static int barrier_smp_intra(MPID_Comm *comm_ptr, mpir_errflag_t *errflag)
{
int mpi_errno=MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPIU_Assert(MPIR_CVAR_ENABLE_SMP_COLLECTIVES && MPIR_CVAR_ENABLE_SMP_BARRIER &&
MPIR_Comm_is_node_aware(comm_ptr));
/* do the intranode barrier on all nodes */
if (comm_ptr->node_comm != NULL)
{
mpi_errno = MPIR_Barrier_impl(comm_ptr->node_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
/* do the barrier across roots of all nodes */
if (comm_ptr->node_roots_comm != NULL) {
mpi_errno = MPIR_Barrier_impl(comm_ptr->node_roots_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
/* release the local processes on each node with a 1-byte
broadcast (0-byte broadcast just returns without doing
anything) */
if (comm_ptr->node_comm != NULL)
{
int i=0;
mpi_errno = MPIR_Bcast_impl(&i, 1, MPI_BYTE, 0, comm_ptr->node_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
fn_exit:
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIU_ERR_SET(mpi_errno, *errflag, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
/**
* \brief MPI-PAMI glue for MPI_Win_create function
*
* Create a window object. Allocates a MPID_Win object and initializes it,
* then allocates the collective info array, initalizes our entry, and
* performs an Allgather to distribute/collect the rest of the array entries.
*
* ON first call, initializes (registers) protocol objects for locking,
* get, and send operations to message layer. Also creates datatype to
* represent the rma_sends element of the collective info array,
* used later to synchronize epoch end events.
*
* \param[in] base Local window buffer
* \param[in] size Local window size
* \param[in] disp_unit Displacement unit size
* \param[in] info Window hints (not used)
* \param[in] comm_ptr Communicator
* \param[out] win_ptr Window
* \return MPI_SUCCESS, MPI_ERR_OTHER, or error returned from
* MPI_Comm_dup or MPI_Allgather.
*/
int
MPID_Win_create(void * base,
MPI_Aint size,
int disp_unit,
MPID_Info * info,
MPID_Comm * comm_ptr,
MPID_Win ** win_ptr)
{
int mpi_errno = MPI_SUCCESS;
int rc = MPI_SUCCESS;
MPID_Win *win;
size_t rank;
MPIDI_Win_info *winfo;
rc=MPIDI_Win_init(size,disp_unit,win_ptr, info, comm_ptr, MPI_WIN_FLAVOR_CREATE, MPI_WIN_UNIFIED);
win = *win_ptr;
win->base = base;
rank = comm_ptr->rank;
winfo = &win->mpid.info[rank];
winfo->base_addr = base;
winfo->win = win;
winfo->disp_unit = disp_unit;
rc= MPIDI_Win_allgather(size,win_ptr);
if (rc != MPI_SUCCESS)
return rc;
mpi_errno = MPIR_Barrier_impl(comm_ptr, &mpi_errno);
return mpi_errno;
}
int
MPID_Win_set_info(MPID_Win *win, MPID_Info *info)
{
int mpi_errno = MPI_SUCCESS;
mpi_errno = MPIDI_Win_set_info(win, info);
MPID_assert(mpi_errno == MPI_SUCCESS);
mpi_errno = MPIR_Barrier_impl(win->comm_ptr, &mpi_errno);
return mpi_errno;
}
/**
* \brief MPI-PAMI glue for MPI_Win_allocate function
*
* Create a window object. Allocates a MPID_Win object and initializes it,
* then allocates the collective info array, initalizes our entry, and
* performs an Allgather to distribute/collect the rest of the array entries.
* On each process, it allocates memory of at least size bytes, returns a
* pointer to it, and returns a window object that can be used by all processes
* in comm to * perform RMA operations. The returned memory consists of size
* bytes local to each process, starting at address base_ptr and is associated
* with the window as if the user called 'MPI_Win_create' on existing memory.
* The size argument may be different at each process and size = 0 is valid;
* however, a library might allocate and expose more memory in order to create
* a fast, globally symmetric allocation.
