summaryrefslogtreecommitdiffstats
path: root/src/H5Dmpio.c
blob: 3bddf0b07dd82b055eb1da8f43c4955297e30adc (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
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
3809
3810
3811
3812
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
3872
3873
3874
3875
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
3954
3955
3956
3957
3958
3959
3960
3961
3962
3963
3964
3965
3966
3967
3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
4010
4011
4012
4013
4014
4015
4016
4017
4018
4019
4020
4021
4022
4023
4024
4025
4026
4027
4028
4029
4030
4031
4032
4033
4034
4035
4036
4037
4038
4039
4040
4041
4042
4043
4044
4045
4046
4047
4048
4049
4050
4051
4052
4053
4054
4055
4056
4057
4058
4059
4060
4061
4062
4063
4064
4065
4066
4067
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
4080
4081
4082
4083
4084
4085
4086
4087
4088
4089
4090
4091
4092
4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
4141
4142
4143
4144
4145
4146
4147
4148
4149
4150
4151
4152
4153
4154
4155
4156
4157
4158
4159
4160
4161
4162
4163
4164
4165
4166
4167
4168
4169
4170
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
4181
4182
4183
4184
4185
4186
4187
4188
4189
4190
4191
4192
4193
4194
4195
4196
4197
4198
4199
4200
4201
4202
4203
4204
4205
4206
4207
4208
4209
4210
4211
4212
4213
4214
4215
4216
4217
4218
4219
4220
4221
4222
4223
4224
4225
4226
4227
4228
4229
4230
4231
4232
4233
4234
4235
4236
4237
4238
4239
4240
4241
4242
4243
4244
4245
4246
4247
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
4263
4264
4265
4266
4267
4268
4269
4270
4271
4272
4273
4274
4275
4276
4277
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
4290
4291
4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
4329
4330
4331
4332
4333
4334
4335
4336
4337
4338
4339
4340
4341
4342
4343
4344
4345
4346
4347
4348
4349
4350
4351
4352
4353
4354
4355
4356
4357
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367
4368
4369
4370
4371
4372
4373
4374
4375
4376
4377
4378
4379
4380
4381
4382
4383
4384
4385
4386
4387
4388
4389
4390
4391
4392
4393
4394
4395
4396
4397
4398
4399
4400
4401
4402
4403
4404
4405
4406
4407
4408
4409
4410
4411
4412
4413
4414
4415
4416
4417
4418
4419
4420
4421
4422
4423
4424
4425
4426
4427
4428
4429
4430
4431
4432
4433
4434
4435
4436
4437
4438
4439
4440
4441
4442
4443
4444
4445
4446
4447
4448
4449
4450
4451
4452
4453
4454
4455
4456
4457
4458
4459
4460
4461
4462
4463
4464
4465
4466
4467
4468
4469
4470
4471
4472
4473
4474
4475
4476
4477
4478
4479
4480
4481
4482
4483
4484
4485
4486
4487
4488
4489
4490
4491
4492
4493
4494
4495
4496
4497
4498
4499
4500
4501
4502
4503
4504
4505
4506
4507
4508
4509
4510
4511
4512
4513
4514
4515
4516
4517
4518
4519
4520
4521
4522
4523
4524
4525
4526
4527
4528
4529
4530
4531
4532
4533
4534
4535
4536
4537
4538
4539
4540
4541
4542
4543
4544
4545
4546
4547
4548
4549
4550
4551
4552
4553
4554
4555
4556
4557
4558
4559
4560
4561
4562
4563
4564
4565
4566
4567
4568
4569
4570
4571
4572
4573
4574
4575
4576
4577
4578
4579
4580
4581
4582
4583
4584
4585
4586
4587
4588
4589
4590
4591
4592
4593
4594
4595
4596
4597
4598
4599
4600
4601
4602
4603
4604
4605
4606
4607
4608
4609
4610
4611
4612
4613
4614
4615
4616
4617
4618
4619
4620
4621
4622
4623
4624
4625
4626
4627
4628
4629
4630
4631
4632
4633
4634
4635
4636
4637
4638
4639
4640
4641
4642
4643
4644
4645
4646
4647
4648
4649
4650
4651
4652
4653
4654
4655
4656
4657
4658
4659
4660
4661
4662
4663
4664
4665
4666
4667
4668
4669
4670
4671
4672
4673
4674
4675
4676
4677
4678
4679
4680
4681
4682
4683
4684
4685
4686
4687
4688
4689
4690
4691
4692
4693
4694
4695
4696
4697
4698
4699
4700
4701
4702
4703
4704
4705
4706
4707
4708
4709
4710
4711
4712
4713
4714
4715
4716
4717
4718
4719
4720
4721
4722
4723
4724
4725
4726
4727
4728
4729
4730
4731
4732
4733
4734
4735
4736
4737
4738
4739
4740
4741
4742
4743
4744
4745
4746
4747
4748
4749
4750
4751
4752
4753
4754
4755
4756
4757
4758
4759
4760
4761
4762
4763
4764
4765
4766
4767
4768
4769
4770
4771
4772
4773
4774
4775
4776
4777
4778
4779
4780
4781
4782
4783
4784
4785
4786
4787
4788
4789
4790
4791
4792
4793
4794
4795
4796
4797
4798
4799
4800
4801
4802
4803
4804
4805
4806
4807
4808
4809
4810
4811
4812
4813
4814
4815
4816
4817
4818
4819
4820
4821
4822
4823
4824
4825
4826
4827
4828
4829
4830
4831
4832
4833
4834
4835
4836
4837
4838
4839
4840
4841
4842
4843
4844
4845
4846
4847
4848
4849
4850
4851
4852
4853
4854
4855
4856
4857
4858
4859
4860
4861
4862
4863
4864
4865
4866
4867
4868
4869
4870
4871
4872
4873
4874
4875
4876
4877
4878
4879
4880
4881
4882
4883
4884
4885
4886
4887
4888
4889
4890
4891
4892
4893
4894
4895
4896
4897
4898
4899
4900
4901
4902
4903
4904
4905
4906
4907
4908
4909
4910
4911
4912
4913
4914
4915
4916
4917
4918
4919
4920
4921
4922
4923
4924
4925
4926
4927
4928
4929
4930
4931
4932
4933
4934
4935
4936
4937
4938
4939
4940
4941
4942
4943
4944
4945
4946
4947
4948
4949
4950
4951
4952
4953
4954
4955
4956
4957
4958
4959
4960
4961
4962
4963
4964
4965
4966
4967
4968
4969
4970
4971
4972
4973
4974
4975
4976
4977
4978
4979
4980
4981
4982
4983
4984
4985
4986
4987
4988
4989
4990
4991
4992
4993
4994
4995
4996
4997
4998
4999
5000
5001
5002
5003
5004
5005
5006
5007
5008
5009
5010
5011
5012
5013
5014
5015
5016
5017
5018
5019
5020
5021
5022
5023
5024
5025
5026
5027
5028
5029
5030
5031
5032
5033
5034
5035
5036
5037
5038
5039
5040
5041
5042
5043
5044
5045
5046
5047
5048
5049
5050
5051
5052
5053
5054
5055
5056
5057
5058
5059
5060
5061
5062
5063
5064
5065
5066
5067
5068
5069
5070
5071
5072
5073
5074
5075
5076
5077
5078
5079
5080
5081
5082
5083
5084
5085
5086
5087
5088
5089
5090
5091
5092
5093
5094
5095
5096
5097
5098
5099
5100
5101
5102
5103
5104
5105
5106
5107
5108
5109
5110
5111
5112
5113
5114
5115
5116
5117
5118
5119
5120
5121
5122
5123
5124
5125
5126
5127
5128
5129
5130
5131
5132
5133
5134
5135
5136
5137
5138
5139
5140
5141
5142
5143
5144
5145
5146
5147
5148
5149
5150
5151
5152
5153
5154
5155
5156
5157
5158
5159
5160
5161
5162
5163
5164
5165
5166
5167
5168
5169
5170
5171
5172
5173
5174
5175
5176
5177
5178
5179
5180
5181
5182
5183
5184
5185
5186
5187
5188
5189
5190
5191
5192
5193
5194
5195
5196
5197
5198
5199
5200
5201
5202
5203
5204
5205
5206
5207
5208
5209
5210
5211
5212
5213
5214
5215
5216
5217
5218
5219
5220
5221
5222
5223
5224
5225
5226
5227
5228
5229
5230
5231
5232
5233
5234
5235
5236
5237
5238
5239
5240
5241
5242
5243
5244
5245
5246
5247
5248
5249
5250
5251
5252
5253
5254
5255
5256
5257
5258
5259
5260
5261
5262
5263
5264
5265
5266
5267
5268
5269
5270
5271
5272
5273
5274
5275
5276
5277
5278
5279
5280
5281
5282
5283
5284
5285
5286
5287
5288
5289
5290
5291
5292
5293
5294
5295
5296
5297
5298
5299
5300
5301
5302
5303
5304
5305
5306
5307
5308
5309
5310
5311
5312
5313
5314
5315
5316
5317
5318
5319
5320
5321
5322
5323
5324
5325
5326
5327
5328
5329
5330
5331
5332
5333
5334
5335
5336
5337
5338
5339
5340
5341
5342
5343
5344
5345
5346
5347
5348
5349
5350
5351
5352
5353
5354
5355
5356
5357
5358
5359
5360
5361
5362
5363
5364
5365
5366
5367
5368
5369
5370
5371
5372
5373
5374
5375
5376
5377
5378
5379
5380
5381
5382
5383
5384
5385
5386
5387
5388
5389
5390
5391
5392
5393
5394
5395
5396
5397
5398
5399
5400
5401
5402
5403
5404
5405
5406
5407
5408
5409
5410
5411
5412
5413
5414
5415
5416
5417
5418
5419
5420
5421
5422
5423
5424
5425
5426
5427
5428
5429
5430
5431
5432
5433
5434
5435
5436
5437
5438
5439
5440
5441
5442
5443
5444
5445
5446
5447
5448
5449
5450
5451
5452
5453
5454
5455
5456
5457
5458
5459
5460
5461
5462
5463
5464
5465
5466
5467
5468
5469
5470
5471
5472
5473
5474
5475
5476
5477
5478
5479
5480
5481
5482
5483
5484
5485
5486
5487
5488
5489
5490
5491
5492
5493
5494
5495
5496
5497
5498
5499
5500
5501
5502
5503
5504
5505
5506
5507
5508
5509
5510
5511
5512
5513
5514
5515
5516
5517
5518
5519
5520
5521
5522
5523
5524
5525
5526
5527
5528
5529
5530
5531
5532
5533
5534
5535
5536
5537
5538
5539
5540
5541
5542
5543
5544
5545
5546
5547
5548
5549
5550
5551
5552
5553
5554
5555
5556
5557
5558
5559
5560
5561
5562
5563
5564
5565
5566
5567
5568
5569
5570
5571
5572
5573
5574
5575
5576
5577
5578
5579
5580
5581
5582
5583
5584
5585
5586
5587
5588
5589
5590
5591
5592
5593
5594
5595
5596
5597
5598
5599
5600
5601
5602
5603
5604
5605
5606
5607
5608
5609
5610
5611
5612
5613
5614
5615
5616
5617
5618
5619
5620
5621
5622
5623
5624
5625
5626
5627
5628
5629
5630
5631
5632
5633
5634
5635
5636
5637
5638
5639
5640
5641
5642
5643
5644
5645
5646
5647
5648
5649
5650
5651
5652
5653
5654
5655
5656
5657
5658
5659
5660
5661
5662
5663
5664
5665
5666
5667
5668
5669
5670
5671
5672
5673
5674
5675
5676
5677
5678
5679
5680
5681
5682
5683
5684
5685
5686
5687
5688
5689
5690
5691
5692
5693
5694
5695
5696
5697
5698
5699
5700
5701
5702
5703
5704
5705
5706
5707
5708
5709
5710
5711
5712
5713
5714
5715
5716
5717
5718
5719
5720
5721
5722
5723
5724
5725
5726
5727
5728
5729
5730
5731
5732
5733
5734
5735
5736
5737
5738
5739
5740
5741
5742
5743
5744
5745
5746
5747
5748
5749
5750
5751
5752
5753
5754
5755
5756
5757
5758
5759
5760
5761
5762
5763
5764
5765
5766
5767
5768
5769
5770
5771
5772
5773
5774
5775
5776
5777
5778
5779
5780
5781
5782
5783
5784
5785
5786
5787
5788
5789
5790
5791
5792
5793
5794
5795
5796
5797
5798
5799
5800
5801
5802
5803
5804
5805
5806
5807
5808
5809
5810
5811
5812
5813
5814
5815
5816
5817
5818
5819
5820
5821
5822
5823
5824
5825
5826
5827
5828
5829
5830
5831
5832
5833
5834
5835
5836
5837
5838
5839
5840
5841
5842
5843
5844
5845
5846
5847
5848
5849
5850
5851
5852
5853
5854
5855
5856
5857
5858
5859
5860
5861
5862
5863
5864
5865
5866
5867
5868
5869
5870
5871
5872
5873
5874
5875
5876
5877
5878
5879
5880
5881
5882
5883
5884
5885
5886
5887
5888
5889
5890
5891
5892
5893
5894
5895
5896
5897
5898
5899
5900
5901
5902
5903
5904
5905
5906
5907
5908
5909
5910
5911
5912
5913
5914
5915
5916
5917
5918
/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
 * 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://www.hdfgroup.org/licenses.               *
 * 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 "H5FLprivate.h" /* Free Lists        */
#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
#define H5D_NO_IO                      4

/***** 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

/*
 * Threshold value for redistributing shared filtered chunks
 * on all MPI ranks, or just MPI rank 0
 */
#define H5D_CHUNK_REDISTRIBUTE_THRES ((size_t)((25 * H5_MB) / sizeof(H5D_chunk_redistribute_info_t)))

/*
 * Initial allocation size for the arrays that hold
 * buffers for chunk modification data that is sent
 * to other ranks and the MPI_Request objects for
 * those send operations
 */
#define H5D_CHUNK_NUM_SEND_MSGS_INIT 64

/*
 * Define a tag value for the MPI messages sent/received for
 * chunk modification data
 */
#define H5D_CHUNK_MOD_DATA_TAG 64

/*
 * Macro to initialize a H5D_chk_idx_info_t
 * structure, given a pointer to a H5D_io_info_t
 * structure
 */
#define H5D_MPIO_INIT_CHUNK_IDX_INFO(index_info, dset)                                                       \
    do {                                                                                                     \
        index_info.f       = (dset)->oloc.file;                                                              \
        index_info.pline   = &((dset)->shared->dcpl_cache.pline);                                            \
        index_info.layout  = &((dset)->shared->layout.u.chunk);                                              \
        index_info.storage = &((dset)->shared->layout.storage.u.chunk);                                      \
    } while (0)

/*
 * Macro to initialize a H5D_chunk_ud_t structure
 * given a pointer to a H5D_chk_idx_info_t structure
 */
#define H5D_MPIO_INIT_CHUNK_UD_INFO(chunk_ud, index_info_ptr)                                                \
    do {                                                                                                     \
        HDmemset(&chunk_ud, 0, sizeof(H5D_chunk_ud_t));                                                      \
        chunk_ud.common.layout  = (index_info_ptr)->layout;                                                  \
        chunk_ud.common.storage = (index_info_ptr)->storage;                                                 \
    } while (0)

/******************/
/* Local Typedefs */
/******************/

/* Combine chunk/piece address and chunk/piece info into a struct for
 * better performance. */
typedef struct H5D_chunk_addr_info_t {
    /* piece for multi-dset */
    haddr_t          piece_addr;
    H5D_piece_info_t piece_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 necessary for re-allocating file space for a chunk
 * during a parallel write of a chunked dataset with filters
 * applied.
 */
typedef struct H5D_chunk_alloc_info_t {
    H5F_block_t chunk_current;
    H5F_block_t chunk_new;
    hsize_t     chunk_idx;
} H5D_chunk_alloc_info_t;

/*
 * Information for a chunk pertaining to the dataset's chunk
 * index entry for the chunk
 */
typedef struct H5D_chunk_index_info_t {
    hsize_t  chunk_idx;
    unsigned filter_mask;
    hbool_t  need_insert;
} H5D_chunk_index_info_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__mpio_collective_filtered_chunk_io_setup(). This
 * struct's fields are as follows:
 *
 * index_info - A structure containing the information needed when collectively
 *              re-inserting the chunk into the dataset's chunk index. The structure
 *              is distributed to all ranks during the re-insertion operation. Its fields
 *              are as follows:
 *
 *     chunk_idx - The index of the chunk in the dataset's chunk index.
 *
 *     filter_mask - A bit-mask that indicates which filters are to be applied to the
 *                   chunk. Each filter in a chunk's filter pipeline has a bit position
 *                   that can be masked to disable that particular filter for the chunk.
 *                   This filter mask is saved alongside the chunk in the file.
 *
 *     need_insert - A flag which determines whether or not a chunk needs to be re-inserted into
 *                   the chunk index after the write operation.
 *
 * chunk_info - A pointer to the chunk's H5D_piece_info_t structure, which contains useful
 *              information like the dataspaces containing the selection in the chunk.
 *
 * 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 MPI ranks and as a parameter to the chunk file
 *                 space reallocation function.
 *
 * chunk_new - 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.
 *
 * need_read - A flag which determines whether or not a chunk needs to be read from the
 *             file. During writes, 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. During reads, this field should generally
 *             be true, but may be false if the chunk isn't allocated, for example.
 *
 * skip_filter_pline - A flag which determines whether to skip calls to the filter pipeline
 *                     for this chunk. This flag is mostly useful for correct handling of
 *                     partial edge chunks when the "don't filter partial edge chunks" flag
 *                     is set on the dataset's DCPL.
 *
 * 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 MPI rank which has the chunk selected and is used to
 *           determine the value for the previous `full_overwrite` flag.
 *
 * chunk_buf_size - The size in bytes of the data buffer allocated for the chunk
 *
 * orig_owner - The MPI rank 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 MPI ranks.
 *
 * new_owner - The MPI rank which has been selected to perform the modifications to this chunk.
 *
 * num_writers - The total number of MPI ranks writing to this chunk. This field is used when
 *               the new owner of a chunk is receiving messages from other MPI ranks that
 *               contain their selections in the chunk and the data to update the chunk with.
 *               The new owner must know how many MPI ranks it should expect messages from so
 *               that it can post an equal number of receive calls.
 *
 * 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.
 *
 * hh - A handle for hash tables provided by the uthash.h header
 *
 */
typedef struct H5D_filtered_collective_io_info_t {
    H5D_chunk_index_info_t index_info;

    H5D_piece_info_t *chunk_info;
    H5F_block_t       chunk_current;
    H5F_block_t       chunk_new;
    hbool_t           need_read;
    hbool_t           skip_filter_pline;
    size_t            io_size;
    size_t            chunk_buf_size;
    int               orig_owner;
    int               new_owner;
    int               num_writers;
    void             *buf;

    UT_hash_handle hh;
} H5D_filtered_collective_io_info_t;

/*
 * Information necessary for redistributing shared chunks during
 * a parallel write of a chunked dataset with filters applied.
 */
typedef struct H5D_chunk_redistribute_info_t {
    H5F_block_t chunk_block;
    hsize_t     chunk_idx;
    int         orig_owner;
    int         new_owner;
    int         num_writers;
} H5D_chunk_redistribute_info_t;

/*
 * Information used when re-inserting a chunk into a dataset's
 * chunk index during a parallel write of a chunked dataset with
 * filters applied.
 */
typedef struct H5D_chunk_insert_info_t {
    H5F_block_t            chunk_block;
    H5D_chunk_index_info_t index_info;
} H5D_chunk_insert_info_t;

/********************/
/* Local Prototypes */
/********************/
static herr_t H5D__piece_io(H5D_io_info_t *io_info);
static herr_t H5D__multi_chunk_collective_io(H5D_io_info_t *io_info, H5D_dset_io_info_t *dset_info,
                                             int mpi_rank, int mpi_size);
static herr_t H5D__multi_chunk_filtered_collective_io(H5D_io_info_t *io_info, H5D_dset_io_info_t *dset_info,
                                                      int mpi_rank, int mpi_size);
static herr_t H5D__link_piece_collective_io(H5D_io_info_t *io_info, int mpi_rank);
static herr_t H5D__link_chunk_filtered_collective_io(H5D_io_info_t *io_info, H5D_dset_io_info_t *dset_info,
                                                     int mpi_rank, int mpi_size);
static herr_t H5D__inter_collective_io(H5D_io_info_t *io_info, const H5D_dset_io_info_t *di,
                                       H5S_t *file_space, H5S_t *mem_space);
static herr_t H5D__final_collective_io(H5D_io_info_t *io_info, hsize_t mpi_buf_count,
                                       MPI_Datatype mpi_file_type, MPI_Datatype mpi_buf_type);
static herr_t H5D__obtain_mpio_mode(H5D_io_info_t *io_info, H5D_dset_io_info_t *di, uint8_t assign_io_mode[],
                                    haddr_t chunk_addr[], int mpi_rank, int mpi_size);
static herr_t H5D__mpio_get_sum_chunk(const H5D_io_info_t *io_info, int *sum_chunkf);
static herr_t H5D__mpio_get_sum_chunk_dset(const H5D_io_info_t *io_info, const H5D_dset_io_info_t *dset_info,
                                           int *sum_chunkf);
static herr_t H5D__mpio_collective_filtered_chunk_io_setup(const H5D_io_info_t                *io_info,
                                                           const H5D_dset_io_info_t           *di,
                                                           H5D_filtered_collective_io_info_t **chunk_list,
                                                           size_t *num_entries, int mpi_rank);
static herr_t H5D__mpio_redistribute_shared_chunks(H5D_filtered_collective_io_info_t *chunk_list,
                                                   size_t                             chunk_list_num_entries,
                                                   const H5D_io_info_t *io_info, int mpi_rank, int mpi_size,
                                                   size_t **rank_chunks_assigned_map);
static herr_t H5D__mpio_redistribute_shared_chunks_int(H5D_filtered_collective_io_info_t *chunk_list,
                                                       size_t              *num_chunks_assigned_map,
                                                       hbool_t              all_ranks_involved,
                                                       const H5D_io_info_t *io_info, int mpi_rank,
                                                       int mpi_size);
static herr_t H5D__mpio_share_chunk_modification_data(H5D_filtered_collective_io_info_t *chunk_list,
                                                      size_t *chunk_list_num_entries, H5D_io_info_t *io_info,
                                                      H5D_dset_io_info_t *dset_info, int mpi_rank,
                                                      int                                 mpi_size,
                                                      H5D_filtered_collective_io_info_t **chunk_hash_table,
                                                      unsigned char                    ***chunk_msg_bufs,
                                                      int                                *chunk_msg_bufs_len);
static herr_t H5D__mpio_collective_filtered_chunk_common_io(H5D_filtered_collective_io_info_t *chunk_list,
                                                            size_t               chunk_list_num_entries,
                                                            const H5D_io_info_t *io_info, int mpi_size);
static herr_t H5D__mpio_collective_filtered_chunk_read(H5D_filtered_collective_io_info_t *chunk_list,
                                                       size_t                    chunk_list_num_entries,
                                                       const H5D_io_info_t      *io_info,
                                                       const H5D_dset_io_info_t *di, int mpi_rank,
                                                       int mpi_size);
static herr_t H5D__mpio_collective_filtered_chunk_update(H5D_filtered_collective_io_info_t *chunk_list,
                                                         size_t chunk_list_num_entries,
                                                         H5D_filtered_collective_io_info_t *chunk_hash_table,
                                                         unsigned char                    **chunk_msg_bufs,
                                                         int chunk_msg_bufs_len, const H5D_io_info_t *io_info,
                                                         const H5D_dset_io_info_t *di, int mpi_rank,
                                                         int mpi_size);
static herr_t H5D__mpio_collective_filtered_chunk_reallocate(H5D_filtered_collective_io_info_t *chunk_list,
                                                             size_t              chunk_list_num_entries,
                                                             size_t             *num_chunks_assigned_map,
                                                             H5D_io_info_t      *io_info,
                                                             H5D_chk_idx_info_t *idx_info, int mpi_rank,
                                                             int mpi_size);
static herr_t H5D__mpio_collective_filtered_chunk_reinsert(H5D_filtered_collective_io_info_t *chunk_list,
                                                           size_t         chunk_list_num_entries,
                                                           size_t        *num_chunks_assigned_map,
                                                           H5D_io_info_t *io_info, H5D_dset_io_info_t *di,
                                                           H5D_chk_idx_info_t *idx_info, int mpi_rank,
                                                           int mpi_size);
static herr_t H5D__mpio_get_chunk_redistribute_info_types(MPI_Datatype *contig_type,
                                                          hbool_t      *contig_type_derived,
                                                          MPI_Datatype *resized_type,
                                                          hbool_t      *resized_type_derived);
static herr_t H5D__mpio_get_chunk_alloc_info_types(MPI_Datatype *contig_type, hbool_t *contig_type_derived,
                                                   MPI_Datatype *resized_type, hbool_t *resized_type_derived);
static herr_t H5D__mpio_get_chunk_insert_info_types(MPI_Datatype *contig_type, hbool_t *contig_type_derived,
                                                    MPI_Datatype *resized_type,
                                                    hbool_t      *resized_type_derived);
static herr_t H5D__mpio_collective_filtered_io_type(H5D_filtered_collective_io_info_t *chunk_list,
                                                    size_t num_entries, H5D_io_op_type_t op_type,
                                                    MPI_Datatype *new_mem_type, hbool_t *mem_type_derived,
                                                    MPI_Datatype *new_file_type, hbool_t *file_type_derived);
static int    H5D__cmp_piece_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);
static int    H5D__cmp_chunk_redistribute_info(const void *entry1, const void *entry2);
static int    H5D__cmp_chunk_redistribute_info_orig_owner(const void *entry1, const void *entry2);

