summaryrefslogtreecommitdiffstats
path: root/src/H5Dmpio.c
blob: 01cf9320c4c4947e29d9fd9da60c94e4e3dc8381 (plain)
1
2
3
4
5
6
7
8
9
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
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
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
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
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
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
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
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
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
 * Copyright by The HDF Group.                                               *
 * Copyright by the Board of Trustees of the University of Illinois.         *
 * All rights reserved.                                                      *
 *                                                                           *
 * This file is part of HDF5.  The full HDF5 copyright notice, including     *
 * terms governing use, modification, and redistribution, is contained in    *
 * the COPYING file, which can be found at the root of the source code       *
 * distribution tree, or in https://support.hdfgroup.org/ftp/HDF5/releases.  *
 * If you do not have access to either file, you may request a copy from     *
 * help@hdfgroup.org.                                                        *
 * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */

/*
 * Programmer:  rky 980813
 * KY 2005 revised the code and made the change to support and optimize
 * collective IO support.
 * Purpose:    Functions to read/write directly between app buffer and file.
 *
 *         Beware of the ifdef'ed print statements.
 *         I didn't make them portable.
 */

/****************/
/* Module Setup */
/****************/

#include "H5Dmodule.h"          /* This source code file is part of the H5D module */


/***********/
/* Headers */
/***********/
#include "H5private.h"        /* Generic Functions */
#include "H5CXprivate.h"      /* API Contexts      */
#include "H5Dpkg.h"           /* Datasets          */
#include "H5Eprivate.h"       /* Error handling    */
#include "H5Fprivate.h"       /* File access       */
#include "H5FDprivate.h"      /* File drivers      */
#include "H5Iprivate.h"       /* IDs               */
#include "H5MMprivate.h"      /* Memory management */
#include "H5Oprivate.h"       /* Object headers    */
#include "H5Pprivate.h"       /* Property lists    */
#include "H5Sprivate.h"       /* Dataspaces        */
#include "H5VMprivate.h"      /* Vector            */

#ifdef H5_HAVE_PARALLEL

/****************/
/* Local Macros */
/****************/

/* Macros to represent different IO options */
#define H5D_ONE_LINK_CHUNK_IO          0
#define H5D_MULTI_CHUNK_IO             1
#define H5D_ONE_LINK_CHUNK_IO_MORE_OPT 2
#define H5D_MULTI_CHUNK_IO_MORE_OPT    3

/***** Macros for One linked collective IO case. *****/
/* The default value to do one linked collective IO for all chunks.
   If the average number of chunks per process is greater than this value,
      the library will create an MPI derived datatype to link all chunks to do collective IO.
      The user can set this value through an API. */

/* Macros to represent options on how to obtain chunk address for one linked-chunk IO case */
#define H5D_OBTAIN_ONE_CHUNK_ADDR_IND 0
#define H5D_OBTAIN_ALL_CHUNK_ADDR_COL 2

/* Macros to define the default ratio of obtaining all chunk addresses for one linked-chunk IO case */
#define H5D_ALL_CHUNK_ADDR_THRES_COL  30
#define H5D_ALL_CHUNK_ADDR_THRES_COL_NUM 10000

/***** Macros for multi-chunk collective IO case. *****/
/* The default value of the threshold to do collective IO for this chunk.
   If the average number of processes per chunk is greater than the default value,
   collective IO is done for this chunk.
*/

/* Macros to represent different IO modes(NONE, Independent or collective)for multiple chunk IO case */
#define H5D_CHUNK_IO_MODE_COL         1

/* Macros to represent the regularity of the selection for multiple chunk IO case. */
#define H5D_CHUNK_SELECT_REG          1


/******************/
/* Local Typedefs */
/******************/
/* Combine chunk address and chunk info into a struct for better performance. */
typedef struct H5D_chunk_addr_info_t {
    haddr_t chunk_addr;
    H5D_chunk_info_t chunk_info;
} H5D_chunk_addr_info_t;

/* Rank 0 Bcast values */
typedef enum H5D_mpio_no_rank0_bcast_cause_t {
    H5D_MPIO_RANK0_BCAST = 0x00,
    H5D_MPIO_RANK0_NOT_H5S_ALL = 0x01,
    H5D_MPIO_RANK0_NOT_CONTIGUOUS = 0x02,
    H5D_MPIO_RANK0_NOT_FIXED_SIZE = 0x04,
    H5D_MPIO_RANK0_GREATER_THAN_2GB = 0x08
} H5D_mpio_no_rank0_bcast_cause_t;


/*
 * Information about a single chunk when performing collective filtered I/O. All
 * of the fields of one of these structs are initialized at the start of collective
 * filtered I/O in the function H5D__construct_filtered_io_info_list().
 *
 * This struct's fields are as follows:
 *
 *   index - The "Index" of the chunk in the dataset. The index of a chunk is used during
 *           the collective re-insertion of chunks into the chunk index after the collective
 *           I/O has been performed.
 *
 *   scaled - The scaled coordinates of the chunk in the dataset's file dataspace. The
 *            coordinates are used in both the collective re-allocation of space in the file
 *            and the collective re-insertion of chunks into the chunk index after the collective
 *            I/O has been performed.
 *
 *   full_overwrite - A flag which determines whether or not a chunk needs to be read from the
 *                    file when being updated. If a chunk is being fully overwritten (the entire
 *                    extent is selected in its file dataspace), then it is not necessary to
 *                    read the chunk from the file. However, if the chunk is not being fully
 *                    overwritten, it has to be read from the file in order to update the chunk
 *                    without trashing the parts of the chunk that are not selected.
 *
 *   num_writers - The total number of processors writing to this chunk. This field is used
 *                 when the new owner of a chunk is receiving messages, which contain selections in
 *                 the chunk and data to update the chunk with, from other processors which have this
 *                 chunk selected in the I/O operation. The new owner must know how many processors it
 *                 should expect messages from so that it can post an equal number of receive calls.
 *
 *   io_size - The total size of I/O to this chunk. This field is an accumulation of the size of
 *             I/O to the chunk from each processor which has the chunk selected and is used to
 *             determine the value for the previous full_overwrite flag.
 *
 *   buf - A pointer which serves the dual purpose of holding either the chunk data which is to be
 *         written to the file or the chunk data which has been read from the file.
 *
 *   chunk_states - In the case of dataset writes only, this struct is used to track a chunk's size and
 *                  address in the file before and after the filtering operation has occurred.
 *
 *                  Its fields are as follows:
 *
 *                  chunk_current - The address in the file and size of this chunk before the filtering
 *                                  operation. When reading a chunk from the file, this field is used to
 *                                  read the correct amount of bytes. It is also used when redistributing
 *                                  shared chunks among processors and as a parameter to the chunk file
 *                                  space reallocation function.
 *
 *                  new_chunk - The address in the file and size of this chunk after the filtering
 *                              operation. This field is relevant when collectively re-allocating space
 *                              in the file for all of the chunks written to in the I/O operation, as
 *                              their sizes may have changed after their data has been filtered.
 *
 *   owners - In the case of dataset writes only, this struct is used to manage which single processor
 *            will ultimately write data out to the chunk. It allows the other processors to act according
 *            to the decision and send their selection in the chunk, as well as the data they wish
 *            to update the chunk with, to the processor which is writing to the chunk.
 *
 *            Its fields are as follows:
 *
 *            original_owner - The processor which originally had this chunk selected at the beginning of
 *                             the collective filtered I/O operation. This field is currently used when
 *                             redistributing shared chunks among processors.
 *
 *            new_owner - The processor which has been selected to perform the write to this chunk.
 *
 *   async_info - In the case of dataset writes only, this struct is used by the owning processor of the
 *                chunk in order to manage the MPI send and receive calls made between it and all of
 *                the other processors which have this chunk selected in the I/O operation.
 *
 *                Its fields are as follows:
 *
 *                receive_requests_array - An array containing one MPI_Request for each of the
 *                                         asynchronous MPI receive calls the owning processor of this
 *                                         chunk makes to another processor in order to receive that
 *                                         processor's chunk modification data and selection in the chunk.
 *
 *                receive_buffer_array - An array of buffers into which the owning processor of this chunk
 *                                       will store chunk modification data and the selection in the chunk
 *                                       received from another processor.
 *
 *                num_receive_requests - The number of entries in the receive_request_array and
 *                                       receive_buffer_array fields.
 */
typedef struct H5D_filtered_collective_io_info_t {
  hsize_t             index;
  hsize_t             scaled[H5O_LAYOUT_NDIMS];
  hbool_t             full_overwrite;
  size_t              num_writers;
  size_t              io_size;
  void               *buf;

  struct {
      H5F_block_t     chunk_current;
      H5F_block_t     new_chunk;
  } chunk_states;

  struct {
      int             original_owner;
      int             new_owner;
  } owners;

  struct {
      MPI_Request    *receive_requests_array;
      unsigned char **receive_buffer_array;
      int             num_receive_requests;
  } async_info;
} H5D_filtered_collective_io_info_t;

/********************/
/* Local Prototypes */
/********************/
static herr_t H5D__chunk_collective_io(H5D_io_info_t *io_info,
    const H5D_type_info_t *type_info, H5D_chunk_map_t *fm);
static herr_t H5D__multi_chunk_collective_io(H5D_io_info_t *io_info,
    const H5D_type_info_t *type_info, H5D_chunk_map_t *fm);
static herr_t H5D__multi_chunk_filtered_collective_io(H5D_io_info_t *io_info,
    const H5D_type_info_t *type_info, H5D_chunk_map_t *fm);
static herr_t H5D__link_chunk_collective_io(H5D_io_info_t *io_info,
    const H5D_type_info_t *type_info, H5D_chunk_map_t *fm, int sum_chunk);
static herr_t H5D__link_chunk_filtered_collective_io(H5D_io_info_t *io_info,
    const H5D_type_info_t *type_info, H5D_chunk_map_t *fm);
static herr_t H5D__inter_collective_io(H5D_io_info_t *io_info,
    const H5D_type_info_t *type_info, const H5S_t *file_space,
    const H5S_t *mem_space);
static herr_t H5D__final_collective_io(H5D_io_info_t *io_info,
    const H5D_type_info_t *type_info, hsize_t nelmts, MPI_Datatype mpi_file_type,
    MPI_Datatype mpi_buf_type);
static herr_t H5D__sort_chunk(H5D_io_info_t *io_info, const H5D_chunk_map_t *fm,
    H5D_chunk_addr_info_t chunk_addr_info_array[], int many_chunk_opt);
static herr_t H5D__obtain_mpio_mode(H5D_io_info_t *io_info, H5D_chunk_map_t *fm,
    uint8_t assign_io_mode[], haddr_t chunk_addr[]);
static herr_t H5D__mpio_get_sum_chunk(const H5D_io_info_t *io_info,
    const H5D_chunk_map_t *fm, int *sum_chunkf);
static herr_t H5D__construct_filtered_io_info_list(const H5D_io_info_t *io_info,
    const H5D_type_info_t *type_info, const H5D_chunk_map_t *fm,
    H5D_filtered_collective_io_info_t **chunk_list, size_t *num_entries);
#if MPI_VERSION >= 3
static herr_t H5D__chunk_redistribute_shared_chunks(const H5D_io_info_t *io_info,
    const H5D_type_info_t *type_info, const H5D_chunk_map_t *fm,
    H5D_filtered_collective_io_info_t *local_chunk_array, size_t *local_chunk_array_num_entries);
#endif
static herr_t H5D__mpio_array_gatherv(void *local_array, size_t local_array_num_entries,
    size_t array_entry_size, void **gathered_array, size_t *gathered_array_num_entries,
    hbool_t allgather, int root, MPI_Comm comm, int (*sort_func)(const void *, const void *));
static herr_t H5D__mpio_filtered_collective_write_type(
    H5D_filtered_collective_io_info_t *chunk_list, size_t num_entries,
    MPI_Datatype *new_mem_type, hbool_t *mem_type_derived,
    MPI_Datatype *new_file_type, hbool_t *file_type_derived);
static herr_t H5D__filtered_collective_chunk_entry_io(H5D_filtered_collective_io_info_t *chunk_entry,
    const H5D_io_info_t *io_info, const H5D_type_info_t *type_info, const H5D_chunk_map_t *fm);
static int H5D__cmp_chunk_addr(const void *chunk_addr_info1, const void *chunk_addr_info2);
static int H5D__cmp_filtered_collective_io_info_entry(const void *filtered_collective_io_info_entry1,
    const void *filtered_collective_io_info_entry2);
#if MPI_VERSION >= 3
static int H5D__cmp_filtered_collective_io_info_entry_owner(const void *filtered_collective_io_info_entry1,
    const void *filtered_collective_io_info_entry2);
#endif


/*********************/
/* Package Variables */
/*********************/


/*******************/
/* Local Variables */
/*******************/


/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_opt_possible
 *
 * Purpose:     Checks if an direct I/O transfer is possible between memory and
 *                  the file.
 *
 * Return:      Success:   Non-negative: TRUE or FALSE
 *              Failure:    Negative
 *
 * Programmer:  Quincey Koziol
 *              Wednesday, April 3, 2002
 *
 *-------------------------------------------------------------------------
 */
htri_t
H5D__mpio_opt_possible(const H5D_io_info_t *io_info, const H5S_t *file_space,
    const H5S_t *mem_space, const H5D_type_info_t *type_info)
{
    H5FD_mpio_xfer_t io_xfer_mode;      /* MPI I/O transfer mode */
    unsigned local_cause[2] = {0,0};    /* [0] Local reason(s) for breaking collective mode */
                                        /* [1] Flag if dataset is both: H5S_ALL and small */
    unsigned global_cause[2] = {0,0};   /* Global reason(s) for breaking collective mode */
    htri_t is_vl_storage;               /* Whether the dataset's datatype is stored in a variable-length form */
    htri_t ret_value = SUCCEED;         /* Return value */

    FUNC_ENTER_PACKAGE

    /* Check args */
    HDassert(io_info);
    HDassert(mem_space);
    HDassert(file_space);
    HDassert(type_info);


    /* For independent I/O, get out quickly and don't try to form consensus */
    if(H5CX_get_io_xfer_mode(&io_xfer_mode) < 0)
        /* Set error flag, but keep going */
        local_cause[0] |= H5D_MPIO_ERROR_WHILE_CHECKING_COLLECTIVE_POSSIBLE;
    if(io_xfer_mode == H5FD_MPIO_INDEPENDENT)
        local_cause[0] |= H5D_MPIO_SET_INDEPENDENT;

    /* Optimized MPI types flag must be set */
    /* (based on 'HDF5_MPI_OPT_TYPES' environment variable) */
    if(!H5FD_mpi_opt_types_g)
        local_cause[0] |= H5D_MPIO_MPI_OPT_TYPES_ENV_VAR_DISABLED;

    /* Don't allow collective operations if datatype conversions need to happen */
    if(!type_info->is_conv_noop)
        local_cause[0] |= H5D_MPIO_DATATYPE_CONVERSION;

    /* Don't allow collective operations if data transform operations should occur */
    if(!type_info->is_xform_noop)
        local_cause[0] |= H5D_MPIO_DATA_TRANSFORMS;

    /* Check whether these are both simple or scalar dataspaces */
    if(!((H5S_SIMPLE == H5S_GET_EXTENT_TYPE(mem_space) || H5S_SCALAR == H5S_GET_EXTENT_TYPE(mem_space))
            && (H5S_SIMPLE == H5S_GET_EXTENT_TYPE(file_space) || H5S_SCALAR == H5S_GET_EXTENT_TYPE(file_space))))
        local_cause[0] |= H5D_MPIO_NOT_SIMPLE_OR_SCALAR_DATASPACES;

    /* Dataset storage must be contiguous or chunked */
    if(!(io_info->dset->shared->layout.type == H5D_CONTIGUOUS ||
            io_info->dset->shared->layout.type == H5D_CHUNKED))
        local_cause[0] |= H5D_MPIO_NOT_CONTIGUOUS_OR_CHUNKED_DATASET;

    /* check if external-file storage is used */
    if(io_info->dset->shared->dcpl_cache.efl.nused > 0)
        local_cause[0] |= H5D_MPIO_NOT_CONTIGUOUS_OR_CHUNKED_DATASET;

    /* The handling of memory space is different for chunking and contiguous
     *  storage.  For contiguous storage, mem_space and file_space won't change
     *  when it it is doing disk IO.  For chunking storage, mem_space will
     *  change for different chunks. So for chunking storage, whether we can
     *  use collective IO will defer until each chunk IO is reached.
     */

#if MPI_VERSION < 3
    /*
     * Don't allow parallel writes to filtered datasets if the MPI version
     * is less than 3. The functions needed (MPI_Mprobe and MPI_Imrecv) will
     * not be available.
     */
    if(io_info->op_type == H5D_IO_OP_WRITE &&
            io_info->dset->shared->layout.type == H5D_CHUNKED &&
            io_info->dset->shared->dcpl_cache.pline.nused > 0)
        local_cause[0] |= H5D_MPIO_PARALLEL_FILTERED_WRITES_DISABLED;
#endif

    /* Check if we are able to do a MPI_Bcast of the data from one rank
     * instead of having all the processes involved in the collective I/O call.
     */

    /* Check to see if the process is reading the entire dataset */
    if(H5S_GET_SELECT_TYPE(file_space) != H5S_SEL_ALL)
        local_cause[1] |= H5D_MPIO_RANK0_NOT_H5S_ALL; 
    /* Only perform this optimization for contigous datasets, currently */
    else if(H5D_CONTIGUOUS != io_info->dset->shared->layout.type)
        /* Flag to do a MPI_Bcast of the data from one proc instead of 
         * having all the processes involved in the collective I/O.
         */
        local_cause[1] |= H5D_MPIO_RANK0_NOT_CONTIGUOUS; 
    else if((is_vl_storage = H5T_is_vl_storage(type_info->dset_type)) < 0)
        local_cause[0] |= H5D_MPIO_ERROR_WHILE_CHECKING_COLLECTIVE_POSSIBLE;
    else if(is_vl_storage)
        local_cause[1] |= H5D_MPIO_RANK0_NOT_FIXED_SIZE;
    else {
        size_t type_size;       /* Size of dataset's datatype */

