ch3u_rma_sync.c 229 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
/*
=== BEGIN_MPI_T_CVAR_INFO_BLOCK ===

categories:
    - name        : CH3
      description : cvars that control behavior of ch3

cvars:
    - name        : MPIR_CVAR_CH3_RMA_ACC_IMMED
      category    : CH3
      type        : boolean
      default     : true
      class       : none
      verbosity   : MPI_T_VERBOSITY_USER_BASIC
      scope       : MPI_T_SCOPE_ALL_EQ
      description : >-
        Use the immediate accumulate optimization

    - name        : MPIR_CVAR_CH3_RMA_NREQUEST_THRESHOLD
      category    : CH3
      type        : int
      default     : 4000
      class       : none
      verbosity   : MPI_T_VERBOSITY_USER_BASIC
      scope       : MPI_T_SCOPE_ALL_EQ
      description : >-
        Threshold at which the RMA implementation attempts to complete requests
        while completing RMA operations and while using the lazy synchonization
        approach.  Change this value if programs fail because they run out of
        requests or other internal resources

    - name        : MPIR_CVAR_CH3_RMA_NREQUEST_NEW_THRESHOLD
      category    : CH3
      type        : int
44
      default     : 0
45
46
47
48
49
50
51
52
53
      class       : none
      verbosity   : MPI_T_VERBOSITY_USER_BASIC
      scope       : MPI_T_SCOPE_ALL_EQ
      description : >-
        Threshold for the number of new requests since the last attempt to
        complete pending requests.  Higher values can increase performance,
        but may run the risk of exceeding the available number of requests
        or other internal resources.

54
55
56
    - name        : MPIR_CVAR_CH3_RMA_GC_NUM_COMPLETED
      category    : CH3
      type        : int
57
      default     : (-1)
58
59
60
61
62
63
64
      class       : none
      verbosity   : MPI_T_VERBOSITY_USER_BASIC
      scope       : MPI_T_SCOPE_ALL_EQ
      description : >-
        Threshold for the number of completed requests the runtime finds
        before it stops trying to find more completed requests in garbage
        collection function.
65
66
67
68
        Note that it works with MPIR_CVAR_CH3_RMA_GC_NUM_TESTED as an OR
        relation, which means runtime will stop checking when either one
        of its following conditions is satisfied or one of conditions of
        MPIR_CVAR_CH3_RMA_GC_NUM_TESTED is satisfied.
69
70
71
72
73
74
75
76
77
78
79
        When it is set to negative value, it means runtime will not stop
        checking the operation list until it reaches the end of the list.
        When it is set to positive value, it means runtime will not stop
        checking the operation list until it finds certain number of
        completed requests. When it is set to zero value, the outcome is
        undefined.
        Note that in garbage collection function, if runtime finds a chain
        of completed RMA requests, it will temporarily ignore this CVAR
        and try to find continuous completed requests as many as possible,
        until it meets an incomplete request.

80
81
82
    - name        : MPIR_CVAR_CH3_RMA_GC_NUM_TESTED
      category    : CH3
      type        : int
83
      default     : 100
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
      class       : none
      verbosity   : MPI_T_VERBOSITY_USER_BASIC
      scope       : MPI_T_SCOPE_ALL_EQ
      description : >-
        Threshold for the number of RMA requests the runtime tests before
        it stops trying to check more requests in garbage collection
        routine.
        Note that it works with MPIR_CVAR_CH3_RMA_GC_NUM_COMPLETED as an
        OR relation, which means runtime will stop checking when either
        one of its following conditions is satisfied or one of conditions
        of MPIR_CVAR_CH3_RMA_GC_NUM_COMPLETED is satisfied.
        When it is set to negative value, runtime will not stop checking
        operation list until runtime reaches the end of the list. It has
        the risk of O(N) traversing overhead if there is no completed
        request in the list. When it is set to positive value, it means
        runtime will not stop checking the operation list until it visits
        such number of requests. Higher values may make more completed
        requests to be found, but it has the risk of visiting too many
        requests, leading to significant performance overhead. When it is
        set to zero value, runtime will stop checking the operation list
        immediately, which may cause weird performance in practice.
        Note that in garbage collection function, if runtime finds a chain
        of completed RMA requests, it will temporarily ignore this CVAR and
        try to find continuous completed requests as many as possible, until
        it meets an incomplete request.

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
    - name        : MPIR_CVAR_CH3_RMA_LOCK_IMMED
      category    : CH3
      type        : boolean
      default     : false
      class       : none
      verbosity   : MPI_T_VERBOSITY_USER_BASIC
      scope       : MPI_T_SCOPE_ALL_EQ
      description : >-
        Issue a request for the passive target RMA lock immediately.  Default
        behavior is to defer the lock request until the call to MPI_Win_unlock.

    - name        : MPIR_CVAR_CH3_RMA_MERGE_LOCK_OP_UNLOCK
      category    : CH3
      type        : boolean
      default     : true
      class       : none
      verbosity   : MPI_T_VERBOSITY_USER_BASIC
      scope       : MPI_T_SCOPE_ALL_EQ
      description : >-
        Enable/disable an optimization that merges lock, op, and unlock
        messages, for single-operation passive target epochs.

=== END_MPI_T_CVAR_INFO_BLOCK ===
*/

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
212
213
214
215
216
217
218
219
220
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
265
266
267
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
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
/* 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.
*/

337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_lockqueue_alloc);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winfence_clearlock);
MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_winfence_clearlock_aux);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winfence_rs);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winfence_issue);
MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_winfence_issue_aux);
MPIR_T_PVAR_ULONG2_HIGHWATERMARK_DECL(RMA, rma_winfence_issue_aux);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winfence_complete);
MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_winfence_complete_aux);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winfence_wait);
MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_winfence_wait_aux);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winfence_block);
MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_winfence_block_aux);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winpost_clearlock);
MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_winpost_clearlock_aux);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winpost_sendsync);
MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_winpost_sendsync_aux);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winstart_clearlock);
MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_winstart_clearlock_aux);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_wincomplete_issue);
MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_wincomplete_issue_aux);
MPIR_T_PVAR_ULONG2_HIGHWATERMARK_DECL(RMA, rma_wincomplete_issue_aux);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_wincomplete_complete);
MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_wincomplete_complete_aux);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_wincomplete_recvsync);
MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_wincomplete_recvsync_aux);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_wincomplete_block);
MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_wincomplete_block_aux);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winwait_wait);
MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_winwait_wait_aux);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winlock_getlocallock);
MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winunlock_getlock);
MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winunlock_issue);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winunlock_complete);
MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_winunlock_complete_aux);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winunlock_block);
MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_winunlock_block_aux);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_rmapkt_acc);
MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_rmapkt_acc_predef);
MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_rmapkt_acc_immed);
MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_rmapkt_acc_immed_op);
MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_rmapkt_cas);
MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_rmapkt_fop);
MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_rmapkt_get_accum);