#ifdef H5Dmpio_DEBUG
static herr_t H5D__mpio_debug_init(void);
static herr_t H5D__mpio_dump_collective_filtered_chunk_list(H5D_filtered_collective_io_info_t *chunk_list,
                                                            size_t chunk_list_num_entries, int mpi_rank);
#endif

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

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

/* Declare extern free list to manage the H5S_sel_iter_t struct */
H5FL_EXTERN(H5S_sel_iter_t);

#ifdef H5Dmpio_DEBUG

/* Flags to control debug actions in this file.
 * (Meant to be indexed by characters)
 *
 * These flags can be set with either (or both) the environment variable
 *      "H5D_mpio_Debug" set to a string containing one or more characters
 *      (flags) or by setting them as a string value for the
 *      "H5D_mpio_debug_key" MPI Info key.
 *
 * Supported characters in 'H5D_mpio_Debug' string:
 *      't' trace function entry and exit
 *      'f' log to file rather than debugging stream
 *      'm' show (rough) memory usage statistics
 *      'c' show critical timing information
 *
 *      To only show output from a particular MPI rank, specify its rank
 *      number as a character, e.g.:
 *
 *      '0' only show output from rank 0
 *
 *      To only show output from a particular range (up to 8 ranks supported
 *      between 0-9) of MPI ranks, specify the start and end ranks separated
 *      by a hyphen, e.g.:
 *
 *      '0-7' only show output from ranks 0 through 7
 *
 */
static int               H5D_mpio_debug_flags_s[256];
static int               H5D_mpio_debug_rank_s[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
static hbool_t           H5D_mpio_debug_inited    = FALSE;
static const char *const trace_in_pre             = "-> ";
static const char *const trace_out_pre            = "<- ";
static int               debug_indent             = 0;
static FILE             *debug_stream             = NULL;

/* Determine if this rank should output debugging info */
#define H5D_MPIO_DEBUG_THIS_RANK(rank)                                                                       \
    (H5D_mpio_debug_rank_s[0] < 0 || rank == H5D_mpio_debug_rank_s[0] || rank == H5D_mpio_debug_rank_s[1] || \
     rank == H5D_mpio_debug_rank_s[2] || rank == H5D_mpio_debug_rank_s[3] ||                                 \
     rank == H5D_mpio_debug_rank_s[4] || rank == H5D_mpio_debug_rank_s[5] ||                                 \
     rank == H5D_mpio_debug_rank_s[6] || rank == H5D_mpio_debug_rank_s[7])

/* Print some debugging string */
#define H5D_MPIO_DEBUG(rank, string)                                                                         \
    do {                                                                                                     \
        if (debug_stream && H5D_MPIO_DEBUG_THIS_RANK(rank)) {                                                \
            HDfprintf(debug_stream, "%*s(Rank %d) " string "\n", debug_indent, "", rank);                    \
            HDfflush(debug_stream);                                                                          \
        }                                                                                                    \
    } while (0)

/* Print some debugging string with printf-style arguments */
#define H5D_MPIO_DEBUG_VA(rank, string, ...)                                                                 \
    do {                                                                                                     \
        if (debug_stream && H5D_MPIO_DEBUG_THIS_RANK(rank)) {                                                \
            HDfprintf(debug_stream, "%*s(Rank %d) " string "\n", debug_indent, "", rank, __VA_ARGS__);       \
            HDfflush(debug_stream);                                                                          \
        }                                                                                                    \
    } while (0)

#define H5D_MPIO_TRACE_ENTER(rank)                                                                           \
    do {                                                                                                     \
        hbool_t trace_flag = H5D_mpio_debug_flags_s[(int)'t'];                                               \
                                                                                                             \
        if (trace_flag) {                                                                                    \
            H5D_MPIO_DEBUG_VA(rank, "%s%s", trace_in_pre, __func__);                                         \
            debug_indent += (int)HDstrlen(trace_in_pre);                                                     \
        }                                                                                                    \
    } while (0)

#define H5D_MPIO_TRACE_EXIT(rank)                                                                            \
    do {                                                                                                     \
        hbool_t trace_flag = H5D_mpio_debug_flags_s[(int)'t'];                                               \
                                                                                                             \
        if (trace_flag) {                                                                                    \
            debug_indent -= (int)HDstrlen(trace_out_pre);                                                    \
            H5D_MPIO_DEBUG_VA(rank, "%s%s", trace_out_pre, __func__);                                        \
        }                                                                                                    \
    } while (0)

#define H5D_MPIO_TIME_START(rank, op_name)                                                                   \
    {                                                                                                        \
        hbool_t           time_flag  = H5D_mpio_debug_flags_s[(int)'c'];                                     \
        double            start_time = 0.0, end_time = 0.0;                                                  \
        const char *const op = op_name;                                                                      \
                                                                                                             \
        if (time_flag) {                                                                                     \
            start_time = MPI_Wtime();                                                                        \
        }

#define H5D_MPIO_TIME_STOP(rank)                                                                             \
    if (time_flag) {                                                                                         \
        end_time = MPI_Wtime();                                                                              \
        H5D_MPIO_DEBUG_VA(rank, "'%s' took %f seconds", op, (end_time - start_time));                        \
    }                                                                                                        \
    }

/*---------------------------------------------------------------------------
 * Function:    H5D__mpio_parse_debug_str
 *
 * Purpose:     Parse a string for H5Dmpio-related debugging flags
 *
 * Returns:     N/A
 *
 *---------------------------------------------------------------------------
 */
static void
H5D__mpio_parse_debug_str(const char *s)
{
    FUNC_ENTER_PACKAGE_NOERR

    HDassert(s);

    while (*s) {
        int c = (int)(*s);

        if (c >= (int)'0' && c <= (int)'9') {
            hbool_t range = FALSE;

            if (*(s + 1) && *(s + 2))
                range = (int)*(s + 1) == '-' && (int)*(s + 2) >= (int)'0' && (int)*(s + 2) <= (int)'9';

            if (range) {
                int start_rank = c - (int)'0';
                int end_rank   = (int)*(s + 2) - '0';
                int num_ranks  = end_rank - start_rank + 1;
                int i;

                if (num_ranks > 8) {
                    end_rank  = start_rank + 7;
                    num_ranks = 8;
                }

                for (i = 0; i < num_ranks; i++)
                    H5D_mpio_debug_rank_s[i] = start_rank++;

                s += 3;
            }
            else
                H5D_mpio_debug_rank_s[0] = c - (int)'0';
        }
        else
            H5D_mpio_debug_flags_s[c]++;

        s++;
    }

    FUNC_LEAVE_NOAPI_VOID
}

static herr_t
H5D__mpio_debug_init(void)
{
    const char *debug_str;
    herr_t      ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE_NOERR

    HDassert(!H5D_mpio_debug_inited);

    /* Clear the debug flag buffer */
    HDmemset(H5D_mpio_debug_flags_s, 0, sizeof(H5D_mpio_debug_flags_s));

    /* Retrieve and parse the H5Dmpio debug string */
    debug_str = HDgetenv("H5D_mpio_Debug");
    if (debug_str)
        H5D__mpio_parse_debug_str(debug_str);

    if (H5DEBUG(D))
        debug_stream = H5DEBUG(D);

    H5D_mpio_debug_inited = TRUE;

    FUNC_LEAVE_NOAPI(ret_value)
}

#endif

/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_opt_possible
 *
 * Purpose:     Checks if an direct I/O transfer is possible between memory and
 *              the file.
 *
 *              This was derived from H5D__mpio_opt_possible for
 *              multi-dset work.
 *
 * Return:      Success:   Non-negative: TRUE or FALSE
 *              Failure:    Negative
 *
 *-------------------------------------------------------------------------
 */
htri_t
H5D__mpio_opt_possible(H5D_io_info_t *io_info)
{
    H5FD_mpio_xfer_t io_xfer_mode; /* MPI I/O transfer mode */
    size_t           i;
    H5D_t           *dset;
    const H5S_t     *file_space;
    const H5S_t     *mem_space;
    H5D_type_info_t *type_info;
    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 = TRUE; /* Return value */

    FUNC_ENTER_PACKAGE

    /* Check args */
    HDassert(io_info);

    for (i = 0; i < io_info->count; i++) {
        HDassert(io_info->dsets_info[i].file_space);
        HDassert(io_info->dsets_info[i].mem_space);
    }

    /* 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;

    for (i = 0; i < io_info->count; i++) {
        /* Check for skipped I/O */
        if (io_info->dsets_info[i].skip_io)
            continue;

        /* Set convenience pointers */
        dset       = io_info->dsets_info[i].dset;
        file_space = io_info->dsets_info[i].file_space;
        mem_space  = io_info->dsets_info[i].mem_space;
        type_info  = &io_info->dsets_info[i].type_info;

        /* 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 (!(dset->shared->layout.type == H5D_CONTIGUOUS || dset->shared->layout.type == H5D_CHUNKED))
            local_cause[0] |= H5D_MPIO_NOT_CONTIGUOUS_OR_CHUNKED_DATASET;

        /* check if external-file storage is used */
        if (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.
             */

#ifndef H5_HAVE_PARALLEL_FILTERED_WRITES
        /* Don't allow writes to filtered datasets if the functionality is disabled */
        if (io_info->op_type == H5D_IO_OP_WRITE && 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 contiguous datasets, currently */
        else if (H5D_CONTIGUOUS != 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 */
    }             /* end for loop */

    /* 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_get_no_coll_cause_strings
 *
 * Purpose:     When collective I/O is broken internally, it can be useful
 *              for users to see a representative string for the reason(s)
 *              why it was broken. This routine inspects the current
 *              "cause" flags from the API context and prints strings into
 *              the caller's buffers for the local and global reasons that
 *              collective I/O was broken.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 *-------------------------------------------------------------------------
 */
herr_t
H5D__mpio_get_no_coll_cause_strings(char *local_cause, size_t local_cause_len, char *global_cause,
                                    size_t global_cause_len)
{
    uint32_t local_no_coll_cause;
    uint32_t global_no_coll_cause;
    size_t   local_cause_bytes_written  = 0;
    size_t   global_cause_bytes_written = 0;
    int      nbits;
    herr_t   ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    HDassert((local_cause && local_cause_len > 0) || (global_cause && global_cause_len > 0));

    /*
     * Use compile-time assertion so this routine is updated
     * when any new "no collective cause" values are added
     */
    HDcompile_assert(H5D_MPIO_NO_COLLECTIVE_MAX_CAUSE == (H5D_mpio_no_collective_cause_t)256);

    /* Initialize output buffers */
    if (local_cause)
        *local_cause = '\0';
    if (global_cause)
        *global_cause = '\0';

    /* Retrieve the local and global cause flags from the API context */
    if (H5CX_get_mpio_local_no_coll_cause(&local_no_coll_cause) < 0)
        HGOTO_ERROR(H5E_CONTEXT, H5E_CANTGET, FAIL, "unable to get local no collective cause value")
    if (H5CX_get_mpio_global_no_coll_cause(&global_no_coll_cause) < 0)
        HGOTO_ERROR(H5E_CONTEXT, H5E_CANTGET, FAIL, "unable to get global no collective cause value")

    /*
     * Append each of the "reason for breaking collective I/O"
     * error messages to the local and global cause string buffers
     */
    nbits = 8 * sizeof(local_no_coll_cause);
    for (int bit_pos = 0; bit_pos < nbits; bit_pos++) {
        H5D_mpio_no_collective_cause_t cur_cause;
        const char                    *cause_str;
        size_t                         buf_space_left;

        cur_cause = (H5D_mpio_no_collective_cause_t)(1 << bit_pos);
        if (cur_cause == H5D_MPIO_NO_COLLECTIVE_MAX_CAUSE)
            break;

        switch (cur_cause) {
            case H5D_MPIO_SET_INDEPENDENT:
                cause_str = "independent I/O was requested";
                break;
            case H5D_MPIO_DATATYPE_CONVERSION:
                cause_str = "datatype conversions were required";
                break;
            case H5D_MPIO_DATA_TRANSFORMS:
                cause_str = "data transforms needed to be applied";
                break;
            case H5D_MPIO_MPI_OPT_TYPES_ENV_VAR_DISABLED:
                cause_str = "optimized MPI types flag wasn't set";
                break;
            case H5D_MPIO_NOT_SIMPLE_OR_SCALAR_DATASPACES:
                cause_str = "one of the dataspaces was neither simple nor scalar";
                break;
            case H5D_MPIO_NOT_CONTIGUOUS_OR_CHUNKED_DATASET:
                cause_str = "dataset was not contiguous or chunked";
                break;
            case H5D_MPIO_PARALLEL_FILTERED_WRITES_DISABLED:
                cause_str = "parallel writes to filtered datasets are disabled";
                break;
            case H5D_MPIO_ERROR_WHILE_CHECKING_COLLECTIVE_POSSIBLE:
                cause_str = "an error occurred while checking if collective I/O was possible";
                break;
            case H5D_MPIO_COLLECTIVE:
            case H5D_MPIO_NO_COLLECTIVE_MAX_CAUSE:
            default:
                HDassert(0 && "invalid no collective cause reason");
                break;
        }

        /*
         * Determine if the local reasons for breaking collective I/O
         * included the current cause
         */
        if (local_cause && (cur_cause & local_no_coll_cause)) {
            buf_space_left = local_cause_len - local_cause_bytes_written;

            /*
             * Check if there were any previous error messages included. If
             * so, prepend a semicolon to separate the messages.
             */
            if (buf_space_left && local_cause_bytes_written) {
                HDstrncat(local_cause, "; ", buf_space_left);
                local_cause_bytes_written += MIN(buf_space_left, 2);
                buf_space_left -= MIN(buf_space_left, 2);
            }

            if (buf_space_left) {
                HDstrncat(local_cause, cause_str, buf_space_left);
                local_cause_bytes_written += MIN(buf_space_left, HDstrlen(cause_str));
            }
        }

        /*
         * Determine if the global reasons for breaking collective I/O
         * included the current cause
         */
        if (global_cause && (cur_cause & global_no_coll_cause)) {
            buf_space_left = global_cause_len - global_cause_bytes_written;

            /*
             * Check if there were any previous error messages included. If
             * so, prepend a semicolon to separate the messages.
             */
            if (buf_space_left && global_cause_bytes_written) {
                HDstrncat(global_cause, "; ", buf_space_left);
                global_cause_bytes_written += MIN(buf_space_left, 2);
                buf_space_left -= MIN(buf_space_left, 2);
            }

            if (buf_space_left) {
                HDstrncat(global_cause, cause_str, buf_space_left);
                global_cause_bytes_written += MIN(buf_space_left, HDstrlen(cause_str));
            }
        }
    }

done:
    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__mpio_get_no_coll_cause_strings() */

/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_select_read
 *
 * Purpose:     MPI-IO function to read directly from app buffer to file.
 *
 *              This was referred from H5D__mpio_select_read for
 *              multi-dset work.
 *
 * Return:      non-negative on success, negative on failure.
 *
 *-------------------------------------------------------------------------
 */
herr_t
H5D__mpio_select_read(const H5D_io_info_t *io_info, hsize_t mpi_buf_count, H5S_t H5_ATTR_UNUSED *file_space,
                      H5S_t H5_ATTR_UNUSED *mem_space)
{
    void  *rbuf      = NULL;
    herr_t ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    /* memory addr from a piece with lowest file addr */
    rbuf = io_info->base_maddr.vp;

    /*OKAY: CAST DISCARDS CONST QUALIFIER*/
    H5_CHECK_OVERFLOW(mpi_buf_count, hsize_t, size_t);
    if (H5F_shared_block_read(io_info->f_sh, H5FD_MEM_DRAW, io_info->store_faddr, (size_t)mpi_buf_count,
                              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.
 *
 *              This was referred from H5D__mpio_select_write for
 *              multi-dset work.
 *
 * Return:      non-negative on success, negative on failure.
 *
 *-------------------------------------------------------------------------
 */
herr_t
H5D__mpio_select_write(const H5D_io_info_t *io_info, hsize_t mpi_buf_count, H5S_t H5_ATTR_UNUSED *file_space,
                       H5S_t H5_ATTR_UNUSED *mem_space)
{
    const void *wbuf      = NULL;
    herr_t      ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    /* memory addr from a piece with lowest file addr */
    wbuf = io_info->base_maddr.cvp;

    /*OKAY: CAST DISCARDS CONST QUALIFIER*/
    H5_CHECK_OVERFLOW(mpi_buf_count, hsize_t, size_t);
    if (H5F_shared_block_write(io_info->f_sh, H5FD_MEM_DRAW, io_info->store_faddr, (size_t)mpi_buf_count,
                               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_get_sum_chunk
 *
 * Purpose:     Routine for obtaining total number of chunks to cover
 *              hyperslab selection selected by all processors.  Operates
 *              on all datasets in the operation.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_get_sum_chunk(const H5D_io_info_t *io_info, 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_PACKAGE

    /* Get the number of chunks to perform I/O on */
    num_chunkf     = 0;
    ori_num_chunkf = io_info->pieces_added;
    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__mpio_get_sum_chunk_dset
 *
 * Purpose:     Routine for obtaining total number of chunks to cover
 *              hyperslab selection selected by all processors.  Operates
 *              on a single dataset.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_get_sum_chunk_dset(const H5D_io_info_t *io_info, const H5D_dset_io_info_t *dset_info,
                             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_PACKAGE

    /* Check for non-chunked dataset, in this case we know the number of "chunks"
     * is simply the mpi size */
    HDassert(dset_info->layout->type == H5D_CHUNKED);

    /* Get the number of chunks to perform I/O on */
    num_chunkf     = 0;
    ori_num_chunkf = H5SL_count(dset_info->layout_io_info.chunk_map->dset_sel_pieces);
    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_dset() */

/*-------------------------------------------------------------------------
 * Function:    H5D__piece_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__piece_io(H5D_io_info_t *io_info)
{
    H5FD_mpio_chunk_opt_t chunk_opt_mode;
#ifdef H5Dmpio_DEBUG
    hbool_t log_file_flag  = FALSE;
    FILE   *debug_log_file = NULL;
#endif
    int      io_option        = H5D_MULTI_CHUNK_IO_MORE_OPT;
    hbool_t  recalc_io_option = FALSE;
    hbool_t  use_multi_dset   = FALSE;
    unsigned one_link_chunk_io_threshold; /* Threshold to use single collective I/O for all chunks */
    int      sum_chunk = -1;
    int      mpi_rank;
    int      mpi_size;
    size_t   i;
    herr_t   ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    /* Sanity checks */
    HDassert(io_info);
    HDassert(io_info->using_mpi_vfd);
    HDassert(io_info->count > 0);

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

#ifdef H5Dmpio_DEBUG
    /* Initialize file-level debugging if not initialized */
    if (!H5D_mpio_debug_inited && H5D__mpio_debug_init() < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "can't initialize H5Dmpio debugging")

    /* Open file for debugging if necessary */
    log_file_flag = H5D_mpio_debug_flags_s[(int)'f'];
    if (log_file_flag) {
        char   debug_log_filename[1024];
        time_t time_now;

        HDsnprintf(debug_log_filename, 1024, "H5Dmpio_debug.rank%d", mpi_rank);

        if (NULL == (debug_log_file = HDfopen(debug_log_filename, "a")))
            HGOTO_ERROR(H5E_IO, H5E_OPENERROR, FAIL, "couldn't open debugging log file")

        /* Print a short header for this I/O operation */
        time_now = HDtime(NULL);
        HDfprintf(debug_log_file, "##### %s", HDasctime(HDlocaltime(&time_now)));

        debug_stream = debug_log_file;
    }
#endif

    /* Check the optional property list for the collective chunk IO optimization option.
     * Only set here if it's a static option, if it needs to be calculated using the
     * number of chunks per process delay that calculation until later. */
    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;
    else
        recalc_io_option = TRUE;

    /* Check if we can and should use multi dataset path */
    if (io_info->count > 1 && (io_option == H5D_ONE_LINK_CHUNK_IO || recalc_io_option)) {
        /* Use multi dataset path for now */
        use_multi_dset = TRUE;

        /* Check for filtered datasets */
        for (i = 0; i < io_info->count; i++)
            if (io_info->dsets_info[i].dset->shared->dcpl_cache.pline.nused > 0) {
                use_multi_dset = FALSE;
                break;
            }

        /* Check if this I/O exceeds one linked chunk threshold */
        if (recalc_io_option && use_multi_dset) {
            /* 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")

            /* If the threshold is 0, no need to check number of chunks */
            if (one_link_chunk_io_threshold > 0) {
                /* Get number of chunks for all processes */
                if (H5D__mpio_get_sum_chunk(io_info, &sum_chunk) < 0)
                    HGOTO_ERROR(H5E_DATASPACE, H5E_CANTSWAP, FAIL,
                                "unable to obtain the total chunk number of all processes");

                /* If the average number of chunk per process is less than the threshold, we will do multi
                 * chunk IO.  If this threshold is not exceeded for all datasets, no need to check it again
                 * for each individual dataset. */
                if ((unsigned)sum_chunk / (unsigned)mpi_size < one_link_chunk_io_threshold) {
                    recalc_io_option = FALSE;
                    use_multi_dset   = FALSE;
                }
            }
        }

        /* Perform multi dataset I/O if appropriate */
        if (use_multi_dset) {
#ifdef H5_HAVE_INSTRUMENTED_LIBRARY
            /*** Set collective chunk user-input optimization API. ***/
            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 */
#endif        /* H5_HAVE_INSTRUMENTED_LIBRARY */

            /* Perform unfiltered link chunk collective IO */
            if (H5D__link_piece_collective_io(io_info, mpi_rank) < 0)
                HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish linked chunk MPI-IO")
        }
    }

    if (!use_multi_dset) {
        /* Loop over datasets */
        for (i = 0; i < io_info->count; i++) {
            if (io_info->dsets_info[i].layout->type == H5D_CONTIGUOUS) {
                /* Contiguous: call H5D__inter_collective_io() directly */
                H5D_mpio_actual_io_mode_t actual_io_mode = H5D_MPIO_CONTIGUOUS_COLLECTIVE;

                io_info->store_faddr = io_info->dsets_info[i].store->contig.dset_addr;
                io_info->base_maddr  = io_info->dsets_info[i].buf;

                if (H5D__inter_collective_io(io_info, &io_info->dsets_info[i],
                                             io_info->dsets_info[i].file_space,
                                             io_info->dsets_info[i].mem_space) < 0)
                    HGOTO_ERROR(H5E_IO, H5E_CANTGET, 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);
            }
            else {
                /* Chunked I/O path */
                HDassert(io_info->dsets_info[i].layout->type == H5D_CHUNKED);

                /* Recalculate io_option if necessary */
                if (recalc_io_option) {
                    /* 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")

                    /* If the threshold is 0, no need to check number of chunks */
                    if (one_link_chunk_io_threshold == 0) {
                        io_option        = H5D_ONE_LINK_CHUNK_IO_MORE_OPT;
                        recalc_io_option = FALSE;
                    }
                    else {
                        /* Get number of chunks for all processes */
                        if (H5D__mpio_get_sum_chunk_dset(io_info, &io_info->dsets_info[i], &sum_chunk) < 0)
                            HGOTO_ERROR(H5E_DATASPACE, H5E_CANTSWAP, FAIL,
                                        "unable to obtain the total chunk number of all processes");

                        /* 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;
                        else
                            io_option = H5D_MULTI_CHUNK_IO_MORE_OPT;
                    }
                }

                /* 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->dsets_info[i].dset->shared->dcpl_cache.pline.nused > 0) {
                            if (H5D__link_chunk_filtered_collective_io(io_info, &io_info->dsets_info[i],
                                                                       mpi_rank, mpi_size) < 0)
                                HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL,
                                            "couldn't finish filtered linked chunk MPI-IO")
                        } /* end if */
                        else {
                            /* If there is more than one dataset we cannot make the multi dataset call here,
                             * fall back to multi chunk */
                            if (io_info->count > 1) {
                                io_option        = H5D_MULTI_CHUNK_IO_MORE_OPT;
                                recalc_io_option = TRUE;

                                if (H5D__multi_chunk_collective_io(io_info, &io_info->dsets_info[i], mpi_rank,
                                                                   mpi_size) < 0)
                                    HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL,
                                                "couldn't finish optimized multiple chunk MPI-IO")
                            }
                            else {
                                /* Perform unfiltered link chunk collective IO */
                                if (H5D__link_piece_collective_io(io_info, mpi_rank) < 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->dsets_info[i].dset->shared->dcpl_cache.pline.nused > 0) {
                            if (H5D__multi_chunk_filtered_collective_io(io_info, &io_info->dsets_info[i],
                                                                        mpi_rank, mpi_size) < 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, &io_info->dsets_info[i], mpi_rank,
                                                               mpi_size) < 0)
                                HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL,
                                            "couldn't finish optimized multiple chunk MPI-IO")
                        }

                        break;
                } /* end switch */

#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 (H5D_MULTI_CHUNK_IO_MORE_OPT == io_option) {
                        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 */
            }
        }
    }

done:
#ifdef H5Dmpio_DEBUG
    /* Close debugging log file */
    if (debug_log_file) {
        HDfprintf(debug_log_file, "##############\n\n");
        if (EOF == HDfclose(debug_log_file))
            HDONE_ERROR(H5E_IO, H5E_CLOSEERROR, FAIL, "couldn't close debugging log file")
        debug_stream = H5DEBUG(D);
    }
#endif

