ch3u_rma_sync.c 66.9 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.
*/

212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232
/*
=== BEGIN_MPI_T_CVAR_INFO_BLOCK ===

cvars:
    - name        : MPIR_CVAR_CH3_RMA_SCALABLE_FENCE_PROCESS_NUM
      category    : CH3
      type        : int
      default     : 1024
      class       : none
      verbosity   : MPI_T_VERBOSITY_USER_BASIC
      scope       : MPI_T_SCOPE_ALL_EQ
      description : >-
          Specify the threshold of switching the algorithm used in
          FENCE from the basic algorithm to the scalable algorithm.
          The value can be nagative, zero or positive.
          When the number of processes is larger than or equal to
          this value, FENCE will use a scalable algorithm which do
          not use O(P) data structure; when the number of processes
          is smaller than the value, FENCE will use a basic but fast
          algorithm which requires an O(P) data structure.

233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255
    - name        : MPIR_CVAR_CH3_RMA_DELAY_ISSUING_FOR_PIGGYBACKING
      category    : CH3
      type        : int
      default     : 0
      class       : none
      verbosity   : MPI_T_VERBOSITY_USER_BASIC
      scope       : MPI_T_SCOPE_ALL_EQ
      description : >-
        Specify if delay issuing of RMA operations for piggybacking
        LOCK/UNLOCK/FLUSH is enabled. It can be either 0 or 1. When
        it is set to 1, the issuing of LOCK message is delayed until
        origin process see the first RMA operation and piggyback
        LOCK with that operation, and the origin process always keeps
        the current last operation until the ending synchronization
        call in order to piggyback UNLOCK/FLUSH with that operation.
        When it is set to 0, in WIN_LOCK/UNLOCK case, the LOCK message
        is sent out as early as possible, in WIN_LOCK_ALL/UNLOCK_ALL
        case, the origin process still tries to piggyback LOCK message
        with the first operation; for UNLOCK/FLUSH message, the origin
        process no longer keeps the current last operation but only
        piggyback UNLOCK/FLUSH if there is an operation avaliable in
        the ending synchronization call.

256 257 258
=== END_MPI_T_CVAR_INFO_BLOCK ===
*/

Xin Zhao's avatar
Xin Zhao committed
259 260 261 262 263 264 265 266
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)
267
{
Xin Zhao's avatar
Xin Zhao committed
268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311
    /* 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)");
312
}
313

314 315
/* These are used to use a common routine to complete lists of RMA
   operations with a single routine, while collecting data that
316 317 318 319
   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.
*/
320

321 322 323
/*
 * These routines provide a default implementation of the MPI RMA operations
 * in terms of the low-level, two-sided channel operations.  A channel
324 325
 * may override these functions, on a per-window basis, by overriding
 * the MPID functions in the RMAFns section of MPID_Win object.
326 327
 */

328 329
#define SYNC_POST_TAG 100

330

331 332 333 334 335
/********************************************************************************/
/* Active Target synchronization (including WIN_FENCE, WIN_POST, WIN_START,     */
/* WIN_COMPLETE, WIN_WAIT, WIN_TEST)                                            */
/********************************************************************************/

336 337 338 339
#undef FUNCNAME
#define FUNCNAME MPIDI_Win_fence
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
340
int MPIDI_Win_fence(int assert, MPID_Win * win_ptr)
341
{
342 343 344
    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
345
    mpir_errflag_t errflag = MPIR_ERR_NONE;
346 347 348
    int comm_size = win_ptr->comm_ptr->local_size;
    int scalable_fence_enabled = 0;
    int *rma_target_marks = NULL;
349
    int mpi_errno = MPI_SUCCESS;
350
    MPIU_CHKLMEM_DECL(1);
351 352 353 354
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_FENCE);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_FENCE);

355 356 357 358
    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
359 360
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");

361 362 363 364 365
    /* Judge if we should switch to scalable FENCE algorithm */
    if (comm_size >= MPIR_CVAR_CH3_RMA_SCALABLE_FENCE_PROCESS_NUM) {
        scalable_fence_enabled = 1;
    }

