ch3u_rma_sync.c 69.4 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.
*/

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
    if (assert & MPI_MODE_NOPRECEDE) {
        if (assert & MPI_MODE_NOSUCCEED) {
313
            goto finish_fence;
314
        }
315 316 317 318 319 320 321 322 323 324 325
        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;
            }
326

327 328
            if (win_ptr->shm_allocated == TRUE) {
                MPID_Comm *node_comm_ptr = win_ptr->comm_ptr->node_comm;
329

330 331
                /* Ensure ordering of load/store operations. */
                OPA_read_write_barrier();
332

333 334 335
                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");
336

337 338 339 340
                /* Mark that we triggered the progress engine
                   in this function call. */
                progress_engine_triggered = 1;

341 342
                /* Ensure ordering of load/store operations. */
                OPA_read_write_barrier();
343 344
            }

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

348 349
            win_ptr->states.access_state = MPIDI_RMA_FENCE_ISSUED;
            num_active_issued_win++;
350

351
            goto finish_fence;
352
        }
353
    }
354

355 356 357
    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();
358 359
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
360 361 362 363

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

367 368 369 370 371 372 373 374 375 376
    /* 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++;
377
            }
378
            curr_target = curr_target->next;
379
        }
380
    }
381

382 383 384 385 386 387 388 389 390 391 392 393 394 395
    /* 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);
396 397 398 399

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

403 404 405
    /* 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
406

407
    MPIU_Assert(win_ptr->non_empty_slots == 0);
408

409 410 411 412
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }
413

414 415 416
    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");
417

418 419 420 421
    /* Mark that we triggered the progress engine
       in this function call. */
    progress_engine_triggered = 1;

422 423 424 425
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }
426

427 428 429 430 431
    if (assert & MPI_MODE_NOSUCCEED) {
        win_ptr->states.access_state = MPIDI_RMA_NONE;
    }
    else {
        win_ptr->states.access_state = MPIDI_RMA_FENCE_GRANTED;
432 433
    }

434 435 436 437 438 439 440 441 442 443
 finish_fence:
    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;
    }

444 445
    MPIU_Assert(win_ptr->active_req_cnt == 0);

446 447 448 449 450 451 452 453 454 455 456 457 458 459 460
    if (!(assert & MPI_MODE_NOPRECEDE)) {
        if (!progress_engine_triggered) {
            /* In some cases (e.g. target is myself, or process on SHM),
               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.  */
            mpi_errno = poke_progress_engine();
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
        }
    }

461
  fn_exit:
462 463 464
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_FENCE);
    return mpi_errno;
    /* --BEGIN ERROR HANDLING-- */
465
  fn_fail:
466 467 468 469 470 471
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}


#undef FUNCNAME
472
#define FUNCNAME MPIDI_Win_post
473 474
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
475
int MPIDI_Win_post(MPID_Group * post_grp_ptr, int assert, MPID_Win * win_ptr)
476
{
477
    int *post_ranks_in_win_grp;
478
    int mpi_errno = MPI_SUCCESS;
479
    MPIU_CHKLMEM_DECL(3);
480
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_POST);
481

482
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_POST);
483

484 485 486 487 488
    /* 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,
489
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
490

491
    win_ptr->states.exposure_state = MPIDI_RMA_PSCW_EXPO;
492

493
    win_ptr->at_completion_counter += post_grp_ptr->size;
494

495 496 497
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
498 499
    }

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

513 514 515 516 517 518 519 520 521
        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");
522

523 524
        /* Send a 0-byte message to the source processes */
        for (i = 0; i < post_grp_size; i++) {
525
            dst = post_ranks_in_win_grp[i];
526

527 528 529 530
            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);
531
                if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
532 533 534 535 536
                req[i] = req_ptr->handle;
            }
            else {
                req[i] = MPI_REQUEST_NULL;
            }
537
        }
538

539 540 541 542 543 544 545 546 547 548 549 550
        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);
                }
            }
551
        }
552
        /* --END ERROR HANDLING-- */
553
    }
554

555
  fn_exit:
556 557
    MPIU_CHKLMEM_FREEALL();
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_POST);
558 559
    return mpi_errno;
    /* --BEGIN ERROR HANDLING-- */
560
  fn_fail:
561 562 563 564
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}

565

566 567 568 569 570
#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)
571
{
572
    int mpi_errno = MPI_SUCCESS;
573
    MPIU_CHKLMEM_DECL(2);
574 575
    MPIU_CHKPMEM_DECL(2);
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_START);
576

577
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_START);
578

579 580 581 582 583 584 585 586 587
    /* 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");
588

589
    win_ptr->start_grp_size = group_ptr->size;
590

591
    if ((assert & MPI_MODE_NOCHECK) == 0) {
Xin Zhao's avatar
Xin Zhao committed
592
        int i, intra_cnt;
593 594 595 596 597 598 599 600 601 602
        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);
603

604 605 606 607
        /* 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");
608

