIST,
NULL,
&request->u.scatter.sync_meh);
if (PTL_OK != ret) {
ret = OMPI_ERROR;
line = __LINE__;
goto err_hdlr;
}
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
"coll:portals4:setup_sync_handles exit rank %d", request->u.scatter.my_rank));
return OMPI_SUCCESS;
err_hdlr:
opal_output(ompi_coll_base_framework.framework_output,
"%s:%4d:%4d\tError occurred ret=%d, rank %2d",
__FILE__, __LINE__, line, ret, request->u.scatter.my_rank);
return ret;
}
int
ompi_mtl_portals4_flowctl_init(void)
{
ptl_me_t me;
int ret;
ompi_mtl_portals4.flowctl.flowctl_active = false;
OBJ_CONSTRUCT(&ompi_mtl_portals4.flowctl.pending_sends, opal_list_t);
OBJ_CONSTRUCT(&ompi_mtl_portals4.flowctl.pending_fl, opal_free_list_t);
opal_free_list_init(&ompi_mtl_portals4.flowctl.pending_fl,
sizeof(ompi_mtl_portals4_pending_request_t),
opal_cache_line_size,
OBJ_CLASS(ompi_mtl_portals4_pending_request_t),
0, 0, 1, -1, 1, NULL, 0, NULL, NULL, NULL);
ompi_mtl_portals4.flowctl.max_send_slots = (ompi_mtl_portals4.send_queue_size - 3) / 3;
ompi_mtl_portals4.flowctl.send_slots = ompi_mtl_portals4.flowctl.max_send_slots;
ompi_mtl_portals4.flowctl.alert_req.type = portals4_req_flowctl;
ompi_mtl_portals4.flowctl.alert_req.event_callback = flowctl_alert_callback;
ompi_mtl_portals4.flowctl.fanout_req.type = portals4_req_flowctl;
ompi_mtl_portals4.flowctl.fanout_req.event_callback = flowctl_fanout_callback;
ompi_mtl_portals4.flowctl.epoch_counter = -1;
ret = PtlPTAlloc(ompi_mtl_portals4.ni_h,
PTL_PT_ONLY_TRUNCATE,
ompi_mtl_portals4.send_eq_h,
REQ_FLOWCTL_TABLE_ID,
&ompi_mtl_portals4.flowctl_idx);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlPTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
ret = PtlCTAlloc(ompi_mtl_portals4.ni_h,
&ompi_mtl_portals4.flowctl.trigger_ct_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlCTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
/* everyone creates the trigger ME, even if the root may be the
only to use it */
me.start = NULL;
me.length = 0;
me.min_free = 0;
me.uid = ompi_mtl_portals4.uid;
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.ignore_bits = 0;
me.options = PTL_ME_OP_PUT |
PTL_ME_ACK_DISABLE |
PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE |
PTL_ME_EVENT_COMM_DISABLE |
PTL_ME_EVENT_CT_COMM;
me.ct_handle = ompi_mtl_portals4.flowctl.trigger_ct_h;
me.match_bits = MTL_PORTALS4_FLOWCTL_TRIGGER;
ret = PtlMEAppend(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.flowctl_idx,
&me,
PTL_PRIORITY_LIST,
NULL,
&ompi_mtl_portals4.flowctl.trigger_me_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
/* Alert CT/ME for broadcasting out alert when root receives a
trigger */
ret = PtlCTAlloc(ompi_mtl_portals4.ni_h,
&ompi_mtl_portals4.flowctl.alert_ct_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlCTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
me.options = PTL_ME_OP_PUT |
PTL_ME_ACK_DISABLE |
PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE |
PTL_ME_EVENT_CT_COMM;
me.ct_handle = ompi_mtl_portals4.flowctl.alert_ct_h;
me.match_bits = MTL_PORTALS4_FLOWCTL_ALERT;
ret = PtlMEAppend(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.flowctl_idx,
&me,
PTL_PRIORITY_LIST,
&ompi_mtl_portals4.flowctl.alert_req,
&ompi_mtl_portals4.flowctl.alert_me_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
/* Fanin CT/ME for receiving fan-in for restart */
ret = PtlCTAlloc(ompi_mtl_portals4.ni_h,
&ompi_mtl_portals4.flowctl.fanin_ct_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlCTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
me.options = PTL_ME_OP_PUT |
PTL_ME_ACK_DISABLE |
PTL_ME_EVENT_COMM_DISABLE |
PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE |
PTL_ME_EVENT_CT_COMM;
me.ct_handle = ompi_mtl_portals4.flowctl.fanin_ct_h;
me.match_bits = MTL_PORTALS4_FLOWCTL_FANIN;
ret = PtlMEAppend(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.flowctl_idx,
&me,
PTL_PRIORITY_LIST,
NULL,
