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
5919
5920
5921
5922
5923
5924
5925
5926
5927
5928
5929
5930
5931
5932
5933
5934
5935
5936
5937
5938
5939
5940
5941
5942
5943
5944
5945
5946
5947
5948
5949
5950
5951
5952
5953
5954
5955
5956
5957
5958
5959
5960
5961
5962
5963
5964
5965
5966
5967
5968
5969
5970
5971
5972
5973
5974
5975
5976
5977
5978
5979
5980
5981
5982
5983
5984
5985
5986
5987
5988
5989
5990
5991
5992
5993
5994
5995
5996
5997
5998
5999
6000
6001
6002
6003
6004
6005
6006
6007
6008
6009
6010
6011
6012
6013
6014
6015
6016
6017
6018
6019
6020
6021
6022
6023
6024
6025
6026
6027
6028
6029
6030
6031
6032
6033
6034
6035
6036
6037
6038
6039
6040
6041
6042
6043
6044
6045
6046
6047
6048
6049
6050
6051
6052
6053
6054
6055
6056
6057
6058
6059
6060
6061
6062
6063
6064
6065
6066
6067
6068
6069
6070
6071
6072
6073
6074
6075
6076
6077
6078
6079
6080
6081
6082
6083
6084
6085
6086
6087
6088
6089
6090
6091
6092
6093
6094
6095
6096
6097
6098
6099
6100
6101
6102
6103
6104
6105
6106
6107
6108
6109
6110
6111
6112
6113
6114
6115
6116
6117
6118
6119
6120
6121
6122
6123
6124
6125
6126
6127
6128
6129
6130
6131
6132
6133
6134
6135
6136
6137
6138
6139
6140
6141
6142
6143
6144
6145
6146
6147
6148
6149
6150
6151
6152
6153
6154
6155
6156
6157
6158
6159
6160
6161
6162
6163
6164
6165
6166
6167
6168
6169
6170
6171
6172
6173
6174
6175
6176
6177
6178
6179
6180
6181
6182
6183
6184
6185
6186
6187
6188
6189
6190
6191
6192
6193
6194
6195
6196
6197
6198
6199
6200
6201
6202
6203
6204
6205
6206
6207
6208
6209
6210
6211
6212
6213
6214
6215
6216
6217
6218
6219
6220
6221
6222
6223
6224
6225
6226
6227
6228
6229
6230
6231
6232
6233
6234
6235
6236
6237
6238
6239
6240
6241
6242
6243
6244
6245
6246
6247
6248
6249
6250
6251
6252
6253
6254
6255
6256
6257
6258
6259
6260
6261
6262
6263
6264
6265
6266
6267
6268
6269
6270
6271
6272
6273
6274
6275
6276
6277
6278
6279
6280
6281
6282
6283
6284
6285
6286
6287
6288
6289
6290
6291
6292
6293
6294
6295
6296
6297
6298
6299
6300
6301
6302
6303
6304
6305
6306
6307
6308
6309
6310
6311
6312
6313
6314
6315
6316
6317
6318
6319
6320
6321
6322
6323
6324
6325
6326
6327
6328
6329
6330
6331
6332
6333
6334
6335
6336
6337
6338
6339
6340
6341
6342
6343
6344
6345
6346
6347
6348
6349
6350
6351
6352
6353
6354
6355
6356
6357
6358
6359
6360
6361
6362
6363
6364
6365
6366
6367
6368
6369
6370
6371
6372
6373
6374
6375
6376
6377
6378
6379
6380
6381
6382
6383
|
/*-
* Copyright (c) 2014-2016 MongoDB, Inc.
* Copyright (c) 2008-2014 WiredTiger, Inc.
* All rights reserved.
*
* See the file LICENSE for redistribution information.
*/
#include "wt_internal.h"
struct __rec_boundary; typedef struct __rec_boundary WT_BOUNDARY;
struct __rec_dictionary; typedef struct __rec_dictionary WT_DICTIONARY;
struct __rec_kv; typedef struct __rec_kv WT_KV;
/*
* Reconciliation is the process of taking an in-memory page, walking each entry
* in the page, building a backing disk image in a temporary buffer representing
* that information, and writing that buffer to disk. What could be simpler?
*
* WT_RECONCILE --
* Information tracking a single page reconciliation.
*/
typedef struct {
WT_REF *ref; /* Page being reconciled */
WT_PAGE *page;
uint32_t flags; /* Caller's configuration */
WT_ITEM disk_image; /* Temporary disk-image buffer */
/*
* Track start/stop write generation to decide if all changes to the
* page are written.
*/
uint32_t orig_write_gen;
/*
* Track start/stop checkpoint generations to decide if lookaside table
* records are correct.
*/
uint64_t orig_btree_checkpoint_gen;
uint64_t orig_txn_checkpoint_gen;
/* Track the page's maximum transaction ID. */
uint64_t max_txn;
/*
* When we can't mark the page clean (for example, checkpoint found some
* uncommitted updates), there's a leave-dirty flag.
*/
bool leave_dirty;
/*
* Raw compression (don't get me started, as if normal reconciliation
* wasn't bad enough). If an application wants absolute control over
* what gets written to disk, we give it a list of byte strings and it
* gives us back an image that becomes a file block. Because we don't
* know the number of items we're storing in a block until we've done
* a lot of work, we turn off most compression: dictionary, copy-cell,
* prefix and row-store internal page suffix compression are all off.
*/
bool raw_compression;
uint32_t raw_max_slots; /* Raw compression array sizes */
uint32_t *raw_entries; /* Raw compression slot entries */
uint32_t *raw_offsets; /* Raw compression slot offsets */
uint64_t *raw_recnos; /* Raw compression recno count */
WT_ITEM raw_destination; /* Raw compression destination buffer */
/*
* Track if reconciliation has seen any overflow items. If a leaf page
* with no overflow items is written, the parent page's address cell is
* set to the leaf-no-overflow type. This means we can delete the leaf
* page without reading it because we don't have to discard any overflow
* items it might reference.
*
* The test test is per-page reconciliation, that is, once we see an
* overflow item on the page, all subsequent leaf pages written for the
* page will not be leaf-no-overflow type, regardless of whether or not
* they contain overflow items. In other words, leaf-no-overflow is not
* guaranteed to be set on every page that doesn't contain an overflow
* item, only that if it is set, the page contains no overflow items.
*
* The reason is because of raw compression: there's no easy/fast way to
* figure out if the rows selected by raw compression included overflow
* items, and the optimization isn't worth another pass over the data.
*/
bool ovfl_items;
/*
* Track if reconciliation of a row-store leaf page has seen empty (zero
* length) values. We don't write out anything for empty values, so if
* there are empty values on a page, we have to make two passes over the
* page when it's read to figure out how many keys it has, expensive in
* the common case of no empty values and (entries / 2) keys. Likewise,
* a page with only empty values is another common data set, and keys on
* that page will be equal to the number of entries. In both cases, set
* a flag in the page's on-disk header.
*
* The test is per-page reconciliation as described above for the
* overflow-item test.
*/
bool all_empty_value, any_empty_value;
/*
* Reconciliation gets tricky if we have to split a page, which happens
* when the disk image we create exceeds the page type's maximum disk
* image size.
*
* First, the sizes of the page we're building. If WiredTiger is doing
* page layout, page_size is the same as page_size_orig. We accumulate
* a "page size" of raw data and when we reach that size, we split the
* page into multiple chunks, eventually compressing those chunks. When
* the application is doing page layout (raw compression is configured),
* page_size can continue to grow past page_size_orig, and we keep
* accumulating raw data until the raw compression callback accepts it.
*/
uint32_t page_size; /* Set page size */
uint32_t page_size_orig; /* Saved set page size */
uint32_t max_raw_page_size; /* Max page size with raw compression */
/*
* Second, the split size: if we're doing the page layout, split to a
* smaller-than-maximum page size when a split is required so we don't
* repeatedly split a packed page.
*/
uint32_t split_size; /* Split page size */
/*
* The problem with splits is we've done a lot of work by the time we
* realize we're going to have to split, we don't want to start over.
*
* To keep from having to start over when we hit the maximum page size,
* we track the page information when we approach a split boundary.
* If we eventually have to split, we walk this structure and pretend
* we were splitting all along. After that, we continue to append to
* this structure, and eventually walk it to create a new internal page
* that references all of our split pages.
*/
struct __rec_boundary {
/*
* Offset is the byte offset in the initial split buffer of the
* first byte of the split chunk, recorded before we decide to
* split the page; the difference between chunk[1]'s offset and
* chunk[0]'s offset is chunk[0]'s length.
*
* Once we split a page, we stop filling in offset values, we're
* writing the split chunks as we find them.
*/
size_t offset; /* Split's first byte */
/*
* The recno and entries fields are the starting record number
* of the split chunk (for column-store splits), and the number
* of entries in the split chunk. These fields are used both
* to write the split chunk, and to create a new internal page
* to reference the split pages.
*/
uint64_t recno; /* Split's starting record */
uint32_t entries; /* Split's entries */
WT_ADDR addr; /* Split's written location */
uint32_t size; /* Split's size */
uint32_t cksum; /* Split's checksum */
void *disk_image; /* Split's disk image */
/*
* Raw compression, the disk image being written is already
* compressed.
*/
bool already_compressed;
/*
* Saved update list, supporting the WT_EVICT_UPDATE_RESTORE and
* WT_EVICT_LOOKASIDE configurations.
*/
WT_SAVE_UPD *supd; /* Saved updates */
uint32_t supd_next;
size_t supd_allocated;
/*
* The key for a row-store page; no column-store key is needed
* because the page's recno, stored in the recno field, is the
* column-store key.
*/
WT_ITEM key; /* Promoted row-store key */
} *bnd; /* Saved boundaries */
uint32_t bnd_next; /* Next boundary slot */
uint32_t bnd_next_max; /* Maximum boundary slots used */
size_t bnd_entries; /* Total boundary slots */
size_t bnd_allocated; /* Bytes allocated */
/*
* We track the total number of page entries copied into split chunks
* so we can easily figure out how many entries in the current split
* chunk.
*/
uint32_t total_entries; /* Total entries in splits */
/*
* And there's state information as to where in this process we are:
* (1) tracking split boundaries because we can still fit more split
* chunks into the maximum page size, (2) tracking the maximum page
* size boundary because we can't fit any more split chunks into the
* maximum page size, (3) not performing boundary checks because it's
* either not useful with the current page size configuration, or
* because we've already been forced to split.
*/
enum { SPLIT_BOUNDARY=0, /* Next: a split page boundary */
SPLIT_MAX=1, /* Next: the maximum page boundary */
SPLIT_TRACKING_OFF=2, /* No boundary checks */
SPLIT_TRACKING_RAW=3 } /* Underlying compression decides */
bnd_state;
/*
* We track current information about the current record number, the
* number of entries copied into the temporary buffer, where we are
* in the temporary buffer, and how much memory remains. Those items
* are packaged here rather than passing pointers to stack locations
* around the code.
*/
uint64_t recno; /* Current record number */
uint32_t entries; /* Current number of entries */
uint8_t *first_free; /* Current first free byte */
size_t space_avail; /* Remaining space in this chunk */
/*
* Saved update list, supporting the WT_EVICT_UPDATE_RESTORE and
* WT_EVICT_LOOKASIDE configurations. While reviewing updates for each
* page, we save WT_UPDATE lists here, and then move them to per-block
* areas as the blocks are defined.
*/
WT_SAVE_UPD *supd; /* Saved updates */
uint32_t supd_next;
size_t supd_allocated;
/*
* We don't need to keep the 0th key around on internal pages, the
* search code ignores them as nothing can sort less by definition.
* There's some trickiness here, see the code for comments on how
* these fields work.
*/
bool cell_zero; /* Row-store internal page 0th key */
/*
* WT_DICTIONARY --
* We optionally build a dictionary of row-store values for leaf
* pages. Where two value cells are identical, only write the value
* once, the second and subsequent copies point to the original cell.
* The dictionary is fixed size, but organized in a skip-list to make
* searches faster.
*/
struct __rec_dictionary {
uint64_t hash; /* Hash value */
void *cell; /* Matching cell */
u_int depth; /* Skiplist */
WT_DICTIONARY *next[0];
} **dictionary; /* Dictionary */
u_int dictionary_next, dictionary_slots; /* Next, max entries */
/* Skiplist head. */
WT_DICTIONARY *dictionary_head[WT_SKIP_MAXDEPTH];
/*
* WT_KV--
* An on-page key/value item we're building.
*/
struct __rec_kv {
WT_ITEM buf; /* Data */
WT_CELL cell; /* Cell and cell's length */
size_t cell_len;
size_t len; /* Total length of cell + data */
} k, v; /* Key/Value being built */
WT_ITEM *cur, _cur; /* Key/Value being built */
WT_ITEM *last, _last; /* Last key/value built */
bool key_pfx_compress; /* If can prefix-compress next key */
bool key_pfx_compress_conf; /* If prefix compression configured */
bool key_sfx_compress; /* If can suffix-compress next key */
bool key_sfx_compress_conf; /* If suffix compression configured */
bool is_bulk_load; /* If it's a bulk load */
WT_SALVAGE_COOKIE *salvage; /* If it's a salvage operation */
bool cache_write_lookaside; /* Used the lookaside table */
bool cache_write_restore; /* Used update/restoration */
uint32_t tested_ref_state; /* Debugging information */
} WT_RECONCILE;
static void __rec_bnd_cleanup(WT_SESSION_IMPL *, WT_RECONCILE *, bool);
static void __rec_cell_build_addr(WT_SESSION_IMPL *,
WT_RECONCILE *, const void *, size_t, u_int, uint64_t);
static int __rec_cell_build_int_key(WT_SESSION_IMPL *,
WT_RECONCILE *, const void *, size_t, bool *);
static int __rec_cell_build_leaf_key(WT_SESSION_IMPL *,
WT_RECONCILE *, const void *, size_t, bool *);
static int __rec_cell_build_ovfl(WT_SESSION_IMPL *,
WT_RECONCILE *, WT_KV *, uint8_t, uint64_t);
static int __rec_cell_build_val(WT_SESSION_IMPL *,
WT_RECONCILE *, const void *, size_t, uint64_t);
static int __rec_col_fix(WT_SESSION_IMPL *, WT_RECONCILE *, WT_REF *);
static int __rec_col_fix_slvg(WT_SESSION_IMPL *,
WT_RECONCILE *, WT_REF *, WT_SALVAGE_COOKIE *);
static int __rec_col_int(WT_SESSION_IMPL *, WT_RECONCILE *, WT_REF *);
static int __rec_col_merge(WT_SESSION_IMPL *, WT_RECONCILE *, WT_PAGE *);
static int __rec_col_var(WT_SESSION_IMPL *,
WT_RECONCILE *, WT_REF *, WT_SALVAGE_COOKIE *);
static int __rec_col_var_helper(WT_SESSION_IMPL *, WT_RECONCILE *,
WT_SALVAGE_COOKIE *, WT_ITEM *, bool, uint8_t, uint64_t);
static int __rec_destroy_session(WT_SESSION_IMPL *);
static int __rec_root_write(WT_SESSION_IMPL *, WT_PAGE *, uint32_t);
static int __rec_row_int(WT_SESSION_IMPL *, WT_RECONCILE *, WT_PAGE *);
static int __rec_row_leaf(WT_SESSION_IMPL *,
WT_RECONCILE *, WT_PAGE *, WT_SALVAGE_COOKIE *);
static int __rec_row_leaf_insert(
WT_SESSION_IMPL *, WT_RECONCILE *, WT_INSERT *);
static int __rec_row_merge(WT_SESSION_IMPL *, WT_RECONCILE *, WT_PAGE *);
static int __rec_split_col(WT_SESSION_IMPL *, WT_RECONCILE *, WT_PAGE *);
static int __rec_split_discard(WT_SESSION_IMPL *, WT_PAGE *);
static int __rec_split_fixup(WT_SESSION_IMPL *, WT_RECONCILE *);
static int __rec_split_row(WT_SESSION_IMPL *, WT_RECONCILE *, WT_PAGE *);
static int __rec_split_row_promote(
WT_SESSION_IMPL *, WT_RECONCILE *, WT_ITEM *, uint8_t);
static int __rec_split_write(WT_SESSION_IMPL *,
WT_RECONCILE *, WT_BOUNDARY *, WT_ITEM *, bool);
static int __rec_update_las(
WT_SESSION_IMPL *, WT_RECONCILE *, uint32_t, WT_BOUNDARY *);
static int __rec_write_init(WT_SESSION_IMPL *,
WT_REF *, uint32_t, WT_SALVAGE_COOKIE *, void *);
static int __rec_write_status(WT_SESSION_IMPL *, WT_RECONCILE *, WT_PAGE *);
static int __rec_write_wrapup(WT_SESSION_IMPL *, WT_RECONCILE *, WT_PAGE *);
static int __rec_write_wrapup_err(
WT_SESSION_IMPL *, WT_RECONCILE *, WT_PAGE *);
static void __rec_dictionary_free(WT_SESSION_IMPL *, WT_RECONCILE *);
static int __rec_dictionary_init(WT_SESSION_IMPL *, WT_RECONCILE *, u_int);
static int __rec_dictionary_lookup(
WT_SESSION_IMPL *, WT_RECONCILE *, WT_KV *, WT_DICTIONARY **);
static void __rec_dictionary_reset(WT_RECONCILE *);
/*
* __wt_reconcile --
* Reconcile an in-memory page into its on-disk format, and write it.
*/
int
__wt_reconcile(WT_SESSION_IMPL *session,
WT_REF *ref, WT_SALVAGE_COOKIE *salvage, uint32_t flags)
{
WT_DECL_RET;
WT_PAGE *page;
WT_PAGE_MODIFY *mod;
WT_RECONCILE *r;
uint64_t oldest_id;
page = ref->page;
mod = page->modify;
WT_RET(__wt_verbose(session,
WT_VERB_RECONCILE, "%s", __wt_page_type_string(page->type)));
/* We shouldn't get called with a clean page, that's an error. */
WT_ASSERT(session, __wt_page_is_modified(page));
/*
* Reconciliation locks the page for three reasons:
* Reconciliation reads the lists of page updates, obsolete updates
* cannot be discarded while reconciliation is in progress;
* The compaction process reads page modification information, which
* reconciliation modifies;
* In-memory splits: reconciliation of an internal page cannot handle
* a child page splitting during the reconciliation.
*/
WT_RET(__wt_fair_lock(session, &page->page_lock));
/*
* Check that transaction time always moves forward for a given page.
* If this check fails, reconciliation can free something that a future
* reconciliation will need.
*/
oldest_id = __wt_txn_oldest_id(session);
WT_ASSERT(session, WT_TXNID_LE(mod->last_oldest_id, oldest_id));
mod->last_oldest_id = oldest_id;
/* Initialize the reconciliation structure for each new run. */
if ((ret = __rec_write_init(
session, ref, flags, salvage, &session->reconcile)) != 0) {
WT_TRET(__wt_fair_unlock(session, &page->page_lock));
return (ret);
}
r = session->reconcile;
/* Reconcile the page. */
switch (page->type) {
case WT_PAGE_COL_FIX:
if (salvage != NULL)
ret = __rec_col_fix_slvg(session, r, ref, salvage);
else
ret = __rec_col_fix(session, r, ref);
break;
case WT_PAGE_COL_INT:
WT_WITH_PAGE_INDEX(session,
ret = __rec_col_int(session, r, ref));
break;
case WT_PAGE_COL_VAR:
ret = __rec_col_var(session, r, ref, salvage);
break;
case WT_PAGE_ROW_INT:
WT_WITH_PAGE_INDEX(session,
ret = __rec_row_int(session, r, page));
break;
case WT_PAGE_ROW_LEAF:
ret = __rec_row_leaf(session, r, page, salvage);
break;
WT_ILLEGAL_VALUE_SET(session);
}
/* Get the final status for the reconciliation. */
if (ret == 0)
ret = __rec_write_status(session, r, page);
/* Wrap up the page reconciliation. */
if (ret == 0)
ret = __rec_write_wrapup(session, r, page);
else
WT_TRET(__rec_write_wrapup_err(session, r, page));
/* Release the reconciliation lock. */
WT_TRET(__wt_fair_unlock(session, &page->page_lock));
/* Update statistics. */
WT_STAT_FAST_CONN_INCR(session, rec_pages);
WT_STAT_FAST_DATA_INCR(session, rec_pages);
if (LF_ISSET(WT_EVICTING)) {
WT_STAT_FAST_CONN_INCR(session, rec_pages_eviction);
WT_STAT_FAST_DATA_INCR(session, rec_pages_eviction);
}
if (r->cache_write_lookaside) {
WT_STAT_FAST_CONN_INCR(session, cache_write_lookaside);
WT_STAT_FAST_DATA_INCR(session, cache_write_lookaside);
}
if (r->cache_write_restore) {
WT_STAT_FAST_CONN_INCR(session, cache_write_restore);
WT_STAT_FAST_DATA_INCR(session, cache_write_restore);
}
/*
* When application threads perform eviction, don't cache block manager
* or reconciliation structures (even across calls), we can have a
* significant number of application threads doing eviction at the same
* time with large items. We ignore checkpoints, once the checkpoint
* completes, all unnecessary session resources will be discarded.
*
* Even in application threads doing checkpoints or in internal threads
* doing any reconciliation, clean up reconciliation resources. Some
* workloads have millions of boundary structures in a reconciliation
* and we don't want to tie that memory down, even across calls.
*/
if (WT_SESSION_IS_CHECKPOINT(session) ||
F_ISSET(session, WT_SESSION_INTERNAL))
__rec_bnd_cleanup(session, r, false);
else {
/*
* Clean up the underlying block manager memory too: it's not
* reconciliation, but threads discarding reconciliation
* structures want to clean up the block manager's structures
* as well, and there's no obvious place to do that.
*/
if (session->block_manager_cleanup != NULL)
WT_TRET(session->block_manager_cleanup(session));
WT_TRET(__rec_destroy_session(session));
}
WT_RET(ret);
/*
* Root pages are special, splits have to be done, we can't put it off
* as the parent's problem any more.
*/
if (__wt_ref_is_root(ref)) {
WT_WITH_PAGE_INDEX(session,
ret = __rec_root_write(session, page, flags));
return (ret);
}
/*
* Otherwise, mark the page's parent dirty.
* Don't mark the tree dirty: if this reconciliation is in service of a
* checkpoint, it's cleared the tree's dirty flag, and we don't want to
* set it again as part of that walk.
*/
return (__wt_page_parent_modify_set(session, ref, true));
}
/*
* __rec_las_checkpoint_test --
* Return if the lookaside table is going to collide with a checkpoint.
*/
static inline bool
__rec_las_checkpoint_test(WT_SESSION_IMPL *session, WT_RECONCILE *r)
{
WT_CONNECTION_IMPL *conn;
WT_BTREE *btree;
conn = S2C(session);
btree = S2BT(session);
/*
* Running checkpoints can collide with the lookaside table because
* reconciliation using the lookaside table writes the key's last
* committed value, which might not be the value checkpoint would write.
* If reconciliation was configured for lookaside table eviction, this
* file participates in checkpoints, and any of the tree or system
* transactional generation numbers don't match, there's a possible
* collision.
*
* It's a complicated test, but the alternative is to have checkpoint
* drain lookaside table reconciliations, and this isn't a problem for
* most workloads.
*/
if (!F_ISSET(r, WT_EVICT_LOOKASIDE))
return (false);
if (F_ISSET(btree, WT_BTREE_NO_CHECKPOINT))
return (false);
if (r->orig_btree_checkpoint_gen == btree->checkpoint_gen &&
r->orig_txn_checkpoint_gen == conn->txn_global.checkpoint_gen &&
r->orig_btree_checkpoint_gen == r->orig_txn_checkpoint_gen)
return (false);
return (true);
}
/*
* __rec_write_status --
* Return the final status for reconciliation.
*/
static int
__rec_write_status(WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_PAGE *page)
{
WT_BTREE *btree;
WT_PAGE_MODIFY *mod;
btree = S2BT(session);
mod = page->modify;
/* Check for a lookaside table and checkpoint collision. */
if (__rec_las_checkpoint_test(session, r))
return (EBUSY);
/*
* Set the page's status based on whether or not we cleaned the page.
*/
if (r->leave_dirty) {
/*
* The page remains dirty.
*
* Any checkpoint call cleared the tree's modified flag before
* writing pages, so we must explicitly reset it. We insert a
* barrier after the change for clarity (the requirement is the
* flag be set before a subsequent checkpoint reads it, and
* as the current checkpoint is waiting on this reconciliation
* to complete, there's no risk of that happening)
*/
btree->modified = 1;
WT_FULL_BARRIER();
/*
* Eviction should only be here if following the save/restore
* eviction path.
*/
WT_ASSERT(session,
!F_ISSET(r, WT_EVICTING) ||
F_ISSET(r, WT_EVICT_UPDATE_RESTORE));
} else {
/*
* Track the page's maximum transaction ID (used to decide if
* we're likely to be able to evict this page in the future).
*/
mod->rec_max_txn = r->max_txn;
/*
* Track the tree's maximum transaction ID (used to decide if
* it's safe to discard the tree). Reconciliation for eviction
* is multi-threaded, only update the tree's maximum transaction
* ID when doing a checkpoint. That's sufficient, we only care
* about the maximum transaction ID of current updates in the
* tree, and checkpoint visits every dirty page in the tree.
