ch3u_rma_sync.c 65.1 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.
*/

212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235
/*
=== 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.

=== END_MPI_T_CVAR_INFO_BLOCK ===
*/

Xin Zhao's avatar
Xin Zhao committed
236 237 238 239 240 241 242 243
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)
244
{
Xin Zhao's avatar
Xin Zhao committed
245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288
    /* 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)");
289
}
290

291 292
/* These are used to use a common routine to complete lists of RMA
   operations with a single routine, while collecting data that
293 294 295 296
   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.
*/
297

298 299 300
/*
 * These routines provide a default implementation of the MPI RMA operations
 * in terms of the low-level, two-sided channel operations.  A channel
301 302
 * may override these functions, on a per-window basis, by overriding
 * the MPID functions in the RMAFns section of MPID_Win object.
303 304
 */

305 306
#define SYNC_POST_TAG 100

307

308 309 310 311 312
/********************************************************************************/
/* Active Target synchronization (including WIN_FENCE, WIN_POST, WIN_START,     */
/* WIN_COMPLETE, WIN_WAIT, WIN_TEST)                                            */
/********************************************************************************/

313 314 315 316
#undef FUNCNAME
#define FUNCNAME MPIDI_Win_fence
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
317
int MPIDI_Win_fence(int assert, MPID_Win * win_ptr)
318
{
319 320 321
    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
322
    mpir_errflag_t errflag = MPIR_ERR_NONE;
323 324 325
    int comm_size = win_ptr->comm_ptr->local_size;
    int scalable_fence_enabled = 0;
    int *rma_target_marks = NULL;
326
    int mpi_errno = MPI_SUCCESS;
327
    MPIU_CHKLMEM_DECL(1);
328 329 330 331
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_FENCE);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_FENCE);

332 333 334 335
    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
336 337
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");

338 339 340 341 342
    /* Judge if we should switch to scalable FENCE algorithm */
    if (comm_size >= MPIR_CVAR_CH3_RMA_SCALABLE_FENCE_PROCESS_NUM) {
        scalable_fence_enabled = 1;
    }

343 344 345 346 347
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }

348 349
    if (assert & MPI_MODE_NOPRECEDE) {
        if (assert & MPI_MODE_NOSUCCEED) {
350
            goto finish_fence;
351
        }
352 353
        else {
            /* It is possible that there is a IBARRIER in MPI_WIN_FENCE with
354 355
             * MODE_NOPRECEDE not being completed, we let the progress engine
             * to delete its request when it is completed. */
356 357 358 359 360 361
            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;
362 363
                MPIDI_CH3I_num_active_issued_win--;
                MPIU_Assert(MPIDI_CH3I_num_active_issued_win >= 0);
364
            }
365

366 367
            if (win_ptr->shm_allocated == TRUE) {
                MPID_Comm *node_comm_ptr = win_ptr->comm_ptr->node_comm;
368

369
                mpi_errno = MPIR_Barrier_impl(node_comm_ptr, &errflag);
370 371
                if (mpi_errno != MPI_SUCCESS)
                    MPIU_ERR_POP(mpi_errno);
372
                MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
373 374
            }

375
            mpi_errno = MPIR_Ibarrier_impl(win_ptr->comm_ptr, &(win_ptr->fence_sync_req));
376 377
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
378

379
            /* Set window access state properly. */
380
            win_ptr->states.access_state = MPIDI_RMA_FENCE_ISSUED;
381
            MPIDI_CH3I_num_active_issued_win++;
382

383
            goto finish_fence;
384
        }
385
    }
386

387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421
    /* 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");
422

423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459
        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) {
                if (curr_target->pending_op_list_head != NULL) {
                    if (curr_target->sync.sync_flag < MPIDI_RMA_SYNC_FLUSH_LOCAL) {
                        curr_target->sync.sync_flag = MPIDI_RMA_SYNC_FLUSH_LOCAL;
                    }
                    /* flag is set in order to decrement complete counter on target */
                    curr_target->win_complete_flag = 1;
                }
                else {
                    mpi_errno = send_decr_at_cnt_msg(curr_target->target_rank, win_ptr);
                    if (mpi_errno != MPI_SUCCESS)
                        MPIU_ERR_POP(mpi_errno);
                }
                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;
460 461
            }
        }
462
    }
463