&ompi_mtl_portals4.flowctl.fanin_me_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
/* Fan-out CT/ME for sending restart messages after fan-in */
ret = PtlCTAlloc(ompi_mtl_portals4.ni_h,
&ompi_mtl_portals4.flowctl.fanout_ct_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlCTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
me.options = PTL_ME_OP_PUT |
PTL_ME_ACK_DISABLE |
PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE |
PTL_ME_EVENT_CT_COMM;
me.ct_handle = ompi_mtl_portals4.flowctl.fanout_ct_h;
me.match_bits = MTL_PORTALS4_FLOWCTL_FANOUT;
ret = PtlMEAppend(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.flowctl_idx,
&me,
PTL_PRIORITY_LIST,
&ompi_mtl_portals4.flowctl.fanout_req,
&ompi_mtl_portals4.flowctl.fanout_me_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
ompi_mtl_portals4.flowctl.num_children = 0;
gettimeofday(&ompi_mtl_portals4.flowctl.tv, NULL);
ompi_mtl_portals4.flowctl.backoff_count = 0;
ret = OMPI_SUCCESS;
error:
return ret;
}
static inline int
ompi_mtl_portals4_long_isend(void *start, size_t length, int contextid, int tag,
int localrank,
ptl_process_t ptl_proc,
ompi_mtl_portals4_isend_request_t *ptl_request)
{
int ret;
ptl_match_bits_t match_bits;
ptl_me_t me;
ptl_hdr_data_t hdr_data;
ptl_size_t put_length;
MTL_PORTALS4_SET_SEND_BITS(match_bits, contextid, localrank, tag,
MTL_PORTALS4_LONG_MSG);
MTL_PORTALS4_SET_HDR_DATA(hdr_data, ptl_request->opcount, length, 0);
me.start = start;
me.length = length;
me.ct_handle = PTL_CT_NONE;
me.min_free = 0;
me.uid = ompi_mtl_portals4.uid;
me.options =
PTL_ME_OP_GET |
PTL_ME_USE_ONCE |
PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE;
me.match_id = ptl_proc;
me.match_bits = hdr_data;
me.ignore_bits = 0;
ret = PtlMEAppend(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.read_idx,
&me,
PTL_PRIORITY_LIST,
ptl_request,
&ptl_request->me_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d",
__FILE__, __LINE__, ret);
return ompi_mtl_portals4_get_error(ret);
}
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_framework.framework_output,
"Send %lu long send with hdr_data 0x%lx (0x%lx)",
ptl_request->opcount, hdr_data, match_bits));
put_length = (rndv == ompi_mtl_portals4.protocol) ?
(ptl_size_t) ompi_mtl_portals4.eager_limit : (ptl_size_t) length;
ret = PtlPut(ompi_mtl_portals4.send_md_h,
(ptl_size_t) start,
put_length,
PTL_ACK_REQ,
ptl_proc,
ompi_mtl_portals4.recv_idx,
match_bits,
0,
ptl_request,
hdr_data);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlPut failed: %d",
__FILE__, __LINE__, ret);
PtlMEUnlink(ptl_request->me_h);
return ompi_mtl_portals4_get_error(ret);
}
return OMPI_SUCCESS;
}
static inline int
ompi_mtl_portals4_short_isend(mca_pml_base_send_mode_t mode,
void *start, int length, int contextid, int tag,
int localrank,
ptl_process_t ptl_proc,
ompi_mtl_portals4_isend_request_t *ptl_request)
{
int ret;
ptl_match_bits_t match_bits;
ptl_me_t me;
ptl_hdr_data_t hdr_data;
MTL_PORTALS4_SET_SEND_BITS(match_bits, contextid, localrank, tag,
MTL_PORTALS4_SHORT_MSG);
MTL_PORTALS4_SET_HDR_DATA(hdr_data, ptl_request->opcount, length,
(MCA_PML_BASE_SEND_SYNCHRONOUS == mode) ? 1 : 0);
if (MCA_PML_BASE_SEND_SYNCHRONOUS == mode) {
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_USE_ONCE |
PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE;
me.match_id = ptl_proc;
me.match_bits = hdr_data;
me.ignore_bits = 0;
ret = PtlMEAppend(ompi_mtl_portals4.ni_h,
ompi_mtl_portals4.read_idx,
&me,
PTL_PRIORITY_LIST,
ptl_request,
&ptl_request->me_h);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d",
__FILE__, __LINE__, ret);
return ompi_mtl_portals4_get_error(ret);
}
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_framework.framework_output,
"Send %lu short sync send with hdr_data 0x%lx (0x%lx)",
ptl_request->opcount, hdr_data, match_bits));
} else {
ptl_request->event_count = 1;
ptl_request->me_h = PTL_INVALID_HANDLE;
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_framework.framework_output,
