ch3u_rma_sync.c 57.4 KB
Newer Older
1
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil ; -*- */
2 3 4 5 6 7 8 9
/*
 *  (C) 2001 by Argonne National Laboratory.
 *      See COPYRIGHT in top-level directory.
 */

#include "mpidimpl.h"
#include "mpidrma.h"

10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211
/* Notes for memory barriers in RMA synchronizations

   When SHM is allocated for RMA window, we need to add memory berriers at proper
   places in RMA synchronization routines to guarantee the ordering of read/write
   operations, so that any operations after synchronization calls will see the
   correct data.

   There are four kinds of operations involved in the following explanation:

   1. Local loads/stores: any operations happening outside RMA epoch and accessing
      each process's own window memory.

   2. SHM operations: any operations happening inside RMA epoch. They may access
      any processes' window memory, which include direct loads/stores, and
      RMA operations that are internally implemented as direct loads/stores in
      MPI implementation.

   3. PROC_SYNC: synchronzations among processes by sending/recving messages.

   4. MEM_SYNC: a full memory barrier. It ensures the ordering of read/write
      operations on each process.

   (1) FENCE synchronization

              RANK 0                           RANK 1

       (local loads/stores)             (local loads/stores)

           WIN_FENCE {                    WIN_FENCE {
               MEM_SYNC                       MEM_SYNC
               PROC_SYNC -------------------- PROC_SYNC
               MEM_SYNC                       MEM_SYNC
           }                              }

        (SHM operations)                  (SHM operations)

           WIN_FENCE {                     WIN_FENCE {
               MEM_SYNC                        MEM_SYNC
               PROC_SYNC --------------------- PROC_SYNC
               MEM_SYNC                        MEM_SYNC
           }                               }

      (local loads/stores)              (local loads/stores)

       We need MEM_SYNC before and after PROC_SYNC for both starting WIN_FENCE
       and ending WIN_FENCE, to ensure the ordering between local loads/stores
       and PROC_SYNC in starting WIN_FENCE (and vice versa in ending WIN_FENCE),
       and the ordering between PROC_SYNC and SHM operations in starting WIN_FENCE
       (and vice versa for ending WIN_FENCE).

       In starting WIN_FENCE, the MEM_SYNC before PROC_SYNC essentially exposes
       previous local loads/stores to other processes; after PROC_SYNC, each
       process knows that everyone else already exposed their local loads/stores;
       the MEM_SYNC after PROC_SYNC ensures that my following SHM operations will
       happen after PROC_SYNC and will see the latest data on other processes.

       In ending WIN_FENCE, the MEM_SYNC before PROC_SYNC essentially exposes
       previous SHM operations to other processes; after PROC_SYNC, each process
       knows everyone else already exposed their SHM operations; the MEM_SYNC
       after PROC_SYNC ensures that my following local loads/stores will happen
       after PROC_SYNC and will see the latest data in my memory region.

   (2) POST-START-COMPLETE-WAIT synchronization

              RANK 0                           RANK 1

                                          (local loads/stores)

           WIN_START {                      WIN_POST {
                                                MEM_SYNC
               PROC_SYNC ---------------------- PROC_SYNC
               MEM_SYNC
           }                                }

         (SHM operations)

           WIN_COMPLETE {                  WIN_WAIT/TEST {
               MEM_SYNC
               PROC_SYNC --------------------- PROC_SYNC
                                               MEM_SYNC
           }                               }

                                          (local loads/stores)

       We need MEM_SYNC before PROC_SYNC for WIN_POST and WIN_COMPLETE, and
       MEM_SYNC after PROC_SYNC in WIN_START and WIN_WAIT/TEST, to ensure the
       ordering between local loads/stores and PROC_SYNC in WIN_POST (and
       vice versa in WIN_WAIT/TEST), and the ordering between PROC_SYNC and SHM
       operations in WIN_START (and vice versa in WIN_COMPLETE).

       In WIN_POST, the MEM_SYNC before PROC_SYNC essentially exposes previous
       local loads/stores to group of origin processes; after PROC_SYNC, origin
       processes knows all target processes already exposed their local
       loads/stores; in WIN_START, the MEM_SYNC after PROC_SYNC ensures that
       following SHM operations will happen after PROC_SYNC and will see the
       latest data on target processes.

       In WIN_COMPLETE, the MEM_SYNC before PROC_SYNC essentailly exposes previous
       SHM operations to group of target processes; after PROC_SYNC, target
       processes knows all origin process already exposed their SHM operations;
       in WIN_WAIT/TEST, the MEM_SYNC after PROC_SYNC ensures that following local
       loads/stores will happen after PROC_SYNC and will see the latest data in
       my memory region.