464 465
    /* Issue out all operations. */
    mpi_errno = MPIDI_CH3I_RMA_Make_progress_win(win_ptr, &made_progress);
466 467
    if (mpi_errno != MPI_SUCCESS)
        MPIU_ERR_POP(mpi_errno);
468

469
    /* Wait for local/remote completion. */
470
    do {
471 472 473
        mpi_errno = MPIDI_CH3I_RMA_Cleanup_ops_win(win_ptr, &local_completed, &remote_completed);
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
474 475
        if ((scalable_fence_enabled && !remote_completed) ||
            (!scalable_fence_enabled && !local_completed)) {
476 477 478
            mpi_errno = wait_progress_engine();
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
479
        }
480 481
    } while ((scalable_fence_enabled && !remote_completed) ||
             (!scalable_fence_enabled && !local_completed));
482

483 484
    /* Cleanup all targets on window. */
    mpi_errno = MPIDI_CH3I_RMA_Cleanup_targets_win(win_ptr);
485 486
    if (mpi_errno != MPI_SUCCESS)
        MPIU_ERR_POP(mpi_errno);
James Dinan's avatar
James Dinan committed
487

488 489 490 491 492
    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");
493

494 495 496 497 498 499 500
        /* 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;
        }
501 502
    }
    else {
503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528
        /* 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");
            }
        }
529 530
    }

531
  finish_fence:
532 533 534
    /* Make sure that all targets are freed. */
    MPIU_Assert(win_ptr->non_empty_slots == 0);

535 536
    MPIU_Assert(win_ptr->active_req_cnt == 0);

537 538 539 540 541
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }

542
  fn_exit:
543
    MPIU_CHKLMEM_FREEALL();
544 545 546
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_FENCE);
    return mpi_errno;
    /* --BEGIN ERROR HANDLING-- */
547
  fn_fail:
548 549 550 551 552 553
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}


#undef FUNCNAME
554
#define FUNCNAME MPIDI_Win_post
555 556
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
557
int MPIDI_Win_post(MPID_Group * post_grp_ptr, int assert, MPID_Win * win_ptr)
558
{
559
    int *post_ranks_in_win_grp;
560
    int mpi_errno = MPI_SUCCESS;
561
    MPIU_CHKLMEM_DECL(3);
562
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_POST);
563

564
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_POST);
565

566
    /* Note that here we cannot distinguish if this exposure epoch is overlapped
567 568 569
     * 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. */
570
    MPIU_ERR_CHKANDJUMP(win_ptr->states.exposure_state != MPIDI_RMA_NONE,
571
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
572

573 574 575
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
576 577
    }

578
    /* Set window exposure state properly. */
579 580 581 582
    win_ptr->states.exposure_state = MPIDI_RMA_PSCW_EXPO;

    win_ptr->at_completion_counter += post_grp_ptr->size;

583 584 585
    if ((assert & MPI_MODE_NOCHECK) == 0) {
        MPI_Request *req;
        MPI_Status *status;
586 587
        int i, post_grp_size, dst, rank;
        MPID_Comm *win_comm_ptr;
588

589 590
        /* NOCHECK not specified. We need to notify the source
         * processes that Post has been called. */
591

592
        post_grp_size = post_grp_ptr->size;
593 594
        win_comm_ptr = win_ptr->comm_ptr;
        rank = win_ptr->comm_ptr->rank;
595

596 597 598
        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);
599 600
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
601 602 603 604 605

        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");
606

607 608
        /* Send a 0-byte message to the source processes */
        for (i = 0; i < post_grp_size; i++) {
609
            dst = post_ranks_in_win_grp[i];
610