MPIR_T_PVAR_ULONG2_LEVEL_DECL(RMA, rma_winfence_reqs);
MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_winfence_reqs);
MPIR_T_PVAR_ULONG2_HIGHWATERMARK_DECL(RMA, rma_winfence_reqs);

MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_winunlock_reqs);
MPIR_T_PVAR_ULONG2_COUNTER_DECL(RMA, rma_wincomplete_reqs);

MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_wincreate_allgather);
MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winfree_rs);
MPIR_T_PVAR_DOUBLE_TIMER_DECL(RMA, rma_winfree_complete);
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_Pvars(void)
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
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
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
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
    /* 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_winfence_clearlock */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_winfence_clearlock,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_FENCE:Clear prior lock (in seconds)");

    MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_winfence_clearlock_aux,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_FENCE:Clear prior lock");

    /* rma_winfence_rs */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_winfence_rs,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_FENCE:ReduceScatterBlock (in seconds)");

    /* rma_winfence_issue and auxiliaries */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_winfence_issue,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_FENCE:Issue RMA ops (in seconds)");

    MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_winfence_issue_aux,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_FENCE:Issue RMA ops");

    MPIR_T_PVAR_HIGHWATERMARK_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_winfence_issue_aux,
        0, /* init value */
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_FENCE:Issue RMA ops");

    /* rma_winfence_complete */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_winfence_complete,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_FENCE:Complete RMA ops (in seconds)");

    MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_winfence_complete_aux,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_FENCE:Complete RMA ops");

    /* rma_winfence_wait */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_winfence_wait,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_FENCE:Wait for ops from other processes (in seconds)");

    MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_winfence_wait_aux,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_FENCE:Wait for ops from other processes");

    /* rma_winfence_block */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_winfence_block,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_FENCE:Wait for any progress (in seconds)");

    MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_winfence_block_aux,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_FENCE:Wait for any progress");


    /* rma_winpost_clearlock */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_winpost_clearlock,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_POST:Clear prior lock  (in seconds)");

    MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_winpost_clearlock_aux,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_POST:Clear prior lock");

    /* rma_winpost_sendsync */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_winpost_sendsync,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_POST:Senc sync messages (in seconds)");

    MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_winpost_sendsync_aux,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_POST:Senc sync messages");

    /* rma_winstart_clearlock */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_winstart_clearlock,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_START:Clear prior lock (in seconds)");

    MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_winstart_clearlock_aux,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_START:Clear prior lock");

    /* rma_wincomplete_issue and auxiliaries */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_wincomplete_issue,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_COMPLETE:Issue RMA ops (in seconds)");

    MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_wincomplete_issue_aux,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_COMPLETE:Issue RMA ops");

    MPIR_T_PVAR_HIGHWATERMARK_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_wincomplete_issue_aux,
        0, /* init value */
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_COMPLETE:Issue RMA ops");

    /* rma_wincomplete_complete */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_wincomplete_complete,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_COMPLETE:Complete RMA ops (in seconds)");

    MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_wincomplete_complete_aux,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_COMPLETE:Complete RMA ops");

    /* rma_wincomplete_recvsync */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_wincomplete_recvsync,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_COMPLETE:Recv sync messages (in seconds)");

    MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_wincomplete_recvsync_aux,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_COMPLETE:Recv sync messages");

    /* rma_wincomplete_block */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_wincomplete_block,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_COMPLETE:Wait for any progress (in seconds)");

    MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_wincomplete_block_aux,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_COMPLETE:Wait for any progress");

    /* rma_winwait_wait */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_winwait_wait,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_WAIT:Wait for ops from other processes (in seconds)");

    MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_winwait_wait_aux,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_WAIT:Wait for ops from other processes");

    /* 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_winunlock_getlock */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_winunlock_getlock,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_UNLOCK:Acquire lock (in seconds)");

    /* rma_winunlock_issue */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_winunlock_issue,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_UNLOCK:Issue RMA ops (in seconds)");

    /* rma_winunlock_complete */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_winunlock_complete,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_UNLOCK:Complete RMA ops (in seconds)");

    MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_winunlock_complete_aux,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_UNLOCK:Complete RMA ops (in seconds)");

    /* rma_winunlock_block */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_winunlock_block,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_UNLOCK:Wait for any progress (in seconds)");

    MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_winunlock_block_aux,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_UNLOCK:Wait for any progress");

    /* rma_rmapkt_acc */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_rmapkt_acc,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "RMA:PKTHANDLER for Accumulate (in seconds)");

    /* rma_rmapkt_acc_predef */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_rmapkt_acc_predef,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "RMA:PKTHANDLER for Accumulate: predef dtype (in seconds)");

    /* rma_rmapkt_acc_immed */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_rmapkt_acc_immed,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "RMA:PKTHANDLER for Accum immed (in seconds)");

    /* rma_rmapkt_acc_immed_op */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_rmapkt_acc_immed_op,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "RMA:PKTHANDLER for Accum immed operation (in seconds)");

    /* rma_rmapkt_cas */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_rmapkt_cas,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "RMA:PKTHANDLER for Compare-and-swap (in seconds)");

    /* rma_rmapkt_fop */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_rmapkt_fop,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "RMA:PKTHANDLER for Fetch-and-op (in seconds)");

    /* rma_rmapkt_get_accum */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_rmapkt_get_accum,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "RMA:PKTHANDLER for Get-Accumulate (in seconds)");

    /* Level, counter and highwatermark for rma_winfence_reqs */
    MPIR_T_PVAR_LEVEL_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_winfence_reqs,
        0, /* init value */
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY | MPIR_T_PVAR_FLAG_CONTINUOUS,
        "RMA",
        "WIN_FENCE:Pending requests");

      MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_winfence_reqs,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_FENCE:Pending requests");

      MPIR_T_PVAR_HIGHWATERMARK_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_winfence_reqs,
        0, /* init value */
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_FENCE:Pending requests");

    /* rma_winunlock_reqs */
    MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_winunlock_reqs,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_UNLOCK:Pending requests");

    /* rma_wincomplete_reqs */
    MPIR_T_PVAR_COUNTER_REGISTER_STATIC(
        RMA,
        MPI_UNSIGNED_LONG_LONG,
        rma_wincomplete_reqs,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_COMPLETE:Pending requests");

    /* 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_winfree_rs */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_winfree_rs,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_FREE:ReduceScatterBlock (in seconds)");

    /* rma_winfree_complete */
    MPIR_T_PVAR_TIMER_REGISTER_STATIC(
        RMA,
        MPI_DOUBLE,
        rma_winfree_complete,
        MPI_T_VERBOSITY_MPIDEV_DETAIL,
        MPI_T_BIND_NO_OBJECT,
        MPIR_T_PVAR_FLAG_READONLY,
        "RMA",
        "WIN_FREE:Complete (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)");
994
}
995
996
997
998
999
1000
1001