   (3) Passive target synchronization

              RANK 0                          RANK 1

                                        WIN_LOCK(target=1) {
                                            PROC_SYNC (lock granted)
                                            MEM_SYNC
                                        }

                                        (SHM operations)

                                        WIN_UNLOCK(target=1) {
                                            MEM_SYNC
                                            PROC_SYNC (lock released)
                                        }

         PROC_SYNC -------------------- PROC_SYNC

         WIN_LOCK (target=1) {
             PROC_SYNC (lock granted)
             MEM_SYNC
         }

         (SHM operations)

         WIN_UNLOCK (target=1) {
             MEM_SYNC
             PROC_SYNC (lock released)
         }

         PROC_SYNC -------------------- PROC_SYNC

                                        WIN_LOCK(target=1) {
                                            PROC_SYNC (lock granted)
                                            MEM_SYNC
                                        }

                                        (SHM operations)

                                        WIN_UNLOCK(target=1) {
                                            MEM_SYNC
                                            PROC_SYNC (lock released)
                                        }

         We need MEM_SYNC after PROC_SYNC in WIN_LOCK, and MEM_SYNC before
         PROC_SYNC in WIN_UNLOCK, to ensure the ordering between SHM operations
         and PROC_SYNC and vice versa.

         In WIN_LOCK, the MEM_SYNC after PROC_SYNC guarantees two things:
         (a) it guarantees that following SHM operations will happen after
         lock is granted; (b) it guarantees that following SHM operations
         will happen after any PROC_SYNC with target before WIN_LOCK is called,
         which means those SHM operations will see the latest data on target
         process.

         In WIN_UNLOCK, the MEM_SYNC before PROC_SYNC also guarantees two
         things: (a) it guarantees that SHM operations will happen before
         lock is released; (b) it guarantees that SHM operations will happen
         before any PROC_SYNC with target after WIN_UNLOCK is returned, which
         means following SHM operations on that target will see the latest data.

         WIN_LOCK_ALL/UNLOCK_ALL are same with WIN_LOCK/UNLOCK.

              RANK 0                          RANK 1

         WIN_LOCK_ALL

         (SHM operations)

         WIN_FLUSH(target=1) {
             MEM_SYNC
         }

         PROC_SYNC ------------------------PROC_SYNC

                                           WIN_LOCK(target=1) {
                                               PROC_SYNC (lock granted)
                                               MEM_SYNC
                                           }

                                           (SHM operations)

                                           WIN_UNLOCK(target=1) {
                                               MEM_SYNC
                                               PROC_SYNC (lock released)
                                           }

         WIN_UNLOCK_ALL

         We need MEM_SYNC in WIN_FLUSH to ensure the ordering between SHM
         operations and PROC_SYNC.

         The MEM_SYNC in WIN_FLUSH guarantees that all SHM operations before
         this WIN_FLUSH will happen before any PROC_SYNC with target after
         this WIN_FLUSH, which means SHM operations on target process after
         PROC_SYNC with origin will see the latest data.
*/

212
void MPIDI_CH3_RMA_Init_Pvars(void)
213 214
{
}
215

216 217
/* These are used to use a common routine to complete lists of RMA
   operations with a single routine, while collecting data that
218 219 220 221
   distinguishes between different synchronization modes.  This is not
   thread-safe; the best choice for thread-safety is to eliminate this
   ability to discriminate between the different types of RMA synchronization.
*/
222

223 224 225
/*
 * These routines provide a default implementation of the MPI RMA operations
 * in terms of the low-level, two-sided channel operations.  A channel
226 227
 * may override these functions, on a per-window basis, by overriding
 * the MPID functions in the RMAFns section of MPID_Win object.
228 229
 */

230 231
#define SYNC_POST_TAG 100

232

233 234 235 236
#undef FUNCNAME
#define FUNCNAME MPIDI_Win_fence
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
237
int MPIDI_Win_fence(int assert, MPID_Win * win_ptr)
238
{
239 240 241
    int i, made_progress = 0;
    int local_completed = 0, remote_completed = 0;
    MPIDI_RMA_Target_t *curr_target = NULL;
242
    int errflag = FALSE;
243
    int mpi_errno = MPI_SUCCESS;
244 245 246 247
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_FENCE);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_FENCE);

248 249 250 251
    MPIU_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,
James Dinan's avatar
James Dinan committed
252 253
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");

254
    win_ptr->posted_ops_cnt = 0;
255

256 257 258
    if (assert & MPI_MODE_NOPRECEDE) {
        if (assert & MPI_MODE_NOSUCCEED) {
            goto fn_exit;
259
        }
260 261 262 263 264 265 266 267 268 269 270
        else {
            /* It is possible that there is a IBARRIER in MPI_WIN_FENCE with
               MODE_NOPRECEDE not being completed, we let the progress engine
               to delete its request when it is completed. */
            if (win_ptr->fence_sync_req != MPI_REQUEST_NULL) {
                MPID_Request *req_ptr;
                MPID_Request_get_ptr(win_ptr->fence_sync_req, req_ptr);
                MPID_Request_release(req_ptr);
                win_ptr->fence_sync_req = MPI_REQUEST_NULL;
                win_ptr->states.access_state = MPIDI_RMA_NONE;
            }
271

