static MPIR_Request * create_request(void * hdr, intptr_t hdr_sz,
size_t nb)
{
MPIR_Request * sreq;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CREATE_REQUEST);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CREATE_REQUEST);
sreq = MPIR_Request_create(MPIR_REQUEST_KIND__UNDEFINED);
/* --BEGIN ERROR HANDLING-- */
if (sreq == NULL)
return NULL;
/* --END ERROR HANDLING-- */
MPIR_Object_set_ref(sreq, 2);
sreq->kind = MPIR_REQUEST_KIND__SEND;
MPIR_Assert(hdr_sz == sizeof(MPIDI_CH3_Pkt_t));
sreq->dev.pending_pkt = *(MPIDI_CH3_Pkt_t *) hdr;
sreq->dev.iov[0].MPL_IOV_BUF =
(MPL_IOV_BUF_CAST)((char *) &sreq->dev.pending_pkt + nb);
sreq->dev.iov[0].MPL_IOV_LEN = hdr_sz - nb;
sreq->dev.iov_count = 1;
sreq->dev.OnDataAvail = 0;
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_CREATE_REQUEST);
return sreq;
}
int MPID_Rget_accumulate(const void *origin_addr, int origin_count,
MPI_Datatype origin_datatype, void *result_addr, int result_count,
MPI_Datatype result_datatype, int target_rank, MPI_Aint target_disp,
int target_count, MPI_Datatype target_datatype, MPI_Op op,
MPIR_Win * win_ptr, MPIR_Request ** request)
{
int mpi_errno = MPI_SUCCESS;
int dt_contig ATTRIBUTE((unused));
MPIR_Datatype*dtp;
MPI_Aint dt_true_lb ATTRIBUTE((unused));
intptr_t data_sz, trg_data_sz;
MPIR_Request *ureq;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPID_RGET_ACCUMULATE);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPID_RGET_ACCUMULATE);
/* request-based RMA operations are only valid within a passive epoch */
MPIR_ERR_CHKANDJUMP(win_ptr->states.access_state != MPIDI_RMA_PER_TARGET &&
win_ptr->states.access_state != MPIDI_RMA_LOCK_ALL_CALLED &&
win_ptr->states.access_state != MPIDI_RMA_LOCK_ALL_ISSUED &&
win_ptr->states.access_state != MPIDI_RMA_LOCK_ALL_GRANTED,
mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
/* Create user request, initially cc=1, ref=1 */
ureq = MPIR_Request_create(MPIR_REQUEST_KIND__RMA);
MPIR_ERR_CHKANDJUMP(ureq == NULL, mpi_errno, MPI_ERR_OTHER, "**nomemreq");
/* This request is referenced by user and ch3 by default. */
MPIR_Object_set_ref(ureq, 2);
/* Note that GACC is only a no-op if no data goes in both directions */
MPIDI_Datatype_get_info(origin_count, origin_datatype, dt_contig, data_sz, dtp, dt_true_lb);
MPIDI_Datatype_get_info(origin_count, origin_datatype, dt_contig, trg_data_sz, dtp, dt_true_lb);
/* Enqueue or perform the RMA operation */
if (target_rank != MPI_PROC_NULL && (data_sz != 0 || trg_data_sz != 0)) {
mpi_errno = MPIDI_CH3I_Get_accumulate(origin_addr, origin_count,
origin_datatype, result_addr,
result_count, result_datatype,
target_rank, target_disp,
target_count, target_datatype, op, win_ptr, ureq);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_POP(mpi_errno);
}
}
else {
mpi_errno = MPID_Request_complete(ureq);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_POP(mpi_errno);
}
}
*request = ureq;
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPID_RGET_ACCUMULATE);
return mpi_errno;
fn_fail:
goto fn_exit;
}
static MPIR_Request *create_request(MPL_IOV * iov, int iov_count, int iov_offset, size_t nb)
{
MPIR_Request *sreq;
int i;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CREATE_REQUEST);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CREATE_REQUEST);
sreq = MPIR_Request_create(MPIR_REQUEST_KIND__SEND);
/* --BEGIN ERROR HANDLING-- */
if (sreq == NULL)
return NULL;
/* --END ERROR HANDLING-- */
MPIR_Object_set_ref(sreq, 2);
for (i = 0; i < iov_count; i++) {
sreq->dev.iov[i] = iov[i];
}
if (iov_offset == 0) {
MPIR_Assert(iov[0].MPL_IOV_LEN == sizeof(MPIDI_CH3_Pkt_t));
sreq->dev.pending_pkt = *(MPIDI_CH3_Pkt_t *) iov[0].MPL_IOV_BUF;
sreq->dev.iov[0].MPL_IOV_BUF = (MPL_IOV_BUF_CAST) & sreq->dev.pending_pkt;
}
sreq->dev.iov[iov_offset].MPL_IOV_BUF =
(MPL_IOV_BUF_CAST) ((char *) sreq->dev.iov[iov_offset].MPL_IOV_BUF + nb);
sreq->dev.iov[iov_offset].MPL_IOV_LEN -= nb;
sreq->dev.iov_count = iov_count;
sreq->dev.OnDataAvail = 0;
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_CREATE_REQUEST);
return sreq;
}
static
int MPID_PSP_persistent_init(const void *buf, MPI_Aint count, MPI_Datatype datatype, int rank, int tag,
MPIR_Comm *comm, int context_offset, MPIR_Request **request,
int (*call)(const void * buf, MPI_Aint count, MPI_Datatype datatype, int rank,
int tag, struct MPIR_Comm * comm, int context_offset, MPIR_Request ** request),
MPIR_Request_kind_t type)
{
MPIR_Request *req;
struct MPID_DEV_Request_persistent *preq;
/*
printf("#%d ps--- %s() called\n", MPIDI_Process.my_pg_rank, __func__);
printf("#%d buf %p, count %d, datatype 0x%0x, rank %d, tag %d, comm %p, off %d\n",
MPIDI_Process.my_pg_rank, buf, count, datatype, rank, tag, comm, context_offset);
printf("#%d ctx.id %d ctx.rank %d, ctx.name %s\n",
MPIDI_Process.my_pg_rank, comm->context_id, comm->rank, comm->name);
*/
req = MPIR_Request_create(type);
if (unlikely(!req)) goto err_request_recv_create;
req->comm = comm;
MPIR_Comm_add_ref(comm);
req->u.persist.real_request = NULL;