611 612 613 614
            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);
615 616
                if (mpi_errno != MPI_SUCCESS)
                    MPIU_ERR_POP(mpi_errno);
617 618 619 620 621
                req[i] = req_ptr->handle;
            }
            else {
                req[i] = MPI_REQUEST_NULL;
            }
622
        }
623

624 625 626 627 628 629 630 631 632 633 634 635
        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);
                }
            }
636
        }
637
        /* --END ERROR HANDLING-- */
638
    }
639

640
  fn_exit:
641 642
    MPIU_CHKLMEM_FREEALL();
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_POST);
643 644
    return mpi_errno;
    /* --BEGIN ERROR HANDLING-- */
645
  fn_fail:
646 647 648 649
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}

650

651 652 653 654 655
#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)
656
{
657
    int mpi_errno = MPI_SUCCESS;
658
    MPIU_CHKLMEM_DECL(2);
659 660
    MPIU_CHKPMEM_DECL(2);
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_START);
661

662
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_START);
663

664
    /* Note that here we cannot distinguish if this access epoch is overlapped
665 666 667 668
     * 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. */
669 670 671 672
    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");
673

674
    win_ptr->start_grp_size = group_ptr->size;
675

676 677 678 679 680
    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);
681 682
    if (mpi_errno)
        MPIU_ERR_POP(mpi_errno);
683

684
    if ((assert & MPI_MODE_NOCHECK) == 0) {
Xin Zhao's avatar
Xin Zhao committed
685
        int i, intra_cnt;
686 687 688 689 690 691
        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 */
692

693 694 695 696
        /* 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");
697

698 699 700
        if (win_ptr->shm_allocated == TRUE) {
            int node_comm_size = comm_ptr->node_comm->local_size;
            MPIU_CHKLMEM_MALLOC(intra_start_req, MPI_Request *,
701
                                node_comm_size * sizeof(MPI_Request), mpi_errno, "intra_start_req");
702 703 704 705
            MPIU_CHKLMEM_MALLOC(intra_start_status, MPI_Status *,
                                node_comm_size * sizeof(MPI_Status),
                                mpi_errno, "intra_start_status");
        }
706

707 708
        intra_cnt = 0;
        for (i = 0; i < win_ptr->start_grp_size; i++) {
709
            MPID_Request *req_ptr;
710 711
            MPIDI_VC_t *orig_vc = NULL, *target_vc = NULL;
            int src = win_ptr->start_ranks_in_win_grp[i];
712

713 714 715
            if (src != rank) {
                MPIDI_Comm_get_vc(comm_ptr, rank, &orig_vc);
                MPIDI_Comm_get_vc(comm_ptr, src, &target_vc);
716

717 718
                mpi_errno = MPID_Irecv(NULL, 0, MPI_INT, src, SYNC_POST_TAG,
                                       comm_ptr, MPID_CONTEXT_INTRA_PT2PT, &req_ptr);
719 720
                if (mpi_errno != MPI_SUCCESS)
                    MPIU_ERR_POP(mpi_errno);
721

722
                if (win_ptr->shm_allocated == TRUE && orig_vc->node_id == target_vc->node_id) {
723 724 725 726 727 728 729 730 731
                    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;
732
            }
733
        }
734

735 736 737 738 739 740 741 742 743 744 745 746
        /* 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);
                    }
747
                }
748
            }
749
            /* --END ERROR HANDLING-- */
750
        }
751 752
    }

753
  finish_start:
754
    /* Set window access state properly. */
755
    win_ptr->states.access_state = MPIDI_RMA_PSCW_ISSUED;
756
    MPIDI_CH3I_num_active_issued_win++;
757

758
    MPIU_Assert(win_ptr->active_req_cnt == 0);
759

760 761 762 763 764
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }

765
  fn_exit:
766 767 768
    MPIU_CHKLMEM_FREEALL();
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_START);
    return mpi_errno;
769
  fn_fail:
770 771
    MPIU_CHKPMEM_REAP();
    goto fn_exit;
772 773 774
}