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__piece_io */

/*-------------------------------------------------------------------------
 * Function:    H5D__collective_read
 *
 * Purpose:     Read directly from pieces (chunks/contig) 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__collective_read(H5D_io_info_t *io_info)
{
    herr_t ret_value = SUCCEED; /* Return value */

    FUNC_ENTER_PACKAGE

    /* Call generic selection operation */
    if (H5D__piece_io(io_info) < 0)
        HGOTO_ERROR(H5E_DATASPACE, H5E_READERROR, FAIL, "read error")

done:
    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__collective_read() */

/*-------------------------------------------------------------------------
 * Function:    H5D__collective_write
 *
 * Purpose:     Write directly to pieces (chunks/contig) 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__collective_write(H5D_io_info_t *io_info)
{
    herr_t ret_value = SUCCEED; /* Return value */

    FUNC_ENTER_PACKAGE

    /* Call generic selection operation */
    if (H5D__piece_io(io_info) < 0)
        HGOTO_ERROR(H5E_DATASPACE, H5E_WRITEERROR, FAIL, "write error")

done:
    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__collective_write() */

/*-------------------------------------------------------------------------
 * Function:    H5D__link_piece_collective_io
 *
 * Purpose:     Routine for single collective IO with one MPI derived datatype
 *              to link with all pieces (chunks + contig)
 *
 *              1. Use the piece addresses and piece info sorted in skiplist
 *              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_piece_collective_io(H5D_io_info_t *io_info, int mpi_rank)
{
    MPI_Datatype  chunk_final_mtype; /* Final memory MPI datatype for all chunks with selection */
    hbool_t       chunk_final_mtype_is_derived = FALSE;
    MPI_Datatype  chunk_final_ftype; /* Final file MPI datatype for all chunks with selection */
    hbool_t       chunk_final_ftype_is_derived = FALSE;
    H5D_storage_t ctg_store; /* Storage info for "fake" contiguous dataset */
    MPI_Datatype *chunk_mtype           = NULL;
    MPI_Datatype *chunk_ftype           = NULL;
    MPI_Aint     *chunk_file_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 */
    H5D_mpio_actual_chunk_opt_mode_t actual_chunk_opt_mode = H5D_MPIO_LINK_CHUNK;
    H5D_mpio_actual_io_mode_t        actual_io_mode        = 0;
    size_t                           i; /* Local index variable */
    herr_t                           ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    /* set actual_io_mode */
    for (i = 0; i < io_info->count; i++) {
        HDassert(io_info->dsets_info[i].dset->shared->dcpl_cache.pline.nused == 0);
        if (io_info->dsets_info[i].layout->type == H5D_CHUNKED)
            actual_io_mode |= H5D_MPIO_CHUNK_COLLECTIVE;
        else if (io_info->dsets_info[i].layout->type == H5D_CONTIGUOUS) {
            actual_io_mode |= H5D_MPIO_CONTIGUOUS_COLLECTIVE;

            /* if only single-dset */
            if (1 == io_info->count)
                actual_chunk_opt_mode = H5D_MPIO_NO_CHUNK_OPTIMIZATION;
        }
        else
            HGOTO_ERROR(H5E_IO, H5E_UNSUPPORTED, FAIL, "unsupported storage layout")
    }

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

    /* 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(actual_io_mode);

    /* Code block for actual actions (Build a MPI Type, IO) */
    {
        hsize_t mpi_buf_count; /* Number of MPI types */
        size_t  num_chunk;     /* Number of chunks for this process */

        H5D_piece_info_t *piece_info;

        /* local variable for base address for buffer */
        H5_flexible_const_ptr_t base_buf_addr;
        base_buf_addr.cvp = NULL;

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

#ifdef H5Dmpio_DEBUG
        H5D_MPIO_DEBUG_VA(mpi_rank, "num_chunk = %zu\n", num_chunk);
#endif

        /* Set up MPI datatype for chunks selected */
        if (num_chunk) {
            hbool_t need_sort = FALSE;

            /* Check if sel_pieces array is sorted */
            HDassert(io_info->sel_pieces[0]->faddr != HADDR_UNDEF);
            for (i = 1; i < num_chunk; i++) {
                HDassert(io_info->sel_pieces[i]->faddr != HADDR_UNDEF);

                if (io_info->sel_pieces[i]->faddr < io_info->sel_pieces[i - 1]->faddr) {
                    need_sort = TRUE;
                    break;
                }
            }

            /* Sort sel_pieces if necessary */
            if (need_sort)
                HDqsort(io_info->sel_pieces, io_info->pieces_added, sizeof(io_info->sel_pieces[0]),
                        H5D__cmp_piece_addr);

            /* Allocate chunking information */
            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_file_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")

            /* save lowest file address */
            ctg_store.contig.dset_addr = io_info->sel_pieces[0]->faddr;

            /* save base mem addr of piece for read/write */
            base_buf_addr = io_info->sel_pieces[0]->dset_info->buf;

#ifdef H5Dmpio_DEBUG
            H5D_MPIO_DEBUG(mpi_rank, "before iterate over selected pieces\n");
#endif

            /* Obtain MPI derived datatype from all individual pieces */
            /* Iterate over selected pieces for this process */
            for (i = 0; i < num_chunk; i++) {
                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;

                /* Assign convenience pointer to piece info */
                piece_info = io_info->sel_pieces[i];

                /* 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(piece_info->fspace, piece_info->dset_info->type_info.src_type_size,
                                        &chunk_ftype[i],                  /* OUT: datatype created */
                                        &chunk_mpi_file_counts[i],        /* OUT */
                                        &(chunk_mft_is_derived_array[i]), /* 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(piece_info->mspace, piece_info->dset_info->type_info.dst_type_size,
                                        &chunk_mtype[i], &chunk_mpi_mem_counts[i],
                                        &(chunk_mbt_is_derived_array[i]), 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);

                /* Piece address relative to the first piece addr
                 * Assign piece address to MPI displacement
                 * (assume MPI_Aint big enough to hold it) */
                chunk_file_disp_array[i] = (MPI_Aint)piece_info->faddr - (MPI_Aint)ctg_store.contig.dset_addr;

                if (io_info->op_type == H5D_IO_OP_WRITE) {
                    chunk_mem_disp_array[i] =
                        (MPI_Aint)piece_info->dset_info->buf.cvp - (MPI_Aint)base_buf_addr.cvp;
                }
                else if (io_info->op_type == H5D_IO_OP_READ) {
                    chunk_mem_disp_array[i] =
                        (MPI_Aint)piece_info->dset_info->buf.vp - (MPI_Aint)base_buf_addr.vp;
                }
            } /* 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_file_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 (i = 0; i < num_chunk; i++) {
                if (chunk_mbt_is_derived_array[i])
                    if (MPI_SUCCESS != (mpi_code = MPI_Type_free(chunk_mtype + i)))
                        HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)

                if (chunk_mft_is_derived_array[i])
                    if (MPI_SUCCESS != (mpi_code = MPI_Type_free(chunk_ftype + i)))
                        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;

            /* just provide a valid mem address. no actual IO occur */
            base_buf_addr = io_info->dsets_info[0].buf;

            /* 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 H5Dmpio_DEBUG
        H5D_MPIO_DEBUG(mpi_rank, "before coming to final collective I/O");
#endif
        /* Set up the base storage address for this piece */
        io_info->store_faddr = ctg_store.contig.dset_addr;
        io_info->base_maddr  = base_buf_addr;

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

done:
#ifdef H5Dmpio_DEBUG
    H5D_MPIO_DEBUG_VA(mpi_rank, "before freeing memory inside H5D_link_collective_io ret_value = %d",
                      ret_value);
#endif

    /* Release resources */
    if (chunk_mtype)
        H5MM_xfree(chunk_mtype);
    if (chunk_ftype)
        H5MM_xfree(chunk_ftype);
    if (chunk_file_disp_array)
        H5MM_xfree(chunk_file_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_piece_collective_io */

/*-------------------------------------------------------------------------
 * Function:    H5D__link_chunk_filtered_collective_io
 *
 * Purpose:     Performs collective I/O on filtered chunks by creating a
 *              single MPI derived datatype to link with all filtered
 *              chunks. The general algorithm is as follows:
 *
 *              1. Construct a list of selected chunks in the collective
 *                 I/O operation
 *              2. If the operation is a read operation
 *                 A. Ensure that the list of chunks is sorted in
 *                    monotonically non-decreasing order of chunk offset
 *                    in the file
 *                 B. Participate in a collective read of chunks from
 *                    the file
 *                 C. Loop through each selected chunk, unfiltering it and
 *                    scattering the data to the application's read buffer
 *              3. If the operation is a write operation
 *                 A. Redistribute any chunks being written by more than 1
 *                    MPI rank, such that the chunk is only owned by 1 MPI
 *                    rank. The rank 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)
 *                 B. Participate in a collective read of chunks from the
 *                    file
 *                 C. Loop through each chunk selected in the operation
 *                    and for each chunk:
 *                    I. If we actually read the chunk from the file (if
 *                       a chunk is being fully overwritten, we skip
 *                       reading it), pass the chunk through the filter
 *                       pipeline in reverse order (unfilter the chunk)
 *                    II. Update the chunk data with the modifications from
 *                        the owning MPI rank
 *                    III. Receive any modification data from other
 *                         ranks and update the chunk data with those
 *                         modifications
 *                    IV. Filter the chunk
 *                 D. Contribute the modified chunks to an array gathered
 *                    by all ranks which contains information for
 *                    re-allocating space in the file for every chunk
 *                    modified. Then, each rank collectively re-allocates
 *                    each chunk from the gathered array with their new
 *                    sizes after the filter operation
 *                 E. Proceed with the collective write operation for all
 *                    the modified chunks
 *                 F. Contribute the modified chunks to an array gathered
 *                    by all ranks which contains information for
 *                    re-inserting every chunk modified into the chunk
 *                    index. Then, each rank collectively re-inserts each
 *                    chunk from the gathered array into the chunk index
 *
 *              TODO: Note that steps D. and F. here are both collective
 *                    operations that partially share data from the
 *                    H5D_filtered_collective_io_info_t structure. To
 *                    try to conserve on memory a bit, the distributed
 *                    arrays these operations create are discarded after
 *                    each operation is performed. If memory consumption
 *                    here proves to not be an issue, the necessary data
 *                    for both operations could be combined into a single
 *                    structure so that only one collective MPI operation
 *                    is needed to carry out both operations, rather than
 *                    two.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__link_chunk_filtered_collective_io(H5D_io_info_t *io_info, H5D_dset_io_info_t *dset_info, int mpi_rank,
                                       int mpi_size)
{
    H5D_filtered_collective_io_info_t *chunk_list          = NULL; /* The list of chunks being read/written */
    H5D_filtered_collective_io_info_t *chunk_hash_table    = NULL;
    unsigned char                    **chunk_msg_bufs      = NULL;
    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                            *rank_chunks_assigned_map = NULL;
    size_t                             chunk_list_num_entries;
    size_t                             i;
    int                                chunk_msg_bufs_len = 0;
    char   fake_buf; /* Used as a fake buffer for ranks with no chunks, thus a NULL buf pointer */
    int    mpi_code;
    herr_t ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE_TAG(dset_info->dset->oloc.addr)

    HDassert(io_info);

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TRACE_ENTER(mpi_rank);
    H5D_MPIO_DEBUG_VA(mpi_rank, "Performing Linked-chunk I/O (%s) with MPI Comm size of %d",
                      io_info->op_type == H5D_IO_OP_WRITE ? "write" : "read", mpi_size);
    H5D_MPIO_TIME_START(mpi_rank, "Linked-chunk I/O");
#endif

    /* 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__mpio_collective_filtered_chunk_io_setup(io_info, dset_info, &chunk_list, &chunk_list_num_entries,
                                                     mpi_rank) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "couldn't construct filtered I/O info list")

    if (io_info->op_type == H5D_IO_OP_READ) { /* Filtered collective read */
        if (H5D__mpio_collective_filtered_chunk_read(chunk_list, chunk_list_num_entries, io_info, dset_info,
                                                     mpi_rank, mpi_size) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "couldn't read filtered chunks")
    }
    else { /* Filtered collective write */
        H5D_chk_idx_info_t index_info;
        hsize_t            mpi_buf_count;

        H5D_MPIO_INIT_CHUNK_IDX_INFO(index_info, dset_info->dset);

        if (mpi_size > 1) {
            /* Redistribute shared chunks being written to */
            if (H5D__mpio_redistribute_shared_chunks(chunk_list, chunk_list_num_entries, io_info, mpi_rank,
                                                     mpi_size, &rank_chunks_assigned_map) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "unable to redistribute shared chunks")

            /* Send any chunk modification messages for chunks this rank no longer owns */
            if (H5D__mpio_share_chunk_modification_data(chunk_list, &chunk_list_num_entries, io_info,
                                                        dset_info, mpi_rank, mpi_size, &chunk_hash_table,
                                                        &chunk_msg_bufs, &chunk_msg_bufs_len) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL,
                            "unable to send chunk modification data between MPI ranks")

            /* Make sure the local chunk list was updated correctly */
            HDassert(chunk_list_num_entries == rank_chunks_assigned_map[mpi_rank]);
        }

        /* Proceed to update all the chunks this rank owns with its own
         * modification data and data from other ranks, before re-filtering
         * the chunks. As chunk reads are done collectively here, all ranks
         * must participate.
         */
        if (H5D__mpio_collective_filtered_chunk_update(chunk_list, chunk_list_num_entries, chunk_hash_table,
                                                       chunk_msg_bufs, chunk_msg_bufs_len, io_info, dset_info,
                                                       mpi_rank, mpi_size) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "couldn't update modified chunks")

        /* Free up resources used by chunk hash table now that we're done updating chunks */
        HASH_CLEAR(hh, chunk_hash_table);

        /* All ranks now collectively re-allocate file space for all chunks */
        if (H5D__mpio_collective_filtered_chunk_reallocate(chunk_list, chunk_list_num_entries,
                                                           rank_chunks_assigned_map, io_info, &index_info,
                                                           mpi_rank, mpi_size) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL,
                        "couldn't collectively re-allocate file space for chunks")

        /* If this rank has any chunks selected, create a MPI type for collectively
         * writing out the chunks to file. Otherwise, the rank contributes to the
         * collective write with a none type.
         */
        if (H5D__mpio_collective_filtered_io_type(chunk_list, chunk_list_num_entries, io_info->op_type,
                                                  &mem_type, &mem_type_is_derived, &file_type,
                                                  &file_type_is_derived) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL,
                        "couldn't create MPI type for writing filtered chunks")

        mpi_buf_count = (file_type_is_derived || mem_type_is_derived) ? 1 : 0;

        /* Setup contig storage info for I/O operation */
        if (chunk_list_num_entries) {
            /*
             * Override the write buffer to point to the first
             * chunk's data buffer
             */
            io_info->base_maddr.cvp = chunk_list[0].buf;

            /*
             * Setup the base storage address for this operation
             * to be the first chunk's file address
             */
            io_info->store_faddr = chunk_list[0].chunk_new.offset;
        }
        else {
            io_info->base_maddr.cvp = &fake_buf;
            io_info->store_faddr    = 0;
        }

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

        /* Free up resources in anticipation of following collective operation */
        for (i = 0; i < chunk_list_num_entries; i++) {
            if (chunk_list[i].buf) {
                H5MM_free(chunk_list[i].buf);
                chunk_list[i].buf = NULL;
            }
        }

        /* Participate in the collective re-insertion of all chunks modified
         * into the chunk index
         */
        if (H5D__mpio_collective_filtered_chunk_reinsert(chunk_list, chunk_list_num_entries,
                                                         rank_chunks_assigned_map, io_info, dset_info,
                                                         &index_info, mpi_rank, mpi_size) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL,
                        "couldn't collectively re-insert modified chunks into chunk index")
    }

done:
    /* 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)

    if (chunk_msg_bufs) {
        for (i = 0; i < (size_t)chunk_msg_bufs_len; i++)
            H5MM_free(chunk_msg_bufs[i]);

        H5MM_free(chunk_msg_bufs);
    }

    HASH_CLEAR(hh, chunk_hash_table);

    /* Free resources used by a rank 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 (rank_chunks_assigned_map)
        H5MM_free(rank_chunks_assigned_map);

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TIME_STOP(mpi_rank);
    H5D_MPIO_TRACE_EXIT(mpi_rank);
#endif

    FUNC_LEAVE_NOAPI_TAG(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, H5D_dset_io_info_t *dset_info, int mpi_rank,
                               int mpi_size)
{
    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
                                  */
    H5FD_mpio_collective_opt_t orig_coll_opt_mode =
        H5FD_MPIO_COLLECTIVE_IO;           /* Original parallel transfer property on entering this function */
    size_t                    total_chunk; /* Total # of chunks in dataset */
    size_t                    num_chunk;   /* Number of chunks for this process */
    H5SL_node_t              *piece_node      = NULL; /* Current node in chunk skip list */
    H5D_piece_info_t         *next_chunk_info = NULL; /* Chunk info for next selected chunk */
    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_PACKAGE_TAG(dset_info->dset->oloc.addr)

    HDassert(dset_info->layout->type == H5D_CHUNKED);

    /* Get the current I/O collective opt mode so we can restore it later */
    if (H5CX_get_mpio_coll_opt(&orig_coll_opt_mode) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get MPI-I/O collective_op property")

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

    /* Retrieve total # of chunks in dataset */
    H5_CHECKED_ASSIGN(total_chunk, size_t, dset_info->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 H5Dmpio_DEBUG
    H5D_MPIO_DEBUG_VA(mpi_rank, "total_chunk %zu", total_chunk);
#endif

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

    /* Set memory buffers */
    io_info->base_maddr = dset_info->buf;

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

    /* Get the number of chunks with a selection */
    num_chunk = H5SL_count(dset_info->layout_io_info.chunk_map->dset_sel_pieces);

    if (num_chunk) {
        /* Start at the beginning of the chunk map skiplist.  Since these chunks are
         * stored in index order and since we're iterating in index order we can
         * just check for each chunk being selected in order */
        if (NULL == (piece_node = H5SL_first(dset_info->layout_io_info.chunk_map->dset_sel_pieces)))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "couldn't get piece node from skip list")
        if (NULL == (next_chunk_info = (H5D_piece_info_t *)H5SL_item(piece_node)))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "couldn't get piece info from skip list")
    }

    /* Loop over _all_ the chunks */
    for (u = 0; u < total_chunk; u++) {
        H5D_piece_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 H5Dmpio_DEBUG
        H5D_MPIO_DEBUG_VA(mpi_rank, "mpi_rank = %d, chunk index = %zu", mpi_rank, u);
#endif

        /* Check if this chunk is the next chunk in the skip list, if there are
         * selected chunks left to process */
        HDassert(!num_chunk || next_chunk_info);
        HDassert(!num_chunk || next_chunk_info->index >= u);
        if (num_chunk && next_chunk_info->index == u) {
            /* Next chunk is this chunk */
            chunk_info = next_chunk_info;

            /* One less chunk to process */
            num_chunk--;

            /* Advance next chunk to next node in skip list, if there are more chunks selected */
            if (num_chunk) {
                if (NULL == (piece_node = H5SL_next(piece_node)))
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "chunk skip list terminated early")
                if (NULL == (next_chunk_info = (H5D_piece_info_t *)H5SL_item(piece_node)))
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "couldn't get piece info from skip list")
            }

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

        /* 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 H5Dmpio_DEBUG
            H5D_MPIO_DEBUG_VA(mpi_rank, "inside collective chunk IO mpi_rank = %d, chunk index = %zu",
                              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 */
            io_info->store_faddr = chunk_addr[u];

            /* Perform the I/O */
            if (H5D__inter_collective_io(io_info, dset_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 H5Dmpio_DEBUG
            H5D_MPIO_DEBUG_VA(mpi_rank, "inside independent IO mpi_rank = %d, chunk index = %zu", 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 */
            io_info->store_faddr = chunk_addr[u];

            /* Perform the I/O */
            if (H5D__inter_collective_io(io_info, dset_info, fspace, mspace) < 0)
                HGOTO_ERROR(H5E_IO, H5E_CANTGET, FAIL, "couldn't finish shared collective MPI-IO")
#ifdef H5Dmpio_DEBUG
            H5D_MPIO_DEBUG(mpi_rank, "after inter collective IO");
#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:
    /* Reset collective opt mode */
    if (H5CX_set_mpio_coll_opt(orig_coll_opt_mode) < 0)
        HDONE_ERROR(H5E_DATASET, H5E_CANTSET, FAIL, "can't reset MPI-I/O collective_op property")

    /* Free memory */
    if (chunk_io_option)
        H5MM_xfree(chunk_io_option);
    if (chunk_addr)
        H5MM_xfree(chunk_addr);

    FUNC_LEAVE_NOAPI_TAG(ret_value)
} /* end H5D__multi_chunk_collective_io */

/*-------------------------------------------------------------------------
 * Function:    H5D__multi_chunk_filtered_collective_io
 *
 * Purpose:     Performs collective I/O on filtered chunks iteratively to
 *              save on memory and potentially get better performance
 *              depending on the average number of chunks per rank. While
 *              linked-chunk I/O will construct and work on a list of all
 *              of the chunks selected in the I/O operation at once, this
 *              function works iteratively on a set of chunks at a time; at
 *              most one chunk per rank per iteration.  The general
 *              algorithm is as follows:
 *
 *              1. Construct a list of selected chunks in the collective
 *                 I/O operation
 *              2. If the operation is a read operation, loop an amount of
 *                 times equal to the maximum number of chunks selected on
 *                 any particular rank and on each iteration:
 *                 A. Participate in a collective read of chunks from
 *                    the file (ranks that run out of chunks still need
 *                    to participate)
 *                 B. Unfilter the chunk that was read (if any)
 *                 C. Scatter the read chunk's data to the application's
 *                    read buffer
 *              3. If the operation is a write operation, redistribute any
 *                 chunks being written to by more than 1 MPI rank, such
 *                 that the chunk is only owned by 1 MPI rank. The rank
 *                 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). Then, loop an amount of times equal to the
 *                 maximum number of chunks selected on any particular
 *                 rank and on each iteration:
 *                 A. Participate in a collective read of chunks from
 *                    the file (ranks that run out of chunks still need
 *                    to participate)
 *                    I. If we actually read a chunk from the file (if
 *                       a chunk is being fully overwritten, we skip
 *                       reading it), pass the chunk through the filter
 *                       pipeline in reverse order (unfilter the chunk)
 *                 B. Update the chunk data with the modifications from
 *                    the owning rank
 *                 C. Receive any modification data from other ranks and
 *                    update the chunk data with those modifications
 *                 D. Filter the chunk
 *                 E. Contribute the chunk to an array gathered by
 *                    all ranks which contains information for
 *                    re-allocating space in the file for every chunk
 *                    modified in this iteration (up to one chunk per
 *                    rank; some ranks may not have a selection/may have
 *                    less chunks to work on than other ranks). Then,
 *                    each rank collectively re-allocates each chunk
 *                    from the gathered array with their new sizes
 *                    after the filter operation
 *                 F. Proceed with the collective write operation
 *                    for the chunks modified on this iteration
 *                 G. Contribute the chunk to an array gathered by
 *                    all ranks which contains information for
 *                    re-inserting every chunk modified on this
 *                    iteration into the chunk index. Then, each rank
 *                    collectively re-inserts each chunk from the
 *                    gathered array into the chunk index
 *
 *              TODO: Note that steps E. and G. here are both collective
 *                    operations that partially share data from the
 *                    H5D_filtered_collective_io_info_t structure. To
 *                    try to conserve on memory a bit, the distributed
 *                    arrays these operations create are discarded after
 *                    each operation is performed. If memory consumption
 *                    here proves to not be an issue, the necessary data
 *                    for both operations could be combined into a single
 *                    structure so that only one collective MPI operation
 *                    is needed to carry out both operations, rather than
 *                    two.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__multi_chunk_filtered_collective_io(H5D_io_info_t *io_info, H5D_dset_io_info_t *dset_info, int mpi_rank,
                                        int mpi_size)
{
    H5D_filtered_collective_io_info_t *chunk_list       = NULL; /* The list of chunks being read/written */
    H5D_filtered_collective_io_info_t *chunk_hash_table = NULL;
    unsigned char                    **chunk_msg_bufs   = NULL;
    H5D_io_info_t                      ctg_io_info; /* Contiguous I/O info object */
    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;
    hbool_t                            have_chunk_to_process;
    size_t                             chunk_list_num_entries;
    size_t                             i;
    size_t                             max_num_chunks;
    int                                chunk_msg_bufs_len = 0;
    int                                mpi_code;
    herr_t                             ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE_TAG(dset_info->dset->oloc.addr)

    HDassert(io_info);

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TRACE_ENTER(mpi_rank);
    H5D_MPIO_DEBUG_VA(mpi_rank, "Performing Multi-chunk I/O (%s) with MPI Comm size of %d",
                      io_info->op_type == H5D_IO_OP_WRITE ? "write" : "read", mpi_size);
    H5D_MPIO_TIME_START(mpi_rank, "Multi-chunk I/O");
#endif

    /* 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__mpio_collective_filtered_chunk_io_setup(io_info, dset_info, &chunk_list, &chunk_list_num_entries,
                                                     mpi_rank) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "couldn't construct filtered I/O info list")

    /* Retrieve the maximum number of chunks selected for any rank */
    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 has anything selected at all, end the operation */
    if (0 == max_num_chunks)
        HGOTO_DONE(SUCCEED);