        /* Retrieve the size of the dataset's datatype */
        if(0 == (type_size = H5T_GET_SIZE(type_info->dset_type)))
            local_cause[0] |= H5D_MPIO_ERROR_WHILE_CHECKING_COLLECTIVE_POSSIBLE;
        else {
            hssize_t snelmts;   /* [Signed] # of elements in dataset's dataspace */

            /* Retrieve the size of the dataset's datatype */
            if((snelmts = H5S_GET_EXTENT_NPOINTS(file_space)) < 0)
                local_cause[0] |= H5D_MPIO_ERROR_WHILE_CHECKING_COLLECTIVE_POSSIBLE;
            else {
                hsize_t dset_size;

                /* Determine dataset size */
                dset_size = ((hsize_t)snelmts) * type_size;

               /* If the size of the dataset is less than 2GB then do an MPI_Bcast
                * of the data from one process instead of having all the processes 
                * involved in the collective I/O.
                */
                if(dset_size > ((hsize_t)(2.0F * H5_GB) - 1))
                    local_cause[1] |= H5D_MPIO_RANK0_GREATER_THAN_2GB;
            } /* end else */
        } /* end else */
    } /* end else */
    
    /* Check for independent I/O */
    if(local_cause[0] & H5D_MPIO_SET_INDEPENDENT)
        global_cause[0] = local_cause[0];
    else {
        int mpi_code;               /* MPI error code */

        /* Form consensus opinion among all processes about whether to perform
         * collective I/O
         */
        if(MPI_SUCCESS != (mpi_code = MPI_Allreduce(&local_cause, &global_cause, 2, MPI_UNSIGNED, MPI_BOR, io_info->comm)))
            HMPI_GOTO_ERROR(FAIL, "MPI_Allreduce failed", mpi_code)
    } /* end else */

    /* Set the local & global values of no-collective-cause in the API context */
    H5CX_set_mpio_local_no_coll_cause(local_cause[0]);
    H5CX_set_mpio_global_no_coll_cause(global_cause[0]);

    /* Set read-with-rank0-and-bcast flag if possible */
    if(global_cause[0] == 0 && global_cause[1] == 0) {
        H5CX_set_mpio_rank0_bcast(TRUE);
#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
        H5CX_test_set_mpio_coll_rank0_bcast(TRUE);
#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */
    } /* end if */

    /* Set the return value, based on the global cause */
    ret_value = global_cause[0] > 0 ? FALSE : TRUE;

done:
    FUNC_LEAVE_NOAPI(ret_value)
} /* H5D__mpio_opt_possible() */


/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_select_read
 *
 * Purpose:     MPI-IO function to read directly from app buffer to file.
 *
 * Return:      non-negative on success, negative on failure.
 *
 * Programmer:
 *
 *-------------------------------------------------------------------------
 */
herr_t
H5D__mpio_select_read(const H5D_io_info_t *io_info, const H5D_type_info_t H5_ATTR_UNUSED *type_info,
    hsize_t mpi_buf_count, const H5S_t H5_ATTR_UNUSED *file_space, const H5S_t H5_ATTR_UNUSED *mem_space)
{
    const H5D_contig_storage_t *store_contig = &(io_info->store->contig);    /* Contiguous storage info for this I/O operation */
    herr_t ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    H5_CHECK_OVERFLOW(mpi_buf_count, hsize_t, size_t);
    if(H5F_block_read(io_info->dset->oloc.file, H5FD_MEM_DRAW, store_contig->dset_addr, (size_t)mpi_buf_count, io_info->u.rbuf) < 0)
        HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "can't finish collective parallel read")

done:
    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__mpio_select_read() */


/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_select_write
 *
 * Purpose:     MPI-IO function to write directly from app buffer to file.
 *
 * Return:      non-negative on success, negative on failure.
 *
 * Programmer:
 *
 *-------------------------------------------------------------------------
 */
herr_t
H5D__mpio_select_write(const H5D_io_info_t *io_info, const H5D_type_info_t H5_ATTR_UNUSED *type_info,
    hsize_t mpi_buf_count, const H5S_t H5_ATTR_UNUSED *file_space, const H5S_t H5_ATTR_UNUSED *mem_space)
{
    const H5D_contig_storage_t *store_contig = &(io_info->store->contig);    /* Contiguous storage info for this I/O operation */
    herr_t ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    /*OKAY: CAST DISCARDS CONST QUALIFIER*/
    H5_CHECK_OVERFLOW(mpi_buf_count, hsize_t, size_t);
    if(H5F_block_write(io_info->dset->oloc.file, H5FD_MEM_DRAW, store_contig->dset_addr, (size_t)mpi_buf_count, io_info->u.wbuf) < 0)
       HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "can't finish collective parallel write")

done:
    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__mpio_select_write() */


/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_array_gatherv
 *
 * Purpose:     Given an array, specified in local_array, by each processor
 *              calling this function, collects each array into a single
 *              array which is then either gathered to the processor
 *              specified by root, when allgather is false, or is
 *              distributed back to all processors when allgather is true.
 *
 *              The number of entries in the array contributed by an
 *              individual processor and the size of each entry should be
 *              specified in local_array_num_entries and array_entry_size,
 *              respectively.
 *
 *              The MPI communicator to use should be specified for comm.
 *
 *              If the sort_func argument is supplied, the array is sorted
 *              before the function returns.
 *
 *              Note: if allgather is specified as true, root is ignored.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Jordan Henderson
 *              Sunday, April 9th, 2017
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_array_gatherv(void *local_array, size_t local_array_num_entries,
    size_t array_entry_size, void **_gathered_array, size_t *_gathered_array_num_entries,
    hbool_t allgather, int root, MPI_Comm comm, int (*sort_func)(const void *, const void *))
{
    size_t  gathered_array_num_entries = 0; /* The size of the newly-constructed array */
    void   *gathered_array = NULL;          /* The newly-constructed array returned to the caller */
    int    *receive_counts_array = NULL;    /* Array containing number of entries each processor is contributing */
    int    *displacements_array = NULL;     /* Array of displacements where each processor places its data in the final array */
    int     mpi_code, mpi_rank, mpi_size;
    int     sendcount;
    herr_t  ret_value = SUCCEED;

    FUNC_ENTER_STATIC

    HDassert(_gathered_array);
    HDassert(_gathered_array_num_entries);

    MPI_Comm_size(comm, &mpi_size);
    MPI_Comm_rank(comm, &mpi_rank);

    /* Determine the size of the end result array by collecting the number
     * of entries contributed by each processor into a single total.
     */
    if (MPI_SUCCESS != (mpi_code = MPI_Allreduce(&local_array_num_entries, &gathered_array_num_entries, 1, MPI_INT, MPI_SUM, comm)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Allreduce failed", mpi_code)

    /* If 0 entries resulted from the collective operation, no processor is contributing anything and there is nothing to do */
    if (gathered_array_num_entries > 0) {
        /*
         * If gathering to all processors, all processors need to allocate space for the resulting array, as well as
         * the receive counts and displacements arrays for the collective MPI_Allgatherv call. Otherwise, only the
         * root processor needs to allocate the space for an MPI_Gatherv call.
         */
        if (allgather || (mpi_rank == root)) {
            if (NULL == (gathered_array = H5MM_malloc(gathered_array_num_entries * array_entry_size)))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate gathered array")

            if (NULL == (receive_counts_array = (int *) H5MM_malloc((size_t) mpi_size * sizeof(int))))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate receive counts array")

            if (NULL == (displacements_array = (int *) H5MM_malloc((size_t) mpi_size * sizeof(int))))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate receive displacements array")
        } /* end if */

        /*
         * If gathering to all processors, inform each processor of how many entries each other processor is
         * contributing to the resulting array by collecting the counts into each processor's "receive counts"
         * array. Otherwise, inform only the root processor of how many entries each other processor is contributing.
         */
        if (allgather) {
            if (MPI_SUCCESS != (mpi_code = MPI_Allgather(&local_array_num_entries, 1, MPI_INT, receive_counts_array, 1, MPI_INT, comm)))
                HMPI_GOTO_ERROR(FAIL, "MPI_Allgather failed", mpi_code)
        } /* end if */
        else {
            if (MPI_SUCCESS != (mpi_code = MPI_Gather(&local_array_num_entries, 1, MPI_INT, receive_counts_array, 1, MPI_INT, root, comm)))
                HMPI_GOTO_ERROR(FAIL, "MPI_Gather failed", mpi_code)
        } /* end else */

        if (allgather || (mpi_rank == root)) {
            size_t i;

            /* Multiply each receive count by the size of the array entry, since the data is sent as bytes. */
            for (i = 0; i < (size_t) mpi_size; i++)
                H5_CHECKED_ASSIGN(receive_counts_array[i], int, (size_t) receive_counts_array[i] * array_entry_size, size_t);

            /* Set receive buffer offsets for the collective MPI_Allgatherv/MPI_Gatherv call. */
            displacements_array[0] = 0;
            for (i = 1; i < (size_t) mpi_size; i++)
                displacements_array[i] = displacements_array[i - 1] + receive_counts_array[i - 1];
        } /* end if */

        /* As the data is sent as bytes, calculate the true sendcount for the data. */
        H5_CHECKED_ASSIGN(sendcount, int, local_array_num_entries * array_entry_size, size_t);

        if (allgather) {
            if (MPI_SUCCESS != (mpi_code = MPI_Allgatherv(local_array, sendcount, MPI_BYTE,
                    gathered_array, receive_counts_array, displacements_array, MPI_BYTE, comm)))
                HMPI_GOTO_ERROR(FAIL, "MPI_Allgatherv failed", mpi_code)
        } /* end if */
        else {
            if (MPI_SUCCESS != (mpi_code = MPI_Gatherv(local_array, sendcount, MPI_BYTE,
                    gathered_array, receive_counts_array, displacements_array, MPI_BYTE, root, comm)))
                HMPI_GOTO_ERROR(FAIL, "MPI_Gatherv failed", mpi_code)
        } /* end else */

        if (sort_func && (allgather || (mpi_rank == root)))
            HDqsort(gathered_array, gathered_array_num_entries, array_entry_size, sort_func);
    } /* end if */

    *_gathered_array = gathered_array;
    *_gathered_array_num_entries = gathered_array_num_entries;

done:
    if (receive_counts_array)
        H5MM_free(receive_counts_array);
    if (displacements_array)
        H5MM_free(displacements_array);

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__mpio_array_gatherv() */


/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_get_sum_chunk
 *
 * Purpose:     Routine for obtaining total number of chunks to cover
 *              hyperslab selection selected by all processors.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Muqun Yang
 *              Monday, Feb. 13th, 2006
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_get_sum_chunk(const H5D_io_info_t *io_info, const H5D_chunk_map_t *fm,
    int *sum_chunkf)
{
    int num_chunkf;             /* Number of chunks to iterate over */
    size_t ori_num_chunkf;
    int mpi_code;               /* MPI return code */
    herr_t ret_value = SUCCEED;

    FUNC_ENTER_STATIC

    /* Get the number of chunks to perform I/O on */
    num_chunkf = 0;
    ori_num_chunkf = H5SL_count(fm->sel_chunks);
    H5_CHECKED_ASSIGN(num_chunkf, int, ori_num_chunkf, size_t);

    /* Determine the summation of number of chunks for all processes */
    if(MPI_SUCCESS != (mpi_code = MPI_Allreduce(&num_chunkf, sum_chunkf, 1, MPI_INT, MPI_SUM, io_info->comm)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Allreduce failed", mpi_code)

done:
    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__mpio_get_sum_chunk() */


/*-------------------------------------------------------------------------
 * Function:    H5D__contig_collective_read
 *
 * Purpose:     Reads directly from contiguous data in file into application
 *              memory using collective I/O.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Quincey Koziol
 *              Tuesday, March  4, 2008
 *
 *-------------------------------------------------------------------------
 */
herr_t
H5D__contig_collective_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
    hsize_t H5_ATTR_UNUSED nelmts, const H5S_t *file_space, const H5S_t *mem_space,
    H5D_chunk_map_t H5_ATTR_UNUSED *fm)
{
    H5D_mpio_actual_io_mode_t actual_io_mode = H5D_MPIO_CONTIGUOUS_COLLECTIVE;
    herr_t ret_value = SUCCEED;         /* Return value */

    FUNC_ENTER_PACKAGE

    /* Sanity check */
    HDassert(H5FD_MPIO == H5F_DRIVER_ID(io_info->dset->oloc.file));

    /* Call generic internal collective I/O routine */
    if(H5D__inter_collective_io(io_info, type_info, file_space, mem_space) < 0)
        HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "couldn't finish shared collective MPI-IO")

    /* Set the actual I/O mode property. internal_collective_io will not break to
     * independent I/O, so we set it here.
     */
    H5CX_set_mpio_actual_io_mode(actual_io_mode);

done:
    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__contig_collective_read() */


/*-------------------------------------------------------------------------
 * Function:    H5D__contig_collective_write
 *
 * Purpose:     Write directly to contiguous data in file from application
 *              memory using collective I/O.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Quincey Koziol
 *              Tuesday, March  4, 2008
 *
 *-------------------------------------------------------------------------
 */
herr_t
H5D__contig_collective_write(H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
    hsize_t H5_ATTR_UNUSED nelmts, const H5S_t *file_space, const H5S_t *mem_space,
    H5D_chunk_map_t H5_ATTR_UNUSED *fm)
{
    H5D_mpio_actual_io_mode_t actual_io_mode = H5D_MPIO_CONTIGUOUS_COLLECTIVE;
    herr_t ret_value = SUCCEED;         /* Return value */

    FUNC_ENTER_PACKAGE

    /* Sanity check */
    HDassert(H5FD_MPIO == H5F_DRIVER_ID(io_info->dset->oloc.file));

    /* Call generic internal collective I/O routine */
    if(H5D__inter_collective_io(io_info, type_info, file_space, mem_space) < 0)
        HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "couldn't finish shared collective MPI-IO")

    /* Set the actual I/O mode property. internal_collective_io will not break to
     * independent I/O, so we set it here.
     */
    H5CX_set_mpio_actual_io_mode(actual_io_mode);

done:
    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__contig_collective_write() */


/*-------------------------------------------------------------------------
 * Function:    H5D__chunk_collective_io
 *
 * Purpose:     Routine for
 *              1) choose an IO option:
 *                    a) One collective IO defined by one MPI derived datatype to link through all chunks
 *              or    b) multiple chunk IOs,to do MPI-IO for each chunk, the IO mode may be adjusted
 *                       due to the selection pattern for each chunk.
 *              For option a)
 *                      1. Sort the chunk address, obtain chunk info according to the sorted chunk address
 *                      2. Build up MPI derived datatype for each chunk
 *                      3. Build up the final MPI derived datatype
 *                      4. Set up collective IO property list
 *                      5. Do IO
 *              For option b)
 *                      1. Use MPI_gather and MPI_Bcast to obtain information of *collective/independent/none*
 *                         IO mode for each chunk of the selection
 *                      2. Depending on whether the IO mode is collective or independent or none,
 *                         Create either MPI derived datatype for each chunk to do collective IO or
 *                         just do independent IO or independent IO with file set view
 *                      3. Set up collective IO property list for collective mode
 *                      4. DO IO
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Muqun Yang
 *              Monday, Feb. 13th, 2006
 *
 * Modification:
 *  - Refctore to remove multi-chunk-without-opimization feature and update for
 *    multi-chunk-io accordingly
 * Programmer: Jonathan Kim
 * Date: 2012-10-10
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
    H5D_chunk_map_t *fm)
{
    H5FD_mpio_chunk_opt_t chunk_opt_mode;
    int         io_option = H5D_MULTI_CHUNK_IO_MORE_OPT;
    int         sum_chunk = -1;
#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
    htri_t      temp_not_link_io = FALSE;
#endif
    herr_t      ret_value = SUCCEED;

    FUNC_ENTER_STATIC

    /* Sanity checks */
    HDassert(io_info);
    HDassert(io_info->using_mpi_vfd);
    HDassert(type_info);
    HDassert(fm);

    /* Disable collective metadata reads for chunked dataset I/O operations
     * in order to prevent potential hangs */
    H5CX_set_coll_metadata_read(FALSE);

    /* Check the optional property list for the collective chunk IO optimization option */
    if(H5CX_get_mpio_chunk_opt_mode(&chunk_opt_mode) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "couldn't get chunk optimization option")

    if(H5FD_MPIO_CHUNK_ONE_IO == chunk_opt_mode)
        io_option = H5D_ONE_LINK_CHUNK_IO;      /*no opt*/
    /* direct request to multi-chunk-io */
    else if(H5FD_MPIO_CHUNK_MULTI_IO == chunk_opt_mode)
        io_option = H5D_MULTI_CHUNK_IO;         
    /* via default path. branch by num threshold */
    else {
        unsigned one_link_chunk_io_threshold;   /* Threshold to use single collective I/O for all chunks */
        int mpi_size;                   /* Number of processes in MPI job */

        if(H5D__mpio_get_sum_chunk(io_info, fm, &sum_chunk) < 0)
            HGOTO_ERROR(H5E_DATASPACE, H5E_CANTSWAP, FAIL, "unable to obtain the total chunk number of all processes");
        if((mpi_size = H5F_mpi_get_size(io_info->dset->oloc.file)) < 0)
            HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi size")

        /* Get the chunk optimization option threshold */
        if(H5CX_get_mpio_chunk_opt_num(&one_link_chunk_io_threshold) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "couldn't get chunk optimization option threshold value")

        /* step 1: choose an IO option */
        /* If the average number of chunk per process is greater than a threshold, we will do one link chunked IO. */
        if((unsigned)sum_chunk / (unsigned)mpi_size >= one_link_chunk_io_threshold)
            io_option = H5D_ONE_LINK_CHUNK_IO_MORE_OPT;
#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
        else
            temp_not_link_io = TRUE;
#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */
    } /* end else */