775

776
#undef FUNCNAME
777
#define FUNCNAME MPIDI_Win_complete
778 779
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
780
int MPIDI_Win_complete(MPID_Win * win_ptr)
781
{
782
    int mpi_errno = MPI_SUCCESS;
783 784 785 786 787
    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;
788
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_COMPLETE);
789

790
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_COMPLETE);
791

792 793 794
    /* 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
795 796
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");

797 798 799 800 801
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }

802 803 804 805 806 807
    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);
        }
808 809
    }

810 811 812 813 814 815
    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;
816
        }
817

818
        if (win_comm_ptr->local_size <= win_ptr->num_slots)
819
            curr_target = win_ptr->slots[dst].target_list_head;
820
        else {
821
            curr_target = win_ptr->slots[dst % win_ptr->num_slots].target_list_head;
822 823 824
            while (curr_target != NULL && curr_target->target_rank != dst)
                curr_target = curr_target->next;
        }
825

826 827
        if (curr_target != NULL) {
            /* set sync_flag in sync struct */
828 829
            if (curr_target->sync.sync_flag < MPIDI_RMA_SYNC_FLUSH_LOCAL) {
                curr_target->sync.sync_flag = MPIDI_RMA_SYNC_FLUSH_LOCAL;
830 831
            }
            curr_target->win_complete_flag = 1;
832 833
        }
        else {
834 835
            /* FIXME: do we need to wait for remote completion? */
            mpi_errno = send_decr_at_cnt_msg(dst, win_ptr);
836 837
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
838
        }
839 840
    }

841 842
    /* issue out all operations */
    mpi_errno = MPIDI_CH3I_RMA_Make_progress_win(win_ptr, &made_progress);
843 844
    if (mpi_errno != MPI_SUCCESS)
        MPIU_ERR_POP(mpi_errno);
845 846 847

    /* wait until all slots are empty */
    do {
848 849 850
        mpi_errno = MPIDI_CH3I_RMA_Cleanup_ops_win(win_ptr, &local_completed, &remote_completed);
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
851
        if (!local_completed) {
852 853 854
            mpi_errno = wait_progress_engine();
            if (mpi_errno != MPI_SUCCESS)
                MPIU_ERR_POP(mpi_errno);
855
        }
856
    } while (!local_completed);
857

858 859
    /* Cleanup all targets on this window. */
    mpi_errno = MPIDI_CH3I_RMA_Cleanup_targets_win(win_ptr);
860 861
    if (mpi_errno != MPI_SUCCESS)
        MPIU_ERR_POP(mpi_errno);
862

863
  finish_complete:
864
    /* Set window access state properly. */
865
    win_ptr->states.access_state = MPIDI_RMA_NONE;
866

867 868 869 870
    /* 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);
871

872 873 874
    /* Make sure that all targets are freed. */
    MPIU_Assert(win_ptr->non_empty_slots == 0);

875
    MPIU_Assert(win_ptr->active_req_cnt == 0);
876

877 878
  fn_exit:
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_COMPLETE);
879
    return mpi_errno;
880 881 882 883
    /* --BEGIN ERROR HANDLING-- */
  fn_fail:
    goto fn_exit;
    /* --END ERROR HANDLING-- */
884
}
885

886 887


888
#undef FUNCNAME
889
#define FUNCNAME MPIDI_Win_wait
890 891
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
892
int MPIDI_Win_wait(MPID_Win * win_ptr)
893
{
894 895 896 897
    int mpi_errno = MPI_SUCCESS;
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_WAIT);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_WAIT);
898

899
    MPIU_ERR_CHKANDJUMP(win_ptr->states.exposure_state != MPIDI_RMA_PSCW_EXPO,
900
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
901

902
    /* wait for all operations from other processes to finish */
903 904 905 906
    while (win_ptr->at_completion_counter) {
        mpi_errno = wait_progress_engine();
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
907 908
    }

