ch3u_rma_sync.c 68.8 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.
*/

Xin Zhao's avatar
Xin Zhao committed
212 213 214 215 216 217 218 219
MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_lockqueue_alloc);
MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winlock_getlocallock);
MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_wincreate_allgather);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_rmaqueue_alloc);
MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_rmaqueue_set);

void MPIDI_CH3_RMA_Init_sync_pvars(void)
220
{
Xin Zhao's avatar
Xin Zhao committed
221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264
    /* rma_lockqueue_alloc */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(RMA,
                                      MPI_DOUBLE,
                                      rma_lockqueue_alloc,
                                      MPI_T_VERBOSITY_MPIDEV_DETAIL,
                                      MPI_T_BIND_NO_OBJECT,
                                      MPIR_T_PVAR_FLAG_READONLY,
                                      "RMA", "Allocate Lock Queue element (in seconds)");

    /* rma_winlock_getlocallock */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(RMA,
                                      MPI_DOUBLE,
                                      rma_winlock_getlocallock,
                                      MPI_T_VERBOSITY_MPIDEV_DETAIL,
                                      MPI_T_BIND_NO_OBJECT,
                                      MPIR_T_PVAR_FLAG_READONLY,
                                      "RMA", "WIN_LOCK:Get local lock (in seconds)");

    /* rma_wincreate_allgather */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(RMA,
                                      MPI_DOUBLE,
                                      rma_wincreate_allgather,
                                      MPI_T_VERBOSITY_MPIDEV_DETAIL,
                                      MPI_T_BIND_NO_OBJECT,
                                      MPIR_T_PVAR_FLAG_READONLY,
                                      "RMA", "WIN_CREATE:Allgather (in seconds)");

    /* rma_rmaqueue_alloc */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(RMA,
                                      MPI_DOUBLE,
                                      rma_rmaqueue_alloc,
                                      MPI_T_VERBOSITY_MPIDEV_DETAIL,
                                      MPI_T_BIND_NO_OBJECT,
                                      MPIR_T_PVAR_FLAG_READONLY,
                                      "RMA", "Allocate RMA Queue element (in seconds)");

    /* rma_rmaqueue_set */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(RMA,
                                      MPI_DOUBLE,
                                      rma_rmaqueue_set,
                                      MPI_T_VERBOSITY_MPIDEV_DETAIL,
                                      MPI_T_BIND_NO_OBJECT,
                                      MPIR_T_PVAR_FLAG_READONLY,
                                      "RMA", "Set fields in RMA Queue element (in seconds)");
265
}
266

267 268
/* These are used to use a common routine to complete lists of RMA
   operations with a single routine, while collecting data that
269 270 271 272
   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.
*/
273

274 275 276
/*
 * These routines provide a default implementation of the MPI RMA operations
 * in terms of the low-level, two-sided channel operations.  A channel
277 278
 * may override these functions, on a per-window basis, by overriding
 * the MPID functions in the RMAFns section of MPID_Win object.
279 280
 */

281 282
#define SYNC_POST_TAG 100

283

284 285 286 287 288
/********************************************************************************/
/* Active Target synchronization (including WIN_FENCE, WIN_POST, WIN_START,     */
/* WIN_COMPLETE, WIN_WAIT, WIN_TEST)                                            */
/********************************************************************************/

289 290 291 292
#undef FUNCNAME
#define FUNCNAME MPIDI_Win_fence
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
293
int MPIDI_Win_fence(int assert, MPID_Win * win_ptr)
294
{
295 296 297
    int i, made_progress = 0;
    int local_completed = 0, remote_completed = 0;
    MPIDI_RMA_Target_t *curr_target = NULL;
Wesley Bland's avatar
Wesley Bland committed
298
    mpir_errflag_t errflag = MPIR_ERR_NONE;
299
    int progress_engine_triggered = 0;
300
    int mpi_errno = MPI_SUCCESS;
301 302 303 304
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_FENCE);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_FENCE);

305 306 307 308
    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
309 310
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");