366 367 368 369 370
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }

371 372
    if (assert & MPI_MODE_NOPRECEDE) {
        if (assert & MPI_MODE_NOSUCCEED) {
373
            goto finish_fence;
374
        }
375 376
        else {
            /* It is possible that there is a IBARRIER in MPI_WIN_FENCE with
377 378
             * MODE_NOPRECEDE not being completed, we let the progress engine
             * to delete its request when it is completed. */
379 380 381 382 383 384
            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;
385 386
                MPIDI_CH3I_num_active_issued_win--;
                MPIU_Assert(MPIDI_CH3I_num_active_issued_win >= 0);
387
            }
388

389 390
            if (win_ptr->shm_allocated == TRUE) {
                MPID_Comm *node_comm_ptr = win_ptr->comm_ptr->node_comm;
391

392
                mpi_errno = MPIR_Barrier_impl(node_comm_ptr, &errflag);
393 394
                if (mpi_errno != MPI_SUCCESS)
                    MPIU_ERR_POP(mpi_errno);
395
                MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
396 397
            }

398
            mpi_errno = MPIR_Ibarrier_impl(win_ptr->comm_ptr, &(win_ptr->fence_sync_req));
399 400
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
401

402
            /* Set window access state properly. */
403
            win_ptr->states.access_state = MPIDI_RMA_FENCE_ISSUED;
404
            MPIDI_CH3I_num_active_issued_win++;
405

406
            goto finish_fence;
407
        }
408
    }
409

410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444
    /* Perform basic algorithm by calling reduce-scatter */
    if (!scalable_fence_enabled) {
        /* If the IBARRIER is not completed, do not need to wait for
         * it since we are going to call reduce-scatter */
        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;
            MPIDI_CH3I_num_active_issued_win--;
            MPIU_Assert(MPIDI_CH3I_num_active_issued_win >= 0);

            win_ptr->states.access_state = MPIDI_RMA_NONE;
        }
        MPIU_CHKLMEM_MALLOC(rma_target_marks, int *, comm_size * sizeof(int),
                            mpi_errno, "rma_target_marks");
        for (i = 0; i < comm_size; i++)
            rma_target_marks[i] = 0;

        for (i = 0; i < win_ptr->num_slots; i++) {
            curr_target = win_ptr->slots[i].target_list_head;
            while (curr_target != NULL) {
                rma_target_marks[curr_target->target_rank] = 1;
                curr_target = curr_target->next;
            }
        }

        win_ptr->at_completion_counter += comm_size;

        mpi_errno = MPIR_Reduce_scatter_block_impl(MPI_IN_PLACE, rma_target_marks, 1,
                                                   MPI_INT, MPI_SUM, 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");
445

446 447 448 449 450 451 452 453 454 455 456 457
        win_ptr->at_completion_counter -= comm_size;
        win_ptr->at_completion_counter += rma_target_marks[0];
        MPIU_Assert(win_ptr->at_completion_counter >= 0);

        win_ptr->states.access_state = MPIDI_RMA_FENCE_GRANTED;
    }

    /* Set sync_flag in target structs. */
    if (!scalable_fence_enabled) {
        for (i = 0; i < win_ptr->num_slots; i++) {
            curr_target = win_ptr->slots[i].target_list_head;
            while (curr_target != NULL) {
458 459
                if (curr_target->sync.sync_flag < MPIDI_RMA_SYNC_FLUSH_LOCAL) {
                    curr_target->sync.sync_flag = MPIDI_RMA_SYNC_FLUSH_LOCAL;
460
                }
461 462 463
                /* flag is set in order to decrement complete counter on target */
                curr_target->win_complete_flag = 1;

464 465 466 467 468 469 470 471 472 473 474 475 476
                curr_target = curr_target->next;
            }
        }
    }
    else {
        for (i = 0; i < win_ptr->num_slots; i++) {
            curr_target = win_ptr->slots[i].target_list_head;
            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 = curr_target->next;
477 478
            }
        }
479
    }
480