*/
if (!F_ISSET(r, WT_EVICTING) &&
WT_TXNID_LT(btree->rec_max_txn, r->max_txn))
btree->rec_max_txn = r->max_txn;
/*
* The page only might be clean; if the write generation is
* unchanged since reconciliation started, it's clean.
*
* If the write generation changed, the page has been written
* since reconciliation started and remains dirty (that can't
* happen when evicting, the page is exclusively locked).
*/
if (__wt_atomic_cas32(&mod->write_gen, r->orig_write_gen, 0))
__wt_cache_dirty_decr(session, page);
else
WT_ASSERT(session, !F_ISSET(r, WT_EVICTING));
}
return (0);
}
/*
* __rec_root_write --
* Handle the write of a root page.
*/
static int
__rec_root_write(WT_SESSION_IMPL *session, WT_PAGE *page, uint32_t flags)
{
WT_DECL_RET;
WT_PAGE *next;
WT_PAGE_INDEX *pindex;
WT_PAGE_MODIFY *mod;
WT_REF fake_ref;
uint32_t i;
mod = page->modify;
/*
* If a single root page was written (either an empty page or there was
* a 1-for-1 page swap), we've written root and checkpoint, we're done.
* If the root page split, write the resulting WT_REF array. We already
* have an infrastructure for writing pages, create a fake root page and
* write it instead of adding code to write blocks based on the list of
* blocks resulting from a multiblock reconciliation.
*/
switch (mod->rec_result) {
case WT_PM_REC_EMPTY: /* Page is empty */
case WT_PM_REC_REPLACE: /* 1-for-1 page swap */
return (0);
case WT_PM_REC_MULTIBLOCK: /* Multiple blocks */
break;
WT_ILLEGAL_VALUE(session);
}
WT_RET(__wt_verbose(session, WT_VERB_SPLIT,
"root page split -> %" PRIu32 " pages", mod->mod_multi_entries));
/*
* Create a new root page, initialize the array of child references,
* mark it dirty, then write it.
*/
switch (page->type) {
case WT_PAGE_COL_INT:
WT_RET(__wt_page_alloc(session,
WT_PAGE_COL_INT, mod->mod_multi_entries, false, &next));
break;
case WT_PAGE_ROW_INT:
WT_RET(__wt_page_alloc(session,
WT_PAGE_ROW_INT, mod->mod_multi_entries, false, &next));
break;
WT_ILLEGAL_VALUE(session);
}
WT_INTL_INDEX_GET(session, next, pindex);
for (i = 0; i < mod->mod_multi_entries; ++i) {
/*
* There's special error handling required when re-instantiating
* pages in memory; it's not needed here, asserted for safety.
*/
WT_ASSERT(session, mod->mod_multi[i].supd == NULL);
WT_ERR(__wt_multi_to_ref(session,
next, &mod->mod_multi[i], &pindex->index[i], NULL, false));
pindex->index[i]->home = next;
}
/*
* We maintain a list of pages written for the root in order to free the
* backing blocks the next time the root is written.
*/
mod->mod_root_split = next;
/*
* Mark the page dirty.
* Don't mark the tree dirty: if this reconciliation is in service of a
* checkpoint, it's cleared the tree's dirty flag, and we don't want to
* set it again as part of that walk.
*/
WT_ERR(__wt_page_modify_init(session, next));
__wt_page_only_modify_set(session, next);
/*
* Fake up a reference structure, and write the next root page.
*/
__wt_root_ref_init(&fake_ref, next, page->type == WT_PAGE_COL_INT);
return (__wt_reconcile(session, &fake_ref, NULL, flags));
err: __wt_page_out(session, &next);
return (ret);
}
/*
* __rec_raw_compression_config --
* Configure raw compression.
*/
static inline bool
__rec_raw_compression_config(
WT_SESSION_IMPL *session, WT_PAGE *page, WT_SALVAGE_COOKIE *salvage)
{
WT_BTREE *btree;
btree = S2BT(session);
/* Check if raw compression configured. */
if (btree->compressor == NULL ||
btree->compressor->compress_raw == NULL)
return (false);
/* Only for row-store and variable-length column-store objects. */
if (page->type == WT_PAGE_COL_FIX)
return (false);
/*
* Raw compression cannot support dictionary compression. (Technically,
* we could still use the raw callback on column-store variable length
* internal pages with dictionary compression configured, because
* dictionary compression only applies to column-store leaf pages, but
* that seems an unlikely use case.)
*/
if (btree->dictionary != 0)
return (false);
/* Raw compression cannot support prefix compression. */
if (btree->prefix_compression)
return (false);
/*
* Raw compression is also turned off during salvage: we can't allow
* pages to split during salvage, raw compression has no point if it
* can't manipulate the page size.
*/
if (salvage != NULL)
return (false);
return (true);
}
/*
* __rec_write_init --
* Initialize the reconciliation structure.
*/
static int
__rec_write_init(WT_SESSION_IMPL *session,
WT_REF *ref, uint32_t flags, WT_SALVAGE_COOKIE *salvage, void *reconcilep)
{
WT_BTREE *btree;
WT_CONNECTION_IMPL *conn;
WT_PAGE *page;
WT_RECONCILE *r;
btree = S2BT(session);
conn = S2C(session);
page = ref->page;
if ((r = *(WT_RECONCILE **)reconcilep) == NULL) {
WT_RET(__wt_calloc_one(session, &r));
*(WT_RECONCILE **)reconcilep = r;
session->reconcile_cleanup = __rec_destroy_session;
/* Connect pointers/buffers. */
r->cur = &r->_cur;
r->last = &r->_last;
/* Disk buffers need to be aligned for writing. */
F_SET(&r->disk_image, WT_ITEM_ALIGNED);
}
/* Reconciliation is not re-entrant, make sure that doesn't happen. */
WT_ASSERT(session, r->ref == NULL);
/* Remember the configuration. */
r->ref = ref;
r->page = page;
/*
* Save the page's write generation before reading the page.
* Save the transaction generations before reading the page.
* These are all ordered reads, but we only need one.
*/
r->orig_btree_checkpoint_gen = btree->checkpoint_gen;
r->orig_txn_checkpoint_gen = conn->txn_global.checkpoint_gen;
WT_ORDERED_READ(r->orig_write_gen, page->modify->write_gen);
/*
* Lookaside table eviction is configured when eviction gets aggressive,
* adjust the flags for cases we don't support.
*/
if (LF_ISSET(WT_EVICT_LOOKASIDE)) {
/*
* Saving lookaside table updates into the lookaside table won't
* work.
*/
if (F_ISSET(btree, WT_BTREE_LOOKASIDE))
LF_CLR(WT_EVICT_LOOKASIDE);
/*
* We don't yet support fixed-length column-store combined with
* the lookaside table. It's not hard to do, but the underlying
* function that reviews which updates can be written to the
* evicted page and which updates need to be written to the
* lookaside table needs access to the original value from the
* page being evicted, and there's no code path for that in the
* case of fixed-length column-store objects. (Row-store and
* variable-width column-store objects provide a reference to
* the unpacked on-page cell for this purpose, but there isn't
* an on-page cell for fixed-length column-store objects.) For
* now, turn it off.
*/
if (page->type == WT_PAGE_COL_FIX)
LF_CLR(WT_EVICT_LOOKASIDE);
/*
* Check for a lookaside table and checkpoint collision, and if
* we find one, turn off the lookaside file (we've gone to all
* the effort of getting exclusive access to the page, might as
* well try and evict it).
*/
if (__rec_las_checkpoint_test(session, r))
LF_CLR(WT_EVICT_LOOKASIDE);
}
r->flags = flags;
/* Track the page's maximum transaction ID. */
r->max_txn = WT_TXN_NONE;
/* Track if the page can be marked clean. */
r->leave_dirty = false;
/* Raw compression. */
r->raw_compression =
__rec_raw_compression_config(session, page, salvage);
r->raw_destination.flags = WT_ITEM_ALIGNED;
/* Track overflow items. */
r->ovfl_items = false;
/* Track empty values. */
r->all_empty_value = true;
r->any_empty_value = false;
/* The list of saved updates. */
r->supd_next = 0;
/*
* Dictionary compression only writes repeated values once. We grow
* the dictionary as necessary, always using the largest size we've
* seen.
*
* Reset the dictionary.
*
* Sanity check the size: 100 slots is the smallest dictionary we use.
*/
if (btree->dictionary != 0 && btree->dictionary > r->dictionary_slots)
WT_RET(__rec_dictionary_init(session,
r, btree->dictionary < 100 ? 100 : btree->dictionary));
__rec_dictionary_reset(r);
/*
* Prefix compression discards repeated prefix bytes from row-store leaf
* page keys.
*/
r->key_pfx_compress_conf = false;
if (btree->prefix_compression && page->type == WT_PAGE_ROW_LEAF)
r->key_pfx_compress_conf = true;
/*
* Suffix compression shortens internal page keys by discarding trailing
* bytes that aren't necessary for tree navigation. We don't do suffix
* compression if there is a custom collator because we don't know what
* bytes a custom collator might use. Some custom collators (for
* example, a collator implementing reverse ordering of strings), won't
* have any problem with suffix compression: if there's ever a reason to
* implement suffix compression for custom collators, we can add a
* setting to the collator, configured when the collator is added, that
* turns on suffix compression.
*
* The raw compression routines don't even consider suffix compression,
* but it doesn't hurt to confirm that.
*/
r->key_sfx_compress_conf = false;
if (btree->collator == NULL &&
btree->internal_key_truncate && !r->raw_compression)
r->key_sfx_compress_conf = true;
r->is_bulk_load = false;
r->salvage = salvage;
r->cache_write_lookaside = r->cache_write_restore = false;
return (0);
}
/*
* __rec_destroy --
* Clean up the reconciliation structure.
*/
static void
__rec_destroy(WT_SESSION_IMPL *session, void *reconcilep)
{
WT_RECONCILE *r;
if ((r = *(WT_RECONCILE **)reconcilep) == NULL)
return;
*(WT_RECONCILE **)reconcilep = NULL;
__wt_buf_free(session, &r->disk_image);
__wt_free(session, r->raw_entries);
__wt_free(session, r->raw_offsets);
__wt_free(session, r->raw_recnos);
__wt_buf_free(session, &r->raw_destination);
__rec_bnd_cleanup(session, r, true);
__wt_free(session, r->supd);
__wt_buf_free(session, &r->k.buf);
__wt_buf_free(session, &r->v.buf);
__wt_buf_free(session, &r->_cur);
__wt_buf_free(session, &r->_last);
__rec_dictionary_free(session, r);
__wt_free(session, r);
}
/*
* __rec_destroy_session --
* Clean up the reconciliation structure, session version.
*/
static int
__rec_destroy_session(WT_SESSION_IMPL *session)
{
__rec_destroy(session, &session->reconcile);
return (0);
}
/*
* __rec_bnd_cleanup --
* Cleanup the boundary structure information.
*/
static void
__rec_bnd_cleanup(WT_SESSION_IMPL *session, WT_RECONCILE *r, bool destroy)
{
WT_BOUNDARY *bnd;
uint32_t i, last_used;
if (r->bnd == NULL)
return;
/* Reconciliation is not re-entrant, make sure that doesn't happen. */
r->ref = NULL;
/*
* Free the boundary structures' memory. In the case of normal cleanup,
* discard any memory we won't reuse in the next reconciliation; in the
* case of destruction, discard everything.
*
* During some big-page evictions we have seen boundary arrays that have
* millions of elements. That should not be a normal event, but if the
* memory is associated with a random application session, it won't be
* discarded until the session is closed or reset. If there are more
* than 10,000 boundary structure elements, discard the boundary array
* entirely and start over next time.
*/
if (destroy || r->bnd_entries > 10 * WT_THOUSAND) {
for (bnd = r->bnd, i = 0; i < r->bnd_entries; ++bnd, ++i) {
__wt_free(session, bnd->addr.addr);
__wt_free(session, bnd->disk_image);
__wt_free(session, bnd->supd);
__wt_buf_free(session, &bnd->key);
}
__wt_free(session, r->bnd);
r->bnd_next = 0;
r->bnd_entries = r->bnd_allocated = 0;
} else {
/*
* The boundary-next field points to the next boundary structure
* we were going to use, but there's no requirement that value
* be incremented before reconciliation updates the structure it
* points to, that is, there's no guarantee elements of the next
* boundary structure are still unchanged. Be defensive, clean
* up the "next" structure as well as the ones we know we used.
*/
last_used = r->bnd_next;
if (last_used < r->bnd_entries)
++last_used;
for (bnd = r->bnd, i = 0; i < last_used; ++bnd, ++i) {
__wt_free(session, bnd->addr.addr);
__wt_free(session, bnd->disk_image);
__wt_free(session, bnd->supd);
}
}
}
/*
* __rec_update_save --
* Save a WT_UPDATE list for later restoration.
*/
static int
__rec_update_save(WT_SESSION_IMPL *session,
WT_RECONCILE *r, WT_INSERT *ins, WT_ROW *rip, uint64_t txnid)
{
WT_RET(__wt_realloc_def(
session, &r->supd_allocated, r->supd_next + 1, &r->supd));
r->supd[r->supd_next].ins = ins;
r->supd[r->supd_next].rip = rip;
r->supd[r->supd_next].onpage_txn = txnid;
++r->supd_next;
return (0);
}
/*
* __rec_update_move --
* Move a saved WT_UPDATE list from the per-page cache to a specific
* block's list.
*/
static int
__rec_update_move(WT_SESSION_IMPL *session, WT_BOUNDARY *bnd, WT_SAVE_UPD *supd)
{
WT_RET(__wt_realloc_def(
session, &bnd->supd_allocated, bnd->supd_next + 1, &bnd->supd));
bnd->supd[bnd->supd_next] = *supd;
++bnd->supd_next;
supd->ins = NULL;
supd->rip = NULL;
return (0);
}
/*
* __rec_txn_read --
* Return the update in a list that should be written (or NULL if none can
* be written).
*/
static int
__rec_txn_read(WT_SESSION_IMPL *session, WT_RECONCILE *r,
WT_INSERT *ins, WT_ROW *rip, WT_CELL_UNPACK *vpack, WT_UPDATE **updp)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_DECL_ITEM(tmp);
WT_PAGE *page;
WT_UPDATE *append, *upd, *upd_list;
size_t notused;
uint64_t max_txn, min_txn, txnid;
bool append_origv, skipped;
*updp = NULL;
append = NULL; /* -Wconditional-uninitialized */
btree = S2BT(session);
page = r->page;
/*
* If called with a WT_INSERT item, use its WT_UPDATE list (which must
* exist), otherwise check for an on-page row-store WT_UPDATE list
* (which may not exist). Return immediately if the item has no updates.
*/
if (ins == NULL) {
if ((upd_list = WT_ROW_UPDATE(page, rip)) == NULL)
return (0);
} else
upd_list = ins->upd;
for (skipped = false,
max_txn = WT_TXN_NONE, min_txn = UINT64_MAX,
upd = upd_list; upd != NULL; upd = upd->next) {
if ((txnid = upd->txnid) == WT_TXN_ABORTED)
continue;
/* Track the largest/smallest transaction IDs on the list. */
if (WT_TXNID_LT(max_txn, txnid))
max_txn = txnid;
if (WT_TXNID_LT(txnid, min_txn))
min_txn = txnid;
/*
* Find the first update we can use.
*/
if (F_ISSET(r, WT_EVICTING)) {
/*
* Eviction can write any committed update.
*
* When reconciling for eviction, track whether any
* uncommitted updates are found.
*/
if (__wt_txn_committed(session, txnid)) {
if (*updp == NULL)
*updp = upd;
} else
skipped = true;
} else {
/*
* Checkpoint can only write updates visible as of its
* snapshot.
*
* When reconciling for a checkpoint, track whether any
* updates were skipped on the way to finding the first
* visible update.
*/
if (*updp == NULL) {
if (__wt_txn_visible(session, txnid))
*updp = upd;
else
skipped = true;
}
}
}
/*
* If all of the updates were aborted, quit. This test is not strictly
* necessary because the above loop exits with skipped not set and the
* maximum transaction left at its initial value of WT_TXN_NONE, so
* the test below will be branch true and return, but it's cheap and a
* little more explicit, and makes Coverity happy.
*/
if (max_txn == WT_TXN_NONE)
return (0);
/*
* Track the maximum transaction ID in the page. We store this in the
* tree at the end of reconciliation in the service of checkpoints, it
* is used to avoid discarding trees from memory when they have changes
* required to satisfy a snapshot read.
*/
if (WT_TXNID_LT(r->max_txn, max_txn))
r->max_txn = max_txn;
/*
* If there are no skipped updates and all updates are globally visible,
* the page can be marked clean and we're done, regardless if evicting
* or checkpointing.
*
* We have to check both: the oldest transaction ID may have moved while
* we were scanning the update list, so it is possible to find a skipped
* update, but then find all updates are stable at the end of the scan.
*
* Skip the visibility check for the lookaside table as a special-case,
* we know there are no older readers of that table.
*/
if (!skipped &&
(F_ISSET(btree, WT_BTREE_LOOKASIDE) ||
__wt_txn_visible_all(session, max_txn))) {
#ifdef HAVE_DIAGNOSTIC
/*
* The checkpoint transaction is special. Make sure we never
* write (metadata) updates from a checkpoint in a concurrent
* session.
*/
txnid = *updp == NULL ? WT_TXN_NONE : (*updp)->txnid;
WT_ASSERT(session, txnid == WT_TXN_NONE ||
txnid != S2C(session)->txn_global.checkpoint_txnid ||
WT_SESSION_IS_CHECKPOINT(session));
#endif
return (0);
}
/*
* In some cases, there had better not be skipped updates or updates not
* yet globally visible.
*/
if (F_ISSET(r, WT_VISIBILITY_ERR))
WT_PANIC_RET(session, EINVAL,
"reconciliation error, uncommitted update or update not "
"globally visible");
/*
* If not trying to evict the page, we know what we'll write and we're
* done. Because some updates were skipped or are not globally visible,
* the page can't be marked clean.
*/
if (!F_ISSET(r, WT_EVICTING)) {
r->leave_dirty = true;
return (0);
}
/*
* Evicting with either uncommitted changes or not-yet-globally-visible
* changes. There are two ways to continue, the save/restore eviction
* path or the lookaside table eviction path. Both cannot be configured
* because the paths track different information. The save/restore path
* can handle both uncommitted and not-yet-globally-visible changes, by
* evicting most of the page and then creating a new, smaller page into
* which we re-instantiate those changes. The lookaside table path can
* only handle not-yet-globally-visible changes by writing those changes
* into the lookaside table and restoring them on demand if and when the
* page is read back into memory.
*
* Both paths are configured outside of reconciliation: the save/restore
* path is the WT_EVICT_UPDATE_RESTORE flag, the lookaside table path is
* the WT_EVICT_LOOKASIDE flag.
*/
if (!F_ISSET(r, WT_EVICT_LOOKASIDE | WT_EVICT_UPDATE_RESTORE))
return (EBUSY);
if (skipped && !F_ISSET(r, WT_EVICT_UPDATE_RESTORE))
return (EBUSY);
append_origv = false;
if (F_ISSET(r, WT_EVICT_UPDATE_RESTORE)) {
/*
* The save/restore eviction path.
*
* Clear the returned update so our caller ignores the key/value
* pair in the case of an insert/append list entry (everything
* we need is in the update list), and otherwise writes the
* original on-page key/value pair to which the update list
* applies.
*/
*updp = NULL;
/* The page can't be marked clean. */
r->leave_dirty = true;
/*
* A special-case for overflow values, where we can't write the
* original on-page value item to disk because it's been updated
* or removed.
*
* What happens is that an overflow value is updated or removed
* and its backing blocks freed. If any reader in the system
* might still want the value, a copy was cached in the page
* reconciliation tracking memory, and the page cell set to
* WT_CELL_VALUE_OVFL_RM. Eviction then chose the page and
* we're splitting it up in order to push parts of it out of
* memory.
*
* We could write the original on-page value item to disk... if
* we had a copy. The cache may not have a copy (a globally
* visible update would have kept a value from being cached), or
* an update that subsequently became globally visible could
* cause a cached value to be discarded. Either way, once there
* is a globally visible update, we may not have the original
* value.
*
* Fortunately, if there's a globally visible update we don't
* care about the original version, so we simply ignore it, no
* transaction can ever try and read it. If there isn't a
* globally visible update, there had better be a cached value.
*
* In the latter case, we could write the value out to disk, but
* (1) we are planning on re-instantiating this page in memory,
* it isn't going to disk, and (2) the value item is eventually
* going to be discarded, that seems like a waste of a write.
* Instead, find the cached value and append it to the update
* list we're saving for later restoration.
*/
if (vpack != NULL &&
vpack->raw == WT_CELL_VALUE_OVFL_RM &&
!__wt_txn_visible_all(session, min_txn))
append_origv = true;
} else {
/*
* The lookaside table eviction path.
*
* If at least one update is globally visible, copy the update
* list and ignore the current on-page value. If no update is
* globally visible, readers require the page's original value.
*/
if (!__wt_txn_visible_all(session, min_txn))
append_origv = true;
}
/*
* We need the original on-page value for some reason: get a copy and
* append it to the end of the update list with a transaction ID that
* guarantees its visibility.
*/
if (append_origv) {
/*
* If we don't have a value cell, it's an insert/append list
* key/value pair which simply doesn't exist for some reader;
* place a deleted record at the end of the update list.
*/
if (vpack == NULL || vpack->type == WT_CELL_DEL)
WT_RET(__wt_update_alloc(
session, NULL, &append, ¬used));
else {
WT_RET(__wt_scr_alloc(session, 0, &tmp));
if ((ret = __wt_page_cell_data_ref(
session, page, vpack, tmp)) == 0)
ret = __wt_update_alloc(
session, tmp, &append, ¬used);
__wt_scr_free(session, &tmp);
WT_RET(ret);
}
/*
* Give the entry an impossibly low transaction ID to ensure its
* global visibility, append it to the update list.
*
* Note the change to the actual reader-accessible update list:
* from now on, the original on-page value appears at the end
* of the update list, even if this reconciliation subsequently
* fails.
*/
append->txnid = WT_TXN_NONE;
for (upd = upd_list; upd->next != NULL; upd = upd->next)
;
upd->next = append;
__wt_cache_page_inmem_incr(
session, page, WT_UPDATE_MEMSIZE(append));
}
/*
* The order of the updates on the list matters, we can't move only the
* unresolved updates, move the entire update list.
*
* If we skipped updates, the transaction value is never used. If we
* didn't skip updates, the list of updates are eventually written to
* the lookaside table, and associated with each update record is the
* transaction ID of the update we wrote in the reconciled page; once
* that transaction ID is globally visible, we know we no longer need
* the lookaside table records, allowing them to be discarded.
*/
return (__rec_update_save(session,
r, ins, rip, (*updp == NULL) ? WT_TXN_NONE : (*updp)->txnid));
}
/*
* WT_CHILD_RELEASE, WT_CHILD_RELEASE_ERR --
* Macros to clean up during internal-page reconciliation, releasing the
* hazard pointer we're holding on child pages.
*/
#define WT_CHILD_RELEASE(session, hazard, ref) do { \
if (hazard) { \
hazard = false; \
WT_TRET( \
__wt_page_release(session, ref, WT_READ_NO_EVICT)); \
} \
} while (0)
#define WT_CHILD_RELEASE_ERR(session, hazard, ref) do { \
WT_CHILD_RELEASE(session, hazard, ref); \
WT_ERR(ret); \
} while (0)
typedef enum {
WT_CHILD_IGNORE, /* Ignored child */
WT_CHILD_MODIFIED, /* Modified child */
WT_CHILD_ORIGINAL, /* Original child */
WT_CHILD_PROXY /* Deleted child: proxy */
} WT_CHILD_STATE;
/*
* __rec_child_deleted --
* Handle pages with leaf pages in the WT_REF_DELETED state.
*/
static int
__rec_child_deleted(WT_SESSION_IMPL *session,
WT_RECONCILE *r, WT_REF *ref, WT_CHILD_STATE *statep)
{
WT_PAGE_DELETED *page_del;
page_del = ref->page_del;
/*
* Internal pages with child leaf pages in the WT_REF_DELETED state are
* a special case during reconciliation. First, if the deletion was a
* result of a session truncate call, the deletion may not be visible to
* us. In that case, we proceed as with any change not visible during
* reconciliation by ignoring the change for the purposes of writing the
* internal page.
*
* In this case, there must be an associated page-deleted structure, and
* it holds the transaction ID we care about.
*
* In some cases, there had better not be any updates we can't see.
*/
if (F_ISSET(r, WT_VISIBILITY_ERR) &&
page_del != NULL && !__wt_txn_visible(session, page_del->txnid))
WT_PANIC_RET(session, EINVAL,
"reconciliation illegally skipped an update");
/*
* Deal with any underlying disk blocks.
*
* First, check to see if there is an address associated with this leaf:
* if there isn't, we're done, the underlying page is already gone. If
* the page still exists, check for any transactions in the system that
* might want to see the page's state before it's deleted.
*
* If any such transactions exist, we cannot discard the underlying leaf
* page to the block manager because the transaction may eventually read
* it. However, this write might be part of a checkpoint, and should we
* recover to that checkpoint, we'll need to delete the leaf page, else
* we'd leak it. The solution is to write a proxy cell on the internal
* page ensuring the leaf page is eventually discarded.