#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
{
    /*** Set collective chunk user-input optimization APIs. ***/
    if(H5D_ONE_LINK_CHUNK_IO == io_option) {
        if(H5CX_test_set_mpio_coll_chunk_link_hard(0) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTSET, FAIL, "unable to set property value")
    } /* end if */
    else if(H5D_MULTI_CHUNK_IO == io_option) {
        if(H5CX_test_set_mpio_coll_chunk_multi_hard(0) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTSET, FAIL, "unable to set property value")
    } /* end else-if */
    else if(H5D_ONE_LINK_CHUNK_IO_MORE_OPT == io_option) {
        if(H5CX_test_set_mpio_coll_chunk_link_num_true(0) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTSET, FAIL, "unable to set property value")
    } /* end if */
    else if(temp_not_link_io) {
        if(H5CX_test_set_mpio_coll_chunk_link_num_false(0) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTSET, FAIL, "unable to set property value")
    } /* end if */
}
#endif /* H5_HAVE_INSTRUMENTED_LIBRARY */

    /* step 2:  Go ahead to do IO.*/
    switch (io_option) {
        case H5D_ONE_LINK_CHUNK_IO:
        case H5D_ONE_LINK_CHUNK_IO_MORE_OPT:
            /* Check if there are any filters in the pipeline */
            if(io_info->dset->shared->dcpl_cache.pline.nused > 0) {
                /* For now, Multi-chunk IO must be forced for parallel filtered read,
                 * so that data can be unfiltered as it is received. There is significant
                 * complexity in unfiltering the data when it is read all at once into a
                 * single buffer.
                 */
                if(io_info->op_type == H5D_IO_OP_READ) {
                    if(H5D__multi_chunk_filtered_collective_io(io_info, type_info, fm) < 0)
                        HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish optimized multiple filtered chunk MPI-IO")
                } /* end if */
                else
                    if(H5D__link_chunk_filtered_collective_io(io_info, type_info, fm) < 0)
                        HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish filtered linked chunk MPI-IO")
            } /* end if */
            else
                /* Perform unfiltered link chunk collective IO */
                if(H5D__link_chunk_collective_io(io_info, type_info, fm, sum_chunk) < 0)
                    HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish linked chunk MPI-IO")
            break;

        case H5D_MULTI_CHUNK_IO: /* direct request to do multi-chunk IO */
        default:                 /* multiple chunk IO via threshold */
            /* Check if there are any filters in the pipeline */
            if(io_info->dset->shared->dcpl_cache.pline.nused > 0) {
                if(H5D__multi_chunk_filtered_collective_io(io_info, type_info, fm) < 0)
                    HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish optimized multiple filtered chunk MPI-IO")
            } /* end if */
            else
                /* Perform unfiltered multi chunk collective IO */
                if(H5D__multi_chunk_collective_io(io_info, type_info, fm) < 0)
                    HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish optimized multiple chunk MPI-IO")
            break;
    } /* end switch */

done:
    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__chunk_collective_io */


/*-------------------------------------------------------------------------
 * Function:    H5D__chunk_collective_read
 *
 * Purpose:     Reads directly from chunks in file into application memory
 *              using collective I/O.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Quincey Koziol
 *              Tuesday, March  4, 2008
 *
 *-------------------------------------------------------------------------
 */
herr_t
H5D__chunk_collective_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
    hsize_t H5_ATTR_UNUSED nelmts, const H5S_t H5_ATTR_UNUSED *file_space, const H5S_t H5_ATTR_UNUSED *mem_space,
    H5D_chunk_map_t *fm)
{
    herr_t ret_value = SUCCEED;         /* Return value */

    FUNC_ENTER_PACKAGE

    /* Call generic selection operation */
    if(H5D__chunk_collective_io(io_info, type_info, fm) < 0)
        HGOTO_ERROR(H5E_DATASPACE, H5E_READERROR, FAIL, "read error")

done:
    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__chunk_collective_read() */


/*-------------------------------------------------------------------------
 * Function:    H5D__chunk_collective_write
 *
 * Purpose:     Write directly to chunks in file from application memory
 *              using collective I/O.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Quincey Koziol
 *              Tuesday, March  4, 2008
 *
 *-------------------------------------------------------------------------
 */
herr_t
H5D__chunk_collective_write(H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
    hsize_t H5_ATTR_UNUSED nelmts, const H5S_t H5_ATTR_UNUSED *file_space, const H5S_t H5_ATTR_UNUSED *mem_space,
    H5D_chunk_map_t *fm)
{
    herr_t ret_value = SUCCEED;         /* Return value */

    FUNC_ENTER_PACKAGE

    /* Call generic selection operation */
    if(H5D__chunk_collective_io(io_info, type_info, fm) < 0)
        HGOTO_ERROR(H5E_DATASPACE, H5E_WRITEERROR, FAIL, "write error")

done:
    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__chunk_collective_write() */


/*-------------------------------------------------------------------------
 * Function:    H5D__link_chunk_collective_io
 *
 * Purpose:     Routine for one collective IO with one MPI derived datatype to link with all chunks
 *
 *                      1. Sort the chunk address and chunk info
 *                      2. Build up MPI derived datatype for each chunk
 *                      3. Build up the final MPI derived datatype
 *                      4. Use common collective IO routine to do MPI-IO
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Muqun Yang
 *              Monday, Feb. 13th, 2006
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__link_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
    H5D_chunk_map_t *fm, int sum_chunk)
{
    H5D_chunk_addr_info_t *chunk_addr_info_array = NULL;
    MPI_Datatype chunk_final_mtype;         /* Final memory MPI datatype for all chunks with seletion */
    hbool_t chunk_final_mtype_is_derived = FALSE;
    MPI_Datatype chunk_final_ftype;         /* Final file MPI datatype for all chunks with seletion */
    hbool_t chunk_final_ftype_is_derived = FALSE;
    H5D_storage_t ctg_store;                /* Storage info for "fake" contiguous dataset */
    size_t              total_chunks;
    MPI_Datatype       *chunk_mtype = NULL;
    MPI_Datatype       *chunk_ftype = NULL;
    MPI_Aint           *chunk_disp_array = NULL;
    MPI_Aint           *chunk_mem_disp_array = NULL;
    hbool_t            *chunk_mft_is_derived_array = NULL;      /* Flags to indicate each chunk's MPI file datatype is derived */
    hbool_t            *chunk_mbt_is_derived_array = NULL;      /* Flags to indicate each chunk's MPI memory datatype is derived */
    int                *chunk_mpi_file_counts = NULL;   /* Count of MPI file datatype for each chunk */
    int                *chunk_mpi_mem_counts = NULL;    /* Count of MPI memory datatype for each chunk */
    int                 mpi_code;           /* MPI return code */
    herr_t              ret_value = SUCCEED;

    FUNC_ENTER_STATIC

    /* Set the actual-chunk-opt-mode property. */
    H5CX_set_mpio_actual_chunk_opt(H5D_MPIO_LINK_CHUNK);

    /* Set the actual-io-mode property.
     * Link chunk I/O does not break to independent, so can set right away */
    H5CX_set_mpio_actual_io_mode(H5D_MPIO_CHUNK_COLLECTIVE);

    /* Get the sum # of chunks, if not already available */
    if(sum_chunk < 0) {
        if(H5D__mpio_get_sum_chunk(io_info, fm, &sum_chunk) < 0)
            HGOTO_ERROR(H5E_DATASPACE, H5E_CANTSWAP, FAIL, "unable to obtain the total chunk number of all processes");
    } /* end if */

    /* Retrieve total # of chunks in dataset */
    H5_CHECKED_ASSIGN(total_chunks, size_t, fm->layout->u.chunk.nchunks, hsize_t);

    /* Handle special case when dataspace dimensions only allow one chunk in
     *  the dataset.  [This sometimes is used by developers who want the
     *  equivalent of compressed contiguous datasets - QAK]
     */
    if(total_chunks == 1) {
        H5SL_node_t *chunk_node;        /* Pointer to chunk node for selection */
        H5S_t *fspace;                  /* Dataspace describing chunk & selection in it */
        H5S_t *mspace;                  /* Dataspace describing selection in memory corresponding to this chunk */

        /* Check for this process having selection in this chunk */
        chunk_node = H5SL_first(fm->sel_chunks);

        if(chunk_node == NULL) {
            /* Set the dataspace info for I/O to NULL, this process doesn't have any I/O to perform */
            fspace = mspace = NULL;

            /* Initialize chunk address */
            ctg_store.contig.dset_addr = 0;
        } /* end if */
        else {
            H5D_chunk_ud_t udata;           /* User data for querying chunk info */
            H5D_chunk_info_t *chunk_info;   /* Info for chunk in skiplist */

            /* Get the chunk info, for the selection in the chunk */
            if(NULL == (chunk_info = (H5D_chunk_info_t *)H5SL_item(chunk_node)))
                HGOTO_ERROR(H5E_STORAGE, H5E_CANTGET, FAIL, "couldn't get chunk info from skip list")

            /* Set the dataspace info for I/O */
            fspace = chunk_info->fspace;
            mspace = chunk_info->mspace;

            /* Look up address of chunk */
            if(H5D__chunk_lookup(io_info->dset, chunk_info->scaled, &udata) < 0)
                HGOTO_ERROR(H5E_STORAGE, H5E_CANTGET, FAIL, "couldn't get chunk address")
            ctg_store.contig.dset_addr = udata.chunk_block.offset;
        } /* end else */

        /* Set up the base storage address for this chunk */
        io_info->store = &ctg_store;

#ifdef H5D_DEBUG
if(H5DEBUG(D))
    HDfprintf(H5DEBUG(D),"before inter_collective_io for total chunk = 1 \n");
#endif

        /* Perform I/O */
        if(H5D__inter_collective_io(io_info, type_info, fspace, mspace) < 0)
            HGOTO_ERROR(H5E_STORAGE, H5E_CANTGET, FAIL, "couldn't finish shared collective MPI-IO")
    } /* end if */
    else {
        hsize_t mpi_buf_count;  /* Number of MPI types */
        size_t num_chunk;       /* Number of chunks for this process */
        size_t u;               /* Local index variable */

        /* Get the number of chunks with a selection */
        num_chunk = H5SL_count(fm->sel_chunks);
        H5_CHECK_OVERFLOW(num_chunk, size_t, int);

#ifdef H5D_DEBUG
if(H5DEBUG(D))
    HDfprintf(H5DEBUG(D),"total_chunks = %Zu, num_chunk = %Zu\n", total_chunks, num_chunk);
#endif

        /* Set up MPI datatype for chunks selected */
        if(num_chunk) {
            /* Allocate chunking information */
            if(NULL == (chunk_addr_info_array = (H5D_chunk_addr_info_t *)H5MM_malloc(num_chunk * sizeof(H5D_chunk_addr_info_t))))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk array buffer")
            if(NULL == (chunk_mtype           = (MPI_Datatype *)H5MM_malloc(num_chunk * sizeof(MPI_Datatype))))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk memory datatype buffer")
            if(NULL == (chunk_ftype           = (MPI_Datatype *)H5MM_malloc(num_chunk * sizeof(MPI_Datatype))))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk file datatype buffer")
            if(NULL == (chunk_disp_array      = (MPI_Aint *)H5MM_malloc(num_chunk * sizeof(MPI_Aint))))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk file displacement buffer")
            if(NULL == (chunk_mem_disp_array  = (MPI_Aint *)H5MM_calloc(num_chunk * sizeof(MPI_Aint))))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk memory displacement buffer")
            if(NULL == (chunk_mpi_mem_counts        = (int *)H5MM_calloc(num_chunk * sizeof(int))))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk memory counts buffer")
            if(NULL == (chunk_mpi_file_counts       = (int *)H5MM_calloc(num_chunk * sizeof(int))))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk file counts buffer")
            if(NULL == (chunk_mbt_is_derived_array  = (hbool_t *)H5MM_calloc(num_chunk * sizeof(hbool_t))))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk memory is derived datatype flags buffer")
            if(NULL == (chunk_mft_is_derived_array  = (hbool_t *)H5MM_calloc(num_chunk * sizeof(hbool_t))))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk file is derived datatype flags buffer")

#ifdef H5D_DEBUG
if(H5DEBUG(D))
    HDfprintf(H5DEBUG(D),"before sorting the chunk address \n");
#endif
            /* Sort the chunk address */
            if(H5D__sort_chunk(io_info, fm, chunk_addr_info_array, sum_chunk) < 0)
                HGOTO_ERROR(H5E_DATASPACE, H5E_CANTSWAP, FAIL, "unable to sort chunk address")
            ctg_store.contig.dset_addr = chunk_addr_info_array[0].chunk_addr;

#ifdef H5D_DEBUG
if(H5DEBUG(D))
    HDfprintf(H5DEBUG(D),"after sorting the chunk address \n");
#endif

            /* Obtain MPI derived datatype from all individual chunks */
            for(u = 0; u < num_chunk; u++) {
                hsize_t *permute_map = NULL; /* array that holds the mapping from the old, 
                                                out-of-order displacements to the in-order 
                                                displacements of the MPI datatypes of the 
                                                point selection of the file space */
                hbool_t is_permuted = FALSE;

                /* Obtain disk and memory MPI derived datatype */
                /* NOTE: The permute_map array can be allocated within H5S_mpio_space_type
                 *              and will be fed into the next call to H5S_mpio_space_type
                 *              where it will be freed.
                 */
                if(H5S_mpio_space_type(chunk_addr_info_array[u].chunk_info.fspace,
                                       type_info->src_type_size, 
                                       &chunk_ftype[u], /* OUT: datatype created */ 
                                       &chunk_mpi_file_counts[u], /* OUT */
                                       &(chunk_mft_is_derived_array[u]), /* OUT */
                                       TRUE, /* this is a file space,
                                                so permute the
                                                datatype if the point
                                                selections are out of
                                                order */
                                       &permute_map,/* OUT: a map to indicate the
                                                       permutation of points
                                                       selected in case they
                                                       are out of order */
                                       &is_permuted /* OUT */) < 0)
                    HGOTO_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL, "couldn't create MPI file type")
                /* Sanity check */
                if(is_permuted)
                    HDassert(permute_map);
                if(H5S_mpio_space_type(chunk_addr_info_array[u].chunk_info.mspace,
                                       type_info->dst_type_size, &chunk_mtype[u], 
                                       &chunk_mpi_mem_counts[u], 
                                       &(chunk_mbt_is_derived_array[u]), 
                                       FALSE, /* this is a memory
                                                 space, so if the file
                                                 space is not
                                                 permuted, there is no
                                                 need to permute the
                                                 datatype if the point
                                                 selections are out of
                                                 order*/
                                       &permute_map, /* IN: the permutation map
                                                        generated by the
                                                        file_space selection
                                                        and applied to the
                                                        memory selection */
                                       &is_permuted /* IN */) < 0)
                    HGOTO_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL, "couldn't create MPI buf type")
                /* Sanity check */
                if(is_permuted)
                    HDassert(!permute_map);

                /* Chunk address relative to the first chunk */
                chunk_addr_info_array[u].chunk_addr -= ctg_store.contig.dset_addr;

                /* Assign chunk address to MPI displacement */
                /* (assume MPI_Aint big enough to hold it) */
                chunk_disp_array[u] = (MPI_Aint)chunk_addr_info_array[u].chunk_addr;
            } /* end for */

            /* Create final MPI derived datatype for the file */
            if(MPI_SUCCESS != (mpi_code = MPI_Type_create_struct((int)num_chunk, chunk_mpi_file_counts, chunk_disp_array, chunk_ftype, &chunk_final_ftype)))
                HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_struct failed", mpi_code)
            if(MPI_SUCCESS != (mpi_code = MPI_Type_commit(&chunk_final_ftype)))
                HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code)
            chunk_final_ftype_is_derived = TRUE;

            /* Create final MPI derived datatype for memory */
            if(MPI_SUCCESS != (mpi_code = MPI_Type_create_struct((int)num_chunk, chunk_mpi_mem_counts, chunk_mem_disp_array, chunk_mtype, &chunk_final_mtype)))
                HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_struct failed", mpi_code)
            if(MPI_SUCCESS != (mpi_code = MPI_Type_commit(&chunk_final_mtype)))
                HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code)
            chunk_final_mtype_is_derived = TRUE;

            /* Free the file & memory MPI datatypes for each chunk */
            for(u = 0; u < num_chunk; u++) {
                if(chunk_mbt_is_derived_array[u])
                    if(MPI_SUCCESS != (mpi_code = MPI_Type_free(chunk_mtype + u)))
                        HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)

                if(chunk_mft_is_derived_array[u])
                    if(MPI_SUCCESS != (mpi_code = MPI_Type_free(chunk_ftype + u)))
                        HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
            } /* end for */

            /* We have a single, complicated MPI datatype for both memory & file */
            mpi_buf_count  = (hsize_t)1;
        } /* end if */
        else {      /* no selection at all for this process */
            ctg_store.contig.dset_addr = 0;

            /* Set the MPI datatype */
            chunk_final_ftype = MPI_BYTE;
            chunk_final_mtype = MPI_BYTE;

            /* No chunks selected for this process */
            mpi_buf_count  = (hsize_t)0;
        } /* end else */
#ifdef H5D_DEBUG
if(H5DEBUG(D))
    HDfprintf(H5DEBUG(D),"before coming to final collective IO\n");
#endif

        /* Set up the base storage address for this chunk */
        io_info->store = &ctg_store;