609 610 611 612 613 614 615 616 617
        if (win_ptr->shm_allocated == TRUE) {
            int node_comm_size = comm_ptr->node_comm->local_size;
            MPIU_CHKLMEM_MALLOC(intra_start_req, MPI_Request *,
                                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");
        }
618

619 620
        intra_cnt = 0;
        for (i = 0; i < win_ptr->start_grp_size; i++) {
621
            MPID_Request *req_ptr;
622 623
            MPIDI_VC_t *orig_vc = NULL, *target_vc = NULL;
            int src = win_ptr->start_ranks_in_win_grp[i];
624

625 626 627
            if (src != rank) {
                MPIDI_Comm_get_vc(comm_ptr, rank, &orig_vc);
                MPIDI_Comm_get_vc(comm_ptr, src, &target_vc);
628

629 630
                mpi_errno = MPID_Irecv(NULL, 0, MPI_INT, src, SYNC_POST_TAG,
                                       comm_ptr, MPID_CONTEXT_INTRA_PT2PT, &req_ptr);
631
                if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
632

633 634 635 636 637 638 639 640 641 642 643
                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;
                }
            }
            else {
                win_ptr->start_req[i] = MPI_REQUEST_NULL;
644
            }
645
        }
646

647 648 649 650 651 652 653 654 655 656 657 658
        /* 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);
                    }
659
                }
660
            }
661
            /* --END ERROR HANDLING-- */
662
        }
663

664 665 666 667
        if (win_ptr->shm_allocated == TRUE) {
            /* Ensure ordering of load/store operations */
            OPA_read_write_barrier();
        }
668 669
    }

670 671
    win_ptr->states.access_state = MPIDI_RMA_PSCW_ISSUED;
    num_active_issued_win++;
672

673 674 675 676
 finish_start:
    /* BEGINNING synchronization: the following counter should be zero. */
    MPIU_Assert(win_ptr->accumulated_ops_cnt == 0);

677
    MPIU_Assert(win_ptr->active_req_cnt == 0);
678

679
 fn_exit:
680 681 682
    MPIU_CHKLMEM_FREEALL();
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_START);
    return mpi_errno;
683 684 685
 fn_fail:
    MPIU_CHKPMEM_REAP();
    goto fn_exit;
686 687 688
}


689

690
#undef FUNCNAME
691
#define FUNCNAME MPIDI_Win_complete
692 693
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
694
int MPIDI_Win_complete(MPID_Win * win_ptr)
695
{
696
    int mpi_errno = MPI_SUCCESS;
697 698 699 700
    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;
701
    int progress_engine_triggered = 0;
702
    int made_progress;
703
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_COMPLETE);
704

705
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_COMPLETE);
706

707 708 709
    /* 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
710 711
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");

712 713 714 715 716
    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);
717 718 719 720

            /* Mark that we triggered the progress engine
               in this function call. */
            progress_engine_triggered = 1;
721
        }
722 723
    }

724 725 726 727 728 729
    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;
730
        }
731

732 733 734 735 736 737 738
        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;
        }
739

740 741 742 743 744 745 746 747
        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;
748 749
        }
        else {
750 751 752
            /* 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);
753
        }
754 755
    }

756 757 758 759 760 761 762 763 764 765 766 767 768
    /* 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);
769 770 771 772

            /* Mark that we triggered the progress engine
               in this function call. */
            progress_engine_triggered = 1;
773
        }
774
    } while (!remote_completed);
775

776 777 778
    /* 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);
779

780
    MPIU_Assert(win_ptr->non_empty_slots == 0);
781

782 783 784
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
785
    }
786

787 788 789 790 791
    /* 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);
792

793
    win_ptr->states.access_state = MPIDI_RMA_NONE;
794

795 796 797 798
 finish_complete:
    /* ENDING synchronization: correctly decrement the following counter. */
    win_ptr->accumulated_ops_cnt = 0;

799
    MPIU_Assert(win_ptr->active_req_cnt == 0);
800

801 802 803 804 805 806 807 808 809 810 811 812 813
    if (!progress_engine_triggered) {
        /* In some cases (e.g. target is myself, or process on SHM),
           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.  */
        mpi_errno = poke_progress_engine();
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
    }

814 815
  fn_exit:
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_COMPLETE);
816
    return mpi_errno;
817 818 819 820
    /* --BEGIN ERROR HANDLING-- */
  fn_fail:
    goto fn_exit;
    /* --END ERROR HANDLING-- */
821
}
822

823 824


825
#undef FUNCNAME
826
#define FUNCNAME MPIDI_Win_wait
827 828
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
829
int MPIDI_Win_wait(MPID_Win * win_ptr)
830
{
831
    int progress_engine_triggered = 0;
832 833 834 835
    int mpi_errno = MPI_SUCCESS;
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_WAIT);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_WAIT);
836

837
    MPIU_ERR_CHKANDJUMP(win_ptr->states.exposure_state != MPIDI_RMA_PSCW_EXPO,
838
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
839