272 273
            if (win_ptr->shm_allocated == TRUE) {
                MPID_Comm *node_comm_ptr = win_ptr->comm_ptr->node_comm;
274

275 276
                /* Ensure ordering of load/store operations. */
                OPA_read_write_barrier();
277

278 279 280
                mpi_errno = MPIR_Barrier_impl(node_comm_ptr, &errflag);
                if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
                MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
281

282 283
                /* Ensure ordering of load/store operations. */
                OPA_read_write_barrier();
284 285
            }

286 287
            mpi_errno = MPIR_Ibarrier_impl(win_ptr->comm_ptr, &(win_ptr->fence_sync_req));
            if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
288

289 290
            win_ptr->states.access_state = MPIDI_RMA_FENCE_ISSUED;
            num_active_issued_win++;
291

292
            goto fn_exit;
293
        }
294
    }
295

296 297 298
    if (win_ptr->states.access_state == MPIDI_RMA_FENCE_ISSUED) {
        while (win_ptr->states.access_state != MPIDI_RMA_FENCE_GRANTED) {
            mpi_errno = wait_progress_engine();
299 300 301
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
        }
302
    }
303

304 305 306 307 308 309 310 311 312 313
    /* Set sync_flag in target structs. */
    for (i = 0; i < win_ptr->num_slots; i++) {
        curr_target = win_ptr->slots[i].target_list;
        while (curr_target != NULL) {

            /* set sync_flag in sync struct */
            if (curr_target->sync.sync_flag < MPIDI_RMA_SYNC_FLUSH) {
                curr_target->sync.sync_flag = MPIDI_RMA_SYNC_FLUSH;
                curr_target->sync.have_remote_incomplete_ops = 0;
                curr_target->sync.outstanding_acks++;
314
            }
315
            curr_target = curr_target->next;
316
        }
317
    }
318

319 320 321 322 323 324 325 326 327 328 329 330 331 332
    /* Issue out all operations. */
    mpi_errno = MPIDI_CH3I_RMA_Make_progress_win(win_ptr, &made_progress);
    if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);

    /* Wait for remote completion. */
    do {
        mpi_errno = MPIDI_CH3I_RMA_Cleanup_ops_win(win_ptr,
                                                   &local_completed,
                                                   &remote_completed);
        if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
        if (!remote_completed) {
            mpi_errno = wait_progress_engine();
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
333
        }
334
    } while (!remote_completed);
335

336 337 338
    /* Cleanup all targets on window. */
    mpi_errno = MPIDI_CH3I_RMA_Cleanup_targets_win(win_ptr);
    if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
James Dinan's avatar
James Dinan committed
339

340
    MPIU_Assert(win_ptr->non_empty_slots == 0);
341

342 343 344 345
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }
346

347 348 349
    mpi_errno = MPIR_Barrier_impl(win_ptr->comm_ptr, &errflag);
    if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
    MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
350

351 352 353 354
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }
355

356 357 358 359 360
    if (assert & MPI_MODE_NOSUCCEED) {
        win_ptr->states.access_state = MPIDI_RMA_NONE;
    }
    else {
        win_ptr->states.access_state = MPIDI_RMA_FENCE_GRANTED;
361 362
    }

363 364 365
    /* There should be no active requests. */
    MPIU_Assert(win_ptr->active_req_cnt == 0);

366
  fn_exit:
367 368 369
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_FENCE);
    return mpi_errno;
    /* --BEGIN ERROR HANDLING-- */
370
  fn_fail:
371 372 373 374 375
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}


376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408
static int fill_ranks_in_win_grp(MPID_Win *win_ptr, MPID_Group *group_ptr, int *ranks_in_win_grp)
{
    int mpi_errno = MPI_SUCCESS;
    int i, *ranks_in_grp;
    MPID_Group *win_grp_ptr;
    MPIU_CHKLMEM_DECL(1);
    MPIDI_STATE_DECL(MPID_STATE_FILL_RANKS_IN_WIN_GRP);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_FILL_RANKS_IN_WIN_GRP);

    MPIU_CHKLMEM_MALLOC(ranks_in_grp, int *, group_ptr->size * sizeof(int),
                        mpi_errno, "ranks_in_grp");
    for (i = 0; i < group_ptr->size; i++) ranks_in_grp[i] = i;

    mpi_errno = MPIR_Comm_group_impl(win_ptr->comm_ptr, &win_grp_ptr);
    if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);

    mpi_errno = MPIR_Group_translate_ranks_impl(group_ptr, group_ptr->size,
                                                ranks_in_grp, win_grp_ptr, ranks_in_win_grp);
    if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);

    mpi_errno = MPIR_Group_free_impl(win_grp_ptr);
    if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);

  fn_exit:
    MPIU_CHKLMEM_FREEALL();
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_FILL_RANKS_IN_WIN_GRP);
    return mpi_errno;
 fn_fail:
    goto fn_exit;
}