* Input Parameters:
* \param[in] size size of window in bytes (nonnegative integer)
* \param[in] disp_unit local unit size for displacements, in bytes (positive integer)
* \param[in] info info argument (handle))
* \param[in] comm_ptr Communicator (handle)
* \param[out] base_ptr - base address of the window in local memory
* \param[out] win_ptr window object returned by the call (handle)
* \return MPI_SUCCESS, MPI_ERR_ARG, MPI_ERR_COMM, MPI_ERR_INFO. MPI_ERR_OTHER,
* MPI_ERR_SIZE
*/
int
MPID_Win_allocate(MPI_Aint size,
int disp_unit,
MPID_Info * info,
MPID_Comm * comm_ptr,
void *base_ptr,
MPID_Win ** win_ptr)
{
int mpi_errno = MPI_SUCCESS;
int rc = MPI_SUCCESS;
mpir_errflag_t errflag = MPIR_ERR_NONE;
void *baseP;
static char FCNAME[] = "MPID_Win_allocate";
MPIDI_Win_info *winfo;
MPID_Win *win;
int rank;
rc=MPIDI_Win_init(size,disp_unit,win_ptr, info, comm_ptr, MPI_WIN_FLAVOR_ALLOCATE, MPI_WIN_UNIFIED);
win = *win_ptr;
if (size > 0) {
baseP = MPIU_Malloc(size);
#ifndef MPIDI_NO_ASSERT
MPID_assert(baseP != NULL);
#else
MPIU_ERR_CHKANDJUMP((baseP == NULL), mpi_errno, MPI_ERR_BUFFER, "**bufnull");
#endif
} else if (size == 0) {
baseP = NULL;
} else {
MPIU_ERR_CHKANDSTMT(size >=0 , mpi_errno, MPI_ERR_SIZE,
return mpi_errno, "**rmasize");
}
win->base = baseP;
rank = comm_ptr->rank;
winfo = &win->mpid.info[rank];
winfo->base_addr = baseP;
winfo->win = win;
winfo->disp_unit = disp_unit;
rc= MPIDI_Win_allgather(size,win_ptr);
if (rc != MPI_SUCCESS)
return rc;
*(void**) base_ptr = (void *) win->base;
mpi_errno = MPIR_Barrier_impl(comm_ptr, &errflag);
fn_fail:
return mpi_errno;
}
int
MPID_Win_fence(int assert,
MPID_Win *win)
{
int mpi_errno = MPI_SUCCESS;
struct MPIDI_Win_sync* sync = &win->mpid.sync;
MPID_PROGRESS_WAIT_WHILE(sync->total != sync->complete);
sync->total = 0;
sync->started = 0;
sync->complete = 0;
mpi_errno = MPIR_Barrier_impl(win->comm_ptr, &mpi_errno);
return mpi_errno;
}
int
MPID_Win_free(MPID_Win **win_ptr)
{
int mpi_errno = MPI_SUCCESS;
MPID_Win *win = *win_ptr;
size_t rank = win->comm_ptr->rank;
mpir_errflag_t errflag = MPIR_ERR_NONE;
if(win->mpid.sync.origin_epoch_type != win->mpid.sync.target_epoch_type ||
(win->mpid.sync.origin_epoch_type != MPID_EPOTYPE_NONE &&
win->mpid.sync.origin_epoch_type != MPID_EPOTYPE_REFENCE)) {
MPIU_ERR_SETANDSTMT(mpi_errno, MPI_ERR_RMA_SYNC, return mpi_errno, "**rmasync");
}
mpi_errno = MPIR_Barrier_impl(win->comm_ptr, &errflag);
MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**mpi_bcast");
if (win->create_flavor == MPI_WIN_FLAVOR_SHARED)
mpi_errno=MPIDI_SHM_Win_free(win_ptr);
if (win->create_flavor == MPI_WIN_FLAVOR_ALLOCATE)
MPIU_Free(win->base);
struct MPIDI_Win_info *winfo = &win->mpid.info[rank];
#ifdef USE_PAMI_RDMA
if (win->size != 0)
{
pami_result_t rc;
static int barrier_smp_intra(MPID_Comm *comm_ptr, MPIR_Errflag_t *errflag)
{
int mpi_errno=MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPIU_Assert(MPIR_CVAR_ENABLE_SMP_COLLECTIVES && MPIR_CVAR_ENABLE_SMP_BARRIER &&
MPIR_Comm_is_node_aware(comm_ptr));
#if defined(FINEGRAIN_MPI)
int colocated_size = -1;
int colocated_sense = -1;
/* do barrier on osproc_colocated_comm */
if (comm_ptr->osproc_colocated_comm != NULL)
{
colocated_size = comm_ptr->osproc_colocated_comm->local_size;
MPIU_Assert( (comm_ptr->osproc_colocated_comm->co_shared_vars != NULL) && (comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars != NULL) );
MPIU_Assert(colocated_size > 1 );
colocated_sense = comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_signal;
if( comm_ptr->osproc_colocated_comm->rank != 0 ) { /* non-leader */
(comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_counter)++;
if (comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_counter == (colocated_size-1)){ /* excluding the leader */
comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->leader_signal = 1;
}
while(comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_signal == colocated_sense) {
FG_Yield();
}
}
else { /* leader */
while(comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->leader_signal == 0) {
FG_Yield();
}
}
#if 0 /* Non-optimized version */
mpi_errno = MPIR_Barrier_impl(comm_ptr->osproc_colocated_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
#endif
}
#endif
/* do the intranode barrier on all nodes */
if (comm_ptr->node_comm != NULL)
{
mpi_errno = MPIR_Barrier_impl(comm_ptr->node_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
/* do the barrier across roots of all nodes */
if (comm_ptr->node_roots_comm != NULL) {
mpi_errno = MPIR_Barrier_impl(comm_ptr->node_roots_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
/* release the local processes on each node with a 1-byte
broadcast (0-byte broadcast just returns without doing
anything) */
if (comm_ptr->node_comm != NULL)
{
int i=0;
mpi_errno = MPIR_Bcast_impl(&i, 1, MPI_BYTE, 0, comm_ptr->node_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
#if defined(FINEGRAIN_MPI)
if (comm_ptr->osproc_colocated_comm != NULL)
{
if (comm_ptr->osproc_colocated_comm->rank == 0) { /* leader */
comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->leader_signal = 0;
comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_counter = 0;
comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_signal = 1 - comm_ptr->osproc_colocated_comm->co_shared_vars->co_barrier_vars->coproclet_signal;
}
#if 0 /* Non-optimized version */
/* release the colocated processes in each OS-process with a 1-byte
broadcast (0-byte broadcast just returns without doing
anything) */
int i=0;
mpi_errno = MPIR_Bcast_impl(&i, 1, MPI_BYTE, 0, comm_ptr->osproc_colocated_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
#endif
}
#endif
fn_exit:
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIR_ERR_SET(mpi_errno, *errflag, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
/*@
MPI_Barrier - Blocks until all processes in the communicator have
reached this routine.