    /* Set up contiguous I/O info object */
    H5MM_memcpy(&ctg_io_info, io_info, sizeof(ctg_io_info));

    if (io_info->op_type == H5D_IO_OP_READ) { /* Filtered collective read */
        for (i = 0; i < max_num_chunks; i++) {
            /* Check if this rank has a chunk to work on for this iteration */
            have_chunk_to_process = (i < chunk_list_num_entries);

            if (H5D__mpio_collective_filtered_chunk_read(have_chunk_to_process ? &chunk_list[i] : NULL,
                                                         have_chunk_to_process ? 1 : 0, io_info, dset_info,
                                                         mpi_rank, mpi_size) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "couldn't read filtered chunks")

            if (have_chunk_to_process && chunk_list[i].buf) {
                H5MM_free(chunk_list[i].buf);
                chunk_list[i].buf = NULL;
            }
        }
    }
    else { /* Filtered collective write */
        H5D_chk_idx_info_t index_info;
        hsize_t            mpi_buf_count;

        /* Construct chunked index info */
        H5D_MPIO_INIT_CHUNK_IDX_INFO(index_info, dset_info->dset);

        if (mpi_size > 1) {
            /* Redistribute shared chunks being written to */
            if (H5D__mpio_redistribute_shared_chunks(chunk_list, chunk_list_num_entries, io_info, mpi_rank,
                                                     mpi_size, NULL) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "unable to redistribute shared chunks")

            /* Send any chunk modification messages for chunks this rank no longer owns */
            if (H5D__mpio_share_chunk_modification_data(chunk_list, &chunk_list_num_entries, io_info,
                                                        dset_info, mpi_rank, mpi_size, &chunk_hash_table,
                                                        &chunk_msg_bufs, &chunk_msg_bufs_len) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL,
                            "unable to send chunk modification data between MPI ranks")
        }

        /* Iterate over the max number of chunks among all ranks, as this rank 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 ranks.
         */
        for (i = 0; i < max_num_chunks; i++) {
            /* Check if this rank has a chunk to work on for this iteration */
            have_chunk_to_process = (i < chunk_list_num_entries) && (mpi_rank == chunk_list[i].new_owner);

            /* Proceed to update the chunk this rank owns (if any left) with its
             * own modification data and data from other ranks, before re-filtering
             * the chunks. As chunk reads are done collectively here, all ranks
             * must participate.
             */
            if (H5D__mpio_collective_filtered_chunk_update(have_chunk_to_process ? &chunk_list[i] : NULL,
                                                           have_chunk_to_process ? 1 : 0, chunk_hash_table,
                                                           chunk_msg_bufs, chunk_msg_bufs_len, io_info,
                                                           dset_info, mpi_rank, mpi_size) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "couldn't update modified chunks")

            /* All ranks now collectively re-allocate file space for all chunks */
            if (H5D__mpio_collective_filtered_chunk_reallocate(have_chunk_to_process ? &chunk_list[i] : NULL,
                                                               have_chunk_to_process ? 1 : 0, NULL, io_info,
                                                               &index_info, mpi_rank, mpi_size) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL,
                            "couldn't collectively re-allocate file space for chunks")

            /*
             * If this rank has a chunk to work on, create a MPI type
             * for writing out the chunk. Otherwise, the rank will
             * use MPI_BYTE for the file and memory type and specify
             * a count of 0.
             */
            if (H5D__mpio_collective_filtered_io_type(
                    have_chunk_to_process ? &chunk_list[i] : NULL, have_chunk_to_process ? 1 : 0,
                    io_info->op_type, &mem_type, &mem_type_is_derived, &file_type, &file_type_is_derived) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL,
                            "couldn't create MPI type for writing filtered chunks")

            mpi_buf_count = (file_type_is_derived || mem_type_is_derived) ? 1 : 0;

            /* Override the write buffer to point to the chunk data buffer */
            if (have_chunk_to_process) {
                /*
                 * Override the write buffer to point to the
                 * chunk's data buffer
                 */
                ctg_io_info.base_maddr.cvp = chunk_list[i].buf;

                /*
                 * Setup the base storage address for this
                 * operation to be the chunk's file address
                 */
                ctg_io_info.store_faddr = chunk_list[i].chunk_new.offset;
            }
            else {
                ctg_io_info.store_faddr = 0;
                ctg_io_info.base_maddr  = dset_info->buf;
            }

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

            /* Free up resources in anticipation of following collective operation */
            if (have_chunk_to_process && chunk_list[i].buf) {
                H5MM_free(chunk_list[i].buf);
                chunk_list[i].buf = NULL;
            }

            /* Participate in the collective re-insertion of all chunks modified
             * in this iteration into the chunk index
             */
            if (H5D__mpio_collective_filtered_chunk_reinsert(have_chunk_to_process ? &chunk_list[i] : NULL,
                                                             have_chunk_to_process ? 1 : 0, NULL, io_info,
                                                             dset_info, &index_info, mpi_rank, mpi_size) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL,
                            "couldn't collectively re-insert modified chunks into chunk index")

            /* Free the MPI types, if they were derived */
            if (mem_type_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&mem_type)))
                HMPI_GOTO_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
            mem_type_is_derived = FALSE;
            if (file_type_is_derived && MPI_SUCCESS != (mpi_code = MPI_Type_free(&file_type)))
                HMPI_GOTO_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
            file_type_is_derived = FALSE;
        } /* end for */
    }

done:
    /* 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)

    if (chunk_msg_bufs) {
        for (i = 0; i < (size_t)chunk_msg_bufs_len; i++)
            H5MM_free(chunk_msg_bufs[i]);

        H5MM_free(chunk_msg_bufs);
    }

    HASH_CLEAR(hh, chunk_hash_table);

    /* Free resources used by a rank 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 */

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TIME_STOP(mpi_rank);
    H5D_MPIO_TRACE_EXIT(mpi_rank);
#endif

    FUNC_LEAVE_NOAPI_TAG(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_dset_io_info_t *di, H5S_t *file_space,
                         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 */
#ifdef H5Dmpio_DEBUG
    int mpi_rank;
#endif
    herr_t ret_value = SUCCEED; /* return value */

    FUNC_ENTER_PACKAGE

#ifdef H5Dmpio_DEBUG
    mpi_rank = H5F_mpi_get_rank(di->dset->oloc.file);
    H5D_MPIO_TRACE_ENTER(mpi_rank);
    H5D_MPIO_TIME_START(mpi_rank, "Inter collective I/O");
    if (mpi_rank < 0)
        HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain MPI rank")
#endif

    HDassert(io_info);

    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;

        HDassert(di);

        /* 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, di->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, di->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 H5Dmpio_DEBUG
    H5D_MPIO_DEBUG(mpi_rank, "before final collective I/O");
#endif

    /* Perform final collective I/O operation */
    if (H5D__final_collective_io(io_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 H5Dmpio_DEBUG
    H5D_MPIO_TIME_STOP(mpi_rank);
    H5D_MPIO_DEBUG_VA(mpi_rank, "before leaving inter_collective_io ret_value = %d", ret_value);
    H5D_MPIO_TRACE_EXIT(mpi_rank);
#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, hsize_t mpi_buf_count, MPI_Datatype mpi_file_type,
                         MPI_Datatype mpi_buf_type)
{
#ifdef H5Dmpio_DEBUG
    int mpi_rank;
#endif
    herr_t ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

#ifdef H5Dmpio_DEBUG
    mpi_rank = H5F_mpi_get_rank(io_info->dsets_info[0].dset->oloc.file);
    H5D_MPIO_TRACE_ENTER(mpi_rank);
    H5D_MPIO_TIME_START(mpi_rank, "Final collective I/O");
    if (mpi_rank < 0)
        HGOTO_ERROR(H5E_IO, H5E_MPI, FAIL, "unable to obtain MPI rank")
#endif

    /* 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->md_io_ops.single_write_md)(io_info, mpi_buf_count, NULL, NULL) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "optimized write failed")
    } /* end if */
    else {
        if ((io_info->md_io_ops.single_read_md)(io_info, mpi_buf_count, NULL, NULL) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "optimized read failed")
    } /* end else */

done:
#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TIME_STOP(mpi_rank);
    H5D_MPIO_DEBUG_VA(mpi_rank, "ret_value before leaving final_collective_io=%d", ret_value);
    H5D_MPIO_TRACE_EXIT(mpi_rank);
#endif

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__final_collective_io */

/*-------------------------------------------------------------------------
 * Function:    H5D__cmp_piece_addr
 *
 * Purpose:     Routine to compare piece addresses
 *
 * Description: Callback for qsort() to compare piece addresses
 *
 * Return:      -1, 0, 1
 *
 *-------------------------------------------------------------------------
 */
static int
H5D__cmp_piece_addr(const void *piece_info1, const void *piece_info2)
{
    haddr_t addr1;
    haddr_t addr2;

    FUNC_ENTER_PACKAGE_NOERR

    addr1 = (*((const H5D_piece_info_t *const *)piece_info1))->faddr;
    addr2 = (*((const H5D_piece_info_t *const *)piece_info2))->faddr;

    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
 *
 *-------------------------------------------------------------------------
 */
static int
H5D__cmp_filtered_collective_io_info_entry(const void *filtered_collective_io_info_entry1,
                                           const void *filtered_collective_io_info_entry2)
{
    const H5D_filtered_collective_io_info_t *entry1;
    const H5D_filtered_collective_io_info_t *entry2;
    haddr_t                                  addr1 = HADDR_UNDEF;
    haddr_t                                  addr2 = HADDR_UNDEF;
    int                                      ret_value;

    FUNC_ENTER_PACKAGE_NOERR

    entry1 = (const H5D_filtered_collective_io_info_t *)filtered_collective_io_info_entry1;
    entry2 = (const H5D_filtered_collective_io_info_t *)filtered_collective_io_info_entry2;

    addr1 = entry1->chunk_new.offset;
    addr2 = entry2->chunk_new.offset;

    /*
     * If both chunk addresses are defined, H5F_addr_cmp is safe to use.
     * Otherwise, if both addresses aren't defined, compared chunk
     * entries based on their chunk index. Finally, if only one chunk
     * address is defined, return the appropriate value based on which
     * is defined.
     */
    if (H5F_addr_defined(addr1) && H5F_addr_defined(addr2)) {
        ret_value = H5F_addr_cmp(addr1, addr2);
    }
    else if (!H5F_addr_defined(addr1) && !H5F_addr_defined(addr2)) {
        hsize_t chunk_idx1 = entry1->index_info.chunk_idx;
        hsize_t chunk_idx2 = entry2->index_info.chunk_idx;

        ret_value = (chunk_idx1 > chunk_idx2) - (chunk_idx1 < chunk_idx2);
    }
    else
        ret_value = H5F_addr_defined(addr1) ? 1 : -1;

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__cmp_filtered_collective_io_info_entry() */

/*-------------------------------------------------------------------------
 * Function:    H5D__cmp_chunk_redistribute_info
 *
 * Purpose:     Routine to compare two H5D_chunk_redistribute_info_t
 *              structures
 *
 * Description: Callback for qsort() to compare two
 *              H5D_chunk_redistribute_info_t structures
 *
 * Return:      -1, 0, 1
 *
 *-------------------------------------------------------------------------
 */
static int
H5D__cmp_chunk_redistribute_info(const void *_entry1, const void *_entry2)
{
    const H5D_chunk_redistribute_info_t *entry1;
    const H5D_chunk_redistribute_info_t *entry2;
    hsize_t                              chunk_index1;
    hsize_t                              chunk_index2;
    int                                  ret_value;

    FUNC_ENTER_PACKAGE_NOERR

    entry1 = (const H5D_chunk_redistribute_info_t *)_entry1;
    entry2 = (const H5D_chunk_redistribute_info_t *)_entry2;

    chunk_index1 = entry1->chunk_idx;
    chunk_index2 = entry2->chunk_idx;

    if (chunk_index1 == chunk_index2) {
        int orig_owner1 = entry1->orig_owner;
        int orig_owner2 = entry2->orig_owner;

        ret_value = (orig_owner1 > orig_owner2) - (orig_owner1 < orig_owner2);
    }
    else
        ret_value = (chunk_index1 > chunk_index2) - (chunk_index1 < chunk_index2);

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__cmp_chunk_redistribute_info() */

/*-------------------------------------------------------------------------
 * Function:    H5D__cmp_chunk_redistribute_info_orig_owner
 *
 * Purpose:     Routine to compare the original owning MPI rank for two
 *              H5D_chunk_redistribute_info_t structures
 *
 * Description: Callback for qsort() to compare the original owning MPI
 *              rank for two H5D_chunk_redistribute_info_t
 *              structures
 *
 * Return:      -1, 0, 1
 *
 *-------------------------------------------------------------------------
 */
static int
H5D__cmp_chunk_redistribute_info_orig_owner(const void *_entry1, const void *_entry2)
{
    const H5D_chunk_redistribute_info_t *entry1;
    const H5D_chunk_redistribute_info_t *entry2;
    int                                  owner1 = -1;
    int                                  owner2 = -1;
    int                                  ret_value;

    FUNC_ENTER_PACKAGE_NOERR

    entry1 = (const H5D_chunk_redistribute_info_t *)_entry1;
    entry2 = (const H5D_chunk_redistribute_info_t *)_entry2;

    owner1 = entry1->orig_owner;
    owner2 = entry2->orig_owner;

    if (owner1 == owner2) {
        haddr_t addr1 = entry1->chunk_block.offset;
        haddr_t addr2 = entry2->chunk_block.offset;

        /*
         * If both chunk addresses are defined, H5F_addr_cmp is safe to use.
         * Otherwise, if both addresses aren't defined, compared chunk
         * entries based on their chunk index. Finally, if only one chunk
         * address is defined, return the appropriate value based on which
         * is defined.
         */
        if (H5F_addr_defined(addr1) && H5F_addr_defined(addr2)) {
            ret_value = H5F_addr_cmp(addr1, addr2);
        }
        else if (!H5F_addr_defined(addr1) && !H5F_addr_defined(addr2)) {
            hsize_t chunk_idx1 = entry1->chunk_idx;
            hsize_t chunk_idx2 = entry2->chunk_idx;

            ret_value = (chunk_idx1 > chunk_idx2) - (chunk_idx1 < chunk_idx2);
        }
        else
            ret_value = H5F_addr_defined(addr1) ? 1 : -1;
    }
    else
        ret_value = (owner1 > owner2) - (owner1 < owner2);

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__cmp_chunk_redistribute_info_orig_owner() */

/*-------------------------------------------------------------------------
 * 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) whether 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_dset_io_info_t *di,(dataset info 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_dset_io_info_t *di, uint8_t assign_io_mode[],
                      haddr_t chunk_addr[], int mpi_rank, int mpi_size)
{
    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_piece_info_t       *chunk_info;
    H5P_coll_md_read_flag_t md_reads_file_flag;
    hbool_t                 md_reads_context_flag;
    hbool_t                 restore_md_reads_state = FALSE;
    MPI_Comm                comm;
    int                     root;
    size_t                  ic;
    int                     mpi_code;
    herr_t                  ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    HDassert(di->layout->type == H5D_CHUNKED);

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

    /* Setup parameters */
    H5_CHECKED_ASSIGN(total_chunks, size_t, di->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(di->dset, 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(di->layout_io_info.chunk_map->dset_sel_pieces);
    while (chunk_node) {
        chunk_info = (H5D_piece_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;

        /*
         * If enabled, disable collective metadata reads here.
         * Since the chunk address mapping is done on rank 0
         * only here, it will cause problems if collective
         * metadata reads are enabled.
         */
        if (H5F_get_coll_metadata_reads(di->dset->oloc.file)) {
            md_reads_file_flag    = H5P_FORCE_FALSE;
            md_reads_context_flag = FALSE;
            H5F_set_coll_metadata_reads(di->dset->oloc.file, &md_reads_file_flag, &md_reads_context_flag);
            restore_md_reads_state = TRUE;
        }

        /* 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(di->dset, 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 */
        H5MM_memcpy(mergebuf, assign_io_mode, total_chunks);
        H5MM_memcpy(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 ((sizeof(haddr_t) + 1) * total_chunks > INT_MAX)
        HGOTO_ERROR(H5E_DATASET, H5E_BADVALUE, FAIL, "result overflow")
    if (MPI_SUCCESS !=
        (mpi_code = MPI_Bcast(mergebuf, (int)((sizeof(haddr_t) + 1) * total_chunks), MPI_BYTE, root, comm)))
        HMPI_GOTO_ERROR(FAIL, "MPI_BCast failed", mpi_code)

    H5MM_memcpy(assign_io_mode, mergebuf, total_chunks);
    H5MM_memcpy(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:
    /* Re-enable collective metadata reads if we disabled them */
    if (restore_md_reads_state)
        H5F_set_coll_metadata_reads(di->dset->oloc.file, &md_reads_file_flag, &md_reads_context_flag);

    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__mpio_collective_filtered_chunk_io_setup
 *
 * 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 MPI rank 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
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_collective_filtered_chunk_io_setup(const H5D_io_info_t *io_info, const H5D_dset_io_info_t *di,
                                             H5D_filtered_collective_io_info_t **chunk_list,
                                             size_t *num_entries, int mpi_rank)
{
    H5D_filtered_collective_io_info_t *local_info_array = NULL;
    H5D_chunk_ud_t                     udata;
    hbool_t                            filter_partial_edge_chunks;
    size_t                             num_chunks_selected;
    size_t                             i;
    herr_t                             ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    HDassert(io_info);
    HDassert(di);
    HDassert(chunk_list);
    HDassert(num_entries);
#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TRACE_ENTER(mpi_rank);
    H5D_MPIO_TIME_START(mpi_rank, "Filtered Collective I/O Setup");
#endif

    HDassert(di->layout->type == H5D_CHUNKED);

    /* Each rank builds a local list of the chunks they have selected */
    if ((num_chunks_selected = H5SL_count(di->layout_io_info.chunk_map->dset_sel_pieces))) {
        H5D_piece_info_t *chunk_info;
        H5SL_node_t      *chunk_node;
        hsize_t           select_npoints;
        hbool_t           need_sort = FALSE;

        /* Determine whether partial edge chunks should be filtered */
        filter_partial_edge_chunks =
            !(di->dset->shared->layout.u.chunk.flags & H5O_LAYOUT_CHUNK_DONT_FILTER_PARTIAL_BOUND_CHUNKS);

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

        chunk_node = H5SL_first(di->layout_io_info.chunk_map->dset_sel_pieces);
        for (i = 0; chunk_node; i++) {
            chunk_info = (H5D_piece_info_t *)H5SL_item(chunk_node);

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

            /* Initialize rank-local chunk info */
            local_info_array[i].chunk_info     = chunk_info;
            local_info_array[i].chunk_buf_size = 0;
            local_info_array[i].num_writers    = 0;
            local_info_array[i].orig_owner     = mpi_rank;
            local_info_array[i].new_owner      = mpi_rank;
            local_info_array[i].buf            = NULL;

            select_npoints              = H5S_GET_SELECT_NPOINTS(chunk_info->fspace);
            local_info_array[i].io_size = (size_t)select_npoints * di->type_info.dst_type_size;

            /*
             * Determine whether this chunk will need to be read from the file. If this is
             * a read operation, the chunk will be read. If this is a write operation, we
             * generally need to read a filtered chunk from the file before modifying it,
             * unless the chunk is being fully overwritten.
             *
             * TODO: Currently the full overwrite status of a chunk is only obtained on a
             * per-rank basis. This means that if the total selection in the chunk, as
             * determined by the combination of selections of all of the ranks 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.
             *
             * To implement this case, a few approaches were considered:
             *
             *  - Keep a running total (distributed to each rank) of the number of chunk
             *    elements selected during chunk redistribution and compare that to the total
             *    number of elements in the chunk once redistribution is finished
             *
             *  - Process all incoming chunk messages before doing I/O (these are currently
             *    processed AFTER doing I/O), combine the owning rank's selection in a chunk
             *    with the selections received from other ranks and check to see whether that
             *    combined selection covers the entire chunk
             *
             * The first approach will be dangerous if the application performs an overlapping
             * write to a chunk, as the number of selected elements can equal or exceed the
             * number of elements in the chunk without the whole chunk selection being covered.
             * While it might be considered erroneous for an application to do an overlapping
             * write, we don't explicitly disallow it.
             *
             * The second approach contains a bit of complexity in that part of the chunk
             * messages will be needed before doing I/O and part will be needed after doing I/O.
             * Since modification data from chunk messages can't be applied until after any I/O
             * is performed (otherwise, we'll overwrite any applied modification data), chunk
             * messages are currently entirely processed after I/O. However, in order to determine
             * if a chunk is being fully overwritten, we need the dataspace portion of the chunk
             * messages before doing I/O. The naive way to do this is to process chunk messages
             * twice, using just the relevant information from the message before and after I/O.
             * The better way would be to avoid processing chunk messages twice by extracting (and
             * keeping around) the dataspace portion of the message before I/O and processing the
             * rest of the chunk message after I/O. Note that the dataspace portion of each chunk
             * message is used to correctly apply chunk modification data from the message, so
             * must be kept around both before and after I/O in this case.
             */
            if (io_info->op_type == H5D_IO_OP_READ)
                local_info_array[i].need_read = TRUE;
            else {
                local_info_array[i].need_read =
                    local_info_array[i].io_size < (size_t)di->dset->shared->layout.u.chunk.size;
            }

            local_info_array[i].skip_filter_pline = FALSE;
            if (!filter_partial_edge_chunks) {
                /*
                 * If this is a partial edge chunk and the "don't filter partial edge
                 * chunks" flag is set, make sure not to apply filters to the chunk.
                 */
                if (H5D__chunk_is_partial_edge_chunk(di->dset->shared->ndims,
                                                     di->dset->shared->layout.u.chunk.dim, chunk_info->scaled,
                                                     di->dset->shared->curr_dims))
                    local_info_array[i].skip_filter_pline = TRUE;
            }

            /* Initialize the chunk's shared info */
            local_info_array[i].chunk_current = udata.chunk_block;
            local_info_array[i].chunk_new     = udata.chunk_block;

            /*
             * Check if the list is not in ascending order of offset in the file
             * or has unallocated chunks. In either case, the list should get
             * sorted.
             */
            if (i) {
                haddr_t curr_chunk_offset = local_info_array[i].chunk_current.offset;
                haddr_t prev_chunk_offset = local_info_array[i - 1].chunk_current.offset;

                if (!H5F_addr_defined(prev_chunk_offset) || !H5F_addr_defined(curr_chunk_offset) ||
                    (curr_chunk_offset < prev_chunk_offset))
                    need_sort = TRUE;
            }

            /*
             * Extensible arrays may calculate a chunk's index a little differently
             * than normal when the dataset's unlimited dimension is not the
             * slowest-changing dimension, so set the index here based on what the
             * extensible array code calculated instead of what was calculated
             * in the chunk file mapping.
             */
            if (di->dset->shared->layout.u.chunk.idx_type == H5D_CHUNK_IDX_EARRAY)
                local_info_array[i].index_info.chunk_idx = udata.chunk_idx;
            else
                local_info_array[i].index_info.chunk_idx = chunk_info->index;

            local_info_array[i].index_info.filter_mask = udata.filter_mask;
            local_info_array[i].index_info.need_insert = FALSE;

            chunk_node = H5SL_next(chunk_node);
        }

        /* Ensure the chunk list is sorted in ascending order of offset in the file */
        if (need_sort)
            HDqsort(local_info_array, num_chunks_selected, sizeof(H5D_filtered_collective_io_info_t),
                    H5D__cmp_filtered_collective_io_info_entry);

#ifdef H5Dmpio_DEBUG
        H5D__mpio_dump_collective_filtered_chunk_list(local_info_array, num_chunks_selected, mpi_rank);
#endif
    }
    else if (H5F_get_coll_metadata_reads(di->dset->oloc.file)) {
        hsize_t scaled[H5O_LAYOUT_NDIMS] = {0};

        /*
         * If this rank has no selection in the dataset and collective
         * metadata reads are enabled, do a fake lookup of a chunk to
         * ensure that this rank has the chunk index opened. Otherwise,
         * only the ranks that had a selection will have opened the
         * chunk index and they will have done so independently. Therefore,
         * when ranks with no selection participate in later collective
         * metadata reads, they will try to open the chunk index collectively
         * and issues will occur since other ranks won't participate.
         *
         * In the future, we should consider having a chunk index "open"
         * callback that can be used to ensure collectivity between ranks
         * in a more natural way, but this hack should suffice for now.
         */
        if (H5D__chunk_lookup(di->dset, scaled, &udata) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "error looking up chunk address")
    }

    *chunk_list  = local_info_array;
    *num_entries = num_chunks_selected;

done:
#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TIME_STOP(mpi_rank);
    H5D_MPIO_TRACE_EXIT(mpi_rank);
#endif

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__mpio_collective_filtered_chunk_io_setup() */

/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_redistribute_shared_chunks
 *
 * Purpose:     When performing a parallel write on a chunked Dataset with
 *              filters applied, we must ensure that any particular chunk
 *              is only written to by a single MPI rank in order to avoid
 *              potential data races on the chunk. This function is used to
 *              redistribute (by assigning ownership to a single rank) any
 *              chunks which are selected by more than one MPI rank.
 *
 *              An initial Allgather is performed to determine how many
 *              chunks each rank has selected in the write operation and
 *              then that number is compared against a threshold value to
 *              determine whether chunk redistribution should be done on
 *              MPI rank 0 only, or on all MPI ranks.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_redistribute_shared_chunks(H5D_filtered_collective_io_info_t *chunk_list,
                                     size_t chunk_list_num_entries, const H5D_io_info_t *io_info,
                                     int mpi_rank, int mpi_size, size_t **rank_chunks_assigned_map)
{
    hbool_t redistribute_on_all_ranks;
    size_t *num_chunks_map       = NULL;
    size_t  coll_chunk_list_size = 0;
    size_t  i;
    int     mpi_code;
    herr_t  ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    HDassert(chunk_list || 0 == chunk_list_num_entries);
    HDassert(io_info);
    HDassert(mpi_size > 1); /* No chunk sharing is possible for MPI Comm size of 1 */