481 482
    /* Issue out all operations. */
    mpi_errno = MPIDI_CH3I_RMA_Make_progress_win(win_ptr, &made_progress);
483 484
    if (mpi_errno != MPI_SUCCESS)
        MPIU_ERR_POP(mpi_errno);
485

486
    /* Wait for local/remote completion. */
487
    do {
488 489 490
        mpi_errno = MPIDI_CH3I_RMA_Cleanup_ops_win(win_ptr, &local_completed, &remote_completed);
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
491 492
        if ((scalable_fence_enabled && !remote_completed) ||
            (!scalable_fence_enabled && !local_completed)) {
493 494 495
            mpi_errno = wait_progress_engine();
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
496
        }
497 498
    } while ((scalable_fence_enabled && !remote_completed) ||
             (!scalable_fence_enabled && !local_completed));
499

500 501
    /* Cleanup all targets on window. */
    mpi_errno = MPIDI_CH3I_RMA_Cleanup_targets_win(win_ptr);
502 503
    if (mpi_errno != MPI_SUCCESS)
        MPIU_ERR_POP(mpi_errno);
James Dinan's avatar
James Dinan committed
504

505 506 507 508 509
    if (scalable_fence_enabled) {
        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");
510

511 512 513 514 515 516 517
        /* Set window access state properly. */
        if (assert & MPI_MODE_NOSUCCEED) {
            win_ptr->states.access_state = MPIDI_RMA_NONE;
        }
        else {
            win_ptr->states.access_state = MPIDI_RMA_FENCE_GRANTED;
        }
518 519
    }
    else {
520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545
        /* Waiting for all operations targeting at me to be finished. */
        while (win_ptr->at_completion_counter) {
            mpi_errno = wait_progress_engine();
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
        }

        if (assert & MPI_MODE_NOSUCCEED) {
            win_ptr->states.access_state = MPIDI_RMA_NONE;
        }
        else {
            /* Prepare for the next possible epoch */
            mpi_errno = MPIR_Ibarrier_impl(win_ptr->comm_ptr, &(win_ptr->fence_sync_req));
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
            MPIDI_CH3I_num_active_issued_win++;
            win_ptr->states.access_state = MPIDI_RMA_FENCE_ISSUED;

            if (win_ptr->shm_allocated == TRUE) {
                MPID_Comm *node_comm_ptr = win_ptr->comm_ptr->node_comm;
                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");
            }
        }
546 547
    }

548
  finish_fence:
549 550 551
    /* Make sure that all targets are freed. */
    MPIU_Assert(win_ptr->non_empty_slots == 0);

552 553
    MPIU_Assert(win_ptr->active_req_cnt == 0);

554 555 556 557 558
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }

559
  fn_exit:
560
    MPIU_CHKLMEM_FREEALL();
561 562 563
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_FENCE);
    return mpi_errno;
    /* --BEGIN ERROR HANDLING-- */
564
  fn_fail:
565 566 567 568 569 570
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}


#undef FUNCNAME
571
#define FUNCNAME MPIDI_Win_post
572 573
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
574
int MPIDI_Win_post(MPID_Group * post_grp_ptr, int assert, MPID_Win * win_ptr)
575
{
576
    int *post_ranks_in_win_grp;
577
    int mpi_errno = MPI_SUCCESS;
578
    MPIU_CHKLMEM_DECL(3);
579
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_POST);
580

581
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_POST);
582

583
    /* Note that here we cannot distinguish if this exposure epoch is overlapped
584 585 586
     * 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. */
587
    MPIU_ERR_CHKANDJUMP(win_ptr->states.exposure_state != MPIDI_RMA_NONE,
588
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
589

590 591 592
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
593 594
    }

595
    /* Set window exposure state properly. */
596 597 598 599
    win_ptr->states.exposure_state = MPIDI_RMA_PSCW_EXPO;

    win_ptr->at_completion_counter += post_grp_ptr->size;

