int oshmpi_trylock(long * lockp)
{
int is_locked = -1, nil = -1;
oshmpi_lock_t *lock = (oshmpi_lock_t *) lockp;
lock->prev = -1;
/* Get the last tail, if -1 replace with me */
MPI_Compare_and_swap (&shmem_world_rank, &nil, &(lock->prev), MPI_INT,
TAIL, TAIL_DISP, oshmpi_lock_win);
MPI_Win_flush (TAIL, oshmpi_lock_win);
/* Find if the last proc is holding lock */
if (lock->prev != -1)
{
MPI_Fetch_and_op (NULL, &is_locked, MPI_INT, lock->prev,
LOCK_DISP, MPI_NO_OP, oshmpi_lock_win);
MPI_Win_flush (lock->prev, oshmpi_lock_win);
if (is_locked)
return 0;
}
/* Add myself in tail */
MPI_Fetch_and_op (&shmem_world_rank, &(lock->prev), MPI_INT, TAIL,
TAIL_DISP, MPI_REPLACE, oshmpi_lock_win);
MPI_Win_flush (TAIL, oshmpi_lock_win);
/* Hold lock */
oshmpi_lock_base[LOCK_DISP] = 1;
MPI_Win_sync (oshmpi_lock_win);
return 1;
}
/** Unlock a mutex.
*
* @param[in] hdl Handle to the mutex
* @return MPI status
*/
int MCS_Mutex_unlock(MCS_Mutex hdl)
{
int next;
/* Read my next pointer. FOP is used since another process may write to
* this location concurrent with this read. */
MPI_Fetch_and_op(NULL, &next, MPI_INT, shmem_world_rank,
MCS_MTX_ELEM_DISP, MPI_NO_OP, hdl->window);
MPI_Win_flush(shmem_world_rank, hdl->window);
if ( next == -1) {
int tail;
int nil = -1;
/* Check if we are the at the tail of the lock queue. If so, we're
* done. If not, we need to send notification. */
MPI_Compare_and_swap(&nil, &shmem_world_rank, &tail, MPI_INT, hdl->tail_rank,
MCS_MTX_TAIL_DISP, hdl->window);
MPI_Win_flush(hdl->tail_rank, hdl->window);
if (tail != shmem_world_rank) {
debug_print("%2d: UNLOCK - waiting for next pointer (tail = %d)\n", shmem_world_rank, tail);
assert(tail >= 0 && tail < shmem_world_size);
for (;;) {
int flag;
MPI_Fetch_and_op(NULL, &next, MPI_INT, shmem_world_rank, MCS_MTX_ELEM_DISP,
MPI_NO_OP, hdl->window);
MPI_Win_flush(shmem_world_rank, hdl->window);
if (next != -1) break;
/* Is this here just to poke progress? If yes, then that is lame. */
MPI_Iprobe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &flag, MPI_STATUS_IGNORE);
}
}
}
/* Notify the next waiting process */
if (next != -1) {
debug_print("%2d: UNLOCK - notifying %d\n", shmem_world_rank, next);
MPI_Send(NULL, 0, MPI_BYTE, next, MCS_MUTEX_TAG, hdl->comm);
}
debug_print("%2d: UNLOCK - lock released\n", shmem_world_rank);
return MPI_SUCCESS;
}
/** Attempt to acquire a mutex.