/* These are used to use a common routine to complete lists of RMA 
   operations with a single routine, while collecting data that 
   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.
*/
1002
1003
1004
static MPIR_T_pvar_timer_t *list_complete_timer;  /* outer */
static unsigned long long *list_complete_counter;
static MPIR_T_pvar_timer_t *list_block_timer;     /* Inner; while waiting */
1005

1006
1007
1008
/*
 * These routines provide a default implementation of the MPI RMA operations
 * in terms of the low-level, two-sided channel operations.  A channel
1009
1010
 * may override these functions, on a per-window basis, by overriding
 * the MPID functions in the RMAFns section of MPID_Win object.
1011
1012
 */

1013
1014
#define SYNC_POST_TAG 100

Xin Zhao's avatar
Xin Zhao committed
1015
1016
1017
1018
1019
static int send_lock_msg(int dest, int lock_type, MPID_Win *win_ptr);
static int send_unlock_msg(int dest, MPID_Win *win_ptr);
/* static int send_flush_msg(int dest, MPID_Win *win_ptr); */
static int wait_for_lock_granted(MPID_Win *win_ptr, int target_rank);
static int acquire_local_lock(MPID_Win *win_ptr, int lock_mode);
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
static int send_rma_msg(MPIDI_RMA_Op_t * rma_op, MPID_Win * win_ptr,
                                   MPIDI_CH3_Pkt_flags_t flags,
				   MPI_Win source_win_handle, 
				   MPI_Win target_win_handle, 
				   MPIDI_RMA_dtype_info * dtype_info, 
				   void ** dataloop, MPID_Request ** request);
static int recv_rma_msg(MPIDI_RMA_Op_t * rma_op, MPID_Win * win_ptr,
                                   MPIDI_CH3_Pkt_flags_t flags,
				   MPI_Win source_win_handle, 
				   MPI_Win target_win_handle, 
				   MPIDI_RMA_dtype_info * dtype_info, 
				   void ** dataloop, MPID_Request ** request); 
static int send_contig_acc_msg(MPIDI_RMA_Op_t *, MPID_Win *,
                                          MPIDI_CH3_Pkt_flags_t flags,
					  MPI_Win, MPI_Win, MPID_Request ** );
static int send_immed_rmw_msg(MPIDI_RMA_Op_t *, MPID_Win *,
                                         MPIDI_CH3_Pkt_flags_t flags,
                                         MPI_Win, MPI_Win, MPID_Request ** );
Xin Zhao's avatar
Xin Zhao committed
1038
static int do_passive_target_rma(MPID_Win *win_ptr, int target_rank,
1039
                                            int *wait_for_rma_done_pkt,
James Dinan's avatar
James Dinan committed
1040
                                            MPIDI_CH3_Pkt_flags_t sync_flags);
Xin Zhao's avatar
Xin Zhao committed
1041
1042
static int send_lock_put_or_acc(MPID_Win *, int);
static int send_lock_get(MPID_Win *, int);
1043
static inline int poke_progress_engine(void);
1044
static inline int rma_list_complete(MPID_Win *win_ptr,
1045
                                      MPIDI_RMA_Ops_list_t *ops_list);
1046
static inline int rma_list_gc(MPID_Win *win_ptr,
1047
1048
                                             MPIDI_RMA_Ops_list_t *ops_list,
                                             MPIDI_RMA_Op_t *last_elm, int *nDone);
1049

1050
1051
static int create_datatype(const MPIDI_RMA_dtype_info *dtype_info,
                           const void *dataloop, MPI_Aint dataloop_sz,
1052
1053
                           const void *o_addr, int o_count,
			   MPI_Datatype o_datatype,
1054
                           MPID_Datatype **combined_dtp);
1055

1056
1057
1058
/* Issue an RMA operation -- Before calling this macro, you must define the
 * MPIDI_CH3I_TRACK_RMA_WRITE helper macro.  This macro defines any extra action
 * that should be taken when a write (put/acc) operation is encountered. */
1059
#define MPIDI_CH3I_ISSUE_RMA_OP(op_ptr_, win_ptr_, flags_, source_win_handle_, target_win_handle_,err_) \
1060
    do {                                                                                        \
1061
    switch ((op_ptr_)->type)                                                                    \
1062
    {                                                                                           \
1063
1064
        case (MPIDI_RMA_PUT):                                                                   \
        case (MPIDI_RMA_ACCUMULATE):                                                            \
1065
            MPIDI_CH3I_TRACK_RMA_WRITE(op_ptr_, win_ptr_);                                      \
1066
1067
1068
            (err_) = send_rma_msg((op_ptr_), (win_ptr_), (flags_), (source_win_handle_),        \
                                                (target_win_handle_), &(op_ptr_)->dtype_info,   \
                                                &(op_ptr_)->dataloop, &(op_ptr_)->request);     \
1069
1070
            if (err_) { MPIU_ERR_POP(err_); }                                                   \
            break;                                                                              \
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
        case (MPIDI_RMA_GET_ACCUMULATE):                                                        \
            if ((op_ptr_)->op == MPI_NO_OP) {                                                   \
                /* Note: Origin arguments are ignored for NO_OP, so we don't                    \
                 * need to release a ref to the origin datatype. */                             \
                                                                                                \
                /* Convert the GAcc to a Get */                                                 \
                (op_ptr_)->type            = MPIDI_RMA_GET;                                     \
                (op_ptr_)->origin_addr     = (op_ptr_)->result_addr;                            \
                (op_ptr_)->origin_count    = (op_ptr_)->result_count;                           \
                (op_ptr_)->origin_datatype = (op_ptr_)->result_datatype;                        \
                                                                                                \
                (err_) = recv_rma_msg((op_ptr_), (win_ptr_), (flags_), (source_win_handle_),    \
                                                    (target_win_handle_), &(op_ptr_)->dtype_info,\
                                                    &(op_ptr_)->dataloop, &(op_ptr_)->request); \
            } else {                                                                            \
                MPIDI_CH3I_TRACK_RMA_WRITE(op_ptr_, win_ptr_);                                  \
                (err_) = send_rma_msg((op_ptr_), (win_ptr_), (flags_), (source_win_handle_),    \
                                                    (target_win_handle_), &(op_ptr_)->dtype_info,\
                                                    &(op_ptr_)->dataloop, &(op_ptr_)->request); \
            }                                                                                   \
            if (err_) { MPIU_ERR_POP(err_); }                                                   \
            break;                                                                              \
        case MPIDI_RMA_ACC_CONTIG:                                                              \
1094
            MPIDI_CH3I_TRACK_RMA_WRITE(op_ptr_, win_ptr_);                                      \
1095
1096
1097
            (err_) = send_contig_acc_msg((op_ptr_), (win_ptr_), (flags_),                       \
                                                       (source_win_handle_), (target_win_handle_),\
                                                       &(op_ptr_)->request );                   \
1098
1099
            if (err_) { MPIU_ERR_POP(err_); }                                                   \
            break;                                                                              \
1100
1101
1102
1103
1104
        case (MPIDI_RMA_GET):                                                                   \
            (err_) = recv_rma_msg((op_ptr_), (win_ptr_), (flags_),                              \
                                                (source_win_handle_), (target_win_handle_),     \
                                                &(op_ptr_)->dtype_info,                         \
                                                &(op_ptr_)->dataloop, &(op_ptr_)->request);     \
1105
1106
            if (err_) { MPIU_ERR_POP(err_); }                                                   \
            break;                                                                              \
1107
1108
        case (MPIDI_RMA_COMPARE_AND_SWAP):                                                      \
        case (MPIDI_RMA_FETCH_AND_OP):                                                          \
1109
            MPIDI_CH3I_TRACK_RMA_WRITE(op_ptr_, win_ptr_);                                      \
1110
1111
1112
            (err_) = send_immed_rmw_msg((op_ptr_), (win_ptr_), (flags_),                        \
                                                      (source_win_handle_), (target_win_handle_),\
                                                      &(op_ptr_)->request );                    \
1113
1114
1115
1116
1117
1118
1119
1120
1121
            if (err_) { MPIU_ERR_POP(err_); }                                                   \
            break;                                                                              \
                                                                                                \
        default:                                                                                \
            MPIU_ERR_SETANDJUMP(err_,MPI_ERR_OTHER,"**winInvalidOp");                           \
    }                                                                                           \
    } while (0)