409
#undef FUNCNAME
410
#define FUNCNAME MPIDI_Win_post
411 412
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
413
int MPIDI_Win_post(MPID_Group * post_grp_ptr, int assert, MPID_Win * win_ptr)
414
{
415
    int *post_ranks_in_win_grp;
416
    int mpi_errno = MPI_SUCCESS;
417
    MPIU_CHKLMEM_DECL(3);
418
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_POST);
419

420
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_POST);
421

422 423 424 425 426
    /* Note that here we cannot distinguish if this exposure epoch is overlapped
       with an exposure epoch of FENCE (which is not allowed), since FENCE may be
       ended up with not unsetting the window state. We can only detect if this
       exposure epoch is overlapped with another exposure epoch of PSCW. */
    MPIU_ERR_CHKANDJUMP(win_ptr->states.exposure_state != MPIDI_RMA_NONE,
427
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
428

429
    win_ptr->states.exposure_state = MPIDI_RMA_PSCW_EXPO;
430

431
    win_ptr->at_completion_counter += post_grp_ptr->size;
432

433 434 435
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
436 437
    }

438 439 440
    if ((assert & MPI_MODE_NOCHECK) == 0) {
        MPI_Request *req;
        MPI_Status *status;
441 442
        int i, post_grp_size, dst, rank;
        MPID_Comm *win_comm_ptr;
443

444 445
        /* NOCHECK not specified. We need to notify the source
         * processes that Post has been called. */
446

447
        post_grp_size = post_grp_ptr->size;
448 449
        win_comm_ptr = win_ptr->comm_ptr;
        rank = win_ptr->comm_ptr->rank;
450

451 452 453 454 455 456 457 458 459
        MPIU_CHKLMEM_MALLOC(post_ranks_in_win_grp, int *,
                            post_grp_size * sizeof(int), mpi_errno, "post_ranks_in_win_grp");
        mpi_errno = fill_ranks_in_win_grp(win_ptr, post_grp_ptr, post_ranks_in_win_grp);
        if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);

        MPIU_CHKLMEM_MALLOC(req, MPI_Request *, post_grp_size * sizeof(MPI_Request),
                            mpi_errno, "req");
        MPIU_CHKLMEM_MALLOC(status, MPI_Status *, post_grp_size * sizeof(MPI_Status),
                            mpi_errno, "status");
460

461 462
        /* Send a 0-byte message to the source processes */
        for (i = 0; i < post_grp_size; i++) {
463
            dst = post_ranks_in_win_grp[i];
464

465 466 467 468
            if (dst != rank) {
                MPID_Request *req_ptr;
                mpi_errno = MPID_Isend(&i, 0, MPI_INT, dst, SYNC_POST_TAG, win_comm_ptr,
                                       MPID_CONTEXT_INTRA_PT2PT, &req_ptr);
469
                if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
470 471 472 473 474
                req[i] = req_ptr->handle;
            }
            else {
                req[i] = MPI_REQUEST_NULL;
            }
475
        }
476

477 478 479 480 481 482 483 484 485 486 487 488
        mpi_errno = MPIR_Waitall_impl(post_grp_size, req, status);
        if (mpi_errno && mpi_errno != MPI_ERR_IN_STATUS)
            MPIU_ERR_POP(mpi_errno);

        /* --BEGIN ERROR HANDLING-- */
        if (mpi_errno == MPI_ERR_IN_STATUS) {
            for (i = 0; i < post_grp_size; i++) {
                if (status[i].MPI_ERROR != MPI_SUCCESS) {
                    mpi_errno = status[i].MPI_ERROR;
                    MPIU_ERR_POP(mpi_errno);
                }
            }
489
        }
490
        /* --END ERROR HANDLING-- */
491
    }
492

493
  fn_exit:
494 495
    MPIU_CHKLMEM_FREEALL();
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_POST);
496 497
    return mpi_errno;
    /* --BEGIN ERROR HANDLING-- */
498
  fn_fail:
499 500 501 502
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}

503

504 505 506 507 508
#undef FUNCNAME
#define FUNCNAME MPIDI_Win_start
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
int MPIDI_Win_start(MPID_Group * group_ptr, int assert, MPID_Win * win_ptr)
509
{
510
    int mpi_errno = MPI_SUCCESS;
511
    MPIU_CHKLMEM_DECL(2);
512 513
    MPIU_CHKPMEM_DECL(2);
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_START);
514

515
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_START);
516

517 518 519 520 521 522 523 524 525
    /* Note that here we cannot distinguish if this access epoch is overlapped
       with an access epoch of FENCE (which is not allowed), since FENCE may be
       ended up with not unsetting the window state. We can only detect if this
       access epoch is overlapped with another access epoch of PSCW or Passive
       Target. */
    MPIU_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,
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
526