600 601 602
    if ((assert & MPI_MODE_NOCHECK) == 0) {
        MPI_Request *req;
        MPI_Status *status;
603 604
        int i, post_grp_size, dst, rank;
        MPID_Comm *win_comm_ptr;
605

606 607
        /* NOCHECK not specified. We need to notify the source
         * processes that Post has been called. */
608

609
        post_grp_size = post_grp_ptr->size;
610 611
        win_comm_ptr = win_ptr->comm_ptr;
        rank = win_ptr->comm_ptr->rank;
612

613 614 615
        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);
616 617
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
618 619 620 621 622

        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");
623

624 625
        /* Send a 0-byte message to the source processes */
        for (i = 0; i < post_grp_size; i++) {
626
            dst = post_ranks_in_win_grp[i];
627

628 629 630 631
            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);
632 633
                if (mpi_errno != MPI_SUCCESS)
                    MPIU_ERR_POP(mpi_errno);
634 635 636 637 638
                req[i] = req_ptr->handle;
            }
            else {
                req[i] = MPI_REQUEST_NULL;
            }
639
        }
640

641 642 643 644 645 646 647 648 649 650 651 652
        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);
                }
            }
653
        }
654
        /* --END ERROR HANDLING-- */
655
    }
656

657
  fn_exit:
658 659
    MPIU_CHKLMEM_FREEALL();
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_POST);
660 661
    return mpi_errno;
    /* --BEGIN ERROR HANDLING-- */
662
  fn_fail:
663 664 665 666
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}

667

668 669 670 671 672
#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)
673
{
674
    int mpi_errno = MPI_SUCCESS;
675
    MPIU_CHKLMEM_DECL(2);
676 677
    MPIU_CHKPMEM_DECL(2);
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_START);
678

679
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_START);
680

681
    /* Note that here we cannot distinguish if this access epoch is overlapped
682 683 684 685
     * 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. */
686 687 688 689
    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");
690

691
    win_ptr->start_grp_size = group_ptr->size;
692

693 694 695 696 697
    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);
698 699
    if (mpi_errno)
        MPIU_ERR_POP(mpi_errno);
700

701
    if ((assert & MPI_MODE_NOCHECK) == 0) {
Xin Zhao's avatar
Xin Zhao committed
702
        int i, intra_cnt;
703 704 705 706 707 708
        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 */
709

710 711 712 713
        /* 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");
714

715 716 717
        if (win_ptr->shm_allocated == TRUE) {
            int node_comm_size = comm_ptr->node_comm->local_size;
            MPIU_CHKLMEM_MALLOC(intra_start_req, MPI_Request *,
718
                                node_comm_size * sizeof(MPI_Request), mpi_errno, "intra_start_req");
719 720 721 722
            MPIU_CHKLMEM_MALLOC(intra_start_status, MPI_Status *,
                                node_comm_size * sizeof(MPI_Status),
                                mpi_errno, "intra_start_status");
        }
723

724 725
        intra_cnt = 0;
        for (i = 0; i < win_ptr->start_grp_size; i++) {
726
            MPID_Request *req_ptr;
727 728
            MPIDI_VC_t *orig_vc = NULL, *target_vc = NULL;
            int src = win_ptr->start_ranks_in_win_grp[i];
729

730 731 732
            if (src != rank) {
                MPIDI_Comm_get_vc(comm_ptr, rank, &orig_vc);
                MPIDI_Comm_get_vc(comm_ptr, src, &target_vc);
733

734 735
                mpi_errno = MPID_Irecv(NULL, 0, MPI_INT, src, SYNC_POST_TAG,
                                       comm_ptr, MPID_CONTEXT_INTRA_PT2PT, &req_ptr);
736 737
                if (mpi_errno != MPI_SUCCESS)
                    MPIU_ERR_POP(mpi_errno);
738

739
                if (win_ptr->shm_allocated == TRUE && orig_vc->node_id == target_vc->node_id) {
740 741 742 743 744 745 746 747 748
                    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;
749
            }
750
        }
751