311 312 313 314 315
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }

316 317
    if (assert & MPI_MODE_NOPRECEDE) {
        if (assert & MPI_MODE_NOSUCCEED) {
318
            goto finish_fence;
319
        }
320 321
        else {
            /* It is possible that there is a IBARRIER in MPI_WIN_FENCE with
322 323
             * MODE_NOPRECEDE not being completed, we let the progress engine
             * to delete its request when it is completed. */
324 325 326 327 328 329 330
            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;
            }
331

332 333
            if (win_ptr->shm_allocated == TRUE) {
                MPID_Comm *node_comm_ptr = win_ptr->comm_ptr->node_comm;
334

335
                mpi_errno = MPIR_Barrier_impl(node_comm_ptr, &errflag);
336 337
                if (mpi_errno != MPI_SUCCESS)
                    MPIU_ERR_POP(mpi_errno);
338
                MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
339

340
                /* Mark that we triggered the progress engine
341
                 * in this function call. */
342
                progress_engine_triggered = 1;
343 344
            }

345
            mpi_errno = MPIR_Ibarrier_impl(win_ptr->comm_ptr, &(win_ptr->fence_sync_req));
346 347
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
348

349
            /* Set window access state properly. */
350
            win_ptr->states.access_state = MPIDI_RMA_FENCE_ISSUED;
351
            MPIDI_CH3I_num_active_issued_win++;
352

353
            goto finish_fence;
354
        }
355
    }
356

357 358 359
    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();
360 361
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
362 363

            /* Mark that we triggered the progress engine
364
             * in this function call. */
365
            progress_engine_triggered = 1;
366
        }
367
    }
368

369 370
    /* Set sync_flag in target structs. */
    for (i = 0; i < win_ptr->num_slots; i++) {
371
        curr_target = win_ptr->slots[i].target_list_head;
372 373 374 375 376
        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;
377
            }
378
            curr_target = curr_target->next;
379
        }
380
    }
381

382 383
    /* Issue out all operations. */
    mpi_errno = MPIDI_CH3I_RMA_Make_progress_win(win_ptr, &made_progress);
384 385
    if (mpi_errno != MPI_SUCCESS)
        MPIU_ERR_POP(mpi_errno);
386 387 388

    /* Wait for remote completion. */
    do {
389 390 391
        mpi_errno = MPIDI_CH3I_RMA_Cleanup_ops_win(win_ptr, &local_completed, &remote_completed);
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
392 393 394 395
        if (!remote_completed) {
            mpi_errno = wait_progress_engine();
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
396 397

            /* Mark that we triggered the progress engine
398
             * in this function call. */
399
            progress_engine_triggered = 1;
400
        }
401
    } while (!remote_completed);
402

403 404
    /* Cleanup all targets on window. */
    mpi_errno = MPIDI_CH3I_RMA_Cleanup_targets_win(win_ptr);
405 406
    if (mpi_errno != MPI_SUCCESS)
        MPIU_ERR_POP(mpi_errno);
James Dinan's avatar
James Dinan committed
407

408
    mpi_errno = MPIR_Barrier_impl(win_ptr->comm_ptr, &errflag);
409 410
    if (mpi_errno != MPI_SUCCESS)
        MPIU_ERR_POP(mpi_errno);
411
    MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
412

413
    /* Mark that we triggered the progress engine
414
     * in this function call. */
415 416
    progress_engine_triggered = 1;

417
    /* Set window access state properly. */
418 419 420 421 422
    if (assert & MPI_MODE_NOSUCCEED) {
        win_ptr->states.access_state = MPIDI_RMA_NONE;
    }
    else {
        win_ptr->states.access_state = MPIDI_RMA_FENCE_GRANTED;
423 424
    }

425
  finish_fence:
426 427 428
    /* Make sure that all targets are freed. */
    MPIU_Assert(win_ptr->non_empty_slots == 0);