*
* @param[in] hdl Handle to the mutex
* @param[out] success Indicates whether the mutex was acquired
* @return MPI status
*/
int MCS_Mutex_trylock(MCS_Mutex hdl, int *success)
{
int tail, nil = -1;
/* This store is safe, since it cannot happen concurrently with a remote
* write */
hdl->base[MCS_MTX_ELEM_DISP] = -1;
MPI_Win_sync(hdl->window);
/* Check if the lock is available and claim it if it is. */
MPI_Compare_and_swap(&shmem_world_rank, &nil, &tail, MPI_INT, hdl->tail_rank,
MCS_MTX_TAIL_DISP, hdl->window);
MPI_Win_flush(hdl->tail_rank, hdl->window);
/* If the old tail was -1, we have claimed the mutex */
*success = (tail == nil) ? 0 : 1;
debug_print("%2d: TRYLOCK - %s\n", shmem_world_rank, (*success) ? "Success" : "Non-success");
return MPI_SUCCESS;
}
dart_ret_t dart_lock_try_acquire (dart_lock_t lock, int32_t *is_acquired)
{
dart_unit_t unitid;
dart_team_myid (lock -> teamid, &unitid);
if (lock -> is_acquired == 1)
{
printf ("Warning: TRYLOCK - %2d has acquired the lock already\n", unitid);
return DART_OK;
}
dart_gptr_t gptr_tail;
int32_t result[1];
int32_t compare[1] = {-1};
DART_GPTR_COPY(gptr_tail, lock -> gptr_tail);
dart_unit_t tail = gptr_tail.unitid;
uint64_t offset = gptr_tail.addr_or_offs.offset;
/* Atomicity: Check if the lock is available and claim it if it is. */
MPI_Compare_and_swap (&unitid, compare, result, MPI_INT32_T, tail, offset, dart_win_local_alloc);
MPI_Win_flush (tail, dart_win_local_alloc);
/* If the old predecessor was -1, we will claim the lock, otherwise, do nothing. */
if (*result == -1)
{
lock -> is_acquired = 1;
*is_acquired = 1;
}
else
{
*is_acquired = 0;
}
DART_LOG_DEBUG("dart_lock_try_acquire: trylock %s in team %d",
((*is_acquired) ? "succeeded" : "failed"),
(lock -> teamid));
return DART_OK;
}
int main(int argc, char **argv) {
int procid, nproc, i;
MPI_Win llist_win;
llist_ptr_t head_ptr, tail_ptr;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &procid);
MPI_Comm_size(MPI_COMM_WORLD, &nproc);
MPI_Win_create_dynamic(MPI_INFO_NULL, MPI_COMM_WORLD, &llist_win);
/* Process 0 creates the head node */
if (procid == 0)
head_ptr.disp = alloc_elem(-1, llist_win);
/* Broadcast the head pointer to everyone */
head_ptr.rank = 0;
MPI_Bcast(&head_ptr.disp, 1, MPI_AINT, 0, MPI_COMM_WORLD);
tail_ptr = head_ptr;
/* All processes concurrently append NUM_ELEMS elements to the list */
for (i = 0; i < NUM_ELEMS; i++) {
llist_ptr_t new_elem_ptr;
int success;
/* Create a new list element and register it with the window */
new_elem_ptr.rank = procid;
new_elem_ptr.disp = alloc_elem(procid, llist_win);
/* Append the new node to the list. This might take multiple attempts if
others have already appended and our tail pointer is stale. */
do {
llist_ptr_t next_tail_ptr = nil;
MPI_Win_lock(MPI_LOCK_SHARED, tail_ptr.rank, MPI_MODE_NOCHECK, llist_win);
MPI_Compare_and_swap((void*) &new_elem_ptr.rank, (void*) &nil.rank,
(void*) &next_tail_ptr.rank, MPI_INT, tail_ptr.rank,
(MPI_Aint) &(((llist_elem_t*)tail_ptr.disp)->next.rank), llist_win);
MPI_Win_unlock(tail_ptr.rank, llist_win);
success = (next_tail_ptr.rank == nil.rank);
if (success) {
int i, flag;