527
    win_ptr->start_grp_size = group_ptr->size;
528

529 530 531 532 533 534 535 536 537 538 539 540
    if ((assert & MPI_MODE_NOCHECK) == 0) {
        int i, intra_cnt, inter_cnt;
        MPI_Request *intra_start_req = NULL;
        MPI_Status *intra_start_status = NULL;
        MPID_Comm *comm_ptr = win_ptr->comm_ptr;
        int rank = comm_ptr->rank;

        /* wait for messages from local processes */
        MPIU_CHKPMEM_MALLOC(win_ptr->start_ranks_in_win_grp, int *, win_ptr->start_grp_size * sizeof(int),
                            mpi_errno, "win_ptr->start_ranks_in_win_grp");
        mpi_errno = fill_ranks_in_win_grp(win_ptr, group_ptr, win_ptr->start_ranks_in_win_grp);
        if (mpi_errno) MPIU_ERR_POP(mpi_errno);
541

542 543 544 545
        /* post IRECVs */
        MPIU_CHKPMEM_MALLOC(win_ptr->start_req, MPI_Request *,
                            win_ptr->start_grp_size * sizeof(MPI_Request),
                            mpi_errno, "win_ptr->start_req");
546

547 548 549 550 551 552 553 554 555
        if (win_ptr->shm_allocated == TRUE) {
            int node_comm_size = comm_ptr->node_comm->local_size;
            MPIU_CHKLMEM_MALLOC(intra_start_req, MPI_Request *,
                                node_comm_size * sizeof(MPI_Request),
                                mpi_errno, "intra_start_req");
            MPIU_CHKLMEM_MALLOC(intra_start_status, MPI_Status *,
                                node_comm_size * sizeof(MPI_Status),
                                mpi_errno, "intra_start_status");
        }
556

557 558
        intra_cnt = 0;
        for (i = 0; i < win_ptr->start_grp_size; i++) {
559
            MPID_Request *req_ptr;
560 561
            MPIDI_VC_t *orig_vc = NULL, *target_vc = NULL;
            int src = win_ptr->start_ranks_in_win_grp[i];
562

563 564 565
            if (src != rank) {
                MPIDI_Comm_get_vc(comm_ptr, rank, &orig_vc);
                MPIDI_Comm_get_vc(comm_ptr, src, &target_vc);
566

567 568
                mpi_errno = MPID_Irecv(NULL, 0, MPI_INT, src, SYNC_POST_TAG,
                                       comm_ptr, MPID_CONTEXT_INTRA_PT2PT, &req_ptr);
569
                if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
570

571 572 573 574 575 576 577 578 579 580 581 582
                if (win_ptr->shm_allocated == TRUE &&
                    orig_vc->node_id == target_vc->node_id) {
                    intra_start_req[intra_cnt++] = req_ptr->handle;
                    win_ptr->start_req[i] = MPI_REQUEST_NULL;
                }
                else {
                    win_ptr->start_req[i] = req_ptr->handle;
                    inter_cnt++;
                }
            }
            else {
                win_ptr->start_req[i] = MPI_REQUEST_NULL;
583
            }
584
        }
585

586 587 588 589 590 591 592 593 594 595 596 597
        /* for targets on SHM, waiting until their IRECVs to be finished */
        if (intra_cnt) {
            mpi_errno = MPIR_Waitall_impl(intra_cnt, intra_start_req, intra_start_status);
            if (mpi_errno && mpi_errno != MPI_ERR_IN_STATUS)
                MPIU_ERR_POP(mpi_errno);
            /* --BEGIN ERROR HANDLING-- */
            if (mpi_errno == MPI_ERR_IN_STATUS) {
                for (i = 0; i < intra_cnt; i++) {
                    if (intra_start_status[i].MPI_ERROR != MPI_SUCCESS) {
                        mpi_errno = intra_start_status[i].MPI_ERROR;
                        MPIU_ERR_POP(mpi_errno);
                    }
598
                }
599
            }
600
            /* --END ERROR HANDLING-- */
601
        }
602

603 604 605 606
        if (win_ptr->shm_allocated == TRUE) {
            /* Ensure ordering of load/store operations */
            OPA_read_write_barrier();
        }
607 608
    }

609 610
    win_ptr->states.access_state = MPIDI_RMA_PSCW_ISSUED;
    num_active_issued_win++;
611

612 613
    MPIU_Assert(win_ptr->posted_ops_cnt == 0);
    MPIU_Assert(win_ptr->active_req_cnt == 0);
614

615
 fn_exit:
616 617 618
    MPIU_CHKLMEM_FREEALL();
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_START);
    return mpi_errno;
619 620 621
 fn_fail:
    MPIU_CHKPMEM_REAP();
    goto fn_exit;
622 623 624
}


625

626
#undef FUNCNAME
627
#define FUNCNAME MPIDI_Win_complete
628 629
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
630
int MPIDI_Win_complete(MPID_Win * win_ptr)
631
{
632
    int mpi_errno = MPI_SUCCESS;
633 634 635 636 637
    int i, dst, rank = win_ptr->comm_ptr->rank;
    int local_completed = 0, remote_completed = 0;
    MPID_Comm *win_comm_ptr = win_ptr->comm_ptr;
    MPIDI_RMA_Target_t *curr_target;
    int made_progress;
638
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_COMPLETE);
639