*
* If no such transactions exist, we can discard the leaf page to the
* block manager and no cell needs to be written at all. We do this
* outside of the underlying tracking routines because this action is
* permanent and irrevocable. (Clearing the address means we've lost
* track of the disk address in a permanent way. This is safe because
* there's no path to reading the leaf page again: if there's ever a
* read into this part of the name space again, the cache read function
* instantiates an entirely new page.)
*/
if (ref->addr != NULL &&
(page_del == NULL ||
__wt_txn_visible_all(session, page_del->txnid)))
WT_RET(__wt_ref_block_free(session, ref));
/*
* If the original page is gone, we can skip the slot on the internal
* page.
*/
if (ref->addr == NULL) {
*statep = WT_CHILD_IGNORE;
/*
* Minor memory cleanup: if a truncate call deleted this page
* and we were ever forced to instantiate the page in memory,
* we would have built a list of updates in the page reference
* in order to be able to abort the truncate. It's a cheap
* test to make that memory go away, we do it here because
* there's really nowhere else we do the checks. In short, if
* we have such a list, and the backing address blocks are
* gone, there can't be any transaction that can abort.
*/
if (page_del != NULL) {
__wt_free(session, ref->page_del->update_list);
__wt_free(session, ref->page_del);
}
return (0);
}
/*
* Internal pages with deletes that aren't stable cannot be evicted, we
* don't have sufficient information to restore the page's information
* if subsequently read (we wouldn't know which transactions should see
* the original page and which should see the deleted page).
*/
if (F_ISSET(r, WT_EVICTING))
return (EBUSY);
/*
* If there are deleted child pages we can't discard immediately, keep
* the page dirty so they are eventually freed.
*/
r->leave_dirty = true;
/*
* If the original page cannot be freed, we need to keep a slot on the
* page to reference it from the parent page.
*
* If the delete is not visible in this checkpoint, write the original
* address normally. Otherwise, we have to write a proxy record.
*/
if (__wt_txn_visible(session, page_del->txnid))
*statep = WT_CHILD_PROXY;
return (0);
}
/*
* __rec_child_modify --
* Return if the internal page's child references any modifications.
*/
static int
__rec_child_modify(WT_SESSION_IMPL *session,
WT_RECONCILE *r, WT_REF *ref, bool *hazardp, WT_CHILD_STATE *statep)
{
WT_DECL_RET;
WT_PAGE_MODIFY *mod;
/* We may acquire a hazard pointer our caller must release. */
*hazardp = false;
/* Default to using the original child address. */
*statep = WT_CHILD_ORIGINAL;
/*
* This function is called when walking an internal page to decide how
* to handle child pages referenced by the internal page.
*
* Internal pages are reconciled for two reasons: first, when evicting
* an internal page, second by the checkpoint code when writing internal
* pages. During eviction, all pages should be in the WT_REF_DISK or
* WT_REF_DELETED state. During checkpoint, eviction that might affect
* review of an internal page is prohibited, however, as the subtree is
* not reserved for our exclusive use, there are other page states that
* must be considered.
*/
for (;; __wt_yield())
switch (r->tested_ref_state = ref->state) {
case WT_REF_DISK:
/* On disk, not modified by definition. */
goto done;
case WT_REF_DELETED:
/*
* The child is in a deleted state.
*
* It's possible the state could change underneath us as
* the page is read in, and we can race between checking
* for a deleted state and looking at the transaction ID
* to see if the delete is visible to us. Lock down the
* structure.
*/
if (!__wt_atomic_casv32(
&ref->state, WT_REF_DELETED, WT_REF_LOCKED))
break;
ret = __rec_child_deleted(session, r, ref, statep);
WT_PUBLISH(ref->state, WT_REF_DELETED);
goto done;
case WT_REF_LOCKED:
/*
* Locked.
*
* We should never be here during eviction, active child
* pages in an evicted page's subtree fails the eviction
* attempt.
*/
if (F_ISSET(r, WT_EVICTING)) {
WT_ASSERT(session, !F_ISSET(r, WT_EVICTING));
return (EBUSY);
}
/*
* If called during checkpoint, the child is being
* considered by the eviction server or the child is a
* fast-delete page being read. The eviction may have
* started before the checkpoint and so we must wait
* for the eviction to be resolved. I suspect we could
* handle fast-delete reads, but we can't distinguish
* between the two and fast-delete reads aren't expected
* to be common.
*/
break;
case WT_REF_MEM:
/*
* In memory.
*
* We should never be here during eviction, active child
* pages in an evicted page's subtree fails the eviction
* attempt.
*/
if (F_ISSET(r, WT_EVICTING)) {
WT_ASSERT(session, !F_ISSET(r, WT_EVICTING));
return (EBUSY);
}
/*
* If called during checkpoint, acquire a hazard pointer
* so the child isn't evicted, it's an in-memory case.
*
* This call cannot return split/restart, we have a lock
* on the parent which prevents a child page split.
*/
ret = __wt_page_in(session, ref,
WT_READ_CACHE | WT_READ_NO_EVICT |
WT_READ_NO_GEN | WT_READ_NO_WAIT);
if (ret == WT_NOTFOUND) {
ret = 0;
break;
}
WT_RET(ret);
*hazardp = true;
goto in_memory;
case WT_REF_READING:
/*
* Being read, not modified by definition.
*
* We should never be here during eviction, active child
* pages in an evicted page's subtree fails the eviction
* attempt.
*/
if (F_ISSET(r, WT_EVICTING)) {
WT_ASSERT(session, !F_ISSET(r, WT_EVICTING));
return (EBUSY);
}
goto done;
case WT_REF_SPLIT:
/*
* The page was split out from under us.
*
* We should never be here during eviction, active child
* pages in an evicted page's subtree fails the eviction
* attempt.
*
* We should never be here during checkpoint, dirty page
* eviction is shutout during checkpoint, all splits in
* process will have completed before we walk any pages
* for checkpoint.
*/
WT_ASSERT(session, WT_REF_SPLIT != WT_REF_SPLIT);
return (EBUSY);
WT_ILLEGAL_VALUE(session);
}
in_memory:
/*
* In-memory states: the child is potentially modified if the page's
* modify structure has been instantiated. If the modify structure
* exists and the page has actually been modified, set that state.
* If that's not the case, we would normally use the original cell's
* disk address as our reference, however there are two special cases,
* both flagged by a missing block address.
*
* First, if forced to instantiate a deleted child page and it's never
* modified, we end up here with a page that has a modify structure, no
* modifications, and no disk address. Ignore those pages, they're not
* modified and there is no reason to write the cell.
*
* Second, insert splits are permitted during checkpoint. When doing the
* final checkpoint pass, we first walk the internal page's page-index
* and write out any dirty pages we find, then we write out the internal
* page in post-order traversal. If we found the split page in the first
* step, it will have an address; if we didn't find the split page in
* the first step, it won't have an address and we ignore it, it's not
* part of the checkpoint.
*/
mod = ref->page->modify;
if (mod != NULL && mod->rec_result != 0)
*statep = WT_CHILD_MODIFIED;
else if (ref->addr == NULL) {
*statep = WT_CHILD_IGNORE;
WT_CHILD_RELEASE(session, *hazardp, ref);
}
done: WT_DIAGNOSTIC_YIELD;
return (ret);
}
/*
* __rec_incr --
* Update the memory tracking structure for a set of new entries.
*/
static inline void
__rec_incr(WT_SESSION_IMPL *session, WT_RECONCILE *r, uint32_t v, size_t size)
{
/*
* The buffer code is fragile and prone to off-by-one errors -- check
* for overflow in diagnostic mode.
*/
WT_ASSERT(session, r->space_avail >= size);
WT_ASSERT(session, WT_BLOCK_FITS(
r->first_free, size, r->disk_image.mem, r->disk_image.memsize));
r->entries += v;
r->space_avail -= size;
r->first_free += size;
}
/*
* __rec_copy_incr --
* Copy a key/value cell and buffer pair into the new image.
*/
static inline void
__rec_copy_incr(WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_KV *kv)
{
size_t len;
uint8_t *p, *t;
/*
* If there's only one chunk of data to copy (because the cell and data
* are being copied from the original disk page), the cell length won't
* be set, the WT_ITEM data/length will reference the data to be copied.
*
* WT_CELLs are typically small, 1 or 2 bytes -- don't call memcpy, do
* the copy in-line.
*/
for (p = (uint8_t *)r->first_free,
t = (uint8_t *)&kv->cell, len = kv->cell_len; len > 0; --len)
*p++ = *t++;
/* The data can be quite large -- call memcpy. */
if (kv->buf.size != 0)
memcpy(p, kv->buf.data, kv->buf.size);
WT_ASSERT(session, kv->len == kv->cell_len + kv->buf.size);
__rec_incr(session, r, 1, kv->len);
}
/*
* __rec_dict_replace --
* Check for a dictionary match.
*/
static int
__rec_dict_replace(
WT_SESSION_IMPL *session, WT_RECONCILE *r, uint64_t rle, WT_KV *val)
{
WT_DICTIONARY *dp;
uint64_t offset;
/*
* We optionally create a dictionary of values and only write a unique
* value once per page, using a special "copy" cell for all subsequent
* copies of the value. We have to do the cell build and resolution at
* this low level because we need physical cell offsets for the page.
*
* Sanity check: short-data cells can be smaller than dictionary-copy
* cells. If the data is already small, don't bother doing the work.
* This isn't just work avoidance: on-page cells can't grow as a result
* of writing a dictionary-copy cell, the reconciliation functions do a
* split-boundary test based on the size required by the value's cell;
* if we grow the cell after that test we'll potentially write off the
* end of the buffer's memory.
*/
if (val->buf.size <= WT_INTPACK32_MAXSIZE)
return (0);
WT_RET(__rec_dictionary_lookup(session, r, val, &dp));
if (dp == NULL)
return (0);
/*
* If the dictionary cell reference is not set, we're creating a new
* entry in the dictionary, update its location.
*
* If the dictionary cell reference is set, we have a matching value.
* Create a copy cell instead.
*/
if (dp->cell == NULL)
dp->cell = r->first_free;
else {
offset = WT_PTRDIFF(r->first_free, dp->cell);
val->len = val->cell_len =
__wt_cell_pack_copy(&val->cell, rle, offset);
val->buf.data = NULL;
val->buf.size = 0;
}
return (0);
}
/*
* __rec_key_state_update --
* Update prefix and suffix compression based on the last key.
*/
static inline void
__rec_key_state_update(WT_RECONCILE *r, bool ovfl_key)
{
WT_ITEM *a;
/*
* If writing an overflow key onto the page, don't update the "last key"
* value, and leave the state of prefix compression alone. (If we are
* currently doing prefix compression, we have a key state which will
* continue to work, we're just skipping the key just created because
* it's an overflow key and doesn't participate in prefix compression.
* If we are not currently doing prefix compression, we can't start, an
* overflow key doesn't give us any state.)
*
* Additionally, if we wrote an overflow key onto the page, turn off the
* suffix compression of row-store internal node keys. (When we split,
* "last key" is the largest key on the previous page, and "cur key" is
* the first key on the next page, which is being promoted. In some
* cases we can discard bytes from the "cur key" that are not needed to
* distinguish between the "last key" and "cur key", compressing the
* size of keys on internal nodes. If we just built an overflow key,
* we're not going to update the "last key", making suffix compression
* impossible for the next key. Alternatively, we could remember where
* the last key was on the page, detect it's an overflow key, read it
* from disk and do suffix compression, but that's too much work for an
* unlikely event.)
*
* If we're not writing an overflow key on the page, update the last-key
* value and turn on both prefix and suffix compression.
*/
if (ovfl_key)
r->key_sfx_compress = false;
else {
a = r->cur;
r->cur = r->last;
r->last = a;
r->key_pfx_compress = r->key_pfx_compress_conf;
r->key_sfx_compress = r->key_sfx_compress_conf;
}
}
/*
* Macros from fixed-length entries to/from bytes.
*/
#define WT_FIX_BYTES_TO_ENTRIES(btree, bytes) \
((uint32_t)((((bytes) * 8) / (btree)->bitcnt)))
#define WT_FIX_ENTRIES_TO_BYTES(btree, entries) \
((uint32_t)WT_ALIGN((entries) * (btree)->bitcnt, 8))
/*
* __rec_leaf_page_max --
* Figure out the maximum leaf page size for the reconciliation.
*/
static inline uint32_t
__rec_leaf_page_max(WT_SESSION_IMPL *session, WT_RECONCILE *r)
{
WT_BTREE *btree;
WT_PAGE *page;
uint32_t page_size;
btree = S2BT(session);
page = r->page;
page_size = 0;
switch (page->type) {
case WT_PAGE_COL_FIX:
/*
* Column-store pages can grow if there are missing records
* (that is, we lost a chunk of the range, and have to write
* deleted records). Fixed-length objects are a problem, if
* there's a big missing range, we could theoretically have to
* write large numbers of missing objects.
*/
page_size = (uint32_t)WT_ALIGN(WT_FIX_ENTRIES_TO_BYTES(btree,
r->salvage->take + r->salvage->missing), btree->allocsize);
break;
case WT_PAGE_COL_VAR:
/*
* Column-store pages can grow if there are missing records
* (that is, we lost a chunk of the range, and have to write
* deleted records). Variable-length objects aren't usually a
* problem because we can write any number of deleted records
* in a single page entry because of the RLE, we just need to
* ensure that additional entry fits.
*/
break;
case WT_PAGE_ROW_LEAF:
default:
/*
* Row-store pages can't grow, salvage never does anything
* other than reduce the size of a page read from disk.
*/
break;
}
/*
* Default size for variable-length column-store and row-store pages
* during salvage is the maximum leaf page size.
*/
if (page_size < btree->maxleafpage)
page_size = btree->maxleafpage;
/*
* The page we read from the disk should be smaller than the page size
* we just calculated, check out of paranoia.
*/
if (page_size < page->dsk->mem_size)
page_size = page->dsk->mem_size;
/*
* Salvage is the backup plan: don't let this fail.
*/
return (page_size * 2);
}
/*
* __rec_split_bnd_init --
* Initialize a single boundary structure.
*/
static void
__rec_split_bnd_init(WT_SESSION_IMPL *session, WT_BOUNDARY *bnd)
{
bnd->offset = 0;
bnd->recno = WT_RECNO_OOB;
bnd->entries = 0;
__wt_free(session, bnd->addr.addr);
WT_CLEAR(bnd->addr);
bnd->size = 0;
bnd->cksum = 0;
__wt_free(session, bnd->disk_image);
__wt_free(session, bnd->supd);
bnd->supd_next = 0;
bnd->supd_allocated = 0;
bnd->already_compressed = false;
/*
* Don't touch the key, we re-use that memory in each new
* reconciliation.
*/
}
/*
* __rec_split_bnd_grow --
* Grow the boundary array as necessary.
*/
static int
__rec_split_bnd_grow(WT_SESSION_IMPL *session, WT_RECONCILE *r)
{
/*
* Make sure there's enough room for another boundary. The calculation
* is +2, because when filling in the current boundary's information,
* we save start information for the next boundary (a byte offset and a
* record number or key), in the (current + 1) slot.
*
* For the same reason, we're always initializing one ahead.
*/
WT_RET(__wt_realloc_def(
session, &r->bnd_allocated, r->bnd_next + 2, &r->bnd));
r->bnd_entries = r->bnd_allocated / sizeof(r->bnd[0]);
__rec_split_bnd_init(session, &r->bnd[r->bnd_next + 1]);
return (0);
}
/*
* __wt_split_page_size --
* Split page size calculation: we don't want to repeatedly split every
* time a new entry is added, so we split to a smaller-than-maximum page size.
*/
uint32_t
__wt_split_page_size(WT_BTREE *btree, uint32_t maxpagesize)
{
uintmax_t a;
uint32_t split_size;
/*
* Ideally, the split page size is some percentage of the maximum page
* size rounded to an allocation unit (round to an allocation unit so
* we don't waste space when we write).
*/
a = maxpagesize; /* Don't overflow. */
split_size = (uint32_t)WT_ALIGN_NEAREST(
(a * (u_int)btree->split_pct) / 100, btree->allocsize);
/*
* Respect the configured split percentage if the calculated split
* size is either zero or a full page. The user has either configured
* an allocation size that matches the page size, or a split
* percentage that is close to zero or one hundred. Rounding is going
* to provide a worse outcome than having a split point that doesn't
* fall on an allocation size boundary in those cases.
*/
if (split_size == 0 || split_size == maxpagesize)
split_size = (uint32_t)((a * (u_int)btree->split_pct) / 100);
return (split_size);
}
/*
* __rec_split_init --
* Initialization for the reconciliation split functions.
*/
static int
__rec_split_init(WT_SESSION_IMPL *session,
WT_RECONCILE *r, WT_PAGE *page, uint64_t recno, uint32_t max)
{
WT_BM *bm;
WT_BTREE *btree;
WT_PAGE_HEADER *dsk;
size_t corrected_page_size;
btree = S2BT(session);
bm = btree->bm;
/*
* The maximum leaf page size governs when an in-memory leaf page splits
* into multiple on-disk pages; however, salvage can't be allowed to
* split, there's no parent page yet. If we're doing salvage, override
* the caller's selection of a maximum page size, choosing a page size
* that ensures we won't split.
*/
if (r->salvage != NULL)
max = __rec_leaf_page_max(session, r);
/*
* Set the page sizes. If we're doing the page layout, the maximum page
* size is the same as the page size. If the application is doing page
* layout (raw compression is configured), we accumulate some amount of
* additional data because we don't know how well it will compress, and
* we don't want to increment our way up to the amount of data needed by
* the application to successfully compress to the target page size.
* Ideally accumulate data several times the page size without
* approaching the memory page maximum, but at least have data worth
* one page.
*
* There are cases when we grow the page size to accommodate large
* records, in those cases we split the pages once they have crossed
* the maximum size for a page with raw compression.
*/
r->page_size = r->page_size_orig = max;
if (r->raw_compression)
r->max_raw_page_size = r->page_size =
(uint32_t)WT_MIN(r->page_size * 10,
WT_MAX(r->page_size, btree->maxmempage / 2));
/*
* Ensure the disk image buffer is large enough for the max object, as
* corrected by the underlying block manager.
*/
corrected_page_size = r->page_size;
WT_RET(bm->write_size(bm, session, &corrected_page_size));
WT_RET(__wt_buf_init(session, &r->disk_image, corrected_page_size));
/*
* Clear the disk page header to ensure all of it is initialized, even
* the unused fields.
*
* In the case of fixed-length column-store, clear the entire buffer:
* fixed-length column-store sets bits in bytes, where the bytes are
* assumed to initially be 0.
*/
memset(r->disk_image.mem, 0, page->type == WT_PAGE_COL_FIX ?
corrected_page_size : WT_PAGE_HEADER_SIZE);
/*
* Set the page type (the type doesn't change, and setting it later
* would require additional code in a few different places).
*/
dsk = r->disk_image.mem;
dsk->type = page->type;
/*
* If we have to split, we want to choose a smaller page size for the
* split pages, because otherwise we could end up splitting one large
* packed page over and over. We don't want to pick the minimum size
* either, because that penalizes an application that did a bulk load
* and subsequently inserted a few items into packed pages. Currently
* defaulted to 75%, but I have no empirical evidence that's "correct".
*
* The maximum page size may be a multiple of the split page size (for
* example, there's a maximum page size of 128KB, but because the table
* is active and we don't want to split a lot, the split size is 20KB).
* The maximum page size may NOT be an exact multiple of the split page
* size.
*
* It's lots of work to build these pages and don't want to start over
* when we reach the maximum page size (it's painful to restart after
* creating overflow items and compacted data, for example, as those
* items have already been written to disk). So, the loop calls the
* helper functions when approaching a split boundary, and we save the
* information at that point. That allows us to go back and split the
* page at the boundary points if we eventually overflow the maximum
* page size.
*
* Finally, all this doesn't matter for fixed-size column-store pages,
* raw compression, and salvage. Fixed-size column store pages can
* split under (very) rare circumstances, but they're allocated at a
* fixed page size, never anything smaller. In raw compression, the
* underlying compression routine decides when we split, so it's not
* our problem. In salvage, as noted above, we can't split at all.
*/
if (r->raw_compression || r->salvage != NULL) {
r->split_size = 0;
r->space_avail = r->page_size - WT_PAGE_HEADER_BYTE_SIZE(btree);
}
else if (page->type == WT_PAGE_COL_FIX) {
r->split_size = r->page_size;
r->space_avail =
r->split_size - WT_PAGE_HEADER_BYTE_SIZE(btree);
} else {
r->split_size = __wt_split_page_size(btree, r->page_size);
r->space_avail =
r->split_size - WT_PAGE_HEADER_BYTE_SIZE(btree);
}
r->first_free = WT_PAGE_HEADER_BYTE(btree, dsk);
/* Initialize the first boundary. */
r->bnd_next = 0;
WT_RET(__rec_split_bnd_grow(session, r));
__rec_split_bnd_init(session, &r->bnd[0]);
r->bnd[0].recno = recno;
r->bnd[0].offset = WT_PAGE_HEADER_BYTE_SIZE(btree);
/*
* If the maximum page size is the same as the split page size, either
* because of the object type or application configuration, there isn't
* any need to maintain split boundaries within a larger page.
*
* No configuration for salvage here, because salvage can't split.
*/
if (r->raw_compression)
r->bnd_state = SPLIT_TRACKING_RAW;
else if (max == r->split_size)
r->bnd_state = SPLIT_TRACKING_OFF;
else
r->bnd_state = SPLIT_BOUNDARY;
/* Initialize the entry counters. */
r->entries = r->total_entries = 0;
/* Initialize the starting record number. */
r->recno = recno;
/* New page, compression off. */
r->key_pfx_compress = r->key_sfx_compress = false;
return (0);
}
/*
* __rec_is_checkpoint --
* Return if we're writing a checkpoint.
*/
static bool
__rec_is_checkpoint(WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_BOUNDARY *bnd)
{
WT_BTREE *btree;
btree = S2BT(session);
/*
* Check to see if we're going to create a checkpoint.
*
* This function exists as a place to hang this comment.
*
* Any time we write the root page of the tree without splitting we are
* creating a checkpoint (and have to tell the underlying block manager
* so it creates and writes the additional information checkpoints
* require). However, checkpoints are completely consistent, and so we
* have to resolve information about the blocks we're expecting to free
* as part of the checkpoint, before writing the checkpoint. In short,
* we don't do checkpoint writes here; clear the boundary information as
* a reminder and create the checkpoint during wrapup.
*/
if (!F_ISSET(btree, WT_BTREE_NO_CHECKPOINT) &&
bnd == &r->bnd[0] && __wt_ref_is_root(r->ref)) {
bnd->addr.addr = NULL;
bnd->addr.size = 0;
bnd->addr.type = 0;
return (true);
}
return (false);
}
/*
* __rec_split_row_promote_cell --
* Get a key from a cell for the purposes of promotion.
*/
static int
__rec_split_row_promote_cell(
WT_SESSION_IMPL *session, WT_PAGE_HEADER *dsk, WT_ITEM *key)
{
WT_BTREE *btree;
WT_CELL *cell;
WT_CELL_UNPACK *kpack, _kpack;
btree = S2BT(session);
kpack = &_kpack;
/*
* The cell had better have a zero-length prefix and not be a copy cell;
* the first cell on a page cannot refer an earlier cell on the page.
*/
cell = WT_PAGE_HEADER_BYTE(btree, dsk);
__wt_cell_unpack(cell, kpack);
WT_ASSERT(session,
kpack->prefix == 0 && kpack->raw != WT_CELL_VALUE_COPY);
WT_RET(__wt_cell_data_copy(session, dsk->type, kpack, key));
return (0);
}
/*
* __rec_split_row_promote --
* Key promotion for a row-store.
*/
static int
__rec_split_row_promote(
WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_ITEM *key, uint8_t type)
{
WT_BTREE *btree;
WT_DECL_ITEM(update);
WT_DECL_RET;
WT_ITEM *max;
WT_SAVE_UPD *supd;
size_t cnt, len, size;
uint32_t i;
const uint8_t *pa, *pb;
int cmp;
/*
* For a column-store, the promoted key is the recno and we already have
* a copy. For a row-store, it's the first key on the page, a variable-
* length byte string, get a copy.
*
* This function is called from the split code at each split boundary,
* but that means we're not called before the first boundary, and we
* will eventually have to get the first key explicitly when splitting
* a page.
*
* For the current slot, take the last key we built, after doing suffix
* compression. The "last key we built" describes some process: before
* calling the split code, we must place the last key on the page before
* the boundary into the "last" key structure, and the first key on the
* page after the boundary into the "current" key structure, we're going
* to compare them for suffix compression.
*
* Suffix compression is a hack to shorten keys on internal pages. We
* only need enough bytes in the promoted key to ensure searches go to
* the correct page: the promoted key has to be larger than the last key
* on the leaf page preceding it, but we don't need any more bytes than
* that. In other words, we can discard any suffix bytes not required
* to distinguish between the key being promoted and the last key on the
* leaf page preceding it. This can only be done for the first level of
* internal pages, you cannot repeat suffix truncation as you split up
* the tree, it loses too much information.