Input Parameters:
. comm - communicator (handle)
Notes:
Blocks the caller until all processes in the communicator have called it;
that is, the call returns at any process only after all members of the
communicator have entered the call.
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_COMM
@*/
int MPI_Barrier( MPI_Comm comm )
{
int mpi_errno = MPI_SUCCESS;
MPID_Comm *comm_ptr = NULL;
MPIR_Errflag_t errflag = MPIR_ERR_NONE;
MPID_MPI_STATE_DECL(MPID_STATE_MPI_BARRIER);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPID_MPI_COLL_FUNC_ENTER(MPID_STATE_MPI_BARRIER);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* Convert MPI object handles to object pointers */
MPID_Comm_get_ptr( comm, comm_ptr );
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
/* Validate communicator */
MPID_Comm_valid_ptr( comm_ptr, mpi_errno, FALSE );
if (mpi_errno) goto fn_fail;
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPIR_Barrier_impl(comm_ptr, &errflag);
if (mpi_errno) goto fn_fail;
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_COLL_FUNC_EXIT(MPID_STATE_MPI_BARRIER);
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER,
"**mpi_barrier", "**mpi_barrier %C", comm);
}
# endif
mpi_errno = MPIR_Err_return_comm( comm_ptr, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/**
* \brief Shut down the system
*
* At this time, no attempt is made to free memory being used for MPI structures.
* \return MPI_SUCCESS
*/
int MPID_Finalize()
{
pami_result_t rc;
int mpierrno = MPI_SUCCESS;
mpir_errflag_t errflag=MPIR_ERR_NONE;
MPIR_Barrier_impl(MPIR_Process.comm_world, &errflag);
#ifdef MPIDI_STATISTICS
if (MPIDI_Process.mp_statistics) {
MPIDI_print_statistics();
}
MPIDI_close_pe_extension();
#endif
#ifdef DYNAMIC_TASKING
mpidi_finalized = 1;
if(mpidi_dynamic_tasking) {
/* Tell the process group code that we're done with the process groups.
This will notify PMI (with PMI_Finalize) if necessary. It
also frees all PG structures, including the PG for COMM_WORLD, whose
pointer is also saved in MPIDI_Process.my_pg */
mpierrno = MPIDI_PG_Finalize();
if (mpierrno) {
TRACE_ERR("MPIDI_PG_Finalize returned with mpierrno=%d\n", mpierrno);
}
MPIDI_FreeParentPort();
}
if(_conn_info_list)
MPIU_Free(_conn_info_list);
MPIDI_free_all_tranid_node();
#endif
/* ------------------------- */
/* shutdown request queues */
/* ------------------------- */
MPIDI_Recvq_finalize();
PAMIX_Finalize(MPIDI_Client);
#ifdef MPID_NEEDS_ICOMM_WORLD
MPIR_Comm_release_always(MPIR_Process.icomm_world, 0);
#endif
MPIR_Comm_release_always(MPIR_Process.comm_self,0);
MPIR_Comm_release_always(MPIR_Process.comm_world,0);
rc = PAMI_Context_destroyv(MPIDI_Context, MPIDI_Process.avail_contexts);
MPID_assert_always(rc == PAMI_SUCCESS);
rc = PAMI_Client_destroy(&MPIDI_Client);
MPID_assert_always(rc == PAMI_SUCCESS);
#ifdef MPIDI_TRACE
{ int i;
for (i=0; i< MPIDI_Process.numTasks; i++) {
if (MPIDI_Trace_buf[i].R)
MPIU_Free(MPIDI_Trace_buf[i].R);
if (MPIDI_Trace_buf[i].PR)
MPIU_Free(MPIDI_Trace_buf[i].PR);
if (MPIDI_Trace_buf[i].S)
MPIU_Free(MPIDI_Trace_buf[i].S);
}
}
MPIU_Free(MPIDI_Trace_buf);
#endif
#ifdef OUT_OF_ORDER_HANDLING
MPIU_Free(MPIDI_In_cntr);
MPIU_Free(MPIDI_Out_cntr);
#endif
if (TOKEN_FLOW_CONTROL_ON)
{
#if TOKEN_FLOW_CONTROL
extern char *EagerLimit;
if (EagerLimit) MPIU_Free(EagerLimit);
MPIU_Free(MPIDI_Token_cntr);
MPIDI_close_mm();
#else
MPID_assert_always(0);
#endif
}
return MPI_SUCCESS;
}
/* This function produces topology aware trees for reduction and broadcasts, with different
* K values. This is a heavy-weight function as it allocates shared memory, generates topology
* information, builds a package-level tree (for package leaders), and a per-package tree.
* These are combined in shared memory for other ranks to read out from.