752 753 754 755 756 757 758 759 760 761 762 763
        /* 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);
                    }
764
                }
765
            }
766
            /* --END ERROR HANDLING-- */
767
        }
768 769
    }

770
  finish_start:
771
    /* Set window access state properly. */
772
    win_ptr->states.access_state = MPIDI_RMA_PSCW_ISSUED;
773
    MPIDI_CH3I_num_active_issued_win++;
774

775
    MPIU_Assert(win_ptr->active_req_cnt == 0);
776

777 778 779 780 781
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }

782
  fn_exit:
783 784 785
    MPIU_CHKLMEM_FREEALL();
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_START);
    return mpi_errno;
786
  fn_fail:
787 788
    MPIU_CHKPMEM_REAP();
    goto fn_exit;
789 790 791
}


792

793
#undef FUNCNAME
794
#define FUNCNAME MPIDI_Win_complete
795 796
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
797
int MPIDI_Win_complete(MPID_Win * win_ptr)
798
{
799
    int mpi_errno = MPI_SUCCESS;
800 801 802 803 804
    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;
805
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_COMPLETE);
806

807
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_COMPLETE);
808

809 810 811
    /* 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
812 813
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");

814 815 816 817 818
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }

819 820 821 822 823 824
    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);
        }
825 826
    }

827 828 829 830 831 832
    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;
833
        }
834

835
        if (win_comm_ptr->local_size <= win_ptr->num_slots)
836
            curr_target = win_ptr->slots[dst].target_list_head;
837
        else {
838
            curr_target = win_ptr->slots[dst % win_ptr->num_slots].target_list_head;
839 840 841
            while (curr_target != NULL && curr_target->target_rank != dst)
                curr_target = curr_target->next;
        }
842

843 844
        if (curr_target != NULL) {
            /* set sync_flag in sync struct */
845 846
            if (curr_target->sync.sync_flag < MPIDI_RMA_SYNC_FLUSH_LOCAL) {
                curr_target->sync.sync_flag = MPIDI_RMA_SYNC_FLUSH_LOCAL;
847 848
            }
            curr_target->win_complete_flag = 1;
849 850
        }
        else {
851
            /* FIXME: do we need to wait for remote completion? */
852
            mpi_errno = send_decr_at_cnt_msg(dst, win_ptr, MPIDI_CH3_PKT_FLAG_NONE);
853 854
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
855
        }
856 857
    }

858 859
    /* issue out all operations */
    mpi_errno = MPIDI_CH3I_RMA_Make_progress_win(win_ptr, &made_progress);
860 861
    if (mpi_errno != MPI_SUCCESS)
        MPIU_ERR_POP(mpi_errno);
862 863 864

    /* wait until all slots are empty */
    do {
865 866 867
        mpi_errno = MPIDI_CH3I_RMA_Cleanup_ops_win(win_ptr, &local_completed, &remote_completed);
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
868
        if (!local_completed) {
869 870 871
            mpi_errno = wait_progress_engine();
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
872
        }
873
    } while (!local_completed);
874

875 876
    /* Cleanup all targets on this window. */
    mpi_errno = MPIDI_CH3I_RMA_Cleanup_targets_win(win_ptr);
877 878
    if (mpi_errno != MPI_SUCCESS)
        MPIU_ERR_POP(mpi_errno);
879

880
  finish_complete:
881
    /* Set window access state properly. */
882
    win_ptr->states.access_state = MPIDI_RMA_NONE;
883

884 885 886 887
    /* 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);
888

889 890 891
    /* Make sure that all targets are freed. */
    MPIU_Assert(win_ptr->non_empty_slots == 0);

892
    MPIU_Assert(win_ptr->active_req_cnt == 0);
893

894 895
  fn_exit:
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_COMPLETE);
896
    return mpi_errno;
897 898 899 900
    /* --BEGIN ERROR HANDLING-- */
  fn_fail:
    goto fn_exit;
    /* --END ERROR HANDLING-- */
901
}
902