*
* Note #1: if the last key on the previous page was an overflow key,
* we don't have the in-memory key against which to compare, and don't
* try to do suffix compression. The code for that case turns suffix
* compression off for the next key, we don't have to deal with it here.
*/
if (type != WT_PAGE_ROW_LEAF || !r->key_sfx_compress)
return (__wt_buf_set(session, key, r->cur->data, r->cur->size));
btree = S2BT(session);
WT_RET(__wt_scr_alloc(session, 0, &update));
/*
* Note #2: if we skipped updates, an update key may be larger than the
* last key stored in the previous block (probable for append-centric
* workloads). If there are skipped updates, check for one larger than
* the last key and smaller than the current key.
*/
max = r->last;
if (F_ISSET(r, WT_EVICT_UPDATE_RESTORE))
for (i = r->supd_next; i > 0; --i) {
supd = &r->supd[i - 1];
if (supd->ins == NULL)
WT_ERR(__wt_row_leaf_key(session,
r->page, supd->rip, update, false));
else {
update->data = WT_INSERT_KEY(supd->ins);
update->size = WT_INSERT_KEY_SIZE(supd->ins);
}
/* Compare against the current key, it must be less. */
WT_ERR(__wt_compare(
session, btree->collator, update, r->cur, &cmp));
if (cmp >= 0)
continue;
/* Compare against the last key, it must be greater. */
WT_ERR(__wt_compare(
session, btree->collator, update, r->last, &cmp));
if (cmp >= 0)
max = update;
/*
* The saved updates are in key-sort order so the entry
* we're looking for is either the last or the next-to-
* last one in the list. Once we've compared an entry
* against the last key on the page, we're done.
*/
break;
}
/*
* The largest key on the last block must sort before the current key,
* so we'll either find a larger byte value in the current key, or the
* current key will be a longer key, and the interesting byte is one
* past the length of the shorter key.
*/
pa = max->data;
pb = r->cur->data;
len = WT_MIN(max->size, r->cur->size);
size = len + 1;
for (cnt = 1; len > 0; ++cnt, --len, ++pa, ++pb)
if (*pa != *pb) {
if (size != cnt) {
WT_STAT_FAST_DATA_INCRV(session,
rec_suffix_compression, size - cnt);
size = cnt;
}
break;
}
ret = __wt_buf_set(session, key, r->cur->data, size);
err: __wt_scr_free(session, &update);
return (ret);
}
/*
* __rec_split_grow --
* Grow the split buffer.
*/
static int
__rec_split_grow(WT_SESSION_IMPL *session, WT_RECONCILE *r, size_t add_len)
{
WT_BM *bm;
WT_BTREE *btree;
size_t corrected_page_size, len;
btree = S2BT(session);
bm = btree->bm;
len = WT_PTRDIFF(r->first_free, r->disk_image.mem);
corrected_page_size = len + add_len;
WT_RET(bm->write_size(bm, session, &corrected_page_size));
WT_RET(__wt_buf_grow(session, &r->disk_image, corrected_page_size));
r->first_free = (uint8_t *)r->disk_image.mem + len;
WT_ASSERT(session, corrected_page_size >= len);
r->space_avail = corrected_page_size - len;
WT_ASSERT(session, r->space_avail >= add_len);
return (0);
}
/*
* __rec_split --
* Handle the page reconciliation bookkeeping. (Did you know "bookkeeper"
* has 3 doubled letters in a row? Sweet-tooth does, too.)
*/
static int
__rec_split(WT_SESSION_IMPL *session, WT_RECONCILE *r, size_t next_len)
{
WT_BOUNDARY *last, *next;
WT_BTREE *btree;
WT_PAGE_HEADER *dsk;
size_t inuse;
btree = S2BT(session);
dsk = r->disk_image.mem;
/*
* We should never split during salvage, and we're about to drop core
* because there's no parent page.
*/
if (r->salvage != NULL)
WT_PANIC_RET(session, WT_PANIC,
"%s page too large, attempted split during salvage",
__wt_page_type_string(r->page->type));
/* Hitting a page boundary resets the dictionary, in all cases. */
__rec_dictionary_reset(r);
inuse = WT_PTRDIFF(r->first_free, dsk);
switch (r->bnd_state) {
case SPLIT_BOUNDARY:
/*
* We can get here if the first key/value pair won't fit.
* Additionally, grow the buffer to contain the current item if
* we haven't already consumed a reasonable portion of a split
* chunk.
*/
if (inuse < r->split_size / 2)
break;
/*
* About to cross a split boundary but not yet forced to split
* into multiple pages. If we have to split, this is one of the
* split points, save information about where we are when the
* split would have happened.
*/
WT_RET(__rec_split_bnd_grow(session, r));
last = &r->bnd[r->bnd_next++];
next = last + 1;
/* Set the number of entries for the just finished chunk. */
last->entries = r->entries - r->total_entries;
r->total_entries = r->entries;
/* Set the key for the next chunk. */
next->recno = r->recno;
if (dsk->type == WT_PAGE_ROW_INT ||
dsk->type == WT_PAGE_ROW_LEAF)
WT_RET(__rec_split_row_promote(
session, r, &next->key, dsk->type));
/*
* Set the starting buffer offset and clear the entries (the
* latter not required, but cleaner).
*/
next->offset = WT_PTRDIFF(r->first_free, dsk);
next->entries = 0;
/* Set the space available to another split-size chunk. */
r->space_avail =
r->split_size - WT_PAGE_HEADER_BYTE_SIZE(btree);
/*
* Adjust the space available to handle two cases:
* - We don't have enough room for another full split-size
* chunk on the page.
* - We chose to fill past a page boundary because of a
* large item.
*/
if (inuse + r->space_avail > r->page_size) {
r->space_avail =
r->page_size > inuse ? (r->page_size - inuse) : 0;
/* There are no further boundary points. */
r->bnd_state = SPLIT_MAX;
}
/*
* Return if the next object fits into this page, else we have
* to split the page.
*/
if (r->space_avail >= next_len)
return (0);
/* FALLTHROUGH */
case SPLIT_MAX:
/*
* We're going to have to split and create multiple pages.
*
* Cycle through the saved split-point information, writing the
* split chunks we have tracked. The underlying fixup function
* sets the space available and other information, and copied
* any unwritten chunk of data to the beginning of the buffer.
*/
WT_RET(__rec_split_fixup(session, r));
/* We're done saving split chunks. */
r->bnd_state = SPLIT_TRACKING_OFF;
break;
case SPLIT_TRACKING_OFF:
/*
* We can get here if the first key/value pair won't fit.
* Additionally, grow the buffer to contain the current item if
* we haven't already consumed a reasonable portion of a split
* chunk.
*/
if (inuse < r->split_size / 2)
break;
/*
* The key/value pairs didn't fit into a single page, but either
* we've already noticed that and are now processing the rest of
* the pairs at split size boundaries, or the split size was the
* same as the page size, and we never bothered with split point
* information at all.
*/
WT_RET(__rec_split_bnd_grow(session, r));
last = &r->bnd[r->bnd_next++];
next = last + 1;
/*
* Set the key for the next chunk (before writing the block, a
* key range is needed in that code).
*/
next->recno = r->recno;
if (dsk->type == WT_PAGE_ROW_INT ||
dsk->type == WT_PAGE_ROW_LEAF)
WT_RET(__rec_split_row_promote(
session, r, &next->key, dsk->type));
/* Clear the entries (not required, but cleaner). */
next->entries = 0;
/* Finalize the header information and write the page. */
dsk->recno = last->recno;
dsk->u.entries = r->entries;
dsk->mem_size = WT_PTRDIFF32(r->first_free, dsk);
r->disk_image.size = dsk->mem_size;
WT_RET(
__rec_split_write(session, r, last, &r->disk_image, false));
/*
* Set the caller's entry count and buffer information for the
* next chunk. We only get here if we're not splitting or have
* already split, so it's split-size chunks from here on out.
*/
r->entries = 0;
r->first_free = WT_PAGE_HEADER_BYTE(btree, dsk);
r->space_avail =
r->split_size - WT_PAGE_HEADER_BYTE_SIZE(btree);
break;
case SPLIT_TRACKING_RAW:
return (__wt_illegal_value(session, NULL));
}
/*
* Overflow values can be larger than the maximum page size but still be
* "on-page". If the next key/value pair is larger than space available
* after a split has happened (in other words, larger than the maximum
* page size), create a page sized to hold that one key/value pair. This
* generally splits the page into key/value pairs before a large object,
* the object, and key/value pairs after the object. It's possible other
* key/value pairs will also be aggregated onto the bigger page before
* or after, if the page happens to hold them, but it won't necessarily
* happen that way.
*/
if (r->space_avail < next_len)
WT_RET(__rec_split_grow(session, r, next_len));
return (0);
}
/*
* __rec_split_raw_worker --
* Handle the raw compression page reconciliation bookkeeping.
*/
static int
__rec_split_raw_worker(WT_SESSION_IMPL *session,
WT_RECONCILE *r, size_t next_len, bool no_more_rows)
{
WT_BM *bm;
WT_BOUNDARY *last, *next;
WT_BTREE *btree;
WT_CELL *cell;
WT_CELL_UNPACK *unpack, _unpack;
WT_COMPRESSOR *compressor;
WT_DECL_RET;
WT_ITEM *dst, *write_ref;
WT_PAGE_HEADER *dsk, *dsk_dst, *disk_image;
WT_SESSION *wt_session;
size_t corrected_page_size, extra_skip, len, result_len;
uint64_t recno;
uint32_t entry, i, max_image_slot, result_slots, slots;
bool last_block;
uint8_t *dsk_start;
wt_session = (WT_SESSION *)session;
btree = S2BT(session);
bm = btree->bm;
unpack = &_unpack;
compressor = btree->compressor;
dst = &r->raw_destination;
dsk = r->disk_image.mem;
WT_RET(__rec_split_bnd_grow(session, r));
last = &r->bnd[r->bnd_next];
next = last + 1;
/*
* We can get here if the first key/value pair won't fit.
*/
if (r->entries == 0)
goto split_grow;
/*
* Build arrays of offsets and cumulative counts of cells and rows in
* the page: the offset is the byte offset to the possible split-point
* (adjusted for an initial chunk that cannot be compressed), entries
* is the cumulative page entries covered by the byte offset, recnos is
* the cumulative rows covered by the byte offset. Allocate to handle
* both column- and row-store regardless of this page type, structures
* are potentially reused for subsequent reconciliations of different
* page types.
*/
if (r->entries >= r->raw_max_slots) {
__wt_free(session, r->raw_entries);
__wt_free(session, r->raw_offsets);
__wt_free(session, r->raw_recnos);
r->raw_max_slots = 0;
i = r->entries + 100;
WT_RET(__wt_calloc_def(session, i, &r->raw_entries));
WT_RET(__wt_calloc_def(session, i, &r->raw_offsets));
WT_RET(__wt_calloc_def(session, i, &r->raw_recnos));
r->raw_max_slots = i;
}
/*
* We're going to walk the disk image, which requires setting the
* number of entries.
*/
dsk->u.entries = r->entries;
/*
* We track the record number at each column-store split point, set an
* initial value.
*/
recno = WT_RECNO_OOB;
if (dsk->type == WT_PAGE_COL_VAR)
recno = last->recno;
entry = max_image_slot = slots = 0;
WT_CELL_FOREACH(btree, dsk, cell, unpack, i) {
++entry;
/*
* Row-store pages can split at keys, but not at values,
* column-store pages can split at values.
*/
__wt_cell_unpack(cell, unpack);
switch (unpack->type) {
case WT_CELL_KEY:
case WT_CELL_KEY_OVFL:
case WT_CELL_KEY_SHORT:
break;
case WT_CELL_ADDR_DEL:
case WT_CELL_ADDR_INT:
case WT_CELL_ADDR_LEAF:
case WT_CELL_ADDR_LEAF_NO:
case WT_CELL_DEL:
case WT_CELL_VALUE:
case WT_CELL_VALUE_OVFL:
case WT_CELL_VALUE_SHORT:
if (dsk->type == WT_PAGE_COL_INT) {
recno = unpack->v;
break;
}
if (dsk->type == WT_PAGE_COL_VAR) {
recno += __wt_cell_rle(unpack);
break;
}
r->raw_entries[slots] = entry;
continue;
WT_ILLEGAL_VALUE(session);
}
/*
* We can't compress the first 64B of the block (it must be
* written without compression), and a possible split point
* may appear in that 64B; keep it simple, ignore the first
* allocation size of data, anybody splitting smaller than
* that (as calculated before compression), is doing it wrong.
*/
if ((len = WT_PTRDIFF(cell, dsk)) > btree->allocsize)
r->raw_offsets[++slots] =
WT_STORE_SIZE(len - WT_BLOCK_COMPRESS_SKIP);
if (dsk->type == WT_PAGE_COL_INT ||
dsk->type == WT_PAGE_COL_VAR)
r->raw_recnos[slots] = recno;
r->raw_entries[slots] = entry;
/*
* Don't create an image so large that any future update will
* cause a split in memory.
*/
if (max_image_slot == 0 && len > (size_t)r->max_raw_page_size)
max_image_slot = slots;
}
/*
* If we haven't managed to find at least one split point, we're done,
* don't bother calling the underlying compression function.
*/
if (slots == 0) {
result_len = 0;
result_slots = 0;
goto no_slots;
}
/* The slot at array's end is the total length of the data. */
r->raw_offsets[++slots] =
WT_STORE_SIZE(WT_PTRDIFF(cell, dsk) - WT_BLOCK_COMPRESS_SKIP);
/*
* Allocate a destination buffer. If there's a pre-size function, call
* it to determine the destination buffer's size, else the destination
* buffer is documented to be at least the source size. (We can't use
* the target page size, any single key/value could be larger than the
* page size. Don't bother figuring out a minimum, just use the source
* size.)
*
* The destination buffer needs to be large enough for the final block
* size, corrected for the requirements of the underlying block manager.
* If the final block size is 8KB, that's a multiple of 512B and so the
* underlying block manager is fine with it. But... we don't control
* what the pre_size method returns us as a required size, and we don't
* want to document the compress_raw method has to skip bytes in the
* buffer because that's confusing, so do something more complicated.
* First, find out how much space the compress_raw function might need,
* either the value returned from pre_size, or the initial source size.
* Add the compress-skip bytes, and then correct that value for the
* underlying block manager. As a result, we have a destination buffer
* that's large enough when calling the compress_raw method, and there
* are bytes in the header just for us.
*/
if (compressor->pre_size == NULL)
result_len = (size_t)r->raw_offsets[slots];
else
WT_RET(compressor->pre_size(compressor, wt_session,
(uint8_t *)dsk + WT_BLOCK_COMPRESS_SKIP,
(size_t)r->raw_offsets[slots], &result_len));
extra_skip = btree->kencryptor == NULL ? 0 :
btree->kencryptor->size_const + WT_ENCRYPT_LEN_SIZE;
corrected_page_size = result_len + WT_BLOCK_COMPRESS_SKIP;
WT_RET(bm->write_size(bm, session, &corrected_page_size));
WT_RET(__wt_buf_init(session, dst, corrected_page_size));
/*
* Copy the header bytes into the destination buffer, then call the
* compression function.
*/
memcpy(dst->mem, dsk, WT_BLOCK_COMPRESS_SKIP);
ret = compressor->compress_raw(compressor, wt_session,
r->page_size_orig, btree->split_pct,
WT_BLOCK_COMPRESS_SKIP + extra_skip,
(uint8_t *)dsk + WT_BLOCK_COMPRESS_SKIP, r->raw_offsets,
max_image_slot == 0 ? slots : max_image_slot,
(uint8_t *)dst->mem + WT_BLOCK_COMPRESS_SKIP,
result_len,
no_more_rows || max_image_slot != 0,
&result_len, &result_slots);
switch (ret) {
case EAGAIN:
/*
* The compression function wants more rows, accumulate and
* retry if possible.
*
* First, reset the resulting slots count, just in case the
* compression function modified it before giving up.
*/
result_slots = 0;
/*
* If the image is too large and there are more rows to gather,
* act as if the compression engine gave up on this chunk of
* data. That doesn't make sense (we flagged the engine that we
* wouldn't give it any more rows, but it's a possible return).
*/
if (no_more_rows || max_image_slot == 0)
break;
/* FALLTHROUGH */
case 0:
/*
* If the compression function returned zero result slots, it's
* giving up and we write the original data. (This is a pretty
* bad result: we've not done compression on a block much larger
* than the maximum page size, but once compression gives up,
* there's not much else we can do.)
*
* If the compression function returned non-zero result slots,
* we were successful and have a block to write.
*/
if (result_slots == 0) {
WT_STAT_FAST_DATA_INCR(session, compress_raw_fail);
/*
* If there are no more rows, we can write the original
* data from the original buffer.
*/
if (no_more_rows)
break;
/*
* Copy the original data to the destination buffer, as
* if the compression function simply copied it. Take
* all but the last row of the original data (the last
* row has to be set as the key for the next block).
*/
result_slots = slots - 1;
result_len = r->raw_offsets[result_slots];
WT_RET(__wt_buf_grow(
session, dst, result_len + WT_BLOCK_COMPRESS_SKIP));
memcpy((uint8_t *)dst->mem + WT_BLOCK_COMPRESS_SKIP,
(uint8_t *)dsk + WT_BLOCK_COMPRESS_SKIP,
result_len);
/*
* Mark it as uncompressed so the standard compression
* function is called before the buffer is written.
*/
last->already_compressed = false;
} else {
WT_STAT_FAST_DATA_INCR(session, compress_raw_ok);
/*
* If there are more rows and the compression function
* consumed all of the current data, there are problems:
* First, with row-store objects, we're potentially
* skipping updates, we must have a key for the next
* block so we know with what block a skipped update is
* associated. Second, if the compression function
* compressed all of the data, we're not pushing it
* hard enough (unless we got lucky and gave it exactly
* the right amount to work with, which is unlikely).
* Handle both problems by accumulating more data any
* time we're not writing the last block and compression
* ate all of the rows.
*/
if (result_slots == slots && !no_more_rows)
result_slots = 0;
else
last->already_compressed = true;
}
break;
default:
return (ret);
}
no_slots:
/*
* Check for the last block we're going to write: if no more rows and
* we failed to compress anything, or we compressed everything, it's
* the last block.
*/
last_block = no_more_rows &&
(result_slots == 0 || result_slots == slots);
if (result_slots != 0) {
/*
* We have a block, finalize the compressed disk image's header
* information.
*/
dst->size = result_len + WT_BLOCK_COMPRESS_SKIP;
dsk_dst = dst->mem;
dsk_dst->recno = last->recno;
dsk_dst->mem_size =
r->raw_offsets[result_slots] + WT_BLOCK_COMPRESS_SKIP;
dsk_dst->u.entries = r->raw_entries[result_slots - 1];
/*
* Optionally keep the disk image in cache. Update the initial
* page-header fields to reflect the actual data being written.
*
* If updates are saved and need to be restored, we have to keep
* a copy of the disk image. Unfortunately, we don't yet know if
* there are updates to restore for the key range covered by the
* disk image just created. If there are any saved updates, take
* a copy of the disk image, it's freed later if not needed.
*/
if (F_ISSET(r, WT_EVICT_SCRUB) ||
(F_ISSET(r, WT_EVICT_UPDATE_RESTORE) && r->supd_next > 0)) {
WT_RET(__wt_strndup(session, dsk,
dsk_dst->mem_size, &last->disk_image));
disk_image = last->disk_image;
disk_image->recno = last->recno;
disk_image->mem_size = dsk_dst->mem_size;
disk_image->u.entries = dsk_dst->u.entries;
}
/*
* There is likely a remnant in the working buffer that didn't
* get compressed; copy it down to the start of the buffer and
* update the starting record number, free space and so on.
* !!!
* Note use of memmove, the source and destination buffers can
* overlap.
*/
len = WT_PTRDIFF(
r->first_free, (uint8_t *)dsk + dsk_dst->mem_size);
dsk_start = WT_PAGE_HEADER_BYTE(btree, dsk);
(void)memmove(dsk_start, (uint8_t *)r->first_free - len, len);
r->entries -= r->raw_entries[result_slots - 1];
r->first_free = dsk_start + len;
r->space_avail += r->raw_offsets[result_slots];
WT_ASSERT(session, r->first_free + r->space_avail <=
(uint8_t *)r->disk_image.mem + r->disk_image.memsize);
/*
* Set the key for the next block (before writing the block, a
* key range is needed in that code).
*/
switch (dsk->type) {
case WT_PAGE_COL_INT:
next->recno = r->raw_recnos[result_slots];
break;
case WT_PAGE_COL_VAR:
next->recno = r->raw_recnos[result_slots - 1];
break;
case WT_PAGE_ROW_INT:
case WT_PAGE_ROW_LEAF:
next->recno = WT_RECNO_OOB;
if (!last_block) {
/*
* Confirm there was uncompressed data remaining
* in the buffer, we're about to read it for the
* next chunk's initial key.
*/
WT_ASSERT(session, len > 0);
WT_RET(__rec_split_row_promote_cell(
session, dsk, &next->key));
}
break;
}
write_ref = dst;
} else if (no_more_rows) {
/*
* Compression failed and there are no more rows to accumulate,
* write the original buffer instead.
*/
WT_STAT_FAST_DATA_INCR(session, compress_raw_fail);
dsk->recno = last->recno;
dsk->mem_size = WT_PTRDIFF32(r->first_free, dsk);
dsk->u.entries = r->entries;
r->disk_image.size = dsk->mem_size;
r->entries = 0;
r->first_free = WT_PAGE_HEADER_BYTE(btree, dsk);
r->space_avail = r->page_size - WT_PAGE_HEADER_BYTE_SIZE(btree);
write_ref = &r->disk_image;
last->already_compressed = false;
} else {
/*
* Compression failed, there are more rows to accumulate and the
* compression function wants to try again; increase the size of
* the "page" and try again after we accumulate some more rows.
*/
WT_STAT_FAST_DATA_INCR(session, compress_raw_fail_temporary);
goto split_grow;
}
/* We have a block, update the boundary counter. */
++r->bnd_next;
/*
* If we are writing the whole page in our first/only attempt, it might
* be a checkpoint (checkpoints are only a single page, by definition).
* Further, checkpoints aren't written here, the wrapup functions do the
* write, and they do the write from the original buffer location. If
* it's a checkpoint and the block isn't in the right buffer, copy it.
*
* If it's not a checkpoint, write the block.
*/
if (r->bnd_next == 1 &&
last_block && __rec_is_checkpoint(session, r, last)) {
if (write_ref == dst)
WT_RET(__wt_buf_set(
session, &r->disk_image, dst->mem, dst->size));
} else
WT_RET(
__rec_split_write(session, r, last, write_ref, last_block));
/*
* We got called because there wasn't enough room in the buffer for the
* next key and we might or might not have written a block. In any case,
* make sure the next key fits into the buffer.
*/
if (r->space_avail < next_len) {
split_grow: /*
* Double the page size and make sure we accommodate at least
* one more record. The reason for the latter is that we may
* be here because there's a large key/value pair that won't
* fit in our initial page buffer, even at its expanded size.
*/
r->page_size *= 2;
return (__rec_split_grow(session, r, r->page_size + next_len));
}
return (0);
}
/*
* __rec_split_raw --
* Raw compression split routine.
*/
static inline int
__rec_split_raw(WT_SESSION_IMPL *session, WT_RECONCILE *r, size_t next_len)
{
return (__rec_split_raw_worker(session, r, next_len, false));
}
/*
* __rec_split_finish_std --
* Finish processing a page, standard version.
*/
static int
__rec_split_finish_std(WT_SESSION_IMPL *session, WT_RECONCILE *r)
{
WT_BOUNDARY *bnd;
WT_PAGE_HEADER *dsk;
/* Adjust the boundary information based on our split status. */
switch (r->bnd_state) {
case SPLIT_BOUNDARY:
case SPLIT_MAX:
/*
* We never split, the reconciled page fit into a maximum page
* size. Change the first boundary slot to represent the full
* page (the first boundary slot is largely correct, just update
* the number of entries).
*/
r->bnd_next = 0;
break;
case SPLIT_TRACKING_OFF:
/*
* If we have already split, or aren't tracking boundaries, put
* the remaining data in the next boundary slot.
*/
WT_RET(__rec_split_bnd_grow(session, r));
break;
case SPLIT_TRACKING_RAW:
/*
* We were configured for raw compression, and either we never
* wrote anything, or there's a remaindered block of data.
*/
break;
}
/*
* We may arrive here with no entries to write if the page was entirely
* empty or if nothing on the page was visible to us.
*/
if (r->entries == 0) {
/*
* Pages with skipped or not-yet-globally visible updates aren't
* really empty; otherwise, the page is truly empty and we will
* merge it into its parent during the parent's reconciliation.
*/
if (r->supd_next == 0)
return (0);
/*
* If using the save/restore eviction path, continue with the
* write, the page will be restored after we finish.
*
* If using the lookaside table eviction path, we can't continue
* (we need a page to be written, otherwise we won't ever find
* the updates for future reads).