* */
int MPIDI_SHM_topology_tree_init(MPIR_Comm * comm_ptr, int root, int bcast_k,
MPIR_Treealgo_tree_t * bcast_tree, int *bcast_topotree_fail,
int reduce_k, MPIR_Treealgo_tree_t * reduce_tree,
int *reduce_topotree_fail, MPIR_Errflag_t * errflag)
{
int *shared_region;
MPL_shm_hnd_t fd;
int num_ranks, rank;
int mpi_errno = MPI_SUCCESS, mpi_errno_ret = MPI_SUCCESS;
size_t shm_size;
int **bind_map = NULL;
int *max_entries_per_level = NULL;
int **ranks_per_package = NULL;
int *package_ctr = NULL;
size_t topo_depth = 0;
int package_level = 0, i, max_ranks_per_package = 0;
bool mapfail_flag = false;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_TOPOLOGY_TREE_INIT);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_SHM_TOPOLOGY_TREE_INIT);
num_ranks = MPIR_Comm_size(comm_ptr);
rank = MPIR_Comm_rank(comm_ptr);
/* Calculate the size of shared memory that would be needed */
shm_size = sizeof(int) * 5 * num_ranks + num_ranks * sizeof(cpu_set_t);
/* STEP 1. Create shared memory region for exchanging topology information (root only) */
mpi_errno = MPIDIU_allocate_shm_segment(comm_ptr, shm_size, &fd, (void **) &shared_region,
&mapfail_flag);
if (mpi_errno || mapfail_flag) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* STEP 2. Collect cpu_sets for each rank at the root */
cpu_set_t my_cpu_set;
CPU_ZERO(&my_cpu_set);
sched_getaffinity(0, sizeof(my_cpu_set), &my_cpu_set);
((cpu_set_t *) (shared_region))[rank] = my_cpu_set;
mpi_errno = MPIR_Barrier_impl(comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* STEP 3. Root has all the cpu_set information, now build tree */
if (rank == root) {
topo_depth = hwloc_topology_get_depth(MPIR_Process.hwloc_topology);
bind_map = (int **) MPL_malloc(num_ranks * sizeof(int *), MPL_MEM_OTHER);
MPIR_ERR_CHKANDJUMP(!bind_map, mpi_errno, MPI_ERR_OTHER, "**nomem");
for (i = 0; i < num_ranks; ++i) {
bind_map[i] = (int *) MPL_calloc(topo_depth, sizeof(int), MPL_MEM_OTHER);
MPIR_ERR_CHKANDJUMP(!bind_map[i], mpi_errno, MPI_ERR_OTHER, "**nomem");
}
MPIDI_SHM_hwloc_init_bindmap(num_ranks, topo_depth, shared_region, bind_map);
/* Done building the topology information */
/* STEP 3.1. Count the maximum entries at each level - used for breaking the tree into
* intra/inter socket */
max_entries_per_level = (int *) MPL_calloc(topo_depth, sizeof(size_t), MPL_MEM_OTHER);
MPIR_ERR_CHKANDJUMP(!max_entries_per_level, mpi_errno, MPI_ERR_OTHER, "**nomem");
package_level =
MPIDI_SHM_topotree_get_package_level(topo_depth, max_entries_per_level, num_ranks,
bind_map);
if (MPIDI_SHM_TOPOTREE_DEBUG)
fprintf(stderr, "Breaking topology at :: %d (default= %d)\n", package_level,
MPIDI_SHM_TOPOTREE_CUTOFF);
/* STEP 3.2. allocate space for the entries that go in each package based on hwloc info */
ranks_per_package =
(int
**) MPL_malloc(max_entries_per_level[package_level] * sizeof(int *), MPL_MEM_OTHER);
MPIR_ERR_CHKANDJUMP(!ranks_per_package, mpi_errno, MPI_ERR_OTHER, "**nomem");
package_ctr =
(int *) MPL_calloc(max_entries_per_level[package_level], sizeof(int), MPL_MEM_OTHER);
MPIR_ERR_CHKANDJUMP(!package_ctr, mpi_errno, MPI_ERR_OTHER, "**nomem");
for (i = 0; i < max_entries_per_level[package_level]; ++i) {
package_ctr[i] = 0;
ranks_per_package[i] = (int *) MPL_calloc(num_ranks, sizeof(int), MPL_MEM_OTHER);
MPIR_ERR_CHKANDJUMP(!ranks_per_package[i], mpi_errno, MPI_ERR_OTHER, "**nomem");
}
/* sort the ranks into packages based on the binding information */
for (i = 0; i < num_ranks; ++i) {
int package = bind_map[i][package_level];
ranks_per_package[package][package_ctr[package]++] = i;
}
max_ranks_per_package = 0;
for (i = 0; i < max_entries_per_level[package_level]; ++i) {
max_ranks_per_package = MPL_MAX(max_ranks_per_package, package_ctr[i]);
}