MPIDI_PSP_Request_set_completed(req); /* an inactive persistent request is a completed request. */
preq = &req->dev.kind.persistent;
preq->buf = (void *)buf;
preq->count = count;
preq->datatype = datatype;
MPID_PSP_Datatype_add_ref(preq->datatype);
preq->rank = rank;
preq->tag = tag;
preq->comm = comm;
// MPIR_Comm_add_ref(comm);
preq->context_offset = context_offset;
preq->call = call;
*request = req;
return MPI_SUCCESS;
/* --- */
err_request_recv_create:
return MPI_ERR_NO_MEM;
}
int MPII_Genutil_sched_start(MPII_Genutil_sched_t * sched, MPIR_Comm * comm, MPIR_Request ** req)
{
int mpi_errno = MPI_SUCCESS;
int is_complete;
int made_progress;
MPIR_Request *reqp;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPII_GENUTIL_SCHED_START);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPII_GENUTIL_SCHED_START);
/* Create a request */
reqp = MPIR_Request_create(MPIR_REQUEST_KIND__COLL);
if (!reqp)
MPIR_ERR_SETANDJUMP(mpi_errno, MPI_ERR_OTHER, "**nomem");
*req = reqp;
MPIR_Request_add_ref(reqp);
/* Make some progress */
mpi_errno = MPII_Genutil_sched_poke(sched, &is_complete, &made_progress);
if (is_complete) {
MPID_Request_complete(reqp);
goto fn_exit;
}
/* Enqueue schedule and activate progress hook if not already activated */
reqp->u.nbc.coll.sched = (void *) sched;
if (coll_queue.head == NULL)
MPID_Progress_activate_hook(MPII_Genutil_progress_hook_id);
DL_APPEND(coll_queue.head, &(reqp->u.nbc.coll));
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPII_GENUTIL_SCHED_START);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_Put_generic(const void *origin_addr, int origin_count, MPI_Datatype origin_datatype,
int target_rank, MPI_Aint target_disp, int target_count,
MPI_Datatype target_datatype, MPIR_Win *win_ptr, MPIR_Request **request)
{
int mpi_error = MPI_SUCCESS;
MPID_PSP_Datatype_info dt_info;
MPID_PSP_packed_msg_t msg;
MPID_Win_rank_info *ri = win_ptr->rank_info + target_rank;
char *target_buf;
#if 0
fprintf(stderr, "int MPID_Put(origin_addr: %p, origin_count: %d, origin_datatype: %08x,"
" target_rank: %d, target_disp: %d, target_count: %d, target_datatype: %08x,"
" *win_ptr: %p)\n",
origin_addr, origin_count, origin_datatype,
target_rank, target_disp, target_count,
target_datatype, win_ptr);
#endif
/* Datatype */
MPID_PSP_Datatype_get_info(target_datatype, &dt_info);
if(request) {
*request = MPIR_Request_create(MPIR_REQUEST_KIND__SEND);
(*request)->comm = win_ptr->comm_ptr;
MPIR_Comm_add_ref(win_ptr->comm_ptr);
}
if (unlikely(target_rank == MPI_PROC_NULL)) {
goto fn_completed;
}
/* Request-based RMA operations are only valid within a passive target epoch! */
if(request && win_ptr->epoch_state != MPID_PSP_EPOCH_LOCK && win_ptr->epoch_state != MPID_PSP_EPOCH_LOCK_ALL) {
mpi_error = MPI_ERR_RMA_SYNC;
goto err_sync_rma;
}
/* Check that we are within an access/exposure epoch: */
if (win_ptr->epoch_state == MPID_PSP_EPOCH_NONE) {
mpi_error = MPI_ERR_RMA_SYNC;
goto err_sync_rma;
}
/* Track access epoch state: */
if (win_ptr->epoch_state == MPID_PSP_EPOCH_FENCE_ISSUED) {
win_ptr->epoch_state = MPID_PSP_EPOCH_FENCE;
}
/* If the put is a local operation, do it here */
if (target_rank == win_ptr->rank || win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) {
void *base;
int disp_unit;
if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) {
MPID_PSP_shm_rma_get_base(win_ptr, target_rank, &disp_unit, &base);
}
else {
base = win_ptr->base;
disp_unit = win_ptr->disp_unit;
}
mpi_error = MPIR_Localcopy(origin_addr, origin_count, origin_datatype,
(char *) base + disp_unit * target_disp,
target_count, target_datatype);
if (mpi_error) {
goto err_local_copy;
}
goto fn_completed;
}
/* Data */
mpi_error = MPID_PSP_packed_msg_prepare(origin_addr, origin_count, origin_datatype, &msg);
if (unlikely(mpi_error != MPI_SUCCESS)) goto err_create_packed_msg;
MPID_PSP_packed_msg_pack(origin_addr, origin_count, origin_datatype, &msg);
target_buf = (char *) ri->base_addr + ri->disp_unit * target_disp;
if (0 && MPID_PSP_Datatype_is_contig(&dt_info)) { /* ToDo: reenable pscom buildin rma_write */
/* Contig message. Use pscom buildin rma */
pscom_request_t *req = pscom_request_create(0, 0);
req->data_len = msg.msg_sz;
req->data = msg.msg;
req->connection = ri->con;
/* ToDo: need a new io_done. inside io_done, call MPID_PSP_packed_msg_cleanup(msg)!!! */
req->ops.io_done = pscom_request_free;
req->xheader.rma_write.dest = target_buf;
pscom_post_rma_write(req);
/* win_ptr->rma_puts_accs[target_rank]++; / ToDo: Howto receive this? */
} else {
unsigned int encode_dt_size = MPID_PSP_Datatype_get_size(&dt_info);
unsigned int xheader_len = sizeof(MPID_PSCOM_XHeader_Rma_put_t) + encode_dt_size;
pscom_request_t *req = pscom_request_create(xheader_len, sizeof(pscom_request_put_send_t));
MPID_PSCOM_XHeader_Rma_put_t *xheader = &req->xheader.user.put;
/* encoded datatype too large for xheader? */
assert(xheader_len < (1<<(8*sizeof(((struct PSCOM_header_net*)0)->xheader_len))));
req->user->type.put_send.msg = msg;