*/
if (F_ISSET(r, WT_EVICT_LOOKASIDE))
return (EBUSY);
}
/* Set the boundary reference and increment the count. */
bnd = &r->bnd[r->bnd_next++];
bnd->entries = r->entries;
/* Finalize the header information. */
dsk = r->disk_image.mem;
dsk->recno = bnd->recno;
dsk->u.entries = r->entries;
dsk->mem_size = WT_PTRDIFF32(r->first_free, dsk);
r->disk_image.size = dsk->mem_size;
/* If this is a checkpoint, we're done, otherwise write the page. */
return (__rec_is_checkpoint(session, r, bnd) ?
0 : __rec_split_write(session, r, bnd, &r->disk_image, true));
}
/*
* __rec_split_finish --
* Finish processing a page.
*/
static int
__rec_split_finish(WT_SESSION_IMPL *session, WT_RECONCILE *r)
{
WT_BTREE *btree;
size_t data_size;
btree = S2BT(session);
/*
* We're done reconciling, write the final page. Call raw compression
* until/unless there's not enough data to compress.
*/
if (r->raw_compression && r->entries != 0) {
while (r->entries != 0) {
data_size =
WT_PTRDIFF(r->first_free, r->disk_image.mem);
if (data_size <= btree->allocsize)
break;
WT_RET(__rec_split_raw_worker(session, r, 0, true));
}
if (r->entries == 0)
return (0);
}
return (__rec_split_finish_std(session, r));
}
/*
* __rec_split_fixup --
* Fix up after crossing the maximum page boundary.
*/
static int
__rec_split_fixup(WT_SESSION_IMPL *session, WT_RECONCILE *r)
{
WT_BOUNDARY *bnd;
WT_BTREE *btree;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_PAGE_HEADER *dsk;
size_t i, len;
uint8_t *dsk_start, *p;
/*
* When we overflow physical limits of the page, we walk the list of
* split chunks we've created and write those pages out, then update
* the caller's information.
*/
btree = S2BT(session);
/*
* The data isn't laid out on a page boundary or nul padded; copy it to
* a clean, aligned, padded buffer before writing it.
*
* Allocate a scratch buffer to hold the new disk image. Copy the disk
* page's header and block-manager space into the scratch buffer, most
* of the header information remains unchanged between the pages.
*/
WT_RET(__wt_scr_alloc(session, r->disk_image.memsize, &tmp));
dsk = tmp->mem;
memcpy(dsk, r->disk_image.mem, WT_PAGE_HEADER_BYTE_SIZE(btree));
/*
* For each split chunk we've created, update the disk image and copy
* it into place.
*/
dsk_start = WT_PAGE_HEADER_BYTE(btree, dsk);
for (i = 0, bnd = r->bnd; i < r->bnd_next; ++i, ++bnd) {
/* Copy the page contents to the temporary buffer. */
len = (bnd + 1)->offset - bnd->offset;
memcpy(dsk_start,
(uint8_t *)r->disk_image.mem + bnd->offset, len);
/* Finalize the header information and write the page. */
dsk->recno = bnd->recno;
dsk->u.entries = bnd->entries;
tmp->size = WT_PAGE_HEADER_BYTE_SIZE(btree) + len;
dsk->mem_size = WT_STORE_SIZE(tmp->size);
WT_ERR(__rec_split_write(session, r, bnd, tmp, false));
}
/*
* There is probably a remnant in the working buffer that didn't get
* written, copy it down to the beginning of the working buffer.
*
* Confirm the remnant is no larger than a split-sized chunk, including
* header. We know that's the maximum sized remnant because we only have
* remnants if split switches from accumulating to a split boundary to
* accumulating to the end of the page (the other path here is when we
* hit a split boundary, there was room for another split chunk in the
* page, and the next item still wouldn't fit, in which case there is no
* remnant). So: we were accumulating to the end of the page and created
* a remnant. We know the remnant cannot be as large as a split-sized
* chunk, including header, because if there was room for that large a
* remnant, we wouldn't have switched from accumulating to a page end.
*/
p = (uint8_t *)r->disk_image.mem + bnd->offset;
len = WT_PTRDIFF(r->first_free, p);
if (len >= r->split_size - WT_PAGE_HEADER_BYTE_SIZE(btree))
WT_PANIC_ERR(session, EINVAL,
"Reconciliation remnant too large for the split buffer");
dsk = r->disk_image.mem;
dsk_start = WT_PAGE_HEADER_BYTE(btree, dsk);
(void)memmove(dsk_start, p, len);
/*
* Fix up our caller's information, including updating the starting
* record number.
*/
r->entries -= r->total_entries;
r->first_free = dsk_start + len;
WT_ASSERT(session,
r->page_size >= (WT_PAGE_HEADER_BYTE_SIZE(btree) + len));
r->space_avail =
r->split_size - (WT_PAGE_HEADER_BYTE_SIZE(btree) + len);
err: __wt_scr_free(session, &tmp);
return (ret);
}
/*
* __rec_split_write --
* Write a disk block out for the split helper functions.
*/
static int
__rec_split_write(WT_SESSION_IMPL *session,
WT_RECONCILE *r, WT_BOUNDARY *bnd, WT_ITEM *buf, bool last_block)
{
WT_BTREE *btree;
WT_DECL_ITEM(key);
WT_DECL_RET;
WT_MULTI *multi;
WT_PAGE *page;
WT_PAGE_HEADER *dsk;
WT_PAGE_MODIFY *mod;
WT_SAVE_UPD *supd;
size_t addr_size;
uint32_t bnd_slot, i, j;
int cmp;
uint8_t addr[WT_BTREE_MAX_ADDR_COOKIE];
bool need_image;
btree = S2BT(session);
dsk = buf->mem;
page = r->page;
mod = page->modify;
/* Set the zero-length value flag in the page header. */
if (dsk->type == WT_PAGE_ROW_LEAF) {
F_CLR(dsk, WT_PAGE_EMPTY_V_ALL | WT_PAGE_EMPTY_V_NONE);
if (r->entries != 0 && r->all_empty_value)
F_SET(dsk, WT_PAGE_EMPTY_V_ALL);
if (r->entries != 0 && !r->any_empty_value)
F_SET(dsk, WT_PAGE_EMPTY_V_NONE);
}
bnd->entries = r->entries;
/* Initialize the address (set the page type for the parent). */
switch (dsk->type) {
case WT_PAGE_COL_FIX:
bnd->addr.type = WT_ADDR_LEAF_NO;
break;
case WT_PAGE_COL_VAR:
case WT_PAGE_ROW_LEAF:
bnd->addr.type = r->ovfl_items ? WT_ADDR_LEAF : WT_ADDR_LEAF_NO;
break;
case WT_PAGE_COL_INT:
case WT_PAGE_ROW_INT:
bnd->addr.type = WT_ADDR_INT;
break;
WT_ILLEGAL_VALUE(session);
}
bnd->size = (uint32_t)buf->size;
bnd->cksum = 0;
/*
* Check if we've saved updates that belong to this block, and move
* any to the per-block structure. Quit as soon as we find a saved
* update that doesn't belong to the block, they're in sorted order.
*
* This code requires a key be filled in for the next block (or the
* last block flag be set, if there's no next block).
*/
if (page->type == WT_PAGE_ROW_LEAF)
WT_RET(__wt_scr_alloc(session, 0, &key));
for (i = 0, supd = r->supd; i < r->supd_next; ++i, ++supd) {
/* The last block gets all remaining saved updates. */
if (last_block) {
WT_ERR(__rec_update_move(session, bnd, supd));
continue;
}
/*
* Get the saved update's key and compare it with this block's
* key range. If the saved update list belongs with the block
* we're about to write, move it to the per-block memory. Check
* only to the first update that doesn't go with the block, they
* must be in sorted order.
*/
switch (page->type) {
case WT_PAGE_COL_FIX:
case WT_PAGE_COL_VAR:
if (WT_INSERT_RECNO(supd->ins) >= (bnd + 1)->recno)
goto supd_check_complete;
break;
case WT_PAGE_ROW_LEAF:
if (supd->ins == NULL)
WT_ERR(__wt_row_leaf_key(
session, page, supd->rip, key, false));
else {
key->data = WT_INSERT_KEY(supd->ins);
key->size = WT_INSERT_KEY_SIZE(supd->ins);
}
WT_ERR(__wt_compare(session,
btree->collator, key, &(bnd + 1)->key, &cmp));
if (cmp >= 0)
goto supd_check_complete;
break;
WT_ILLEGAL_VALUE_ERR(session);
}
WT_ERR(__rec_update_move(session, bnd, supd));
}
supd_check_complete:
/*
* If there are updates that weren't moved to the block, shuffle them to
* the beginning of the cached list (we maintain the saved updates in
* sorted order, new saved updates must be appended to the list).
*/
for (j = 0; i < r->supd_next; ++j, ++i)
r->supd[j] = r->supd[i];
r->supd_next = j;
/*
* If using the lookaside table eviction path and we found updates that
* weren't globally visible when reconciling this page, note that in the
* page header.
*/
if (F_ISSET(r, WT_EVICT_LOOKASIDE) && bnd->supd != NULL) {
F_SET(dsk, WT_PAGE_LAS_UPDATE);
r->cache_write_lookaside = true;
}
/*
* If configured for an in-memory database, or using the save/restore
* eviction path and we had to skip updates in order to build this disk
* image, we can't actually write it. Instead, we will re-instantiate
* the page using the disk image and any list of updates we skipped.
*/
if (F_ISSET(r, WT_EVICT_IN_MEMORY))
goto copy_image;
if (F_ISSET(r, WT_EVICT_UPDATE_RESTORE) && bnd->supd != NULL) {
r->cache_write_restore = true;
goto copy_image;
}
/*
* If we wrote this block before, re-use it. Pages get written in the
* same block order every time, only check the appropriate slot. The
* expensive part of this test is the checksum, only do that work when
* there has been or will be a reconciliation of this page involving
* split pages. This test isn't perfect: we're doing a checksum if a
* previous reconciliation of the page split or if we will split this
* time, but that test won't calculate a checksum on the first block
* the first time the page splits.
*/
bnd_slot = (uint32_t)(bnd - r->bnd);
if (bnd_slot > 1 ||
(mod->rec_result == WT_PM_REC_MULTIBLOCK &&
mod->mod_multi != NULL)) {
/*
* There are page header fields which need to be cleared to get
* consistent checksums: specifically, the write generation and
* the memory owned by the block manager. We are reusing the
* same buffer space each time, clear it before calculating the
* checksum.
*/
dsk->write_gen = 0;
memset(WT_BLOCK_HEADER_REF(dsk), 0, btree->block_header);
bnd->cksum = __wt_cksum(buf->data, buf->size);
/*
* One last check: don't reuse blocks if compacting, the reason
* for compaction is to move blocks to different locations. We
* do this check after calculating the checksums, hopefully the
* next write can be skipped.
*/
if (session->compact_state == WT_COMPACT_NONE &&
mod->rec_result == WT_PM_REC_MULTIBLOCK &&
mod->mod_multi_entries > bnd_slot) {
multi = &mod->mod_multi[bnd_slot];
if (multi->size == bnd->size &&
multi->cksum == bnd->cksum) {
multi->addr.reuse = 1;
bnd->addr = multi->addr;
WT_STAT_FAST_DATA_INCR(session, rec_page_match);
goto copy_image;
}
}
}
#ifdef HAVE_VERBOSE
/* Output a verbose message if we create a page without many entries */
if (WT_VERBOSE_ISSET(session, WT_VERB_SPLIT) && r->entries < 6)
WT_ERR(__wt_verbose(session, WT_VERB_SPLIT,
"Reconciliation creating a page with %" PRIu32
" entries, memory footprint %" PRIu64
", page count %" PRIu32 ", %s, split state: %d\n",
r->entries, r->page->memory_footprint, r->bnd_next,
F_ISSET(r, WT_EVICTING) ? "evict" : "checkpoint",
r->bnd_state));
#endif
WT_ERR(__wt_bt_write(session, buf, addr, &addr_size,
false, F_ISSET(r, WT_CHECKPOINTING), bnd->already_compressed));
WT_ERR(__wt_strndup(session, addr, addr_size, &bnd->addr.addr));
bnd->addr.size = (uint8_t)addr_size;
/*
* If using the lookaside table eviction path and we found updates that
* weren't globally visible when reconciling this page, copy them into
* the database's lookaside store.
*/
if (F_ISSET(r, WT_EVICT_LOOKASIDE) && bnd->supd != NULL)
WT_ERR(__rec_update_las(session, r, btree->id, bnd));
copy_image:
/*
* If re-instantiating this page in memory (either because eviction
* wants to, or because we skipped updates to build the disk image),
* save a copy of the disk image.
*
* Raw compression might have already saved a copy of the disk image
* before we could know if we skipped updates to create it, and now
* we know if we're going to need it.
*
* Copy the disk image if we need a copy and don't already have one,
* discard any already saved copy we don't need.
*/
need_image = F_ISSET(r, WT_EVICT_SCRUB) ||
(F_ISSET(r, WT_EVICT_UPDATE_RESTORE) && bnd->supd != NULL);
if (need_image && bnd->disk_image == NULL)
WT_ERR(__wt_strndup(
session, buf->data, buf->size, &bnd->disk_image));
if (!need_image)
__wt_free(session, bnd->disk_image);
err: __wt_scr_free(session, &key);
return (ret);
}
/*
* __rec_update_las --
* Copy a set of updates into the database's lookaside buffer.
*/
static int
__rec_update_las(WT_SESSION_IMPL *session,
WT_RECONCILE *r, uint32_t btree_id, WT_BOUNDARY *bnd)
{
WT_CURSOR *cursor;
WT_DECL_ITEM(key);
WT_DECL_RET;
WT_ITEM las_addr, las_value;
WT_PAGE *page;
WT_SAVE_UPD *list;
WT_UPDATE *upd;
uint64_t las_counter;
int64_t insert_cnt;
uint32_t i, session_flags, slot;
uint8_t *p;
cursor = NULL;
WT_CLEAR(las_addr);
WT_CLEAR(las_value);
page = r->page;
insert_cnt = 0;
/*
* We're writing lookaside records: start instantiating them on pages
* we read (with the right flag set), and start sweeping the file.
*/
__wt_las_set_written(session);
WT_ERR(__wt_las_cursor(session, &cursor, &session_flags));
/* Ensure enough room for a column-store key without checking. */
WT_ERR(__wt_scr_alloc(session, WT_INTPACK64_MAXSIZE, &key));
/*
* Each key in the lookaside table is associated with a block, and those
* blocks are freed and reallocated to other pages as pages in the tree
* are modified and reconciled. We want to be sure we don't add records
* to the lookaside table, then discard the block to which they apply,
* then write a new block to the same address, and then apply the old
* records to the new block when it's read. We don't want to clean old
* records out of the lookaside table every time we free a block because
* that happens a lot and would be costly; instead, we clean out the old
* records when adding new records into the lookaside table. This works
* because we only read from the lookaside table for pages marked with
* the WT_PAGE_LAS_UPDATE flag: that flag won't be set if we rewrite a
* block with no lookaside records, so the lookaside table won't be
* checked when the block is read, even if there are lookaside table
* records matching that block. If we rewrite a block that has lookaside
* records, we'll run this code, discarding any old records that might
* exist.
*/
WT_ERR(__wt_las_remove_block(
session, cursor, btree_id, bnd->addr.addr, bnd->addr.size));
/* Lookaside table key component: block address. */
las_addr.data = bnd->addr.addr;
las_addr.size = bnd->addr.size;
/* Enter each update in the boundary's list into the lookaside store. */
for (las_counter = 0, i = 0,
list = bnd->supd; i < bnd->supd_next; ++i, ++list) {
/* Lookaside table key component: source key. */
switch (page->type) {
case WT_PAGE_COL_FIX:
case WT_PAGE_COL_VAR:
p = key->mem;
WT_ERR(
__wt_vpack_uint(&p, 0, WT_INSERT_RECNO(list->ins)));
key->size = WT_PTRDIFF(p, key->data);
break;
case WT_PAGE_ROW_LEAF:
if (list->ins == NULL)
WT_ERR(__wt_row_leaf_key(
session, page, list->rip, key, false));
else {
key->data = WT_INSERT_KEY(list->ins);
key->size = WT_INSERT_KEY_SIZE(list->ins);
}
break;
WT_ILLEGAL_VALUE_ERR(session);
}
/* Lookaside table value component: update reference. */
switch (page->type) {
case WT_PAGE_COL_FIX:
case WT_PAGE_COL_VAR:
upd = list->ins->upd;
break;
case WT_PAGE_ROW_LEAF:
if (list->ins == NULL) {
slot = WT_ROW_SLOT(page, list->rip);
upd = page->modify->mod_row_update[slot];
} else
upd = list->ins->upd;
break;
WT_ILLEGAL_VALUE_ERR(session);
}
/*
* Walk the list of updates, storing each key/value pair into
* the lookaside table.
*/
do {
cursor->set_key(cursor, btree_id,
&las_addr, ++las_counter, list->onpage_txn, key);
if (WT_UPDATE_DELETED_ISSET(upd))
las_value.size = 0;
else {
las_value.data = WT_UPDATE_DATA(upd);
las_value.size = upd->size;
}
cursor->set_value(
cursor, upd->txnid, upd->size, &las_value);
WT_ERR(cursor->insert(cursor));
++insert_cnt;
} while ((upd = upd->next) != NULL);
}
err: WT_TRET(__wt_las_cursor_close(session, &cursor, session_flags));
if (insert_cnt > 0)
(void)__wt_atomic_addi64(
&S2C(session)->las_record_cnt, insert_cnt);
__wt_scr_free(session, &key);
return (ret);
}
/*
* __wt_bulk_init --
* Bulk insert initialization.
*/
int
__wt_bulk_init(WT_SESSION_IMPL *session, WT_CURSOR_BULK *cbulk)
{
WT_BTREE *btree;
WT_PAGE_INDEX *pindex;
WT_RECONCILE *r;
uint64_t recno;
btree = S2BT(session);
/*
* Bulk-load is only permitted on newly created files, not any empty
* file -- see the checkpoint code for a discussion.
*/
if (!btree->bulk_load_ok)
WT_RET_MSG(session, EINVAL,
"bulk-load is only possible for newly created trees");
/*
* Get a reference to the empty leaf page; we have exclusive access so
* we can take a copy of the page, confident the parent won't split.
*/
pindex = WT_INTL_INDEX_GET_SAFE(btree->root.page);
cbulk->ref = pindex->index[0];
cbulk->leaf = cbulk->ref->page;
WT_RET(
__rec_write_init(session, cbulk->ref, 0, NULL, &cbulk->reconcile));
r = cbulk->reconcile;
r->is_bulk_load = true;
recno = WT_RECNO_OOB; /* -Werror=maybe-uninitialized */
switch (btree->type) {
case BTREE_COL_FIX:
case BTREE_COL_VAR:
recno = 1;
break;
case BTREE_ROW:
recno = WT_RECNO_OOB;
break;
}
return (__rec_split_init(
session, r, cbulk->leaf, recno, btree->maxleafpage));
}
/*
* __wt_bulk_wrapup --
* Bulk insert cleanup.
*/
int
__wt_bulk_wrapup(WT_SESSION_IMPL *session, WT_CURSOR_BULK *cbulk)
{
WT_BTREE *btree;
WT_PAGE *parent;
WT_RECONCILE *r;
btree = S2BT(session);
if ((r = cbulk->reconcile) == NULL)
return (0);
switch (btree->type) {
case BTREE_COL_FIX:
if (cbulk->entry != 0)
__rec_incr(session, r, cbulk->entry,
__bitstr_size(
(size_t)cbulk->entry * btree->bitcnt));
break;
case BTREE_COL_VAR:
if (cbulk->rle != 0)
WT_RET(__wt_bulk_insert_var(session, cbulk, false));
break;
case BTREE_ROW:
break;
}
WT_RET(__rec_split_finish(session, r));
WT_RET(__rec_write_wrapup(session, r, r->page));
WT_RET(__rec_write_status(session, r, r->page));
/* Mark the page's parent and the tree dirty. */
parent = r->ref->home;
WT_RET(__wt_page_modify_init(session, parent));
__wt_page_modify_set(session, parent);
__rec_destroy(session, &cbulk->reconcile);
return (0);
}
/*
* __wt_bulk_insert_row --
* Row-store bulk insert.
*/
int
__wt_bulk_insert_row(WT_SESSION_IMPL *session, WT_CURSOR_BULK *cbulk)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_KV *key, *val;
WT_RECONCILE *r;
bool ovfl_key;
r = cbulk->reconcile;
btree = S2BT(session);
cursor = &cbulk->cbt.iface;
key = &r->k;
val = &r->v;
WT_RET(__rec_cell_build_leaf_key(session, r, /* Build key cell */
cursor->key.data, cursor->key.size, &ovfl_key));
WT_RET(__rec_cell_build_val(session, r, /* Build value cell */
cursor->value.data, cursor->value.size, (uint64_t)0));
/* Boundary: split or write the page. */
if (key->len + val->len > r->space_avail) {
if (r->raw_compression)
WT_RET(
__rec_split_raw(session, r, key->len + val->len));
else {
/*
* Turn off prefix compression until a full key written
* to the new page, and (unless already working with an
* overflow key), rebuild the key without compression.
*/
if (r->key_pfx_compress_conf) {
r->key_pfx_compress = false;
if (!ovfl_key)
WT_RET(__rec_cell_build_leaf_key(
session, r, NULL, 0, &ovfl_key));
}
WT_RET(__rec_split(session, r, key->len + val->len));
}
}
/* Copy the key/value pair onto the page. */
__rec_copy_incr(session, r, key);
if (val->len == 0)
r->any_empty_value = true;
else {
r->all_empty_value = false;
if (btree->dictionary)
WT_RET(__rec_dict_replace(session, r, 0, val));
__rec_copy_incr(session, r, val);
}
/* Update compression state. */
__rec_key_state_update(r, ovfl_key);
return (0);
}
/*
* __rec_col_fix_bulk_insert_split_check --
* Check if a bulk-loaded fixed-length column store page needs to split.
*/
static inline int
__rec_col_fix_bulk_insert_split_check(WT_CURSOR_BULK *cbulk)
{
WT_BTREE *btree;
WT_RECONCILE *r;
WT_SESSION_IMPL *session;
session = (WT_SESSION_IMPL *)cbulk->cbt.iface.session;
r = cbulk->reconcile;
btree = S2BT(session);
if (cbulk->entry == cbulk->nrecs) {
if (cbulk->entry != 0) {
/*
* If everything didn't fit, update the counters and
* split.
*
* Boundary: split or write the page.
*/
__rec_incr(session, r, cbulk->entry,
__bitstr_size(
(size_t)cbulk->entry * btree->bitcnt));
WT_RET(__rec_split(session, r, 0));
}
cbulk->entry = 0;
cbulk->nrecs = WT_FIX_BYTES_TO_ENTRIES(btree, r->space_avail);
}
return (0);
}
/*
* __wt_bulk_insert_fix --
* Fixed-length column-store bulk insert.
*/
int
__wt_bulk_insert_fix(
WT_SESSION_IMPL *session, WT_CURSOR_BULK *cbulk, bool deleted)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_RECONCILE *r;
r = cbulk->reconcile;
btree = S2BT(session);
cursor = &cbulk->cbt.iface;
WT_RET(__rec_col_fix_bulk_insert_split_check(cbulk));
__bit_setv(r->first_free, cbulk->entry,
btree->bitcnt, deleted ? 0 : ((uint8_t *)cursor->value.data)[0]);
++cbulk->entry;
++r->recno;
return (0);
}
/*
* __wt_bulk_insert_fix_bitmap --
* Fixed-length column-store bulk insert.
*/
int
__wt_bulk_insert_fix_bitmap(WT_SESSION_IMPL *session, WT_CURSOR_BULK *cbulk)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_RECONCILE *r;
uint32_t entries, offset, page_entries, page_size;
const uint8_t *data;
r = cbulk->reconcile;
btree = S2BT(session);
cursor = &cbulk->cbt.iface;
if (((r->recno - 1) * btree->bitcnt) & 0x7)
WT_RET_MSG(session, EINVAL,
"Bulk bitmap load not aligned on a byte boundary");
for (data = cursor->value.data,
entries = (uint32_t)cursor->value.size;
entries > 0;
entries -= page_entries, data += page_size) {
WT_RET(__rec_col_fix_bulk_insert_split_check(cbulk));
page_entries = WT_MIN(entries, cbulk->nrecs - cbulk->entry);
page_size = __bitstr_size(page_entries * btree->bitcnt);
offset = __bitstr_size(cbulk->entry * btree->bitcnt);
memcpy(r->first_free + offset, data, page_size);
cbulk->entry += page_entries;
r->recno += page_entries;
}
return (0);
}
/*
* __wt_bulk_insert_var --
* Variable-length column-store bulk insert.
*/
int
__wt_bulk_insert_var(
WT_SESSION_IMPL *session, WT_CURSOR_BULK *cbulk, bool deleted)
{
WT_BTREE *btree;
WT_KV *val;
WT_RECONCILE *r;
r = cbulk->reconcile;
btree = S2BT(session);
val = &r->v;
if (deleted) {
val->cell_len = __wt_cell_pack_del(&val->cell, cbulk->rle);
val->buf.data = NULL;
val->buf.size = 0;
val->len = val->cell_len;
} else
/*
* Store the bulk cursor's last buffer, not the current value,
* we're tracking duplicates, which means we want the previous
* value seen, not the current value.