MPI_Aint result;
MPI_Win_lock(MPI_LOCK_SHARED, tail_ptr.rank, MPI_MODE_NOCHECK, llist_win);
MPI_Fetch_and_op(&new_elem_ptr.disp, &result, MPI_AINT, tail_ptr.rank,
(MPI_Aint) &(((llist_elem_t*)tail_ptr.disp)->next.disp),
MPI_REPLACE, llist_win);
/* Note: accumulate is faster, since we don't need the result. Replacing with
Fetch_and_op to create a more complete test case. */
/*
MPI_Accumulate(&new_elem_ptr.disp, 1, MPI_AINT, tail_ptr.rank,
(MPI_Aint) &(((llist_elem_t*)tail_ptr.disp)->next.disp), 1,
MPI_AINT, MPI_REPLACE, llist_win);
*/
MPI_Win_unlock(tail_ptr.rank, llist_win);
tail_ptr = new_elem_ptr;
/* For implementations that use pt-to-pt messaging, force progress for other threads'
RMA operations. */
for (i = 0; i < NPROBE; i++)
MPI_Iprobe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &flag, MPI_STATUS_IGNORE);
} else {
/* Tail pointer is stale, fetch the displacement. May take multiple tries
if it is being updated. */
do {
MPI_Aint junk = 0;
MPI_Win_lock(MPI_LOCK_SHARED, tail_ptr.rank, MPI_MODE_NOCHECK, llist_win);
MPI_Fetch_and_op(NULL, &next_tail_ptr.disp, MPI_AINT, tail_ptr.rank,
(MPI_Aint) &(((llist_elem_t*)tail_ptr.disp)->next.disp),
MPI_NO_OP, llist_win);
MPI_Win_unlock(tail_ptr.rank, llist_win);
} while (next_tail_ptr.disp == nil.disp);
tail_ptr = next_tail_ptr;
}
} while (!success);
}
MPI_Barrier(MPI_COMM_WORLD);
/* Traverse the list and verify that all processes inserted exactly the correct
number of elements. */
if (procid == 0) {
int have_root = 0;
int errors = 0;
int *counts, count = 0;
counts = (int*) malloc(sizeof(int) * nproc);
assert(counts != NULL);
for (i = 0; i < nproc; i++)
counts[i] = 0;
tail_ptr = head_ptr;
/* Walk the list and tally up the number of elements inserted by each rank */
while (tail_ptr.disp != nil.disp) {
llist_elem_t elem;
MPI_Win_lock(MPI_LOCK_SHARED, tail_ptr.rank, MPI_MODE_NOCHECK, llist_win);
MPI_Get(&elem, sizeof(llist_elem_t), MPI_BYTE,
tail_ptr.rank, tail_ptr.disp, sizeof(llist_elem_t), MPI_BYTE, llist_win);
MPI_Win_unlock(tail_ptr.rank, llist_win);
tail_ptr = elem.next;
/* This is not the root */
if (have_root) {
assert(elem.value >= 0 && elem.value < nproc);
counts[elem.value]++;
count++;
if (verbose) {
int last_elem = tail_ptr.disp == nil.disp;
printf("%2d%s", elem.value, last_elem ? "" : " -> ");
if (count % ELEM_PER_ROW == 0 && !last_elem)
printf("\n");
}
}
/* This is the root */
else {
assert(elem.value == -1);
have_root = 1;
}
}
if (verbose)
printf("\n\n");
/* Verify the counts we collected */
for (i = 0; i < nproc; i++) {
int expected = NUM_ELEMS;
if (counts[i] != expected) {
printf("Error: Rank %d inserted %d elements, expected %d\n", i, counts[i], expected);
errors++;
}
}
printf("%s\n", errors == 0 ? " No Errors" : "FAIL");
free(counts);
}
MPI_Win_free(&llist_win);
/* Free all the elements in the list */
for ( ; my_elems_count > 0; my_elems_count--)
MPI_Free_mem(my_elems[my_elems_count-1]);
MPI_Finalize();
return 0;
}
int main(int argc, char *argv[])
{
int rank, size, i, j, k;
int errors = 0;
int origin_shm, origin_am, dest;
int *orig_buf = NULL, *result_buf = NULL, *compare_buf = NULL,