req->user->type.put_send.win_ptr = win_ptr;
MPID_PSP_Datatype_encode(&dt_info, &xheader->encoded_type);
xheader->common.tag = 0;
xheader->common.context_id = 0;
xheader->common.type = MPID_PSP_MSGTYPE_RMA_PUT;
xheader->common._reserved_ = 0;
xheader->common.src_rank = win_ptr->rank;
/* xheader->target_disp = target_disp; */
xheader->target_count = target_count;
xheader->target_buf = target_buf;
/* xheader->epoch = ri->epoch_origin; */
xheader->win_ptr = ri->win_ptr; /* remote win_ptr */
req->xheader_len = xheader_len;
req->data = msg.msg;
req->data_len = msg.msg_sz;
req->ops.io_done = rma_put_done;
req->user->type.put_send.target_rank = target_rank;
req->connection = ri->con;
win_ptr->rma_local_pending_cnt++;
win_ptr->rma_local_pending_rank[target_rank]++;
win_ptr->rma_puts_accs[target_rank]++;
if(request) {
MPIR_Request *mpid_req = *request;
/* TODO: Use a new and 'put_send'-dedicated MPID_DEV_Request_create() */
/* instead of allocating and overloading a common send request. */
pscom_request_free(mpid_req->dev.kind.common.pscom_req);
mpid_req->dev.kind.common.pscom_req = req;
MPIR_Request_add_ref(mpid_req);
req->user->type.put_send.mpid_req = mpid_req;
} else {
req->user->type.put_send.mpid_req = NULL;
}
pscom_post_send(req);
}
fn_exit:
return MPI_SUCCESS;
fn_completed:
if(request) {
MPIDI_PSP_Request_set_completed(*request);
}
return MPI_SUCCESS;
/* --- */
err_exit:
if(request) {
MPIDI_PSP_Request_set_completed(*request);
MPIR_Request_free(*request);
}
return mpi_error;
/* --- */
err_create_packed_msg:
goto err_exit;
err_local_copy:
goto err_exit;
err_sync_rma:
goto err_exit;
}
int MPIR_Init_thread(int *argc, char ***argv, int required, int *provided)
{
int mpi_errno = MPI_SUCCESS;
int has_args;
int has_env;
int thread_provided = 0;
int exit_init_cs_on_failure = 0;
MPIR_Info *info_ptr;
#if defined(MPICH_IS_THREADED)
bool cs_initialized = false;
#endif
/* The threading library must be initialized at the very beginning because
* it manages all synchronization objects (e.g., mutexes) that will be
* initialized later */
{
int thread_err;
MPL_thread_init(&thread_err);
if (thread_err)
goto fn_fail;
}
#ifdef HAVE_HWLOC
MPIR_Process.bindset = hwloc_bitmap_alloc();
hwloc_topology_init(&MPIR_Process.hwloc_topology);
MPIR_Process.bindset_is_valid = 0;
hwloc_topology_set_io_types_filter(MPIR_Process.hwloc_topology, HWLOC_TYPE_FILTER_KEEP_ALL);
if (!hwloc_topology_load(MPIR_Process.hwloc_topology)) {
MPIR_Process.bindset_is_valid =
!hwloc_get_proc_cpubind(MPIR_Process.hwloc_topology, getpid(), MPIR_Process.bindset,
HWLOC_CPUBIND_PROCESS);
}
#endif
#ifdef HAVE_NETLOC
MPIR_Process.network_attr.u.tree.node_levels = NULL;
MPIR_Process.network_attr.network_endpoint = NULL;
MPIR_Process.netloc_topology = NULL;
MPIR_Process.network_attr.type = MPIR_NETLOC_NETWORK_TYPE__INVALID;
if (strlen(MPIR_CVAR_NETLOC_NODE_FILE)) {
mpi_errno =
netloc_parse_topology(&MPIR_Process.netloc_topology, MPIR_CVAR_NETLOC_NODE_FILE);
if (mpi_errno == NETLOC_SUCCESS) {
MPIR_Netloc_parse_topology(MPIR_Process.netloc_topology, &MPIR_Process.network_attr);
}
}
#endif
/* For any code in the device that wants to check for runtime
* decisions on the value of isThreaded, set a provisional
* value here. We could let the MPID_Init routine override this */
#if defined MPICH_IS_THREADED
MPIR_ThreadInfo.isThreaded = required == MPI_THREAD_MULTIPLE;
#endif /* MPICH_IS_THREADED */
#if defined(MPICH_IS_THREADED)
mpi_errno = thread_cs_init();
cs_initialized = true;
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
#endif
/* FIXME: Move to os-dependent interface? */
#ifdef HAVE_WINDOWS_H
/* prevent the process from bringing up an error message window if mpich
* asserts */
_CrtSetReportMode(_CRT_ASSERT, _CRTDBG_MODE_FILE);
_CrtSetReportFile(_CRT_ASSERT, _CRTDBG_FILE_STDERR);
_CrtSetReportHook2(_CRT_RPTHOOK_INSTALL, assert_hook);
#ifdef _WIN64
{
/* FIXME: (Windows) This severly degrades performance but fixes alignment
* issues with the datatype code. */
/* Prevent misaligned faults on Win64 machines */
UINT mode, old_mode;
old_mode = SetErrorMode(SEM_NOALIGNMENTFAULTEXCEPT);
mode = old_mode | SEM_NOALIGNMENTFAULTEXCEPT;
SetErrorMode(mode);
}
#endif
#endif
/* We need this inorder to implement IS_THREAD_MAIN */
#if (MPICH_THREAD_LEVEL >= MPI_THREAD_SERIALIZED) && defined(MPICH_IS_THREADED)
{
MPID_Thread_self(&MPIR_ThreadInfo.master_thread);
}
#endif
#ifdef HAVE_ERROR_CHECKING
/* Because the PARAM system has not been initialized, temporarily
* uncondtionally enable error checks. Once the PARAM system is
* initialized, this may be reset */
MPIR_Process.do_error_checks = 1;
#else
MPIR_Process.do_error_checks = 0;
#endif
/* Initialize necessary subsystems and setup the predefined attribute
* values. Subsystems may change these values. */
MPIR_Process.attrs.appnum = -1;
MPIR_Process.attrs.host = MPI_PROC_NULL;
MPIR_Process.attrs.io = MPI_PROC_NULL;
MPIR_Process.attrs.lastusedcode = MPI_ERR_LASTCODE;
MPIR_Process.attrs.universe = MPIR_UNIVERSE_SIZE_NOT_SET;