/* At this point we have done the common work in extracting topology information
* and restructuring it to our needs. Now we generate the tree. */
/* For Bcast, package leaders are added before the package local ranks, and the per_package
* tree is left_skewed */
mpi_errno = MPIDI_SHM_gen_tree(bcast_k, shared_region, max_entries_per_level,
ranks_per_package, max_ranks_per_package, package_ctr,
package_level, num_ranks, 1 /*package_leaders_first */ ,
0 /*left_skewed */ , errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
mpi_errno = MPIR_Barrier_impl(comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* Every rank copies their tree out from shared memory */
MPIDI_SHM_copy_tree(shared_region, num_ranks, rank, bcast_tree, bcast_topotree_fail);
if (MPIDI_SHM_TOPOTREE_DEBUG)
MPIDI_SHM_print_topotree_file("BCAST", comm_ptr->context_id, rank, bcast_tree);
/* Wait until shared memory is available */
mpi_errno = MPIR_Barrier_impl(comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* Generate the reduce tree */
/* For Reduce, package leaders are added after the package local ranks, and the per_package
* tree is right_skewed (children are added in the reverse order */
if (rank == root) {
memset(shared_region, 0, shm_size);
mpi_errno = MPIDI_SHM_gen_tree(reduce_k, shared_region, max_entries_per_level,
ranks_per_package, max_ranks_per_package, package_ctr,
package_level, num_ranks, 0 /*package_leaders_last */ ,
1 /*right_skewed */ , errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
mpi_errno = MPIR_Barrier_impl(comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* each rank copy the reduce tree out */
MPIDI_SHM_copy_tree(shared_region, num_ranks, rank, reduce_tree, reduce_topotree_fail);
if (MPIDI_SHM_TOPOTREE_DEBUG)
MPIDI_SHM_print_topotree_file("REDUCE", comm_ptr->context_id, rank, reduce_tree);
/* Wait for all ranks to copy out the tree */
mpi_errno = MPIR_Barrier_impl(comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag =
MPIX_ERR_PROC_FAILED ==
MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* Cleanup */
if (rank == root) {
for (i = 0; i < max_entries_per_level[package_level]; ++i) {
MPL_free(ranks_per_package[i]);
}
MPL_free(ranks_per_package);
MPL_free(package_ctr);
if (MPIDI_SHM_TOPOTREE_DEBUG)
for (i = 0; i < topo_depth; ++i) {
fprintf(stderr, "Level :: %d, Max :: %d\n", i, max_entries_per_level[i]);
}
for (i = 0; i < num_ranks; ++i) {
MPL_free(bind_map[i]);
}
MPL_free(max_entries_per_level);
MPL_free(bind_map);
}
MPIDIU_destroy_shm_segment(shm_size, &fd, (void **) &shared_region);
fn_exit:
if (rank == root && MPIDI_SHM_TOPOTREE_DEBUG)
fprintf(stderr, "Done creating tree for %d\n", num_ranks);
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_SHM_TOPOLOGY_TREE_INIT);
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int MPIDI_CH3I_Win_allocate_shm(MPI_Aint size, int disp_unit, MPIR_Info * info,
MPIR_Comm * comm_ptr, void *base_ptr, MPIR_Win ** win_ptr)
{
int mpi_errno = MPI_SUCCESS;
void **base_pp = (void **) base_ptr;
int i, node_size, node_rank;
MPIR_Comm *node_comm_ptr;
MPI_Aint *node_sizes;
MPIR_Errflag_t errflag = MPIR_ERR_NONE;
int noncontig = FALSE;
MPIR_CHKPMEM_DECL(1);
MPIR_CHKLMEM_DECL(1);
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_CH3I_WIN_ALLOCATE_SHM);
MPIR_FUNC_VERBOSE_RMA_ENTER(MPID_STATE_MPIDI_CH3I_WIN_ALLOCATE_SHM);
if ((*win_ptr)->comm_ptr->node_comm == NULL) {
mpi_errno =
MPIDI_CH3U_Win_allocate_no_shm(size, disp_unit, info, comm_ptr, base_ptr, win_ptr);
goto fn_exit;
}
/* see if we can allocate all windows contiguously */
noncontig = (*win_ptr)->info_args.alloc_shared_noncontig;
(*win_ptr)->shm_allocated = TRUE;
/* When allocating shared memory region segment, we need comm of processes
* that are on the same node as this process (node_comm).