*/
WT_RET(__rec_cell_build_val(session,
r, cbulk->last.data, cbulk->last.size, cbulk->rle));
/* Boundary: split or write the page. */
if (val->len > r->space_avail)
WT_RET(r->raw_compression ?
__rec_split_raw(session, r, val->len) :
__rec_split(session, r, val->len));
/* Copy the value onto the page. */
if (btree->dictionary)
WT_RET(__rec_dict_replace(session, r, cbulk->rle, val));
__rec_copy_incr(session, r, val);
/* Update the starting record number in case we split. */
r->recno += cbulk->rle;
return (0);
}
/*
* __rec_vtype --
* Return a value cell's address type.
*/
static inline u_int
__rec_vtype(WT_ADDR *addr)
{
if (addr->type == WT_ADDR_INT)
return (WT_CELL_ADDR_INT);
if (addr->type == WT_ADDR_LEAF)
return (WT_CELL_ADDR_LEAF);
return (WT_CELL_ADDR_LEAF_NO);
}
/*
* __rec_col_int --
* Reconcile a column-store internal page.
*/
static int
__rec_col_int(WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_REF *pageref)
{
WT_ADDR *addr;
WT_BTREE *btree;
WT_CELL_UNPACK *vpack, _vpack;
WT_CHILD_STATE state;
WT_DECL_RET;
WT_KV *val;
WT_PAGE *child, *page;
WT_REF *ref;
bool hazard;
btree = S2BT(session);
page = pageref->page;
child = NULL;
hazard = false;
val = &r->v;
vpack = &_vpack;
WT_RET(__rec_split_init(
session, r, page, pageref->ref_recno, btree->maxintlpage));
/* For each entry in the in-memory page... */
WT_INTL_FOREACH_BEGIN(session, page, ref) {
/* Update the starting record number in case we split. */
r->recno = ref->ref_recno;
/*
* Modified child.
* The page may be emptied or internally created during a split.
* Deleted/split pages are merged into the parent and discarded.
*/
WT_ERR(__rec_child_modify(session, r, ref, &hazard, &state));
addr = NULL;
child = ref->page;
switch (state) {
case WT_CHILD_IGNORE:
/* Ignored child. */
WT_CHILD_RELEASE_ERR(session, hazard, ref);
continue;
case WT_CHILD_MODIFIED:
/*
* Modified child. Empty pages are merged into the
* parent and discarded.
*/
switch (child->modify->rec_result) {
case WT_PM_REC_EMPTY:
/*
* Column-store pages are almost never empty, as
* discarding a page would remove a chunk of the
* name space. The exceptions are pages created
* when the tree is created, and never filled.
*/
WT_CHILD_RELEASE_ERR(session, hazard, ref);
continue;
case WT_PM_REC_MULTIBLOCK:
WT_ERR(__rec_col_merge(session, r, child));
WT_CHILD_RELEASE_ERR(session, hazard, ref);
continue;
case WT_PM_REC_REPLACE:
addr = &child->modify->mod_replace;
break;
WT_ILLEGAL_VALUE_ERR(session);
}
break;
case WT_CHILD_ORIGINAL:
/* Original child. */
break;
case WT_CHILD_PROXY:
/*
* Deleted child where we write a proxy cell, not
* yet supported for column-store.
*/
ret = __wt_illegal_value(session, NULL);
goto err;
}
/*
* Build the value cell. The child page address is in one of 3
* places: if the page was replaced, the page's modify structure
* references it and we built the value cell just above in the
* switch statement. Else, the WT_REF->addr reference points to
* an on-page cell or an off-page WT_ADDR structure: if it's an
* on-page cell and we copy it from the page, else build a new
* cell.
*/
if (addr == NULL && __wt_off_page(page, ref->addr))
addr = ref->addr;
if (addr == NULL) {
__wt_cell_unpack(ref->addr, vpack);
val->buf.data = ref->addr;
val->buf.size = __wt_cell_total_len(vpack);
val->cell_len = 0;
val->len = val->buf.size;
} else
__rec_cell_build_addr(session, r,
addr->addr, addr->size,
__rec_vtype(addr), ref->ref_recno);
WT_CHILD_RELEASE_ERR(session, hazard, ref);
/* Boundary: split or write the page. */
if (val->len > r->space_avail)
WT_ERR(r->raw_compression ?
__rec_split_raw(session, r, val->len) :
__rec_split(session, r, val->len));
/* Copy the value onto the page. */
__rec_copy_incr(session, r, val);
} WT_INTL_FOREACH_END;
/* Write the remnant page. */
return (__rec_split_finish(session, r));
err: WT_CHILD_RELEASE(session, hazard, ref);
return (ret);
}
/*
* __rec_col_merge --
* Merge in a split page.
*/
static int
__rec_col_merge(WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_PAGE *page)
{
WT_ADDR *addr;
WT_KV *val;
WT_MULTI *multi;
WT_PAGE_MODIFY *mod;
uint32_t i;
mod = page->modify;
val = &r->v;
/* For each entry in the split array... */
for (multi = mod->mod_multi,
i = 0; i < mod->mod_multi_entries; ++multi, ++i) {
/* Update the starting record number in case we split. */
r->recno = multi->key.recno;
/* Build the value cell. */
addr = &multi->addr;
__rec_cell_build_addr(session, r,
addr->addr, addr->size, __rec_vtype(addr), r->recno);
/* Boundary: split or write the page. */
if (val->len > r->space_avail)
WT_RET(r->raw_compression ?
__rec_split_raw(session, r, val->len) :
__rec_split(session, r, val->len));
/* Copy the value onto the page. */
__rec_copy_incr(session, r, val);
}
return (0);
}
/*
* __rec_col_fix --
* Reconcile a fixed-width, column-store leaf page.
*/
static int
__rec_col_fix(WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_REF *pageref)
{
WT_BTREE *btree;
WT_INSERT *ins;
WT_PAGE *page;
WT_UPDATE *upd;
uint64_t recno;
uint32_t entry, nrecs;
btree = S2BT(session);
page = pageref->page;
WT_RET(__rec_split_init(
session, r, page, pageref->ref_recno, btree->maxleafpage));
/* Copy the original, disk-image bytes into place. */
memcpy(r->first_free, page->pg_fix_bitf,
__bitstr_size((size_t)page->pg_fix_entries * btree->bitcnt));
/* Update any changes to the original on-page data items. */
WT_SKIP_FOREACH(ins, WT_COL_UPDATE_SINGLE(page)) {
WT_RET(__rec_txn_read(session, r, ins, NULL, NULL, &upd));
if (upd != NULL)
__bit_setv(r->first_free,
WT_INSERT_RECNO(ins) - pageref->ref_recno,
btree->bitcnt, *(uint8_t *)WT_UPDATE_DATA(upd));
}
/* Calculate the number of entries per page remainder. */
entry = page->pg_fix_entries;
nrecs = WT_FIX_BYTES_TO_ENTRIES(
btree, r->space_avail) - page->pg_fix_entries;
r->recno += entry;
/* Walk any append list. */
for (ins =
WT_SKIP_FIRST(WT_COL_APPEND(page));; ins = WT_SKIP_NEXT(ins)) {
if (ins == NULL) {
/*
* If the page split, instantiate any missing records in
* the page's name space. (Imagine record 98 is
* transactionally visible, 99 wasn't created or is not
* yet visible, 100 is visible. Then the page splits and
* record 100 moves to another page. When we reconcile
* the original page, we write record 98, then we don't
* see record 99 for whatever reason. If we've moved
* record 100, we don't know to write a deleted record
* 99 on the page.)
*
* The record number recorded during the split is the
* first key on the split page, that is, one larger than
* the last key on this page, we have to decrement it.
*/
if ((recno =
page->modify->mod_col_split_recno) == WT_RECNO_OOB)
break;
recno -= 1;
/*
* The following loop assumes records to write, and the
* previous key might have been visible.
*/
if (r->recno > recno)
break;
upd = NULL;
} else {
WT_RET(
__rec_txn_read(session, r, ins, NULL, NULL, &upd));
recno = WT_INSERT_RECNO(ins);
}
for (;;) {
/*
* The application may have inserted records which left
* gaps in the name space.
*/
for (;
nrecs > 0 && r->recno < recno;
--nrecs, ++entry, ++r->recno)
__bit_setv(
r->first_free, entry, btree->bitcnt, 0);
if (nrecs > 0) {
__bit_setv(r->first_free, entry, btree->bitcnt,
upd == NULL ? 0 :
*(uint8_t *)WT_UPDATE_DATA(upd));
--nrecs;
++entry;
++r->recno;
break;
}
/*
* If everything didn't fit, update the counters and
* split.
*
* Boundary: split or write the page.
*/
__rec_incr(session, r, entry,
__bitstr_size((size_t)entry * btree->bitcnt));
WT_RET(__rec_split(session, r, 0));
/* Calculate the number of entries per page. */
entry = 0;
nrecs = WT_FIX_BYTES_TO_ENTRIES(btree, r->space_avail);
}
/*
* Execute this loop once without an insert item to catch any
* missing records due to a split, then quit.
*/
if (ins == NULL)
break;
}
/* Update the counters. */
__rec_incr(
session, r, entry, __bitstr_size((size_t)entry * btree->bitcnt));
/* Write the remnant page. */
return (__rec_split_finish(session, r));
}
/*
* __rec_col_fix_slvg --
* Reconcile a fixed-width, column-store leaf page created during salvage.
*/
static int
__rec_col_fix_slvg(WT_SESSION_IMPL *session,
WT_RECONCILE *r, WT_REF *pageref, WT_SALVAGE_COOKIE *salvage)
{
WT_BTREE *btree;
WT_PAGE *page;
uint64_t page_start, page_take;
uint32_t entry, nrecs;
btree = S2BT(session);
page = pageref->page;
/*
* !!!
* It's vanishingly unlikely and probably impossible for fixed-length
* column-store files to have overlapping key ranges. It's possible
* for an entire key range to go missing (if a page is corrupted and
* lost), but because pages can't split, it shouldn't be possible to
* find pages where the key ranges overlap. That said, we check for
* it during salvage and clean up after it here because it doesn't
* cost much and future column-store formats or operations might allow
* for fixed-length format ranges to overlap during salvage, and I
* don't want to have to retrofit the code later.
*/
WT_RET(__rec_split_init(
session, r, page, pageref->ref_recno, btree->maxleafpage));
/* We may not be taking all of the entries on the original page. */
page_take = salvage->take == 0 ? page->pg_fix_entries : salvage->take;
page_start = salvage->skip == 0 ? 0 : salvage->skip;
/* Calculate the number of entries per page. */
entry = 0;
nrecs = WT_FIX_BYTES_TO_ENTRIES(btree, r->space_avail);
for (; nrecs > 0 && salvage->missing > 0;
--nrecs, --salvage->missing, ++entry)
__bit_setv(r->first_free, entry, btree->bitcnt, 0);
for (; nrecs > 0 && page_take > 0;
--nrecs, --page_take, ++page_start, ++entry)
__bit_setv(r->first_free, entry, btree->bitcnt,
__bit_getv(page->pg_fix_bitf,
(uint32_t)page_start, btree->bitcnt));
r->recno += entry;
__rec_incr(session, r, entry,
__bitstr_size((size_t)entry * btree->bitcnt));
/*
* We can't split during salvage -- if everything didn't fit, it's
* all gone wrong.
*/
if (salvage->missing != 0 || page_take != 0)
WT_PANIC_RET(session, WT_PANIC,
"%s page too large, attempted split during salvage",
__wt_page_type_string(page->type));
/* Write the page. */
return (__rec_split_finish(session, r));
}
/*
* __rec_col_var_helper --
* Create a column-store variable length record cell and write it onto a
* page.
*/
static int
__rec_col_var_helper(WT_SESSION_IMPL *session, WT_RECONCILE *r,
WT_SALVAGE_COOKIE *salvage,
WT_ITEM *value, bool deleted, uint8_t overflow_type, uint64_t rle)
{
WT_BTREE *btree;
WT_KV *val;
btree = S2BT(session);
val = &r->v;
/*
* Occasionally, salvage needs to discard records from the beginning or
* end of the page, and because the items may be part of a RLE cell, do
* the adjustments here. It's not a mistake we don't bother telling
* our caller we've handled all the records from the page we care about,
* and can quit processing the page: salvage is a rare operation and I
* don't want to complicate our caller's loop.
*/
if (salvage != NULL) {
if (salvage->done)
return (0);
if (salvage->skip != 0) {
if (rle <= salvage->skip) {
salvage->skip -= rle;
return (0);
}
rle -= salvage->skip;
salvage->skip = 0;
}
if (salvage->take != 0) {
if (rle <= salvage->take)
salvage->take -= rle;
else {
rle = salvage->take;
salvage->take = 0;
}
if (salvage->take == 0)
salvage->done = true;
}
}
if (deleted) {
val->cell_len = __wt_cell_pack_del(&val->cell, rle);
val->buf.data = NULL;
val->buf.size = 0;
val->len = val->cell_len;
} else if (overflow_type) {
val->cell_len = __wt_cell_pack_ovfl(
&val->cell, overflow_type, rle, value->size);
val->buf.data = value->data;
val->buf.size = value->size;
val->len = val->cell_len + value->size;
} else
WT_RET(__rec_cell_build_val(
session, r, value->data, value->size, rle));
/* Boundary: split or write the page. */
if (val->len > r->space_avail)
WT_RET(r->raw_compression ?
__rec_split_raw(session, r, val->len) :
__rec_split(session, r, val->len));
/* Copy the value onto the page. */
if (!deleted && !overflow_type && btree->dictionary)
WT_RET(__rec_dict_replace(session, r, rle, val));
__rec_copy_incr(session, r, val);
/* Update the starting record number in case we split. */
r->recno += rle;
return (0);
}
/*
* __rec_col_var --
* Reconcile a variable-width column-store leaf page.
*/
static int
__rec_col_var(WT_SESSION_IMPL *session,
WT_RECONCILE *r, WT_REF *pageref, WT_SALVAGE_COOKIE *salvage)
{
enum { OVFL_IGNORE, OVFL_UNUSED, OVFL_USED } ovfl_state;
WT_BTREE *btree;
WT_CELL *cell;
WT_CELL_UNPACK *vpack, _vpack;
WT_COL *cip;
WT_DECL_ITEM(orig);
WT_DECL_RET;
WT_INSERT *ins;
WT_ITEM *last;
WT_PAGE *page;
WT_UPDATE *upd;
uint64_t n, nrepeat, repeat_count, rle, skip, src_recno;
uint32_t i, size;
bool deleted, last_deleted, orig_deleted, update_no_copy;
const void *data;
btree = S2BT(session);
page = pageref->page;
last = r->last;
vpack = &_vpack;
WT_RET(__rec_split_init(
session, r, page, pageref->ref_recno, btree->maxleafpage));
WT_RET(__wt_scr_alloc(session, 0, &orig));
data = NULL;
size = 0;
upd = NULL;
/*
* The salvage code may be calling us to reconcile a page where there
* were missing records in the column-store name space. If taking the
* first record from on the page, it might be a deleted record, so we
* have to give the RLE code a chance to figure that out. Else, if
* not taking the first record from the page, write a single element
* representing the missing records onto a new page. (Don't pass the
* salvage cookie to our helper function in this case, we're handling
* one of the salvage cookie fields on our own, and we don't need the
* helper function's assistance.)
*/
rle = 0;
last_deleted = false;
if (salvage != NULL && salvage->missing != 0) {
if (salvage->skip == 0) {
rle = salvage->missing;
last_deleted = true;
/*
* Correct the number of records we're going to "take",
* pretending the missing records were on the page.
*/
salvage->take += salvage->missing;
} else
WT_ERR(__rec_col_var_helper(session,
r, NULL, NULL, true, false, salvage->missing));
}
/*
* We track two data items through this loop: the previous (last) item
* and the current item: if the last item is the same as the current
* item, we increment the RLE count for the last item; if the last item
* is different from the current item, we write the last item onto the
* page, and replace it with the current item. The r->recno counter
* tracks records written to the page, and is incremented by the helper
* function immediately after writing records to the page. The record
* number of our source record, that is, the current item, is maintained
* in src_recno.
*/
src_recno = r->recno + rle;
/* For each entry in the in-memory page... */
WT_COL_FOREACH(page, cip, i) {
ovfl_state = OVFL_IGNORE;
if ((cell = WT_COL_PTR(page, cip)) == NULL) {
nrepeat = 1;
ins = NULL;
orig_deleted = true;
} else {
__wt_cell_unpack(cell, vpack);
nrepeat = __wt_cell_rle(vpack);
ins = WT_SKIP_FIRST(WT_COL_UPDATE(page, cip));
/*
* If the original value is "deleted", there's no value
* to compare, we're done.
*/
orig_deleted = vpack->type == WT_CELL_DEL;
if (orig_deleted)
goto record_loop;
/*
* Overflow items are tricky: we don't know until we're
* finished processing the set of values if we need the
* overflow value or not. If we don't use the overflow
* item at all, we have to discard it from the backing
* file, otherwise we'll leak blocks on the checkpoint.
* That's safe because if the backing overflow value is
* still needed by any running transaction, we'll cache
* a copy in the reconciliation tracking structures.
*
* Regardless, we avoid copying in overflow records: if
* there's a WT_INSERT entry that modifies a reference
* counted overflow record, we may have to write copies
* of the overflow record, and in that case we'll do the
* comparisons, but we don't read overflow items just to
* see if they match records on either side.
*/
if (vpack->ovfl) {
ovfl_state = OVFL_UNUSED;
goto record_loop;
}
/*
* If data is Huffman encoded, we have to decode it in
* order to compare it with the last item we saw, which
* may have been an update string. This guarantees we
* find every single pair of objects we can RLE encode,
* including applications updating an existing record
* where the new value happens (?) to match a Huffman-
* encoded value in a previous or next record.
*/
WT_ERR(__wt_dsk_cell_data_ref(
session, WT_PAGE_COL_VAR, vpack, orig));
}
record_loop: /*
* Generate on-page entries: loop repeat records, looking for
* WT_INSERT entries matching the record number. The WT_INSERT
* lists are in sorted order, so only need check the next one.
*/
for (n = 0;
n < nrepeat; n += repeat_count, src_recno += repeat_count) {
upd = NULL;
if (ins != NULL && WT_INSERT_RECNO(ins) == src_recno) {
WT_ERR(__rec_txn_read(
session, r, ins, NULL, vpack, &upd));
ins = WT_SKIP_NEXT(ins);
}
if (upd != NULL) {
update_no_copy = true; /* No data copy */
repeat_count = 1; /* Single record */
deleted = WT_UPDATE_DELETED_ISSET(upd);
if (!deleted) {
data = WT_UPDATE_DATA(upd);
size = upd->size;
}
} else if (vpack->raw == WT_CELL_VALUE_OVFL_RM) {
update_no_copy = true; /* No data copy */
repeat_count = 1; /* Single record */
deleted = false;
/*
* If doing update save and restore, there's an
* update that's not globally visible, and the
* underlying value is a removed overflow value,
* we end up here.
*
* When the update save/restore code noticed the
* removed overflow value, it appended a copy of
* the cached, original overflow value to the
* update list being saved (ensuring the on-page
* item will never be accessed after the page is
* re-instantiated), then returned a NULL update
* to us.
*
* Assert the case: if we remove an underlying
* overflow object, checkpoint reconciliation
* should never see it again, there should be a
* visible update in the way.
*
* Write a placeholder.
*/
WT_ASSERT(session,
F_ISSET(r, WT_EVICT_UPDATE_RESTORE));
data = "@";
size = 1;
} else {
update_no_copy = false; /* Maybe data copy */
/*
* The repeat count is the number of records up
* to the next WT_INSERT record, or up to the
* end of the entry if we have no more WT_INSERT
* records.
*/
if (ins == NULL)
repeat_count = nrepeat - n;
else
repeat_count =
WT_INSERT_RECNO(ins) - src_recno;
deleted = orig_deleted;
if (deleted)
goto compare;
/*
* If we are handling overflow items, use the
* overflow item itself exactly once, after
* which we have to copy it into a buffer and
* from then on use a complete copy because we
* are re-creating a new overflow record each
* time.
*/
switch (ovfl_state) {
case OVFL_UNUSED:
/*
* An as-yet-unused overflow item.
*
* We're going to copy the on-page cell,
* write out any record we're tracking.
*/
if (rle != 0) {
WT_ERR(__rec_col_var_helper(
session, r, salvage, last,
last_deleted, 0, rle));
rle = 0;
}
last->data = vpack->data;
last->size = vpack->size;
WT_ERR(__rec_col_var_helper(
session, r, salvage, last, false,
WT_CELL_VALUE_OVFL, repeat_count));
/* Track if page has overflow items. */
r->ovfl_items = true;
ovfl_state = OVFL_USED;
continue;
case OVFL_USED:
/*
* Original is an overflow item; we used
* it for a key and now we need another
* copy; read it into memory.
*/
WT_ERR(__wt_dsk_cell_data_ref(session,
WT_PAGE_COL_VAR, vpack, orig));
ovfl_state = OVFL_IGNORE;
/* FALLTHROUGH */
case OVFL_IGNORE:
/*
* Original is an overflow item and we
* were forced to copy it into memory,
* or the original wasn't an overflow
* item; use the data copied into orig.
*/
data = orig->data;
size = (uint32_t)orig->size;
break;
}
}
compare: /*
* If we have a record against which to compare, and
* the records compare equal, increment the rle counter
* and continue. If the records don't compare equal,
* output the last record and swap the last and current
* buffers: do NOT update the starting record number,
* we've been doing that all along.
*/
if (rle != 0) {
if ((deleted && last_deleted) ||
(!last_deleted && !deleted &&
last->size == size &&
memcmp(last->data, data, size) == 0)) {
rle += repeat_count;
continue;
}
WT_ERR(__rec_col_var_helper(session, r,
salvage, last, last_deleted, 0, rle));
}
/*
* Swap the current/last state.
*
* Reset RLE counter and turn on comparisons.
*/
if (!deleted) {
/*
* We can't simply assign the data values into
* the last buffer because they may have come
* from a copy built from an encoded/overflow
* cell and creating the next record is going
* to overwrite that memory. Check, because
* encoded/overflow cells aren't that common
* and we'd like to avoid the copy. If data
* was taken from the current unpack structure
* (which points into the page), or was taken
* from an update structure, we can just use
* the pointers, they're not moving.
*/
if (data == vpack->data || update_no_copy) {
last->data = data;
last->size = size;
} else
WT_ERR(__wt_buf_set(
session, last, data, size));
}
last_deleted = deleted;
rle = repeat_count;
}
/*
* If we had a reference to an overflow record we never used,
* discard the underlying blocks, they're no longer useful.
*
* One complication: we must cache a copy before discarding the
* on-disk version if there's a transaction in the system that
* might read the original value.
*/
if (ovfl_state == OVFL_UNUSED &&
vpack->raw != WT_CELL_VALUE_OVFL_RM)
WT_ERR(__wt_ovfl_cache(session, page, upd, vpack));
}
/* Walk any append list. */
for (ins =
WT_SKIP_FIRST(WT_COL_APPEND(page));; ins = WT_SKIP_NEXT(ins)) {
if (ins == NULL) {
/*
* If the page split, instantiate any missing records in
* the page's name space. (Imagine record 98 is
* transactionally visible, 99 wasn't created or is not
* yet visible, 100 is visible. Then the page splits and
* record 100 moves to another page. When we reconcile
* the original page, we write record 98, then we don't
* see record 99 for whatever reason. If we've moved
* record 100, we don't know to write a deleted record
* 99 on the page.)
*
* Assert the recorded record number is past the end of
* the page.
*
* The record number recorded during the split is the
* first key on the split page, that is, one larger than
* the last key on this page, we have to decrement it.
*/
if ((n = page->
modify->mod_col_split_recno) == WT_RECNO_OOB)
break;
WT_ASSERT(session, n >= src_recno);
n -= 1;
upd = NULL;
} else {
WT_ERR(
__rec_txn_read(session, r, ins, NULL, NULL, &upd));
n = WT_INSERT_RECNO(ins);
}
while (src_recno <= n) {
/*
* The application may have inserted records which left
* gaps in the name space, and these gaps can be huge.
* If we're in a set of deleted records, skip the boring
* part.
*/
if (src_recno < n) {
deleted = true;
if (last_deleted) {
/*
* The record adjustment is decremented
* by one so we can naturally fall into
* the RLE accounting below, where we
* increment rle by one, then continue
* in the outer loop, where we increment
* src_recno by one.
*/
skip = (n - src_recno) - 1;
rle += skip;
src_recno += skip;
}
} else {
deleted = upd == NULL ||
WT_UPDATE_DELETED_ISSET(upd);
if (!deleted) {
data = WT_UPDATE_DATA(upd);
size = upd->size;
}
}
/*
* Handle RLE accounting and comparisons -- see comment
* above, this code fragment does the same thing.
*/
if (rle != 0) {
if ((deleted && last_deleted) ||
(!last_deleted && !deleted &&
last->size == size &&
memcmp(last->data, data, size) == 0)) {
++rle;
goto next;
}
WT_ERR(__rec_col_var_helper(session, r,
salvage, last, last_deleted, 0, rle));
}
/*
* Swap the current/last state. We always assign the
* data values to the buffer because they can only be
* the data from a WT_UPDATE structure.
*
* Reset RLE counter and turn on comparisons.