MPIR_Process.attrs.wtime_is_global = 0;
/* Set the functions used to duplicate attributes. These are
* when the first corresponding keyval is created */
MPIR_Process.attr_dup = 0;
MPIR_Process.attr_free = 0;
#ifdef HAVE_CXX_BINDING
/* Set the functions used to call functions in the C++ binding
* for reductions and attribute operations. These are null
* until a C++ operation is defined. This allows the C code
* that implements these operations to not invoke a C++ code
* directly, which may force the inclusion of symbols known only
* to the C++ compiler (e.g., under more non-GNU compilers, including
* Solaris and IRIX). */
MPIR_Process.cxx_call_op_fn = 0;
#endif
#ifdef HAVE_F08_BINDING
MPIR_C_MPI_UNWEIGHTED = MPI_UNWEIGHTED;
MPIR_C_MPI_WEIGHTS_EMPTY = MPI_WEIGHTS_EMPTY;
#endif
/* This allows the device to select an alternative function for
* dimsCreate */
MPIR_Process.dimsCreate = 0;
/* "Allocate" from the reserved space for builtin communicators and
* (partially) initialize predefined communicators. comm_parent is
* intially NULL and will be allocated by the device if the process group
* was started using one of the MPI_Comm_spawn functions. */
MPIR_Process.comm_world = MPIR_Comm_builtin + 0;
MPII_Comm_init(MPIR_Process.comm_world);
MPIR_Process.comm_world->handle = MPI_COMM_WORLD;
MPIR_Process.comm_world->context_id = 0 << MPIR_CONTEXT_PREFIX_SHIFT;
MPIR_Process.comm_world->recvcontext_id = 0 << MPIR_CONTEXT_PREFIX_SHIFT;
MPIR_Process.comm_world->comm_kind = MPIR_COMM_KIND__INTRACOMM;
/* This initialization of the comm name could be done only when
* comm_get_name is called */
MPL_strncpy(MPIR_Process.comm_world->name, "MPI_COMM_WORLD", MPI_MAX_OBJECT_NAME);
MPIR_Process.comm_self = MPIR_Comm_builtin + 1;
MPII_Comm_init(MPIR_Process.comm_self);
MPIR_Process.comm_self->handle = MPI_COMM_SELF;
MPIR_Process.comm_self->context_id = 1 << MPIR_CONTEXT_PREFIX_SHIFT;
MPIR_Process.comm_self->recvcontext_id = 1 << MPIR_CONTEXT_PREFIX_SHIFT;
MPIR_Process.comm_self->comm_kind = MPIR_COMM_KIND__INTRACOMM;
MPL_strncpy(MPIR_Process.comm_self->name, "MPI_COMM_SELF", MPI_MAX_OBJECT_NAME);
#ifdef MPID_NEEDS_ICOMM_WORLD
MPIR_Process.icomm_world = MPIR_Comm_builtin + 2;
MPII_Comm_init(MPIR_Process.icomm_world);
MPIR_Process.icomm_world->handle = MPIR_ICOMM_WORLD;
MPIR_Process.icomm_world->context_id = 2 << MPIR_CONTEXT_PREFIX_SHIFT;
MPIR_Process.icomm_world->recvcontext_id = 2 << MPIR_CONTEXT_PREFIX_SHIFT;
MPIR_Process.icomm_world->comm_kind = MPIR_COMM_KIND__INTRACOMM;
MPL_strncpy(MPIR_Process.icomm_world->name, "MPI_ICOMM_WORLD", MPI_MAX_OBJECT_NAME);
/* Note that these communicators are not ready for use - MPID_Init
* will setup self and world, and icomm_world if it desires it. */
#endif
MPIR_Process.comm_parent = NULL;
/* Setup the initial communicator list in case we have
* enabled the debugger message-queue interface */
MPII_COMML_REMEMBER(MPIR_Process.comm_world);
MPII_COMML_REMEMBER(MPIR_Process.comm_self);
/* MPIU_Timer_pre_init(); */
/* Wait for debugger to attach if requested. */
if (MPIR_CVAR_DEBUG_HOLD) {
volatile int hold = 1;
while (hold)
#ifdef HAVE_USLEEP
usleep(100);
#endif
;
}
#if defined(HAVE_ERROR_CHECKING) && (HAVE_ERROR_CHECKING == MPID_ERROR_LEVEL_RUNTIME)
MPIR_Process.do_error_checks = MPIR_CVAR_ERROR_CHECKING;
#endif
/* define MPI as initialized so that we can use MPI functions within
* MPID_Init if necessary */
OPA_store_int(&MPIR_Process.mpich_state, MPICH_MPI_STATE__IN_INIT);
/* We can't acquire any critical sections until this point. Any
* earlier the basic data structures haven't been initialized */
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
exit_init_cs_on_failure = 1;
/* create MPI_INFO_NULL object */
/* FIXME: Currently this info object is empty, we need to add data to this
* as defined by the standard. */
info_ptr = MPIR_Info_builtin + 1;
info_ptr->handle = MPI_INFO_ENV;
MPIR_Object_set_ref(info_ptr, 1);
info_ptr->next = NULL;
info_ptr->key = NULL;
info_ptr->value = NULL;
#ifdef USE_MEMORY_TRACING
MPL_trinit();
#endif
/* Set the number of tag bits. The device may override this value. */
MPIR_Process.tag_bits = MPIR_TAG_BITS_DEFAULT;
/* Create complete request to return in the event of immediately complete
* operations. Use a SEND request to cover all possible use-cases. */
MPIR_Process.lw_req = MPIR_Request_create(MPIR_REQUEST_KIND__SEND);
MPIR_ERR_CHKANDSTMT(MPIR_Process.lw_req == NULL, mpi_errno, MPIX_ERR_NOREQ, goto fn_fail,
"**nomemreq");
MPIR_cc_set(&MPIR_Process.lw_req->cc, 0);
mpi_errno = MPID_Init(argc, argv, required, &thread_provided, &has_args, &has_env);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* Initialize collectives infrastructure */
mpi_errno = MPII_Coll_init();
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* Set tag_ub as function of tag_bits set by the device */
MPIR_Process.attrs.tag_ub = MPIR_TAG_USABLE_BITS;
/* Assert: tag_ub should be a power of 2 minus 1 */
MPIR_Assert(((unsigned) MPIR_Process.