        /* Perform final collective I/O operation */
        if(H5D__final_collective_io(io_info, type_info, mpi_buf_count, chunk_final_ftype, chunk_final_mtype) < 0)
            HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish MPI-IO")
    } /* end else */

done:
#ifdef H5D_DEBUG
if(H5DEBUG(D))
    HDfprintf(H5DEBUG(D),"before freeing memory inside H5D_link_collective_io ret_value = %d\n", ret_value);
#endif
    /* Release resources */
    if(chunk_addr_info_array)
        H5MM_xfree(chunk_addr_info_array);
    if(chunk_mtype)
        H5MM_xfree(chunk_mtype);
    if(chunk_ftype)
        H5MM_xfree(chunk_ftype);
    if(chunk_disp_array)
        H5MM_xfree(chunk_disp_array);
    if(chunk_mem_disp_array)
        H5MM_xfree(chunk_mem_disp_array);
    if(chunk_mpi_mem_counts)
        H5MM_xfree(chunk_mpi_mem_counts);
    if(chunk_mpi_file_counts)
        H5MM_xfree(chunk_mpi_file_counts);
    if(chunk_mbt_is_derived_array)
        H5MM_xfree(chunk_mbt_is_derived_array);
    if(chunk_mft_is_derived_array)
        H5MM_xfree(chunk_mft_is_derived_array);

    /* Free the MPI buf and file types, if they were derived */
    if(chunk_final_mtype_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&chunk_final_mtype)))
        HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
    if(chunk_final_ftype_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&chunk_final_ftype)))
        HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__link_chunk_collective_io */


/*-------------------------------------------------------------------------
 * Function:    H5D__link_chunk_filtered_collective_io
 *
 * Purpose:     Routine for one collective IO with one MPI derived datatype
 *              to link with all filtered chunks
 *
 *              1. Construct a list of selected chunks in the collective IO
 *                 operation
 *                 A. If any chunk is being written to by more than 1
 *                    process, the process writing to the chunk which
 *                    currently has the least amount of chunks assigned
 *                    to it becomes the new owner (in the case of ties,
 *                    the lowest MPI rank becomes the new owner)
 *              2. If the operation is a write operation
 *                 A. Loop through each chunk in the operation
 *                    I. If this is not a full overwrite of the chunk
 *                       a) Read the chunk from file and pass the chunk
 *                          through the filter pipeline in reverse order
 *                          (Unfilter the chunk)
 *                    II. Update the chunk data with the modifications from
 *                        the owning process
 *                    III. Receive any modification data from other
 *                         processes and update the chunk data with these
 *                         modifications
 *                    IV. Filter the chunk
 *                 B. Contribute the modified chunks to an array gathered
 *                    by all processes which contains the new sizes of
 *                    every chunk modified in the collective IO operation
 *                 C. All processes collectively re-allocate each chunk
 *                    from the gathered array with their new sizes after
 *                    the filter operation
 *                 D. If this process has any chunks selected in the IO
 *                    operation, create an MPI derived type for memory and
 *                    file to write out the process' selected chunks to the
 *                    file
 *                 E. Perform the collective write
 *                 F. All processes collectively re-insert each modified
 *                    chunk from the gathered array into the chunk index
 *
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Jordan Henderson
 *              Friday, Nov. 4th, 2016
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__link_chunk_filtered_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
    H5D_chunk_map_t *fm)
{
    H5D_filtered_collective_io_info_t *chunk_list = NULL; /* The list of chunks being read/written */
    H5D_filtered_collective_io_info_t *collective_chunk_list = NULL; /* The list of chunks used during collective operations */
    H5D_storage_t                      ctg_store;                        /* Chunk storage information as contiguous dataset */
    MPI_Datatype                       mem_type = MPI_BYTE;
    MPI_Datatype                       file_type = MPI_BYTE;
    hbool_t                            mem_type_is_derived = FALSE;
    hbool_t                            file_type_is_derived = FALSE;
    size_t                             chunk_list_num_entries;
    size_t                             collective_chunk_list_num_entries;
    size_t                            *num_chunks_selected_array = NULL; /* Array of number of chunks selected on each process */
    size_t                             i;                                /* Local index variable */
    int                                mpi_rank, mpi_size, mpi_code;
    herr_t                             ret_value = SUCCEED;

    FUNC_ENTER_STATIC

    HDassert(io_info);
    HDassert(type_info);
    HDassert(fm);

    /* Obtain the current rank of the process and the number of processes */
    if ((mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file)) < 0)
        HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi rank")
    if ((mpi_size = H5F_mpi_get_size(io_info->dset->oloc.file)) < 0)
        HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi size")

    /* Set the actual-chunk-opt-mode property. */
    H5CX_set_mpio_actual_chunk_opt(H5D_MPIO_LINK_CHUNK);

    /* Set the actual-io-mode property.
     * Link chunk filtered I/O does not break to independent, so can set right away
     */
    H5CX_set_mpio_actual_io_mode(H5D_MPIO_CHUNK_COLLECTIVE);

    /* Build a list of selected chunks in the collective io operation */
    if (H5D__construct_filtered_io_info_list(io_info, type_info, fm, &chunk_list, &chunk_list_num_entries) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "couldn't construct filtered I/O info list")

    if (io_info->op_type == H5D_IO_OP_WRITE) { /* Filtered collective write */
        H5D_chk_idx_info_t index_info;
        H5D_chunk_ud_t     udata;
        hsize_t            mpi_buf_count;

        /* Construct chunked index info */
        index_info.f = io_info->dset->oloc.file;
        index_info.pline = &(io_info->dset->shared->dcpl_cache.pline);
        index_info.layout = &(io_info->dset->shared->layout.u.chunk);
        index_info.storage = &(io_info->dset->shared->layout.storage.u.chunk);

        /* Set up chunk information for insertion to chunk index */
        udata.common.layout = index_info.layout;
        udata.common.storage = index_info.storage;
        udata.filter_mask = 0;

        /* Iterate through all the chunks in the collective write operation,
         * updating each chunk with the data modifications from other processes,
         * then re-filtering the chunk.
         */
        for (i = 0; i < chunk_list_num_entries; i++)
            if (mpi_rank == chunk_list[i].owners.new_owner)
                if (H5D__filtered_collective_chunk_entry_io(&chunk_list[i], io_info, type_info, fm) < 0)
                    HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "couldn't process chunk entry")

        /* Gather the new chunk sizes to all processes for a collective reallocation
         * of the chunks in the file.
         */
        if (H5D__mpio_array_gatherv(chunk_list, chunk_list_num_entries, sizeof(H5D_filtered_collective_io_info_t),
                (void **) &collective_chunk_list, &collective_chunk_list_num_entries, true, 0, io_info->comm, NULL) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTGATHER, FAIL, "couldn't gather new chunk sizes")

        /* Collectively re-allocate the modified chunks (from each process) in the file */
        for (i = 0; i < collective_chunk_list_num_entries; i++) {
            hbool_t insert;

            if (H5D__chunk_file_alloc(&index_info, &collective_chunk_list[i].chunk_states.chunk_current,
                    &collective_chunk_list[i].chunk_states.new_chunk, &insert, collective_chunk_list[i].scaled) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate chunk")
        } /* end for */

        if (NULL == (num_chunks_selected_array = (size_t *) H5MM_malloc((size_t) mpi_size * sizeof(size_t))))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate num chunks selected array")

        if (MPI_SUCCESS != (mpi_code = MPI_Allgather(&chunk_list_num_entries, 1, MPI_UNSIGNED_LONG_LONG, num_chunks_selected_array,
                1, MPI_UNSIGNED_LONG_LONG, io_info->comm)))
            HMPI_GOTO_ERROR(FAIL, "MPI_Allgather failed", mpi_code)

        /* If this process has any chunks selected, create a MPI type for collectively
         * writing out the chunks to file. Otherwise, the process contributes to the
         * collective write with a none type.
         */
        if (chunk_list_num_entries) {
            size_t offset;

            /* During the collective re-allocation of chunks in the file, the record for each
             * chunk is only updated in the collective array, not in the local copy of chunks on each
             * process. However, each process needs the updated chunk records so that they can create
             * a MPI type for the collective write that will write to the chunk's possible new locations
             * in the file instead of the old ones. This ugly hack seems to be the best solution to
             * copy the information back to the local array and avoid having to modify the collective
             * write type function in an ugly way so that it will accept the collective array instead
             * of the local array. This works correctly because the array gather function guarantees
             * that the chunk data in the collective array is ordered in blocks by rank.
             */
            for (i = 0, offset = 0; i < (size_t) mpi_rank; i++)
                offset += num_chunks_selected_array[i];

            HDmemcpy(chunk_list, &collective_chunk_list[offset], num_chunks_selected_array[mpi_rank] * sizeof(H5D_filtered_collective_io_info_t));

            /* Create single MPI type encompassing each selection in the dataspace */
            if (H5D__mpio_filtered_collective_write_type(chunk_list, chunk_list_num_entries,
                    &mem_type, &mem_type_is_derived, &file_type, &file_type_is_derived) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_BADTYPE, FAIL, "couldn't create MPI link chunk I/O type")

            /* Override the write buffer to point to the address of the first
             * chunk data buffer
             */
            io_info->u.wbuf = chunk_list[0].buf;
        } /* end if */

        /* We have a single, complicated MPI datatype for both memory & file */
        mpi_buf_count = (mem_type_is_derived && file_type_is_derived) ? (hsize_t) 1 : (hsize_t) 0;

        /* Set up the base storage address for this operation */
        ctg_store.contig.dset_addr = 0; /* Write address must be set to address 0 */
        io_info->store = &ctg_store;

        /* Perform I/O */
        if (H5D__final_collective_io(io_info, type_info, mpi_buf_count, file_type, mem_type) < 0)
            HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish MPI-IO")

        /* Participate in the collective re-insertion of all chunks modified
         * in this iteration into the chunk index
         */
        for (i = 0; i < collective_chunk_list_num_entries; i++) {
            udata.chunk_block = collective_chunk_list[i].chunk_states.new_chunk;
            udata.common.scaled = collective_chunk_list[i].scaled;
            udata.chunk_idx = collective_chunk_list[i].index;

            if ((index_info.storage->ops->insert)(&index_info, &udata, io_info->dset) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTINSERT, FAIL, "unable to insert chunk address into index")
        } /* end for */
    } /* end if */

done:
    /* Free resources used by a process which had some selection */
    if (chunk_list) {
        for (i = 0; i < chunk_list_num_entries; i++)
            if (chunk_list[i].buf)
                H5MM_free(chunk_list[i].buf);

        H5MM_free(chunk_list);
    } /* end if */

    if (num_chunks_selected_array)
        H5MM_free(num_chunks_selected_array);
    if (collective_chunk_list)
        H5MM_free(collective_chunk_list);

    /* Free the MPI buf and file types, if they were derived */
    if (mem_type_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&mem_type)))
        HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
    if (file_type_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&file_type)))
        HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__link_chunk_filtered_collective_io() */


/*-------------------------------------------------------------------------
 * Function:    H5D__multi_chunk_collective_io
 *
 * Purpose:     To do IO per chunk according to IO mode(collective/independent/none)
 *
 *              1. Use MPI_gather and MPI_Bcast to obtain IO mode in each chunk(collective/independent/none)
 *              2. Depending on whether the IO mode is collective or independent or none,
 *                 Create either MPI derived datatype for each chunk or just do independent IO
 *              3. Use common collective IO routine to do MPI-IO
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Muqun Yang
 *              Monday, Feb. 13th, 2006
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__multi_chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
    H5D_chunk_map_t *fm)
{
    H5D_io_info_t       ctg_io_info;          /* Contiguous I/O info object */
    H5D_storage_t       ctg_store;            /* Chunk storage information as contiguous dataset */
    H5D_io_info_t       cpt_io_info;          /* Compact I/O info object */
    H5D_storage_t       cpt_store;            /* Chunk storage information as compact dataset */
    hbool_t             cpt_dirty;            /* Temporary placeholder for compact storage "dirty" flag */
    uint8_t            *chunk_io_option = NULL;
    haddr_t            *chunk_addr = NULL;
    H5D_storage_t       store;                /* union of EFL and chunk pointer in file space */
    H5FD_mpio_collective_opt_t last_coll_opt_mode = H5FD_MPIO_COLLECTIVE_IO; /* Last parallel transfer with independent IO or collective IO with this mode */
    size_t              total_chunk;          /* Total # of chunks in dataset */
#ifdef H5Dmpio_DEBUG
    int mpi_rank;
#endif
    size_t              u;                    /* Local index variable */
    H5D_mpio_actual_io_mode_t actual_io_mode = H5D_MPIO_NO_COLLECTIVE; /* Local variable for tracking the I/O mode used. */
    herr_t              ret_value = SUCCEED;

    FUNC_ENTER_STATIC

    /* Set the actual chunk opt mode property */
    H5CX_set_mpio_actual_chunk_opt(H5D_MPIO_MULTI_CHUNK);

#ifdef H5Dmpio_DEBUG
    mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file);
#endif

    /* Retrieve total # of chunks in dataset */
    H5_CHECKED_ASSIGN(total_chunk, size_t, fm->layout->u.chunk.nchunks, hsize_t);
    HDassert(total_chunk != 0);

    /* Allocate memories */
    chunk_io_option = (uint8_t *)H5MM_calloc(total_chunk);
    chunk_addr      = (haddr_t *)H5MM_calloc(total_chunk * sizeof(haddr_t));
#ifdef H5D_DEBUG
if(H5DEBUG(D))
    HDfprintf(H5DEBUG(D), "total_chunk %Zu\n", total_chunk);
#endif

    /* Obtain IO option for each chunk */
    if(H5D__obtain_mpio_mode(io_info, fm, chunk_io_option, chunk_addr) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTRECV, FAIL, "unable to obtain MPIO mode")

    /* Set up contiguous I/O info object */
    HDmemcpy(&ctg_io_info, io_info, sizeof(ctg_io_info));
    ctg_io_info.store = &ctg_store;
    ctg_io_info.layout_ops = *H5D_LOPS_CONTIG;

    /* Initialize temporary contiguous storage info */
    ctg_store.contig.dset_size = (hsize_t)io_info->dset->shared->layout.u.chunk.size;

    /* Set up compact I/O info object */
    HDmemcpy(&cpt_io_info, io_info, sizeof(cpt_io_info));
    cpt_io_info.store = &cpt_store;
    cpt_io_info.layout_ops = *H5D_LOPS_COMPACT;

    /* Initialize temporary compact storage info */
    cpt_store.compact.dirty = &cpt_dirty;

    /* Set dataset storage for I/O info */
    io_info->store = &store;

    /* Loop over _all_ the chunks */
    for(u = 0; u < total_chunk; u++) {
        H5D_chunk_info_t *chunk_info;    /* Chunk info for current chunk */
        H5S_t *fspace;              /* Dataspace describing chunk & selection in it */
        H5S_t *mspace;              /* Dataspace describing selection in memory corresponding to this chunk */

#ifdef H5D_DEBUG
if(H5DEBUG(D))
    HDfprintf(H5DEBUG(D),"mpi_rank = %d, chunk index = %Zu\n", mpi_rank, u);
#endif
        /* Get the chunk info for this chunk, if there are elements selected */
        chunk_info = fm->select_chunk[u];

        /* Set the storage information for chunks with selections */
        if(chunk_info) {
            HDassert(chunk_info->index == u);

            /* Pass in chunk's coordinates in a union. */
            store.chunk.scaled  = chunk_info->scaled;
        } /* end if */

        /* Collective IO for this chunk,
         * Note: even there is no selection for this process, the process still
         *      needs to contribute MPI NONE TYPE.
         */
        if(chunk_io_option[u] == H5D_CHUNK_IO_MODE_COL) {
#ifdef H5D_DEBUG
if(H5DEBUG(D))
    HDfprintf(H5DEBUG(D),"inside collective chunk IO mpi_rank = %d, chunk index = %Zu\n", mpi_rank, u);
#endif

            /* Set the file & memory dataspaces */
            if(chunk_info) {
                fspace = chunk_info->fspace;
                mspace = chunk_info->mspace;

                /* Update the local variable tracking the actual io mode property.
                 *
                 * Note: H5D_MPIO_COLLECTIVE_MULTI | H5D_MPIO_INDEPENDENT = H5D_MPIO_MIXED
                 *      to ease switching between to mixed I/O without checking the current
                 *      value of the property. You can see the definition in H5Ppublic.h
                 */
                actual_io_mode = (H5D_mpio_actual_io_mode_t) (actual_io_mode | H5D_MPIO_CHUNK_COLLECTIVE);

            } /* end if */
            else {
                fspace = mspace = NULL;
            } /* end else */

            /* Switch back to collective I/O */
            if(last_coll_opt_mode != H5FD_MPIO_COLLECTIVE_IO) {
                if(H5CX_set_mpio_coll_opt(H5FD_MPIO_COLLECTIVE_IO) < 0)
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTSET, FAIL, "can't switch to collective I/O")
                last_coll_opt_mode = H5FD_MPIO_COLLECTIVE_IO;
            } /* end if */

            /* Initialize temporary contiguous storage address */
            ctg_store.contig.dset_addr = chunk_addr[u];

            /* Perform the I/O */
            if(H5D__inter_collective_io(&ctg_io_info, type_info, fspace, mspace) < 0)
                HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish shared collective MPI-IO")
        } /* end if */
        else {  /* possible independent IO for this chunk */
#ifdef H5D_DEBUG
if(H5DEBUG(D))
    HDfprintf(H5DEBUG(D),"inside independent IO mpi_rank = %d, chunk index = %Zu\n", mpi_rank, u);
#endif

            HDassert(chunk_io_option[u] == 0);

            /* Set the file & memory dataspaces */
            if(chunk_info) {
                fspace = chunk_info->fspace;
                mspace = chunk_info->mspace;

                /* Update the local variable tracking the actual io mode. */
                actual_io_mode = (H5D_mpio_actual_io_mode_t) (actual_io_mode | H5D_MPIO_CHUNK_INDEPENDENT);
            } /* end if */
            else {
                fspace = mspace = NULL;
            } /* end else */

            /* Using independent I/O with file setview.*/
            if(last_coll_opt_mode != H5FD_MPIO_INDIVIDUAL_IO) {
                if(H5CX_set_mpio_coll_opt(H5FD_MPIO_INDIVIDUAL_IO) < 0)
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTSET, FAIL, "can't switch to individual I/O")
                last_coll_opt_mode = H5FD_MPIO_INDIVIDUAL_IO;
            } /* end if */