* If node_comm == NULL, this process is the only one on this node, therefore
* we use comm_self as node comm. */
node_comm_ptr = (*win_ptr)->comm_ptr->node_comm;
MPIR_Assert(node_comm_ptr != NULL);
node_size = node_comm_ptr->local_size;
node_rank = node_comm_ptr->rank;
MPIR_T_PVAR_TIMER_START(RMA, rma_wincreate_allgather);
/* allocate memory for the base addresses, disp_units, and
* completion counters of all processes */
MPIR_CHKPMEM_MALLOC((*win_ptr)->shm_base_addrs, void **,
node_size * sizeof(void *), mpi_errno, "(*win_ptr)->shm_base_addrs");
/* get the sizes of the windows and window objectsof
* all processes. allocate temp. buffer for communication */
MPIR_CHKLMEM_MALLOC(node_sizes, MPI_Aint *, node_size * sizeof(MPI_Aint), mpi_errno,
"node_sizes");
/* FIXME: This needs to be fixed for heterogeneous systems */
node_sizes[node_rank] = (MPI_Aint) size;
mpi_errno = MPIR_Allgather_impl(MPI_IN_PLACE, 0, MPI_DATATYPE_NULL,
node_sizes, sizeof(MPI_Aint), MPI_BYTE,
node_comm_ptr, &errflag);
MPIR_T_PVAR_TIMER_END(RMA, rma_wincreate_allgather);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
(*win_ptr)->shm_segment_len = 0;
for (i = 0; i < node_size; i++) {
if (noncontig)
/* Round up to next page size */
(*win_ptr)->shm_segment_len += MPIDI_CH3_ROUND_UP_PAGESIZE(node_sizes[i]);
else
(*win_ptr)->shm_segment_len += node_sizes[i];
}
if ((*win_ptr)->shm_segment_len == 0) {
(*win_ptr)->base = NULL;
}
else {
mpi_errno = MPL_shm_hnd_init(&(*win_ptr)->shm_segment_handle);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
if (node_rank == 0) {
char *serialized_hnd_ptr = NULL;
/* create shared memory region for all processes in win and map */
mpi_errno =
MPL_shm_seg_create_and_attach((*win_ptr)->shm_segment_handle,
(*win_ptr)->shm_segment_len,
(char **) &(*win_ptr)->shm_base_addr, 0);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* serialize handle and broadcast it to the other processes in win */
mpi_errno =
MPL_shm_hnd_get_serialized_by_ref((*win_ptr)->shm_segment_handle,
&serialized_hnd_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno =
MPIR_Bcast_impl(serialized_hnd_ptr, MPL_SHM_GHND_SZ, MPI_CHAR, 0, node_comm_ptr,
&errflag);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
/* wait for other processes to attach to win */
mpi_errno = MPIR_Barrier_impl(node_comm_ptr, &errflag);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
/* unlink shared memory region so it gets deleted when all processes exit */
mpi_errno = MPL_shm_seg_remove((*win_ptr)->shm_segment_handle);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
}
else {
char serialized_hnd[MPL_SHM_GHND_SZ] = { 0 };
/* get serialized handle from rank 0 and deserialize it */
mpi_errno =
MPIR_Bcast_impl(serialized_hnd, MPL_SHM_GHND_SZ, MPI_CHAR, 0, node_comm_ptr,
&errflag);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
mpi_errno =
MPL_shm_hnd_deserialize((*win_ptr)->shm_segment_handle, serialized_hnd,
strlen(serialized_hnd));
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* attach to shared memory region created by rank 0 */
mpi_errno =
MPL_shm_seg_attach((*win_ptr)->shm_segment_handle, (*win_ptr)->shm_segment_len,
(char **) &(*win_ptr)->shm_base_addr, 0);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Barrier_impl(node_comm_ptr, &errflag);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
}
/* Allocated the interprocess mutex segment. */
mpi_errno = MPL_shm_hnd_init(&(*win_ptr)->shm_mutex_segment_handle);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
if (node_rank == 0) {
char *serialized_hnd_ptr = NULL;
/* create shared memory region for all processes in win and map */
mpi_errno =
MPL_shm_seg_create_and_attach((*win_ptr)->shm_mutex_segment_handle,
sizeof(MPIDI_CH3I_SHM_MUTEX),
(char **) &(*win_ptr)->shm_mutex, 0);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIDI_CH3I_SHM_MUTEX_INIT(*win_ptr);
/* serialize handle and broadcast it to the other processes in win */
mpi_errno =
MPL_shm_hnd_get_serialized_by_ref((*win_ptr)->shm_mutex_segment_handle,
&serialized_hnd_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno =
MPIR_Bcast_impl(serialized_hnd_ptr, MPL_SHM_GHND_SZ, MPI_CHAR, 0, node_comm_ptr,
&errflag);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
/* wait for other processes to attach to win */
mpi_errno = MPIR_Barrier_impl(node_comm_ptr, &errflag);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
/* unlink shared memory region so it gets deleted when all processes exit */
mpi_errno = MPL_shm_seg_remove((*win_ptr)->shm_mutex_segment_handle);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
}
else {
char serialized_hnd[MPL_SHM_GHND_SZ] = { 0 };
/* get serialized handle from rank 0 and deserialize it */
mpi_errno =
MPIR_Bcast_impl(serialized_hnd, MPL_SHM_GHND_SZ, MPI_CHAR, 0, node_comm_ptr,