attrs.tag_ub & ((unsigned) MPIR_Process.attrs.tag_ub + 1)) == 0);
/* Assert: tag_ub is at least the minimum asked for in the MPI spec */
MPIR_Assert(MPIR_Process.attrs.tag_ub >= 32767);
/* Capture the level of thread support provided */
MPIR_ThreadInfo.thread_provided = thread_provided;
if (provided)
*provided = thread_provided;
#if defined MPICH_IS_THREADED
MPIR_ThreadInfo.isThreaded = (thread_provided == MPI_THREAD_MULTIPLE);
#endif /* MPICH_IS_THREADED */
/* FIXME: Define these in the interface. Does Timer init belong here? */
MPII_Timer_init(MPIR_Process.comm_world->rank, MPIR_Process.comm_world->local_size);
#ifdef USE_MEMORY_TRACING
#ifdef MPICH_IS_THREADED
MPL_trconfig(MPIR_Process.comm_world->rank, MPIR_ThreadInfo.isThreaded);
#else
MPL_trconfig(MPIR_Process.comm_world->rank, 0);
#endif
/* Indicate that we are near the end of the init step; memory
* allocated already will have an id of zero; this helps
* separate memory leaks in the initialization code from
* leaks in the "active" code */
#endif
#ifdef MPL_USE_DBG_LOGGING
/* FIXME: This is a hack to handle the common case of two worlds.
* If the parent comm is not NULL, we always give the world number
* as "1" (false). */
#ifdef MPICH_IS_THREADED
MPL_dbg_init(argc, argv, has_args, has_env,
MPIR_Process.comm_parent != NULL, MPIR_Process.comm_world->rank,
MPIR_ThreadInfo.isThreaded);
#else
MPL_dbg_init(argc, argv, has_args, has_env,
MPIR_Process.comm_parent != NULL, MPIR_Process.comm_world->rank, 0);
#endif
MPIR_DBG_INIT = MPL_dbg_class_alloc("INIT", "init");
MPIR_DBG_PT2PT = MPL_dbg_class_alloc("PT2PT", "pt2pt");
MPIR_DBG_THREAD = MPL_dbg_class_alloc("THREAD", "thread");
MPIR_DBG_DATATYPE = MPL_dbg_class_alloc("DATATYPE", "datatype");
MPIR_DBG_HANDLE = MPL_dbg_class_alloc("HANDLE", "handle");
MPIR_DBG_COMM = MPL_dbg_class_alloc("COMM", "comm");
MPIR_DBG_BSEND = MPL_dbg_class_alloc("BSEND", "bsend");
MPIR_DBG_ERRHAND = MPL_dbg_class_alloc("ERRHAND", "errhand");
MPIR_DBG_OTHER = MPL_dbg_class_alloc("OTHER", "other");
MPIR_DBG_REQUEST = MPL_dbg_class_alloc("REQUEST", "request");
MPIR_DBG_COLL = MPL_dbg_class_alloc("COLL", "coll");
MPIR_DBG_ASSERT = MPL_dbg_class_alloc("ASSERT", "assert");
MPIR_DBG_STRING = MPL_dbg_class_alloc("STRING", "string");
#endif
/* Initialize the C versions of the Fortran link-time constants.
*
* We now initialize the Fortran symbols from within the Fortran
* interface in the routine that first needs the symbols.
* This fixes a problem with symbols added by a Fortran compiler that
* are not part of the C runtime environment (the Portland group
* compilers would do this)
*/
#if defined(HAVE_FORTRAN_BINDING) && defined(HAVE_MPI_F_INIT_WORKS_WITH_C)
mpirinitf_();
#endif
/* FIXME: Does this need to come before the call to MPID_InitComplete?
* For some debugger support, MPII_Wait_for_debugger may want to use
* MPI communication routines to collect information for the debugger */
#ifdef HAVE_DEBUGGER_SUPPORT
MPII_Wait_for_debugger();
#endif
/* Let the device know that the rest of the init process is completed */
if (mpi_errno == MPI_SUCCESS)
mpi_errno = MPID_InitCompleted();
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
/* Make fields of MPIR_Process global visible and set mpich_state
* atomically so that MPI_Initialized() etc. are thread safe */
OPA_write_barrier();
OPA_store_int(&MPIR_Process.mpich_state, MPICH_MPI_STATE__POST_INIT);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
/* signal to error handling routines that core services are unavailable */
OPA_store_int(&MPIR_Process.mpich_state, MPICH_MPI_STATE__PRE_INIT);
if (exit_init_cs_on_failure) {
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
}
#if defined(MPICH_IS_THREADED)
if (cs_initialized) {
MPIR_Thread_CS_Finalize();
}
#endif
return mpi_errno;
/* --END ERROR HANDLING-- */
}
int MPIDI_CH3_iStartMsgv(MPIDI_VC_t * vc, MPL_IOV * iov, int n_iov, MPIR_Request ** sreq_ptr)
{
MPIR_Request *sreq = NULL;
MPIDI_CH3I_VC *vcch = &vc->ch;
int mpi_errno = MPI_SUCCESS;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_CH3_ISTARTMSGV);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_CH3_ISTARTMSGV);
MPIR_Assert(n_iov <= MPL_IOV_LIMIT);
/* The SOCK channel uses a fixed length header, the size of which is the
* maximum of all possible packet headers */
iov[0].MPL_IOV_LEN = sizeof(MPIDI_CH3_Pkt_t);
MPL_DBG_STMT(MPIDI_CH3_DBG_CHANNEL, VERBOSE,
MPIDI_DBG_Print_packet((MPIDI_CH3_Pkt_t *) iov[0].MPL_IOV_BUF));
if (vcch->state == MPIDI_CH3I_VC_STATE_CONNECTED) { /* MT */
/* Connection already formed. If send queue is empty attempt to send
* data, queuing any unsent data. */
if (MPIDI_CH3I_SendQ_empty(vcch)) { /* MT */
int rc;
size_t nb;
MPL_DBG_MSG(MPIDI_CH3_DBG_CHANNEL, VERBOSE, "send queue empty, attempting to write");