840
    /* wait for all operations from other processes to finish */
841 842 843 844
    while (win_ptr->at_completion_counter) {
        mpi_errno = wait_progress_engine();
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
845 846 847 848

        /* Mark that we triggered the progress engine
           in this function call. */
        progress_engine_triggered = 1;
849 850
    }

851 852 853
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
854 855
    }

856 857
    win_ptr->states.exposure_state = MPIDI_RMA_NONE;

858 859 860 861 862 863 864 865 866 867 868 869 870 871
 finish_wait:
    if (!progress_engine_triggered) {
        /* In some cases (e.g. target is myself, or process on SHM),
           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.  */
        mpi_errno = poke_progress_engine();
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
    }

872 873
  fn_exit:
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_WAIT);
874
    return mpi_errno;
875 876 877 878
    /* --BEGIN ERROR HANDLING-- */
  fn_fail:
    goto fn_exit;
    /* --END ERROR HANDLING-- */
879 880
}

881

882
#undef FUNCNAME
883
#define FUNCNAME MPIDI_Win_test
884 885
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
886
int MPIDI_Win_test(MPID_Win * win_ptr, int *flag)
887 888
{
    int mpi_errno = MPI_SUCCESS;
889
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_TEST);
890

891
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_TEST);
892

893
    MPIU_ERR_CHKANDJUMP(win_ptr->states.exposure_state != MPIDI_RMA_PSCW_EXPO,
894
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
895

896 897
    mpi_errno = MPID_Progress_test();
    if (mpi_errno != MPI_SUCCESS) {
898
	MPIU_ERR_POP(mpi_errno);
899 900
    }

901 902 903 904 905
    *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();
906
        }
907 908

        win_ptr->states.exposure_state = MPIDI_RMA_NONE;
909 910
    }

911
  fn_exit:
912
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_TEST);
913
    return mpi_errno;
914
    /* --BEGIN ERROR HANDLING-- */
915
  fn_fail:
916
    goto fn_exit;
917
    /* --END ERROR HANDLING-- */
918 919
}

920

921 922 923 924 925 926
/********************************************************************************/
/* 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)                                               */
/********************************************************************************/

927
#undef FUNCNAME
928
#define FUNCNAME MPIDI_Win_lock
929 930
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
931
int MPIDI_Win_lock(int lock_type, int dest, int assert, MPID_Win * win_ptr)
932
{
933 934 935 936 937
    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;
938
    int mpi_errno = MPI_SUCCESS;
939
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_LOCK);
940

941
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_LOCK);
942

943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960
    /* 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");
    }
961

962 963 964 965 966 967
    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");
    }
968

969
    /* Error handling is finished. */
970

971 972 973 974 975
    if (win_ptr->lock_epoch_count == 0) {
        win_ptr->states.access_state = MPIDI_RMA_PER_TARGET;
        num_passive_win++;
    }
    win_ptr->lock_epoch_count++;
976

977
    if (dest == MPI_PROC_NULL)
978
        goto finish_lock;
979

980 981 982 983 984
    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;
985
    }
986

987 988 989
    /* 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);
990

991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004
    /* 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)
1005 1006
            MPIU_ERR_POP(mpi_errno);

1007 1008 1009
        while (target->access_state != MPIDI_RMA_LOCK_GRANTED) {
            mpi_errno = wait_progress_engine();
            if (mpi_errno != MPI_SUCCESS)
1010
                MPIU_ERR_POP(mpi_errno);
1011
        }
1012
    }
1013

1014 1015 1016 1017 1018 1019
 finish_lock:
    if (win_ptr->lock_epoch_count == 1) {
        /* BEGINNING synchronization: the following counter should be zero. */
        MPIU_Assert(win_ptr->accumulated_ops_cnt == 0);
    }

1020 1021 1022
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
1023 1024
    }

1025 1026 1027 1028 1029 1030 1031
  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-- */
1032 1033 1034
}

#undef FUNCNAME
1035
#define FUNCNAME MPIDI_Win_unlock
1036 1037
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
1038
int MPIDI_Win_unlock(int dest, MPID_Win *win_ptr)
1039
{
1040 1041 1042 1043
    int made_progress = 0;
    int local_completed = 0, remote_completed = 0;
    MPIDI_RMA_Target_t *target = NULL;
    enum MPIDI_RMA_sync_types sync_flag;
1044
    int progress_engine_triggered = 0;
1045
    int mpi_errno = MPI_SUCCESS;
1046
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_UNLOCK);
1047

1048
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_UNLOCK);
1049

1050
    MPIU_ERR_CHKANDJUMP(win_ptr->states.access_state != MPIDI_RMA_PER_TARGET,
1051
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
1052

1053
    /* Ensure ordering of load/store operations. */
1054
    if (win_ptr->shm_allocated) {
1055 1056
        OPA_read_write_barrier();
    }
1057

1058 1059 1060 1061 1062 1063 1064