429 430 431 432 433 434 435 436 437
    if (assert & MPI_MODE_NOPRECEDE) {
        /* BEGINNING synchronization: the following counter should be zero. */
        MPIU_Assert(win_ptr->accumulated_ops_cnt == 0);
    }
    else {
        /* ENDING synchronization: correctly decrement the following counter. */
        win_ptr->accumulated_ops_cnt = 0;
    }

438 439
    MPIU_Assert(win_ptr->active_req_cnt == 0);

440 441 442
    if (!(assert & MPI_MODE_NOPRECEDE)) {
        if (!progress_engine_triggered) {
            /* In some cases (e.g. target is myself, or process on SHM),
443 444 445 446 447 448
             * this function call does not go through the progress engine.
             * Therefore, it is possible that this process never process
             * events coming from other processes. This may cause deadlock in
             * applications where the program execution on this process depends
             * on the happening of events from other processes. Here we poke
             * the progress engine once to avoid such issue.  */
449 450 451 452 453 454
            mpi_errno = poke_progress_engine();
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
        }
    }

455 456 457 458 459
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }

460
  fn_exit:
461 462 463
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_FENCE);
    return mpi_errno;
    /* --BEGIN ERROR HANDLING-- */
464
  fn_fail:
465 466 467 468 469 470
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}


#undef FUNCNAME
471
#define FUNCNAME MPIDI_Win_post
472 473
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
474
int MPIDI_Win_post(MPID_Group * post_grp_ptr, int assert, MPID_Win * win_ptr)
475
{
476
    int *post_ranks_in_win_grp;
477
    int mpi_errno = MPI_SUCCESS;
478
    MPIU_CHKLMEM_DECL(3);
479
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_POST);
480

481
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_POST);
482

483
    /* Note that here we cannot distinguish if this exposure epoch is overlapped
484 485 486
     * 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. */
487
    MPIU_ERR_CHKANDJUMP(win_ptr->states.exposure_state != MPIDI_RMA_NONE,
488
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
489

490 491 492
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
493 494
    }

495
    /* Set window exposure state properly. */
496 497 498 499
    win_ptr->states.exposure_state = MPIDI_RMA_PSCW_EXPO;

    win_ptr->at_completion_counter += post_grp_ptr->size;

500 501 502
    if ((assert & MPI_MODE_NOCHECK) == 0) {
        MPI_Request *req;
        MPI_Status *status;
503 504
        int i, post_grp_size, dst, rank;
        MPID_Comm *win_comm_ptr;
505

506 507
        /* NOCHECK not specified. We need to notify the source
         * processes that Post has been called. */
508

509
        post_grp_size = post_grp_ptr->size;
510 511
        win_comm_ptr = win_ptr->comm_ptr;
        rank = win_ptr->comm_ptr->rank;
512

513 514 515
        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);
516 517
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
518 519 520 521 522

        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");
523

524 525
        /* Send a 0-byte message to the source processes */
        for (i = 0; i < post_grp_size; i++) {
526
            dst = post_ranks_in_win_grp[i];
527

528 529 530 531
            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);
532 533
                if (mpi_errno != MPI_SUCCESS)
                    MPIU_ERR_POP(mpi_errno);
534 535 536 537 538
                req[i] = req_ptr->handle;
            }
            else {
                req[i] = MPI_REQUEST_NULL;
            }
539
        }
540

541 542 543 544 545 546 547 548 549 550 551 552
        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);
                }
            }
553
        }
554
        /* --END ERROR HANDLING-- */
555
    }
556

557
  fn_exit:
558 559
    MPIU_CHKLMEM_FREEALL();
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_POST);
560 561
    return mpi_errno;
    /* --BEGIN ERROR HANDLING-- */
562
  fn_fail:
563 564 565 566
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}

567

568 569 570 571 572
#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)
573
{
574
    int mpi_errno = MPI_SUCCESS;
575
    MPIU_CHKLMEM_DECL(2);
576 577
    MPIU_CHKPMEM_DECL(2);
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_START);
578