640
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_COMPLETE);
641

642 643 644
    /* Access epochs on the same window must be disjoint. */
    MPIU_ERR_CHKANDJUMP(win_ptr->states.access_state != MPIDI_RMA_PSCW_ISSUED &&
                        win_ptr->states.access_state != MPIDI_RMA_PSCW_GRANTED,
James Dinan's avatar
James Dinan committed
645 646
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");

647 648 649 650 651 652
    if (win_ptr->states.access_state == MPIDI_RMA_PSCW_ISSUED) {
        while (win_ptr->states.access_state != MPIDI_RMA_PSCW_GRANTED) {
            mpi_errno = wait_progress_engine();
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
        }
653 654
    }

655 656 657 658 659 660
    for (i = 0; i < win_ptr->start_grp_size; i++) {
        dst = win_ptr->start_ranks_in_win_grp[i];
        if (dst == rank) {
            win_ptr->at_completion_counter--;
            MPIU_Assert(win_ptr->at_completion_counter >= 0);
            continue;
661
        }
662

663 664 665 666 667 668 669
        if (win_comm_ptr->local_size <= win_ptr->num_slots)
            curr_target = win_ptr->slots[dst].target_list;
        else {
            curr_target = win_ptr->slots[dst % win_ptr->num_slots].target_list;
            while (curr_target != NULL && curr_target->target_rank != dst)
                curr_target = curr_target->next;
        }
670

671 672 673 674 675 676 677 678
        if (curr_target != NULL) {
            /* set sync_flag in sync struct */
            if (curr_target->sync.sync_flag < MPIDI_RMA_SYNC_FLUSH) {
                curr_target->sync.sync_flag = MPIDI_RMA_SYNC_FLUSH;
                curr_target->sync.have_remote_incomplete_ops = 0;
                curr_target->sync.outstanding_acks++;
            }
            curr_target->win_complete_flag = 1;
679 680
        }
        else {
681 682 683
            /* FIXME: do we need to wait for remote completion? */
            mpi_errno = send_decr_at_cnt_msg(dst, win_ptr);
            if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
684
        }
685 686
    }

687 688 689 690 691 692 693 694 695 696 697 698 699
    /* issue out all operations */
    mpi_errno = MPIDI_CH3I_RMA_Make_progress_win(win_ptr, &made_progress);
    if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);

    /* wait until all slots are empty */
    do {
        mpi_errno = MPIDI_CH3I_RMA_Cleanup_ops_win(win_ptr, &local_completed,
                                                   &remote_completed);
        if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
        if (!remote_completed) {
            mpi_errno = wait_progress_engine();
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
700
        }
701
    } while (!remote_completed);
702

703 704 705
    /* Cleanup all targets on this window. */
    mpi_errno = MPIDI_CH3I_RMA_Cleanup_targets_win(win_ptr);
    if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
706

707
    MPIU_Assert(win_ptr->non_empty_slots == 0);
708

709 710 711
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
712
    }
713

714 715 716 717 718 719 720
    /* free start group stored in window */
    MPIU_Free(win_ptr->start_ranks_in_win_grp);
    win_ptr->start_ranks_in_win_grp = NULL;

    win_ptr->posted_ops_cnt = 0;
    MPIU_Assert(win_ptr->active_req_cnt == 0);
    MPIU_Assert(win_ptr->start_req == NULL);
721

722
    win_ptr->states.access_state = MPIDI_RMA_NONE;
723

724 725
  fn_exit:
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_COMPLETE);
726
    return mpi_errno;
727 728 729 730
    /* --BEGIN ERROR HANDLING-- */
  fn_fail:
    goto fn_exit;
    /* --END ERROR HANDLING-- */
731
}
732

733 734


735
#undef FUNCNAME
736
#define FUNCNAME MPIDI_Win_wait
737 738
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
739
int MPIDI_Win_wait(MPID_Win * win_ptr)
740
{
741 742 743 744
    int mpi_errno = MPI_SUCCESS;
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_WAIT);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_WAIT);
745

746
    MPIU_ERR_CHKANDJUMP(win_ptr->states.exposure_state != MPIDI_RMA_PSCW_EXPO,
747
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
748

749
    /* wait for all operations from other processes to finish */
750 751 752 753
    while (win_ptr->at_completion_counter) {
        mpi_errno = wait_progress_engine();
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
754 755
    }

756 757 758
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
759 760
    }

761 762
    win_ptr->states.exposure_state = MPIDI_RMA_NONE;

763 764
  fn_exit:
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_WAIT);
765
    return mpi_errno;
766 767 768 769
    /* --BEGIN ERROR HANDLING-- */
  fn_fail:
    goto fn_exit;
    /* --END ERROR HANDLING-- */
770 771
}