"Send %lu short send with hdr_data 0x%lx (0x%lx)",
ptl_request->opcount, hdr_data, match_bits));
}
OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_framework.framework_output,
"Send %lu, start: %p",
ptl_request->opcount, start));
ret = PtlPut(ompi_mtl_portals4.send_md_h,
(ptl_size_t) start,
length,
PTL_ACK_REQ,
ptl_proc,
ompi_mtl_portals4.recv_idx,
match_bits,
0,
ptl_request,
hdr_data);
if (OPAL_UNLIKELY(PTL_OK != ret)) {
opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
"%s:%d: PtlPut failed: %d",
__FILE__, __LINE__, ret);
if (MCA_PML_BASE_SEND_SYNCHRONOUS == mode) {
PtlMEUnlink(ptl_request->me_h);
}
return ompi_mtl_portals4_get_error(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;
}
/*
/!\ 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;
}
int
ompi_coll_portals4_barrier_intra(struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_portals4_module_t *portals4_module = (mca_coll_portals4_module_t*) module;
int ret, i, dim, hibit, mask, num_msgs;
int size = ompi_comm_size(comm);
int rank = ompi_comm_rank(comm);
ptl_ct_event_t ct;
ptl_handle_ct_t ct_h;
ptl_handle_me_t me_h;
ptl_me_t me;
size_t count;
ptl_match_bits_t match_bits;
ptl_handle_md_t md_h;
void *base;
ompi_coll_portals4_get_md(0, &md_h, &base);
count = opal_atomic_add_size_t(&portals4_module->barrier_count, 1);
ret = PtlCTAlloc(mca_coll_portals4_component.ni_h,
&ct_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlCTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
}
COLL_PORTALS4_SET_BITS(match_bits, ompi_comm_get_cid(comm),
0, COLL_PORTALS4_BARRIER, count);
/* Build "tree" out of hypercube */
dim = comm->c_cube_dim;
hibit = opal_hibit(rank, dim);
--dim;
/* receive space */
me.start = NULL;
me.length = 0;
me.ct_handle = ct_h;
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_LINK_DISABLE | PTL_ME_EVENT_UNLINK_DISABLE |
PTL_ME_EVENT_CT_COMM | PTL_ME_EVENT_CT_OVERFLOW;
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
me.match_bits = match_bits;
me.ignore_bits = 0;
ret = PtlMEAppend(mca_coll_portals4_component.ni_h,
mca_coll_portals4_component.pt_idx,
&me,
PTL_PRIORITY_LIST,
NULL,
&me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* calculate number of children to receive from */
num_msgs = ompi_coll_portals4_get_nchildren(dim + 1, hibit, rank, size);
/* send to parent when children have sent to us */
if (rank > 0) {
int parent = rank & ~(1 << hibit);
ret = PtlTriggeredPut(md_h,
0,
0,
PTL_NO_ACK_REQ,
ompi_coll_portals4_get_peer(comm, parent),
mca_coll_portals4_component.pt_idx,
match_bits,
0,
NULL,
0,
ct_h,
num_msgs);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlTriggeredPut failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* we'll need to wait for the parent response before the next set of comms */
num_msgs++;
}
/* send to children when parent (or all children if root) has sent to us */
for (i = hibit + 1, mask = 1 << i; i <= dim; ++i, mask <<= 1) {
int peer = rank | mask;
if (peer < size) {
ret = PtlTriggeredPut(md_h,
0,
0,
PTL_NO_ACK_REQ,
ompi_coll_portals4_get_peer(comm, peer),
mca_coll_portals4_component.pt_idx,
match_bits,
0,
NULL,
0,
ct_h,
num_msgs);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlTriggeredPut failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
}
}
/* Wait for all incoming messages */
ret = PtlCTWait(ct_h, num_msgs, &ct);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlCTWait failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* cleanup */
ret = PtlMEUnlink(me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlMEUnlink failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
ret = PtlCTFree(ct_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlCTFree failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
return OMPI_SUCCESS;
}
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;
}
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;
}