579
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_START);
580

581
    /* Note that here we cannot distinguish if this access epoch is overlapped
582 583 584 585
     * 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. */
586 587 588 589
    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");
590

591
    win_ptr->start_grp_size = group_ptr->size;
592

593 594 595 596 597
    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);
598 599
    if (mpi_errno)
        MPIU_ERR_POP(mpi_errno);
600

601
    if ((assert & MPI_MODE_NOCHECK) == 0) {
Xin Zhao's avatar
Xin Zhao committed
602
        int i, intra_cnt;
603 604 605 606 607 608
        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 */
609

610 611 612 613
        /* 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");
614

615 616 617
        if (win_ptr->shm_allocated == TRUE) {
            int node_comm_size = comm_ptr->node_comm->local_size;
            MPIU_CHKLMEM_MALLOC(intra_start_req, MPI_Request *,
618
                                node_comm_size * sizeof(MPI_Request), mpi_errno, "intra_start_req");
619 620 621 622
            MPIU_CHKLMEM_MALLOC(intra_start_status, MPI_Status *,
                                node_comm_size * sizeof(MPI_Status),
                                mpi_errno, "intra_start_status");
        }
623

624 625
        intra_cnt = 0;
        for (i = 0; i < win_ptr->start_grp_size; i++) {
626
            MPID_Request *req_ptr;
627 628
            MPIDI_VC_t *orig_vc = NULL, *target_vc = NULL;
            int src = win_ptr->start_ranks_in_win_grp[i];
629

630 631 632
            if (src != rank) {
                MPIDI_Comm_get_vc(comm_ptr, rank, &orig_vc);
                MPIDI_Comm_get_vc(comm_ptr, src, &target_vc);
633

634 635
                mpi_errno = MPID_Irecv(NULL, 0, MPI_INT, src, SYNC_POST_TAG,
                                       comm_ptr, MPID_CONTEXT_INTRA_PT2PT, &req_ptr);
636 637
                if (mpi_errno != MPI_SUCCESS)
                    MPIU_ERR_POP(mpi_errno);
638

639
                if (win_ptr->shm_allocated == TRUE && orig_vc->node_id == target_vc->node_id) {
640 641 642 643 644 645 646 647 648
                    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;
                }
            }
            else {
                win_ptr->start_req[i] = MPI_REQUEST_NULL;
649
            }
650
        }
651

652 653 654 655 656 657 658 659 660 661 662 663
        /* 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);
                    }
664
                }
665
            }
666
            /* --END ERROR HANDLING-- */
667
        }
668 669
    }

670
  finish_start:
671
    /* Set window access state properly. */
672
    win_ptr->states.access_state = MPIDI_RMA_PSCW_ISSUED;
673
    MPIDI_CH3I_num_active_issued_win++;
674

675 676 677
    /* BEGINNING synchronization: the following counter should be zero. */
    MPIU_Assert(win_ptr->accumulated_ops_cnt == 0);

678
    MPIU_Assert(win_ptr->active_req_cnt == 0);
679

680 681 682 683 684
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }

685
  fn_exit:
686 687 688
    MPIU_CHKLMEM_FREEALL();
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_START);
    return mpi_errno;
689
  fn_fail:
690 691
    MPIU_CHKPMEM_REAP();
    goto fn_exit;
692 693 694
}


695

696
#undef FUNCNAME
697
#define FUNCNAME MPIDI_Win_complete
698 699
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
700
int MPIDI_Win_complete(MPID_Win * win_ptr)
701
{
702
    int mpi_errno = MPI_SUCCESS;
703 704 705 706
    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;
707
    int progress_engine_triggered = 0;
708
    int made_progress;
709
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_COMPLETE);
710

711
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_COMPLETE);
712

713 714 715
    /* 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
716 717
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");

718 719 720 721 722
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }

723 724 725 726 727
    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);
728 729

            /* Mark that we triggered the progress engine
730
             * in this function call. */
731
            progress_engine_triggered = 1;
732
        }
733 734
    }