            /* Initialize temporary contiguous storage address */
            ctg_store.contig.dset_addr = chunk_addr[u];

            /* Perform the I/O */
            if(H5D__inter_collective_io(&ctg_io_info, type_info, fspace, mspace) < 0)
                HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish shared collective MPI-IO")
#ifdef H5D_DEBUG
  if(H5DEBUG(D))
    HDfprintf(H5DEBUG(D),"after inter collective IO\n");
#endif
        } /* end else */
    } /* end for */

    /* Write the local value of actual io mode to the API context. */
    H5CX_set_mpio_actual_io_mode(actual_io_mode);

done:
    if(chunk_io_option)
        H5MM_xfree(chunk_io_option);
    if(chunk_addr)
        H5MM_xfree(chunk_addr);

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__multi_chunk_collective_io */


/*-------------------------------------------------------------------------
 * Function:    H5D__multi_chunk_filtered_collective_io
 *
 * Purpose:     To do filtered collective IO iteratively to save on memory.
 *              While link_chunk_filtered_collective_io will construct and
 *              work on a list of all of the chunks selected in the IO
 *              operation at once, this function works iteratively on a set
 *              of chunks at a time; at most one chunk per rank per
 *              iteration.
 *
 *              1. Construct a list of selected chunks in the collective IO
 *                 operation
 *                 A. If any chunk is being written to by more than 1
 *                    process, the process writing to the chunk which
 *                    currently has the least amount of chunks assigned
 *                    to it becomes the new owner (in the case of ties,
 *                    the lowest MPI rank becomes the new owner)
 *              2. If the operation is a read operation
 *                 A. Loop through each chunk in the operation
 *                    I. Read the chunk from the file
 *                    II. Unfilter the chunk
 *                    III. Scatter the read chunk data to the user's buffer
 *              3. If the operation is a write operation
 *                 A. Loop through each chunk in the operation
 *                    I. If this is not a full overwrite of the chunk
 *                       a) Read the chunk from file and pass the chunk
 *                          through the filter pipeline in reverse order
 *                          (Unfilter the chunk)
 *                    II. Update the chunk data with the modifications from
 *                        the owning process
 *                    III. Receive any modification data from other
 *                         processes and update the chunk data with these
 *                         modifications
 *                    IV. Filter the chunk
 *                    V. Contribute the chunk to an array gathered by
 *                        all processes which contains every chunk
 *                        modified in this iteration (up to one chunk
 *                        per process, some processes may not have a
 *                        selection/may have less chunks to work on than
 *                        other processes)
 *                    VI. All processes collectively re-allocate each
 *                        chunk from the gathered array with their new
 *                        sizes after the filter operation
 *                    VII. Proceed with the collective write operation
 *                        for the chunks modified on this iteration
 *                    VIII. All processes collectively re-insert each
 *                       chunk from the gathered array into the chunk
 *                       index
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Jordan Henderson
 *              Friday, Dec. 2nd, 2016
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__multi_chunk_filtered_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
    H5D_chunk_map_t *fm)
{
    H5D_filtered_collective_io_info_t *chunk_list = NULL; /* The list of chunks being read/written */
    H5D_filtered_collective_io_info_t *collective_chunk_list = NULL; /* The list of chunks used during collective operations */
    H5D_storage_t                      store;                /* union of EFL and chunk pointer in file space */
    H5D_io_info_t                      ctg_io_info;          /* Contiguous I/O info object */
    H5D_storage_t                      ctg_store;            /* Chunk storage information as contiguous dataset */
    MPI_Datatype                      *file_type_array = NULL;
    MPI_Datatype                      *mem_type_array = NULL;
    hbool_t                           *file_type_is_derived_array = NULL;
    hbool_t                           *mem_type_is_derived_array = NULL;
    hbool_t                           *has_chunk_selected_array = NULL; /* Array of whether or not each process is contributing a chunk to each iteration */
    size_t                             chunk_list_num_entries;
    size_t                             collective_chunk_list_num_entries;
    size_t                             i, j;                    /* Local index variable */
    int                                mpi_rank, mpi_size, mpi_code;
    herr_t                             ret_value = SUCCEED;

    FUNC_ENTER_STATIC

    HDassert(io_info);
    HDassert(type_info);
    HDassert(fm);

    /* Obtain the current rank of the process and the number of processes */
    if ((mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file)) < 0)
        HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi rank")
    if ((mpi_size = H5F_mpi_get_size(io_info->dset->oloc.file)) < 0)
        HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi size")

    /* Set the actual chunk opt mode property */
    H5CX_set_mpio_actual_chunk_opt(H5D_MPIO_MULTI_CHUNK);

    /* Set the actual_io_mode property.
     * Multi chunk I/O does not break to independent, so can set right away
     */
    H5CX_set_mpio_actual_io_mode(H5D_MPIO_CHUNK_COLLECTIVE);

    /* Build a list of selected chunks in the collective IO operation */
    if (H5D__construct_filtered_io_info_list(io_info, type_info, fm, &chunk_list, &chunk_list_num_entries) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "couldn't construct filtered I/O info list")

    /* Set up contiguous I/O info object */
    HDmemcpy(&ctg_io_info, io_info, sizeof(ctg_io_info));
    ctg_io_info.store = &ctg_store;
    ctg_io_info.layout_ops = *H5D_LOPS_CONTIG;

    /* Initialize temporary contiguous storage info */
    ctg_store.contig.dset_size = (hsize_t) io_info->dset->shared->layout.u.chunk.size;
    ctg_store.contig.dset_addr = 0;

    /* Set dataset storage for I/O info */
    io_info->store = &store;

    if (io_info->op_type == H5D_IO_OP_READ) { /* Filtered collective read */
        for (i = 0; i < chunk_list_num_entries; i++)
            if (H5D__filtered_collective_chunk_entry_io(&chunk_list[i], io_info, type_info, fm) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "couldn't process chunk entry")
    } /* end if */
    else { /* Filtered collective write */
        H5D_chk_idx_info_t index_info;
        H5D_chunk_ud_t     udata;
        size_t             max_num_chunks;
        hsize_t            mpi_buf_count;

        /* Construct chunked index info */
        index_info.f = io_info->dset->oloc.file;
        index_info.pline = &(io_info->dset->shared->dcpl_cache.pline);
        index_info.layout = &(io_info->dset->shared->layout.u.chunk);
        index_info.storage = &(io_info->dset->shared->layout.storage.u.chunk);

        /* Set up chunk information for insertion to chunk index */
        udata.common.layout = index_info.layout;
        udata.common.storage = index_info.storage;
        udata.filter_mask = 0;

        /* Retrieve the maximum number of chunks being written among all processes */
        if (MPI_SUCCESS != (mpi_code = MPI_Allreduce(&chunk_list_num_entries, &max_num_chunks,
                1, MPI_UNSIGNED_LONG_LONG, MPI_MAX, io_info->comm)))
            HMPI_GOTO_ERROR(FAIL, "MPI_Allreduce failed", mpi_code)

        /* If no one is writing anything at all, end the operation */
        if (!(max_num_chunks > 0)) HGOTO_DONE(SUCCEED);

        /* Allocate arrays for storing MPI file and mem types and whether or not the
         * types were derived.
         */
        if (NULL == (file_type_array = (MPI_Datatype *) H5MM_malloc(max_num_chunks * sizeof(MPI_Datatype))))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate file type array")

        if (NULL == (file_type_is_derived_array = (hbool_t *) H5MM_calloc(max_num_chunks * sizeof(hbool_t))))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate file type is derived array")

        if (NULL == (mem_type_array = (MPI_Datatype *) H5MM_malloc(max_num_chunks * sizeof(MPI_Datatype))))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate mem type array")

        if (NULL == (mem_type_is_derived_array = (hbool_t *) H5MM_calloc(max_num_chunks * sizeof(hbool_t))))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate mem type is derived array")

        /* Iterate over the max number of chunks among all processes, as this process could
         * have no chunks left to work on, but it still needs to participate in the collective
         * re-allocation and re-insertion of chunks modified by other processes.
         */
        for (i = 0; i < max_num_chunks; i++) {
            /* Check if this process has a chunk to work on for this iteration */
            hbool_t have_chunk_to_process = (i < chunk_list_num_entries) && (mpi_rank == chunk_list[i].owners.new_owner);

            if (have_chunk_to_process)
                if (H5D__filtered_collective_chunk_entry_io(&chunk_list[i], io_info, type_info, fm) < 0)
                    HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "couldn't process chunk entry")

            /* Gather the new chunk sizes to all processes for a collective re-allocation
             * of the chunks in the file
             */
            if (H5D__mpio_array_gatherv(&chunk_list[i], have_chunk_to_process ? 1 : 0, sizeof(H5D_filtered_collective_io_info_t),
                    (void **) &collective_chunk_list, &collective_chunk_list_num_entries, true, 0, io_info->comm, NULL) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTGATHER, FAIL, "couldn't gather new chunk sizes")

            /* Participate in the collective re-allocation of all chunks modified
             * in this iteration.
             */
            for (j = 0; j < collective_chunk_list_num_entries; j++) {
                hbool_t insert = FALSE;

                if (H5D__chunk_file_alloc(&index_info, &collective_chunk_list[j].chunk_states.chunk_current,
                        &collective_chunk_list[j].chunk_states.new_chunk, &insert, chunk_list[j].scaled) < 0)
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate chunk")
            } /* end for */

            if (NULL == (has_chunk_selected_array = (hbool_t *) H5MM_malloc((size_t) mpi_size * sizeof(hbool_t))))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate num chunks selected array")

            if (MPI_SUCCESS != (mpi_code = MPI_Allgather(&have_chunk_to_process, 1, MPI_C_BOOL, has_chunk_selected_array,
                    1, MPI_C_BOOL, io_info->comm)))
                HMPI_GOTO_ERROR(FAIL, "MPI_Allgather failed", mpi_code)

            /* If this process has a chunk to work on, create a MPI type for the
             * memory and file for writing out the chunk
             */
            if (have_chunk_to_process) {
                size_t offset;
                int    mpi_type_count;

                for (j = 0, offset = 0; j < (size_t) mpi_rank; j++)
                    offset += has_chunk_selected_array[j];

                /* Collect the new chunk info back to the local copy, since only the record in the
                 * collective array gets updated by the chunk re-allocation */
                HDmemcpy(&chunk_list[i].chunk_states.new_chunk, &collective_chunk_list[offset].chunk_states.new_chunk, sizeof(chunk_list[i].chunk_states.new_chunk));

                H5_CHECKED_ASSIGN(mpi_type_count, int, chunk_list[i].chunk_states.new_chunk.length, hsize_t);

                /* Create MPI memory type for writing to chunk */
                if (MPI_SUCCESS != (mpi_code = MPI_Type_contiguous(mpi_type_count, MPI_BYTE, &mem_type_array[i])))
                    HMPI_GOTO_ERROR(FAIL, "MPI_Type_contiguous failed", mpi_code)
                if(MPI_SUCCESS != (mpi_code = MPI_Type_commit(&mem_type_array[i])))
                    HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code)
                mem_type_is_derived_array[i] = TRUE;

                /* Create MPI file type for writing to chunk */
                if (MPI_SUCCESS != (mpi_code = MPI_Type_contiguous(mpi_type_count, MPI_BYTE, &file_type_array[i])))
                    HMPI_GOTO_ERROR(FAIL, "MPI_Type_contiguous failed", mpi_code)
                if(MPI_SUCCESS != (mpi_code = MPI_Type_commit(&file_type_array[i])))
                    HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code)
                file_type_is_derived_array[i] = TRUE;

                mpi_buf_count = 1;

                /* Set up the base storage address for this operation */
                ctg_store.contig.dset_addr = chunk_list[i].chunk_states.new_chunk.offset;

                /* Override the write buffer to point to the address of the
                 * chunk data buffer
                 */
                ctg_io_info.u.wbuf = chunk_list[i].buf;
            } /* end if */
            else {
                mem_type_array[i] = file_type_array[i] = MPI_BYTE;
                mpi_buf_count = 0;
            } /* end else */

            /* Perform the I/O */
            if (H5D__final_collective_io(&ctg_io_info, type_info, mpi_buf_count, file_type_array[i], mem_type_array[i]) < 0)
                HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish MPI-IO")

            /* Participate in the collective re-insertion of all chunks modified
             * in this iteration into the chunk index
             */
            for (j = 0; j < collective_chunk_list_num_entries; j++) {
                udata.chunk_block = collective_chunk_list[j].chunk_states.new_chunk;
                udata.common.scaled = collective_chunk_list[j].scaled;
                udata.chunk_idx = collective_chunk_list[j].index;

                if ((index_info.storage->ops->insert)(&index_info, &udata, io_info->dset) < 0)
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTINSERT, FAIL, "unable to insert chunk address into index")
            } /* end for */

            if (collective_chunk_list){
                H5MM_free(collective_chunk_list);
                collective_chunk_list = NULL;
            } /* end if */
            if (has_chunk_selected_array){
                H5MM_free(has_chunk_selected_array);
                has_chunk_selected_array = NULL;
            } /* end if */
        } /* end for */

        /* Free the MPI file and memory types, if they were derived */
        for (i = 0; i < max_num_chunks; i++) {
            if (file_type_is_derived_array[i])
                if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&file_type_array[i])))
                    HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)

            if (mem_type_is_derived_array[i])
                if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&mem_type_array[i])))
                    HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
        } /* end for */
    } /* end else */

done:
    if (chunk_list) {
        for (i = 0; i < chunk_list_num_entries; i++)
            if (chunk_list[i].buf)
                H5MM_free(chunk_list[i].buf);

        H5MM_free(chunk_list);
    } /* end if */

    if (collective_chunk_list)
        H5MM_free(collective_chunk_list);
    if (file_type_array)
        H5MM_free(file_type_array);
    if (mem_type_array)
        H5MM_free(mem_type_array);
    if (file_type_is_derived_array)
        H5MM_free(file_type_is_derived_array);
    if (mem_type_is_derived_array)
        H5MM_free(mem_type_is_derived_array);

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__multi_chunk_filtered_collective_io() */


/*-------------------------------------------------------------------------
 * Function:    H5D__inter_collective_io
 *
 * Purpose:     Routine for the shared part of collective IO between multiple chunk
 *              collective IO and contiguous collective IO
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Muqun Yang
 *              Monday, Feb. 13th, 2006
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__inter_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
    const H5S_t *file_space, const H5S_t *mem_space)
{
    int mpi_buf_count;                  /* # of MPI types */
    hbool_t mbt_is_derived = FALSE;
    hbool_t mft_is_derived = FALSE;
    MPI_Datatype        mpi_file_type, mpi_buf_type;
    int                 mpi_code;       /* MPI return code */
    herr_t       ret_value = SUCCEED;   /* return value */

    FUNC_ENTER_STATIC

    if((file_space != NULL) && (mem_space != NULL)) {
        int  mpi_file_count;         /* Number of file "objects" to transfer */
        hsize_t *permute_map = NULL; /* array that holds the mapping from the old, 
                                        out-of-order displacements to the in-order 
                                        displacements of the MPI datatypes of the 
                                        point selection of the file space */
        hbool_t is_permuted = FALSE;

        /* Obtain disk and memory MPI derived datatype */
        /* NOTE: The permute_map array can be allocated within H5S_mpio_space_type
         *              and will be fed into the next call to H5S_mpio_space_type
         *              where it will be freed.
         */
        if(H5S_mpio_space_type(file_space, type_info->src_type_size, 
                               &mpi_file_type, &mpi_file_count, &mft_is_derived, /* OUT: datatype created */  
                               TRUE, /* this is a file space, so
                                        permute the datatype if the
                                        point selection is out of
                                        order */
                               &permute_map, /* OUT: a map to indicate
                                                the permutation of
                                                points selected in
                                                case they are out of
                                                order */ 
                               &is_permuted /* OUT */) < 0)
            HGOTO_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL, "couldn't create MPI file type")
        /* Sanity check */
        if(is_permuted)
            HDassert(permute_map);
        if(H5S_mpio_space_type(mem_space, type_info->src_type_size, 
                               &mpi_buf_type, &mpi_buf_count, &mbt_is_derived, /* OUT: datatype created */
                               FALSE, /* this is a memory space, so if
                                         the file space is not
                                         permuted, there is no need to
                                         permute the datatype if the
                                         point selections are out of
                                         order*/
                               &permute_map /* IN: the permutation map
                                               generated by the
                                               file_space selection
                                               and applied to the
                                               memory selection */, 
                               &is_permuted /* IN */) < 0)
            HGOTO_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL, "couldn't create MPI buffer type")
        /* Sanity check */
        if(is_permuted)
            HDassert(!permute_map);
    } /* end if */
    else {
        /* For non-selection, participate with a none MPI derived datatype, the count is 0.  */
        mpi_buf_type   = MPI_BYTE;
        mpi_file_type  = MPI_BYTE;
        mpi_buf_count  = 0;
        mbt_is_derived = FALSE;
        mft_is_derived = FALSE;
    } /* end else */

#ifdef H5D_DEBUG
if(H5DEBUG(D))
    HDfprintf(H5DEBUG(D),"before final collective IO \n");
#endif

    /* Perform final collective I/O operation */
    if(H5D__final_collective_io(io_info, type_info, (hsize_t)mpi_buf_count, mpi_file_type, mpi_buf_type) < 0)
        HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish collective MPI-IO")

done:
    /* Free the MPI buf and file types, if they were derived */
    if(mbt_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&mpi_buf_type)))
        HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
    if(mft_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&mpi_file_type)))
        HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)

#ifdef H5D_DEBUG
if(H5DEBUG(D))
    HDfprintf(H5DEBUG(D),"before leaving inter_collective_io ret_value = %d\n",ret_value);
#endif