1122
1123
1124
1125
1126
1127
1128
#undef FUNCNAME
#define FUNCNAME MPIDI_Win_fence
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
int MPIDI_Win_fence(int assert, MPID_Win *win_ptr)
{
    int mpi_errno = MPI_SUCCESS;
1129
    int comm_size;
1130
    int *rma_target_proc, *nops_to_proc, i, total_op_count, *curr_ops_cnt;
1131
    MPIDI_RMA_Op_t *curr_ptr;
1132
    MPIDI_RMA_Ops_list_t *ops_list;
1133
    MPID_Comm *comm_ptr;
1134
    MPI_Win source_win_handle, target_win_handle;
1135
    MPID_Progress_state progress_state;
1136
    int errflag = FALSE;
1137
    MPIU_CHKLMEM_DECL(3);
1138
1139
1140
1141
    MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_FENCE);

    MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_FENCE);

James Dinan's avatar
James Dinan committed
1142
1143
1144
1145
    MPIU_ERR_CHKANDJUMP(win_ptr->epoch_state != MPIDI_EPOCH_NONE &&
                        win_ptr->epoch_state != MPIDI_EPOCH_FENCE,
                        mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");

1146
1147
1148
1149
1150
1151
1152
1153
1154
    /* In case this process was previously the target of passive target rma
     * operations, we need to take care of the following...
     * Since we allow MPI_Win_unlock to return without a done ack from
     * the target in the case of multiple rma ops and exclusive lock,
     * we need to check whether there is a lock on the window, and if
     * there is a lock, poke the progress engine until the operartions
     * have completed and the lock is released. */
    if (win_ptr->current_lock_type != MPID_LOCK_NONE)
    {
1155
	MPIR_T_PVAR_TIMER_START(RMA, rma_winfence_clearlock);
1156
1157
1158
1159
1160
1161
	MPID_Progress_start(&progress_state);
	while (win_ptr->current_lock_type != MPID_LOCK_NONE)
	{
	    /* poke the progress engine */
	    mpi_errno = MPID_Progress_wait(&progress_state);
	    /* --BEGIN ERROR HANDLING-- */
1162
	    if (mpi_errno != MPI_SUCCESS) {
1163
		MPID_Progress_end(&progress_state);
1164
		MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**winnoprogress");
1165
1166
	    }
	    /* --END ERROR HANDLING-- */
1167
	    MPIR_T_PVAR_COUNTER_INC(RMA, rma_winfence_clearlock_aux, 1);
1168
1169
	}
	MPID_Progress_end(&progress_state);
1170
	MPIR_T_PVAR_TIMER_END(RMA, rma_winfence_clearlock);
1171
    }
1172

1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
    if (!(assert & MPI_MODE_NOSUCCEED) &&
        (assert & MPI_MODE_NOPRECEDE || win_ptr->fence_issued == 0)) {

        /* In the FENCE that opens an epoch but does not close an epoch,
           if SHM is allocated, perform a barrier among processes on the
           same node, to prevent one process modifying another process's
           memory before that process starts an epoch. */

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

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

            mpi_errno = MPIR_Barrier_impl(node_comm_ptr, &errflag);
            if (mpi_errno) {goto fn_fail;}

            /* Ensure ordering of load/store operations. */
            OPA_read_write_barrier();
        }
    }

1195
1196
    /* Note that the NOPRECEDE and NOSUCCEED must be specified by all processes
       in the window's group if any specify it */
1197
1198
    if (assert & MPI_MODE_NOPRECEDE)
    {
1199
1200
1201
1202
        /* Error: Operations were issued and the user claimed NOPRECEDE */
        MPIU_ERR_CHKANDJUMP(win_ptr->epoch_state == MPIDI_EPOCH_FENCE,
                            mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");

1203
	win_ptr->fence_issued = (assert & MPI_MODE_NOSUCCEED) ? 0 : 1;
1204
1205
1206
	goto fn_exit;
    }
    
1207
    if (win_ptr->fence_issued == 0)
1208
    {
1209
	/* win_ptr->fence_issued == 0 means either this is the very first
1210
1211
	   call to fence or the preceding fence had the
	   MPI_MODE_NOSUCCEED assert. 
1212
1213
1214
1215

           If this fence has MPI_MODE_NOSUCCEED, do nothing and return.
	   Otherwise just increment the fence count and return. */

1216
	if (!(assert & MPI_MODE_NOSUCCEED)) win_ptr->fence_issued = 1;
1217
1218
1219
    }
    else
    {
1220
1221
	int nRequest = 0;
	int nRequestNew = 0;
1222
1223

        /* Ensure ordering of load/store operations. */
1224
        if (win_ptr->shm_allocated == TRUE) {
1225
1226
1227
           OPA_read_write_barrier();
        }

1228
	MPIR_T_PVAR_TIMER_START(RMA, rma_winfence_rs);
1229
	/* This is the second or later fence. Do all the preceding RMA ops. */
1230
	comm_ptr = win_ptr->comm_ptr;
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
	/* First inform every process whether it is a target of RMA
	   ops from this process */
	comm_size = comm_ptr->local_size;

	MPIU_CHKLMEM_MALLOC(rma_target_proc, int *, comm_size*sizeof(int),
			    mpi_errno, "rma_target_proc");
	for (i=0; i<comm_size; i++) rma_target_proc[i] = 0;
	