735 736 737 738 739 740
    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;
741
        }
742

743
        if (win_comm_ptr->local_size <= win_ptr->num_slots)
744
            curr_target = win_ptr->slots[dst].target_list_head;
745
        else {
746
            curr_target = win_ptr->slots[dst % win_ptr->num_slots].target_list_head;
747 748 749
            while (curr_target != NULL && curr_target->target_rank != dst)
                curr_target = curr_target->next;
        }
750

751 752 753 754 755 756
        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->win_complete_flag = 1;
757 758
        }
        else {
759 760
            /* FIXME: do we need to wait for remote completion? */
            mpi_errno = send_decr_at_cnt_msg(dst, win_ptr);
761 762
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
763
        }
764 765
    }

766 767
    /* issue out all operations */
    mpi_errno = MPIDI_CH3I_RMA_Make_progress_win(win_ptr, &made_progress);
768 769
    if (mpi_errno != MPI_SUCCESS)
        MPIU_ERR_POP(mpi_errno);
770 771 772

    /* wait until all slots are empty */
    do {
773 774 775
        mpi_errno = MPIDI_CH3I_RMA_Cleanup_ops_win(win_ptr, &local_completed, &remote_completed);
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
776 777 778 779
        if (!remote_completed) {
            mpi_errno = wait_progress_engine();
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
780 781

            /* Mark that we triggered the progress engine
782
             * in this function call. */
783
            progress_engine_triggered = 1;
784
        }
785
    } while (!remote_completed);
786

787 788
    /* Cleanup all targets on this window. */
    mpi_errno = MPIDI_CH3I_RMA_Cleanup_targets_win(win_ptr);
789 790
    if (mpi_errno != MPI_SUCCESS)
        MPIU_ERR_POP(mpi_errno);
791

792
  finish_complete:
793
    /* Set window access state properly. */
794
    win_ptr->states.access_state = MPIDI_RMA_NONE;
795

796 797 798 799
    /* free start group stored in window */
    MPIU_Free(win_ptr->start_ranks_in_win_grp);
    win_ptr->start_ranks_in_win_grp = NULL;
    MPIU_Assert(win_ptr->start_req == NULL);
800

801 802 803
    /* Make sure that all targets are freed. */
    MPIU_Assert(win_ptr->non_empty_slots == 0);

804 805 806
    /* ENDING synchronization: correctly decrement the following counter. */
    win_ptr->accumulated_ops_cnt = 0;

807
    MPIU_Assert(win_ptr->active_req_cnt == 0);
808

809 810
    if (!progress_engine_triggered) {
        /* In some cases (e.g. target is myself, or process on SHM),
811 812 813 814 815 816
         * this function call does not go through the progress engine.
         * Therefore, it is possible that this process never process
         * events coming from other processes. This may cause deadlock in
         * applications where the program execution on this process depends
         * on the happening of events from other processes. Here we poke
         * the progress engine once to avoid such issue.  */
817 818 819 820 821
        mpi_errno = poke_progress_engine();
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
    }

822 823
  fn_exit:
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_COMPLETE);
824
    return mpi_errno;
825 826 827 828
    /* --BEGIN ERROR HANDLING-- */
  fn_fail:
    goto fn_exit;
    /* --END ERROR HANDLING-- */
829
}
830

831 832


833
#undef FUNCNAME
834
#define FUNCNAME MPIDI_Win_wait
835 836
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
837
int MPIDI_Win_wait(MPID_Win * win_ptr)
838
{
839
    int progress_engine_triggered = 0;
840 841 842 843
    int mpi_errno = MPI_SUCCESS;
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_WAIT);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_WAIT);
844

845
    MPIU_ERR_CHKANDJUMP(win_ptr->states.exposure_state != MPIDI_RMA_PSCW_EXPO,
846
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
847

848
    /* wait for all operations from other processes to finish */
849 850 851 852
    while (win_ptr->at_completion_counter) {
        mpi_errno = wait_progress_engine();
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
853 854