MPL_DBG_PKT(vcch->conn, (MPIDI_CH3_Pkt_t *) iov[0].MPL_IOV_BUF, "isend");
/* MT - need some signalling to lock down our right to use the
* channel, thus insuring that the progress engine does
* also try to write */
rc = MPIDI_CH3I_Sock_writev(vcch->sock, iov, n_iov, &nb);
if (rc == MPI_SUCCESS) {
int offset = 0;
MPL_DBG_MSG_D(MPIDI_CH3_DBG_CHANNEL, VERBOSE,
"wrote %ld bytes", (unsigned long) nb);
while (offset < n_iov) {
if (nb >= (int) iov[offset].MPL_IOV_LEN) {
nb -= iov[offset].MPL_IOV_LEN;
offset++;
} else {
MPL_DBG_MSG(MPIDI_CH3_DBG_CHANNEL, VERBOSE,
"partial write, request enqueued at head");
sreq = create_request(iov, n_iov, offset, nb);
if (sreq == NULL) {
MPIR_ERR_SETANDJUMP(mpi_errno, MPI_ERR_OTHER, "**nomem");
}
MPIDI_CH3I_SendQ_enqueue_head(vcch, sreq);
MPL_DBG_MSG_FMT(MPIDI_CH3_DBG_CHANNEL, VERBOSE,
(MPL_DBG_FDEST, "posting writev, vc=0x%p, sreq=0x%08x", vc,
sreq->handle));
vcch->conn->send_active = sreq;
mpi_errno =
MPIDI_CH3I_Sock_post_writev(vcch->conn->sock, sreq->dev.iov + offset,
sreq->dev.iov_count - offset, NULL);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS) {
mpi_errno =
MPIR_Err_create_code(mpi_errno, MPIR_ERR_FATAL, __func__, __LINE__,
MPI_ERR_OTHER, "**ch3|sock|postwrite",
"ch3|sock|postwrite %p %p %p", sreq,
vcch->conn, vc);
}
/* --END ERROR HANDLING-- */
break;
}
}
if (offset == n_iov) {
MPL_DBG_MSG(MPIDI_CH3_DBG_CHANNEL, VERBOSE, "entire write complete");
}
}
/* --BEGIN ERROR HANDLING-- */
else {
MPL_DBG_MSG_D(MPIDI_CH3_DBG_CHANNEL, TYPICAL,
"ERROR - MPIDI_CH3I_Sock_writev failed, rc=%d", rc);
sreq = MPIR_Request_create(MPIR_REQUEST_KIND__SEND);
if (sreq == NULL) {
MPIR_ERR_SETANDJUMP(mpi_errno, MPI_ERR_OTHER, "**nomem");
}
MPIR_cc_set(&(sreq->cc), 0);
sreq->status.MPI_ERROR = MPIR_Err_create_code(rc,
MPIR_ERR_RECOVERABLE, __func__,
__LINE__, MPI_ERR_INTERN,
"**ch3|sock|writefailed",
"**ch3|sock|writefailed %d", rc);
/* Make sure that the caller sees this error */
mpi_errno = sreq->status.MPI_ERROR;
}
/* --END ERROR HANDLING-- */
} else {
MPL_DBG_MSG(MPIDI_CH3_DBG_CHANNEL, VERBOSE, "send in progress, request enqueued");
sreq = create_request(iov, n_iov, 0, 0);
if (sreq == NULL) {
MPIR_ERR_SETANDJUMP(mpi_errno, MPI_ERR_OTHER, "**nomem");
}
MPIDI_CH3I_SendQ_enqueue(vcch, sreq);
}
} else if (vcch->state == MPIDI_CH3I_VC_STATE_CONNECTING) {
MPL_DBG_VCUSE(vc, "connecting. enqueuing request");
/* queue the data so it can be sent after the connection is formed */
sreq = create_request(iov, n_iov, 0, 0);
if (sreq == NULL) {
MPIR_ERR_SETANDJUMP(mpi_errno, MPI_ERR_OTHER, "**nomem");
}
MPIDI_CH3I_SendQ_enqueue(vcch, sreq);
} else if (vcch->state == MPIDI_CH3I_VC_STATE_UNCONNECTED) {
MPL_DBG_VCUSE(vc, "unconnected. posting connect and enqueuing request");
/* queue the data so it can be sent after the connection is formed */
sreq = create_request(iov, n_iov, 0, 0);
if (sreq == NULL) {
MPIR_ERR_SETANDJUMP(mpi_errno, MPI_ERR_OTHER, "**nomem");
}
MPIDI_CH3I_SendQ_enqueue(vcch, sreq);
/* Form a new connection */
MPIDI_CH3I_VC_post_connect(vc);
} else if (vcch->state != MPIDI_CH3I_VC_STATE_FAILED) {
/* Unable to send data at the moment, so queue it for later */
MPL_DBG_VCUSE(vc, "forming connection, request enqueued");
sreq = create_request(iov, n_iov, 0, 0);
if (sreq == NULL) {
MPIR_ERR_SETANDJUMP(mpi_errno, MPI_ERR_OTHER, "**nomem");
}
MPIDI_CH3I_SendQ_enqueue(vcch, sreq);
}
/* --BEGIN ERROR HANDLING-- */
else {
/* Connection failed, so allocate a request and return an error. */
MPL_DBG_VCUSE(vc, "ERROR - connection failed");
sreq = MPIR_Request_create(MPIR_REQUEST_KIND__SEND);
if (sreq == NULL) {
MPIR_ERR_SETANDJUMP(mpi_errno, MPI_ERR_OTHER, "**nomem");
}
MPIR_cc_set(&(sreq->cc), 0);
sreq->status.MPI_ERROR = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE, __func__, __LINE__,
MPI_ERR_INTERN, "**ch3|sock|connectionfailed",
0);
/* Make sure that the caller sees this error */
mpi_errno = sreq->status.MPI_ERROR;
}
/* --END ERROR HANDLING-- */
fn_fail:
*sreq_ptr = sreq;
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_CH3_ISTARTMSGV);
return mpi_errno;
}
int MPIDI_CH3_iStartMsg(MPIDI_VC_t * vc, void * hdr, intptr_t hdr_sz,
MPIR_Request ** sreq_ptr)
{
MPIR_Request * sreq = NULL;
MPIDI_CH3I_VC *vcch = &vc->ch;
int mpi_errno = MPI_SUCCESS;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_CH3_ISTARTMSG);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_CH3_ISTARTMSG);
MPIR_Assert( hdr_sz <= sizeof(MPIDI_CH3_Pkt_t));
/* The SOCK channel uses a fixed length header, the size of which is the
maximum of all possible packet headers */
hdr_sz = sizeof(MPIDI_CH3_Pkt_t);
MPL_DBG_STMT(MPIDI_CH3_DBG_CHANNEL,VERBOSE,
MPIDI_DBG_Print_packet((MPIDI_CH3_Pkt_t*)hdr));