909
  finish_wait:
910
    /* Set window exposure state properly. */
911 912
    win_ptr->states.exposure_state = MPIDI_RMA_NONE;

913 914 915 916 917
    /* Ensure ordering of load/store operations. */
    if (win_ptr->shm_allocated == TRUE) {
        OPA_read_write_barrier();
    }

918 919
  fn_exit:
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_WAIT);
920
    return mpi_errno;
921 922 923 924
    /* --BEGIN ERROR HANDLING-- */
  fn_fail:
    goto fn_exit;
    /* --END ERROR HANDLING-- */
925 926
}

927

928
#undef FUNCNAME
929
#define FUNCNAME MPIDI_Win_test
930 931
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
932
int MPIDI_Win_test(MPID_Win * win_ptr, int *flag)
933 934
{
    int mpi_errno = MPI_SUCCESS;
935
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_TEST);
936

937
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_TEST);
938

939
    MPIU_ERR_CHKANDJUMP(win_ptr->states.exposure_state != MPIDI_RMA_PSCW_EXPO,
940
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
941

942 943
    mpi_errno = MPID_Progress_test();
    if (mpi_errno != MPI_SUCCESS) {
944
        MPIU_ERR_POP(mpi_errno);
945 946
    }

947 948
    *flag = (win_ptr->at_completion_counter) ? 0 : 1;
    if (*flag) {
949
        /* Set window exposure state properly. */
950 951
        win_ptr->states.exposure_state = MPIDI_RMA_NONE;

952 953 954
        /* Ensure ordering of load/store operations. */
        if (win_ptr->shm_allocated == TRUE) {
            OPA_read_write_barrier();
955 956 957
        }
    }

958
  fn_exit:
959
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_TEST);
960
    return mpi_errno;
961
    /* --BEGIN ERROR HANDLING-- */
962
  fn_fail:
963
    goto fn_exit;
964
    /* --END ERROR HANDLING-- */
965 966
}

967

968 969 970 971 972 973
/********************************************************************************/
/* 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)                                               */
/********************************************************************************/

974
#undef FUNCNAME
975
#define FUNCNAME MPIDI_Win_lock
976 977
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
978
int MPIDI_Win_lock(int lock_type, int dest, int assert, MPID_Win * win_ptr)
979
{
980 981 982 983 984
    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;
985
    int mpi_errno = MPI_SUCCESS;
986
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_LOCK);
987

988
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_LOCK);
989

990
    /* Note that here we cannot distinguish if this access epoch is overlapped
991 992 993 994
     * 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. */
995 996 997 998 999 1000 1001
    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
1002
        MPIU_ERR_CHKANDJUMP(win_ptr->states.access_state != MPIDI_RMA_PER_TARGET,
1003 1004
                            mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
    }
1005

1006 1007 1008
    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);
1009 1010
        if (mpi_errno != MPI_SUCCESS)
            MPIU_ERR_POP(mpi_errno);
1011 1012
        MPIU_ERR_CHKANDJUMP(target != NULL, mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
    }
1013

1014
    /* Error handling is finished. */
1015

1016
    if (win_ptr->lock_epoch_count == 0) {
1017
        /* Set window access state properly. */
1018
        win_ptr->states.access_state = MPIDI_RMA_PER_TARGET;
1019
        MPIDI_CH3I_num_passive_win++;
1020 1021
    }
    win_ptr->lock_epoch_count++;
1022

1023
    if (dest == MPI_PROC_NULL)
1024
        goto finish_lock;
1025

1026 1027 1028 1029 1030
    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;
1031
    }
1032

1033 1034
    /* Create a new target. */
    mpi_errno = MPIDI_CH3I_Win_create_target(win_ptr, dest, &target);
1035 1036
    if (mpi_errno != MPI_SUCCESS)
        MPIU_ERR_POP(mpi_errno);
1037

1038
    /* Store lock_state (CALLED/ISSUED/GRANTED), lock_type (SHARED/EXCLUSIVE),