903 904


905
#undef FUNCNAME
906
#define FUNCNAME MPIDI_Win_wait
907 908
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
909
int MPIDI_Win_wait(MPID_Win * win_ptr)
910
{
911 912 913 914
    int mpi_errno = MPI_SUCCESS;
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_WAIT);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_WAIT);
915

916
    MPIU_ERR_CHKANDJUMP(win_ptr->states.exposure_state != MPIDI_RMA_PSCW_EXPO,
917
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
918

919
    /* wait for all operations from other processes to finish */
920 921 922 923
    while (win_ptr->at_completion_counter) {
        mpi_errno = wait_progress_engine();
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
924 925
    }

926
  finish_wait:
927
    /* Set window exposure state properly. */
928 929
    win_ptr->states.exposure_state = MPIDI_RMA_NONE;

930 931 932 933 934
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }

935 936
  fn_exit:
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_WAIT);
937
    return mpi_errno;
938 939 940 941
    /* --BEGIN ERROR HANDLING-- */
  fn_fail:
    goto fn_exit;
    /* --END ERROR HANDLING-- */
942 943
}

944

945
#undef FUNCNAME
946
#define FUNCNAME MPIDI_Win_test
947 948
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
949
int MPIDI_Win_test(MPID_Win * win_ptr, int *flag)
950 951
{
    int mpi_errno = MPI_SUCCESS;
952
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_TEST);
953

954
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_TEST);
955

956
    MPIU_ERR_CHKANDJUMP(win_ptr->states.exposure_state != MPIDI_RMA_PSCW_EXPO,
957
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
958

959 960
    mpi_errno = MPID_Progress_test();
    if (mpi_errno != MPI_SUCCESS) {
961
        MPIU_ERR_POP(mpi_errno);
962 963
    }

964 965
    *flag = (win_ptr->at_completion_counter) ? 0 : 1;
    if (*flag) {
966
        /* Set window exposure state properly. */
967 968
        win_ptr->states.exposure_state = MPIDI_RMA_NONE;

969 970 971
        /* Ensure ordering of load/store operations. */
        if (win_ptr->shm_allocated == TRUE) {
            OPA_read_write_barrier();
972 973 974
        }
    }

975
  fn_exit:
976
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_TEST);
977
    return mpi_errno;
978
    /* --BEGIN ERROR HANDLING-- */
979
  fn_fail:
980
    goto fn_exit;
981
    /* --END ERROR HANDLING-- */
982 983
}

984

985 986 987 988 989 990
/********************************************************************************/
/* 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)                                               */
/********************************************************************************/

991
#undef FUNCNAME
992
#define FUNCNAME MPIDI_Win_lock
993 994
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
995
int MPIDI_Win_lock(int lock_type, int dest, int assert, MPID_Win * win_ptr)
996
{
997 998 999 1000 1001
    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;
1002
    int mpi_errno = MPI_SUCCESS;
1003
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_LOCK);
1004

1005
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_LOCK);
1006

1007
    /* Note that here we cannot distinguish if this access epoch is overlapped
1008 1009 1010 1011
     * 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. */
1012 1013 1014 1015 1016 1017 1018
    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
1019
        MPIU_ERR_CHKANDJUMP(win_ptr->states.access_state != MPIDI_RMA_PER_TARGET,
1020 1021
                            mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
    }
1022

1023 1024 1025
    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);
1026 1027
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
1028 1029
        MPIU_ERR_CHKANDJUMP(target != NULL, mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
    }
1030

1031
    /* Error handling is finished. */
1032

1033
    if (win_ptr->lock_epoch_count == 0) {
1034
        /* Set window access state properly. */
1035
        win_ptr->states.access_state = MPIDI_RMA_PER_TARGET;
1036
        MPIDI_CH3I_num_passive_win++;
1037 1038
    }
    win_ptr->lock_epoch_count++;
1039

1040
    if (dest == MPI_PROC_NULL)
1041
        goto finish_lock;
1042

Xin Zhao's avatar