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TRACE_ENTER(mpi_rank);
    H5D_MPIO_TIME_START(mpi_rank, "Redistribute shared chunks");
#endif

    /*
     * Allocate an array for each rank to keep track of the number of
     * chunks assigned to any other rank in order to cut down on future
     * MPI communication.
     */
    if (NULL == (num_chunks_map = H5MM_malloc((size_t)mpi_size * sizeof(*num_chunks_map))))
        HGOTO_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL, "couldn't allocate assigned chunks array")

    /* Perform initial Allgather to determine the collective chunk list size */
    if (MPI_SUCCESS != (mpi_code = MPI_Allgather(&chunk_list_num_entries, 1, H5_SIZE_T_AS_MPI_TYPE,
                                                 num_chunks_map, 1, H5_SIZE_T_AS_MPI_TYPE, io_info->comm)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Allgather failed", mpi_code)

    for (i = 0; i < (size_t)mpi_size; i++)
        coll_chunk_list_size += num_chunks_map[i];

    /*
     * Determine whether we should perform chunk redistribution on all
     * ranks or just rank 0. For a relatively small number of chunks,
     * we redistribute on all ranks to cut down on MPI communication
     * overhead. For a larger number of chunks, we redistribute on
     * rank 0 only to cut down on memory usage.
     */
    redistribute_on_all_ranks = coll_chunk_list_size < H5D_CHUNK_REDISTRIBUTE_THRES;

    if (H5D__mpio_redistribute_shared_chunks_int(chunk_list, num_chunks_map, redistribute_on_all_ranks,
                                                 io_info, mpi_rank, mpi_size) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTREDISTRIBUTE, FAIL, "can't redistribute shared chunks")

    /*
     * If the caller provided a pointer for the mapping from
     * rank value -> number of chunks assigned, return that
     * mapping here.
     */
    if (rank_chunks_assigned_map) {
        /*
         * If we performed chunk redistribution on rank 0 only, distribute
         * the rank value -> number of chunks assigned mapping back to all
         * ranks.
         */
        if (!redistribute_on_all_ranks) {
            if (MPI_SUCCESS !=
                (mpi_code = MPI_Bcast(num_chunks_map, mpi_size, H5_SIZE_T_AS_MPI_TYPE, 0, io_info->comm)))
                HMPI_GOTO_ERROR(FAIL, "couldn't broadcast chunk mapping to other ranks", mpi_code)
        }

        *rank_chunks_assigned_map = num_chunks_map;
    }

done:
    if (!rank_chunks_assigned_map || (ret_value < 0)) {
        num_chunks_map = H5MM_xfree(num_chunks_map);
    }

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TIME_STOP(mpi_rank);
    H5D_MPIO_TRACE_EXIT(mpi_rank);
#endif

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__mpio_redistribute_shared_chunks() */

/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_redistribute_shared_chunks_int
 *
 * Purpose:     Routine to perform redistribution of shared chunks during
 *              parallel writes to datasets with filters applied.
 *
 *              If `all_ranks_involved` is TRUE, chunk redistribution
 *              occurs on all MPI ranks. This is usually done when there
 *              is a relatively small number of chunks involved in order to
 *              cut down on MPI communication overhead while increasing
 *              total memory usage a bit.
 *
 *              If `all_ranks_involved` is FALSE, only rank 0 will perform
 *              chunk redistribution. This is usually done when there is
 *              a relatively large number of chunks involved in order to
 *              cut down on total memory usage at the cost of increased
 *              overhead from MPI communication.
 *
 *              This implementation is as follows:
 *
 *              - All MPI ranks send their list of selected chunks to the
 *                ranks involved in chunk redistribution. Then, the
 *                involved ranks sort this new list in order of chunk
 *                index.
 *
 *              - The involved ranks scan the list looking for matching
 *                runs of chunk index values (corresponding to a shared
 *                chunk which has been selected by more than one rank in
 *                the I/O operation) and for each shared chunk,
 *                redistribute the chunk to the MPI rank writing to the
 *                chunk which currently has the least amount of chunks
 *                assigned to it. This is done by modifying the "new_owner"
 *                field in each of the list entries corresponding to that
 *                chunk. The involved ranks then re-sort the list in order
 *                of original chunk owner so that each rank's section of
 *                contributed chunks is contiguous in the collective chunk
 *                list.
 *
 *              - If chunk redistribution occurred on all ranks, each rank
 *                scans through the collective chunk list to find their
 *                contributed section of chunks and uses that to update
 *                their local chunk list with the newly-updated "new_owner"
 *                and "num_writers" fields. If chunk redistribution
 *                occurred only on rank 0, an MPI_Scatterv operation will
 *                be used to scatter the segments of the collective chunk
 *                list from rank 0 back to the corresponding ranks.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_redistribute_shared_chunks_int(H5D_filtered_collective_io_info_t *chunk_list,
                                         size_t *num_chunks_assigned_map, hbool_t all_ranks_involved,
                                         const H5D_io_info_t *io_info, int mpi_rank, int mpi_size)
{
    MPI_Datatype struct_type;
    MPI_Datatype packed_type;
    hbool_t      struct_type_derived = FALSE;
    hbool_t      packed_type_derived = FALSE;
    size_t       i;
    size_t       coll_chunk_list_num_entries = 0;
    void        *coll_chunk_list             = NULL;
    int         *counts_disps_array          = NULL;
    int         *counts_ptr                  = NULL;
    int         *displacements_ptr           = NULL;
    int          num_chunks_int;
    int          mpi_code;
    herr_t       ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    HDassert(num_chunks_assigned_map);
    HDassert(chunk_list || 0 == num_chunks_assigned_map[mpi_rank]);
    HDassert(io_info);
    HDassert(mpi_size > 1);

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TRACE_ENTER(mpi_rank);
    H5D_MPIO_TIME_START(mpi_rank, "Redistribute shared chunks (internal)");
#endif

    /*
     * Make sure it's safe to cast this rank's number
     * of chunks to be sent into an int for MPI
     */
    H5_CHECKED_ASSIGN(num_chunks_int, int, num_chunks_assigned_map[mpi_rank], size_t);

    /*
     * Phase 1 - Participate in collective gathering of every rank's
     * list of chunks to the ranks which are performing the redistribution
     * operation.
     */

    if (all_ranks_involved || (mpi_rank == 0)) {
        /*
         * Allocate array to store the receive counts of each rank, as well as
         * the displacements into the final array where each rank will place
         * their data. The first half of the array contains the receive counts
         * (in rank order), while the latter half contains the displacements
         * (also in rank order).
         */
        if (NULL == (counts_disps_array = H5MM_malloc(2 * (size_t)mpi_size * sizeof(*counts_disps_array)))) {
            /* Push an error, but still participate in collective gather operation */
            HDONE_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL,
                        "couldn't allocate receive counts and displacements array")
        }
        else {
            /* Set the receive counts from the assigned chunks map */
            counts_ptr = counts_disps_array;

            for (i = 0; i < (size_t)mpi_size; i++)
                H5_CHECKED_ASSIGN(counts_ptr[i], int, num_chunks_assigned_map[i], size_t);

            /* Set the displacements into the receive buffer for the gather operation */
            displacements_ptr = &counts_disps_array[mpi_size];

            *displacements_ptr = 0;
            for (i = 1; i < (size_t)mpi_size; i++)
                displacements_ptr[i] = displacements_ptr[i - 1] + counts_ptr[i - 1];
        }
    }

    /*
     * Construct MPI derived types for extracting information
     * necessary for MPI communication
     */
    if (H5D__mpio_get_chunk_redistribute_info_types(&packed_type, &packed_type_derived, &struct_type,
                                                    &struct_type_derived) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL,
                    "can't create derived datatypes for chunk redistribution info")

    /* Perform gather operation */
    if (H5_mpio_gatherv_alloc(chunk_list, num_chunks_int, struct_type, counts_ptr, displacements_ptr,
                              packed_type, all_ranks_involved, 0, io_info->comm, mpi_rank, mpi_size,
                              &coll_chunk_list, &coll_chunk_list_num_entries) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTGATHER, FAIL,
                    "can't gather chunk redistribution info to involved ranks")

    /*
     * If all ranks are redistributing shared chunks, we no
     * longer need the receive counts and displacements array
     */
    if (all_ranks_involved) {
        counts_disps_array = H5MM_xfree(counts_disps_array);
    }

    /*
     * Phase 2 - Involved ranks now redistribute any shared chunks to new
     * owners as necessary.
     */

    if (all_ranks_involved || (mpi_rank == 0)) {
        H5D_chunk_redistribute_info_t *chunk_entry;
        hsize_t                        curr_chunk_idx;
        size_t                         set_begin_index;
        int                            num_writers;
        int                            new_chunk_owner;

        /* Clear the mapping from rank value -> number of assigned chunks */
        HDmemset(num_chunks_assigned_map, 0, (size_t)mpi_size * sizeof(*num_chunks_assigned_map));

        /* Sort collective chunk list according to chunk index */
        HDqsort(coll_chunk_list, coll_chunk_list_num_entries, sizeof(H5D_chunk_redistribute_info_t),
                H5D__cmp_chunk_redistribute_info);

        /*
         * Process all chunks in the collective chunk list.
         * Note that the loop counter is incremented by both
         * the outer loop (while processing each entry in
         * the collective chunk list) and the inner loop
         * (while processing duplicate entries for shared
         * chunks).
         */
        chunk_entry = &((H5D_chunk_redistribute_info_t *)coll_chunk_list)[0];
        for (i = 0; i < coll_chunk_list_num_entries;) {
            /* Set chunk's initial new owner to its original owner */
            new_chunk_owner = chunk_entry->orig_owner;

            /*
             * Set the current chunk index so we know when we've processed
             * all duplicate entries for a particular shared chunk
             */
            curr_chunk_idx = chunk_entry->chunk_idx;

            /* Reset the initial number of writers to this chunk */
            num_writers = 0;

            /* Set index for the beginning of this section of duplicate chunk entries */
            set_begin_index = i;

            /*
             * Process each chunk entry in the set for the current
             * (possibly shared) chunk and increment the loop counter
             * while doing so.
             */
            do {
                /*
                 * The new owner of the chunk is determined by the rank
                 * writing to the chunk which currently has the least amount
                 * of chunks assigned to it
                 */
                if (num_chunks_assigned_map[chunk_entry->orig_owner] <
                    num_chunks_assigned_map[new_chunk_owner])
                    new_chunk_owner = chunk_entry->orig_owner;

                /* Update the number of writers to this particular chunk */
                num_writers++;

                chunk_entry++;
            } while (++i < coll_chunk_list_num_entries && chunk_entry->chunk_idx == curr_chunk_idx);

            /* We should never have more writers to a chunk than the number of MPI ranks */
            HDassert(num_writers <= mpi_size);

            /* Set all processed chunk entries' "new_owner" and "num_writers" fields */
            for (; set_begin_index < i; set_begin_index++) {
                H5D_chunk_redistribute_info_t *entry;

                entry = &((H5D_chunk_redistribute_info_t *)coll_chunk_list)[set_begin_index];

                entry->new_owner   = new_chunk_owner;
                entry->num_writers = num_writers;
            }

            /* Update the number of chunks assigned to the MPI rank that now owns this chunk */
            num_chunks_assigned_map[new_chunk_owner]++;
        }

        /*
         * Re-sort the collective chunk list in order of original chunk owner
         * so that each rank's section of contributed chunks is contiguous in
         * the collective chunk list.
         *
         * NOTE: this re-sort is frail in that it needs to sort the collective
         *       chunk list so that each rank's section of contributed chunks
         *       is in the exact order it was contributed in, or things will
         *       be scrambled when each rank's local chunk list is updated.
         *       Therefore, the sorting algorithm here is tied to the one
         *       used during the I/O setup operation. Specifically, chunks
         *       are first sorted by ascending order of offset in the file and
         *       then by chunk index. In the future, a better redistribution
         *       algorithm may be devised that doesn't rely on frail sorting,
         *       but the current implementation is a quick and naive approach.
         */
        HDqsort(coll_chunk_list, coll_chunk_list_num_entries, sizeof(H5D_chunk_redistribute_info_t),
                H5D__cmp_chunk_redistribute_info_orig_owner);
    }

    if (all_ranks_involved) {
        /*
         * If redistribution occurred on all ranks, search for the section
         * in the collective chunk list corresponding to this rank's locally
         * selected chunks and update the local list after redistribution.
         */
        for (i = 0; i < coll_chunk_list_num_entries; i++)
            if (mpi_rank == ((H5D_chunk_redistribute_info_t *)coll_chunk_list)[i].orig_owner)
                break;

        for (size_t j = 0; j < (size_t)num_chunks_int; j++) {
            H5D_chunk_redistribute_info_t *coll_entry;

            coll_entry = &((H5D_chunk_redistribute_info_t *)coll_chunk_list)[i++];

            chunk_list[j].new_owner   = coll_entry->new_owner;
            chunk_list[j].num_writers = coll_entry->num_writers;
        }
    }
    else {
        /*
         * If redistribution occurred only on rank 0, scatter the segments
         * of the collective chunk list back to each rank so that their
         * local chunk lists get updated
         */
        if (MPI_SUCCESS !=
            (mpi_code = MPI_Scatterv(coll_chunk_list, counts_ptr, displacements_ptr, packed_type, chunk_list,
                                     num_chunks_int, struct_type, 0, io_info->comm)))
            HMPI_GOTO_ERROR(FAIL, "unable to scatter shared chunks info buffer", mpi_code)
    }

#ifdef H5Dmpio_DEBUG
    H5D__mpio_dump_collective_filtered_chunk_list(chunk_list, num_chunks_assigned_map[mpi_rank], mpi_rank);
#endif

done:
    H5MM_free(coll_chunk_list);

    if (struct_type_derived) {
        if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&struct_type)))
            HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
    }
    if (packed_type_derived) {
        if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&packed_type)))
            HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
    }

    H5MM_free(counts_disps_array);

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TIME_STOP(mpi_rank);
    H5D_MPIO_TRACE_EXIT(mpi_rank);
#endif

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__mpio_redistribute_shared_chunks_int() */

/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_share_chunk_modification_data
 *
 * Purpose:     When performing a parallel write on a chunked dataset with
 *              filters applied, we must first ensure that any particular
 *              chunk is only written to by a single MPI rank in order to
 *              avoid potential data races on the chunk. Once dataset
 *              chunks have been redistributed in a suitable manner, each
 *              MPI rank must send its chunk data to other ranks for each
 *              chunk it no longer owns.
 *
 *              The current implementation here follows the Nonblocking
 *              Consensus algorithm described in:
 *              http://unixer.de/publications/img/hoefler-dsde-protocols.pdf
 *
 *              First, each MPI rank scans through its list of selected
 *              chunks and does the following for each chunk:
 *
 *               * If a chunk in the MPI rank's chunk list is still owned
 *                 by that rank, the rank checks how many messages are
 *                 incoming for that chunk and adds that to its running
 *                 total. Then, the rank updates its local chunk list so
 *                 that any previous chunk entries for chunks that are no
 *                 longer owned by the rank get overwritten by chunk
 *                 entries for chunks the rank still owns. Since the data
 *                 for the chunks no longer owned will have already been
 *                 sent, those chunks can effectively be discarded.
 *               * If a chunk in the MPI rank's chunk list is no longer
 *                 owned by that rank, the rank sends the data it wishes to
 *                 update the chunk with to the MPI rank that now has
 *                 ownership of that chunk. To do this, it encodes the
 *                 chunk's index, its selection in the chunk and its
 *                 modification data into a buffer and then posts a
 *                 non-blocking MPI_Issend to the owning rank.
 *
 *              Once this step is complete, all MPI ranks allocate arrays
 *              to hold chunk message receive buffers and MPI request
 *              objects for each non-blocking receive they will post for
 *              incoming chunk modification messages. Then, all MPI ranks
 *              enter a loop that alternates between non-blocking
 *              MPI_Iprobe calls to probe for incoming messages and
 *              MPI_Testall calls to see if all send requests have
 *              completed. As chunk modification messages arrive,
 *              non-blocking MPI_Irecv calls will be posted for each
 *              message.
 *
 *              Once all send requests have completed, an MPI_Ibarrier is
 *              posted and the loop then alternates between MPI_Iprobe
 *              calls and MPI_Test calls to check if all ranks have reached
 *              the non-blocking barrier. Once all ranks have reached the
 *              barrier, processing can move on to updating the selected
 *              chunks that are owned in the operation.
 *
 *              Any chunk messages that were received from other ranks
 *              will be returned through the `chunk_msg_bufs` array and
 *              `chunk_msg_bufs_len` will be set appropriately.
 *
 *              NOTE: The use of non-blocking sends and receives of chunk
 *                    data here may contribute to large amounts of memory
 *                    usage/MPI request overhead if the number of shared
 *                    chunks is high. If this becomes a problem, it may be
 *                    useful to split the message receiving loop away so
 *                    that chunk modification messages can be received and
 *                    processed immediately (MPI_Recv) using a single chunk
 *                    message buffer. However, it's possible this may
 *                    degrade performance since the chunk message sends
 *                    are synchronous (MPI_Issend) in the Nonblocking
 *                    Consensus algorithm.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_share_chunk_modification_data(H5D_filtered_collective_io_info_t *chunk_list,
                                        size_t *chunk_list_num_entries, H5D_io_info_t *io_info,
                                        H5D_dset_io_info_t *dset_info, int mpi_rank,
                                        int H5_ATTR_NDEBUG_UNUSED           mpi_size,
                                        H5D_filtered_collective_io_info_t **chunk_hash_table,
                                        unsigned char ***chunk_msg_bufs, int *chunk_msg_bufs_len)
{
#if H5_CHECK_MPI_VERSION(3, 0)
    H5D_filtered_collective_io_info_t *chunk_table       = NULL;
    H5S_sel_iter_t                    *mem_iter          = NULL;
    unsigned char                    **msg_send_bufs     = NULL;
    unsigned char                    **msg_recv_bufs     = NULL;
    MPI_Request                       *send_requests     = NULL;
    MPI_Request                       *recv_requests     = NULL;
    MPI_Request                        ibarrier          = MPI_REQUEST_NULL;
    hbool_t                            mem_iter_init     = FALSE;
    hbool_t                            ibarrier_posted   = FALSE;
    size_t                             send_bufs_nalloc  = 0;
    size_t                             num_send_requests = 0;
    size_t                             num_recv_requests = 0;
    size_t                             num_msgs_incoming = 0;
    size_t                             last_assigned_idx;
    size_t                             i;
    int                                mpi_code;
    herr_t                             ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    HDassert(chunk_list_num_entries);
    HDassert(chunk_list || 0 == *chunk_list_num_entries);
    HDassert(io_info);
    HDassert(dset_info);
    HDassert(mpi_size > 1);
    HDassert(chunk_msg_bufs);
    HDassert(chunk_msg_bufs_len);

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TRACE_ENTER(mpi_rank);
    H5D_MPIO_TIME_START(mpi_rank, "Share chunk modification data");
#endif

    /* Set to latest format for encoding dataspace */
    H5CX_set_libver_bounds(NULL);

    if (*chunk_list_num_entries) {
        /* Allocate a selection iterator for iterating over chunk dataspaces */
        if (NULL == (mem_iter = H5FL_MALLOC(H5S_sel_iter_t)))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate dataspace selection iterator")

        /*
         * Allocate send buffer and MPI_Request arrays for non-blocking
         * sends of outgoing chunk messages
         */
        send_bufs_nalloc = H5D_CHUNK_NUM_SEND_MSGS_INIT;
        if (NULL == (msg_send_bufs = H5MM_malloc(send_bufs_nalloc * sizeof(*msg_send_bufs))))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL,
                        "couldn't allocate chunk modification message buffer array")

        if (NULL == (send_requests = H5MM_malloc(send_bufs_nalloc * sizeof(*send_requests))))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate send requests array")
    }

    /*
     * For each chunk this rank owns, add to the total number of
     * incoming MPI messages, then update the local chunk list to
     * overwrite any previous chunks no longer owned by this rank.
     * Since the data for those chunks will have already been sent,
     * this rank should no longer be interested in them and they
     * can effectively be discarded. This bookkeeping also makes
     * the code for the collective file space re-allocation and
     * chunk re-insertion operations a bit simpler.
     *
     * For each chunk this rank doesn't own, use non-blocking
     * synchronous sends to send the data this rank is writing to
     * the rank that does own the chunk.
     */
    for (i = 0, last_assigned_idx = 0; i < *chunk_list_num_entries; i++) {
        H5D_filtered_collective_io_info_t *chunk_entry = &chunk_list[i];

        if (mpi_rank == chunk_entry->new_owner) {
            num_msgs_incoming += (size_t)(chunk_entry->num_writers - 1);

            /*
             * Overwrite chunk entries this rank doesn't own with entries that it
             * does own, since it has sent the necessary data and is no longer
             * interested in the chunks it doesn't own.
             */
            chunk_list[last_assigned_idx] = chunk_list[i];

            /*
             * Since, at large scale, a chunk's index value may be larger than
             * the maximum value that can be stored in an int, we cannot rely
             * on using a chunk's index value as the tag for the MPI messages
             * sent/received for a chunk. Therefore, add this chunk to a hash
             * table with the chunk's index as a key so that we can quickly find
             * the chunk when processing chunk messages that were received. The
             * message itself will contain the chunk's index so we can update
             * the correct chunk with the received data.
             */
            HASH_ADD(hh, chunk_table, index_info.chunk_idx, sizeof(hsize_t), &chunk_list[last_assigned_idx]);

            last_assigned_idx++;
        }
        else {
            H5D_piece_info_t *chunk_info = chunk_entry->chunk_info;
            unsigned char    *mod_data_p = NULL;
            hsize_t           iter_nelmts;
            size_t            mod_data_size = 0;
            size_t            space_size    = 0;

            /* Add the size of the chunk index to the encoded size */
            mod_data_size += sizeof(hsize_t);

            /* Determine size of serialized chunk file dataspace */
            if (H5S_encode(chunk_info->fspace, &mod_data_p, &space_size) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "unable to get encoded dataspace size")
            mod_data_size += space_size;

            /* Determine size of data being written */
            iter_nelmts = H5S_GET_SELECT_NPOINTS(chunk_info->mspace);
            H5_CHECK_OVERFLOW(iter_nelmts, hsize_t, size_t);

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

            if (NULL == (msg_send_bufs[num_send_requests] = H5MM_malloc(mod_data_size)))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL,
                            "couldn't allocate chunk modification message buffer")

            mod_data_p = msg_send_bufs[num_send_requests];

            /* Store the chunk's index into the buffer */
            HDmemcpy(mod_data_p, &chunk_entry->index_info.chunk_idx, sizeof(hsize_t));
            mod_data_p += sizeof(hsize_t);

            /* Serialize the chunk's file dataspace into the buffer */
            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, dset_info->type_info.src_type_size,
                                     H5S_SEL_ITER_SHARE_WITH_DATASPACE) < 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 (0 == H5D__gather_mem(dset_info->buf.cvp, mem_iter, (size_t)iter_nelmts, mod_data_p))
                HGOTO_ERROR(H5E_IO, H5E_CANTGATHER, FAIL, "couldn't gather from write buffer")

            /*
             * Ensure that the size of the chunk data being sent can be
             * safely cast to an int for MPI. Note that this should
             * generally be OK for now (unless a rank is sending a
             * whole 32-bit-sized chunk of data + its encoded selection),
             * but if we allow larger than 32-bit-sized chunks in the
             * future, this may become a problem and derived datatypes
             * will need to be used.
             */
            H5_CHECK_OVERFLOW(mod_data_size, size_t, int)

            /* Send modification data to new owner */
            if (MPI_SUCCESS !=
                (mpi_code = MPI_Issend(msg_send_bufs[num_send_requests], (int)mod_data_size, MPI_BYTE,
                                       chunk_entry->new_owner, H5D_CHUNK_MOD_DATA_TAG, io_info->comm,
                                       &send_requests[num_send_requests])))
                HMPI_GOTO_ERROR(FAIL, "MPI_Issend failed", mpi_code)

            num_send_requests++;

            /* Resize send buffer and send request arrays if necessary */
            if (num_send_requests == send_bufs_nalloc) {
                void *tmp_alloc;

                send_bufs_nalloc = (size_t)((double)send_bufs_nalloc * 1.5);

                if (NULL ==
                    (tmp_alloc = H5MM_realloc(msg_send_bufs, send_bufs_nalloc * sizeof(*msg_send_bufs))))
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL,
                                "couldn't resize chunk modification message buffer array")
                msg_send_bufs = tmp_alloc;

                if (NULL ==
                    (tmp_alloc = H5MM_realloc(send_requests, send_bufs_nalloc * sizeof(*send_requests))))
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't resize send requests array")
                send_requests = tmp_alloc;
            }

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

    /* Check if the number of send or receive requests will overflow an int (MPI requirement) */
    if (num_send_requests > INT_MAX || num_msgs_incoming > INT_MAX)
        HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL,
                    "too many shared chunks in parallel filtered write operation")