&errflag);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
mpi_errno =
MPL_shm_hnd_deserialize((*win_ptr)->shm_mutex_segment_handle, serialized_hnd,
strlen(serialized_hnd));
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* attach to shared memory region created by rank 0 */
mpi_errno =
MPL_shm_seg_attach((*win_ptr)->shm_mutex_segment_handle,
sizeof(MPIDI_CH3I_SHM_MUTEX), (char **) &(*win_ptr)->shm_mutex,
0);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Barrier_impl(node_comm_ptr, &errflag);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
}
/* compute the base addresses of each process within the shared memory segment */
{
char *cur_base;
int cur_rank;
cur_base = (*win_ptr)->shm_base_addr;
cur_rank = 0;
((*win_ptr)->shm_base_addrs)[0] = (*win_ptr)->shm_base_addr;
for (i = 1; i < node_size; ++i) {
if (node_sizes[i]) {
/* For the base addresses, we track the previous
* process that has allocated non-zero bytes of shared
* memory. We can not simply use "i-1" for the
* previous process because rank "i-1" might not have
* allocated any memory. */
if (noncontig) {
((*win_ptr)->shm_base_addrs)[i] =
cur_base + MPIDI_CH3_ROUND_UP_PAGESIZE(node_sizes[cur_rank]);
}
else {
((*win_ptr)->shm_base_addrs)[i] = cur_base + node_sizes[cur_rank];
}
cur_base = ((*win_ptr)->shm_base_addrs)[i];
cur_rank = i;
}
else {
((*win_ptr)->shm_base_addrs)[i] = NULL;
}
}
}
(*win_ptr)->base = (*win_ptr)->shm_base_addrs[node_rank];
}
*base_pp = (*win_ptr)->base;
/* gather window information among processes via shared memory region. */
mpi_errno = MPIDI_CH3I_Win_gather_info((*base_pp), size, disp_unit, info, comm_ptr, win_ptr);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
/* Cache SHM windows */
MPIDI_CH3I_SHM_Wins_append(&shm_wins_list, (*win_ptr));
fn_exit:
MPIR_CHKLMEM_FREEALL();
MPIR_FUNC_VERBOSE_RMA_EXIT(MPID_STATE_MPIDI_CH3I_WIN_ALLOCATE_SHM);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
MPIR_CHKPMEM_REAP();
goto fn_exit;
/* --END ERROR HANDLING-- */
}
static int MPIDI_CH3I_Win_gather_info(void *base, MPI_Aint size, int disp_unit, MPIR_Info * info,
MPIR_Comm * comm_ptr, MPIR_Win ** win_ptr)
{
MPIR_Comm *node_comm_ptr = NULL;
int node_rank;
int comm_rank, comm_size;
MPI_Aint *tmp_buf = NULL;
int i, k;
MPIR_Errflag_t errflag = MPIR_ERR_NONE;
int mpi_errno = MPI_SUCCESS;
MPIR_CHKLMEM_DECL(1);
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_CH3I_WIN_GATHER_INFO);
MPIR_FUNC_VERBOSE_RMA_ENTER(MPID_STATE_MPIDI_CH3I_WIN_GATHER_INFO);
if ((*win_ptr)->comm_ptr->node_comm == NULL) {
mpi_errno = MPIDI_CH3U_Win_gather_info(base, size, disp_unit, info, comm_ptr, win_ptr);
goto fn_exit;
}
comm_size = (*win_ptr)->comm_ptr->local_size;
comm_rank = (*win_ptr)->comm_ptr->rank;
node_comm_ptr = (*win_ptr)->comm_ptr->node_comm;
MPIR_Assert(node_comm_ptr != NULL);
node_rank = node_comm_ptr->rank;
(*win_ptr)->info_shm_segment_len = comm_size * sizeof(MPIDI_Win_basic_info_t);
mpi_errno = MPL_shm_hnd_init(&(*win_ptr)->info_shm_segment_handle);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
if (node_rank == 0) {
char *serialized_hnd_ptr = NULL;
/* create shared memory region for all processes in win and map. */
mpi_errno = MPL_shm_seg_create_and_attach((*win_ptr)->info_shm_segment_handle,
(*win_ptr)->info_shm_segment_len,
(char **) &(*win_ptr)->info_shm_base_addr, 0);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* serialize handle and broadcast it to the other processes in win */
mpi_errno =
MPL_shm_hnd_get_serialized_by_ref((*win_ptr)->info_shm_segment_handle,
&serialized_hnd_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno =
MPIR_Bcast_impl(serialized_hnd_ptr, MPL_SHM_GHND_SZ, MPI_CHAR, 0, node_comm_ptr,
&errflag);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
/* wait for other processes to attach to win */
mpi_errno = MPIR_Barrier_impl(node_comm_ptr, &errflag);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
/* unlink shared memory region so it gets deleted when all processes exit */
mpi_errno = MPL_shm_seg_remove((*win_ptr)->info_shm_segment_handle);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
}
else {
char serialized_hnd[MPL_SHM_GHND_SZ] = { 0 };
/* get serialized handle from rank 0 and deserialize it */
mpi_errno =
MPIR_Bcast_impl(serialized_hnd, MPL_SHM_GHND_SZ, MPI_CHAR, 0, node_comm_ptr,
&errflag);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
mpi_errno = MPL_shm_hnd_deserialize((*win_ptr)->info_shm_segment_handle, serialized_hnd,
strlen(serialized_hnd));
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* attach to shared memory region created by rank 0 */
mpi_errno =
MPL_shm_seg_attach((*win_ptr)->info_shm_segment_handle,
(*win_ptr)->info_shm_segment_len,