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__inter_collective_io() */


/*-------------------------------------------------------------------------
 * Function:    H5D__final_collective_io
 *
 * Purpose:     Routine for the common part of collective IO with different storages.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Muqun Yang
 *              Monday, Feb. 13th, 2006
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__final_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
    hsize_t mpi_buf_count, MPI_Datatype mpi_file_type, MPI_Datatype mpi_buf_type)
{
    herr_t      ret_value = SUCCEED;

    FUNC_ENTER_STATIC

    /* Pass buf type, file type to the file driver.  */
    if(H5CX_set_mpi_coll_datatypes(mpi_buf_type, mpi_file_type) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTSET, FAIL, "can't set MPI-I/O collective I/O datatypes")

    if(io_info->op_type == H5D_IO_OP_WRITE) {
        if((io_info->io_ops.single_write)(io_info, type_info, mpi_buf_count, NULL, NULL) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "optimized write failed")
    } /* end if */
    else {
        if((io_info->io_ops.single_read)(io_info, type_info, mpi_buf_count, NULL, NULL) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "optimized read failed")
    } /* end else */

done:
#ifdef H5D_DEBUG
if(H5DEBUG(D))
    HDfprintf(H5DEBUG(D),"ret_value before leaving final_collective_io=%d\n",ret_value);
#endif
      FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__final_collective_io */


/*-------------------------------------------------------------------------
 * Function:    H5D__cmp_chunk_addr
 *
 * Purpose:     Routine to compare chunk addresses
 *
 * Description: Callback for qsort() to compare chunk addresses
 *
 * Return:      -1, 0, 1
 *
 * Programmer:  Muqun Yang
 *              Monday, Feb. 13th, 2006
 *
 *-------------------------------------------------------------------------
 */
static int
H5D__cmp_chunk_addr(const void *chunk_addr_info1, const void *chunk_addr_info2)
{
   haddr_t addr1 = HADDR_UNDEF, addr2 = HADDR_UNDEF;

   FUNC_ENTER_STATIC_NOERR

   addr1 = ((const H5D_chunk_addr_info_t *)chunk_addr_info1)->chunk_addr;
   addr2 = ((const H5D_chunk_addr_info_t *)chunk_addr_info2)->chunk_addr;

   FUNC_LEAVE_NOAPI(H5F_addr_cmp(addr1, addr2))
} /* end H5D__cmp_chunk_addr() */


/*-------------------------------------------------------------------------
 * Function:    H5D__cmp_filtered_collective_io_info_entry
 *
 * Purpose:     Routine to compare filtered collective chunk io info
 *              entries
 *
 * Description: Callback for qsort() to compare filtered collective chunk
 *              io info entries
 *
 * Return:      -1, 0, 1
 *
 * Programmer:  Jordan Henderson
 *              Wednesday, Nov. 30th, 2016
 *
 *-------------------------------------------------------------------------
 */
static int
H5D__cmp_filtered_collective_io_info_entry(const void *filtered_collective_io_info_entry1, const void *filtered_collective_io_info_entry2)
{
    haddr_t addr1 = HADDR_UNDEF, addr2 = HADDR_UNDEF;

    FUNC_ENTER_STATIC_NOERR

    addr1 = ((const H5D_filtered_collective_io_info_t *) filtered_collective_io_info_entry1)->chunk_states.new_chunk.offset;
    addr2 = ((const H5D_filtered_collective_io_info_t *) filtered_collective_io_info_entry2)->chunk_states.new_chunk.offset;

    FUNC_LEAVE_NOAPI(H5F_addr_cmp(addr1, addr2))
} /* end H5D__cmp_filtered_collective_io_info_entry() */

#if MPI_VERSION >= 3

/*-------------------------------------------------------------------------
 * Function:    H5D__cmp_filtered_collective_io_info_entry_owner
 *
 * Purpose:     Routine to compare filtered collective chunk io info
 *              entries's original owner fields
 *
 * Description: Callback for qsort() to compare filtered collective chunk
 *              io info entries's original owner fields
 *
 * Return:      The difference between the two
 *              H5D_filtered_collective_io_info_t's original owner fields
 *
 * Programmer:  Jordan Henderson
 *              Monday, Apr. 10th, 2017
 *
 *-------------------------------------------------------------------------
 */
static int
H5D__cmp_filtered_collective_io_info_entry_owner(const void *filtered_collective_io_info_entry1, const void *filtered_collective_io_info_entry2)
{
    int owner1 = -1, owner2 = -1;

    FUNC_ENTER_STATIC_NOERR

    owner1 = ((const H5D_filtered_collective_io_info_t *) filtered_collective_io_info_entry1)->owners.original_owner;
    owner2 = ((const H5D_filtered_collective_io_info_t *) filtered_collective_io_info_entry2)->owners.original_owner;

    FUNC_LEAVE_NOAPI(owner1 - owner2)
} /* end H5D__cmp_filtered_collective_io_info_entry_owner() */
#endif


/*-------------------------------------------------------------------------
 * Function:    H5D__sort_chunk
 *
 * Purpose:     Routine to sort chunks in increasing order of chunk address
 *              Each chunk address is also obtained.
 *
 * Description:
 *              For most cases, the chunk address has already been sorted in increasing order.
 *              The special sorting flag is used to optimize this common case.
 *              quick sort is used for necessary sorting.
 *
 * Parameters:
 *              Input: H5D_io_info_t* io_info,
 *                      H5D_chunk_map_t *fm(global chunk map struct)
 *              Input/Output:  H5D_chunk_addr_info_t chunk_addr_info_array[]   : array to store chunk address and information
 *                     many_chunk_opt                         : flag to optimize the way to obtain chunk addresses
 *                                                              for many chunks
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Muqun Yang
 *              Monday, Feb. 13th, 2006
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__sort_chunk(H5D_io_info_t *io_info, const H5D_chunk_map_t *fm,
    H5D_chunk_addr_info_t chunk_addr_info_array[], int sum_chunk)
{
    H5SL_node_t    *chunk_node;         /* Current node in chunk skip list */
    H5D_chunk_info_t *chunk_info;         /* Current chunking info. of this node. */
    haddr_t         chunk_addr;         /* Current chunking address of this node */
    haddr_t        *total_chunk_addr_array = NULL; /* The array of chunk address for the total number of chunk */
    hbool_t         do_sort = FALSE;    /* Whether the addresses need to be sorted */
    int             bsearch_coll_chunk_threshold;
    int             many_chunk_opt = H5D_OBTAIN_ONE_CHUNK_ADDR_IND;
    int             mpi_size;                   /* Number of MPI processes */
    int             mpi_code;                   /* MPI return code */
    int             i;                          /* Local index variable */
    herr_t          ret_value = SUCCEED;        /* Return value */

    FUNC_ENTER_STATIC

    /* Retrieve # of MPI processes */
    if((mpi_size = H5F_mpi_get_size(io_info->dset->oloc.file)) < 0)
        HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi size")

    /* Calculate the actual threshold to obtain all chunk addresses collectively
     *  The bigger this number is, the more possible the use of obtaining chunk
     * address collectively.
     */
    /* For non-optimization one-link IO, actual bsearch threshold is always
     *   0, we would always want to obtain the chunk addresses individually
     *   for each process.
     */
    bsearch_coll_chunk_threshold = (sum_chunk * 100) / ((int)fm->layout->u.chunk.nchunks * mpi_size);
    if((bsearch_coll_chunk_threshold > H5D_ALL_CHUNK_ADDR_THRES_COL)
            && ((sum_chunk / mpi_size) >= H5D_ALL_CHUNK_ADDR_THRES_COL_NUM))
        many_chunk_opt = H5D_OBTAIN_ALL_CHUNK_ADDR_COL;

#ifdef H5D_DEBUG
if(H5DEBUG(D))
    HDfprintf(H5DEBUG(D), "many_chunk_opt= %d\n", many_chunk_opt);
#endif

    /* If we need to optimize the way to obtain the chunk address */
    if(many_chunk_opt != H5D_OBTAIN_ONE_CHUNK_ADDR_IND) {
        int mpi_rank;

#ifdef H5D_DEBUG
if(H5DEBUG(D))
    HDfprintf(H5DEBUG(D), "Coming inside H5D_OBTAIN_ALL_CHUNK_ADDR_COL\n");
#endif
        /* Allocate array for chunk addresses */
        if(NULL == (total_chunk_addr_array = (haddr_t *)H5MM_malloc(sizeof(haddr_t) * (size_t)fm->layout->u.chunk.nchunks)))
            HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "unable to allocate memory chunk address array")

        /* Retrieve all the chunk addresses with process 0 */
        if((mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file)) < 0)
            HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi rank")

        if(mpi_rank == 0) {
            if(H5D__chunk_addrmap(io_info, total_chunk_addr_array) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get chunk address")
        } /* end if */

        /* Broadcasting the MPI_IO option info. and chunk address info. */
        if(MPI_SUCCESS != (mpi_code = MPI_Bcast(total_chunk_addr_array, (int)(sizeof(haddr_t) * fm->layout->u.chunk.nchunks), MPI_BYTE, (int)0, io_info->comm)))
            HMPI_GOTO_ERROR(FAIL, "MPI_BCast failed", mpi_code)
    } /* end if */

    /* Start at first node in chunk skip list */
    i = 0;
    if(NULL == (chunk_node = H5SL_first(fm->sel_chunks)))
        HGOTO_ERROR(H5E_STORAGE, H5E_CANTGET, FAIL,"couldn't get chunk node from skipped list")

    /* Iterate over all chunks for this process */
    while(chunk_node) {
        if(NULL == (chunk_info = (H5D_chunk_info_t *)H5SL_item(chunk_node)))
            HGOTO_ERROR(H5E_STORAGE, H5E_CANTGET, FAIL,"couldn't get chunk info from skipped list")

        if(many_chunk_opt == H5D_OBTAIN_ONE_CHUNK_ADDR_IND) {
            H5D_chunk_ud_t udata;   /* User data for querying chunk info */

            /* Get address of chunk */
            if(H5D__chunk_lookup(io_info->dset, chunk_info->scaled, &udata) < 0)
                HGOTO_ERROR(H5E_STORAGE, H5E_CANTGET, FAIL, "couldn't get chunk info from skipped list")
            chunk_addr = udata.chunk_block.offset;
        } /* end if */
        else
            chunk_addr = total_chunk_addr_array[chunk_info->index];

        /* Check if chunk addresses are not in increasing order in the file */
        if(i > 0 && chunk_addr < chunk_addr_info_array[i - 1].chunk_addr)
            do_sort = TRUE;

        /* Set the address & info for this chunk */
        chunk_addr_info_array[i].chunk_addr = chunk_addr;
        chunk_addr_info_array[i].chunk_info = *chunk_info;

        /* Advance to next chunk in list */
        i++;
        chunk_node = H5SL_next(chunk_node);
    } /* end while */

#ifdef H5D_DEBUG
if(H5DEBUG(D))
    HDfprintf(H5DEBUG(D), "before Qsort\n");
#endif
    if(do_sort) {
        size_t num_chunks = H5SL_count(fm->sel_chunks);

        HDqsort(chunk_addr_info_array, num_chunks, sizeof(chunk_addr_info_array[0]), H5D__cmp_chunk_addr);
    } /* end if */

done:
    if(total_chunk_addr_array)
        H5MM_xfree(total_chunk_addr_array);

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__sort_chunk() */


/*-------------------------------------------------------------------------
 * Function:    H5D__obtain_mpio_mode
 *
 * Purpose:     Routine to obtain each io mode(collective,independent or none) for each chunk;
 *              Each chunk address is also obtained.
 *
 * Description:
 *
 *              1) Each process provides two piece of information for all chunks having selection
 *                 a) chunk index
 *                 b) wheather this chunk is regular(for MPI derived datatype not working case)
 *
 *              2) Gather all the information to the root process
 *
 *              3) Root process will do the following:
 *                 a) Obtain chunk addresses for all chunks in this dataspace
 *                 b) With the consideration of the user option, calculate IO mode for each chunk
 *                 c) Build MPI derived datatype to combine "chunk address" and "assign_io" information
 *                      in order to do MPI Bcast only once
 *                 d) MPI Bcast the IO mode and chunk address information for each chunk.
 *              4) Each process then retrieves IO mode and chunk address information to assign_io_mode and chunk_addr.
 *
 * Parameters:
 *
 *              Input: H5D_io_info_t* io_info,
 *                      H5D_chunk_map_t *fm,(global chunk map struct)
 *              Output: uint8_t assign_io_mode[], : IO mode, collective, independent or none
 *                      haddr_t chunk_addr[],     : chunk address array for each chunk
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Muqun Yang
 *              Monday, Feb. 13th, 2006
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__obtain_mpio_mode(H5D_io_info_t* io_info, H5D_chunk_map_t *fm,
    uint8_t assign_io_mode[], haddr_t chunk_addr[])
{
    size_t            total_chunks;
    unsigned          percent_nproc_per_chunk, threshold_nproc_per_chunk;
    uint8_t*          io_mode_info = NULL;
    uint8_t*          recv_io_mode_info = NULL;
    uint8_t*          mergebuf = NULL;
    uint8_t*          tempbuf;
    H5SL_node_t*      chunk_node;
    H5D_chunk_info_t* chunk_info;
    int               mpi_size, mpi_rank;
    MPI_Comm          comm;
    int               root;
    size_t            ic;
    int               mpi_code;
    herr_t            ret_value = SUCCEED;

    FUNC_ENTER_STATIC

    /* Assign the rank 0 to the root */
    root              = 0;
    comm              = io_info->comm;

    /* Obtain the number of process and the current rank of the process */
    if((mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file)) < 0)
        HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi rank")
    if((mpi_size = H5F_mpi_get_size(io_info->dset->oloc.file)) < 0)
        HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi size")

    /* Setup parameters */
    H5_CHECKED_ASSIGN(total_chunks, size_t, fm->layout->u.chunk.nchunks, hsize_t);
    if(H5CX_get_mpio_chunk_opt_ratio(&percent_nproc_per_chunk) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "couldn't get percent nproc per chunk")
    /* if ratio is 0, perform collective io */
    if(0 == percent_nproc_per_chunk) {
        if(H5D__chunk_addrmap(io_info, chunk_addr) < 0)
           HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get chunk address");
        for(ic = 0; ic < total_chunks; ic++)
           assign_io_mode[ic] = H5D_CHUNK_IO_MODE_COL;

        HGOTO_DONE(SUCCEED)
    } /* end if */

    threshold_nproc_per_chunk = (unsigned)mpi_size * percent_nproc_per_chunk/100;

    /* Allocate memory */
    if(NULL == (io_mode_info = (uint8_t *)H5MM_calloc(total_chunks)))
        HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate I/O mode info buffer")
    if(NULL == (mergebuf = (uint8_t *)H5MM_malloc((sizeof(haddr_t) + 1) * total_chunks)))
        HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate mergebuf buffer")
    tempbuf           = mergebuf + total_chunks;
    if(mpi_rank == root)
        if(NULL == (recv_io_mode_info = (uint8_t *)H5MM_malloc(total_chunks * (size_t)mpi_size)))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate recv I/O mode info buffer")

    /* Obtain the regularity and selection information for all chunks in this process. */
    chunk_node        = H5SL_first(fm->sel_chunks);
    while(chunk_node) {
        chunk_info    = (H5D_chunk_info_t *)H5SL_item(chunk_node);

        io_mode_info[chunk_info->index] = H5D_CHUNK_SELECT_REG; /* this chunk is selected and is "regular" */
        chunk_node = H5SL_next(chunk_node);
    } /* end while */

    /* Gather all the information */
    H5_CHECK_OVERFLOW(total_chunks, size_t, int)
    if(MPI_SUCCESS != (mpi_code = MPI_Gather(io_mode_info, (int)total_chunks, MPI_BYTE,
            recv_io_mode_info, (int)total_chunks, MPI_BYTE, root, comm)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Gather failed", mpi_code)

    /* Calculate the mode for IO(collective, independent or none) at root process */
    if(mpi_rank == root) {
        size_t            nproc;
        unsigned*         nproc_per_chunk;

        /* pre-computing: calculate number of processes and
            regularity of the selection occupied in each chunk */
        if(NULL == (nproc_per_chunk = (unsigned*)H5MM_calloc(total_chunks * sizeof(unsigned))))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate nproc_per_chunk buffer")

        /* calculating the chunk address */
        if(H5D__chunk_addrmap(io_info, chunk_addr) < 0) {
            H5MM_free(nproc_per_chunk);
            HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get chunk address")
        } /* end if */

        /* checking for number of process per chunk and regularity of the selection*/
        for(nproc = 0; nproc < (size_t)mpi_size; nproc++) {
            uint8_t *tmp_recv_io_mode_info = recv_io_mode_info + (nproc * total_chunks);

            /* Calculate the number of process per chunk and adding irregular selection option */
            for(ic = 0; ic < total_chunks; ic++, tmp_recv_io_mode_info++) {
                if(*tmp_recv_io_mode_info != 0) {
                    nproc_per_chunk[ic]++;
                } /* end if */
            } /* end for */
        } /* end for */

        /* Calculating MPIO mode for each chunk (collective, independent, none) */
        for(ic = 0; ic < total_chunks; ic++) {
            if(nproc_per_chunk[ic] > MAX(1, threshold_nproc_per_chunk)) {
                assign_io_mode[ic] = H5D_CHUNK_IO_MODE_COL;
            } /* end if */
        } /* end for */


        /* merge buffer io_mode info and chunk addr into one */
        HDmemcpy(mergebuf, assign_io_mode, total_chunks);
        HDmemcpy(tempbuf, chunk_addr, sizeof(haddr_t) * total_chunks);

        H5MM_free(nproc_per_chunk);
    } /* end if */

    /* Broadcasting the MPI_IO option info. and chunk address info. */
    if(MPI_SUCCESS != (mpi_code = MPI_Bcast(mergebuf, ((sizeof(haddr_t) + 1) * total_chunks), MPI_BYTE, root, comm)))
        HMPI_GOTO_ERROR(FAIL, "MPI_BCast failed", mpi_code)