    H5_CHECK_OVERFLOW(num_send_requests, size_t, int)
    H5_CHECK_OVERFLOW(num_msgs_incoming, size_t, int)

    /*
     * Allocate receive buffer and MPI_Request arrays for non-blocking
     * receives of incoming chunk messages
     */
    if (num_msgs_incoming) {
        if (NULL == (msg_recv_bufs = H5MM_malloc(num_msgs_incoming * sizeof(*msg_recv_bufs))))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL,
                        "couldn't allocate chunk modification message buffer array")

        if (NULL == (recv_requests = H5MM_malloc(num_msgs_incoming * sizeof(*recv_requests))))
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate receive requests array")
    }

    /* Process any incoming messages until everyone is done */
    do {
        MPI_Status status;
        int        msg_flag;

        /* Probe for an incoming message from any rank */
        if (MPI_SUCCESS != (mpi_code = MPI_Iprobe(MPI_ANY_SOURCE, H5D_CHUNK_MOD_DATA_TAG, io_info->comm,
                                                  &msg_flag, &status)))
            HMPI_GOTO_ERROR(FAIL, "MPI_Iprobe failed", mpi_code)

        /*
         * If a message was found, allocate a buffer for the message and
         * post a non-blocking receive to receive it
         */
        if (msg_flag) {
#if H5_CHECK_MPI_VERSION(3, 0)
            MPI_Count msg_size = 0;

            if (MPI_SUCCESS != (mpi_code = MPI_Get_elements_x(&status, MPI_BYTE, &msg_size)))
                HMPI_GOTO_ERROR(FAIL, "MPI_Get_elements_x failed", mpi_code)

            H5_CHECK_OVERFLOW(msg_size, MPI_Count, int)
#else
            int msg_size = 0;

            if (MPI_SUCCESS != (mpi_code = MPI_Get_elements(&status, MPI_BYTE, &msg_size)))
                HMPI_GOTO_ERROR(FAIL, "MPI_Get_elements failed", mpi_code)
#endif

            if (msg_size <= 0)
                HGOTO_ERROR(H5E_DATASET, H5E_BADVALUE, FAIL, "invalid chunk modification message size")

            HDassert((num_recv_requests + 1) <= num_msgs_incoming);
            if (NULL ==
                (msg_recv_bufs[num_recv_requests] = H5MM_malloc((size_t)msg_size * sizeof(unsigned char))))
                HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL,
                            "couldn't allocate chunk modification message receive buffer")

            if (MPI_SUCCESS != (mpi_code = MPI_Irecv(msg_recv_bufs[num_recv_requests], (int)msg_size,
                                                     MPI_BYTE, status.MPI_SOURCE, H5D_CHUNK_MOD_DATA_TAG,
                                                     io_info->comm, &recv_requests[num_recv_requests])))
                HMPI_GOTO_ERROR(FAIL, "MPI_Irecv failed", mpi_code)

            num_recv_requests++;
        }

        if (ibarrier_posted) {
            int ibarrier_completed;

            if (MPI_SUCCESS != (mpi_code = MPI_Test(&ibarrier, &ibarrier_completed, MPI_STATUS_IGNORE)))
                HMPI_GOTO_ERROR(FAIL, "MPI_Test failed", mpi_code)

            if (ibarrier_completed)
                break;
        }
        else {
            int all_sends_completed;

            /* Determine if all send requests have completed
             *
             * gcc 11 complains about passing MPI_STATUSES_IGNORE as an MPI_Status
             * array. See the discussion here:
             *
             * https://github.com/pmodels/mpich/issues/5687
             */
            H5_GCC_DIAG_OFF("stringop-overflow")
            if (MPI_SUCCESS != (mpi_code = MPI_Testall((int)num_send_requests, send_requests,
                                                       &all_sends_completed, MPI_STATUSES_IGNORE)))
                HMPI_GOTO_ERROR(FAIL, "MPI_Testall failed", mpi_code)
            H5_GCC_DIAG_ON("stringop-overflow")

            if (all_sends_completed) {
                /* Post non-blocking barrier */
                if (MPI_SUCCESS != (mpi_code = MPI_Ibarrier(io_info->comm, &ibarrier)))
                    HMPI_GOTO_ERROR(FAIL, "MPI_Ibarrier failed", mpi_code)
                ibarrier_posted = TRUE;

                /*
                 * Now that all send requests have completed, free up the
                 * send buffers used in the non-blocking operations
                 */
                if (msg_send_bufs) {
                    for (i = 0; i < num_send_requests; i++) {
                        if (msg_send_bufs[i])
                            H5MM_free(msg_send_bufs[i]);
                    }

                    msg_send_bufs = H5MM_xfree(msg_send_bufs);
                }
            }
        }
    } while (1);

    /*
     * Ensure all receive requests have completed before moving on.
     * For linked-chunk I/O, more overlap with computation could
     * theoretically be achieved by returning the receive requests
     * array and postponing this wait until during chunk updating
     * when the data is really needed. However, multi-chunk I/O
     * only updates a chunk at a time and the messages may not come
     * in the order that chunks are processed. So, the safest way to
     * support both I/O modes is to simply make sure all messages
     * are available.
     *
     * gcc 11 complains about passing MPI_STATUSES_IGNORE as an MPI_Status
     * array. See the discussion here:
     *
     * https://github.com/pmodels/mpich/issues/5687
     */
    H5_GCC_DIAG_OFF("stringop-overflow")
    if (MPI_SUCCESS != (mpi_code = MPI_Waitall((int)num_recv_requests, recv_requests, MPI_STATUSES_IGNORE)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Waitall failed", mpi_code)
    H5_GCC_DIAG_ON("stringop-overflow")

    /* Set the new number of locally-selected chunks */
    *chunk_list_num_entries = last_assigned_idx;

    /* Return chunk message buffers if any were received */
    *chunk_hash_table   = chunk_table;
    *chunk_msg_bufs     = msg_recv_bufs;
    *chunk_msg_bufs_len = (int)num_recv_requests;

done:
    if (ret_value < 0) {
        /* If this rank failed, make sure to participate in collective barrier */
        if (!ibarrier_posted) {
            if (MPI_SUCCESS != (mpi_code = MPI_Ibarrier(io_info->comm, &ibarrier)))
                HMPI_GOTO_ERROR(FAIL, "MPI_Ibarrier failed", mpi_code)
        }

        if (num_send_requests) {
            for (i = 0; i < num_send_requests; i++) {
                MPI_Cancel(&send_requests[i]);
            }
        }

        if (recv_requests) {
            for (i = 0; i < num_recv_requests; i++) {
                MPI_Cancel(&recv_requests[i]);
            }
        }

        if (msg_recv_bufs) {
            for (i = 0; i < num_recv_requests; i++) {
                H5MM_free(msg_recv_bufs[i]);
            }

            H5MM_free(msg_recv_bufs);
        }

        HASH_CLEAR(hh, chunk_table);
    }

    if (recv_requests)
        H5MM_free(recv_requests);
    if (send_requests)
        H5MM_free(send_requests);

    if (msg_send_bufs) {
        for (i = 0; i < num_send_requests; i++) {
            if (msg_send_bufs[i])
                H5MM_free(msg_send_bufs[i]);
        }

        H5MM_free(msg_send_bufs);
    }

    if (mem_iter) {
        if (mem_iter_init && H5S_SELECT_ITER_RELEASE(mem_iter) < 0)
            HDONE_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release dataspace selection iterator")
        mem_iter = H5FL_FREE(H5S_sel_iter_t, mem_iter);
    }

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TIME_STOP(mpi_rank);
    H5D_MPIO_TRACE_EXIT(mpi_rank);
#endif

    FUNC_LEAVE_NOAPI(ret_value)
#else
    FUNC_ENTER_PACKAGE
    HERROR(
        H5E_DATASET, H5E_WRITEERROR,
        "unable to send chunk modification data between MPI ranks - MPI version < 3 (MPI_Ibarrier missing)")
    FUNC_LEAVE_NOAPI(FAIL)
#endif
} /* end H5D__mpio_share_chunk_modification_data() */

/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_collective_filtered_chunk_common_io
 *
 * Purpose:     This routine performs the common part of collective I/O
 *              when reading or writing filtered chunks collectively.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_collective_filtered_chunk_common_io(H5D_filtered_collective_io_info_t *chunk_list,
                                              size_t chunk_list_num_entries, const H5D_io_info_t *io_info,
                                              int mpi_size)
{
    H5D_io_info_t coll_io_info;
    MPI_Datatype  file_type            = MPI_DATATYPE_NULL;
    MPI_Datatype  mem_type             = MPI_DATATYPE_NULL;
    hbool_t       mem_type_is_derived  = FALSE;
    hbool_t       file_type_is_derived = FALSE;
    hsize_t       mpi_buf_count;
    haddr_t       base_read_offset = HADDR_UNDEF;
    size_t        num_chunks;
    size_t        i;
    char          fake_buf; /* Used as a fake buffer for ranks with no chunks, thus a NULL buf pointer */
    int           mpi_code;
    herr_t        ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    HDassert(chunk_list || 0 == chunk_list_num_entries);
    HDassert(io_info);

    /* Initialize temporary I/O info */
    coll_io_info = *io_info;

    /*
     * Construct MPI derived datatype for collective I/O on chunks
     */
    if (H5D__mpio_collective_filtered_io_type(chunk_list, chunk_list_num_entries, io_info->op_type, &mem_type,
                                              &mem_type_is_derived, &file_type, &file_type_is_derived) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_BADTYPE, FAIL, "couldn't create MPI I/O type for chunk I/O")

    /*
     * For reads, determine how many chunks are actually being read.
     * Note that if this is a read during a write operation
     * (read chunk -> unfilter -> modify -> write back), some
     * chunks may not need to be read if they're being fully
     * overwritten during a write operation.
     */
    if (io_info->op_type == H5D_IO_OP_READ) {
        for (i = 0, num_chunks = 0; i < chunk_list_num_entries; i++) {
            HDassert(chunk_list[i].buf);

            if (chunk_list[i].need_read) {
                if (!H5F_addr_defined(base_read_offset))
                    base_read_offset = chunk_list[i].chunk_current.offset;

                num_chunks++;
            }
        }
    }
    else
        num_chunks = chunk_list_num_entries;

    /*
     * If this rank doesn't have a selection, it can
     * skip I/O if the MPI communicator size is 1.
     *
     * Otherwise, this rank has to participate in
     * collective I/O, but probably has a NULL buf
     * pointer, so override to a fake buffer since our
     * write/read function expects one.
     */
    if (num_chunks == 0) {
        if (mpi_size == 1)
            HGOTO_DONE(SUCCEED)
        else {
            if (io_info->op_type == H5D_IO_OP_WRITE)
                coll_io_info.base_maddr.cvp = &fake_buf;
            else
                coll_io_info.base_maddr.vp = &fake_buf;
        }
    }

    /*
     * Setup for I/O operation
     */

    mpi_buf_count = (num_chunks) ? 1 : 0;

    if (num_chunks) {
        /*
         * Setup the base storage address for this operation
         * to be the first chunk's file address
         */
        if (io_info->op_type == H5D_IO_OP_WRITE)
            coll_io_info.store_faddr = chunk_list[0].chunk_new.offset;
        else
            coll_io_info.store_faddr = base_read_offset;
    }
    else
        coll_io_info.store_faddr = 0;

    /* Perform I/O */
    if (H5D__final_collective_io(&coll_io_info, mpi_buf_count, file_type, mem_type) < 0)
        HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "couldn't finish MPI I/O")

done:
    /* 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__mpio_collective_filtered_chunk_common_io() */

/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_collective_filtered_chunk_read
 *
 * Purpose:     This routine coordinates a collective read across all ranks
 *              of the chunks they have selected. Each rank will then go
 *              and
 *
 * Return:      Non-negative on success/Negative on failure
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_collective_filtered_chunk_read(H5D_filtered_collective_io_info_t *chunk_list,
                                         size_t chunk_list_num_entries, const H5D_io_info_t *io_info,
                                         const H5D_dset_io_info_t *di, int mpi_rank, int mpi_size)
{
    H5D_fill_buf_info_t fb_info;
    H5D_piece_info_t   *chunk_info = NULL;
    H5D_io_info_t       coll_io_info;
    H5Z_EDC_t           err_detect; /* Error detection info */
    H5Z_cb_t            filter_cb;  /* I/O filter callback function */
    hsize_t             file_chunk_size = 0;
    hsize_t             iter_nelmts; /* Number of points to iterate over for the chunk IO operation */
    hbool_t             should_fill  = FALSE;
    hbool_t             fb_info_init = FALSE;
    hbool_t             index_empty  = FALSE;
    size_t              i;
    H5S_t              *fill_space    = NULL;
    void               *base_read_buf = NULL;
    herr_t              ret_value     = SUCCEED;

    FUNC_ENTER_PACKAGE

    HDassert(chunk_list || 0 == chunk_list_num_entries);
    HDassert(io_info);
    HDassert(di);

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TRACE_ENTER(mpi_rank);
    H5D_MPIO_TIME_START(mpi_rank, "Filtered collective chunk read");
#else
    (void)mpi_rank;
#endif

    /* Initialize temporary I/O info */
    coll_io_info               = *io_info;
    coll_io_info.base_maddr.vp = NULL;

    if (chunk_list_num_entries) {
        /* 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")

        /* Set size of full chunks in dataset */
        file_chunk_size = di->dset->shared->layout.u.chunk.size;

        /* Determine if fill values should be "read" for unallocated chunks */
        should_fill = (di->dset->shared->dcpl_cache.fill.fill_time == H5D_FILL_TIME_ALLOC) ||
                      ((di->dset->shared->dcpl_cache.fill.fill_time == H5D_FILL_TIME_IFSET) &&
                       di->dset->shared->dcpl_cache.fill.fill_defined);
    }

    /*
     * Allocate memory buffers for all chunks being read. Chunk data buffers are of
     * the largest size between the chunk's current filtered size and the chunk's true
     * size, as calculated by the number of elements in the chunk's file space extent
     * multiplied by the datatype size. This tries to ensure that:
     *
     *  * If we're reading the chunk and the filter normally reduces the chunk size,
     *    the unfiltering operation won't need to grow the buffer.
     *  * If we're reading the chunk and the filter normally grows the chunk size,
     *    we make sure to read into a buffer of size equal to the filtered chunk's
     *    size; reading into a (smaller) buffer of size equal to the unfiltered
     *    chunk size would of course be bad.
     */
    for (i = 0; i < chunk_list_num_entries; i++) {
        HDassert(chunk_list[i].need_read);

        chunk_list[i].chunk_buf_size = MAX(chunk_list[i].chunk_current.length, file_chunk_size);

        if (NULL == (chunk_list[i].buf = H5MM_malloc(chunk_list[i].chunk_buf_size))) {
            /* Push an error, but participate in collective read */
            HDONE_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk data buffer")
            break;
        }

        /*
         * Check if chunk is currently allocated. If not, don't try to
         * read it from the file. Instead, just fill the chunk buffer
         * with the fill value if necessary.
         */
        if (H5F_addr_defined(chunk_list[i].chunk_current.offset)) {
            /* Set first read buffer */
            if (!base_read_buf)
                base_read_buf = chunk_list[i].buf;

            /* Set chunk's new length for eventual filter pipeline calls */
            if (chunk_list[i].skip_filter_pline)
                chunk_list[i].chunk_new.length = file_chunk_size;
            else
                chunk_list[i].chunk_new.length = chunk_list[i].chunk_current.length;
        }
        else {
            chunk_list[i].need_read = FALSE;

            /* Set chunk's new length for eventual filter pipeline calls */
            chunk_list[i].chunk_new.length = file_chunk_size;

            if (should_fill) {
                /* Initialize fill value buffer if not already initialized */
                if (!fb_info_init) {
                    hsize_t chunk_dims[H5S_MAX_RANK];

                    HDassert(di->dset->shared->ndims == di->dset->shared->layout.u.chunk.ndims - 1);
                    for (size_t j = 0; j < di->dset->shared->layout.u.chunk.ndims - 1; j++)
                        chunk_dims[j] = (hsize_t)di->dset->shared->layout.u.chunk.dim[j];

                    /* Get a dataspace for filling chunk memory buffers */
                    if (NULL == (fill_space = H5S_create_simple(di->dset->shared->layout.u.chunk.ndims - 1,
                                                                chunk_dims, NULL)))
                        HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "unable to create chunk fill dataspace")

                    /* Initialize fill value buffer */
                    if (H5D__fill_init(
                            &fb_info, NULL, (H5MM_allocate_t)H5D__chunk_mem_alloc,
                            (void *)&di->dset->shared->dcpl_cache.pline, (H5MM_free_t)H5D__chunk_mem_free,
                            (void *)&di->dset->shared->dcpl_cache.pline, &di->dset->shared->dcpl_cache.fill,
                            di->dset->shared->type, di->dset->shared->type_id, 0, file_chunk_size) < 0)
                        HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "can't initialize fill value buffer")

                    fb_info_init = TRUE;
                }

                /* Write fill value to memory buffer */
                HDassert(fb_info.fill_buf);
                if (H5D__fill(fb_info.fill_buf, di->dset->shared->type, chunk_list[i].buf,
                              di->type_info.mem_type, fill_space) < 0)
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "couldn't fill chunk buffer with fill value")
            }
        }
    }

    /*
     * If dataset is incrementally allocated and hasn't been written to
     * yet, the chunk index should be empty. In this case, a collective
     * read of chunks is essentially a no-op, so avoid it here.
     */
    index_empty = FALSE;
    if (di->dset->shared->dcpl_cache.fill.alloc_time == H5D_ALLOC_TIME_INCR)
        if (H5D__chunk_index_empty(di->dset, &index_empty) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "couldn't determine if chunk index is empty")

    if (!index_empty) {
        /*
         * Override the read buffer to point to the address of
         * the first chunk data buffer being read into
         */
        if (base_read_buf)
            coll_io_info.base_maddr.vp = base_read_buf;

        /* Perform collective chunk read */
        if (H5D__mpio_collective_filtered_chunk_common_io(chunk_list, chunk_list_num_entries, &coll_io_info,
                                                          mpi_size) < 0)
            HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "couldn't finish collective filtered chunk read")
    }

    /*
     * Iterate through all the read chunks, unfiltering them and scattering their
     * data out to the application's read buffer.
     */
    for (i = 0; i < chunk_list_num_entries; i++) {
        chunk_info = chunk_list[i].chunk_info;

        /* Unfilter the chunk, unless we didn't read it from the file */
        if (chunk_list[i].need_read && !chunk_list[i].skip_filter_pline) {
            if (H5Z_pipeline(&di->dset->shared->dcpl_cache.pline, H5Z_FLAG_REVERSE,
                             &(chunk_list[i].index_info.filter_mask), err_detect, filter_cb,
                             (size_t *)&chunk_list[i].chunk_new.length, &chunk_list[i].chunk_buf_size,
                             &chunk_list[i].buf) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTFILTER, FAIL, "couldn't unfilter chunk for modifying")
        }

        /* Scatter the chunk data to the read buffer */
        iter_nelmts = H5S_GET_SELECT_NPOINTS(chunk_info->fspace);

        if (H5D_select_io_mem(di->buf.vp, chunk_info->mspace, chunk_list[i].buf, chunk_info->fspace,
                              di->type_info.src_type_size, (size_t)iter_nelmts) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "couldn't copy chunk data to read buffer")
    }

done:
    /* Free all resources used by entries in the chunk list */
    for (i = 0; i < chunk_list_num_entries; i++) {
        if (chunk_list[i].buf) {
            H5MM_free(chunk_list[i].buf);
            chunk_list[i].buf = NULL;
        }
    }

    /* Release the fill buffer info, if it's been initialized */
    if (fb_info_init && H5D__fill_term(&fb_info) < 0)
        HDONE_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "Can't release fill buffer info")
    if (fill_space && (H5S_close(fill_space) < 0))
        HDONE_ERROR(H5E_DATASET, H5E_CLOSEERROR, FAIL, "can't close fill space")

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TIME_STOP(mpi_rank);
    H5D_MPIO_TRACE_EXIT(mpi_rank);
#endif

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__mpio_collective_filtered_chunk_read() */

/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_collective_filtered_chunk_update
 *
 * Purpose:     When performing a parallel write on a chunked dataset with
 *              filters applied, all ranks must update their owned chunks
 *              with their own modification data and data from other ranks.
 *              This routine is responsible for coordinating that process.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_collective_filtered_chunk_update(H5D_filtered_collective_io_info_t *chunk_list,
                                           size_t                             chunk_list_num_entries,
                                           H5D_filtered_collective_io_info_t *chunk_hash_table,
                                           unsigned char **chunk_msg_bufs, int chunk_msg_bufs_len,
                                           const H5D_io_info_t *io_info, const H5D_dset_io_info_t *di,
                                           int H5_ATTR_NDEBUG_UNUSED mpi_rank, int mpi_size)
{
    const H5D_type_info_t *type_info = NULL;
    H5D_fill_buf_info_t    fb_info;
    H5D_piece_info_t      *chunk_info = NULL;
    H5S_sel_iter_t        *sel_iter   = NULL; /* Dataspace selection iterator for H5D__scatter_mem */
    H5D_io_info_t          coll_io_info;
    H5Z_EDC_t              err_detect; /* Error detection info */
    H5Z_cb_t               filter_cb;  /* I/O filter callback function */
    hsize_t                file_chunk_size = 0;
    hsize_t                iter_nelmts; /* Number of points to iterate over for the chunk IO operation */
    hbool_t                should_fill   = FALSE;
    hbool_t                fb_info_init  = FALSE;
    hbool_t                sel_iter_init = FALSE;
    hbool_t                index_empty   = FALSE;
    size_t                 i;
    H5S_t                 *dataspace     = NULL;
    H5S_t                 *fill_space    = NULL;
    void                  *base_read_buf = NULL;
    herr_t                 ret_value     = SUCCEED;

    FUNC_ENTER_PACKAGE

    HDassert(chunk_list || 0 == chunk_list_num_entries);
    HDassert((chunk_msg_bufs && chunk_hash_table) || 0 == chunk_msg_bufs_len);
    HDassert(io_info);
    HDassert(di);

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TRACE_ENTER(mpi_rank);
    H5D_MPIO_TIME_START(mpi_rank, "Filtered collective chunk update");
#endif

    /* Set convenience pointers */
    type_info = &(di->type_info);
    HDassert(type_info);

    if (chunk_list_num_entries) {
        /* 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")

        /* Set size of full chunks in dataset */
        file_chunk_size = di->dset->shared->layout.u.chunk.size;

        /* Determine if fill values should be written to chunks */
        should_fill = (di->dset->shared->dcpl_cache.fill.fill_time == H5D_FILL_TIME_ALLOC) ||
                      ((di->dset->shared->dcpl_cache.fill.fill_time == H5D_FILL_TIME_IFSET) &&
                       di->dset->shared->dcpl_cache.fill.fill_defined);
    }

    /*
     * Allocate memory buffers for all owned chunks. Chunk data buffers are of the
     * largest size between the chunk's current filtered size and the chunk's true
     * size, as calculated by the number of elements in the chunk's file space extent
     * multiplied by the datatype size. This tries to ensure that:
     *
     *  * If we're fully overwriting the chunk and the filter normally reduces the
     *    chunk size, we simply have the exact buffer size required to hold the
     *    unfiltered chunk data.
     *  * If we're fully overwriting the chunk and the filter normally grows the
     *    chunk size (e.g., fletcher32 filter), the final filtering operation
     *    (hopefully) won't need to grow the buffer.
     *  * If we're reading the chunk and the filter normally reduces the chunk size,
     *    the unfiltering operation won't need to grow the buffer.
     *  * If we're reading the chunk and the filter normally grows the chunk size,
     *    we make sure to read into a buffer of size equal to the filtered chunk's
     *    size; reading into a (smaller) buffer of size equal to the unfiltered
     *    chunk size would of course be bad.
     */
    for (i = 0; i < chunk_list_num_entries; i++) {
        HDassert(mpi_rank == chunk_list[i].new_owner);

        chunk_list[i].chunk_buf_size = MAX(chunk_list[i].chunk_current.length, file_chunk_size);

        /*
         * If this chunk hasn't been allocated yet and we aren't writing
         * out fill values to it, make sure to 0-fill its memory buffer
         * so we don't use uninitialized memory.
         */
        if (!H5F_addr_defined(chunk_list[i].chunk_current.offset) && !should_fill)
            chunk_list[i].buf = H5MM_calloc(chunk_list[i].chunk_buf_size);
        else
            chunk_list[i].buf = H5MM_malloc(chunk_list[i].chunk_buf_size);

        if (NULL == chunk_list[i].buf) {
            /* Push an error, but participate in collective read */
            HDONE_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate chunk data buffer")
            break;
        }

        /* Set chunk's new length for eventual filter pipeline calls */
        if (chunk_list[i].need_read) {
            /*
             * Check if chunk is currently allocated. If not, don't try to
             * read it from the file. Instead, just fill the chunk buffer
             * with the fill value if fill values are to be written.
             */
            if (H5F_addr_defined(chunk_list[i].chunk_current.offset)) {
                /* Set first read buffer */
                if (!base_read_buf)
                    base_read_buf = chunk_list[i].buf;

                /* Set chunk's new length for eventual filter pipeline calls */
                if (chunk_list[i].skip_filter_pline)
                    chunk_list[i].chunk_new.length = file_chunk_size;
                else
                    chunk_list[i].chunk_new.length = chunk_list[i].chunk_current.length;
            }
            else {
                chunk_list[i].need_read = FALSE;