772

773
#undef FUNCNAME
774
#define FUNCNAME MPIDI_Win_test
775 776
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
777
int MPIDI_Win_test(MPID_Win * win_ptr, int *flag)
778 779
{
    int mpi_errno = MPI_SUCCESS;
780
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_TEST);
781

782
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_TEST);
783

784
    MPIU_ERR_CHKANDJUMP(win_ptr->states.exposure_state != MPIDI_RMA_PSCW_EXPO,
785
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
786

787 788
    mpi_errno = MPID_Progress_test();
    if (mpi_errno != MPI_SUCCESS) {
789
	MPIU_ERR_POP(mpi_errno);
790 791
    }

792 793 794 795 796
    *flag = (win_ptr->at_completion_counter) ? 0 : 1;
    if (*flag) {
        /* Ensure ordering of load/store operations. */
        if (win_ptr->shm_allocated == TRUE) {
            OPA_read_write_barrier();
797
        }
798 799

        win_ptr->states.exposure_state = MPIDI_RMA_NONE;
800 801
    }

802
  fn_exit:
803
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_TEST);
804
    return mpi_errno;
805
    /* --BEGIN ERROR HANDLING-- */
806
  fn_fail:
807
    goto fn_exit;
808
    /* --END ERROR HANDLING-- */
809 810
}

811

812
#undef FUNCNAME
813
#define FUNCNAME MPIDI_Win_lock
814 815
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
816
int MPIDI_Win_lock(int lock_type, int dest, int assert, MPID_Win * win_ptr)
817
{
818 819 820 821 822
    int made_progress = 0;
    int shm_target = FALSE;
    int rank = win_ptr->comm_ptr->rank;
    MPIDI_RMA_Target_t *target = NULL;
    MPIDI_VC_t *orig_vc = NULL, *target_vc = NULL;
823
    int mpi_errno = MPI_SUCCESS;
824
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_LOCK);
825

826
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_LOCK);
827

828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845
    /* Note that here we cannot distinguish if this access epoch is overlapped
       with an access epoch of FENCE (which is not allowed), since FENCE may be
       ended up with not unsetting the window state. We can only detect if this
       access epoch is overlapped with another access epoch of PSCW or Passive
       Target. */
    if (win_ptr->lock_epoch_count == 0) {
        MPIU_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,
                            mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
    }
    else {
        MPIU_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.access_state != MPIDI_RMA_PER_TARGET,
                            mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
    }
846

847 848 849 850 851 852
    if (dest != MPI_PROC_NULL) {
        /* check if we lock the same target window more than once. */
        mpi_errno = MPIDI_CH3I_Win_find_target(win_ptr, dest, &target);
        if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
        MPIU_ERR_CHKANDJUMP(target != NULL, mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
    }
853

854
    /* Error handling is finished. */
855

856 857 858 859 860
    if (win_ptr->lock_epoch_count == 0) {
        win_ptr->states.access_state = MPIDI_RMA_PER_TARGET;
        num_passive_win++;
    }
    win_ptr->lock_epoch_count++;
861

862 863
    if (dest == MPI_PROC_NULL)
        goto fn_exit;
864

865 866 867 868 869
    if (win_ptr->shm_allocated == TRUE) {
        MPIDI_Comm_get_vc(win_ptr->comm_ptr, rank, &orig_vc);
        MPIDI_Comm_get_vc(win_ptr->comm_ptr, dest, &target_vc);
        if (orig_vc->node_id == target_vc->node_id)
            shm_target = TRUE;
870
    }
871

872 873 874
    /* Create a new target. */
    mpi_errno = MPIDI_CH3I_Win_create_target(win_ptr, dest, &target);
    if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
875

876 877 878 879 880 881 882 883 884 885 886 887 888 889
    /* Store lock_state (CALLED/ISSUED/GRANTED), lock_type (SHARED/EXCLUSIVE),
       lock_mode (MODE_NOCHECK). */
    if (assert & MPI_MODE_NOCHECK)
        target->access_state = MPIDI_RMA_LOCK_GRANTED;
    else
        target->access_state = MPIDI_RMA_LOCK_CALLED;
    target->lock_type = lock_type;
    target->lock_mode = assert;

    /* If Destination is myself or process on SHM, acquire the lock,
       wait until lock is granted. */
    if (!(assert & MPI_MODE_NOCHECK) && (dest == rank || shm_target)) {
        mpi_errno = MPIDI_CH3I_RMA_Make_progress_target(win_ptr, dest, &made_progress);
        if (mpi_errno != MPI_SUCCESS)
890 891
            MPIU_ERR_POP(mpi_errno);

892 893 894
        while (target->access_state != MPIDI_RMA_LOCK_GRANTED) {
            mpi_errno = wait_progress_engine();
            if (mpi_errno != MPI_SUCCESS)
895
                MPIU_ERR_POP(mpi_errno);
896
        }
897
    }
898