*/
if (!deleted) {
last->data = data;
last->size = size;
}
last_deleted = deleted;
rle = 1;
/*
* Move to the next record. It's not a simple increment
* because if it's the maximum record, incrementing it
* wraps to 0 and this turns into an infinite loop.
*/
next: if (src_recno == UINT64_MAX)
break;
++src_recno;
}
/*
* Execute this loop once without an insert item to catch any
* missing records due to a split, then quit.
*/
if (ins == NULL)
break;
}
/* If we were tracking a record, write it. */
if (rle != 0)
WT_ERR(__rec_col_var_helper(
session, r, salvage, last, last_deleted, 0, rle));
/* Write the remnant page. */
ret = __rec_split_finish(session, r);
err: __wt_scr_free(session, &orig);
return (ret);
}
/*
* __rec_row_int --
* Reconcile a row-store internal page.
*/
static int
__rec_row_int(WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_PAGE *page)
{
WT_ADDR *addr;
WT_BTREE *btree;
WT_CELL *cell;
WT_CELL_UNPACK *kpack, _kpack, *vpack, _vpack;
WT_CHILD_STATE state;
WT_DECL_RET;
WT_IKEY *ikey;
WT_KV *key, *val;
WT_PAGE *child;
WT_REF *ref;
size_t size;
u_int vtype;
bool hazard, key_onpage_ovfl, ovfl_key;
const void *p;
btree = S2BT(session);
child = NULL;
hazard = false;
key = &r->k;
kpack = &_kpack;
WT_CLEAR(*kpack); /* -Wuninitialized */
val = &r->v;
vpack = &_vpack;
WT_CLEAR(*vpack); /* -Wuninitialized */
ikey = NULL; /* -Wuninitialized */
cell = NULL;
key_onpage_ovfl = false;
WT_RET(__rec_split_init(session, r, page, 0ULL, btree->maxintlpage));
/*
* Ideally, we'd never store the 0th key on row-store internal pages
* because it's never used during tree search and there's no reason
* to waste the space. The problem is how we do splits: when we split,
* we've potentially picked out several "split points" in the buffer
* which is overflowing the maximum page size, and when the overflow
* happens, we go back and physically split the buffer, at those split
* points, into new pages. It would be both difficult and expensive
* to re-process the 0th key at each split point to be an empty key,
* so we don't do that. However, we are reconciling an internal page
* for whatever reason, and the 0th key is known to be useless. We
* truncate the key to a single byte, instead of removing it entirely,
* it simplifies various things in other parts of the code (we don't
* have to special case transforming the page from its disk image to
* its in-memory version, for example).
*/
r->cell_zero = true;
/* For each entry in the in-memory page... */
WT_INTL_FOREACH_BEGIN(session, page, ref) {
/*
* There are different paths if the key is an overflow item vs.
* a straight-forward on-page value. If an overflow item, we
* would have instantiated it, and we can use that fact to set
* things up.
*
* Note the cell reference and unpacked key cell are available
* only in the case of an instantiated, off-page key, we don't
* bother setting them if that's not possible.
*/
if (F_ISSET_ATOMIC(page, WT_PAGE_OVERFLOW_KEYS)) {
cell = NULL;
key_onpage_ovfl = false;
ikey = __wt_ref_key_instantiated(ref);
if (ikey != NULL && ikey->cell_offset != 0) {
cell =
WT_PAGE_REF_OFFSET(page, ikey->cell_offset);
__wt_cell_unpack(cell, kpack);
key_onpage_ovfl = kpack->ovfl &&
kpack->raw != WT_CELL_KEY_OVFL_RM;
}
}
WT_ERR(__rec_child_modify(session, r, ref, &hazard, &state));
addr = ref->addr;
child = ref->page;
switch (state) {
case WT_CHILD_IGNORE:
/*
* Ignored child.
*
* Overflow keys referencing pages we're not writing are
* no longer useful, schedule them for discard. Don't
* worry about instantiation, internal page keys are
* always instantiated. Don't worry about reuse,
* reusing this key in this reconciliation is unlikely.
*/
if (key_onpage_ovfl)
WT_ERR(__wt_ovfl_discard_add(
session, page, kpack->cell));
WT_CHILD_RELEASE_ERR(session, hazard, ref);
continue;
case WT_CHILD_MODIFIED:
/*
* Modified child. Empty pages are merged into the
* parent and discarded.
*/
switch (child->modify->rec_result) {
case WT_PM_REC_EMPTY:
/*
* Overflow keys referencing empty pages are no
* longer useful, schedule them for discard.
* Don't worry about instantiation, internal
* page keys are always instantiated. Don't
* worry about reuse, reusing this key in this
* reconciliation is unlikely.
*/
if (key_onpage_ovfl)
WT_ERR(__wt_ovfl_discard_add(
session, page, kpack->cell));
WT_CHILD_RELEASE_ERR(session, hazard, ref);
continue;
case WT_PM_REC_MULTIBLOCK:
/*
* Overflow keys referencing split pages are no
* longer useful (the split page's key is the
* interesting key); schedule them for discard.
* Don't worry about instantiation, internal
* page keys are always instantiated. Don't
* worry about reuse, reusing this key in this
* reconciliation is unlikely.
*/
if (key_onpage_ovfl)
WT_ERR(__wt_ovfl_discard_add(
session, page, kpack->cell));
WT_ERR(__rec_row_merge(session, r, child));
WT_CHILD_RELEASE_ERR(session, hazard, ref);
continue;
case WT_PM_REC_REPLACE:
/*
* If the page is replaced, the page's modify
* structure has the page's address.
*/
addr = &child->modify->mod_replace;
break;
WT_ILLEGAL_VALUE_ERR(session);
}
break;
case WT_CHILD_ORIGINAL:
/* Original child. */
break;
case WT_CHILD_PROXY:
/* Deleted child where we write a proxy cell. */
break;
}
/*
* Build the value cell, the child page's address. Addr points
* to an on-page cell or an off-page WT_ADDR structure. There's
* a special cell type in the case of page deletion requiring
* a proxy cell, otherwise use the information from the addr or
* original cell.
*/
if (__wt_off_page(page, addr)) {
p = addr->addr;
size = addr->size;
vtype = state == WT_CHILD_PROXY ?
WT_CELL_ADDR_DEL : __rec_vtype(addr);
} else {
__wt_cell_unpack(ref->addr, vpack);
p = vpack->data;
size = vpack->size;
vtype = state == WT_CHILD_PROXY ?
WT_CELL_ADDR_DEL : (u_int)vpack->raw;
}
__rec_cell_build_addr(session, r, p, size, vtype, WT_RECNO_OOB);
WT_CHILD_RELEASE_ERR(session, hazard, ref);
/*
* Build key cell.
* Truncate any 0th key, internal pages don't need 0th keys.
*/
if (key_onpage_ovfl) {
key->buf.data = cell;
key->buf.size = __wt_cell_total_len(kpack);
key->cell_len = 0;
key->len = key->buf.size;
ovfl_key = true;
} else {
__wt_ref_key(page, ref, &p, &size);
WT_ERR(__rec_cell_build_int_key(
session, r, p, r->cell_zero ? 1 : size, &ovfl_key));
}
r->cell_zero = false;
/* Boundary: split or write the page. */
if (key->len + val->len > r->space_avail) {
if (r->raw_compression)
WT_ERR(__rec_split_raw(
session, r, key->len + val->len));
else {
/*
* In one path above, we copied address blocks
* from the page rather than building the actual
* key. In that case, we have to build the key
* now because we are about to promote it.
*/
if (key_onpage_ovfl) {
WT_ERR(__wt_buf_set(session, r->cur,
WT_IKEY_DATA(ikey), ikey->size));
key_onpage_ovfl = false;
}
WT_ERR(__rec_split(
session, r, key->len + val->len));
}
}
/* Copy the key and value onto the page. */
__rec_copy_incr(session, r, key);
__rec_copy_incr(session, r, val);
/* Update compression state. */
__rec_key_state_update(r, ovfl_key);
} WT_INTL_FOREACH_END;
/* Write the remnant page. */
return (__rec_split_finish(session, r));
err: WT_CHILD_RELEASE(session, hazard, ref);
return (ret);
}
/*
* __rec_row_merge --
* Merge in a split page.
*/
static int
__rec_row_merge(WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_PAGE *page)
{
WT_ADDR *addr;
WT_KV *key, *val;
WT_MULTI *multi;
WT_PAGE_MODIFY *mod;
uint32_t i;
bool ovfl_key;
mod = page->modify;
key = &r->k;
val = &r->v;
/* For each entry in the split array... */
for (multi = mod->mod_multi,
i = 0; i < mod->mod_multi_entries; ++multi, ++i) {
/* Build the key and value cells. */
WT_RET(__rec_cell_build_int_key(session, r,
WT_IKEY_DATA(multi->key.ikey),
r->cell_zero ? 1 : multi->key.ikey->size, &ovfl_key));
r->cell_zero = false;
addr = &multi->addr;
__rec_cell_build_addr(session, r,
addr->addr, addr->size, __rec_vtype(addr), WT_RECNO_OOB);
/* Boundary: split or write the page. */
if (key->len + val->len > r->space_avail)
WT_RET(r->raw_compression ?
__rec_split_raw(session, r, key->len + val->len) :
__rec_split(session, r, key->len + val->len));
/* Copy the key and value onto the page. */
__rec_copy_incr(session, r, key);
__rec_copy_incr(session, r, val);
/* Update compression state. */
__rec_key_state_update(r, ovfl_key);
}
return (0);
}
/*
* __rec_row_leaf --
* Reconcile a row-store leaf page.
*/
static int
__rec_row_leaf(WT_SESSION_IMPL *session,
WT_RECONCILE *r, WT_PAGE *page, WT_SALVAGE_COOKIE *salvage)
{
WT_BTREE *btree;
WT_CELL *cell, *val_cell;
WT_CELL_UNPACK *kpack, _kpack, *vpack, _vpack;
WT_DECL_ITEM(tmpkey);
WT_DECL_ITEM(tmpval);
WT_DECL_RET;
WT_IKEY *ikey;
WT_INSERT *ins;
WT_KV *key, *val;
WT_ROW *rip;
WT_UPDATE *upd;
size_t size;
uint64_t slvg_skip;
uint32_t i;
bool dictionary, key_onpage_ovfl, ovfl_key;
const void *p;
void *copy;
btree = S2BT(session);
slvg_skip = salvage == NULL ? 0 : salvage->skip;
key = &r->k;
val = &r->v;
WT_RET(__rec_split_init(session, r, page, 0ULL, btree->maxleafpage));
/*
* Write any K/V pairs inserted into the page before the first from-disk
* key on the page.
*/
if ((ins = WT_SKIP_FIRST(WT_ROW_INSERT_SMALLEST(page))) != NULL)
WT_RET(__rec_row_leaf_insert(session, r, ins));
/*
* Temporary buffers in which to instantiate any uninstantiated keys
* or value items we need.
*/
WT_ERR(__wt_scr_alloc(session, 0, &tmpkey));
WT_ERR(__wt_scr_alloc(session, 0, &tmpval));
/* For each entry in the page... */
WT_ROW_FOREACH(page, rip, i) {
/*
* The salvage code, on some rare occasions, wants to reconcile
* a page but skip some leading records on the page. Because
* the row-store leaf reconciliation function copies keys from
* the original disk page, this is non-trivial -- just changing
* the in-memory pointers isn't sufficient, we have to change
* the WT_CELL structures on the disk page, too. It's ugly, but
* we pass in a value that tells us how many records to skip in
* this case.
*/
if (slvg_skip != 0) {
--slvg_skip;
continue;
}
/*
* Figure out the key: set any cell reference (and unpack it),
* set any instantiated key reference.
*/
copy = WT_ROW_KEY_COPY(rip);
(void)__wt_row_leaf_key_info(
page, copy, &ikey, &cell, NULL, NULL);
if (cell == NULL)
kpack = NULL;
else {
kpack = &_kpack;
__wt_cell_unpack(cell, kpack);
}
/* Unpack the on-page value cell, and look for an update. */
if ((val_cell =
__wt_row_leaf_value_cell(page, rip, NULL)) == NULL)
vpack = NULL;
else {
vpack = &_vpack;
__wt_cell_unpack(val_cell, vpack);
}
WT_ERR(__rec_txn_read(session, r, NULL, rip, vpack, &upd));
/* Build value cell. */
dictionary = false;
if (upd == NULL) {
/*
* When the page was read into memory, there may not
* have been a value item.
*
* If there was a value item, check if it's a dictionary
* cell (a copy of another item on the page). If it's a
* copy, we have to create a new value item as the old
* item might have been discarded from the page.
*/
if (vpack == NULL) {
val->buf.data = NULL;
val->cell_len = val->len = val->buf.size = 0;
} else if (vpack->raw == WT_CELL_VALUE_COPY) {
/* If the item is Huffman encoded, decode it. */
if (btree->huffman_value == NULL) {
p = vpack->data;
size = vpack->size;
} else {
WT_ERR(__wt_huffman_decode(session,
btree->huffman_value,
vpack->data, vpack->size,
tmpval));
p = tmpval->data;
size = tmpval->size;
}
WT_ERR(__rec_cell_build_val(
session, r, p, size, (uint64_t)0));
dictionary = true;
} else if (vpack->raw == WT_CELL_VALUE_OVFL_RM) {
/*
* If doing update save and restore in service
* of eviction, there's an update that's not
* globally visible, and the underlying value
* is a removed overflow value, we end up here.
*
* When the update save/restore code noticed the
* removed overflow value, it appended a copy of
* the cached, original overflow value to the
* update list being saved (ensuring any on-page
* item will never be accessed after the page is
* re-instantiated), then returned a NULL update
* to us.
*
* Assert the case.
*/
WT_ASSERT(session,
F_ISSET(r, WT_EVICT_UPDATE_RESTORE));
/*
* If the key is also a removed overflow item,
* don't write anything at all.
*
* We don't have to write anything because the
* code re-instantiating the page gets the key
* to match the saved list of updates from the
* original page. By not putting the key on
* the page, we'll move the key/value set from
* a row-store leaf page slot to an insert list,
* but that shouldn't matter.
*
* The reason we bother with the test is because
* overflows are expensive to write. It's hard
* to imagine a real workload where this test is
* worth the effort, but it's a simple test.
*/
if (kpack != NULL &&
kpack->raw == WT_CELL_KEY_OVFL_RM)
goto leaf_insert;
/*
* The on-page value will never be accessed,
* write a placeholder record.
*/
WT_ERR(__rec_cell_build_val(
session, r, "@", 1, (uint64_t)0));
} else {
val->buf.data = val_cell;
val->buf.size = __wt_cell_total_len(vpack);
val->cell_len = 0;
val->len = val->buf.size;
/* Track if page has overflow items. */
if (vpack->ovfl)
r->ovfl_items = true;
}
} else {
/*
* If the original value was an overflow and we've not
* already done so, discard it. One complication: we
* must cache a copy before discarding the on-disk
* version if there's a transaction in the system that
* might read the original value.
*/
if (vpack != NULL &&
vpack->ovfl && vpack->raw != WT_CELL_VALUE_OVFL_RM)
WT_ERR(
__wt_ovfl_cache(session, page, rip, vpack));
/* If this key/value pair was deleted, we're done. */
if (WT_UPDATE_DELETED_ISSET(upd)) {
/*
* Overflow keys referencing discarded values
* are no longer useful, discard the backing
* blocks. Don't worry about reuse, reusing
* keys from a row-store page reconciliation
* seems unlikely enough to ignore.
*/
if (kpack != NULL && kpack->ovfl &&
kpack->raw != WT_CELL_KEY_OVFL_RM) {
/*
* Keys are part of the name-space, we
* can't remove them from the in-memory
* tree; if an overflow key was deleted
* without being instantiated (for
* example, cursor-based truncation), do
* it now.
*/
if (ikey == NULL)
WT_ERR(__wt_row_leaf_key(
session,
page, rip, tmpkey, true));
WT_ERR(__wt_ovfl_discard_add(
session, page, kpack->cell));
}
/*
* We aren't actually creating the key so we
* can't use bytes from this key to provide
* prefix information for a subsequent key.
*/
tmpkey->size = 0;
/* Proceed with appended key/value pairs. */
goto leaf_insert;
}
/*
* If no value, nothing needs to be copied. Otherwise,
* build the value's WT_CELL chunk from the most recent
* update value.
*/
if (upd->size == 0) {
val->buf.data = NULL;
val->cell_len = val->len = val->buf.size = 0;
} else {
WT_ERR(__rec_cell_build_val(session, r,
WT_UPDATE_DATA(upd), upd->size,
(uint64_t)0));
dictionary = true;
}
}
/*
* Build key cell.
*
* If the key is an overflow key that hasn't been removed, use
* the original backing blocks.
*/
key_onpage_ovfl = kpack != NULL &&
kpack->ovfl && kpack->raw != WT_CELL_KEY_OVFL_RM;
if (key_onpage_ovfl) {
key->buf.data = cell;
key->buf.size = __wt_cell_total_len(kpack);
key->cell_len = 0;
key->len = key->buf.size;
ovfl_key = true;
/*
* We aren't creating a key so we can't use this key as
* a prefix for a subsequent key.
*/
tmpkey->size = 0;
/* Track if page has overflow items. */
r->ovfl_items = true;
} else {
/*
* Get the key from the page or an instantiated key, or
* inline building the key from a previous key (it's a
* fast path for simple, prefix-compressed keys), or by
* by building the key from scratch.
*/
if (__wt_row_leaf_key_info(page, copy,
NULL, &cell, &tmpkey->data, &tmpkey->size))
goto build;
kpack = &_kpack;
__wt_cell_unpack(cell, kpack);
if (btree->huffman_key == NULL &&
kpack->type == WT_CELL_KEY &&
tmpkey->size >= kpack->prefix) {
/*
* The previous clause checked for a prefix of
* zero, which means the temporary buffer must
* have a non-zero size, and it references a
* valid key.
*/
WT_ASSERT(session, tmpkey->size != 0);
/*
* Grow the buffer as necessary, ensuring data
* data has been copied into local buffer space,
* then append the suffix to the prefix already
* in the buffer.
*
* Don't grow the buffer unnecessarily or copy
* data we don't need, truncate the item's data
* length to the prefix bytes.
*/
tmpkey->size = kpack->prefix;
WT_ERR(__wt_buf_grow(session,
tmpkey, tmpkey->size + kpack->size));
memcpy((uint8_t *)tmpkey->mem + tmpkey->size,
kpack->data, kpack->size);
tmpkey->size += kpack->size;
} else
WT_ERR(__wt_row_leaf_key_copy(
session, page, rip, tmpkey));
build:
WT_ERR(__rec_cell_build_leaf_key(session, r,
tmpkey->data, tmpkey->size, &ovfl_key));
}
/* Boundary: split or write the page. */
if (key->len + val->len > r->space_avail) {
if (r->raw_compression)
WT_ERR(__rec_split_raw(
session, r, key->len + val->len));
else {
/*
* In one path above, we copied address blocks
* from the page rather than building the actual
* key. In that case, we have to build the key
* now because we are about to promote it.
*/
if (key_onpage_ovfl) {
WT_ERR(__wt_dsk_cell_data_ref(session,
WT_PAGE_ROW_LEAF, kpack, r->cur));
key_onpage_ovfl = false;
}
/*
* Turn off prefix compression until a full key
* written to the new page, and (unless already
* working with an overflow key), rebuild the
* key without compression.
*/
if (r->key_pfx_compress_conf) {
r->key_pfx_compress = false;
if (!ovfl_key)
WT_ERR(
__rec_cell_build_leaf_key(
session,
r, NULL, 0, &ovfl_key));
}
WT_ERR(__rec_split(
session, r, key->len + val->len));
}
}
/* Copy the key/value pair onto the page. */
__rec_copy_incr(session, r, key);
if (val->len == 0)
r->any_empty_value = true;
else {
r->all_empty_value = false;
if (dictionary && btree->dictionary)
WT_ERR(__rec_dict_replace(session, r, 0, val));
__rec_copy_incr(session, r, val);
}
/* Update compression state. */
__rec_key_state_update(r, ovfl_key);
leaf_insert: /* Write any K/V pairs inserted into the page after this key. */
if ((ins = WT_SKIP_FIRST(WT_ROW_INSERT(page, rip))) != NULL)
WT_ERR(__rec_row_leaf_insert(session, r, ins));
}
/* Write the remnant page. */
ret = __rec_split_finish(session, r);
err: __wt_scr_free(session, &tmpkey);
__wt_scr_free(session, &tmpval);
return (ret);
}
/*
* __rec_row_leaf_insert --
* Walk an insert chain, writing K/V pairs.
*/
static int
__rec_row_leaf_insert(WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_INSERT *ins)
{
WT_BTREE *btree;
WT_KV *key, *val;
WT_UPDATE *upd;
bool ovfl_key;
btree = S2BT(session);
key = &r->k;
val = &r->v;
for (; ins != NULL; ins = WT_SKIP_NEXT(ins)) {
/* Look for an update. */
WT_RET(__rec_txn_read(session, r, ins, NULL, NULL, &upd));
if (upd == NULL || WT_UPDATE_DELETED_ISSET(upd))
continue;
if (upd->size == 0) /* Build value cell. */
val->len = 0;
else
WT_RET(__rec_cell_build_val(session, r,
WT_UPDATE_DATA(upd), upd->size, (uint64_t)0));
/* Build key cell. */
WT_RET(__rec_cell_build_leaf_key(session, r,
WT_INSERT_KEY(ins), WT_INSERT_KEY_SIZE(ins), &ovfl_key));
/* Boundary: split or write the page. */
if (key->len + val->len > r->space_avail) {
if (r->raw_compression)
WT_RET(__rec_split_raw(
session, r, key->len + val->len));
else {
/*
* Turn off prefix compression until a full key
* written to the new page, and (unless already
* working with an overflow key), rebuild the
* key without compression.
*/
if (r->key_pfx_compress_conf) {
r->key_pfx_compress = false;
if (!ovfl_key)
WT_RET(
__rec_cell_build_leaf_key(
session,
r, NULL, 0, &ovfl_key));
}
WT_RET(__rec_split(
session, r, key->len + val->len));
}
}
/* Copy the key/value pair onto the page. */
__rec_copy_incr(session, r, key);
if (val->len == 0)
r->any_empty_value = true;
else {
r->all_empty_value = false;
if (btree->dictionary)
WT_RET(__rec_dict_replace(session, r, 0, val));
__rec_copy_incr(session, r, val);
}
/* Update compression state. */
__rec_key_state_update(r, ovfl_key);
}
return (0);
}
/*
* __rec_split_discard --
* Discard the pages resulting from a previous split.
*/
static int
__rec_split_discard(WT_SESSION_IMPL *session, WT_PAGE *page)
{
WT_DECL_RET;
WT_PAGE_MODIFY *mod;
WT_MULTI *multi;
uint32_t i;
mod = page->modify;
/*
* A page that split is being reconciled for the second, or subsequent
* time; discard underlying block space used in the last reconciliation
* that is not being reused for this reconciliation.
*/
for (multi = mod->mod_multi,
i = 0; i < mod->mod_multi_entries; ++multi, ++i) {
switch (page->type) {
case WT_PAGE_ROW_INT:
case WT_PAGE_ROW_LEAF:
__wt_free(session, multi->key.ikey);
break;
}
/*
* If the page was re-written free the backing disk blocks used
* in the previous write (unless the blocks were reused in this
* write). The page may instead have been a disk image with
* associated saved updates: ownership of the disk image is
* transferred when rewriting the page in-memory and there may
* not have been saved updates. We've gotten this wrong a few
* times, so use the existence of an address to confirm backing
* blocks we care about, and free any disk image/saved updates.
*/
if (multi->addr.addr != NULL && !multi->addr.reuse) {
WT_RET(__wt_btree_block_free(
session, multi->addr.addr, multi->addr.size));
__wt_free(session, multi->addr.addr);
}
__wt_free(session, multi->supd);
__wt_free(session, multi->disk_image);
}
__wt_free(session, mod->mod_multi);
mod->mod_multi_entries = 0;
/*
* This routine would be trivial, and only walk a single page freeing
* any blocks written to support the split, except for root splits.
* In the case of root splits, we have to cope with multiple pages in
* a linked list, and we also have to discard overflow items written
* for the page.
*/
switch (page->type) {
case WT_PAGE_COL_INT:
case WT_PAGE_ROW_INT:
if (mod->mod_root_split == NULL)
break;
WT_RET(__rec_split_discard(session, mod->mod_root_split));
WT_RET(__wt_ovfl_track_wrapup(session, mod->mod_root_split));
__wt_page_out(session, &mod->mod_root_split);
break;
}
return (ret);
}
/*
* __rec_split_dump_keys --
* Dump out the split keys in verbose mode.