                /* Set chunk's new length for eventual filter pipeline calls */
                chunk_list[i].chunk_new.length = file_chunk_size;

                if (should_fill) {
                    /* Initialize fill value buffer if not already initialized */
                    if (!fb_info_init) {
                        hsize_t chunk_dims[H5S_MAX_RANK];

                        HDassert(di->dset->shared->ndims == di->dset->shared->layout.u.chunk.ndims - 1);
                        for (size_t j = 0; j < di->dset->shared->layout.u.chunk.ndims - 1; j++)
                            chunk_dims[j] = (hsize_t)di->dset->shared->layout.u.chunk.dim[j];

                        /* Get a dataspace for filling chunk memory buffers */
                        if (NULL == (fill_space = H5S_create_simple(
                                         di->dset->shared->layout.u.chunk.ndims - 1, chunk_dims, NULL)))
                            HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL,
                                        "unable to create chunk fill dataspace")

                        /* Initialize fill value buffer */
                        if (H5D__fill_init(&fb_info, NULL, (H5MM_allocate_t)H5D__chunk_mem_alloc,
                                           (void *)&di->dset->shared->dcpl_cache.pline,
                                           (H5MM_free_t)H5D__chunk_mem_free,
                                           (void *)&di->dset->shared->dcpl_cache.pline,
                                           &di->dset->shared->dcpl_cache.fill, di->dset->shared->type,
                                           di->dset->shared->type_id, 0, file_chunk_size) < 0)
                            HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "can't initialize fill value buffer")

                        fb_info_init = TRUE;
                    }

                    /* Write fill value to memory buffer */
                    HDassert(fb_info.fill_buf);
                    if (H5D__fill(fb_info.fill_buf, di->dset->shared->type, chunk_list[i].buf,
                                  type_info->mem_type, fill_space) < 0)
                        HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL,
                                    "couldn't fill chunk buffer with fill value")
                }
            }
        }
        else
            chunk_list[i].chunk_new.length = file_chunk_size;
    }

    /*
     * If dataset is incrementally allocated and hasn't been written to
     * yet, the chunk index should be empty. In this case, a collective
     * read of chunks is essentially a no-op, so avoid it here.
     */
    index_empty = FALSE;
    if (di->dset->shared->dcpl_cache.fill.alloc_time == H5D_ALLOC_TIME_INCR)
        if (H5D__chunk_index_empty(di->dset, &index_empty) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "couldn't determine if chunk index is empty")

    if (!index_empty) {
        /*
         * Setup for I/O operation
         */

        /* Initialize temporary I/O info */
        coll_io_info         = *io_info;
        coll_io_info.op_type = H5D_IO_OP_READ;

        /* Override the read buffer to point to the address of the first
         * chunk data buffer being read into
         */
        if (base_read_buf) {
            coll_io_info.base_maddr.vp = base_read_buf;
        }

        /* Read all chunks that need to be read from the file */
        if (H5D__mpio_collective_filtered_chunk_common_io(chunk_list, chunk_list_num_entries, &coll_io_info,
                                                          mpi_size) < 0)
            HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "couldn't finish collective filtered chunk read")
    }

    /*
     * Now that all owned chunks have been read, update the chunks
     * with modification data from the owning rank and other ranks.
     */

    /* Process all chunks with data from the owning rank first */
    for (i = 0; i < chunk_list_num_entries; i++) {
        HDassert(mpi_rank == chunk_list[i].new_owner);

        chunk_info = chunk_list[i].chunk_info;

        /*
         * If this chunk wasn't being fully overwritten, we read it from
         * the file, so we need to unfilter it
         */
        if (chunk_list[i].need_read && !chunk_list[i].skip_filter_pline) {
            if (H5Z_pipeline(&di->dset->shared->dcpl_cache.pline, H5Z_FLAG_REVERSE,
                             &(chunk_list[i].index_info.filter_mask), err_detect, filter_cb,
                             (size_t *)&chunk_list[i].chunk_new.length, &chunk_list[i].chunk_buf_size,
                             &chunk_list[i].buf) < 0)
                HGOTO_ERROR(H5E_DATASET, H5E_CANTFILTER, FAIL, "couldn't unfilter chunk for modifying")
        }

        iter_nelmts = H5S_GET_SELECT_NPOINTS(chunk_info->mspace);

        if (H5D_select_io_mem(chunk_list[i].buf, chunk_info->fspace, di->buf.cvp, chunk_info->mspace,
                              type_info->dst_type_size, (size_t)iter_nelmts) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "couldn't copy chunk data to write buffer")
    }

    /* Allocate iterator for memory selection */
    if (NULL == (sel_iter = H5FL_MALLOC(H5S_sel_iter_t)))
        HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "couldn't allocate memory iterator")

    /* Now process all received chunk message buffers */
    for (i = 0; i < (size_t)chunk_msg_bufs_len; i++) {
        H5D_filtered_collective_io_info_t *chunk_entry = NULL;
        const unsigned char               *msg_ptr     = chunk_msg_bufs[i];
        hsize_t                            chunk_idx;

        if (msg_ptr) {
            /* Retrieve the chunk's index value */
            HDmemcpy(&chunk_idx, msg_ptr, sizeof(hsize_t));
            msg_ptr += sizeof(hsize_t);

            /* Find the chunk entry according to its chunk index */
            HASH_FIND(hh, chunk_hash_table, &chunk_idx, sizeof(hsize_t), chunk_entry);
            HDassert(chunk_entry);
            HDassert(mpi_rank == chunk_entry->new_owner);

            /*
             * Only process the chunk if its data buffer is allocated.
             * In the case of multi-chunk I/O, we're only working on
             * a chunk at a time, so we need to skip over messages
             * that aren't for the chunk we're currently working on.
             */
            if (!chunk_entry->buf)
                continue;
            else {
                /* Decode the chunk file dataspace from the message */
                if (NULL == (dataspace = H5S_decode(&msg_ptr)))
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTDECODE, FAIL, "unable to decode dataspace")

                if (H5S_select_iter_init(sel_iter, dataspace, type_info->dst_type_size,
                                         H5S_SEL_ITER_SHARE_WITH_DATASPACE) < 0)
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL,
                                "unable to initialize memory selection information")
                sel_iter_init = TRUE;

                iter_nelmts = H5S_GET_SELECT_NPOINTS(dataspace);

                /* Update the chunk data with the received modification data */
                if (H5D__scatter_mem(msg_ptr, sel_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(sel_iter) < 0)
                    HGOTO_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release selection iterator")
                sel_iter_init = FALSE;

                if (dataspace) {
                    if (H5S_close(dataspace) < 0)
                        HGOTO_ERROR(H5E_DATASPACE, H5E_CANTFREE, FAIL, "can't close dataspace")
                    dataspace = NULL;
                }

                H5MM_free(chunk_msg_bufs[i]);
                chunk_msg_bufs[i] = NULL;
            }
        }
    }

    /* Finally, filter all the chunks */
    for (i = 0; i < chunk_list_num_entries; i++) {
        if (!chunk_list[i].skip_filter_pline) {
            if (H5Z_pipeline(&di->dset->shared->dcpl_cache.pline, 0, &(chunk_list[i].index_info.filter_mask),
                             err_detect, filter_cb, (size_t *)&chunk_list[i].chunk_new.length,
                             &chunk_list[i].chunk_buf_size, &chunk_list[i].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_list[i].chunk_new.length > ((size_t)0xffffffff))
            HGOTO_ERROR(H5E_DATASET, H5E_BADRANGE, FAIL, "chunk too large for 32-bit length")
#endif
    }

done:
    if (sel_iter) {
        if (sel_iter_init && H5S_SELECT_ITER_RELEASE(sel_iter) < 0)
            HDONE_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "couldn't release selection iterator")
        sel_iter = H5FL_FREE(H5S_sel_iter_t, sel_iter);
    }
    if (dataspace && (H5S_close(dataspace) < 0))
        HDONE_ERROR(H5E_DATASPACE, H5E_CANTFREE, FAIL, "can't close dataspace")
    if (fill_space && (H5S_close(fill_space) < 0))
        HDONE_ERROR(H5E_DATASET, H5E_CLOSEERROR, FAIL, "can't close fill space")

    /* Release the fill buffer info, if it's been initialized */
    if (fb_info_init && H5D__fill_term(&fb_info) < 0)
        HDONE_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "Can't release fill buffer info")

    /* On failure, try to free all resources used by entries in the chunk list */
    if (ret_value < 0) {
        for (i = 0; i < chunk_list_num_entries; i++) {
            if (chunk_list[i].buf) {
                H5MM_free(chunk_list[i].buf);
                chunk_list[i].buf = NULL;
            }
        }
    }

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TIME_STOP(mpi_rank);
    H5D_MPIO_TRACE_EXIT(mpi_rank);
#endif

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__mpio_collective_filtered_chunk_update() */

/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_collective_filtered_chunk_reallocate
 *
 * Purpose:     When performing a parallel write on a chunked dataset with
 *              filters applied, all ranks must eventually get together and
 *              perform a collective reallocation of space in the file for
 *              all chunks that were modified on all ranks. This routine is
 *              responsible for coordinating that process.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_collective_filtered_chunk_reallocate(H5D_filtered_collective_io_info_t *chunk_list,
                                               size_t chunk_list_num_entries, size_t *num_chunks_assigned_map,
                                               H5D_io_info_t *io_info, H5D_chk_idx_info_t *idx_info,
                                               int mpi_rank, int mpi_size)
{
    H5D_chunk_alloc_info_t *collective_list = NULL;
    MPI_Datatype            send_type;
    MPI_Datatype            recv_type;
    hbool_t                 send_type_derived          = FALSE;
    hbool_t                 recv_type_derived          = FALSE;
    hbool_t                 need_sort                  = FALSE;
    size_t                  collective_num_entries     = 0;
    size_t                  num_local_chunks_processed = 0;
    size_t                  i;
    void                   *gathered_array     = NULL;
    int                    *counts_disps_array = NULL;
    int                    *counts_ptr         = NULL;
    int                    *displacements_ptr  = NULL;
    int                     mpi_code;
    herr_t                  ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    HDassert(chunk_list || 0 == chunk_list_num_entries);
    HDassert(io_info);
    HDassert(idx_info);
    HDassert(idx_info->storage->idx_type != H5D_CHUNK_IDX_NONE);

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TRACE_ENTER(mpi_rank);
    H5D_MPIO_TIME_START(mpi_rank, "Reallocation of chunk file space");
#endif

    /*
     * Make sure it's safe to cast this rank's number
     * of chunks to be sent into an int for MPI
     */
    H5_CHECK_OVERFLOW(chunk_list_num_entries, size_t, int);

    /* Create derived datatypes for the chunk file space info needed */
    if (H5D__mpio_get_chunk_alloc_info_types(&recv_type, &recv_type_derived, &send_type, &send_type_derived) <
        0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL,
                    "can't create derived datatypes for chunk file space info")

    /*
     * Gather the new chunk sizes to all ranks for a collective reallocation
     * of the chunks in the file.
     */
    if (num_chunks_assigned_map) {
        /*
         * If a mapping between rank value -> number of assigned chunks has
         * been provided (usually during linked-chunk I/O), we can use this
         * to optimize MPI overhead a bit since MPI ranks won't need to
         * first inform each other about how many chunks they're contributing.
         */
        if (NULL == (counts_disps_array = H5MM_malloc(2 * (size_t)mpi_size * sizeof(*counts_disps_array)))) {
            /* Push an error, but still participate in collective gather operation */
            HDONE_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL,
                        "couldn't allocate receive counts and displacements array")
        }
        else {
            /* Set the receive counts from the assigned chunks map */
            counts_ptr = counts_disps_array;

            for (i = 0; i < (size_t)mpi_size; i++)
                H5_CHECKED_ASSIGN(counts_ptr[i], int, num_chunks_assigned_map[i], size_t);

            /* Set the displacements into the receive buffer for the gather operation */
            displacements_ptr = &counts_disps_array[mpi_size];

            *displacements_ptr = 0;
            for (i = 1; i < (size_t)mpi_size; i++)
                displacements_ptr[i] = displacements_ptr[i - 1] + counts_ptr[i - 1];
        }

        /* Perform gather operation */
        if (H5_mpio_gatherv_alloc(chunk_list, (int)chunk_list_num_entries, send_type, counts_ptr,
                                  displacements_ptr, recv_type, TRUE, 0, io_info->comm, mpi_rank, mpi_size,
                                  &gathered_array, &collective_num_entries) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTGATHER, FAIL, "can't gather chunk file space info to/from ranks")
    }
    else {
        /*
         * If no mapping between rank value -> number of assigned chunks has
         * been provided (usually during multi-chunk I/O), all MPI ranks will
         * need to first inform other ranks about how many chunks they're
         * contributing before performing the actual gather operation. Use
         * the 'simple' MPI_Allgatherv wrapper for this.
         */
        if (H5_mpio_gatherv_alloc_simple(chunk_list, (int)chunk_list_num_entries, send_type, recv_type, TRUE,
                                         0, io_info->comm, mpi_rank, mpi_size, &gathered_array,
                                         &collective_num_entries) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTGATHER, FAIL, "can't gather chunk file space info to/from ranks")
    }

    /* Collectively re-allocate the modified chunks (from each rank) in the file */
    collective_list = (H5D_chunk_alloc_info_t *)gathered_array;
    for (i = 0, num_local_chunks_processed = 0; i < collective_num_entries; i++) {
        H5D_chunk_alloc_info_t *coll_entry = &collective_list[i];
        hbool_t                 need_insert;
        hbool_t                 update_local_chunk;

        if (H5D__chunk_file_alloc(idx_info, &coll_entry->chunk_current, &coll_entry->chunk_new, &need_insert,
                                  NULL) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate chunk")

        /*
         * If we just re-allocated a chunk that is local to this
         * rank, make sure to update the chunk entry in the local
         * chunk list
         */
        update_local_chunk =
            (num_local_chunks_processed < chunk_list_num_entries) &&
            (coll_entry->chunk_idx == chunk_list[num_local_chunks_processed].index_info.chunk_idx);

        if (update_local_chunk) {
            H5D_filtered_collective_io_info_t *local_chunk;

            local_chunk = &chunk_list[num_local_chunks_processed];

            /* Sanity check that this chunk is actually local */
            HDassert(mpi_rank == local_chunk->orig_owner);
            HDassert(mpi_rank == local_chunk->new_owner);

            local_chunk->chunk_new              = coll_entry->chunk_new;
            local_chunk->index_info.need_insert = need_insert;

            /*
             * Since chunk reallocation can move chunks around, check if
             * the local chunk list is still in ascending offset of order
             * in the file
             */
            if (num_local_chunks_processed) {
                haddr_t curr_chunk_offset = local_chunk->chunk_new.offset;
                haddr_t prev_chunk_offset = chunk_list[num_local_chunks_processed - 1].chunk_new.offset;

                HDassert(H5F_addr_defined(prev_chunk_offset) && H5F_addr_defined(curr_chunk_offset));
                if (curr_chunk_offset < prev_chunk_offset)
                    need_sort = TRUE;
            }

            num_local_chunks_processed++;
        }
    }

    HDassert(chunk_list_num_entries == num_local_chunks_processed);

    /*
     * Ensure this rank's local chunk list is sorted in
     * ascending order of offset in the file
     */
    if (need_sort)
        HDqsort(chunk_list, chunk_list_num_entries, sizeof(H5D_filtered_collective_io_info_t),
                H5D__cmp_filtered_collective_io_info_entry);

done:
    H5MM_free(gathered_array);
    H5MM_free(counts_disps_array);

    if (send_type_derived) {
        if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&send_type)))
            HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
    }
    if (recv_type_derived) {
        if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&recv_type)))
            HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
    }

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TIME_STOP(mpi_rank);
    H5D_MPIO_TRACE_EXIT(mpi_rank);
#endif

    FUNC_LEAVE_NOAPI(ret_value)
} /* H5D__mpio_collective_filtered_chunk_reallocate() */

/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_collective_filtered_chunk_reinsert
 *
 * Purpose:     When performing a parallel write on a chunked dataset with
 *              filters applied, all ranks must eventually get together and
 *              perform a collective reinsertion into the dataset's chunk
 *              index of chunks that were modified. This routine is
 *              responsible for coordinating that process.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_collective_filtered_chunk_reinsert(H5D_filtered_collective_io_info_t *chunk_list,
                                             size_t chunk_list_num_entries, size_t *num_chunks_assigned_map,
                                             H5D_io_info_t *io_info, H5D_dset_io_info_t *di,
                                             H5D_chk_idx_info_t *idx_info, int mpi_rank, int mpi_size)
{
    H5D_chunk_ud_t chunk_ud;
    MPI_Datatype   send_type;
    MPI_Datatype   recv_type;
    hbool_t        send_type_derived = FALSE;
    hbool_t        recv_type_derived = FALSE;
    hsize_t        scaled_coords[H5O_LAYOUT_NDIMS];
    size_t         collective_num_entries = 0;
    size_t         i;
    void          *gathered_array     = NULL;
    int           *counts_disps_array = NULL;
    int           *counts_ptr         = NULL;
    int           *displacements_ptr  = NULL;
    int            mpi_code;
    herr_t         ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    HDassert(chunk_list || 0 == chunk_list_num_entries);
    HDassert(io_info);
    HDassert(di);
    HDassert(idx_info);

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TRACE_ENTER(mpi_rank);
    H5D_MPIO_TIME_START(mpi_rank, "Reinsertion of modified chunks into chunk index");
#endif

    /* Only re-insert chunks if index has an insert method */
    if (!idx_info->storage->ops->insert)
        HGOTO_DONE(SUCCEED);

    /*
     * Make sure it's safe to cast this rank's number
     * of chunks to be sent into an int for MPI
     */
    H5_CHECK_OVERFLOW(chunk_list_num_entries, size_t, int);

    /* Create derived datatypes for the chunk re-insertion info needed */
    if (H5D__mpio_get_chunk_insert_info_types(&recv_type, &recv_type_derived, &send_type,
                                              &send_type_derived) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL,
                    "can't create derived datatypes for chunk re-insertion info")

    /*
     * Gather information to all ranks for a collective re-insertion
     * of the modified chunks into the chunk index
     */
    if (num_chunks_assigned_map) {
        /*
         * If a mapping between rank value -> number of assigned chunks has
         * been provided (usually during linked-chunk I/O), we can use this
         * to optimize MPI overhead a bit since MPI ranks won't need to
         * first inform each other about how many chunks they're contributing.
         */
        if (NULL == (counts_disps_array = H5MM_malloc(2 * (size_t)mpi_size * sizeof(*counts_disps_array)))) {
            /* Push an error, but still participate in collective gather operation */
            HDONE_ERROR(H5E_RESOURCE, H5E_CANTALLOC, FAIL,
                        "couldn't allocate receive counts and displacements array")
        }
        else {
            /* Set the receive counts from the assigned chunks map */
            counts_ptr = counts_disps_array;

            for (i = 0; i < (size_t)mpi_size; i++)
                H5_CHECKED_ASSIGN(counts_ptr[i], int, num_chunks_assigned_map[i], size_t);

            /* Set the displacements into the receive buffer for the gather operation */
            displacements_ptr = &counts_disps_array[mpi_size];

            *displacements_ptr = 0;
            for (i = 1; i < (size_t)mpi_size; i++)
                displacements_ptr[i] = displacements_ptr[i - 1] + counts_ptr[i - 1];
        }

        /* Perform gather operation */
        if (H5_mpio_gatherv_alloc(chunk_list, (int)chunk_list_num_entries, send_type, counts_ptr,
                                  displacements_ptr, recv_type, TRUE, 0, io_info->comm, mpi_rank, mpi_size,
                                  &gathered_array, &collective_num_entries) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTGATHER, FAIL,
                        "can't gather chunk index re-insertion info to/from ranks")
    }
    else {
        /*
         * If no mapping between rank value -> number of assigned chunks has
         * been provided (usually during multi-chunk I/O), all MPI ranks will
         * need to first inform other ranks about how many chunks they're
         * contributing before performing the actual gather operation. Use
         * the 'simple' MPI_Allgatherv wrapper for this.
         */
        if (H5_mpio_gatherv_alloc_simple(chunk_list, (int)chunk_list_num_entries, send_type, recv_type, TRUE,
                                         0, io_info->comm, mpi_rank, mpi_size, &gathered_array,
                                         &collective_num_entries) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTGATHER, FAIL,
                        "can't gather chunk index re-insertion info to/from ranks")
    }

    /* Initialize static chunk udata fields from chunk index info */
    H5D_MPIO_INIT_CHUNK_UD_INFO(chunk_ud, idx_info);

    for (i = 0; i < collective_num_entries; i++) {
        H5D_chunk_insert_info_t *coll_entry = &((H5D_chunk_insert_info_t *)gathered_array)[i];

        /*
         * We only need to reinsert this chunk if we had to actually
         * allocate or reallocate space in the file for it
         */
        if (!coll_entry->index_info.need_insert)
            continue;

        chunk_ud.chunk_block   = coll_entry->chunk_block;
        chunk_ud.chunk_idx     = coll_entry->index_info.chunk_idx;
        chunk_ud.filter_mask   = coll_entry->index_info.filter_mask;
        chunk_ud.common.scaled = scaled_coords;

        /* Calculate scaled coordinates for the chunk */
        if (idx_info->layout->idx_type == H5D_CHUNK_IDX_EARRAY && idx_info->layout->u.earray.unlim_dim > 0) {
            /*
             * Extensible arrays where the unlimited dimension is not
             * the slowest-changing dimension "swizzle" the coordinates
             * to move the unlimited dimension value to offset 0. Therefore,
             * we use the "swizzled" down chunks to calculate the "swizzled"
             * scaled coordinates and then we undo the "swizzle" operation.
             *
             * TODO: In the future, this is something that should be handled
             *       by the particular chunk index rather than manually
             *       here. Likely, the chunk index ops should get a new
             *       callback that accepts a chunk index and provides the
             *       caller with the scaled coordinates for that chunk.
             */
            H5VM_array_calc_pre(chunk_ud.chunk_idx, di->dset->shared->ndims,
                                idx_info->layout->u.earray.swizzled_down_chunks, scaled_coords);

            H5VM_unswizzle_coords(hsize_t, scaled_coords, idx_info->layout->u.earray.unlim_dim);
        }
        else {
            H5VM_array_calc_pre(chunk_ud.chunk_idx, di->dset->shared->ndims,
                                di->dset->shared->layout.u.chunk.down_chunks, scaled_coords);
        }

        scaled_coords[di->dset->shared->ndims] = 0;

#ifndef NDEBUG
        /*
         * If a matching local chunk entry is found, the
         * `chunk_info` structure (which contains the chunk's
         * pre-computed scaled coordinates) will be valid
         * for this rank. Compare those coordinates against
         * the calculated coordinates above to make sure
         * they match.
         */
        for (size_t dbg_idx = 0; dbg_idx < chunk_list_num_entries; dbg_idx++) {
            if (coll_entry->index_info.chunk_idx == chunk_list[dbg_idx].index_info.chunk_idx) {
                hbool_t coords_match = !HDmemcmp(scaled_coords, chunk_list[dbg_idx].chunk_info->scaled,
                                                 di->dset->shared->ndims * sizeof(hsize_t));

                HDassert(coords_match && "Calculated scaled coordinates for chunk didn't match "
                                         "chunk's actual scaled coordinates!");
                break;
            }
        }
#endif

        if ((idx_info->storage->ops->insert)(idx_info, &chunk_ud, di->dset) < 0)
            HGOTO_ERROR(H5E_DATASET, H5E_CANTINSERT, FAIL, "unable to insert chunk address into index")
    }

done:
    H5MM_free(gathered_array);
    H5MM_free(counts_disps_array);

    if (send_type_derived) {
        if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&send_type)))
            HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
    }
    if (recv_type_derived) {
        if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&recv_type)))
            HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
    }

#ifdef H5Dmpio_DEBUG
    H5D_MPIO_TIME_STOP(mpi_rank);
    H5D_MPIO_TRACE_EXIT(mpi_rank);
#endif

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__mpio_collective_filtered_chunk_reinsert() */

/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_get_chunk_redistribute_info_types
 *
 * Purpose:     Constructs MPI derived datatypes for communicating the
 *              info from a H5D_filtered_collective_io_info_t structure
 *              that is necessary for redistributing shared chunks during a
 *              collective write of filtered chunks.
 *
 *              The datatype returned through `contig_type` has an extent
 *              equal to the size of an H5D_chunk_redistribute_info_t
 *              structure and is suitable for communicating that structure
 *              type.
 *
 *              The datatype returned through `resized_type` has an extent
 *              equal to the size of an H5D_filtered_collective_io_info_t
 *              structure. This makes it suitable for sending an array of
 *              those structures, while extracting out just the info
 *              necessary for the chunk redistribution operation during
 *              communication.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_get_chunk_redistribute_info_types(MPI_Datatype *contig_type, hbool_t *contig_type_derived,
                                            MPI_Datatype *resized_type, hbool_t *resized_type_derived)
{
    MPI_Datatype struct_type              = MPI_DATATYPE_NULL;
    hbool_t      struct_type_derived      = FALSE;
    MPI_Datatype chunk_block_type         = MPI_DATATYPE_NULL;
    hbool_t      chunk_block_type_derived = FALSE;
    MPI_Datatype types[5];
    MPI_Aint     displacements[5];
    int          block_lengths[5];
    int          field_count;
    int          mpi_code;
    herr_t       ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    HDassert(contig_type);
    HDassert(contig_type_derived);
    HDassert(resized_type);
    HDassert(resized_type_derived);