*target_buf = NULL, *check_buf = NULL;
MPI_Win win;
MPI_Status status;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
if (size != 3) {
/* run this test with three processes */
goto exit_test;
}
/* this works when MPIR_PARAM_CH3_ODD_EVEN_CLIQUES is set */
dest = 2;
origin_shm = 0;
origin_am = 1;
if (rank != dest) {
MPI_Alloc_mem(sizeof(int), MPI_INFO_NULL, &orig_buf);
MPI_Alloc_mem(sizeof(int), MPI_INFO_NULL, &result_buf);
MPI_Alloc_mem(sizeof(int), MPI_INFO_NULL, &compare_buf);
}
MPI_Win_allocate(sizeof(int), sizeof(int), MPI_INFO_NULL, MPI_COMM_WORLD, &target_buf, &win);
for (k = 0; k < LOOP_SIZE; k++) {
/* init buffers */
if (rank == origin_shm) {
orig_buf[0] = 1;
compare_buf[0] = 0;
result_buf[0] = 0;
}
else if (rank == origin_am) {
orig_buf[0] = 0;
compare_buf[0] = 1;
result_buf[0] = 0;
}
else {
MPI_Win_lock(MPI_LOCK_SHARED, rank, 0, win);
target_buf[0] = 0;
MPI_Win_unlock(rank, win);
}
MPI_Barrier(MPI_COMM_WORLD);
/* perform FOP */
MPI_Win_lock_all(0, win);
if (rank != dest) {
MPI_Compare_and_swap(orig_buf, compare_buf, result_buf, MPI_INT, dest, 0, win);
MPI_Win_flush(dest, win);
}
MPI_Win_unlock_all(win);
MPI_Barrier(MPI_COMM_WORLD);
/* check results */
if (rank != dest) {
MPI_Gather(result_buf, 1, MPI_INT, check_buf, 1, MPI_INT, dest, MPI_COMM_WORLD);
}
else {
MPI_Alloc_mem(sizeof(int) * 3, MPI_INFO_NULL, &check_buf);
MPI_Gather(target_buf, 1, MPI_INT, check_buf, 1, MPI_INT, dest, MPI_COMM_WORLD);
if (!(check_buf[dest] == 0 && check_buf[origin_shm] == 0 && check_buf[origin_am] == 1)
&& !(check_buf[dest] == 1 && check_buf[origin_shm] == 0 &&
check_buf[origin_am] == 0)) {
printf
("Wrong results: target result = %d, origin_shm result = %d, origin_am result = %d\n",
check_buf[dest], check_buf[origin_shm], check_buf[origin_am]);
printf
("Expected results (1): target result = 1, origin_shm result = 0, origin_am result = 0\n");
printf
("Expected results (2): target result = 0, origin_shm result = 0, origin_am result = 1\n");
errors++;
}
MPI_Free_mem(check_buf);
}
}
MPI_Win_free(&win);
if (rank == origin_am || rank == origin_shm) {
MPI_Free_mem(orig_buf);
MPI_Free_mem(result_buf);
MPI_Free_mem(compare_buf);
}
exit_test:
if (rank == dest && errors == 0)
printf(" No Errors\n");
MPI_Finalize();
return 0;
}
void IMB_rma_compare_and_swap (struct comm_info* c_info, int size,
struct iter_schedule* iterations,
MODES run_mode, double* time)
{
double res_time = -1.;
int root = c_info->pair1;
int s_size;
int i;
void *comp_b, *orig_b, *res_b;
MPI_Datatype data_type = MPI_INT;
ierr = 0;
if (c_info->rank < 0)
{
*time = res_time;
return;
}
MPI_Type_size(data_type,&s_size);
for(i=0; i<N_BARR; i++) MPI_Barrier(c_info->communicator);
if (c_info->rank == c_info->pair0)
{
/* use r_buffer for all buffers required by compare_and_swap, because
* on all ranks r_buffer is zero-initialized in IMB_set_buf function */
orig_b = (char*)c_info->r_buffer + s_size*2;
comp_b = (char*)c_info->r_buffer + s_size;
res_b = c_info->r_buffer;
MPI_Win_lock(MPI_LOCK_SHARED, root, 0, c_info->WIN);
if (run_mode->AGGREGATE)
{
res_time = MPI_Wtime();
for (i = 0; i < iterations->n_sample; i++)
{
ierr = MPI_Compare_and_swap(