(char **) &(*win_ptr)->info_shm_base_addr, 0);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Barrier_impl(node_comm_ptr, &errflag);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
}
(*win_ptr)->basic_info_table = (MPIDI_Win_basic_info_t *) ((*win_ptr)->info_shm_base_addr);
MPIR_CHKLMEM_MALLOC(tmp_buf, MPI_Aint *, 4 * comm_size * sizeof(MPI_Aint),
mpi_errno, "tmp_buf");
tmp_buf[4 * comm_rank] = MPIR_Ptr_to_aint(base);
tmp_buf[4 * comm_rank + 1] = size;
tmp_buf[4 * comm_rank + 2] = (MPI_Aint) disp_unit;
tmp_buf[4 * comm_rank + 3] = (MPI_Aint) (*win_ptr)->handle;
mpi_errno = MPIR_Allgather_impl(MPI_IN_PLACE, 0, MPI_DATATYPE_NULL, tmp_buf, 4, MPI_AINT,
(*win_ptr)->comm_ptr, &errflag);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
if (node_rank == 0) {
/* only node_rank == 0 writes results to basic_info_table on shared memory region. */
k = 0;
for (i = 0; i < comm_size; i++) {
(*win_ptr)->basic_info_table[i].base_addr = MPIR_Aint_to_ptr(tmp_buf[k++]);
(*win_ptr)->basic_info_table[i].size = tmp_buf[k++];
(*win_ptr)->basic_info_table[i].disp_unit = (int) tmp_buf[k++];
(*win_ptr)->basic_info_table[i].win_handle = (MPI_Win) tmp_buf[k++];
}
}
/* Make sure that all local processes see the results written by node_rank == 0 */
mpi_errno = MPIR_Barrier_impl(node_comm_ptr, &errflag);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
fn_exit:
MPIR_CHKLMEM_FREEALL();
MPIR_FUNC_VERBOSE_RMA_EXIT(MPID_STATE_MPIDI_CH3I_WIN_GATHER_INFO);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPID_Win_free(MPIR_Win ** win_ptr)
{
int mpi_errno = MPI_SUCCESS;
int in_use;
MPIR_Comm *comm_ptr;
MPIR_Errflag_t errflag = MPIR_ERR_NONE;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPID_WIN_FREE);
MPIR_FUNC_VERBOSE_RMA_ENTER(MPID_STATE_MPID_WIN_FREE);
MPIR_ERR_CHKANDJUMP(((*win_ptr)->states.access_state != MPIDI_RMA_NONE &&
(*win_ptr)->states.access_state != MPIDI_RMA_FENCE_ISSUED &&
(*win_ptr)->states.access_state != MPIDI_RMA_FENCE_GRANTED) ||
((*win_ptr)->states.exposure_state != MPIDI_RMA_NONE),
mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
/* 1. Here we must wait until all passive locks are released on this target,
* because for some UNLOCK messages, we do not send ACK back to origin,
* we must wait until lock is released so that we can free window.
* 2. We also need to wait until AT completion counter being zero, because
* this counter is increment everytime we meet a GET-like operation, it is
* possible that when target entering Win_free, passive epoch is not finished
* yet and there are still GETs doing on this target.
* 3. We also need to wait until lock queue becomes empty. It is possible
* that some lock requests is still waiting in the queue when target is
* entering Win_free. */
while ((*win_ptr)->current_lock_type != MPID_LOCK_NONE ||
(*win_ptr)->at_completion_counter != 0 ||
(*win_ptr)->target_lock_queue_head != NULL ||
(*win_ptr)->current_target_lock_data_bytes != 0 || (*win_ptr)->sync_request_cnt != 0) {
mpi_errno = wait_progress_engine();
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
mpi_errno = MPIR_Barrier_impl((*win_ptr)->comm_ptr, &errflag);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* Free window resources in lower layer. */
if (MPIDI_CH3U_Win_hooks.win_free != NULL) {
mpi_errno = MPIDI_CH3U_Win_hooks.win_free(win_ptr);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
/* dequeue window from the global list */
MPIR_Assert((*win_ptr)->active == FALSE);
MPL_DL_DELETE(MPIDI_RMA_Win_inactive_list_head, (*win_ptr));
if (MPIDI_RMA_Win_inactive_list_head == NULL && MPIDI_RMA_Win_active_list_head == NULL) {
/* this is the last window, de-register RMA progress hook */
mpi_errno = MPID_Progress_deregister_hook(MPIDI_CH3I_RMA_Progress_hook_id);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_POP(mpi_errno);
}
}
comm_ptr = (*win_ptr)->comm_ptr;
mpi_errno = MPIR_Comm_free_impl(comm_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
if ((*win_ptr)->basic_info_table != NULL)
MPL_free((*win_ptr)->basic_info_table);
MPL_free((*win_ptr)->op_pool_start);
MPL_free((*win_ptr)->target_pool_start);
MPL_free((*win_ptr)->slots);
MPL_free((*win_ptr)->target_lock_entry_pool_start);
MPIR_Assert((*win_ptr)->current_target_lock_data_bytes == 0);
/* Free the attached buffer for windows created with MPI_Win_allocate() */
if ((*win_ptr)->create_flavor == MPI_WIN_FLAVOR_ALLOCATE ||
(*win_ptr)->create_flavor == MPI_WIN_FLAVOR_SHARED) {
if ((*win_ptr)->shm_allocated == FALSE && (*win_ptr)->size > 0) {
MPL_free((*win_ptr)->base);
}
}
MPIR_Object_release_ref(*win_ptr, &in_use);
/* MPI windows don't have reference count semantics, so this should always be true */
MPIR_Assert(!in_use);
MPIR_Handle_obj_free(&MPIR_Win_mem, *win_ptr);
fn_exit:
MPIR_FUNC_VERBOSE_RMA_EXIT(MPID_STATE_MPID_WIN_FREE);
return mpi_errno;
fn_fail:
goto fn_exit;
}