	/* keep track of no. of ops to each proc. Needed for knowing
	   whether or not to decrement the completion counter. The
	   completion counter is decremented only on the last
	   operation. */
	MPIU_CHKLMEM_MALLOC(nops_to_proc, int *, comm_size*sizeof(int),
			    mpi_errno, "nops_to_proc");
	for (i=0; i<comm_size; i++) nops_to_proc[i] = 0;

1247
1248
1249
        /* Note, active target uses the following ops list, and passive
           target uses win_ptr->targets[..] */
        ops_list = &win_ptr->at_rma_ops_list;
1250

1251
1252
1253
	/* set rma_target_proc[i] to 1 if rank i is a target of RMA
	   ops from this process */
	total_op_count = 0;
1254
        curr_ptr = MPIDI_CH3I_RMA_Ops_head(ops_list);
1255
1256
	while (curr_ptr != NULL)
	{
Xin Zhao's avatar
Xin Zhao committed
1257
1258
1259
	    total_op_count++;
	    rma_target_proc[curr_ptr->target_rank] = 1;
	    nops_to_proc[curr_ptr->target_rank]++;
1260
1261
1262
1263
1264
1265
	    curr_ptr = curr_ptr->next;
	}
	
	MPIU_CHKLMEM_MALLOC(curr_ops_cnt, int *, comm_size*sizeof(int),
			    mpi_errno, "curr_ops_cnt");
	for (i=0; i<comm_size; i++) curr_ops_cnt[i] = 0;
1266
1267
	/* do a reduce_scatter_block (with MPI_SUM) on rma_target_proc. 
	   As a result,
1268
1269
1270
1271
1272
1273
	   each process knows how many other processes will be doing
	   RMA ops on its window */  
            
	/* first initialize the completion counter. */
	win_ptr->my_counter = comm_size;
            
1274
	mpi_errno = MPIR_Reduce_scatter_block_impl(MPI_IN_PLACE, rma_target_proc, 1,
1275
                                                   MPI_INT, MPI_SUM, comm_ptr, &errflag);
1276
	MPIR_T_PVAR_TIMER_END(RMA, rma_winfence_rs);
1277
1278
	/* result is stored in rma_target_proc[0] */
	if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
1279
        MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
1280

1281
1282
1283
1284
1285
        /* Ensure ordering of load/store operations. */
        if (win_ptr->shm_allocated == TRUE) {
            OPA_read_write_barrier();
        }

1286
1287
1288
1289
1290
	/* Set the completion counter */
	/* FIXME: MT: this needs to be done atomically because other
	   procs have the address and could decrement it. */
	win_ptr->my_counter = win_ptr->my_counter - comm_size + 
	    rma_target_proc[0];  
1291

1292
1293
1294
1295
1296
1297
1298
1299
    MPIR_T_PVAR_TIMER_START(RMA, rma_winfence_issue);
    MPIR_T_PVAR_COUNTER_INC(RMA, rma_winfence_issue_aux, total_op_count);
    MPIR_T_PVAR_ULONG2_HIGHWATERMARK_UPDATE(RMA, rma_winfence_issue_aux, total_op_count);

    MPIR_T_PVAR_ULONG2_HIGHWATERMARK_UPDATE(RMA, rma_winfence_reqs, MPIR_T_PVAR_LEVEL_GET(rma_winfence_reqs));
    MPIR_T_PVAR_LEVEL_SET(RMA, rma_winfence_reqs, 0); /* reset the level */
    MPIR_T_PVAR_COUNTER_INC(RMA, rma_winfence_reqs, 1);

1300
	i = 0;
1301
        curr_ptr = MPIDI_CH3I_RMA_Ops_head(ops_list);
1302
1303
	while (curr_ptr != NULL)
	{
1304
1305
            MPIDI_CH3_Pkt_flags_t flags = MPIDI_CH3_PKT_FLAG_NONE;

1306
	    /* The completion counter at the target is decremented only on 
1307
	       the last RMA operation. */
1308
	    if (curr_ops_cnt[curr_ptr->target_rank] ==
1309
1310
1311
1312
                nops_to_proc[curr_ptr->target_rank] - 1) {
                flags = MPIDI_CH3_PKT_FLAG_RMA_AT_COMPLETE;
            }

1313
1314
1315
            source_win_handle = win_ptr->handle;
	    target_win_handle = win_ptr->all_win_handles[curr_ptr->target_rank];

1316
#define MPIDI_CH3I_TRACK_RMA_WRITE(op_ptr_, win_ptr_) /* Not used by active mode */
1317
1318
            MPIDI_CH3I_ISSUE_RMA_OP(curr_ptr, win_ptr, flags,
                                    source_win_handle, target_win_handle, mpi_errno);
1319
#undef MPIDI_CH3I_TRACK_RMA_WRITE
1320

1321
1322
	    i++;
	    curr_ops_cnt[curr_ptr->target_rank]++;
1323
1324
	    /* If the request is null, we can remove it immediately */
	    if (!curr_ptr->request) {
1325
                MPIDI_CH3I_RMA_Ops_free_and_next(ops_list, &curr_ptr);
1326
1327
	    }
	    else  {
1328
		nRequest++;
1329
		MPIR_T_PVAR_LEVEL_INC(RMA, rma_winfence_reqs, 1);
1330
		curr_ptr    = curr_ptr->next;
1331
1332
1333
1334
1335
		/* The test on the difference is to reduce the number
		   of times the partial complete routine is called. Without
		   this, significant overhead is added once the
		   number of requests exceeds the threshold, since the
		   number that are completed in a call may be small. */
1336
1337
		if (nRequest > MPIR_CVAR_CH3_RMA_NREQUEST_THRESHOLD &&
		    nRequest - nRequestNew > MPIR_CVAR_CH3_RMA_NREQUEST_NEW_THRESHOLD) {
1338
		    int nDone = 0;
1339
1340
                    mpi_errno = poke_progress_engine();
                    if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
1341
1342
1343
            MPIR_T_PVAR_STMT(RMA, list_complete_timer=MPIR_T_PVAR_TIMER_ADDR(rma_winfence_complete));
            MPIR_T_PVAR_STMT(RMA, list_complete_counter=MPIR_T_PVAR_COUNTER_ADDR(rma_winfence_complete_aux));

1344
1345
                    mpi_errno = rma_list_gc(win_ptr, ops_list, curr_ptr, &nDone);
                    if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
1346
1347
1348
1349
		    /* if (nDone > 0) printf( "nDone = %d\n", nDone ); */
		    nRequest -= nDone;
		    nRequestNew = nRequest;
		}
1350
	    }
1351
	}
1352
	MPIR_T_PVAR_TIMER_END(RMA, rma_winfence_issue);
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364

	/* We replaced a loop over an array of requests with a list of the
	   incomplete requests.  The reason to do 
	   that is for long lists - processing the entire list until
	   all are done introduces a potentially n^2 time.  In 
	   testing with test/mpi/perf/manyrma.c , the number of iterations
	   within the "while (total_op_count) was O(total_op_count).
	   