(char*)orig_b + i%iterations->r_cache_iter*iterations->r_offs,
(char*)comp_b + i%iterations->r_cache_iter*iterations->r_offs,
(char*)res_b + i%iterations->r_cache_iter*iterations->r_offs,
data_type, root, i%iterations->r_cache_iter*iterations->r_offs,
c_info->WIN );
MPI_ERRHAND(ierr);
}
ierr = MPI_Win_flush(root, c_info->WIN);
res_time = (MPI_Wtime() - res_time)/iterations->n_sample;
}
else if ( !run_mode->AGGREGATE )
{
res_time = MPI_Wtime();
for (i = 0; i < iterations->n_sample; i++)
{
ierr = MPI_Compare_and_swap(
(char*)orig_b + i%iterations->s_cache_iter*iterations->s_offs,
(char*)comp_b + i%iterations->s_cache_iter*iterations->s_offs,
(char*)res_b + i%iterations->r_cache_iter*iterations->r_offs,
data_type, root, i%iterations->r_cache_iter*iterations->r_offs,
c_info->WIN );
MPI_ERRHAND(ierr);
ierr = MPI_Win_flush(root, c_info->WIN);
MPI_ERRHAND(ierr);
}
res_time = (MPI_Wtime() - res_time)/iterations->n_sample;
}
MPI_Win_unlock(root, c_info->WIN);
}
MPI_Barrier(c_info->communicator);
*time = res_time;
return;
}
dart_ret_t dart_lock_release (dart_lock_t lock)
{
dart_unit_t unitid;
dart_team_myid (lock -> teamid, &unitid);
if (lock -> is_acquired == 0)
{
printf ("Warning: RELEASE - %2d has not yet required the lock\n", unitid);
return DART_OK;
}
dart_gptr_t gptr_tail;
dart_gptr_t gptr_list;
MPI_Win win;
int32_t *addr2, next, result[1];
MPI_Aint disp_list;
int32_t origin[1] = {-1};
DART_GPTR_COPY(gptr_tail, lock -> gptr_tail);
DART_GPTR_COPY(gptr_list, lock -> gptr_list);
uint64_t offset_tail = gptr_tail.addr_or_offs.offset;
int16_t seg_id = gptr_list.segid;
dart_unit_t tail = gptr_tail.unitid;
uint16_t index = gptr_list.flags;
dart_gptr_getaddr(gptr_list, (void*)&addr2);
win = dart_win_lists[index];
/* Atomicity: Check if we are at the tail of this lock queue, if so, we are done.
* Otherwise, we still need to send notification. */
MPI_Compare_and_swap (origin, &unitid, result, MPI_INT32_T, tail, offset_tail, dart_win_local_alloc);
MPI_Win_flush (tail, dart_win_local_alloc);
/* We are not at the tail of this lock queue. */
if (*result != unitid)
{
DART_LOG_DEBUG ("%2d: UNLOCK - waiting for next pointer (tail = %d) in team %d",
unitid, *result, (lock -> teamid));
if (dart_adapt_transtable_get_disp (seg_id, unitid, &disp_list) == -1)
{
return DART_ERR_INVAL;
}
/* Waiting for the update of my next pointer finished. */
while (1)
{
MPI_Fetch_and_op (NULL, &next, MPI_INT, unitid, disp_list, MPI_NO_OP, win);
MPI_Win_flush (unitid, win);
if (next != -1) break;
}
DART_LOG_DEBUG ("%2d: UNLOCK - notifying %d in team %d", unitid, next, (lock -> teamid));
/* Notifying the next unit waiting on the lock queue. */
MPI_Send (NULL, 0, MPI_INT, next, 0, dart_teams[index]);
*addr2 = -1;
MPI_Win_sync (win);
}
lock -> is_acquired = 0;
DART_LOG_DEBUG ("%2d: UNLOCK - release lock in team %d", unitid, (lock -> teamid));
return DART_OK;
}
FORT_DLL_SPEC void FORT_CALL mpi_compare_and_swap_ ( void*v1, void*v2, void*v3, MPI_Fint *v4, MPI_Fint *v5, MPI_Aint * v6, MPI_Fint *v7, MPI_Fint *ierr ){
*ierr = MPI_Compare_and_swap( v1, v2, v3, (MPI_Datatype)(*v4), (int)*v5, *v6, (MPI_Win)*v7 );
}