if (vcch->state == MPIDI_CH3I_VC_STATE_CONNECTED) /* MT */
{
/* Connection already formed. If send queue is empty attempt to send
data, queuing any unsent data. */
if (MPIDI_CH3I_SendQ_empty(vcch)) /* MT */
{
size_t nb;
int rc;
MPL_DBG_MSG(MPIDI_CH3_DBG_CHANNEL,VERBOSE,
"send queue empty, attempting to write");
MPL_DBG_PKT(vcch->conn,hdr,"istartmsg");
/* MT: need some signalling to lock down our right to use the
channel, thus insuring that the progress engine does
not also try to write */
rc = MPIDI_CH3I_Sock_write(vcch->sock, hdr, hdr_sz, &nb);
if (rc == MPI_SUCCESS)
{
MPL_DBG_MSG_D(MPIDI_CH3_DBG_CHANNEL,VERBOSE,
"wrote %ld bytes", (unsigned long) nb);
if (nb == hdr_sz)
{
MPL_DBG_MSG_D(MPIDI_CH3_DBG_CHANNEL,VERBOSE,
"entire write complete, %" PRIdPTR " bytes", nb);
/* done. get us out of here as quickly as possible. */
}
else
{
MPL_DBG_MSG_D(MPIDI_CH3_DBG_CHANNEL,VERBOSE,
"partial write of %" PRIdPTR " bytes, request enqueued at head", nb);
sreq = create_request(hdr, hdr_sz, nb);
if (!sreq) {
MPIR_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomem");
}
MPIDI_CH3I_SendQ_enqueue_head(vcch, sreq);
MPL_DBG_MSG_FMT(MPIDI_CH3_DBG_CHANNEL,VERBOSE,
(MPL_DBG_FDEST,"posting write, vc=0x%p, sreq=0x%08x", vc, sreq->handle));
vcch->conn->send_active = sreq;
mpi_errno = MPIDI_CH3I_Sock_post_write(vcch->conn->sock, sreq->dev.iov[0].MPL_IOV_BUF,
sreq->dev.iov[0].MPL_IOV_LEN, sreq->dev.iov[0].MPL_IOV_LEN, NULL);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER,
"**ch3|sock|postwrite", "ch3|sock|postwrite %p %p %p",
sreq, vcch->conn, vc);
goto fn_fail;
}
/* --END ERROR HANDLING-- */
}
}
/* --BEGIN ERROR HANDLING-- */
else
{
MPL_DBG_MSG_D(MPIDI_CH3_DBG_CHANNEL,TYPICAL,
"ERROR - MPIDI_CH3I_Sock_write failed, rc=%d", rc);
sreq = MPIR_Request_create(MPIR_REQUEST_KIND__UNDEFINED);
if (!sreq) {
MPIR_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomem");
}
sreq->kind = MPIR_REQUEST_KIND__SEND;
MPIR_cc_set(&(sreq->cc), 0);
sreq->status.MPI_ERROR = MPIR_Err_create_code( rc,
MPIR_ERR_RECOVERABLE, FCNAME, __LINE__,
MPI_ERR_INTERN, "**ch3|sock|writefailed",
"**ch3|sock|writefailed %d", rc );
/* Make sure that the caller sees this error */
mpi_errno = sreq->status.MPI_ERROR;
}
/* --END ERROR HANDLING-- */
}
else
{
MPL_DBG_MSG(MPIDI_CH3_DBG_CHANNEL,VERBOSE,
"send in progress, request enqueued");
sreq = create_request(hdr, hdr_sz, 0);
if (!sreq) {
MPIR_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomem");
}
MPIDI_CH3I_SendQ_enqueue(vcch, sreq);
}
}
else if (vcch->state == MPIDI_CH3I_VC_STATE_CONNECTING) /* MT */
{
MPL_DBG_VCUSE(vc,
"connecteding. enqueuing request");
/* queue the data so it can be sent after the connection is formed */
sreq = create_request(hdr, hdr_sz, 0);
if (!sreq) {
MPIR_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomem");
}
MPIDI_CH3I_SendQ_enqueue(vcch, sreq);
}
else if (vcch->state == MPIDI_CH3I_VC_STATE_UNCONNECTED) /* MT */
{
MPL_DBG_VCUSE(vc,
"unconnected. posting connect and enqueuing request");
/* queue the data so it can be sent after the connection is formed */
sreq = create_request(hdr, hdr_sz, 0);
if (!sreq) {
MPIR_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomem");
}
MPIDI_CH3I_SendQ_enqueue(vcch, sreq);
/* Form a new connection */
MPIDI_CH3I_VC_post_connect(vc);
}
else if (vcch->state != MPIDI_CH3I_VC_STATE_FAILED)
{
/* Unable to send data at the moment, so queue it for later */
MPL_DBG_VCUSE(vc,"forming connection, request enqueued");
sreq = create_request(hdr, hdr_sz, 0);
if (!sreq) {
MPIR_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomem");
}
MPIDI_CH3I_SendQ_enqueue(vcch, sreq);
}
/* --BEGIN ERROR HANDLING-- */
else
{
/* Connection failed, so allocate a request and return an error. */
MPL_DBG_VCUSE(vc,"ERROR - connection failed");
sreq = MPIR_Request_create(MPIR_REQUEST_KIND__UNDEFINED);
if (!sreq) {
MPIR_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomem");
}
sreq->kind = MPIR_REQUEST_KIND__SEND;
MPIR_cc_set(&sreq->cc, 0);
sreq->status.MPI_ERROR = MPIR_Err_create_code( MPI_SUCCESS,
MPIR_ERR_RECOVERABLE, FCNAME, __LINE__,
MPI_ERR_INTERN, "**ch3|sock|connectionfailed",0 );
/* Make sure that the caller sees this error */
mpi_errno = sreq->status.MPI_ERROR;
}
/* --END ERROR HANDLING-- */
fn_fail:
*sreq_ptr = sreq;
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_CH3_ISTARTMSG);
return mpi_errno;
}
int MPID_nem_send_iov(MPIDI_VC_t *vc, MPIR_Request **sreq_ptr, MPL_IOV *iov, int n_iov)
{
int mpi_errno = MPI_SUCCESS;
intptr_t data_sz;
int i;
int iov_data_copied;
MPIR_Request *sreq = *sreq_ptr;
MPL_IOV *data_iov = &iov[1]; /* iov of just the data, not the header */
int data_n_iov = n_iov - 1;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPID_NEM_SEND_IOV);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPID_NEM_SEND_IOV);