	   Another alternative is to create a more compressed list (storing
	   only the necessary information, reducing the number of cache lines
	   needed while looping through the requests.
	*/
1365
1366
	if (total_op_count)
	{ 
1367
1368
1369
        MPIR_T_PVAR_STMT(RMA, list_complete_timer=MPIR_T_PVAR_TIMER_ADDR(rma_winfence_complete));
        MPIR_T_PVAR_STMT(RMA, list_complete_counter=MPIR_T_PVAR_COUNTER_ADDR(rma_winfence_complete_aux));
        MPIR_T_PVAR_STMT(RMA, list_block_timer=MPIR_T_PVAR_TIMER_ADDR(rma_winfence_block));
1370
1371
            mpi_errno = rma_list_complete(win_ptr, ops_list);
            if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno);
1372
	}
1373

1374
1375
1376
1377
1378
        /* MT: avoid processing unissued operations enqueued by other threads
           in rma_list_complete() */
        curr_ptr = MPIDI_CH3I_RMA_Ops_head(ops_list);
        if (curr_ptr && !curr_ptr->request)
            goto finish_up;
1379
        MPIU_Assert(MPIDI_CH3I_RMA_Ops_isempty(ops_list));
1380
	
1381
 finish_up:
1382
1383
1384
	/* wait for all operations from other processes to finish */
	if (win_ptr->my_counter)
	{
1385
	    MPIR_T_PVAR_TIMER_START(RMA, rma_winfence_wait);
1386
1387
1388
1389
1390
	    MPID_Progress_start(&progress_state);
	    while (win_ptr->my_counter)
	    {
		mpi_errno = MPID_Progress_wait(&progress_state);
		/* --BEGIN ERROR HANDLING-- */
1391
		if (mpi_errno != MPI_SUCCESS) {
1392
		    MPID_Progress_end(&progress_state);
1393
		    MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**winnoprogress");
1394
1395
		}
		/* --END ERROR HANDLING-- */
1396
		MPIR_T_PVAR_COUNTER_INC(RMA, rma_winfence_wait_aux, 1);
1397
1398
	    }
	    MPID_Progress_end(&progress_state);
1399
	    MPIR_T_PVAR_TIMER_END(RMA, rma_winfence_wait);
1400
1401
1402
1403
	} 
	
	if (assert & MPI_MODE_NOSUCCEED)
	{
1404
	    win_ptr->fence_issued = 0;
1405
1406
	}

James Dinan's avatar
James Dinan committed
1407
1408
1409
        win_ptr->epoch_state = MPIDI_EPOCH_NONE;
    }

1410
 fn_exit:
1411
1412
1413
1414
1415
1416
1417
1418
1419
    MPIU_CHKLMEM_FREEALL();
    MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_FENCE);
    return mpi_errno;
    /* --BEGIN ERROR HANDLING-- */
 fn_fail:
    goto fn_exit;
    /* --END ERROR HANDLING-- */
}

1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
/* create_datatype() creates a new struct datatype for the dtype_info
   and the dataloop of the target datatype together with the user data */
#undef FUNCNAME
#define FUNCNAME create_datatype
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
static int create_datatype(const MPIDI_RMA_dtype_info *dtype_info,
                           const void *dataloop, MPI_Aint dataloop_sz,
                           const void *o_addr, int o_count, MPI_Datatype o_datatype,
                           MPID_Datatype **combined_dtp)
{
    int mpi_errno = MPI_SUCCESS;
    /* datatype_set_contents wants an array 'ints' which is the
       blocklens array with count prepended to it.  So blocklens
       points to the 2nd element of ints to avoid having to copy
       blocklens into ints later. */
    int ints[4];
    int *blocklens = &ints[1];
    MPI_Aint displaces[3];
    MPI_Datatype datatypes[3];
    const int count = 3;
    MPI_Datatype combined_datatype;
    MPIDI_STATE_DECL(MPID_STATE_CREATE_DATATYPE);

    MPIDI_FUNC_ENTER(MPID_STATE_CREATE_DATATYPE);

    /* create datatype */
    displaces[0] = MPIU_PtrToAint(dtype_info);
    blocklens[0] = sizeof(*dtype_info);
    datatypes[0] = MPI_BYTE;
    
    displaces[1] = MPIU_PtrToAint(dataloop);
1452
    MPIU_Assign_trunc(blocklens[1], dataloop_sz, int);
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
    datatypes[1] = MPI_BYTE;
    
    displaces[2] = MPIU_PtrToAint(o_addr);
    blocklens[2] = o_count;
    datatypes[2] = o_datatype;
    
    mpi_errno = MPID_Type_struct(count,
                                 blocklens,
                                 displaces,
                                 datatypes,
                                 &combined_datatype);
1464
1465
    if (mpi_errno) MPIU_ERR_POP(mpi_errno);
   
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
    ints[0] = count;

    MPID_Datatype_get_ptr(combined_datatype, *combined_dtp);    
    mpi_errno = MPID_Datatype_set_contents(*combined_dtp,
				           MPI_COMBINER_STRUCT,
				           count+1, /* ints (cnt,blklen) */
				           count, /* aints (disps) */
				           count, /* types */
				           ints,
				           displaces,
				           datatypes);
    if (mpi_errno) MPIU_ERR_POP(mpi_errno);

    /* Commit datatype */
    
    MPID_Dataloop_create(combined_datatype,
                         &(*combined_dtp)->dataloop,
                         &(*combined_dtp)->dataloop_size,
                         &(*combined_dtp)->dataloop_depth,
                         MPID_DATALOOP_HOMOGENEOUS);
    
    /* create heterogeneous dataloop */
    MPID_Dataloop_create(combined_datatype,
                         &(*combined_dtp)->hetero_dloop,
                         &(*combined_dtp)->hetero_dloop_size,
                         &(*combined_dtp)->hetero_dloop_depth,
                         MPID_DATALOOP_HETEROGENEOUS);
 
 fn_exit:
    MPIDI_FUNC_EXIT(MPID_STATE_CREATE_DATATYPE);
    return mpi_errno;
 fn_fail:
    goto fn_exit;
}