        /* Mark that we triggered the progress engine
855
         * in this function call. */
856
        progress_engine_triggered = 1;
857 858
    }

859
  finish_wait:
860
    /* Set window exposure state properly. */
861 862
    win_ptr->states.exposure_state = MPIDI_RMA_NONE;

863 864
    if (!progress_engine_triggered) {
        /* In some cases (e.g. target is myself, or process on SHM),
865 866 867 868 869 870
         * this function call does not go through the progress engine.
         * Therefore, it is possible that this process never process
         * events coming from other processes. This may cause deadlock in
         * applications where the program execution on this process depends
         * on the happening of events from other processes. Here we poke
         * the progress engine once to avoid such issue.  */
871 872 873 874 875
        mpi_errno = poke_progress_engine();
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
    }

876 877 878 879 880
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }

881 882
  fn_exit:
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_WAIT);
883
    return mpi_errno;
884 885 886 887
    /* --BEGIN ERROR HANDLING-- */
  fn_fail:
    goto fn_exit;
    /* --END ERROR HANDLING-- */
888 889
}

890

891
#undef FUNCNAME
892
#define FUNCNAME MPIDI_Win_test
893 894
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
895
int MPIDI_Win_test(MPID_Win * win_ptr, int *flag)
896 897
{
    int mpi_errno = MPI_SUCCESS;
898
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_TEST);
899

900
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_TEST);
901

902
    MPIU_ERR_CHKANDJUMP(win_ptr->states.exposure_state != MPIDI_RMA_PSCW_EXPO,
903
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
904

905 906
    mpi_errno = MPID_Progress_test();
    if (mpi_errno != MPI_SUCCESS) {
907
        MPIU_ERR_POP(mpi_errno);
908 909
    }

910 911
    *flag = (win_ptr->at_completion_counter) ? 0 : 1;
    if (*flag) {
912
        /* Set window exposure state properly. */
913 914
        win_ptr->states.exposure_state = MPIDI_RMA_NONE;

915 916 917
        /* Ensure ordering of load/store operations. */
        if (win_ptr->shm_allocated == TRUE) {
            OPA_read_write_barrier();
918 919 920
        }
    }

921
  fn_exit:
922
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_TEST);
923
    return mpi_errno;
924
    /* --BEGIN ERROR HANDLING-- */
925
  fn_fail:
926
    goto fn_exit;
927
    /* --END ERROR HANDLING-- */
928 929
}

930

931 932 933 934 935 936
/********************************************************************************/
/* Passive Target synchronization (including WIN_LOCK, WIN_UNLOCK, WIN_FLUSH,   */
/* WIN_FLUSH_LOCAL, WIN_LOCK_ALL, WIN_UNLOCK_ALL, WIN_FLUSH_ALL,                */
/* WIN_FLUSH_LOCAL_ALL, WIN_SYNC)                                               */
/********************************************************************************/

937
#undef FUNCNAME
938
#define FUNCNAME MPIDI_Win_lock
939 940
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
941
int MPIDI_Win_lock(int lock_type, int dest, int assert, MPID_Win * win_ptr)
942
{
943 944 945 946 947
    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;
948
    int mpi_errno = MPI_SUCCESS;
949
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_LOCK);
950

951
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_LOCK);
952

953
    /* Note that here we cannot distinguish if this access epoch is overlapped
954 955 956 957
     * 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. */
958 959 960 961 962 963 964
    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 {
Xin Zhao's avatar
Xin Zhao committed
965
        MPIU_ERR_CHKANDJUMP(win_ptr->states.access_state != MPIDI_RMA_PER_TARGET,
966 967
                            mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
    }
968

969 970 971
    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);
972 973
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
974 975
        MPIU_ERR_CHKANDJUMP(target != NULL, mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
    }
976

977
    /* Error handling is finished. */
978

979
    if (win_ptr->lock_epoch_count == 0) {
980
        /* Set window access state properly. */
981
        win_ptr->states.access_state = MPIDI_RMA_PER_TA