899 900 901
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
902 903
    }

904 905 906 907 908 909 910
  fn_exit:
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_LOCK);
    return mpi_errno;
    /* --BEGIN ERROR HANDLING-- */
  fn_fail:
    goto fn_exit;
    /* --END ERROR HANDLING-- */
911 912 913
}

#undef FUNCNAME
914
#define FUNCNAME MPIDI_Win_unlock
915 916
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
917
int MPIDI_Win_unlock(int dest, MPID_Win *win_ptr)
918
{
919 920 921 922
    int made_progress = 0;
    int local_completed = 0, remote_completed = 0;
    MPIDI_RMA_Target_t *target = NULL;
    enum MPIDI_RMA_sync_types sync_flag;
923
    int mpi_errno = MPI_SUCCESS;
924
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_UNLOCK);
925

926
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_UNLOCK);
927

928
    MPIU_ERR_CHKANDJUMP(win_ptr->states.access_state != MPIDI_RMA_PER_TARGET,
929
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
930

931
    /* Ensure ordering of load/store operations. */
932
    if (win_ptr->shm_allocated) {
933 934
        OPA_read_write_barrier();
    }
935

936 937 938 939 940 941 942 943 944 945 946 947 948
    if (dest == MPI_PROC_NULL)
        goto finish_unlock;

    /* When the process tries to acquire the lock on itself, it does not
       go through the progress engine. Therefore, it is possible that
       one process always grants the lock to itself but never process
       events coming from other processes. This may cause deadlock in
       applications where the program execution on target process depends
       on the happening of events from other processes. Here we poke
       the progress engine once to avoid such issue.  */
    mpi_errno = poke_progress_engine();
    if (mpi_errno != MPI_SUCCESS)
        MPIU_ERR_POP(mpi_errno);
949

950 951 952 953 954 955 956
    /* Find or recreate target. */
    mpi_errno = MPIDI_CH3I_Win_find_target(win_ptr, dest, &target);
    if (mpi_errno != MPI_SUCCESS)
        MPIU_ERR_POP(mpi_errno);
    if (target == NULL) {
        mpi_errno = MPIDI_CH3I_Win_create_target(win_ptr, dest, &target);
        if (mpi_errno != MPI_SUCCESS)
957
            MPIU_ERR_POP(mpi_errno);
958
        target->access_state = MPIDI_RMA_LOCK_GRANTED;
959 960
    }

961 962 963 964 965 966 967 968 969
    /* Set sync_flag in sync struct. */
    if (target->lock_mode & MPI_MODE_NOCHECK)
        sync_flag = MPIDI_RMA_SYNC_FLUSH;
    else
        sync_flag = MPIDI_RMA_SYNC_UNLOCK;
    if (target->sync.sync_flag < sync_flag) {
        target->sync.sync_flag = sync_flag;
        target->sync.have_remote_incomplete_ops = 0;
        target->sync.outstanding_acks++;
970
    }
971

972 973 974 975 976
    /* Issue out all operations. */
    mpi_errno = MPIDI_CH3I_RMA_Make_progress_target(win_ptr, dest,
                                                    &made_progress);
    if (mpi_errno != MPI_SUCCESS)
        MPIU_ERR_POP(mpi_errno);
977

978 979 980 981 982 983 984 985 986 987
    /* Wait for remote completion. */
    do {
        mpi_errno = MPIDI_CH3I_RMA_Cleanup_ops_target(win_ptr, target,
                                                      &local_completed,
                                                      &remote_completed);
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
        if (!remote_completed) {
            mpi_errno = wait_progress_engine();
            if (mpi_errno != MPI_SUCCESS)
988
                MPIU_ERR_POP(mpi_errno);
989
        }
990
    } while (!remote_completed);
991

992 993 994
    /* Cleanup the target. */
    mpi_errno = MPIDI_CH3I_RMA_Cleanup_single_target(win_ptr, target);
    if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
995

996 997 998
 finish_unlock:
    win_ptr->posted_ops_cnt = 0;
    MPIU_Assert(win_ptr->active_req_cnt == 0);
999

1000 1001 1002 1003 1004 1005
    win_ptr->lock_epoch_count--;
    if (win_ptr->lock_epoch_count == 0) {
        win_ptr->states.access_state = MPIDI_RMA_NONE;
        num_passive_win--;
        MPIU_Assert(num_passive_win >= 0);
    }
1006

1007
  fn_exit:
1008
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_UNLOCK);
1009
    return mpi_errno;
1010
    /* --BEGIN ERROR HANDLING-- */
1011
  fn_fail:
1012
    goto fn_exit;
1013
    /* --END ERROR HANDLING-- */
1014 1015
}

1016

1017
#undef FUNCNAME
1018
#define FUNCNAME MPIDI_Win_flush_all
1019 1020
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
1021
int MPIDI_Win_flush_all(MPID_Win * win_ptr)
1022
{
1023 1024