1501
1502

#undef FUNCNAME
Xin Zhao's avatar
Xin Zhao committed
1503
#define FUNCNAME send_rma_msg
1504
1505
#undef FCNAME
#define FCNAME MPIDI_QUOTE(FUNCNAME)
1506
1507
1508
1509
1510
1511
static int send_rma_msg(MPIDI_RMA_Op_t *rma_op, MPID_Win *win_ptr,
                                   MPIDI_CH3_Pkt_flags_t flags,
				   MPI_Win source_win_handle, 
				   MPI_Win target_win_handle, 
				   MPIDI_RMA_dtype_info *dtype_info, 
				   void **dataloop, MPID_Request **request) 
1512
{
1513
1514
1515
    MPIDI_CH3_Pkt_t upkt;
    MPIDI_CH3_Pkt_put_t *put_pkt = &upkt.put;
    MPIDI_CH3_Pkt_accum_t *accum_pkt = &upkt.accum;
1516
    MPID_IOV iov[MPID_IOV_LIMIT];
1517
    int mpi_errno=MPI_SUCCESS;
1518
1519
    int origin_dt_derived, target_dt_derived, iovcnt;
    MPI_Aint origin_type_size;
1520
1521
1522
    MPIDI_VC_t * vc;
    MPID_Comm *comm_ptr;
    MPID_Datatype *target_dtp=NULL, *origin_dtp=NULL;
1523
    MPID_Request *resp_req=NULL;
1524
    MPIU_CHKPMEM_DECL(1);
Xin Zhao's avatar
Xin Zhao committed
1525
    MPIDI_STATE_DECL(MPID_STATE_SEND_RMA_MSG);
1526
1527
    MPIDI_STATE_DECL(MPID_STATE_MEMCPY);

Xin Zhao's avatar
Xin Zhao committed
1528
    MPIDI_RMA_FUNC_ENTER(MPID_STATE_SEND_RMA_MSG);
1529

1530
    *request = NULL;
1531

1532
    if (rma_op->type == MPIDI_RMA_PUT)
1533
    {
1534
1535
1536
        MPIDI_Pkt_init(put_pkt, MPIDI_CH3_PKT_PUT);
        put_pkt->addr = (char *) win_ptr->base_addrs[rma_op->target_rank] +
            win_ptr->disp_units[rma_op->target_rank] * rma_op->target_disp;
1537
        put_pkt->flags = flags;
1538
1539
1540
1541
1542
1543
        put_pkt->count = rma_op->target_count;
        put_pkt->datatype = rma_op->target_datatype;
        put_pkt->dataloop_size = 0;
        put_pkt->target_win_handle = target_win_handle;
        put_pkt->source_win_handle = source_win_handle;
        
1544
1545
1546
        iov[0].MPID_IOV_BUF = (MPID_IOV_BUF_CAST) put_pkt;
        iov[0].MPID_IOV_LEN = sizeof(*put_pkt);
    }
1547
    else if (rma_op->type == MPIDI_RMA_GET_ACCUMULATE)
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
    {
        /* Create a request for the GACC response.  Store the response buf, count, and
           datatype in it, and pass the request's handle in the GACC packet. When the
           response comes from the target, it will contain the request handle. */
        resp_req = MPID_Request_create();
        MPIU_ERR_CHKANDJUMP(resp_req == NULL, mpi_errno, MPI_ERR_OTHER, "**nomemreq");

        MPIU_Object_set_ref(resp_req, 2);

        resp_req->dev.user_buf = rma_op->result_addr;
        resp_req->dev.user_count = rma_op->result_count;
        resp_req->dev.datatype = rma_op->result_datatype;
1560
1561
        resp_req->dev.target_win_handle = target_win_handle;
        resp_req->dev.source_win_handle = source_win_handle;
1562

1563
        if (!MPIR_DATATYPE_IS_PREDEFINED(resp_req->dev.datatype)) {
1564
1565
1566
1567
1568
1569
1570
1571
            MPID_Datatype *result_dtp = NULL;
            MPID_Datatype_get_ptr(resp_req->dev.datatype, result_dtp);
            resp_req->dev.datatype_ptr = result_dtp;
            /* this will cause the datatype to be freed when the
               request is freed. */
        }

        /* Note: Get_accumulate uses the same packet type as accumulate */
1572
1573
1574
        MPIDI_Pkt_init(accum_pkt, MPIDI_CH3_PKT_GET_ACCUM);
        accum_pkt->addr = (char *) win_ptr->base_addrs[rma_op->target_rank] +
            win_ptr->disp_units[rma_op->target_rank] * rma_op->target_disp;
Pavan Balaji's avatar
Pavan Balaji committed
1575
        accum_pkt->flags = flags;
1576
1577
1578
1579
1580
1581
1582
1583
        accum_pkt->count = rma_op->target_count;
        accum_pkt->datatype = rma_op->target_datatype;
        accum_pkt->dataloop_size = 0;
        accum_pkt->op = rma_op->op;
        accum_pkt->target_win_handle = target_win_handle;
        accum_pkt->source_win_handle = source_win_handle;
        accum_pkt->request_handle = resp_req->handle;

1584
1585
1586
        iov[0].MPID_IOV_BUF = (MPID_IOV_BUF_CAST) accum_pkt;
        iov[0].MPID_IOV_LEN = sizeof(*accum_pkt);
    }
1587
1588
    else
    {
1589
1590
1591
        MPIDI_Pkt_init(accum_pkt, MPIDI_CH3_PKT_ACCUMULATE);
        accum_pkt->addr = (char *) win_ptr->base_addrs[rma_op->target_rank] +
            win_ptr->disp_units[rma_op->target_rank] * rma_op->target_disp;
1592
        accum_pkt->flags = flags;
1593
1594
1595
1596
1597
1598
1599
        accum_pkt->count = rma_op->target_count;
        accum_pkt->datatype = rma_op->target_datatype;
        accum_pkt->dataloop_size = 0;
        accum_pkt->op = rma_op->op;
        accum_pkt->target_win_handle = target_win_handle;
        accum_pkt->source_win_handle = source_win_handle;

1600
1601
1602
1603
        iov[0].MPID_IOV_BUF = (MPID_IOV_BUF_CAST) accum_pkt;
        iov[0].MPID_IOV_LEN = sizeof(*accum_pkt);
    }

1604
1605
1606
1607
    /*    printf("send pkt: type %d, addr %d, count %d, base %d\n", rma_pkt->type,
          rma_pkt->addr, rma_pkt->count, win_ptr->base_addrs[rma_op->target_rank]);
          fflush(stdout);
    */
1608

1609
    comm_ptr = win_ptr->comm_ptr;
1610
    MPIDI_Comm_get_vc_set_active(comm_ptr, rma_op->target_rank, &vc);
1611

1612
    if (!MPIR_DATATYPE_IS_PREDEFINED(rma_op->origin_datatype))
1613
1614
1615
1616
1617
1618
1619
1620
1621
    {
        origin_dt_derived = 1;
        MPID_Datatype_get_ptr(rma_op->origin_datatype, origin_dtp);
    }
    else
    {
        origin_dt_derived = 0;
    }