    *contig_type_derived  = FALSE;
    *resized_type_derived = FALSE;

    /* Create struct type for the inner H5F_block_t structure */
    if (H5F_mpi_get_file_block_type(FALSE, &chunk_block_type, &chunk_block_type_derived) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't create derived type for chunk file description")

    field_count = 5;
    HDassert(field_count == (sizeof(types) / sizeof(MPI_Datatype)));

    /*
     * Create structure type to pack chunk H5F_block_t structure
     * next to chunk_idx, orig_owner, new_owner and num_writers
     * fields
     */
    block_lengths[0] = 1;
    block_lengths[1] = 1;
    block_lengths[2] = 1;
    block_lengths[3] = 1;
    block_lengths[4] = 1;
    displacements[0] = offsetof(H5D_chunk_redistribute_info_t, chunk_block);
    displacements[1] = offsetof(H5D_chunk_redistribute_info_t, chunk_idx);
    displacements[2] = offsetof(H5D_chunk_redistribute_info_t, orig_owner);
    displacements[3] = offsetof(H5D_chunk_redistribute_info_t, new_owner);
    displacements[4] = offsetof(H5D_chunk_redistribute_info_t, num_writers);
    types[0]         = chunk_block_type;
    types[1]         = HSIZE_AS_MPI_TYPE;
    types[2]         = MPI_INT;
    types[3]         = MPI_INT;
    types[4]         = MPI_INT;
    if (MPI_SUCCESS !=
        (mpi_code = MPI_Type_create_struct(field_count, block_lengths, displacements, types, contig_type)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_struct failed", mpi_code)
    *contig_type_derived = TRUE;

    if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(contig_type)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code)

    /* Create struct type to extract the chunk_current, chunk_idx, orig_owner,
     * new_owner and num_writers fields from a H5D_filtered_collective_io_info_t
     * structure
     */
    block_lengths[0] = 1;
    block_lengths[1] = 1;
    block_lengths[2] = 1;
    block_lengths[3] = 1;
    block_lengths[4] = 1;
    displacements[0] = offsetof(H5D_filtered_collective_io_info_t, chunk_current);
    displacements[1] = offsetof(H5D_filtered_collective_io_info_t, index_info.chunk_idx);
    displacements[2] = offsetof(H5D_filtered_collective_io_info_t, orig_owner);
    displacements[3] = offsetof(H5D_filtered_collective_io_info_t, new_owner);
    displacements[4] = offsetof(H5D_filtered_collective_io_info_t, num_writers);
    types[0]         = chunk_block_type;
    types[1]         = HSIZE_AS_MPI_TYPE;
    types[2]         = MPI_INT;
    types[3]         = MPI_INT;
    types[4]         = MPI_INT;
    if (MPI_SUCCESS !=
        (mpi_code = MPI_Type_create_struct(field_count, block_lengths, displacements, types, &struct_type)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_struct failed", mpi_code)
    struct_type_derived = TRUE;

    if (MPI_SUCCESS != (mpi_code = MPI_Type_create_resized(
                            struct_type, 0, sizeof(H5D_filtered_collective_io_info_t), resized_type)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_resized failed", mpi_code)
    *resized_type_derived = TRUE;

    if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(resized_type)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code)

done:
    if (struct_type_derived) {
        if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&struct_type)))
            HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
    }
    if (chunk_block_type_derived) {
        if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&chunk_block_type)))
            HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
    }

    if (ret_value < 0) {
        if (*resized_type_derived) {
            if (MPI_SUCCESS != (mpi_code = MPI_Type_free(resized_type)))
                HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
            *resized_type_derived = FALSE;
        }
        if (*contig_type_derived) {
            if (MPI_SUCCESS != (mpi_code = MPI_Type_free(contig_type)))
                HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
            *contig_type_derived = FALSE;
        }
    }

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__mpio_get_chunk_redistribute_info_types() */

/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_get_chunk_alloc_info_types
 *
 * Purpose:     Constructs MPI derived datatypes for communicating the info
 *              from a H5D_filtered_collective_io_info_t structure that is
 *              necessary for re-allocating file space during a collective
 *              write of filtered chunks.
 *
 *              The datatype returned through `contig_type` has an extent
 *              equal to the size of an H5D_chunk_alloc_info_t structure
 *              and is suitable for communicating that structure type.
 *
 *              The datatype returned through `resized_type` has an extent
 *              equal to the size of an H5D_filtered_collective_io_info_t
 *              structure. This makes it suitable for sending an array of
 *              those structures, while extracting out just the info
 *              necessary for the chunk file space reallocation operation
 *              during communication.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_get_chunk_alloc_info_types(MPI_Datatype *contig_type, hbool_t *contig_type_derived,
                                     MPI_Datatype *resized_type, hbool_t *resized_type_derived)
{
    MPI_Datatype struct_type              = MPI_DATATYPE_NULL;
    hbool_t      struct_type_derived      = FALSE;
    MPI_Datatype chunk_block_type         = MPI_DATATYPE_NULL;
    hbool_t      chunk_block_type_derived = FALSE;
    MPI_Datatype types[3];
    MPI_Aint     displacements[3];
    int          block_lengths[3];
    int          field_count;
    int          mpi_code;
    herr_t       ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    HDassert(contig_type);
    HDassert(contig_type_derived);
    HDassert(resized_type);
    HDassert(resized_type_derived);

    *contig_type_derived  = FALSE;
    *resized_type_derived = FALSE;

    /* Create struct type for the inner H5F_block_t structure */
    if (H5F_mpi_get_file_block_type(FALSE, &chunk_block_type, &chunk_block_type_derived) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't create derived type for chunk file description")

    field_count = 3;
    HDassert(field_count == (sizeof(types) / sizeof(MPI_Datatype)));

    /*
     * Create structure type to pack both chunk H5F_block_t structures
     * next to chunk_idx field
     */
    block_lengths[0] = 1;
    block_lengths[1] = 1;
    block_lengths[2] = 1;
    displacements[0] = offsetof(H5D_chunk_alloc_info_t, chunk_current);
    displacements[1] = offsetof(H5D_chunk_alloc_info_t, chunk_new);
    displacements[2] = offsetof(H5D_chunk_alloc_info_t, chunk_idx);
    types[0]         = chunk_block_type;
    types[1]         = chunk_block_type;
    types[2]         = HSIZE_AS_MPI_TYPE;
    if (MPI_SUCCESS !=
        (mpi_code = MPI_Type_create_struct(field_count, block_lengths, displacements, types, contig_type)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_struct failed", mpi_code)
    *contig_type_derived = TRUE;

    if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(contig_type)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code)

    /*
     * Create struct type to extract the chunk_current, chunk_new and chunk_idx
     * fields from a H5D_filtered_collective_io_info_t structure
     */
    block_lengths[0] = 1;
    block_lengths[1] = 1;
    block_lengths[2] = 1;
    displacements[0] = offsetof(H5D_filtered_collective_io_info_t, chunk_current);
    displacements[1] = offsetof(H5D_filtered_collective_io_info_t, chunk_new);
    displacements[2] = offsetof(H5D_filtered_collective_io_info_t, index_info.chunk_idx);
    types[0]         = chunk_block_type;
    types[1]         = chunk_block_type;
    types[2]         = HSIZE_AS_MPI_TYPE;
    if (MPI_SUCCESS !=
        (mpi_code = MPI_Type_create_struct(field_count, block_lengths, displacements, types, &struct_type)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_struct failed", mpi_code)
    struct_type_derived = TRUE;

    if (MPI_SUCCESS != (mpi_code = MPI_Type_create_resized(
                            struct_type, 0, sizeof(H5D_filtered_collective_io_info_t), resized_type)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_resized failed", mpi_code)
    *resized_type_derived = TRUE;

    if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(resized_type)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code)

done:
    if (struct_type_derived) {
        if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&struct_type)))
            HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
    }
    if (chunk_block_type_derived) {
        if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&chunk_block_type)))
            HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
    }

    if (ret_value < 0) {
        if (*resized_type_derived) {
            if (MPI_SUCCESS != (mpi_code = MPI_Type_free(resized_type)))
                HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
            *resized_type_derived = FALSE;
        }
        if (*contig_type_derived) {
            if (MPI_SUCCESS != (mpi_code = MPI_Type_free(contig_type)))
                HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
            *contig_type_derived = FALSE;
        }
    }

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__mpio_get_chunk_alloc_info_types() */

/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_get_chunk_insert_info_types
 *
 * Purpose:     Constructs MPI derived datatypes for communicating the
 *              information necessary when reinserting chunks into a
 *              dataset's chunk index. This includes the chunk's new offset
 *              and size (H5F_block_t) and the inner `index_info` structure
 *              of a H5D_filtered_collective_io_info_t structure.
 *
 *              The datatype returned through `contig_type` has an extent
 *              equal to the size of an H5D_chunk_insert_info_t structure
 *              and is suitable for communicating that structure type.
 *
 *              The datatype returned through `resized_type` has an extent
 *              equal to the size of the encompassing
 *              H5D_filtered_collective_io_info_t structure. This makes it
 *              suitable for sending an array of
 *              H5D_filtered_collective_io_info_t structures, while
 *              extracting out just the information needed during
 *              communication.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_get_chunk_insert_info_types(MPI_Datatype *contig_type, hbool_t *contig_type_derived,
                                      MPI_Datatype *resized_type, hbool_t *resized_type_derived)
{
    MPI_Datatype struct_type              = MPI_DATATYPE_NULL;
    hbool_t      struct_type_derived      = FALSE;
    MPI_Datatype chunk_block_type         = MPI_DATATYPE_NULL;
    hbool_t      chunk_block_type_derived = FALSE;
    MPI_Aint     contig_type_extent;
    MPI_Datatype types[4];
    MPI_Aint     displacements[4];
    int          block_lengths[4];
    int          field_count;
    int          mpi_code;
    herr_t       ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    HDassert(contig_type);
    HDassert(contig_type_derived);
    HDassert(resized_type);
    HDassert(resized_type_derived);

    *contig_type_derived  = FALSE;
    *resized_type_derived = FALSE;

    /* Create struct type for an H5F_block_t structure */
    if (H5F_mpi_get_file_block_type(FALSE, &chunk_block_type, &chunk_block_type_derived) < 0)
        HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't create derived type for chunk file description")

    field_count = 4;
    HDassert(field_count == (sizeof(types) / sizeof(MPI_Datatype)));

    /*
     * Create struct type to pack information into memory as follows:
     *
     * Chunk's new Offset/Size (H5F_block_t) ->
     * Chunk Index Info (H5D_chunk_index_info_t)
     */
    block_lengths[0] = 1;
    block_lengths[1] = 1;
    block_lengths[2] = 1;
    block_lengths[3] = 1;
    displacements[0] = offsetof(H5D_chunk_insert_info_t, chunk_block);
    displacements[1] = offsetof(H5D_chunk_insert_info_t, index_info.chunk_idx);
    displacements[2] = offsetof(H5D_chunk_insert_info_t, index_info.filter_mask);
    displacements[3] = offsetof(H5D_chunk_insert_info_t, index_info.need_insert);
    types[0]         = chunk_block_type;
    types[1]         = HSIZE_AS_MPI_TYPE;
    types[2]         = MPI_UNSIGNED;
    types[3]         = MPI_C_BOOL;
    if (MPI_SUCCESS !=
        (mpi_code = MPI_Type_create_struct(field_count, block_lengths, displacements, types, &struct_type)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_struct failed", mpi_code)
    struct_type_derived = TRUE;

    contig_type_extent = (MPI_Aint)(sizeof(H5F_block_t) + sizeof(H5D_chunk_index_info_t));

    if (MPI_SUCCESS != (mpi_code = MPI_Type_create_resized(struct_type, 0, contig_type_extent, contig_type)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_resized failed", mpi_code)
    *contig_type_derived = TRUE;

    if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(contig_type)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code)

    struct_type_derived = FALSE;
    if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&struct_type)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Type_free failed", mpi_code)

    /*
     * Create struct type to correctly extract all needed
     * information from a H5D_filtered_collective_io_info_t
     * structure.
     */
    displacements[0] = offsetof(H5D_filtered_collective_io_info_t, chunk_new);
    displacements[1] = offsetof(H5D_filtered_collective_io_info_t, index_info.chunk_idx);
    displacements[2] = offsetof(H5D_filtered_collective_io_info_t, index_info.filter_mask);
    displacements[3] = offsetof(H5D_filtered_collective_io_info_t, index_info.need_insert);
    if (MPI_SUCCESS !=
        (mpi_code = MPI_Type_create_struct(field_count, block_lengths, displacements, types, &struct_type)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_struct failed", mpi_code)
    struct_type_derived = TRUE;

    if (MPI_SUCCESS != (mpi_code = MPI_Type_create_resized(
                            struct_type, 0, sizeof(H5D_filtered_collective_io_info_t), resized_type)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Type_create_resized failed", mpi_code)
    *resized_type_derived = TRUE;

    if (MPI_SUCCESS != (mpi_code = MPI_Type_commit(resized_type)))
        HMPI_GOTO_ERROR(FAIL, "MPI_Type_commit failed", mpi_code)

done:
    if (struct_type_derived) {
        if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&struct_type)))
            HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
    }
    if (chunk_block_type_derived) {
        if (MPI_SUCCESS != (mpi_code = MPI_Type_free(&chunk_block_type)))
            HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
    }

    if (ret_value < 0) {
        if (*resized_type_derived) {
            if (MPI_SUCCESS != (mpi_code = MPI_Type_free(resized_type)))
                HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
            *resized_type_derived = FALSE;
        }
        if (*contig_type_derived) {
            if (MPI_SUCCESS != (mpi_code = MPI_Type_free(contig_type)))
                HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
            *contig_type_derived = FALSE;
        }
    }

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__mpio_get_chunk_insert_info_types() */

/*-------------------------------------------------------------------------
 * Function:    H5D__mpio_collective_filtered_io_type
 *
 * Purpose:     Constructs a MPI derived datatype for both the memory and
 *              the file for a collective I/O operation on filtered chunks.
 *              The datatype contains the chunk offsets and lengths in the
 *              file and the locations of the chunk data buffers to read
 *              into/write from.
 *
 * Return:      Non-negative on success/Negative on failure
 *
 *-------------------------------------------------------------------------
 */
static herr_t
H5D__mpio_collective_filtered_io_type(H5D_filtered_collective_io_info_t *chunk_list, size_t num_entries,
                                      H5D_io_op_type_t op_type, MPI_Datatype *new_mem_type,
                                      hbool_t *mem_type_derived, MPI_Datatype *new_file_type,
                                      hbool_t *file_type_derived)
{
    MPI_Aint *io_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 */
    int       mpi_code;
    herr_t    ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE

    HDassert(chunk_list || 0 == num_entries);
    HDassert(new_mem_type);
    HDassert(mem_type_derived);
    HDassert(new_file_type);
    HDassert(file_type_derived);

    *mem_type_derived  = FALSE;
    *file_type_derived = FALSE;
    *new_mem_type      = MPI_BYTE;
    *new_file_type     = MPI_BYTE;

    if (num_entries > 0) {
        H5F_block_t *chunk_block;
        size_t       last_valid_idx = 0;
        size_t       i;
        int          chunk_count;

        /*
         * Determine number of chunks for I/O operation and
         * setup for derived datatype creation if I/O operation
         * includes multiple chunks
         */
        if (num_entries == 1) {
            /* Set last valid index to 0 for contiguous datatype creation */
            last_valid_idx = 0;

            if (op_type == H5D_IO_OP_WRITE)
                chunk_count = 1;
            else
                chunk_count = chunk_list[0].need_read ? 1 : 0;
        }
        else {
            MPI_Aint chunk_buf;
            MPI_Aint base_buf;
            haddr_t  base_offset = HADDR_UNDEF;

            H5_CHECK_OVERFLOW(num_entries, size_t, int);

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

            /*
             * If doing a write, we can set the base chunk offset
             * and base chunk data buffer right away.
             *
             * If doing a read, some chunks may be skipped over
             * for reading if they aren't yet allocated in the
             * file. Therefore, we have to find the first chunk
             * actually being read in order to set the base chunk
             * offset and base chunk data buffer.
             */
            if (op_type == H5D_IO_OP_WRITE) {
#if H5_CHECK_MPI_VERSION(3, 0)
                if (MPI_SUCCESS != (mpi_code = MPI_Get_address(chunk_list[0].buf, &base_buf)))
                    HMPI_GOTO_ERROR(FAIL, "MPI_Get_address failed", mpi_code)
#else
                base_buf = (MPI_Aint)chunk_list[0].buf;
#endif

                base_offset = chunk_list[0].chunk_new.offset;
            }

            for (i = 0, chunk_count = 0; i < num_entries; i++) {
                if (op_type == H5D_IO_OP_READ) {
                    /*
                     * If this chunk isn't being read, don't add it
                     * to the MPI type we're building up for I/O
                     */
                    if (!chunk_list[i].need_read)
                        continue;

                    /*
                     * If this chunk is being read, go ahead and
                     * set the base chunk offset and base chunk
                     * data buffer if we haven't already
                     */
                    if (!H5F_addr_defined(base_offset)) {
#if H5_CHECK_MPI_VERSION(3, 0)
                        if (MPI_SUCCESS != (mpi_code = MPI_Get_address(chunk_list[i].buf, &base_buf)))
                            HMPI_GOTO_ERROR(FAIL, "MPI_Get_address failed", mpi_code)
#else
                        base_buf = (MPI_Aint)chunk_list[i].buf;
#endif

                        base_offset = chunk_list[i].chunk_current.offset;
                    }
                }

                /* Set convenience pointer for current chunk block */
                chunk_block =
                    (op_type == H5D_IO_OP_READ) ? &chunk_list[i].chunk_current : &chunk_list[i].chunk_new;

                /*
                 * Set the current chunk entry's offset in the file, relative to
                 * the first chunk entry
                 */
                HDassert(H5F_addr_defined(chunk_block->offset));
                file_offset_array[chunk_count] = (MPI_Aint)(chunk_block->offset - base_offset);

                /*
                 * Ensure the chunk list is sorted in ascending ordering of
                 * offset in the file
                 */
                if (chunk_count)
                    HDassert(file_offset_array[chunk_count] > file_offset_array[chunk_count - 1]);

                /* Set the current chunk entry's size for the I/O operation */
                H5_CHECK_OVERFLOW(chunk_block->length, hsize_t, int);
                length_array[chunk_count] = (int)chunk_block->length;

                /*
                 * Set the displacement of the chunk entry's chunk data buffer,
                 * relative to the first entry's data buffer
                 */
#if H5_CHECK_MPI_VERSION(3, 1)
                if (MPI_SUCCESS != (mpi_code = MPI_Get_address(chunk_list[i].buf, &chunk_buf)))
                    HMPI_GOTO_ERROR(FAIL, "MPI_Get_address failed", mpi_code)

                io_buf_array[chunk_count] = MPI_Aint_diff(chunk_buf, base_buf);
#else
                chunk_buf                 = (MPI_Aint)chunk_list[i].buf;
                io_buf_array[chunk_count] = chunk_buf - base_buf;
#endif

                /*
                 * Set last valid index in case only a single chunk will
                 * be involved in the I/O operation
                 */
                last_valid_idx = i;

                chunk_count++;
            } /* end for */
        }

        /*
         * Create derived datatypes for the chunk list if this
         * rank has any chunks to work on
         */
        if (chunk_count > 0) {
            if (chunk_count == 1) {
                int chunk_len;

                /* Single chunk - use a contiguous type for both memory and file */

                /* Ensure that we can cast chunk size to an int for MPI */
                chunk_block = (op_type == H5D_IO_OP_READ) ? &chunk_list[last_valid_idx].chunk_current
                                                          : &chunk_list[last_valid_idx].chunk_new;
                H5_CHECKED_ASSIGN(chunk_len, int, chunk_block->length, hsize_t);

                if (MPI_SUCCESS != (mpi_code = MPI_Type_contiguous(chunk_len, MPI_BYTE, new_file_type)))
                    HMPI_GOTO_ERROR(FAIL, "MPI_Type_contiguous failed", mpi_code)
                *new_mem_type = *new_file_type;

                /*
                 * Since we use the same datatype for both memory and file, only
                 * mark the file type as derived so the caller doesn't try to
                 * free the same type twice
                 */
                *mem_type_derived  = FALSE;
                *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)
            }
            else {
                HDassert(file_offset_array);
                HDassert(length_array);
                HDassert(io_buf_array);

                /* Multiple chunks - use an hindexed type for both memory and file */

                /* Create memory MPI type */
                if (MPI_SUCCESS != (mpi_code = MPI_Type_create_hindexed(
                                        chunk_count, length_array, io_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(chunk_count, 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 (file_offset_array)
        H5MM_free(file_offset_array);
    if (io_buf_array)
        H5MM_free(io_buf_array);
    if (length_array)
        H5MM_free(length_array);

    if (ret_value < 0) {
        if (*file_type_derived) {
            if (MPI_SUCCESS != (mpi_code = MPI_Type_free(new_file_type)))
                HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
            *file_type_derived = FALSE;
        }
        if (*mem_type_derived) {
            if (MPI_SUCCESS != (mpi_code = MPI_Type_free(new_mem_type)))
                HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code)
            *mem_type_derived = FALSE;
        }
    }

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__mpio_collective_filtered_io_type() */

#ifdef H5Dmpio_DEBUG

static herr_t
H5D__mpio_dump_collective_filtered_chunk_list(H5D_filtered_collective_io_info_t *chunk_list,
                                              size_t chunk_list_num_entries, int mpi_rank)
{
    H5D_filtered_collective_io_info_t *chunk_entry;
    size_t                             i;
    herr_t                             ret_value = SUCCEED;

    FUNC_ENTER_PACKAGE_NOERR

    H5D_MPIO_DEBUG(mpi_rank, "CHUNK LIST: [");
    for (i = 0; i < chunk_list_num_entries; i++) {
        unsigned chunk_rank;

        chunk_entry = &chunk_list[i];

        HDassert(chunk_entry->chunk_info);
        chunk_rank = (unsigned)H5S_GET_EXTENT_NDIMS(chunk_entry->chunk_info->fspace);

        H5D_MPIO_DEBUG(mpi_rank, " {");
        H5D_MPIO_DEBUG_VA(mpi_rank, "   - Entry %zu -", i);

        H5D_MPIO_DEBUG(mpi_rank, "   - Chunk Fspace Info -");
        H5D_MPIO_DEBUG_VA(mpi_rank,
                          "     Chunk Current Info: { Offset: %" PRIuHADDR ", Length: %" PRIuHADDR " }",
                          chunk_entry->chunk_current.offset, chunk_entry->chunk_current.length);
        H5D_MPIO_DEBUG_VA(mpi_rank, "     Chunk New Info: { Offset: %" PRIuHADDR ", Length: %" PRIuHADDR " }",
                          chunk_entry->chunk_new.offset, chunk_entry->chunk_new.length);

        H5D_MPIO_DEBUG(mpi_rank, "   - Chunk Insert Info -");
        H5D_MPIO_DEBUG_VA(mpi_rank,
                          "     Chunk Scaled Coords (4-d): { %" PRIuHSIZE ", %" PRIuHSIZE ", %" PRIuHSIZE
                          ", %" PRIuHSIZE " }",
                          chunk_rank < 1 ? 0 : chunk_entry->chunk_info->scaled[0],
                          chunk_rank < 2 ? 0 : chunk_entry->chunk_info->scaled[1],
                          chunk_rank < 3 ? 0 : chunk_entry->chunk_info->scaled[2],
                          chunk_rank < 4 ? 0 : chunk_entry->chunk_info->scaled[3]);
        H5D_MPIO_DEBUG_VA(mpi_rank, "     Chunk Index: %" PRIuHSIZE, chunk_entry->index_info.chunk_idx);
        H5D_MPIO_DEBUG_VA(mpi_rank, "     Filter Mask: %u", chunk_entry->index_info.filter_mask);
        H5D_MPIO_DEBUG_VA(mpi_rank, "     Need Insert: %s",
                          chunk_entry->index_info.need_insert ? "YES" : "NO");

        H5D_MPIO_DEBUG(mpi_rank, "   - Other Info -");
        H5D_MPIO_DEBUG_VA(mpi_rank, "     Chunk Info Ptr: %p", (void *)chunk_entry->chunk_info);
        H5D_MPIO_DEBUG_VA(mpi_rank, "     Need Read: %s", chunk_entry->need_read ? "YES" : "NO");
        H5D_MPIO_DEBUG_VA(mpi_rank, "     Chunk I/O Size: %zu", chunk_entry->io_size);
        H5D_MPIO_DEBUG_VA(mpi_rank, "     Chunk Buffer Size: %zu", chunk_entry->chunk_buf_size);
        H5D_MPIO_DEBUG_VA(mpi_rank, "     Original Owner: %d", chunk_entry->orig_owner);
        H5D_MPIO_DEBUG_VA(mpi_rank, "     New Owner: %d", chunk_entry->new_owner);
        H5D_MPIO_DEBUG_VA(mpi_rank, "     # of Writers: %d", chunk_entry->num_writers);
        H5D_MPIO_DEBUG_VA(mpi_rank, "     Chunk Data Buffer Ptr: %p", (void *)chunk_entry->buf);

        H5D_MPIO_DEBUG(mpi_rank, " }");
    }
    H5D_MPIO_DEBUG(mpi_rank, "]");

    FUNC_LEAVE_NOAPI(ret_value)
} /* end H5D__mpio_dump_collective_filtered_chunk_list() */

#endif

#endif /* H5_HAVE_PARALLEL */