    HDmemcpy(assign_io_mode, mergebuf, total_chunks);
    HDmemcpy(chunk_addr, tempbuf, sizeof(haddr_t) * total_chunks);

#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
{
    hbool_t coll_op = FALSE;

    for(ic = 0; ic < total_chunks; ic++)
        if(assign_io_mode[ic] == H5D_CHUNK_IO_MODE_COL) {
            if(H5CX_test_set_mpio_coll_chunk_multi_ratio_coll(0) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTSET, FAIL, "unable to set property value")
            coll_op = TRUE;
            break;
        } /* end if */

    if(!coll_op)
        if(H5CX_test_set_mpio_coll_chunk_multi_ratio_ind(0) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTSET, FAIL, "unable to set property value")
}
#endif

done:
    if(io_mode_info)
        H5MM_free(io_mode_info);
    if(mergebuf)
        H5MM_free(mergebuf);
    if(recv_io_mode_info) {
        HDassert(mpi_rank == root);
        H5MM_free(recv_io_mode_info);
    } /* end if */

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__obtain_mpio_mode() */


/*-------------------------------------------------------------------------
 * Function:    H5D__construct_filtered_io_info_list
 *
 * Purpose:     Constructs a list of entries which contain the necessary
 *              information for inter-process communication when performing
 *              collective io on filtered chunks. This list is used by
 *              each process when performing I/O on locally selected chunks
 *              and also in operations that must be collectively done
 *              on every chunk, such as chunk re-allocation, insertion of
 *              chunks into the chunk index, etc.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Jordan Henderson
 *              Tuesday, January 10th, 2017
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__construct_filtered_io_info_list(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
    const H5D_chunk_map_t *fm, H5D_filtered_collective_io_info_t **chunk_list, size_t *num_entries)
{
    H5D_filtered_collective_io_info_t *local_info_array = NULL; /* The list of initially selected chunks for this process */
    size_t                             num_chunks_selected;
    size_t                             i;
    int                                mpi_rank;
    herr_t                             ret_value = SUCCEED;

    FUNC_ENTER_STATIC

    HDassert(io_info);
    HDassert(type_info);
    HDassert(fm);
    HDassert(chunk_list);
    HDassert(num_entries);

    if ((mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file)) < 0)
        HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi rank")

    /* Each process builds a local list of the chunks they have selected */
    if ((num_chunks_selected = H5SL_count(fm->sel_chunks))) {
        H5D_chunk_info_t *chunk_info;
        H5D_chunk_ud_t    udata;
        H5SL_node_t      *chunk_node;
        hssize_t          select_npoints;
        hssize_t          chunk_npoints;

        if(NULL == (local_info_array = (H5D_filtered_collective_io_info_t *) H5MM_malloc(num_chunks_selected * sizeof(H5D_filtered_collective_io_info_t))))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate local io info array buffer")

        chunk_node = H5SL_first(fm->sel_chunks);
        for(i = 0; chunk_node; i++) {
            chunk_info = (H5D_chunk_info_t *) H5SL_item(chunk_node);

            /* Obtain this chunk's address */
            if(H5D__chunk_lookup(io_info->dset, chunk_info->scaled, &udata) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "error looking up chunk address")

            local_info_array[i].index = chunk_info->index;
            local_info_array[i].chunk_states.chunk_current = local_info_array[i].chunk_states.new_chunk = udata.chunk_block;
            local_info_array[i].num_writers = 0;
            local_info_array[i].owners.original_owner = local_info_array[i].owners.new_owner = mpi_rank;
            local_info_array[i].buf = NULL;
            
            local_info_array[i].async_info.num_receive_requests = 0;
            local_info_array[i].async_info.receive_buffer_array = NULL;
            local_info_array[i].async_info.receive_requests_array = NULL;

            HDmemcpy(local_info_array[i].scaled, chunk_info->scaled, sizeof(chunk_info->scaled));

            if ((select_npoints = H5S_GET_SELECT_NPOINTS(chunk_info->mspace)) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTCOUNT, FAIL, "dataspace is invalid")
            local_info_array[i].io_size = (size_t) select_npoints * type_info->src_type_size;

            /* Currently the full overwrite status of a chunk is only obtained on a per-process
             * basis. This means that if the total selection in the chunk, as determined by the combination
             * of selections of all of the processes interested in the chunk, covers the entire chunk,
             * the performance optimization of not reading the chunk from the file is still valid, but
             * is not applied in the current implementation. Something like an appropriately placed
             * MPI_Allreduce or a running total of the number of chunk points selected during chunk
             * redistribution should suffice for implementing this case - JTH.
             */
            if ((chunk_npoints = H5S_GET_EXTENT_NPOINTS(chunk_info->fspace)) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTCOUNT, FAIL, "dataspace is invalid")
            local_info_array[i].full_overwrite =
                    (local_info_array[i].io_size >= (hsize_t) chunk_npoints * type_info->dst_type_size) ? TRUE : FALSE;

            chunk_node = H5SL_next(chunk_node);
        } /* end for */
    } /* end if */

    /* Redistribute shared chunks to new owners as necessary */
    if (io_info->op_type == H5D_IO_OP_WRITE)
#if MPI_VERSION >= 3
        if (H5D__chunk_redistribute_shared_chunks(io_info, type_info, fm, local_info_array, &num_chunks_selected) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "unable to redistribute shared chunks")
#else
        HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "unable to redistribute shared chunks - MPI version < 3 (MPI_Mprobe and MPI_Imrecv missing)")
#endif

    *chunk_list = local_info_array;
    *num_entries = num_chunks_selected;

done:
    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__construct_filtered_io_info_list() */

#if MPI_VERSION >= 3

/*-------------------------------------------------------------------------
 * Function:    H5D__chunk_redistribute_shared_chunks
 *
 * Purpose:     When performing a collective write on a Dataset with
 *              filters applied, this function is used to redistribute any
 *              chunks which are selected by more than one process, so as
 *              to preserve file integrity after the write by ensuring
 *              that any shared chunks are only modified by one process.
 *
 *              The current implementation follows this 3-phase process:
 *
 *              - Collect everyone's list of chunks into one large list,
 *                sort the list in increasing order of chunk offset in the
 *                file and hand the list off to rank 0
 *
 *              - Rank 0 scans the list looking for matching runs of chunk
 *                offset in the file (corresponding to a shared chunk which
 *                has been selected by more than one rank in the I/O
 *                operation) and for each shared chunk, it redistributes
 *                the chunk to the process writing to the chunk which
 *                currently has the least amount of chunks assigned to it
 *                by modifying the "new_owner" field in each of the list
 *                entries corresponding to that chunk
 *
 *              - After the chunks have been redistributed, rank 0 re-sorts
 *                the list in order of previous owner so that each rank
 *                will get back exactly the array that they contributed to
 *                the redistribution operation, with the "new_owner" field
 *                of each chunk they are modifying having possibly been
 *                modified. Rank 0 then scatters each segment of the list
 *                back to its corresponding rank
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Jordan Henderson
 *              Monday, May 1, 2017
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__chunk_redistribute_shared_chunks(const H5D_io_info_t *io_info, const H5D_type_info_t *type_info,
    const H5D_chunk_map_t *fm, H5D_filtered_collective_io_info_t *local_chunk_array, size_t *local_chunk_array_num_entries)
{
    H5D_filtered_collective_io_info_t  *shared_chunks_info_array = NULL; /* The list of all chunks selected in the operation by all processes */
    H5S_sel_iter_t                     *mem_iter = NULL; /* Memory iterator for H5D__gather_mem */
    unsigned char                     **mod_data = NULL; /* Array of chunk modification data buffers sent by a process to new chunk owners */
    MPI_Request                        *send_requests = NULL; /* Array of MPI_Isend chunk modification data send requests */
    MPI_Status                         *send_statuses = NULL; /* Array of MPI_Isend chunk modification send statuses */
    hbool_t                             mem_iter_init = FALSE;
    size_t                              shared_chunks_info_array_num_entries = 0;
    size_t                              num_send_requests = 0;
    size_t                             *num_assigned_chunks_array = NULL;
    size_t                              i, last_assigned_idx;
    int                                *send_counts = NULL;
    int                                *send_displacements = NULL;
    int                                 scatter_recvcount_int;
    int                                 mpi_rank, mpi_size, mpi_code;
    herr_t                              ret_value = SUCCEED;

    FUNC_ENTER_STATIC

    HDassert(io_info);
    HDassert(type_info);
    HDassert(fm);
    HDassert(local_chunk_array_num_entries);

    if ((mpi_rank = H5F_mpi_get_rank(io_info->dset->oloc.file)) < 0)
        HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi rank")
    if ((mpi_size = H5F_mpi_get_size(io_info->dset->oloc.file)) < 0)
        HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain mpi size")

    if (*local_chunk_array_num_entries)
        if (NULL == (send_requests = (MPI_Request *) H5MM_malloc(*local_chunk_array_num_entries * sizeof(MPI_Request))))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate send requests buffer")

    if (NULL == (mem_iter = (H5S_sel_iter_t *) H5MM_malloc(sizeof(H5S_sel_iter_t))))
        HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate memory iterator")

    /* Gather every rank's list of chunks to rank 0 to allow it to perform the redistribution operation. After this
     * call, the gathered list will initially be sorted in increasing order of chunk offset in the file.
     */
    if (H5D__mpio_array_gatherv(local_chunk_array, *local_chunk_array_num_entries, sizeof(H5D_filtered_collective_io_info_t),
            (void **) &shared_chunks_info_array, &shared_chunks_info_array_num_entries, false, 0,
            io_info->comm, H5D__cmp_filtered_collective_io_info_entry) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTGATHER, FAIL, "couldn't gather array")

    /* Rank 0 redistributes any shared chunks to new owners as necessary */
    if (mpi_rank == 0) {
        if (NULL == (send_counts = (int *) H5MM_calloc((size_t) mpi_size * sizeof(int))))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate send counts buffer")

        if (NULL == (send_displacements = (int *) H5MM_malloc((size_t) mpi_size * sizeof(int))))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate send displacements buffer")

        if (NULL == (num_assigned_chunks_array = (size_t *) H5MM_calloc((size_t) mpi_size * sizeof(size_t))))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate number of assigned chunks array")

        for (i = 0; i < shared_chunks_info_array_num_entries;) {
            H5D_filtered_collective_io_info_t chunk_entry;
            haddr_t                           last_seen_addr = shared_chunks_info_array[i].chunk_states.chunk_current.offset;
            size_t                            set_begin_index = i;
            size_t                            num_writers = 0;
            int                               new_chunk_owner = shared_chunks_info_array[i].owners.original_owner;

            /* Process each set of duplicate entries caused by another process writing to the same chunk */
            do {
                chunk_entry = shared_chunks_info_array[i];

                send_counts[chunk_entry.owners.original_owner] += (int) sizeof(chunk_entry);

                /* The new owner of the chunk is determined by the process
                 * writing to the chunk which currently has the least amount
                 * of chunks assigned to it
                 */
                if (num_assigned_chunks_array[chunk_entry.owners.original_owner] < num_assigned_chunks_array[new_chunk_owner])
                    new_chunk_owner = chunk_entry.owners.original_owner;

                num_writers++;
            } while (++i < shared_chunks_info_array_num_entries && shared_chunks_info_array[i].chunk_states.chunk_current.offset == last_seen_addr);

            /* Set all of the chunk entries' "new_owner" fields */
            for (; set_begin_index < i; set_begin_index++) {
                shared_chunks_info_array[set_begin_index].owners.new_owner = new_chunk_owner;
                shared_chunks_info_array[set_begin_index].num_writers = num_writers;
            } /* end for */

            num_assigned_chunks_array[new_chunk_owner]++;
        } /* end for */

        /* Sort the new list in order of previous owner so that each original owner of a chunk
         * entry gets that entry back, with the possibly newly-modified "new_owner" field
         */
        HDqsort(shared_chunks_info_array, shared_chunks_info_array_num_entries,
                sizeof(H5D_filtered_collective_io_info_t), H5D__cmp_filtered_collective_io_info_entry_owner);

        send_displacements[0] = 0;
        for (i = 1; i < (size_t) mpi_size; i++)
            send_displacements[i] = send_displacements[i - 1] + send_counts[i - 1];
    } /* end if */

    /* Scatter the segments of the list back to each process */
    H5_CHECKED_ASSIGN(scatter_recvcount_int, int, *local_chunk_array_num_entries * sizeof(H5D_filtered_collective_io_info_t), size_t);
    if (MPI_SUCCESS != (mpi_code = MPI_Scatterv(shared_chunks_info_array, send_counts, send_displacements,
            MPI_BYTE, local_chunk_array, scatter_recvcount_int, MPI_BYTE, 0, io_info->comm)))
        HMPI_GOTO_ERROR(FAIL, "unable to scatter shared chunks info buffer", mpi_code)

    if (shared_chunks_info_array) {
        H5MM_free(shared_chunks_info_array);
        shared_chunks_info_array = NULL;
    } /* end if */

    /* Now that the chunks have been redistributed, each process must send its modification data
     * to the new owners of any of the chunks it previously possessed. Accordingly, each process
     * must also issue asynchronous receives for any messages it may receive for each of the
     * chunks it is assigned, in order to avoid potential deadlocking issues.
     */
    if (*local_chunk_array_num_entries)
        if (NULL == (mod_data = (unsigned char **) H5MM_malloc(*local_chunk_array_num_entries * sizeof(unsigned char *))))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate modification data buffer array")

    for (i = 0, last_assigned_idx = 0; i < *local_chunk_array_num_entries; i++) {
        H5D_filtered_collective_io_info_t *chunk_entry = &local_chunk_array[i];

        if (mpi_rank != chunk_entry->owners.new_owner) {
            H5D_chunk_info_t *chunk_info = NULL;
            unsigned char    *mod_data_p = NULL;
            hssize_t          iter_nelmts;
            size_t            mod_data_size;

            /* Look up the chunk and get its file and memory dataspaces */
            if(NULL == (chunk_info = (H5D_chunk_info_t *) H5SL_search(fm->sel_chunks, &chunk_entry->index)))
                HGOTO_ERROR(H5E_DATASPACE, H5E_NOTFOUND, FAIL, "can't locate chunk in skip list")

            /* Determine size of serialized chunk file dataspace, plus the size of
             * the data being written
             */
            if(H5S_encode(chunk_info->fspace, &mod_data_p, &mod_data_size) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTENCODE, FAIL, "unable to get encoded dataspace size")

            if((iter_nelmts = H5S_GET_SELECT_NPOINTS(chunk_info->mspace)) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTCOUNT, FAIL, "dataspace is invalid")

            mod_data_size += (size_t) iter_nelmts * type_info->src_type_size;

            if(NULL == (mod_data[num_send_requests] = (unsigned char *) H5MM_malloc(mod_data_size)))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk modification send buffer")

            /* Serialize the chunk's file dataspace into the buffer */
            mod_data_p = mod_data[num_send_requests];
            if(H5S_encode(chunk_info->fspace, &mod_data_p, &mod_data_size) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTENCODE, FAIL, "unable to encode dataspace")

            /* Initialize iterator for memory selection */
            if(H5S_select_iter_init(mem_iter, chunk_info->mspace, type_info->src_type_size) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "unable to initialize memory selection information")
            mem_iter_init = TRUE;

            /* Collect the modification data into the buffer */
            if(!H5D__gather_mem(io_info->u.wbuf, chunk_info->mspace, mem_iter, (size_t)iter_nelmts, mod_data_p))
                HGOTO_ERROR(H5E_IO, H5E_CANTGATHER, FAIL, "couldn't gather from write buffer")

            /* Send modification data to new owner */
            H5_CHECK_OVERFLOW(mod_data_size, size_t, int)
            H5_CHECK_OVERFLOW(chunk_entry->index, hsize_t, int)
            if(MPI_SUCCESS != (mpi_code = MPI_Isend(mod_data[num_send_requests], (int) mod_data_size, MPI_BYTE,
                    chunk_entry->owners.new_owner, (int) chunk_entry->index, io_info->comm, &send_requests[num_send_requests])))
                HMPI_GOTO_ERROR(FAIL, "MPI_Isend failed", mpi_code)

            if(mem_iter_init && H5S_SELECT_ITER_RELEASE(mem_iter) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release memory selection iterator")
            mem_iter_init = FALSE;

            num_send_requests++;
        } /* end if */
        else {
            /* Allocate all necessary buffers for an asynchronous receive operation */
            if (chunk_entry->num_writers > 1) {
                MPI_Message message;
                MPI_Status  status;
                size_t      j;

                chunk_entry->async_info.num_receive_requests = (int) chunk_entry->num_writers - 1;
                if (NULL == (chunk_entry->async_info.receive_requests_array = (MPI_Request *) H5MM_malloc((size_t) chunk_entry->async_info.num_receive_requests * sizeof(MPI_Request))))
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate async requests array")

                if (NULL == (chunk_entry->async_info.receive_buffer_array = (unsigned char **) H5MM_malloc((size_t) chunk_entry->async_info.num_receive_requests * sizeof(unsigned char *))))
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate async receive buffers")

                for (j = 0; j < chunk_entry->num_writers - 1; j++) {
                    int count = 0;

                    /* Probe for a particular message from any process, removing that message
                     * from the receive queue in the process and allocating that much memory
                     * for the asynchronous receive
                     */
                    if (MPI_SUCCESS != (mpi_code = MPI_Mprobe(MPI_ANY_SOURCE, (int) chunk_entry->index, io_info->comm, &message, &status)))
                        HMPI_GOTO_ERROR(FAIL, "MPI_Mprobe failed", mpi_code)

                    if (MPI_SUCCESS != (mpi_code = MPI_Get_count(&status, MPI_BYTE, &count)))
                        HMPI_GOTO_ERROR(FAIL, "MPI_Get_count failed", mpi_code)