1622
    if (!MPIR_DATATYPE_IS_PREDEFINED(rma_op->target_datatype))
1623
1624
    {
        target_dt_derived = 1;
1625
        MPID_Datatype_get_ptr(rma_op->target_datatype, target_dtp);
1626
1627
1628
1629
1630
1631
1632
1633
1634
    }
    else
    {
        target_dt_derived = 0;
    }

    if (target_dt_derived)
    {
        /* derived datatype on target. fill derived datatype info */
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
        dtype_info->is_contig = target_dtp->is_contig;
        dtype_info->max_contig_blocks = target_dtp->max_contig_blocks;
        dtype_info->size = target_dtp->size;
        dtype_info->extent = target_dtp->extent;
        dtype_info->dataloop_size = target_dtp->dataloop_size;
        dtype_info->dataloop_depth = target_dtp->dataloop_depth;
        dtype_info->eltype = target_dtp->eltype;
        dtype_info->dataloop = target_dtp->dataloop;
        dtype_info->ub = target_dtp->ub;
        dtype_info->lb = target_dtp->lb;
        dtype_info->true_ub = target_dtp->true_ub;
        dtype_info->true_lb = target_dtp->true_lb;
        dtype_info->has_sticky_ub = target_dtp->has_sticky_ub;
        dtype_info->has_sticky_lb = target_dtp->has_sticky_lb;

	MPIU_CHKPMEM_MALLOC(*dataloop, void *, target_dtp->dataloop_size, 
1651
1652
1653
			    mpi_errno, "dataloop");

	MPIDI_FUNC_ENTER(MPID_STATE_MEMCPY);
1654
        MPIU_Memcpy(*dataloop, target_dtp->dataloop, target_dtp->dataloop_size);
1655
	MPIDI_FUNC_EXIT(MPID_STATE_MEMCPY);
1656
1657
        /* the dataloop can have undefined padding sections, so we need to let
         * valgrind know that it is OK to pass this data to writev later on */
1658
        MPL_VG_MAKE_MEM_DEFINED(*dataloop, target_dtp->dataloop_size);
1659

1660
        if (rma_op->type == MPIDI_RMA_PUT)
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
	{
            put_pkt->dataloop_size = target_dtp->dataloop_size;
	}
        else
	{
            accum_pkt->dataloop_size = target_dtp->dataloop_size;
	}
    }

    MPID_Datatype_get_size_macro(rma_op->origin_datatype, origin_type_size);

1672
    if (!target_dt_derived)
1673
    {
1674
1675
1676
1677
        /* basic datatype on target */
        if (!origin_dt_derived)
        {
            /* basic datatype on origin */
1678
1679
1680
            iov[1].MPID_IOV_BUF = (MPID_IOV_BUF_CAST)rma_op->origin_addr;
            iov[1].MPID_IOV_LEN = rma_op->origin_count * origin_type_size;
            iovcnt = 2;
1681
	    MPIU_THREAD_CS_ENTER(CH3COMM,vc);
1682
            mpi_errno = MPIDI_CH3_iStartMsgv(vc, iov, iovcnt, request);
1683
	    MPIU_THREAD_CS_EXIT(CH3COMM,vc);
1684
            MPIU_ERR_CHKANDJUMP(mpi_errno, mpi_errno, MPI_ERR_OTHER, "**ch3|rmamsg");
1685
1686
        }
        else
1687
1688
        {
            /* derived datatype on origin */
1689
1690
            *request = MPID_Request_create();
            MPIU_ERR_CHKANDJUMP(*request == NULL,mpi_errno,MPI_ERR_OTHER,"**nomemreq");
1691
            
1692
1693
            MPIU_Object_set_ref(*request, 2);
            (*request)->kind = MPID_REQUEST_SEND;
1694
            
1695
1696
            (*request)->dev.segment_ptr = MPID_Segment_alloc( );
            MPIU_ERR_CHKANDJUMP1((*request)->dev.segment_ptr == NULL, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment_alloc");
1697

1698
            (*request)->dev.datatype_ptr = origin_dtp;
1699
1700
1701
1702
            /* this will cause the datatype to be freed when the request
               is freed. */
            MPID_Segment_init(rma_op->origin_addr, rma_op->origin_count,
                              rma_op->origin_datatype,
1703
1704
1705
                              (*request)->dev.segment_ptr, 0);
            (*request)->dev.segment_first = 0;
            (*request)->dev.segment_size = rma_op->origin_count * origin_type_size;
1706

1707
1708
            (*request)->dev.OnFinal = 0;
            (*request)->dev.OnDataAvail = 0;
1709

1710
	    MPIU_THREAD_CS_ENTER(CH3COMM,vc);
1711
            mpi_errno = vc->sendNoncontig_fn(vc, *request, iov[0].MPID_IOV_BUF, iov[0].MPID_IOV_LEN);
1712
	    MPIU_THREAD_CS_EXIT(CH3COMM,vc);
1713
            MPIU_ERR_CHKANDJUMP(mpi_errno, mpi_errno, MPI_ERR_OTHER, "**ch3|rmamsg");
1714
1715
1716
1717
        }
    }
    else
    {
1718
        /* derived datatype on target */
1719
        MPID_Datatype *combined_dtp = NULL;
1720

1721
1722
        *request = MPID_Request_create();
        if (*request == NULL) {
1723
	    MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomemreq");
1724
1725
        }

1726
1727
        MPIU_Object_set_ref(*request, 2);
        (*request)->kind = MPID_REQUEST_SEND;
1728

1729
1730
	(*request)->dev.segment_ptr = MPID_Segment_alloc( );
        MPIU_ERR_CHKANDJUMP1((*request)->dev.segment_ptr == NULL, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment_alloc");
1731
1732
1733

        /* create a new datatype containing the dtype_info, dataloop, and origin data */

1734
        mpi_errno = create_datatype(dtype_info, *dataloop, target_dtp->dataloop_size, rma_op->origin_addr,
1735
                                    rma_op->origin_count, rma_op->origin_datatype, &combined_dtp);
1736
        if (mpi_errno) MPIU_ERR_POP(mpi_errno);
1737

1738
        (*request)->dev.datatype_ptr = combined_dtp;
1739
1740
1741
        /* combined_datatype will be freed when request is freed */

        MPID_Segment_init(MPI_BOTTOM, 1, combined_dtp->handle,
1742
1743
1744
                          (*request)->dev.segment_ptr, 0);
        (*request)->dev.segment_first = 0;
        (*request)->dev.segment_size = combined_dtp->size;
1745

1746
1747
        (*request)->dev.OnFinal = 0;
        (*request)->dev.OnDataAvail = 0;
1748

1749
	MPIU_THREAD_CS_ENTER(CH3COMM,vc);
1750
        mpi_errno = vc->sendNoncontig_fn(vc, *request, iov[0].MPID_IOV_BUF, iov[0].MPID_IOV_LEN);
1751
	MPIU_THREAD_CS_EXIT(CH3COMM,vc);
1752
        MPIU_ERR_CHKANDJUMP(mpi_errno, mpi_errno, MPI_ERR_OTHER, "**ch3|rmamsg");
1753

1754
1755
1756
        /* we're done with the datatypes */
        if (origin_dt_derived)
            MPID_Datatype_release(origin_dtp);
1757
        MPID_Datatype_release(target_dtp);
1758
1759
1760
1761
1762
    }

    /* This operation can generate two requests; one for inbound and one for
       outbound data. */
    if (resp_req != NULL) {