*/
static int
__rec_split_dump_keys(WT_SESSION_IMPL *session, WT_PAGE *page, WT_RECONCILE *r)
{
WT_BOUNDARY *bnd;
WT_DECL_ITEM(tkey);
WT_DECL_RET;
uint32_t i;
if (page->type == WT_PAGE_ROW_INT || page->type == WT_PAGE_ROW_LEAF)
WT_RET(__wt_scr_alloc(session, 0, &tkey));
WT_ERR(__wt_verbose(
session, WT_VERB_SPLIT, "split: %" PRIu32 " pages", r->bnd_next));
for (bnd = r->bnd, i = 0; i < r->bnd_next; ++bnd, ++i)
switch (page->type) {
case WT_PAGE_ROW_INT:
case WT_PAGE_ROW_LEAF:
WT_ERR(__wt_verbose(session, WT_VERB_SPLIT,
"starting key %s",
__wt_buf_set_printable(
session, bnd->key.data, bnd->key.size, tkey)));
break;
case WT_PAGE_COL_FIX:
case WT_PAGE_COL_INT:
case WT_PAGE_COL_VAR:
WT_ERR(__wt_verbose(session, WT_VERB_SPLIT,
"starting recno %" PRIu64, bnd->recno));
break;
WT_ILLEGAL_VALUE_ERR(session);
}
err: __wt_scr_free(session, &tkey);
return (ret);
}
/*
* __rec_write_wrapup --
* Finish the reconciliation.
*/
static int
__rec_write_wrapup(WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_PAGE *page)
{
WT_BM *bm;
WT_BOUNDARY *bnd;
WT_BTREE *btree;
WT_PAGE_MODIFY *mod;
WT_REF *ref;
btree = S2BT(session);
bm = btree->bm;
mod = page->modify;
ref = r->ref;
/*
* This page may have previously been reconciled, and that information
* is now about to be replaced. Make sure it's discarded at some point,
* and clear the underlying modification information, we're creating a
* new reality.
*/
switch (mod->rec_result) {
case 0: /*
* The page has never been reconciled before, free the original
* address blocks (if any). The "if any" is for empty trees
* created when a new tree is opened or previously deleted pages
* instantiated in memory.
*
* The exception is root pages are never tracked or free'd, they
* are checkpoints, and must be explicitly dropped.
*/
if (__wt_ref_is_root(ref))
break;
WT_RET(__wt_ref_block_free(session, ref));
break;
case WT_PM_REC_EMPTY: /* Page deleted */
break;
case WT_PM_REC_MULTIBLOCK: /* Multiple blocks */
/*
* Discard the multiple replacement blocks.
*/
WT_RET(__rec_split_discard(session, page));
break;
case WT_PM_REC_REPLACE: /* 1-for-1 page swap */
/*
* Discard the replacement leaf page's blocks.
*
* The exception is root pages are never tracked or free'd, they
* are checkpoints, and must be explicitly dropped.
*/
if (!__wt_ref_is_root(ref))
WT_RET(__wt_btree_block_free(session,
mod->mod_replace.addr, mod->mod_replace.size));
/* Discard the replacement page's address and disk image. */
__wt_free(session, mod->mod_replace.addr);
mod->mod_replace.size = 0;
__wt_free(session, mod->mod_disk_image);
break;
WT_ILLEGAL_VALUE(session);
}
/* Reset the reconciliation state. */
mod->rec_result = 0;
/*
* Wrap up overflow tracking. If we are about to create a checkpoint,
* the system must be entirely consistent at that point (the underlying
* block manager is presumably going to do some action to resolve the
* list of allocated/free/whatever blocks that are associated with the
* checkpoint).
*/
WT_RET(__wt_ovfl_track_wrapup(session, page));
switch (r->bnd_next) {
case 0: /* Page delete */
WT_RET(__wt_verbose(
session, WT_VERB_RECONCILE, "page %p empty", page));
WT_STAT_FAST_CONN_INCR(session, rec_page_delete);
WT_STAT_FAST_DATA_INCR(session, rec_page_delete);
/* If this is the root page, we need to create a sync point. */
ref = r->ref;
if (__wt_ref_is_root(ref))
WT_RET(bm->checkpoint(
bm, session, NULL, btree->ckpt, false));
/*
* If the page was empty, we want to discard it from the tree
* by discarding the parent's key when evicting the parent.
* Mark the page as deleted, then return success, leaving the
* page in memory. If the page is subsequently modified, that
* is OK, we'll just reconcile it again.
*/
mod->rec_result = WT_PM_REC_EMPTY;
break;
case 1: /* 1-for-1 page swap */
/*
* Because WiredTiger's pages grow without splitting, we're
* replacing a single page with another single page most of
* the time.
*/
bnd = &r->bnd[0];
/*
* If in-memory, or saving/restoring changes for this page and
* there's only one block, there's nothing to write. Set up
* a single block as if to split, then use that disk image to
* rewrite the page in memory. This is separate from simple
* replacements where eviction has decided to retain the page
* in memory because the latter can't handle update lists and
* splits can.
*/
if (F_ISSET(r, WT_EVICT_IN_MEMORY) ||
(F_ISSET(r, WT_EVICT_UPDATE_RESTORE) && bnd->supd != NULL))
goto split;
/*
* A root page, we don't have an address and we have to create
* a sync point. The address was cleared when we were about to
* write the buffer so we know what to do here.
*/
if (bnd->addr.addr == NULL)
WT_RET(__wt_bt_write(session, &r->disk_image,
NULL, NULL, true, F_ISSET(r, WT_CHECKPOINTING),
bnd->already_compressed));
else {
mod->mod_replace = bnd->addr;
bnd->addr.addr = NULL;
mod->mod_disk_image = bnd->disk_image;
bnd->disk_image = NULL;
}
mod->rec_result = WT_PM_REC_REPLACE;
break;
default: /* Page split */
WT_RET(__wt_verbose(session, WT_VERB_RECONCILE,
"page %p reconciled into %" PRIu32 " pages",
page, r->bnd_next));
switch (page->type) {
case WT_PAGE_COL_INT:
case WT_PAGE_ROW_INT:
WT_STAT_FAST_DATA_INCR(
session, rec_multiblock_internal);
break;
case WT_PAGE_COL_FIX:
case WT_PAGE_COL_VAR:
case WT_PAGE_ROW_LEAF:
WT_STAT_FAST_DATA_INCR(session, rec_multiblock_leaf);
break;
WT_ILLEGAL_VALUE(session);
}
/* Optionally display the actual split keys in verbose mode. */
if (WT_VERBOSE_ISSET(session, WT_VERB_SPLIT))
WT_RET(__rec_split_dump_keys(session, page, r));
/* Track the largest set of page-splits. */
if (r->bnd_next > r->bnd_next_max) {
r->bnd_next_max = r->bnd_next;
WT_STAT_FAST_DATA_SET(
session, rec_multiblock_max, r->bnd_next_max);
}
split: switch (page->type) {
case WT_PAGE_ROW_INT:
case WT_PAGE_ROW_LEAF:
WT_RET(__rec_split_row(session, r, page));
break;
case WT_PAGE_COL_INT:
case WT_PAGE_COL_FIX:
case WT_PAGE_COL_VAR:
WT_RET(__rec_split_col(session, r, page));
break;
WT_ILLEGAL_VALUE(session);
}
mod->rec_result = WT_PM_REC_MULTIBLOCK;
break;
}
return (0);
}
/*
* __rec_write_wrapup_err --
* Finish the reconciliation on error.
*/
static int
__rec_write_wrapup_err(WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_PAGE *page)
{
WT_BOUNDARY *bnd;
WT_DECL_RET;
WT_MULTI *multi;
WT_PAGE_MODIFY *mod;
uint32_t i;
mod = page->modify;
/*
* Clear the address-reused flag from the multiblock reconciliation
* information (otherwise we might think the backing block is being
* reused on a subsequent reconciliation where we want to free it).
*/
if (mod->rec_result == WT_PM_REC_MULTIBLOCK)
for (multi = mod->mod_multi,
i = 0; i < mod->mod_multi_entries; ++multi, ++i)
multi->addr.reuse = 0;
/*
* On error, discard blocks we've written, they're unreferenced by the
* tree. This is not a question of correctness, we're avoiding block
* leaks.
*
* Don't discard backing blocks marked for reuse, they remain part of
* a previous reconciliation.
*/
WT_TRET(__wt_ovfl_track_wrapup_err(session, page));
for (bnd = r->bnd, i = 0; i < r->bnd_next; ++bnd, ++i)
if (bnd->addr.addr != NULL) {
if (bnd->addr.reuse)
bnd->addr.addr = NULL;
else {
WT_TRET(__wt_btree_block_free(session,
bnd->addr.addr, bnd->addr.size));
__wt_free(session, bnd->addr.addr);
}
}
return (ret);
}
/*
* __rec_split_row --
* Split a row-store page into a set of replacement blocks.
*/
static int
__rec_split_row(WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_PAGE *page)
{
WT_BOUNDARY *bnd;
WT_MULTI *multi;
WT_PAGE_MODIFY *mod;
WT_REF *ref;
uint32_t i;
size_t size;
void *p;
mod = page->modify;
/* We never set the first page's key, grab it from the original page. */
ref = r->ref;
if (__wt_ref_is_root(ref))
WT_RET(__wt_buf_set(session, &r->bnd[0].key, "", 1));
else {
__wt_ref_key(ref->home, ref, &p, &size);
WT_RET(__wt_buf_set(session, &r->bnd[0].key, p, size));
}
/* Allocate, then initialize the array of replacement blocks. */
WT_RET(__wt_calloc_def(session, r->bnd_next, &mod->mod_multi));
for (multi = mod->mod_multi,
bnd = r->bnd, i = 0; i < r->bnd_next; ++multi, ++bnd, ++i) {
WT_RET(__wt_row_ikey_alloc(session, 0,
bnd->key.data, bnd->key.size, &multi->key.ikey));
/*
* Copy any disk image. Don't take saved updates without a
* disk image (which happens if they have been saved to the
* lookaside table): they should be discarded along with the
* original page.
*/
multi->disk_image = bnd->disk_image;
bnd->disk_image = NULL;
if (multi->disk_image != NULL) {
multi->supd = bnd->supd;
multi->supd_entries = bnd->supd_next;
bnd->supd = NULL;
}
/* Copy any address. */
multi->addr = bnd->addr;
multi->addr.reuse = 0;
multi->size = bnd->size;
multi->cksum = bnd->cksum;
bnd->addr.addr = NULL;
}
mod->mod_multi_entries = r->bnd_next;
return (0);
}
/*
* __rec_split_col --
* Split a column-store page into a set of replacement blocks.
*/
static int
__rec_split_col(WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_PAGE *page)
{
WT_BOUNDARY *bnd;
WT_MULTI *multi;
WT_PAGE_MODIFY *mod;
uint32_t i;
mod = page->modify;
/* Allocate, then initialize the array of replacement blocks. */
WT_RET(__wt_calloc_def(session, r->bnd_next, &mod->mod_multi));
for (multi = mod->mod_multi,
bnd = r->bnd, i = 0; i < r->bnd_next; ++multi, ++bnd, ++i) {
multi->key.recno = bnd->recno;
/*
* Copy any disk image. Don't take saved updates without a
* disk image (which happens if they have been saved to the
* lookaside table): they should be discarded along with the
* original page.
*/
multi->disk_image = bnd->disk_image;
bnd->disk_image = NULL;
if (multi->disk_image != NULL) {
multi->supd = bnd->supd;
multi->supd_entries = bnd->supd_next;
bnd->supd = NULL;
}
/* Copy any address. */
multi->addr = bnd->addr;
multi->addr.reuse = 0;
multi->size = bnd->size;
multi->cksum = bnd->cksum;
bnd->addr.addr = NULL;
}
mod->mod_multi_entries = r->bnd_next;
return (0);
}
/*
* __rec_cell_build_int_key --
* Process a key and return a WT_CELL structure and byte string to be
* stored on a row-store internal page.
*/
static int
__rec_cell_build_int_key(WT_SESSION_IMPL *session,
WT_RECONCILE *r, const void *data, size_t size, bool *is_ovflp)
{
WT_BTREE *btree;
WT_KV *key;
*is_ovflp = false;
btree = S2BT(session);
key = &r->k;
/* Copy the bytes into the "current" and key buffers. */
WT_RET(__wt_buf_set(session, r->cur, data, size));
WT_RET(__wt_buf_set(session, &key->buf, data, size));
/* Create an overflow object if the data won't fit. */
if (size > btree->maxintlkey) {
WT_STAT_FAST_DATA_INCR(session, rec_overflow_key_internal);
*is_ovflp = true;
return (__rec_cell_build_ovfl(
session, r, key, WT_CELL_KEY_OVFL, (uint64_t)0));
}
key->cell_len = __wt_cell_pack_int_key(&key->cell, key->buf.size);
key->len = key->cell_len + key->buf.size;
return (0);
}
/*
* __rec_cell_build_leaf_key --
* Process a key and return a WT_CELL structure and byte string to be
* stored on a row-store leaf page.
*/
static int
__rec_cell_build_leaf_key(WT_SESSION_IMPL *session,
WT_RECONCILE *r, const void *data, size_t size, bool *is_ovflp)
{
WT_BTREE *btree;
WT_KV *key;
size_t pfx_max;
uint8_t pfx;
const uint8_t *a, *b;
*is_ovflp = false;
btree = S2BT(session);
key = &r->k;
pfx = 0;
if (data == NULL)
/*
* When data is NULL, our caller has a prefix compressed key
* they can't use (probably because they just crossed a split
* point). Use the full key saved when last called, instead.
*/
WT_RET(__wt_buf_set(
session, &key->buf, r->cur->data, r->cur->size));
else {
/*
* Save a copy of the key for later reference: we use the full
* key for prefix-compression comparisons, and if we are, for
* any reason, unable to use the compressed key we generate.
*/
WT_RET(__wt_buf_set(session, r->cur, data, size));
/*
* Do prefix compression on the key. We know by definition the
* previous key sorts before the current key, which means the
* keys must differ and we just need to compare up to the
* shorter of the two keys.
*/
if (r->key_pfx_compress) {
/*
* We can't compress out more than 256 bytes, limit the
* comparison to that.
*/
pfx_max = UINT8_MAX;
if (size < pfx_max)
pfx_max = size;
if (r->last->size < pfx_max)
pfx_max = r->last->size;
for (a = data, b = r->last->data; pfx < pfx_max; ++pfx)
if (*a++ != *b++)
break;
/*
* Prefix compression may cost us CPU and memory when
* the page is re-loaded, don't do it unless there's
* reasonable gain.
*/
if (pfx < btree->prefix_compression_min)
pfx = 0;
else
WT_STAT_FAST_DATA_INCRV(
session, rec_prefix_compression, pfx);
}
/* Copy the non-prefix bytes into the key buffer. */
WT_RET(__wt_buf_set(
session, &key->buf, (uint8_t *)data + pfx, size - pfx));
}
/* Optionally compress the key using the Huffman engine. */
if (btree->huffman_key != NULL)
WT_RET(__wt_huffman_encode(session, btree->huffman_key,
key->buf.data, (uint32_t)key->buf.size, &key->buf));
/* Create an overflow object if the data won't fit. */
if (key->buf.size > btree->maxleafkey) {
/*
* Overflow objects aren't prefix compressed -- rebuild any
* object that was prefix compressed.
*/
if (pfx == 0) {
WT_STAT_FAST_DATA_INCR(session, rec_overflow_key_leaf);
*is_ovflp = true;
return (__rec_cell_build_ovfl(
session, r, key, WT_CELL_KEY_OVFL, (uint64_t)0));
}
return (
__rec_cell_build_leaf_key(session, r, NULL, 0, is_ovflp));
}
key->cell_len = __wt_cell_pack_leaf_key(&key->cell, pfx, key->buf.size);
key->len = key->cell_len + key->buf.size;
return (0);
}
/*
* __rec_cell_build_addr --
* Process an address reference and return a cell structure to be stored
* on the page.
*/
static void
__rec_cell_build_addr(WT_SESSION_IMPL *session, WT_RECONCILE *r,
const void *addr, size_t size, u_int cell_type, uint64_t recno)
{
WT_KV *val;
val = &r->v;
WT_ASSERT(session, size != 0 || cell_type == WT_CELL_ADDR_DEL);
/*
* We don't check the address size because we can't store an address on
* an overflow page: if the address won't fit, the overflow page's
* address won't fit either. This possibility must be handled by Btree
* configuration, we have to disallow internal page sizes that are too
* small with respect to the largest address cookie the underlying block
* manager might return.
*/
/*
* We don't copy the data into the buffer, it's not necessary; just
* re-point the buffer's data/length fields.
*/
val->buf.data = addr;
val->buf.size = size;
val->cell_len =
__wt_cell_pack_addr(&val->cell, cell_type, recno, val->buf.size);
val->len = val->cell_len + val->buf.size;
}
/*
* __rec_cell_build_val --
* Process a data item and return a WT_CELL structure and byte string to
* be stored on the page.
*/
static int
__rec_cell_build_val(WT_SESSION_IMPL *session,
WT_RECONCILE *r, const void *data, size_t size, uint64_t rle)
{
WT_BTREE *btree;
WT_KV *val;
btree = S2BT(session);
val = &r->v;
/*
* We don't copy the data into the buffer, it's not necessary; just
* re-point the buffer's data/length fields.
*/
val->buf.data = data;
val->buf.size = size;
/* Handle zero-length cells quickly. */
if (size != 0) {
/* Optionally compress the data using the Huffman engine. */
if (btree->huffman_value != NULL)
WT_RET(__wt_huffman_encode(
session, btree->huffman_value,
val->buf.data, (uint32_t)val->buf.size, &val->buf));
/* Create an overflow object if the data won't fit. */
if (val->buf.size > btree->maxleafvalue) {
WT_STAT_FAST_DATA_INCR(session, rec_overflow_value);
return (__rec_cell_build_ovfl(
session, r, val, WT_CELL_VALUE_OVFL, rle));
}
}
val->cell_len = __wt_cell_pack_data(&val->cell, rle, val->buf.size);
val->len = val->cell_len + val->buf.size;
return (0);
}
/*
* __rec_cell_build_ovfl --
* Store overflow items in the file, returning the address cookie.
*/
static int
__rec_cell_build_ovfl(WT_SESSION_IMPL *session,
WT_RECONCILE *r, WT_KV *kv, uint8_t type, uint64_t rle)
{
WT_BM *bm;
WT_BTREE *btree;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_PAGE *page;
WT_PAGE_HEADER *dsk;
size_t size;
uint8_t *addr, buf[WT_BTREE_MAX_ADDR_COOKIE];
btree = S2BT(session);
bm = btree->bm;
page = r->page;
/* Track if page has overflow items. */
r->ovfl_items = true;
/*
* See if this overflow record has already been written and reuse it if
* possible, otherwise write a new overflow record.
*/
WT_RET(__wt_ovfl_reuse_search(
session, page, &addr, &size, kv->buf.data, kv->buf.size));
if (addr == NULL) {
/* Allocate a buffer big enough to write the overflow record. */
size = kv->buf.size;
WT_RET(bm->write_size(bm, session, &size));
WT_RET(__wt_scr_alloc(session, size, &tmp));
/* Initialize the buffer: disk header and overflow record. */
dsk = tmp->mem;
memset(dsk, 0, WT_PAGE_HEADER_SIZE);
dsk->type = WT_PAGE_OVFL;
dsk->u.datalen = (uint32_t)kv->buf.size;
memcpy(WT_PAGE_HEADER_BYTE(btree, dsk),
kv->buf.data, kv->buf.size);
dsk->mem_size =
WT_PAGE_HEADER_BYTE_SIZE(btree) + (uint32_t)kv->buf.size;
tmp->size = dsk->mem_size;
/* Write the buffer. */
addr = buf;
WT_ERR(__wt_bt_write(session, tmp,
addr, &size, false, F_ISSET(r, WT_CHECKPOINTING), false));
/*
* Track the overflow record (unless it's a bulk load, which
* by definition won't ever reuse a record.
*/
if (!r->is_bulk_load)
WT_ERR(__wt_ovfl_reuse_add(session, page,
addr, size, kv->buf.data, kv->buf.size));
}
/* Set the callers K/V to reference the overflow record's address. */
WT_ERR(__wt_buf_set(session, &kv->buf, addr, size));
/* Build the cell and return. */
kv->cell_len = __wt_cell_pack_ovfl(&kv->cell, type, rle, kv->buf.size);
kv->len = kv->cell_len + kv->buf.size;
err: __wt_scr_free(session, &tmp);
return (ret);
}
/*
* __rec_dictionary_skip_search --
* Search a dictionary skiplist.
*/
static WT_DICTIONARY *
__rec_dictionary_skip_search(WT_DICTIONARY **head, uint64_t hash)
{
WT_DICTIONARY **e;
int i;
/*
* Start at the highest skip level, then go as far as possible at each
* level before stepping down to the next.
*/
for (i = WT_SKIP_MAXDEPTH - 1, e = &head[i]; i >= 0;) {
if (*e == NULL) { /* Empty levels */
--i;
--e;
continue;
}
/*
* Return any exact matches: we don't care in what search level
* we found a match.
*/
if ((*e)->hash == hash) /* Exact match */
return (*e);
if ((*e)->hash > hash) { /* Drop down a level */
--i;
--e;
} else /* Keep going at this level */
e = &(*e)->next[i];
}
return (NULL);
}
/*
* __rec_dictionary_skip_search_stack --
* Search a dictionary skiplist, returning an insert/remove stack.
*/
static void
__rec_dictionary_skip_search_stack(
WT_DICTIONARY **head, WT_DICTIONARY ***stack, uint64_t hash)
{
WT_DICTIONARY **e;
int i;
/*
* Start at the highest skip level, then go as far as possible at each
* level before stepping down to the next.
*/
for (i = WT_SKIP_MAXDEPTH - 1, e = &head[i]; i >= 0;)
if (*e == NULL || (*e)->hash > hash)
stack[i--] = e--; /* Drop down a level */
else
e = &(*e)->next[i]; /* Keep going at this level */
}
/*
* __rec_dictionary_skip_insert --
* Insert an entry into the dictionary skip-list.
*/
static void
__rec_dictionary_skip_insert(
WT_DICTIONARY **head, WT_DICTIONARY *e, uint64_t hash)
{
WT_DICTIONARY **stack[WT_SKIP_MAXDEPTH];
u_int i;
/* Insert the new entry into the skiplist. */
__rec_dictionary_skip_search_stack(head, stack, hash);
for (i = 0; i < e->depth; ++i) {
e->next[i] = *stack[i];
*stack[i] = e;
}
}
/*
* __rec_dictionary_init --
* Allocate and initialize the dictionary.
*/
static int
__rec_dictionary_init(WT_SESSION_IMPL *session, WT_RECONCILE *r, u_int slots)
{
u_int depth, i;
/* Free any previous dictionary. */
__rec_dictionary_free(session, r);
r->dictionary_slots = slots;
WT_RET(__wt_calloc(session,
r->dictionary_slots, sizeof(WT_DICTIONARY *), &r->dictionary));
for (i = 0; i < r->dictionary_slots; ++i) {
depth = __wt_skip_choose_depth(session);
WT_RET(__wt_calloc(session, 1,
sizeof(WT_DICTIONARY) + depth * sizeof(WT_DICTIONARY *),
&r->dictionary[i]));
r->dictionary[i]->depth = depth;
}
return (0);
}
/*
* __rec_dictionary_free --
* Free the dictionary.
*/
static void
__rec_dictionary_free(WT_SESSION_IMPL *session, WT_RECONCILE *r)
{
u_int i;
if (r->dictionary == NULL)
return;
/*
* We don't correct dictionary_slots when we fail during allocation,
* but that's OK, the value is either NULL or a memory reference to
* be free'd.
*/
for (i = 0; i < r->dictionary_slots; ++i)
__wt_free(session, r->dictionary[i]);
__wt_free(session, r->dictionary);
}
/*
* __rec_dictionary_reset --
* Reset the dictionary when reconciliation restarts and when crossing a
* page boundary (a potential split).
*/
static void
__rec_dictionary_reset(WT_RECONCILE *r)
{
if (r->dictionary_slots) {
r->dictionary_next = 0;
memset(r->dictionary_head, 0, sizeof(r->dictionary_head));
}
}
/*
* __rec_dictionary_lookup --
* Check the dictionary for a matching value on this page.
*/
static int
__rec_dictionary_lookup(
WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_KV *val, WT_DICTIONARY **dpp)
{
WT_DICTIONARY *dp, *next;
uint64_t hash;
bool match;
*dpp = NULL;
/* Search the dictionary, and return any match we find. */
hash = __wt_hash_fnv64(val->buf.data, val->buf.size);
for (dp = __rec_dictionary_skip_search(r->dictionary_head, hash);
dp != NULL && dp->hash == hash; dp = dp->next[0]) {
WT_RET(__wt_cell_pack_data_match(
dp->cell, &val->cell, val->buf.data, &match));
if (match) {
WT_STAT_FAST_DATA_INCR(session, rec_dictionary);
*dpp = dp;
return (0);
}
}
/*
* We're not doing value replacement in the dictionary. We stop adding
* new entries if we run out of empty dictionary slots (but continue to
* use the existing entries). I can't think of any reason a leaf page
* value is more likely to be seen because it was seen more recently
* than some other value: if we find working sets where that's not the
* case, it shouldn't be too difficult to maintain a pointer which is
* the next dictionary slot to re-use.
*/
if (r->dictionary_next >= r->dictionary_slots)
return (0);
/*
* Set the hash value, we'll add this entry into the dictionary when we
* write it into the page's disk image buffer (because that's when we
* know where on the page it will be written).
*/
next = r->dictionary[r->dictionary_next++];
next->cell = NULL; /* Not necessary, just cautious. */
next->hash = hash;
__rec_dictionary_skip_insert(r->dictionary_head, next, hash);
*dpp = next;
return (0);
}
|