                    HDassert(count >= 0);
                    if (NULL == (chunk_entry->async_info.receive_buffer_array[j] = (unsigned char *) H5MM_malloc((size_t) count * sizeof(char *))))
                        HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate modification data receive buffer")

                    if (MPI_SUCCESS != (mpi_code = MPI_Imrecv(chunk_entry->async_info.receive_buffer_array[j], count, MPI_BYTE,
                            &message, &chunk_entry->async_info.receive_requests_array[j])))
                        HMPI_GOTO_ERROR(FAIL, "MPI_Imrecv failed", mpi_code)
                } /* end for */
            } /* end if */

            local_chunk_array[last_assigned_idx++] = local_chunk_array[i];
        } /* end else */
    } /* end for */

    *local_chunk_array_num_entries = last_assigned_idx;

    /* Wait for all async send requests to complete before returning */
    if (num_send_requests) {
        if (NULL == (send_statuses = (MPI_Status *) H5MM_malloc(num_send_requests * sizeof(MPI_Status))))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate send statuses buffer")

        H5_CHECK_OVERFLOW(num_send_requests, size_t, int);
        if (MPI_SUCCESS != (mpi_code = MPI_Waitall((int) num_send_requests, send_requests, send_statuses)))
            HMPI_GOTO_ERROR(FAIL, "MPI_Waitall failed", mpi_code)
    } /* end if */

done:
    /* Now that all async send requests have completed, free up the send
     * buffers used in the async operations
     */
    for (i = 0; i < num_send_requests; i++) {
        if (mod_data[i])
            H5MM_free(mod_data[i]);
    } /* end for */

    if (send_requests)
        H5MM_free(send_requests);
    if (send_statuses)
        H5MM_free(send_statuses);
    if (send_counts)
        H5MM_free(send_counts);
    if (send_displacements)
        H5MM_free(send_displacements);
    if (mod_data)
        H5MM_free(mod_data);
    if (mem_iter_init && H5S_SELECT_ITER_RELEASE(mem_iter) < 0)
        HDONE_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release selection iterator")
    if (mem_iter)
        H5MM_free(mem_iter);
    if (num_assigned_chunks_array)
        H5MM_free(num_assigned_chunks_array);
    if (shared_chunks_info_array)
        H5MM_free(shared_chunks_info_array);

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__chunk_redistribute_shared_chunks() */
#endif


/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_filtered_collective_write_type
 *
 * Purpose:     Constructs a MPI derived datatype for both the memory and
 *              the file for a collective write of filtered chunks. The
 *              datatype contains the offsets in the file and the locations
 *              of the filtered chunk data buffers.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Jordan Henderson
 *              Tuesday, November 22, 2016
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_filtered_collective_write_type(H5D_filtered_collective_io_info_t *chunk_list,
    size_t num_entries, MPI_Datatype *new_mem_type, hbool_t *mem_type_derived,
    MPI_Datatype *new_file_type, hbool_t *file_type_derived)
{
    MPI_Aint *write_buf_array = NULL;   /* Relative displacements of filtered chunk data buffers */
    MPI_Aint *file_offset_array = NULL; /* Chunk offsets in the file */
    int      *length_array = NULL;      /* Filtered Chunk lengths */
    herr_t    ret_value = SUCCEED;

    FUNC_ENTER_STATIC

    HDassert(chunk_list);
    HDassert(new_mem_type);
    HDassert(mem_type_derived);
    HDassert(new_file_type);
    HDassert(file_type_derived);

    if (num_entries > 0) {
        size_t i;
        int    mpi_code;
        void  *base_buf;

        H5_CHECK_OVERFLOW(num_entries, size_t, int);

        /* Allocate arrays */
        if (NULL == (length_array = (int *) H5MM_malloc((size_t) num_entries * sizeof(int))))
            HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "memory allocation failed for filtered collective write length array")
        if (NULL == (write_buf_array = (MPI_Aint *) H5MM_malloc((size_t) num_entries * sizeof(MPI_Aint))))
            HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "memory allocation failed for filtered collective write buf length array")
        if (NULL == (file_offset_array = (MPI_Aint *) H5MM_malloc((size_t) num_entries * sizeof(MPI_Aint))))
            HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "memory allocation failed for collective write offset array")

        /* Ensure the list is sorted in ascending order of offset in the file */
        HDqsort(chunk_list, num_entries, sizeof(H5D_filtered_collective_io_info_t), H5D__cmp_filtered_collective_io_info_entry);

        base_buf = chunk_list[0].buf;
        for (i = 0; i < num_entries; i++) {
            /* Set up the offset in the file, the length of the chunk data, and the relative
             * displacement of the chunk data write buffer
             */
            file_offset_array[i] = (MPI_Aint) chunk_list[i].chunk_states.new_chunk.offset;
            length_array[i] = (int) chunk_list[i].chunk_states.new_chunk.length;
            write_buf_array[i] = (MPI_Aint) chunk_list[i].buf - (MPI_Aint) base_buf;
        } /* end for */

        /* Create memory MPI type */
        if (MPI_SUCCESS != (mpi_code = MPI_Type_create_hindexed((int) num_entries, length_array, write_buf_array, MPI_BYTE, new_mem_type)))
            HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_hindexed failed", mpi_code)
        *mem_type_derived = TRUE;
        if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(new_mem_type)))
            HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code)

        /* Create file MPI type */
        if (MPI_SUCCESS != (mpi_code = MPI_Type_create_hindexed((int) num_entries, length_array, file_offset_array, MPI_BYTE, new_file_type)))
            HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_hindexed failed", mpi_code)
        *file_type_derived = TRUE;
        if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(new_file_type)))
            HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code)
    } /* end if */

done:
    if(write_buf_array)
        H5MM_free(write_buf_array);
    if(file_offset_array)
        H5MM_free(file_offset_array);
    if(length_array)
        H5MM_free(length_array);

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__mpio_filtered_collective_write_type() */


/*-------------------------------------------------------------------------
 * Function:    H5D__filtered_collective_chunk_entry_io
 *
 * Purpose:     Given an entry for a filtered chunk, performs the necessary
 *              steps for updating the chunk data during a collective
 *              write, or for reading the chunk from file during a
 *              collective read.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 * Programmer:  Jordan Henderson
 *              Wednesday, January 18, 2017
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__filtered_collective_chunk_entry_io(H5D_filtered_collective_io_info_t *chunk_entry,
    const H5D_io_info_t *io_info, const H5D_type_info_t *type_info, const H5D_chunk_map_t *fm)
{
    H5D_chunk_info_t *chunk_info = NULL;
    H5S_sel_iter_t   *mem_iter = NULL; /* Memory iterator for H5D__scatter_mem/H5D__gather_mem */
    H5S_sel_iter_t   *file_iter = NULL;
    H5Z_EDC_t         err_detect;       /* Error detection info */
    H5Z_cb_t          filter_cb;        /* I/O filter callback function */
    unsigned          filter_mask = 0;
    hssize_t          iter_nelmts;     /* Number of points to iterate over for the chunk IO operation */
    hssize_t          extent_npoints;
    hsize_t           true_chunk_size;
    hbool_t           mem_iter_init = FALSE;
    hbool_t           file_iter_init = FALSE;
    size_t            buf_size;
    size_t            i;
    H5S_t            *dataspace = NULL; /* Other process' dataspace for the chunk */
    void             *tmp_gath_buf = NULL; /* Temporary gather buffer to gather into from application buffer
                                              before scattering out to the chunk data buffer (when writing data),
                                              or vice versa (when reading data) */
    int               mpi_code;
    herr_t            ret_value = SUCCEED;

    FUNC_ENTER_STATIC

    HDassert(chunk_entry);
    HDassert(io_info);
    HDassert(type_info);
    HDassert(fm);

    /* Retrieve filter settings from API context */
    if(H5CX_get_err_detect(&err_detect) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get error detection info")
    if(H5CX_get_filter_cb(&filter_cb) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get I/O filter callback function")

    /* Look up the chunk and get its file and memory dataspaces */
    if (NULL == (chunk_info = (H5D_chunk_info_t *) H5SL_search(fm->sel_chunks, &chunk_entry->index)))
        HGOTO_ERROR(H5E_DATASPACE, H5E_NOTFOUND, FAIL, "can't locate chunk in skip list")

    if ((extent_npoints = H5S_GET_EXTENT_NPOINTS(chunk_info->fspace)) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTCOUNT, FAIL, "dataspace is invalid")
    true_chunk_size = (hsize_t) extent_npoints * type_info->src_type_size;

    /* If the size of the filtered chunk is larger than the number of points in the
     * chunk file space extent times the datatype size, allocate enough space to hold the
     * whole filtered chunk. Otherwise, allocate a buffer equal to the size of the
     * chunk so that the unfiltering operation doesn't have to grow the buffer.
     */
    buf_size = MAX(chunk_entry->chunk_states.chunk_current.length, true_chunk_size);

    if (NULL == (chunk_entry->buf = H5MM_malloc(buf_size)))
        HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk data buffer")

    /* If this is not a full chunk overwrite or this is a read operation, the chunk must be
     * read from the file and unfiltered.
     */
    if(!chunk_entry->full_overwrite || io_info->op_type == H5D_IO_OP_READ) {
        H5FD_mpio_xfer_t xfer_mode; /* Parallel transfer for this request */

        chunk_entry->chunk_states.new_chunk.length = chunk_entry->chunk_states.chunk_current.length;

        /* Currently, these chunk reads are done independently and will likely
         * cause issues with collective metadata reads enabled. In the future,
         * this should be refactored to use collective chunk reads - JTH */

        /* Get the original state of parallel I/O transfer mode */
        if(H5CX_get_io_xfer_mode(&xfer_mode) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get MPI-I/O transfer mode")

        /* Change the xfer_mode to independent for handling the I/O */
        if(H5CX_set_io_xfer_mode(H5FD_MPIO_INDEPENDENT) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTSET, FAIL, "can't set MPI-I/O transfer mode")

        if(H5F_block_read(io_info->dset->oloc.file, H5FD_MEM_DRAW, chunk_entry->chunk_states.chunk_current.offset,
                chunk_entry->chunk_states.new_chunk.length, chunk_entry->buf) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "unable to read raw data chunk")

        /* Return to the original I/O transfer mode setting */
        if(H5CX_set_io_xfer_mode(xfer_mode) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTSET, FAIL, "can't set MPI-I/O transfer mode")

        if(H5Z_pipeline(&io_info->dset->shared->dcpl_cache.pline, H5Z_FLAG_REVERSE,
                &filter_mask, err_detect, filter_cb, (size_t *)&chunk_entry->chunk_states.new_chunk.length,
                &buf_size, &chunk_entry->buf) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTFILTER, FAIL, "couldn't unfilter chunk for modifying")
    } /* end if */
    else {
        chunk_entry->chunk_states.new_chunk.length = true_chunk_size;
    } /* end else */

    /* Initialize iterator for memory selection */
    if (NULL == (mem_iter = (H5S_sel_iter_t *) H5MM_malloc(sizeof(H5S_sel_iter_t))))
        HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate memory iterator")

    if(H5S_select_iter_init(mem_iter, chunk_info->mspace, type_info->src_type_size) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "unable to initialize memory selection information")
    mem_iter_init = TRUE;

    /* If this is a read operation, scatter the read chunk data to the user's buffer.
     *
     * If this is a write operation, update the chunk data buffer with the modifications
     * from the current process, then apply any modifications from other processes. Finally,
     * filter the newly-updated chunk.
     */
    switch(io_info->op_type) {
        case H5D_IO_OP_READ:
            if(NULL == (file_iter = (H5S_sel_iter_t *) H5MM_malloc(sizeof(H5S_sel_iter_t))))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate file iterator")

            if(H5S_select_iter_init(file_iter, chunk_info->fspace, type_info->src_type_size) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "unable to initialize memory selection information")
            file_iter_init = TRUE;

            if((iter_nelmts = H5S_GET_SELECT_NPOINTS(chunk_info->fspace)) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTCOUNT, FAIL, "dataspace is invalid")

            if(NULL == (tmp_gath_buf = H5MM_malloc((hsize_t) iter_nelmts * type_info->src_type_size)))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate temporary gather buffer")

            if(!H5D__gather_mem(chunk_entry->buf, chunk_info->fspace, file_iter, (size_t) iter_nelmts, tmp_gath_buf))
                HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "couldn't gather from chunk buffer")

            if((iter_nelmts = H5S_GET_SELECT_NPOINTS(chunk_info->mspace)) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTCOUNT, FAIL, "dataspace is invalid")

            if(H5D__scatter_mem(tmp_gath_buf, chunk_info->mspace, mem_iter, (size_t) iter_nelmts, io_info->u.rbuf) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "couldn't scatter to read buffer")

            break;

        case H5D_IO_OP_WRITE:
            if((iter_nelmts = H5S_GET_SELECT_NPOINTS(chunk_info->mspace)) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTCOUNT, FAIL, "dataspace is invalid")

            if(NULL == (tmp_gath_buf = H5MM_malloc((hsize_t) iter_nelmts * type_info->src_type_size)))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate temporary gather buffer")

            /* Gather modification data from the application write buffer into a temporary buffer */
            if(!H5D__gather_mem(io_info->u.wbuf, chunk_info->mspace, mem_iter, (size_t) iter_nelmts, tmp_gath_buf))
                HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "couldn't gather from write buffer")

            if(H5S_SELECT_ITER_RELEASE(mem_iter) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release selection iterator")
            mem_iter_init = FALSE;

            /* Initialize iterator for file selection */
            if(H5S_select_iter_init(mem_iter, chunk_info->fspace, type_info->dst_type_size) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "unable to initialize file selection information")
            mem_iter_init = TRUE;

            if((iter_nelmts = H5S_GET_SELECT_NPOINTS(chunk_info->fspace)) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTCOUNT, FAIL, "dataspace is invalid")

            /* Scatter the owner's modification data into the chunk data buffer according to
             * the file space.
             */
            if(H5D__scatter_mem(tmp_gath_buf, chunk_info->fspace, mem_iter, (size_t) iter_nelmts, chunk_entry->buf) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "couldn't scatter to chunk data buffer")

            if(H5S_SELECT_ITER_RELEASE(mem_iter) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release selection iterator")
            mem_iter_init = FALSE;

            if(MPI_SUCCESS != (mpi_code = MPI_Waitall(chunk_entry->async_info.num_receive_requests,
                    chunk_entry->async_info.receive_requests_array, MPI_STATUSES_IGNORE)))
                HMPI_GOTO_ERROR(FAIL, "MPI_Waitall failed", mpi_code)

            /* For each asynchronous receive call previously posted, receive the chunk modification
             * buffer from another rank and update the chunk data
             */
            for(i = 0; i < (size_t) chunk_entry->async_info.num_receive_requests; i++) {
                const unsigned char *mod_data_p;

                /* Decode the process' chunk file dataspace */
                mod_data_p = chunk_entry->async_info.receive_buffer_array[i];
                if(NULL == (dataspace = H5S_decode(&mod_data_p)))
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTDECODE, FAIL, "unable to decode dataspace")

                if(H5S_select_iter_init(mem_iter, dataspace, type_info->dst_type_size) < 0)
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "unable to initialize memory selection information")
                mem_iter_init = TRUE;

                if((iter_nelmts = H5S_GET_SELECT_NPOINTS(dataspace)) < 0)
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTCOUNT, FAIL, "dataspace is invalid")

                /* Update the chunk data with the received modification data */
                if(H5D__scatter_mem(mod_data_p, dataspace, mem_iter, (size_t) iter_nelmts, chunk_entry->buf) < 0)
                    HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "couldn't scatter to write buffer")

                if(H5S_SELECT_ITER_RELEASE(mem_iter) < 0)
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release selection iterator")
                mem_iter_init = FALSE;
                if(dataspace) {
                    if(H5S_close(dataspace) < 0)
                        HGOTO_ERROR(H5E_DATASPACE, H5E_CANTFREE, FAIL, "can't close dataspace")
                    dataspace = NULL;
                } /* end if */
                H5MM_free(chunk_entry->async_info.receive_buffer_array[i]);
            } /* end for */

            /* Filter the chunk */
            if(H5Z_pipeline(&io_info->dset->shared->dcpl_cache.pline, 0, &filter_mask,
                    err_detect, filter_cb, (size_t *)&chunk_entry->chunk_states.new_chunk.length,
                    &buf_size, &chunk_entry->buf) < 0)
                HGOTO_ERROR(H5E_PLINE, H5E_CANTFILTER, FAIL, "output pipeline failed")

#if H5_SIZEOF_SIZE_T > 4
            /* Check for the chunk expanding too much to encode in a 32-bit value */
            if (chunk_entry->chunk_states.new_chunk.length > ((size_t) 0xffffffff))
                HGOTO_ERROR(H5E_DATASET, H5E_BADRANGE, FAIL, "chunk too large for 32-bit length")
#endif
            break;

        default:
            HGOTO_ERROR(H5E_DATASET, H5E_BADVALUE, FAIL, "invalid I/O operation")
    } /* end switch */

done:
    if (chunk_entry->async_info.receive_buffer_array)
        H5MM_free(chunk_entry->async_info.receive_buffer_array);
    if (chunk_entry->async_info.receive_requests_array)
        H5MM_free(chunk_entry->async_info.receive_requests_array);
    if (tmp_gath_buf)
        H5MM_free(tmp_gath_buf);
    if (file_iter_init && H5S_SELECT_ITER_RELEASE(file_iter) < 0)
        HDONE_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release selection iterator")
    if (file_iter)
        H5MM_free(file_iter);
    if (mem_iter_init && H5S_SELECT_ITER_RELEASE(mem_iter) < 0)
        HDONE_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release selection iterator")
    if (mem_iter)
        H5MM_free(mem_iter);
    if (dataspace)
        if (H5S_close(dataspace) < 0)
            HDONE_ERROR(H5E_DATASPACE, H5E_CANTFREE, FAIL, "can't close dataspace")

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__filtered_collective_chunk_entry_io() */
#endif  /* H5_HAVE_PARALLEL */