if (*sreq_ptr == NULL)
{
/* create a request */
sreq = MPIR_Request_create(MPIR_REQUEST_KIND__UNDEFINED);
MPIR_Assert(sreq != NULL);
MPIR_Object_set_ref(sreq, 2);
sreq->kind = MPIR_REQUEST_KIND__SEND;
sreq->dev.OnDataAvail = 0;
}
data_sz = 0;
for (i = 0; i < data_n_iov; ++i)
data_sz += data_iov[i].MPL_IOV_LEN;
if (!MPIDI_Request_get_srbuf_flag(sreq))
{
MPIDI_CH3U_SRBuf_alloc(sreq, data_sz);
/* --BEGIN ERROR HANDLING-- */
if (sreq->dev.tmpbuf_sz == 0)
{
MPL_DBG_MSG(MPIDI_CH3_DBG_CHANNEL,TYPICAL,"SRBuf allocation failure");
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL,
FCNAME, __LINE__, MPI_ERR_OTHER, "**nomem", 0);
sreq->status.MPI_ERROR = mpi_errno;
goto fn_exit;
}
/* --END ERROR HANDLING-- */
}
MPIR_Assert(sreq->dev.tmpbuf_sz >= data_sz);
iov_data_copied = 0;
for (i = 0; i < data_n_iov; ++i) {
MPIR_Memcpy((char*) sreq->dev.tmpbuf + iov_data_copied, data_iov[i].MPL_IOV_BUF, data_iov[i].MPL_IOV_LEN);
iov_data_copied += data_iov[i].MPL_IOV_LEN;
}
mpi_errno = vc->ch.iSendContig(vc, sreq, iov[0].MPL_IOV_BUF, iov[0].MPL_IOV_LEN, sreq->dev.tmpbuf, data_sz);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
*sreq_ptr = sreq;
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPID_NEM_SEND_IOV);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_nem_ptl_improbe(MPIDI_VC_t *vc, int source, int tag, MPIR_Comm *comm, int context_offset, int *flag,
MPIR_Request **message, MPI_Status *status)
{
int mpi_errno = MPI_SUCCESS;
MPID_nem_ptl_vc_area *const vc_ptl = VC_PTL(vc);
int ret;
ptl_process_t id_any;
ptl_me_t me;
MPIR_Request *req;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPID_NEM_PTL_IMPROBE);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPID_NEM_PTL_IMPROBE);
id_any.phys.nid = PTL_NID_ANY;
id_any.phys.pid = PTL_PID_ANY;
/* create a request */
req = MPIR_Request_create(MPIR_REQUEST_KIND__MPROBE);
MPID_nem_ptl_init_req(req);
MPIR_ERR_CHKANDJUMP1(!req, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPIR_Request_create");
MPIR_Object_set_ref(req, 2); /* 1 ref for progress engine and 1 ref for us */
REQ_PTL(req)->event_handler = handle_mprobe;
/* create a dummy ME to use for searching the list */
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_USE_ONCE );
me.min_free = 0;
me.match_bits = NPTL_MATCH(tag, comm->context_id + context_offset, source);
if (source == MPI_ANY_SOURCE)
me.match_id = id_any;
else {
if (!vc_ptl->id_initialized) {
mpi_errno = MPID_nem_ptl_init_id(vc);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
me.match_id = vc_ptl->id;
}
if (tag == MPI_ANY_TAG)
me.ignore_bits = NPTL_MATCH_IGNORE_ANY_TAG;
else
me.ignore_bits = NPTL_MATCH_IGNORE;
/* submit a search request */
ret = PtlMESearch(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_pt, &me, PTL_SEARCH_DELETE, req);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmesearch", "**ptlmesearch %s", MPID_nem_ptl_strerror(ret));
DBG_MSG_MESearch("REG", vc ? vc->pg_rank : 0, me, req);
/* wait for search request to complete */
do {
mpi_errno = MPID_nem_ptl_poll(FALSE);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
} while (!MPIR_Request_is_complete(req));
*flag = REQ_PTL(req)->found;
if (*flag) {
req->comm = comm;
MPIR_Comm_add_ref(comm);
MPIR_Request_extract_status(req, status);
*message = req;
}
else {
MPIR_Request_free(req);
}
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPID_NEM_PTL_IMPROBE);
return mpi_errno;
fn_fail:
goto fn_exit;
}
{
int ret, mpi_errno = MPI_SUCCESS;
MPIDI_nem_ptl_pkt_cancel_send_req_t *req_pkt = (MPIDI_nem_ptl_pkt_cancel_send_req_t *)pkt;
MPID_PKT_DECL_CAST(upkt, MPIDI_nem_ptl_pkt_cancel_send_resp_t, resp_pkt);
MPIR_Request *search_req, *resp_req;
ptl_me_t me;
MPID_nem_ptl_vc_area *const vc_ptl = VC_PTL(vc);
MPL_DBG_MSG_FMT(MPIDI_CH3_DBG_OTHER,VERBOSE,(MPL_DBG_FDEST,
"received cancel send req pkt, sreq=0x%08x, rank=%d, tag=%d, context=%d",
req_pkt->sender_req_id, req_pkt->match.parts.rank,
req_pkt->match.parts.tag, req_pkt->match.parts.context_id));
/* create a dummy request and search for the message */
/* create a request */
search_req = MPIR_Request_create(MPIR_REQUEST_KIND__MPROBE);
MPID_nem_ptl_init_req(search_req);
MPIR_ERR_CHKANDJUMP1(!search_req, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPIR_Request_create");
MPIR_Object_set_ref(search_req, 2); /* 1 ref for progress engine and 1 ref for us */
/* create a dummy ME to use for searching the list */
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_USE_ONCE );
me.min_free = 0;
me.match_bits = NPTL_MATCH(req_pkt->match.parts.tag, req_pkt->match.parts.context_id, req_pkt->match.parts.rank);
me.match_id = vc_ptl->id;
me.ignore_bits = NPTL_MATCH_IGNORE;
